This essay will appear in our forthcoming book, “Making the Modern Laboratory.”

Published research papers are far from literal accounts of the process of scientific discovery. In contemporary scientific practice, once publishable results are obtained, the actual path taken to reach them becomes more or less irrelevant. Dead ends and false trails are omitted, and out of the messy process of raw research emerges a coherent narrative following clean, linear lines of argument.

But in the space between the hands-on, physical reality of experimental science and the structured narratives fit for printed journals, sits a special genre of scientific writing: lab notebooks. They are the closest witness to “science in the making” (short of live video recordings, which only became available at scale recently).

Historically, scientists recorded ideas and experiments in their lab notebooks with a very restricted audience in mind, sometimes just their colleagues within a research group. For this reason, though some are distinguished by a more literary style and read almost like diaries, most of these records are highly abbreviated and undecipherable to outsiders.

The origins of lab notebooks in experimental science can be traced back to the Renaissance humanist practices1 of copying excerpts from texts to create repositories of proverbs, quotations and miscellaneous facts in personal, thematically organized “commonplace” notebooks. In grammar schools, students were encouraged to develop their notetaking skills, collecting extracts from classical Latin authors. Natural philosophers such as Robert Boyle, John Aubrey, John Ray, and Robert Hooke adopted and repurposed these practices, making meticulous records of their own empirical investigations, while also keeping traditional commonplace books.

The naturalist John Ray’s Collection of English Proverbs (1670) was based on copious notebooks of proverbs extracted from printed catalogues, his own observations of “familiar discourse” and contribution sent to him by “learned and intelligent persons.” Some of the proverbs were accompanied by Ray’s own empirical observations contradicting the proverbs’ claims. For example, the proverb…

If the grass grows in Janiveer, it grows the worse sor’t all year

…is followed by Ray’s qualifier:

There is no general rule without some exception: for in the year 1677 the winter was so mild, that the pastures were very green in January, yet was there scarce ever known a more plentiful crop of hay then the summer following.2

The rise of early modern science was thus deeply influenced by humanist inquiry. Notebooks were used in both traditional and novel ways, as memory aids and as records of information to be communicated later. In Notebooks, English Virtuosi, and Early Modern Science, historian of science Richard Yeo writes:

In the seventeenth century there was a conviction that, to a large extent, copious knowledge could be reliably stored and manipulated in memory. However, during the Scientific Revolution a contrary view was emerging: namely, that the advancement of natural knowledge entailed a reconfiguration of the balance between memory and other ways of storing information. It was accepted that the empirical sciences demanded large quantities of detailed information that needed to be recorded with precision, and kept as durable records to be shared and communicated.3

Isaac Newton’s famous Waste Book,4 currently kept at Cambridge University Library, is a rare example of a physical continuity between the two cultures of notetaking: humanist and scientific.5 It starts out as a commonplace book of excerpted scriptural commentary collected by his stepfather, reverend Barnabas Smith. In 1664, on a visit home in Lincolnshire, Newton found the deceased Smith’s partially used commonplace book and began adding his own prolific and inventive notes on mathematical problems and derivations and sketches of physical experiments.

In contrast to his stepfather, Newton didn’t just collect facts and excerpts: he used them as seeds of his own theoretical explorations. This way, the Waste Book served as an extension of his mind, rather than merely a memory aid, later becoming the foundation of his magnum opus Principia Mathematica. Throughout his life, Newton kept returning to the Waste Book again and again, and the notebook that reached us is quite decrepit from such abundant use.

Newton’s later notebooks, from the 1670s to the 1690s, document his optical investigations in a series of mostly unbound notes.6 These epitomize the gap between private and public research records. Newton seemingly didn’t intend the notebooks to be a lasting record of his experiments as barely any raw data survives, except for some of his late experiments on diffraction. It appears that Newton discarded most of his raw experimental records after completing and writing up each study. In his experiments on thin films, the colors of thick plates, and diffraction, he proceeded from a hypothesis expressed as a mathematical model, to experimental design, to deducing general laws, then back to new drafts.

Newton’s “hypothesis-driven” (his term) experiments on colored circles in thin films are described in his notes under the title “Of y^e coloured circles twixt two contiguous glasses,” likely from 1671. Newton’s rings, as they are now known, are concentric, alternating bright and dark circles formed in the gap between a spherical lens and a flat glass surface, which are caused by the interference of light. Newton first wrote down a series of propositions about the properties of the colored circles, deduced by postulating the existence of hypothetical entities — light corpuscules. The first such proposition on the colored circles reads:

Prop 1. That their areas are in arithmeticall proportion, & soe thicknesse of interjected [film.] Or the spaces rays pass through twixt circle & circ[l]e are in arithm prop[ortion].

He then recorded the measurements of the diameters of the concentric circles and showed that their squares (and therefore, the areas of the circles) increase by a constant quantity — that is, they make up an arithmetic progression, just as stated in the first proposition.

In these and subsequent experiments, Newton made use of averages, a practice almost unheard of in seventeenth century experimental physics, though already in use in astronomy and navigation. The historian of science Richard S. Westfall noted how Newton elevated “quantitative science to a wholly new level of precision … He boldly transported the precision of the heavens into mundane physics … ”

In 1704, Newton published the results of these investigations in his monumental Opticks — in a highly polished form, however, omitting his workings through physical models and the relentless pursuit of precise measurements that populate his research notes. Like Galileo, Newton believed that mathematics was a source of greater certainty than natural philosophy and that natural laws were best expressed in a mathematical language. But his raw experimental data didn’t perfectly align with those laws, even though he managed to achieve remarkably high precision for his time (within 1 to 2 percent). Most intermediate steps of his research thus remain hidden from the readers of Opticks.

Newton also left extended commentary on a famous alchemical text, Introitus apertus ad occlusum regis palatium (An Open Entrance to the Closed Palace of the King).7 This book is attributed to George Starkey,8 a colonial American alchemist who moved from New England to London at age 22 and worked under the tutelage of Robert Boyle. The book is written in a veiled and heavily symbolic language featuring fiery dragons, rabid dogs, and Diana’s doves — traditional alchemical cover-names referring to specific chemical substances. This florid imagery, however, stands in stark contrast to Starkey’s private “chymical” notebooks,9 which are considered models of scholarly clarity. In their laboratories, alchemists seem to have preferred dry recipes with precise annotations, keeping the spectacle and symbolism for public presentation.

Starkey, a Harvard graduate, made extensive use of his scholastic training in documenting his alchemical experiments. Throughout his notebooks, recurring tags mark sequential steps in each experiment: Processus conjecturalis (conjectural process), Conclusio probabilis (probable conclusion), Quaere (search), Observatio (observation), Animadversio (animadversion, criticism), igne refutata (refuted by fire (!), that is, rejected by empirical testing). These are the kinds of annotations that he inherited from the educational culture of early Harvard.

At the center of Starkey’s investigations was, of course, the Philosopher’s Stone, or the Great Bezoar, the legendary and elusive alchemical substance that could turn any “base” metal like lead or copper into a precious one like gold or silver (it additionally was believed to serve as the elixir of life, granting eternal youth and immortality). In his Magnum Opus — in the original alchemical sense of the actual process of creating the Philosopher’s Stone — Starkey worked persistently to achieve higher efficiency in obtaining its supposed precursors, with recurring concerns about the cost of reagents he used. In spirit, Starkey’s work is quite close to modern pharmaceutical and industrial chemistry, and his notebooks attest to his clear-headedness and pragmatism as a practicing (al)chemist.

Two centuries later, historians of science observe at least two distinct lab notebook styles emerging: narrative and numerical. These are best illustrated by the notebooks of two pioneering English physicists, Michael Faraday and James Joule.

Michael Faraday started out as a bookbinder and then worked as a “chemical assistant” and amanuensis to Sir Humphry Davy at the Royal Institution of London.10 After Davy’s death in 1831, Faraday took over as the director of one of the most well-equipped laboratories in Europe, dedicating himself to the study of electromagnetism.

Faraday kept a detailed, narrative lab diary as a series of volumes he bound himself, spanning 42 years — 1820 to 1862.11 They contain records of about 30,000 experiments, both successful and unsuccessful.12 Entries between August 25, 1832, and March 6, 1860, are numbered from 1 to 16041. Each record includes a date, a description of the experimental setup, and results. Helpfully for historians of science, Faraday had the habit of marking with a vertical line the paragraphs in the lab diaries that made it into published papers, often unchanged. To some of the lab notes, he would add his interpretations of the results, new ideas to pursue, or a sign of excitement (as an exclamation mark). For raw ideas and speculations, he kept separate “idea books,” mostly dating from before 1830, at the beginning of his career.

Faraday’s copious notes served as compensation for his famously faulty memory. Even so, he repeated a previously completed experiment that he had apparently forgotten about more than once. (Such amusing occurrences of cryptomnesia are not uncommon among scientists).13

There are some signs that the notebook entries were made with some distance from the immediate action in the lab: Faraday’s handwriting is very neat, there are few corrections, and no chemical stains indicate contact with lab events. As the historian of science H. Otto Sibum puts it, many entries look like “the diary of a Victorian gentleman, written at the conclusion of an exciting day.”14 But Faraday also diligently recorded experimental failures and inaccurate measurements, so the notes appear to reflect the raw reality of his lab investigations.

When his lab diary eventually became too expansive, he added directories and indices to track its contents. This systematic and detailed approach to notetaking could be traced to Faraday’s past engagements as a bookbinder and businessman, as well as to his early quantitative chemical research, all of which required meticulous record keeping.

Though his own lab notes were most likely taken with some delay after execution of the experiments, Faraday encouraged his students to be expedient in notetaking. In one of the earliest experimental manuals for students, Chemical Manipulation, being Instructions to Students in Chemistry, on the Methods of Performing Experiments of Demonstration or of Research, with Accuracy and Success, he writes:

The Laboratory notebook, intended to receive the account of the results of experiments, should always be at hand, as should also pen and ink. All the results worthy of record should be entered at the time the experiments are made, whilst the things themselves are under the eye, and can be re-examined if doubt or difficulty arise. The practice of delaying to note until the end of a train of experiments or to the conclusion of the day, is a bad one, as it then becomes difficult accurately to remember that succession of events. There is a probability also that some important point which may suggest itself during writing, cannot be ascertained by reference to experiment, because of its occurrence to the mind at too late a period.

Faraday was keen on establishing notebooks as a consistent and reliable research practice but, alas, his manual didn’t reach a wide audience at the time. It took a long time before lab notebook practices were standardized.

In contrast, another English physicist, James Prescott Joule, practiced a more quantitatively-oriented, or numerical, way of keeping research notebooks. His major contributions to physics include the mechanical theory of heat and the heat effects of electricity (the SI unit of work, Joule, is named after him).15 Unlike Faraday’s, Joule’s lab notebook entries (from 1843–1858, and over 400 pages in total) seem to have been created in real time as he was taking measurements. They are very terse, containing mostly numerical records and calculations, with little commentary on experimental design. The metadata in each entry usually includes only the date, weather conditions, and a brief description of the experiment’s purpose.

Curiously, Joule had prior experience of working in the brewing business, and his biographers suspect that the accounting skills needed to run such a business and ensure quality control might have shaped his habit of numerical record-keeping in later scientific experiments. Indeed, Joule’s notebooks bear a striking resemblance to brewers’ excise books:

Before putting any water upon his malt for brewing, the brewer is to enter in an excise book or paper, the date of such entry, the quantity of malt intended to be used, and the date of the brewing … 16

That Joule’s notebooks remain mostly silent about the details of measurements suggests he kept them for himself alone. Still, his style wasn’t completely idiosyncratic but indicative of a broader methodological change unfolding in scientific practice at the time.

The nineteenth century saw a tangible improvement in the precision of scientific measurements, and a corresponding shift in judgement where dry numbers came to be trusted more than subjective, narrative descriptions of fallible, all-too-human scientists. Likewise, with the rise of “mechanical objectivity,” photographic images started displacing artistic drawings as illustrations for scientific texts. Some scientists, like the French physiologist and chronophotographer Étienne-Jules Marey, went so far as to declare images to be “the language of the phenomena themselves” and advocated for replacing language with photographs and polygraphs in scientific texts.17

The newly invented instruments like kymographs (tracing spatial position in time) and barographs (tracing atmospheric pressure readings) recorded their own data by generating paper traces as a new type of lab documentation. The lab notebooks increasingly became a place to index and annotate instrument-generated records, along with tabular data and more standardized forms of experiment annotation.

Another shift took place in lab organization, marked by a growth in both lab size and the complexity of coordinated lab operations. The scientific career of the Russian physiologist Ivan Pavlov is an illustration of how lab notetaking practices evolved in response to these changes.18

Pavlov enjoyed a long and prolific life in science. In the 1870s and 80s, he worked in the physiology of digestion and blood circulation, defending his doctoral dissertation on the nerves of the heart in 1883. Next he switched his research focus to digestive physiology, where his work on conditional reflexes (now known as Pavlovian conditioning) in dogs earned him the Nobel Prize in Physiology and Medicine in 1904.

He started his scientific career as a “workshop physiologist,” an independent investigator working in labs at the Veterinary Institute in Saint-Petersburg and later in Breslau (now Wrocław in Poland). During this time, Pavlov designed and conducted his own experiments and analyzed and wrote up his investigations. His lab notebook from this period is a large, thick volume, written in his own hand and reflecting his lab activities: experimental protocols, comments, sketches, and first drafts of research articles.

But in 1891, when Pavlov was appointed as the director of the Physiology Division of the newly established Institute of Experimental Medicine in Saint-Petersburg, he became a “factory physiologist” — the head of a large, hierarchically organized lab.19 He was now in charge of many lab assistants and students (“pracititioners”) whom he regarded as his “skilled hands,” almost like extensions of his own body, conducting and keeping records of experiments in pursuit of his own research agenda. Each practitioner was assigned a research question and a subject dog to experiment on.

As lab head, Pavlov introduced stringent notetaking protocols, with each lab notebook following the fate of a particular dog’s surgeries and treatments. The lab notebooks remained in the lab, where Pavlov could always access them. He instructed his practitioners to record procedure descriptions, notes on the dogs’ behavior, and quantitative data. Practitioners were to abstain from adding their own interpretations, leaving that task to Pavlov himself. He would communicate his analysis of experimental data in lab meetings and more casual conversations with colleagues and, eventually, in published articles.

Though Pavlov didn’t maintain his own notebook, entirely relying on his prodigious memory (and the lab notes of his students), in later years he started to delegate some of his thinking to pocket calendar books repurposed as personal notebooks. His archives contain five such notebooks, dating from 1909 to 1918 and from the late 1920s and 30s, when his research interests shifted to higher nervous activity (that is, the activity of the central nervous system). In addition to addresses, reminders, political comments, and philosophical musings, these eclectic notebooks contain notes on research happenings in his lab, ideas for new experiments, and outlines of articles:

“How will reflexes of time change under the influence of exciter substances: caffeine and so forth?”

“An interesting episode with Kal’m, that impudent and aggressive dog.”

“We consider all so-called psychic activity to be a function of the brain mass, of a defined mechanism, that is, of an object conceived spatially. But how can one place in this mechanism an activity that is conceived psychologically, that is, non-spatially [?]”

“I do not know what exactly we have done, in what way we have broken through, but it is clear to me that there now exists a union of thought, a mixing and unification of the ideas of all participants in the intellectual work [of the laboratory].”

“Some thoughts and dreams about the current war [World War I]: And the example of Germany and England in this war shows that the idea of a world government is not a true resolution of the land question, but rather a human weakness, originating, so to speak, from the inertia of human nature.”

Pavlov’s was one of the large, almost factory-style laboratories that revolutionized the social and material conditions of scientific research from the late nineteenth century onward. Similar in ambition and scale were those led by the chemist Justus von Liebig, microbiologists Robert Koch and Louis Pasteur, and immunologist Paul Ehrlich. These labs were expensive to maintain, had a purpose-designed workspace, clear division of labor, and an additional layer of lab management, which, among other things, took care of the reliable research record keeping in the lab.

In the twentieth century, scientific institutions continued scaling up, as did the pressure for standardization and reproducibility in science communications, including lab notekeeping.20 With the onset of the digital era, scientific data started moving from physical to digital formats that required large memory storage. Electronic lab notebooks (ELNs) emerged to address these changes, yet their history, in fact, goes back much farther than one might think.

One of the first published records of using computers for lab notekeeping was a 1958 paper titled “An Electronic Computer as a Research Assistant.”21 It lists several applications for using computers in lab work: mathematical calculations, copying and storage of large volumes of data, and data analysis and interpretation. These were tasks that entailed “computation volume or complexity, which otherwise would have meant thousands of man-hours for calculation.” The article also mentions “routine report preparation” by computers based on paper-based lab records. Lab notebooks thus evolved from paper notebooks to computer-assisted report generation, followed by digitized laboratory databases and, finally, ELNs themselves.

In the 1980s, a chemistry professor at Virginia Polytechnic Institute, Dr. Raymond Dessy, started advocating for the development of ELNs. In 1985, RS/1, a version of an ELN repurposed from a data analysis and statistical software system, was developed by BBN (Bolt, Beranek and Newman, of the ARPANET fame).22 Dessy later created another ELN prototype from scratch in 1994.

Interestingly, ELNs were first enthusiastically welcomed and adopted by the pharmaceutical industry, whereas their acceptance by academic communities took much longer. By 1997, several pharmaceutical and chemical companies supported a new consortium called Collaborative Electronic Notebook Systems Association (CENSA) which worked with scientific software and hardware vendors to assist with the development of ELNs that met the scientific and regulatory needs of the member organizations.

The University of Oregon introduced one of the first web-based ELNs, Virtual Notebook Environment (ViNE), in 1998. By the 2010s, a range of universities started offering institutional ELN subscriptions, but academic adoption as a whole is still patchy. This state of affairs, however, is likely to change in response to the 2024 NIH IRP Electronic Lab Notebook Policy which mandates researchers to “use only electronic resources to document new and ongoing research.”

ELNs facilitate lab record-keeping by enabling version control, timestamping, search function, hierarchical organization of information, the ability to point to external databases and to manipulate diverse data types (numerical, images, and sequences, among others). One may ask, then, why academia has been so reluctant to adopt them.

Besides the general friction towards adopting a new technology, it could be argued that handwriting is more flexible than typing and more conducive to thinking as a result: one can write how and wherever one wants and draw diagrams and sketches alongside it. ELN templates are more rigid and only allow linear text (though it can be richly formatted). The freedom of a blank sheet of paper cannot be surpassed by the already structured space of an empty digital template. Indeed, drawing and writing have historically remained as valuable, and perhaps indispensable, research techniques in their own right.

When pressing for the adoption of ELNs, the emphasis is on standardization, reproducibility, and regulatory compliance — concepts far from lab notebooks’ original use as a space for working through research questions. Perhaps paper notebooks will remain as equivalents of the waste book used by the bookkeepers of yore, while ELNs will serve as ledgers where final, more organized notes on experimental procedures will be recorded.

Ulkar Aghayeva is a science writer and a columnist at Asimov Press. She also writes about science history on her blog Measure for Measure and about music history and cognition on The Bass Line.

Cite: Aghayeva, U. “A Brief History of Lab Notebooks.” Asimov Press (2026). DOI: 10.62211/52wg-76ye

This essay will appear in our forthcoming book, “Making the Modern Laboratory.”

When protein sequencing was invented in the late 1950s, biologists found themselves faced with the enormous task of managing and analyzing long strings of apparently random amino acids. The difficulty was that most humans can only remember a string of random values around seven items long, which is 67 times shorter than the average protein sequence. Fortunately, scientists’ growing need for sequencing coincided with the development of computers, which helped to make sense of the overwhelming influx of data.

Although the term “bioinformatics” was coined in 1970 by Dutch theoretical biologists Paulien Hogeweg and Ben Hesper, the first bona fide bioinformatician was a quantum chemist, Margaret Belle (née Oakley) Dayhoff. Born in 1925 in Philadelphia to Ruth Clark, a high school math teacher, and Kenneth W. Oakley, a small business owner, Margaret was academically gifted and flourished in the sciences, a notable achievement for a woman in her day.

Dayhoff completed her PhD at Columbia University in just three years, receiving her diploma in 1948. She was one of the first to use computers to solve quantum chemistry problems, a process she described in her thesis, “Punched Card Calculation of Resonance Energies.”1 In the decade or so that followed, Dayhoff focused on raising her children, interrupting this professional hiatus only briefly to take a postdoctoral position in computational chemistry from 1957-1959 at the University of Maryland.

When she tried to resume full-time work in 1960, Dayhoff was surprised to have her funding application rejected by the NIH, which cited her “time off.” This led her to work with an acquaintance of her husband, Robert Ledley, a biophysicist who had just established the National Biomedical Research Foundation (NBRF), an organization in Silver Spring, Maryland dedicated to promoting the use of computers in biomedical research.

Ledley was an early believer in “computerizing” biology and medicine. He explained his motivation as a hunch that computers would one day be “analogous to the staff of a laboratory” and that each program would have “a function to perform just as a laboratory has people with each a job to perform: cleaning people, technicians, senior research workers, a librarian, a machinist, etc. The programmer and the protein chemist have been upgraded to the chief of the computer staff.” Though we take this idea for granted, it was unusual in Ledley’s day. Many biologists at the time rejected computers, and some were outright hostile. Dayhoff later recalled one biochemist who refused to share data or collaborate, stating “I am not a theorizer.”

Still, Ledley and Dayhoff pursued this opportunity to merge biology and computation. Combining their expertise, the two endeavored to create a computer program that could assemble full protein sequences based on results from the Edman Degradation reaction, a sequencing technique developed in 1950. The Edman Degradation reaction was limited in that it could only be used to sequence strings of about 60 amino acids before the reaction stalled out. To handle longer proteins, these shorter strings were broken up and sequenced as a set of overlapping peptide fragments. Assembling the full sequence required a researcher to compare the fragments one by one to piece them together.

Dayhoff’s program was published in 1962. Named COMPROTEIN, it freed scientists from the drudgery of protein alignment. When tested against ribonuclease, a sequence that had taken scientists months to solve manually, the program correctly arrived at the sequence in a matter of minutes.

The ability to automate Edman Degradation reactions with programs like COMPROTEIN led to a rapid leap in the number of sequenced proteins. In 1965, Dayhoff, Richard Eck, and colleagues published the Atlas of Protein Sequence and Structure, the first ever biological sequence database. Notably, the Atlas also contained the first published use of the single-letter amino acid abbreviations (Tryptophan: Y, Glycine: G, Lysine: K, etc.), still employed today.2 By the end of the decade, the collection contained about 1,000 full protein sequences.

These protein sequences allowed scientists to interrogate evolution and phylogeny like never before. While building phylogenetic trees, a branching graph that organizes creatures by how closely related they are, dates back to Charles Darwin himself, access to molecular data and computation allowed scientists to compare organisms based on sequence data rather than observation alone. The Atlas data led to the fundamental realization that sequence similarity was proportional to evolutionary relatedness: as organisms accrue mutations, they separate into distinct species. Closely related organisms will share more protein sequences, having had less time to drift apart from one another genetically.

The first sequence-based phylogenetic trees relied upon closely related sequences that were easy to align and compare visually. In 1966, Dayhoff and Eck published one of the first computer-deduced phylogenies from molecular sequences for a highly conserved iron-sulfur protein called ferredoxin. Dayhoff presented the phylogenetic tree to the public in Scientific American in 1969, writing, “Each protein sequence that is established, each evolutionary mechanism that is illuminated, each major innovation in phylogenetic history that is revealed will improve our understanding of the history of life.”

However, this strategy quickly proved insufficient for more distantly related sequences. This pushed scientists to develop multiple sequence alignment (MSA) algorithms that could compare three or more sequences and handle differences in sequence length arising from mutations like insertions or deletions. Their challenge was to create functional algorithms that were memory and time efficient for the hardware of the era. After alignment, they also needed to find a way to calculate sequence similarity to reflect how these mutations correlated to evolutionary relatedness.

In 1970, Saul Needleman and Christian Wunsch presented a dynamic programming approach that could align two sequences of different lengths. They represented these differences by using gaps to shift the frame of each sequence to accommodate insertions or deletions and maximize the number of amino acids that matched one another. However, their algorithm was computationally too slow to apply to multiple sequences.3

Dayhoff and Eck continued to work on this challenge. In 1978, they published a mathematical framework for calculating sequence similarity after the completion of alignments, a method conceptually similar to the algorithms used today. This “similarity score” was more biologically relevant than previous attempts, capturing evolutionary distance by the number of mutations between two sequences. And finally, in 1987, Da-Fei Feng and Russell F. Doolittle published the first truly practical approach to MSA. They used a “progressive sequence alignment” that initially performed a Needleman-Wunsch alignment for all sequence pairs, then extracted similarity scores for each alignment and built a phylogenetic tree based on the comparison of those scores.4

Even during the early days of bioinformatics, Dayhoff and contemporaries realized that these programs could one day be applied to genes once DNA sequencing was developed. Ironically, the man who would finally unlock ribonucleic acid sequencing, Frederick Sanger, was self-admittedly “rather reluctant” to adopt computers. For Sanger, sequence data was an enjoyable puzzle, and employing computers might “take [away] some of the pleasure” that he got from “looking through the sequences and seeing what could be made of them.”

By the 1970s, though, Sanger had accepted that computation was necessary to handle increasingly lengthy sequences. As his group at the Laboratory of Molecular Biology in Cambridge pursued the sequence of the ΦX174 virus, data was split among nine different researchers’ lab notebooks. A British-Canadian biochemist, Michael Smith, was tasked with collecting and organizing these sequences. Since most of Sanger’s group lacked computational experience, Smith recruited his brother-in-law, Duncan McCallum, who routinely used computers to process administrative data in the management division of the chemical multinational Ciba-Geigy. The pair wrote the first programs to analyze DNA sequence data. Their programs were also used to compile the first full DNA genome sequence, ΦX174, for Sanger’s seminal 1977 publication, considered the beginning of the DNA sequencing era.

When Smith returned to British Columbia, Sanger turned to LMB colleague Rodger Staden to help continue the computational work. Staden, a mathematical physicist by training, expanded and adapted McCallum and Smith’s original programs into the “Staden Package,” able to run on minicomputers rather than mainframes. Staden and colleagues continued to develop this package until 2005, and their software is still available to download.

Our DNA and protein sequence databases grew alongside our ability to manipulate genes with techniques like the polymerase chain reaction and restriction enzyme cloning. As these developments required easier ways to share sequences and manage cloning experiments, it was fortunate that the internet emerged at just this time, in 1969. The first wave of ready-to-use microcomputers followed less than a decade later, released to the public in 1977.

Biologists were among the first to use the internet to share information. Soon, several online databases were established, including the Protein Data Bank (PDB) in 1971, and the Protein Information Resource (PIR) in 1984, led by none other than Dayhoff herself at the NBRF. (Regrettably, she passed away shortly before the project’s completion.)

Also in 1984, the University of Wisconsin Genetics Computer Group published the eponymous “GCG” or “Wisconsin Package,” which initially contained 33 command-line tools for the manipulation of DNA, RNA, and protein sequences. While the Staden Package had focused on DNA assembly, the Wisconsin Package included programs for a variety of tasks, like sequence alignment, identifying protein-coding regions in DNA (called “open reading frames,” or ORFs), translating DNA to the corresponding protein sequence, and finding restriction enzyme cut sites for cloning.

Bioinformatics software continued to evolve steadily over the next 40 years, though this stage was mostly defined by improvements in speed, scale, and user experience. Commercial entities, like Geneious, SnapGene, and Benchling, began to offer programs that dramatically reduced the computer expertise needed for simple bioinformatics tasks like sequence alignment and DNA manipulation, core skills for the modern molecular biologist. However, the field has also welcomed numerous public and open-source programs, such as BLAST and BioPython.

It is hard to believe that in under 70 years we have gone from a handful of painstakingly gathered protein sequences to databases like Genbank, RefSeq, UniProt, and PDB that boast billions of unique sequences. These structural datasets, gathered by hundreds of scientists over decades, were crucial for the development of AlphaFold, an AI model for predicting 3D structure from amino acid sequences that won the 2024 Nobel Prize in Chemistry. Much as the Atlas helped to unlock a deeper understanding of phylogeny and evolution in 1965, these sequences are ushering in a new era of bioinformatics driven by machine learning.

Subscribe now

Ella Watkins-Dulaney holds a PhD in bioengineering from the California Institute of Technology. She is now a sci-comm freelancer and the Art Director for Asimov Press.

Cite: Watkins-Dulaney, E. “A Brief History of Bioinformatics Software.” Asimov Press (2026). DOI: 10.62211/72hw-48jh

To determine which is faster, natural or artificial selection, one might select examples to compare: while wolves and wild jackals branched off from a common ancestor 3.5 million years ago (natural selection), arctic foxes were domesticated within a couple of decades (artificial selection). The oldest moth species appeared in the fossil record around 200 million years ago (natural selection), while the peppered moth changed from a light-colored to a primarily dark-colored one in just 47 years (artificial selection).

But this becomes harder when we consider questions like: Is COVID well-adapted to human hosts? What animal or plant has had as much time to adapt to a new environment as COVID has to humans? Is the persistence of dominant genetic disorders best explained by a lack of purification time (i.e. that humans haven’t lived in their current environment long enough to purify out harmful alleles that might have once been adaptive) or something else?

Very quickly, two major issues arise. First, “years” seems an inaccurate unit for expressing evolutionary rates because evolution considers changes between one generation and the next. In absolute terms, dog, moth, and COVID-19 generations occur over different spans of time; yet from the perspective of evolution, all three should probably be treated similarly.

A second important consideration is that “the time it takes to evolve [X]” is a parameter with large, inherent error bars. It’s difficult to measure rates of evolutionary change accurately, and biological evolution is an inherently stochastic process, contingent on events like a cell accidentally skipping a base in DNA replication or a tortoise, swept away in a storm, happening to land on an island full of edible ferns. Without such random events, an evolutionary process that took a million years might have taken half or twice as long.1

Still, we can get a feel for how long different kinds of evolutionary events take. Let’s standardize generation time and imagine one generation happens per second. Choose an organism and start a stopwatch. Each second, replace the individual in your head with its offspring. Imagine that offspring a tad different from its parents. Do this for thirty seconds, then a minute. How does that span feel? Can you picture evolution happening?

Let’s calibrate this clock against human history:

• At one generation2 per second, the industrial revolution began ten seconds ago. There were fewer than a billion humans, the vast majority of whom were farmers.

• If we go back seven minutes further, humans have just begun implementing large-scale agriculture.

• Go back half an hour and behaviorally modern humans appeared. They possessed paints, composite tools, art, burial, complex hunting and fishing methods, and other common features of modern “primitive” civilizations.

Now let’s look at evolution itself, starting with some familiar examples and expanding to cover much of the eukaryotic tree of life. Then we’ll contrast these with examples of artificial selection to get a feel for their relative speeds.

Natural Selection

Human Evolution

• Anatomically modern humans appeared about three hours ago. These are hominids who are essentially physiologically indistinguishable from contemporary humans.

• Ancestors of Homo sapiens diverged from the other great apes (specifically, from the ancestors of modern gorillas and chimpanzees) some three and a half days ago. These were still primarily quadripedal apes.

• The dinosaurs died out and mammalian diversity began to explode about one hundred days ago.3

Human-Pathogenic Viruses

Using the time from initial infection of a host to the successful propagation to the next host as “one generation”:4

Flu

• Human influenza viruses experience approximately 19 minutes of evolution each season.

• The influenza virus responsible for the devastating 1918 epidemic that killed some 1-5% of the world’s population still exists, having evolved over some 33 hours to become today’s H1N1 seasonal flu virus.

SARS-CoV-2

Under some simplifying assumptions5 we can estimate the time it took to evolve the major branches of SARS-CoV-2:

• The Alpha strain was first detected 42 seconds after the first reported COVID-19 case. It peaked as a proportion of global cases at 72 seconds, then was gradually replaced and effectively went extinct after about 110 seconds.

• Delta appeared about 66 seconds into the pandemic. It was a parallel lineage, not related to Alpha, which became the dominant strain worldwide after 2 minutes and rendered Alpha functionally extinct by the 2.25 minute mark.

• The ancestor of all Omicron lineages probably first appeared 67 seconds into the pandemic. It was not derived from either Alpha or Delta, and wasn’t particularly successful itself, but it did give rise to both Omicron variant 21K and Omicron variant 21L (both first detected at 3 minutes). 21K displaced Delta and peaked at about 3.15 minutes and was displaced in turn by 21L another 10 seconds later.

• In total, SARS-CoV-2 has existed in humans for approximately 10 minutes.

Evolution of Major Clades

• The time to go from single-celled organisms to multi-cellular organisms is tough to estimate — single-celled organisms have generation times as fast as tens of minutes or as slow as decades. Assuming the most historically-important single-celled organisms were like today’s marine phytoplankton,6 multicellularity took 26,000 years to first evolve.

• To understand the evolutionary age of the vertebrates, we might use generation times taken from one of the oldest extant vertebrate lineages — sharks. One study of shark reproductive parameters found average generation times for 18 shark species ranging from 3 years to 42 years, which implies a metaphoric evolutionary age of 140 days to 5.5 years.

• Fast-maturing plants, reproducing annually, have been around for 15 years of evolution. Flowering grasses and other annuals have been around for 4.4 years, while typical flowering trees have evolved for only 1.5 years.7

• What about fungi? The fastest-growing eukaryotes known to science (as of 2009) is an ascomycetous yeast called Kluyveromyces marxianus, which grows a bit faster than one generation per non-metaphorical hour. This sets an upper bound of the metaphorical age of all fungi at about 375,000 years (fungi are quite ancient). Other single-celled fungi like yeasts and slime molds have probably evolved for a few years; more familiar mushroom-producing fungi may have “only” been around for 40-odd years.

Artificial Selection

So far we’ve only covered examples of natural selection. But anything that selectively influences which organisms do and don’t reproduce can force an evolutionary change. Humans do this all the time by selectively breeding animals and plants for traits we like.

How does the speed of artificial selection compare to that of natural selection? We can start with a classic textbook example of evolution that bridges the two: the evolution of the peppered moth under unintentional selection pressures caused by the industrial revolution.

Before the industrial revolution, peppered moths living in Manchester, England were light-colored moths with a dark, splotchy pattern perfect for blending into forests. During the industrial revolution, coal soot and sulfur dioxide emissions from the city blackened the trunks of those trees, making peppered moths stand out. Under intense bird predation, the peppered moth switched from dark splotches on a light background to a totally dark coloration that allowed them to hide on soot-blackened trees. This rapid evolution took about 47 seconds.

Agriculture

The longest-running experiment in artificial selection is the domestication of crop plants.8 Crops are a particularly rich example of evolution because domesticated plants tend to pick up the same cluster of traits — non-shattering seed hulls, enlarged seed or fruit, shorter height, synchronized flowering — so the collection of all agricultural crops can be viewed as replicates in a single giant experiment in artificial selection.

To get a sense of how crop domestication works, let’s look at the recent evolutionary history of a crop for which we have particularly complete evidence: barley. Barley was among the most utilized crops in early agricultural old-world societies, grown across nearly all of Eurasia and North Africa. Critically, it is one for which we have an abundance of archaeological evidence. It seems to have been an unusually slow crop to cultivate, so the evolutionary times given here for barley are more representative of the relative pace of domestication than of its absolute time.

A brief and vastly simplified9 history of the domestication of barley,10 in our metaphorical time, would be that:

• The oldest evidence of cultivation of phenotypically wild barley appeared approximately 20,000 years ago, and serves as our zero second mark.

• The first recognizably cultivated barley appears at around 2.7 hours, after almost 10,000 generations. These seeds look a lot like wild barley, just with bigger grains.

• The first recognizably cultivated barley slowly spreads throughout the fertile crescent over the next 50 minutes.

• Meanwhile, barley gains a “non-shattering hull” just 7 minutes after first displaying a larger grain. This variety has a more solid seed husk, less prone to spontaneously breaking open and prematurely scattering its seeds before harvest.

• The first 6-row barley appears just 10 minutes after the first non-shattering hulls. Wild barley bears two seeds at each repeating subunit; this variety bears six instead, greatly increasing yield.11 (We are now about 3 hours into domestication).

• A second, independent domestication event occurs in Asia around 3 hours and 12 minutes in.

• The first hulless barley appears in Iran just a few minutes later, at the 3 hour and 15 minute mark. “Hulless” is something of a misnomer as this barley still has a hull. However, it is much, much easier to remove. Both hulless and hulled barley will continue to be used for different purposes up through the present day.

• A “dense 6-row” barley appears at about 4 hours and seven minutes, bearing more seeds per ear.

• At about 4 hours and 20 minutes, Hulless barley begins to appear much more broadly across Europe.

• Barley cultivation continues with shifting balances and mixes of existing varieties until close to the present day, around five and a half hours in.

• In the last minute of barley cultivation, humans have modified barley to grow a shorter stalk (less prone to breaking under heavy seed load) with greater seed-to-vegetation weight ratio, and increased resistance to several important diseases.

Barley domestication teaches us that the process occurs in many steps, spread out over thousands of generations; both wild and cultivated varieties were used virtually unmodified for thousands of generations before undergoing multiple major changes over just a few hundred. New traits take almost as long to spread and take over cultivated populations as they take to arise in the first place. Individual domestication events seem to require on the order of thousands of generations, not hundreds and rarely tens of thousands.

But the real value of agriculture as a case study in evolution is the number of replicates it provides. Ideally, to get the best feel for the timescale of plant evolution under artificial selection, we would re-visualize domestication for each of the thousands of species humans have domesticated as crops.

Luckily for us, Meyer, DuVal, and Jensen have collected evidence of domestication for 203 food crops from around the world into a handy spreadsheet. Limiting ourselves to crops with high-quality evidence, we can see that:

• The median crop becomes domesticated after about 24 minutes of (metaphorical-time) exploitation by humans, while mean domestication time is a bit longer, at 33 minutes.

• The fastest-domesticated crop is the common bean, for which the archaeological evidence for first exploitation and first domestication are indistinguishable (though this may say more about the quality of our evidence than the friendliness of the common bean).

• The most recalcitrant crop to eventually yield to domestication is, ironically enough, barley, at its rather lengthy 2.7 hours between first exploitation and eventual domestication.

• Annuals — crops with a one-year lifecycle — take somewhat longer to domesticate, with a median domestication time of 33 minutes.

• Perennials — crops that live more than one year — domesticate more quickly, with a median domestication time of just over 9 minutes.

We can do a lot with this data. For example, if we split up domesticated crops by their gross morphology,12 it appears that trees (and possibly cactuses and succulents) may take fewer generations to domesticate than grasses, vines, and other herbaceous crops.13

Specifically, on average:

• Trees take about 10.5 minutes to domesticate.

• Cactuses and succulents take 8 minutes to domesticate (but N=2).

• Grasses take a bit over an hour to domesticate.

• Vines take about 40 minutes to domesticate, while herbaceous plants take about 35 minutes to domesticate.

Domesticated Mammals

The evolution of the dog from wolf-like ancestors bridges a gap between crop domestication and pure artificial selection. It began as an unintentional and gradual process, but advances in selective breeding (and in the scientific understanding of evolution itself) led to a recent acceleration in dog “speciation.” Its history is also complex, speculative, and contentious, but we have enough combined genetic and archaeological evidence to broadly estimate that:

• The ancestors of dogs split from modern wolves anywhere from 2.5 hours ago to 50 minutes ago by wolf generations,14 or 2-5.5 hours ago by dog generations.15

• The earliest recognizably domesticated dog remains date to approximately 2 hours ago.

• Most modern dog breeds — especially hunting breeds like greyhounds, schnauzers, terriers, beagles, and bloodhounds — are thought to have split from a common ancestor only 1-2 minutes ago.

We also have much more intentional, directed examples of mammal domestication. In 1952, Soviet geneticist Dmitry Belyayev and his graduate student Lyudmila Trut began a multi-decade experiment to attempt to replicate the domestication of dogs. Instead of wolves, Belyayev chose silver foxes as the base species for domestication, starting with foxes bought from commercial fur farms.

Generation after generation, Belyayev and Trut separated out the young foxes who responded most positively to human handlers, as well as the foxes who responded most aggressively. By breeding the most sociable foxes together, the scientists obtained increasingly friendly, loyal, and loving individuals; conversely, the antisocial-selected line quickly evolved heightened aggression responses.

The domestic fox experiment continues to this day. Some key observations are that while the researchers initially categorized their foxes into three “classes” based on their level of tolerance to approach by humans, foxes soon appeared that actively sought human attention, requiring a fourth “elite” category.

• The first four of these “elite” foxes were born after 6 seconds.

• After 30 seconds, fully half of the foxes were born elite.

• After 42 seconds, almost three quarters of friendly foxes were elite.

• As early as 20 seconds into the experiment, some of the domesticated foxes started to be born with dog-like physiological traits, including floppy ears, piebald coat patterns, shortened or upward-curled tails, and shorter legs.16

• Domestication was also associated with a dramatic decrease in corticosteroid production, which the experimenters hypothesize is responsible for the foxes lack of fear response towards humans. By 30 seconds, the domestic foxes had half the baseline cortisol blood concentration of un-selected controls (a ratio which hasn’t changed much since).

The same experiment has been repeated in other species:

• River otters showed early signs of domestication after one or two generations, but breeding them proved too difficult to sustain and their breeding was cut short.

• Wild Norwegian rats (“laboratory rats”) adapted so quickly over their first 35 seconds of breeding that they maxed out the researchers’ initial behavioral scale and forced them to define a second one.17

• American minks began showing active interest in human handling after 18 seconds of selection.

Evolution in the Lab

The most direct and well-quantified examples of human-induced evolution come from laboratory experiments on single-celled bacteria and protists. Fast-growing, single-celled organisms make amazing test platforms for evolution, due to their ease of genetic manipulation and characterization and incredibly fast reproduction. One test tube, incubated overnight, can easily hold tens of billions of E. coli, representing millions of de novo genomic mutations. On the other hand, bacteria reproduce asexually, without sexual recombination, which makes even huge populations of bacteria less genetically diverse than, say, a small tribe of monkeys.18

With those features in mind, let’s look at how scientists have measured — and forced — evolution in microorganisms. In a famous long-term evolution experiment in the lab of Richard Lenski, twelve parallel lineages of laboratory-bred E. coli have been successively grown in simple, glucose-limiting media for over 35 real-world years and 80,000 generations — a bit over 22 hours of metaphorical time. A few highlights from that experiment, in metaphorical timescales are that:

• The E. coli evolved very quickly at first, then slowed considerably (but never stopped). In the first half-hour of evolution, each of the 12 strains increased their fitness in their new environment by about 40 percent.

• One of the first real surprises of the experiment was the sudden appearance of extreme mutation rates in 6 out of 12 strains. These “hypermutator” strains are estimated to be about 1 percent less fit than their more stable ancestors when they first appear (because they are much more likely to replicate genomes with harmful mutations), but were able to evolve more quickly and therefore were able to outcompete their ancestors in the long term.

• The hypermutators took over their respective populations at about 40 minutes, 45 minutes, 1.4 hours, 2.4 hours, 7.5 hours, and 9.7 hours into the experiment.

• One — and only one — of the twelve E. coli strains evolved the ability to eat citrate from their growth media after about 31,000 generations, or 8.6 (metaphorical) hours of adaptation to lab conditions. This trait was dependent on a set of background mutations on acetate metabolism, which arose somewhere around 7 (metaphorical) hours into the experiment.19

• These first citrate-eaters were pretty messy eaters that spewed carbon-rich metabolites20 as a by-product of citrate metabolism. Within about 15 minutes (900 generations) of their appearance, another non-citrate-metabolizing lineage evolved a greatly-enhanced ability to consume those by-products.

• The citrate-metabolizing and by-product-metabolizing strains co-existed for about 2.7 hours of metaphorical time, until the citrate-eating E. coli got good enough at consuming their own waste products to out-compete the secondary strains and drive them to extinction.

• The E. coli evolved to grow steadily larger over time, roughly doubling in cell volume in the first metaphorical half hour of the experiment. They continued to grow in size over the rest of the experiment, but more slowly, expanding to 2.5x their ancestors’ original volume after about 2.7 hours and 3.5x the ancestral volume after 14 hours.

Other examples of long-term evolution experiments abound, too. In 1998, for example, Boraas, Seale, and Boxhorn forced (or at least strongly encouraged) the single-celled green algae Vulgaris chlorella to evolve to grow in multi-celled clumps by exposing them to heavy predation by a flagellate predator. Under this intense selection for larger, harder-to-eat forms, V. chlorella evolved into eight-celled clumps in 10-20 (metaphorical) seconds.

And in a more recent example, four scientists at the University of Minnesota used an even simpler, more ingenious method to select for multicellularity — they propagated media containing Saccharomyces cerevisiae (baker’s yeast) from the bottoms of test tubes over many generations.21 This artificially selected for yeast growing in many-celled “snowflakes” that could sink quickly.22 In our metaphorical time scale, out of ten parallel replicates of evolved yeast, one population evolved noticeable numbers of such snowflake yeast in about 1.5 minutes, half evolved them by 2.5 minutes, and all ten were “multicellular” in 6-7 minutes. The same experiment has been replicated a number of times on a number of different single-celled eukaryotes, with different rates of evolution:

• 70 seconds for Kluyveromyces lactis (another budding yeast that diverged from S. cerevisiae about 100 million non-metaphorical years ago).

• 2.5 minutes for Schizosaccharomyces pombe (a fission yeast).

• Between 1.5 and 7 minutes for Sphaeroforma arctica (a single-celled protist relatively closely-related to animals).

• In slightly over 5 minutes, one out of twelve replicates of Chlamydomonas rheinhardtii (a single-celled green algae closely-related to plants) evolved multicellular clumping.

• Trichormus variabilis (a cyanobacteria) increased in size (through a combination of larger cell size and clumping) by about 3-fold over 2.5 minutes and >30-fold over 6 minutes.

Feeling Evolutionary Time

We’ve argued in previous articles that metaphor helps with quantitative understanding of the sizes and speeds of molecular biology. But metaphor really becomes useful when considering evolutionary time. Relating the number of generations required for evolutionary events to everyday time scales helps us see patterns in what would otherwise seem like totally incomparable evolutionary facts.

Our examples have shown us that intense artificial selection can drive changes far faster than natural selection. “Natural” domestication of both plants and mammals seems to act on the metaphorical scale of hours (with some exceptions, like trees and cacti!), but intense selective breeding can create dozens of new dog breeds in minutes, or even recapitulate dog-like domestication in foxes in tens of seconds. The intensity of selection is just as, or more, important than the size of the difference produced, when it comes to how long it takes to evolve a change.

We have also seen that bacteria evolve relatively slowly. Half a dozen different single-celled eukaryotic species were capable of evolving primitive multicellularity in a matter of metaphorical minutes; E. coli, in contrast, adapted to experimental conditions more on the order of tens of minutes or hours—even though they could draw on absolutely massive population sizes! Speculatively, this may be because eukaryotes use sex (and other powerful DNA-recombination tricks) to generate lots of genetic diversity even from relatively small populations.

And we’ve learned that evolutionary speeds are trends, not exact rates. Several examples (plant domestication, hypermutators in E. coli, multicellularity in single-celled eukaryotes) reveal that roughly the same change occurs multiple times in different populations or species. When this happens, the speeds of the individual examples vary, but only about 3- to 5-fold around the median speed.

We’ve seen that the speed of evolution is complicated. It depends on the size of the evolving population, the strength and stringency of selection, and how the species reproduces—and how quickly. The “one generation per second” quantitative metaphor can’t eliminate all of that complexity, but it at least puts species with radically different life cycles on an even footing, which helps clarify where and when other factors like population size or replication fidelity come into play.

More importantly, using a single scale of “one second per generation” lets us place concrete evolutionary examples onto a more intuitive reference frame. The next time you learn that a weird animal lived so many millions of years ago, or a bacteria evolved antibiotic resistance in so many days, you can place it on that same common scale — simply divide the total time by the length of one generation, convert that number of seconds into something convenient like minutes or days, and you’ll be able to compare it directly to any of the examples given in this brief introduction.

Hopefully, this technique has given you a better sense of how evolution works.

Subscribe now

Samuel Clamons is a bioinformatics scientist at Illumina, Inc. with a PhD in Bioengineering and training in applied mathematics and computer science. Outside of his day job, he writes science fiction and researches theoretical questions in biology at Asimov Press.

Many thanks to Dr. Scott Biering and Dr. Alistair Russell of UCSD for their consultations on viral reproduction and evolution. Header image by Ella Watkins-Dulaney. Image Credits: M. Garde and Tkgd2007.

Cite: Clamons, S. “Metaphors for Biology: Evolution.” Asimov Press (2026). DOI: 10.62211/29je-83ut

Asimov Press is going on hiatus. While a few more articles will appear over the next month — and our stunning hardcover book, Making the Modern Laboratory, will come out this summer — operations will pause in April. All articles will remain freely accessible online.

As we prepare to pause, I want to celebrate what we’ve accomplished. In June 2023, Asimov Press was just a vague idea. (I remember posting the announcement from my phone in the Amsterdam metro, feeling a deep sense of fear about what we were starting.) A few months later, I brought on Xander Balwit. The two of us were its only editors and only full-time staff. We launched with about 7,000 subscribers, ported over from my personal blog, Codon. Our first article appeared in January 2024.

While I had attended journalism school and thought I knew about publishing, I soon discovered that literally everything — from managing people, to recruiting and paying writers, to fact-checking stories, printing books, and much else — doesn’t play out as it’s taught in school. By trial and error, Xander and I had to learn how to develop an eye for writers with exceptional promise, separate good pitches from vague ones, figure out which parts of a draft were most compelling, and make sure stories were told deeply enough to satisfy scientists, while remaining accessible to all “ambitious” readers.

Devon Balwit served as Copy Editor, helping us to convey these complex ideas in readable prose. In 2025, we hired Ella Watkins-Dulaney as our Art Director, and she promptly improved our header images and style. In the last several months, she has stepped into other roles as well, helping fact-check and compile the latest book. Ulkar Aghayeva composed and recorded the music for our podcasts, and also wrote several excellent articles for our pages. We’re also immensely grateful to the dozens of fact-checkers, artists, proofreaders, and writers who have contributed to maintaining a high bar of excellence at Asimov Press.

We’ve come a long way since our inception. Today, we have about 42,000 subscribers between Substack and our website. The press has published 149 original articles, reaching about half a million readers each month.

These articles have ranged from short to long, and from technical reports to history to speculative fiction. Some personal favorites have been The Penicillin Myth, about how Alexander Fleming’s discovery story may have been partly fabricated, How Nature Became a ‘Prestige’ Journal, What Makes an Experiment Beautiful?, and The Making of a Gene Circuit.

While many of our pieces were widely discussed, a handful made an even more tangible impact in the world. Our two recent essays on clinical trials — Clinic-in-the-Loop and AI Won’t Automatically Accelerate Clinical Trials — have been widely discussed in policy circles, as has the essay on China’s Clinical Trial Boom. The essay Mitochondria Are Alive became a bit of a rallying cry for funders considering new tools to engineer these organelles. An essay on artemisinin’s discovery was cited in discussions that led to the Research Revival Fund, a new nonprofit seeking to restore “neglected, illegible, or prematurely dismissed research.”

The Press has published and sold out of two anthologies. When we first tried our hand at printing books, we assumed we would lose a lot of money as printing is expensive, and around 96 percent of all books sell fewer than 1,000 copies. But in both cases, we sold thousands of copies, recouped our costs, and donated the profit to charities. I’m especially proud of our second book, the first commercially available book published in both print and DNA. It was mentioned in The New Yorker and featured in WIRED magazine. Thanks again to the companies that made it possible.

Our decision to pause is not related to funding. Asimov has supported us for the last two years, and we’ve received generous grants from Astera Institute and Stripe. But new projects have called away Xander and I, making it a good time to recalibrate.

For me, the greatest part of writing and editing is drawing closer to ideas and the people who generate them. Essays are a forcing function toward both. Great writers often find an exciting idea and decide to pursue it, only to confront a gap that nothing on the internet can fill — a question too deep or too specific to answer alone. And so the work begins: emailing and calling people, visiting where they work, hosting dinners, and sometimes flying around the world to understand it. The search for great ideas, and a deep understanding of them, is why I’ll keep writing, and helping others do so, for the rest of my life.

When Asimov Press reemerges, I know that those who lead it will be just as curious about the history and future of science as we have been. Thank you for taking this journey with us.

Until next time,

— Niko

In On Exactitude in Science, the writer Jorge Luis Borges imagines an empire so devoted to cartography that its mapmakers draw a map as large and detailed as the empire itself. “In the Deserts of the West, still today, there are Tattered Ruins of that Map,” Borges writes, “inhabited by Animals and Beggars.” Borges’s map is a parable for knowledge, and one of its lessons is that too much detail can quickly become impractical — a map at that scale would be perfect but useless.

But with today’s AI systems, one might wonder if such a map is so absurd after all. Computers and the Internet have already helped us to digitize much of human knowledge, and AI enables us to scan it quickly and easily. For instance, large language models are trained on trillions of words spanning much of recorded human knowledge. In biology, systems like AlphaFold learn from large databases to predict a protein’s folded structure from its amino acid sequence.

This means that, in some domains, something resembling Borges’s life-sized map has become extremely useful. And given the rate of progress on this front, it may seem like advancing science now simply requires building ever larger and more navigable versions of such AI systems, effectively mapping every field.

A lack of practicality, however, was never the sole flaw of Borges’s map. The deeper problem is that adding detail only gives you more of the same kind of information — more roads, more mountains, more villages — when what you might need is a completely different schematic.

Consider the map of the London Underground. Until 1933, the map plotted stations at geographically accurate locations in London. But this made central London, where most stations clustered, an unreadable tangle, while the outer suburbs, devoid of relevant data, took up most of the space. The draughtsman Harry Beck solved this inefficiency by abandoning geographic accuracy and instead redrawing the network as a circuit diagram of colored lines and evenly spaced stations.

A scientific paradigm can also be thought of as a kind of map, but unlike Beck, scientists do not usually know in advance what their maps will be used for. Instead, new paradigms are driven by the desire to explain complex phenomena with a simple and unified set of principles. Such principles tend to have knock-on implications that stretch far beyond the phenomena that inspired them.

For instance, by the mid-nineteenth century, electricity and magnetism were described by a patchwork of separately discovered laws, each explaining a different phenomenon. The physicist James Clerk Maxwell simplified the field by replacing this patchwork with four short equations. But they also implied the existence of electromagnetic waves that could travel through space, including low-frequency waves no one had yet detected. These waves eventually became the basis for radio.

Current AI, by contrast, is not set up to do this. It excels at prediction within existing frameworks, but paradigm shifts require replacing these with simpler alternatives whose implications haven’t yet been explored. A computational system trained on electromagnetic measurements may have predicted these results perfectly, but would never have found radio.

Seen in that light, even as AI becomes more central to scientific work, we risk falling into what one might call hypernormal science, where we get ever better at prediction within current models, alongside a weakening capacity to ask completely new categories of questions. Much like Borges’s empire of cartographers, we risk confusing more detail for a true understanding of the territory.

To avoid this kind of myopia, we must deliberately build AI that helps us devise new conceptual vocabularies. In other words, we must build visionary machines rather than merely predictive ones.

How Paradigms Work

Before exploring how to build such visionary AI, it helps to look more closely at how paradigm shifts in science actually happen. Science usually progresses by adding facts within an existing paradigm, which functions like a rulebook for a field. But over time, evidence accumulates that an existing paradigm cannot explain, requiring a new one.

One might expect that a new paradigm would immediately replace the old one as it better explains the facts. Instead, they tend to gain preeminence only after becoming useful for new applications.

One example is the development of special relativity. In the late nineteenth century, physicists could describe light with wave equations. Because every familiar wave (like sound or water) seemed to need a material carrier, the scientific consensus was that light must also travel through an invisible medium, dubbed the luminiferous ether. The academic establishment was profoundly attached to this concept; Lord Kelvin, the elder statesman of British physics, even declared the ether was the only thing in physics we could be absolutely certain existed.

Albert A. Michelson and Edward W. Morley reasoned that if the ether existed, the Earth’s motion would create an “ether wind,” making light traveling along that wind move at a slightly different effective speed than light traveling across it. Michelson and Morley sent light along two perpendicular paths and expected that because of this speed difference, one beam would come back slightly later than the other. But their experiment revealed no detectable difference.

This result didn’t immediately convince the academic community to abandon the concept of ether. Many physicists (including Michelson) instead adopted an interim position, that the ether’s effects must be hidden in some way. Most prominently, Hendrik Lorentz proposed that the ether existed, but that objects moving through it would shorten in the direction of travel, canceling the expected signal.

An alternative paradigm was ultimately offered by Albert Einstein, then a 26-year-old patent clerk in Bern, Switzerland. His theory of special relativity posited two principles: that the laws of physics are the same in every uniformly moving frame, and the speed of light in a vacuum is the same for all such observers. Through these principles, Einstein was attempting to introduce “a simple and consistent theory,” by which “the introduction of a ‘luminiferous ether’ will prove to be superfluous.”

Initially, Lorentz’s and Einstein’s theories both explained the known experimental data similarly well. But Einstein’s theory proved far more fruitful over time. If light’s speed were genuinely constant, then space and time could not be absolute. This eventually led to the demonstration that mass and energy had to be equivalent, as per Einstein’s famous equation, E=mc², which now underpins technologies from nuclear power to medical imaging.

A paradigm can take hold even if incomplete, provided its core idea is sufficiently useful. For instance, Charles Darwin’s theory of natural selection offered a single principle that could explain the diversity of living species even though it still lacked an explanation for how traits actually passed from parent to offspring. In the late 1860s, he posited the missing mechanism, the erroneous notion of “pangenesis.” In it, he speculated that every cell in the body sheds tiny particles called “gemmules” that collect in the reproductive organs and transmit traits to offspring. Despite this error, Darwin’s core vision survived, and spread amongst biologists, before genetics supplied the necessary, physical mechanisms.

Thus, Einstein and Darwin were both able to generate simple and elegant theories that could make predictions beyond current evidence, even when some details were missing or wrong. In both cases, their decisive advantage was not technical skill within the paradigm, but rather a willingness to step outside it: Einstein benefited from being an outsider to the academic establishment, as this freed him from attachment to the idea of the ether,1 while Darwin appropriated concepts from Charles Lyell’s geology and resource competition from Thomas Malthus’s economics.

Current AI Training Risks Hypernormal Science

If paradigm shifts require stepping outside the prevailing logic, we should ask whether current AI is set up to do this.

Consider an early attempt. In the late 1970s, computer scientist Douglas Lenat built the Automated Mathematician, a program designed to discover not just new facts but entire mathematical concepts. It would start with simple ideas, combine and vary them, and keep the results that seemed interesting. It appeared to work, reportedly rediscovering prime numbers and Goldbach’s conjecture. But its creativity turned out to be limited, because many of the concepts it “discovered” were already implicit in the way mathematics was written inside the program.

While today’s AI has vastly more power than the Automated Mathematician, a similar constraint applies. Most machine-learning systems are trained by minimizing prediction error against a dataset whose inputs and labels are defined in advance. This makes them very good at predicting current data, but locks them into the conceptual vocabulary of the data they learn from.

Consider medicine before germ theory. In the mid-nineteenth century, doctors thought that illness was caused by noxious air, and kept meticulous records accordingly. The physician William Farr mapped cholera deaths across London and found they correlated strongly with low elevation, which he thought was because noxious vapors accumulated in low-lying areas. He was actually picking up a real signal: low-lying districts were closer to the contaminated Thames River. But because his data was organized around air quality, he could not find the true cause.

An AI trained on Farr’s records could have found even subtler correlations, and would have been genuinely useful for predicting which neighborhoods would be hit hardest in the next outbreak. But it would not be able to derive the concept of a waterborne microorganism, as this was not a variable anyone had yet recorded. It took researchers like Louis Pasteur and Robert Koch, working with microscopes and culture dishes rather than statistical registers, to establish germ theory and open the door to antibiotics, antiseptic surgery, and modern public health.

In 2023, Google DeepMind used a graph neural network called GNoME to predict the stability of crystal structures at an enormous scale, discovering 2.2 million new materials. But the vast majority were substitutions within already-known structure types, for instance swapping one element for a neighboring one on the periodic table. The system optimized impressively for thermodynamic stability relative to known structures, but could not venture far from these.

Some newer AI models, known as foundation models, sidestep this problem partly by learning directly from raw data rather than human-curated labels. For instance, the protein generation model ESM3 was trained on protein sequences and structures, and managed to design a novel fluorescent protein highly different from those found in nature.

While impressive, generating novel proteins in this way is analogous to filling in unexplored spaces on a map, but not to creating new maps entirely. ESM3 does not, for instance, ask whether an amino acid sequence is the right level of description, or whether some other organizing principle might unify protein behavior with phenomena outside biology.

When researchers trained a foundation model on ten million simulated solar systems, it learned to predict planetary orbits with high accuracy, but acquired no representation of gravity. Instead, it had assembled a patchwork of statistical regularities that happened to produce correct trajectories. Of course, it is possible that such models already contain, buried in their weights, patterns that point beyond current theories; but extracting them would require deliberate study in its own right.

This grows in importance as scientists use AI more often in their work. On the far end of this spectrum, researchers have begun to develop “AI scientists,” end-to-end pipelines that aim to execute the whole scientific workflow by chaining together literature search, idea generation, code writing, experiment execution, and paper drafting. If these systems are successful, they might be able to work much faster than human scientists, and thereby may constitute the majority of scientific work that gets done.

But these systems have to evaluate the quality of the new ideas they generate, and it is hard to do so without reference to the existing paradigm. When the system proposes a new hypothesis or experiment, the only available proxy for what constitutes a good idea is consistency with existing science. This often involves passing simulated peer review, aligning with established results, and looking like a plausible contribution to the field. A genuinely novel reframing would likely score poorly on all of these measures, for the same reasons that paradigm-shifting work has always faced resistance from reviewers trained in the paradigm it aims to replace.

Simply put, optimizing performance on extant benchmarks makes it difficult for alternatives to emerge. And this, in turn, risks hypernormal science.

Consider a thought experiment in genetics. For centuries, breeders kept detailed records of which animals were mated, what their offspring looked like, and which traits appeared in which family lines. An AI trained on such data could learn to predict what size, color, or yield the offspring of any given two parents might have, which would surely prove immensely useful.

But such a predictor would never discover the gene as a discrete unit of inheritance, nor DNA as its carrier. Without that insight, while farmers may have been thrilled with their better breeding predictions, we would not have had the ability to create genetically modified organisms or targeted gene therapies. Indeed, a prolonged state of hypernormal science is dangerous precisely because its practical implications would not be immediately obvious.

There are early indications that this is already happening. A recent study of 41 million research papers found that scientists using AI publish more and receive more citations, but collectively, AI-augmented research covers around five percent less topical ground. This appears to be because AI gravitates toward problems rich in existing data, where the current paradigm is most established. The result is that AI induces authors to converge on known solutions, rather than search for new ones.

The Bitter Lesson for Science

In a 2019 essay entitled The Bitter Lesson, the computer scientist Richard Sutton observed that methods which try to build in human knowledge consistently lose out, over time, to methods that simply scale search and learning. “We want AI agents that can discover like we can,” he writes, “not which contain what we have discovered.” This applies naturally to paradigm shifts, which by definition move beyond existing knowledge. It might seem, then, that the path to paradigm-shifting AI is to get out of the way and let computation run.

Indeed, open-ended systems like AlphaZero (which learned chess entirely through self-play) are capable of creating both powerful and highly original results. Starting from nothing but the rules, it played millions of games against itself, generating strategies and discarding the ones that lost. Within a day, it was playing at a superhuman level, and did so while playing extremely original moves. The grandmaster Peter Heine Nielsen compared it to a superior species landing on Earth and showing us how to play chess.

Science might seem like a more complex version of the same problem, requiring more powerful computation to solve. In some senses, though, the opposite is true. Chess has simple rules, but a winning strategy is extraordinarily complex. In science, the winning paradigms can be extraordinarily simple; special relativity only has two postulates. But without the benefit of hindsight, there is no clear way to pick a winning paradigm.

To design AI for disruptive science, we would need to understand what “rules” make one paradigm better than another, and build systems that optimize for these. This turns out to be a harder problem than scaling compute. The answer cannot simply be experimental success, since experiments are slow and do not always reliably distinguish between paradigms (as was the case with Lorentz and Einstein). And there are other plausible candidates, but none yet offer a sufficient formulation.

One rule might be that good paradigms are simple. There are early attempts to make AI optimize for this. For example, in physics, symbolic regression systems such as AI Feynman try to discover the simplest equation that explains the data, instead of doing a black-box mapping. On benchmarks drawn from the Feynman Lectures, the method discovered all 100 test equations, while prior software found only 71. One can even formalize a drive towards simple theories using the Minimum Description Length principle, which effectively penalizes unnecessary complexity.2

But these systems currently work on clean data with pre-selected variables, and only search one dataset at a time. So far, they have been tested on rediscovering equations we already know, but haven’t yet demonstrated that they can find new ones.

Another rule may be that good paradigms draw effective analogies. Feynman devoted a chapter of his Lectures to the observation that heat flow, fluid flow, diffusion, and electrostatics all share the same equations, treating this as a deep fact about nature. This makes intuitive sense: an idea that already works in two domains has a better chance of working in a third.

The most obvious kind of analogy is across disciplines, as when Darwin borrowed the logic of competitive scarcity from economics and applied it to biology. In principle, AI could search for these connections at a scale no individual researcher could match, trawling across fields for ideas that seem structurally similar. Early systems have been built that find functional analogies across large databases of patents and product descriptions.

But some analogies are not between written theories, but rather between an idea and a sensory intuition. At sixteen, Einstein imagined riding alongside a beam of light and asked what he would see. Maxwell’s equations modelled light as a wave. But if this were true, if one travelled at the speed of light, the wave would appear frozen, hanging motionless in space. To Einstein, this picture felt physically wrong; this visceral discomfort ultimately inspired him to develop special relativity.3

Some research programs are trying to ground AI reasoning in physical experience. There are early attempts at multimodal architectures that can jointly process vision, language, and action. Self-driving laboratories couple AI to robotic instruments that manipulate real materials, which could in principle ground abstract reasoning in physical feedback (though most operate within a single experimental domain). But these are early efforts, and the gap between a robotic arm pipetting reagents and multisensory human experience is vast.

In the meantime, the fastest path to getting effective analogies may be leveraging both human and AI abilities together. Humans have breadth across modes — we see, hear, touch, move through space, and read, all at once — which grounds our capacity for structural analogy. AI has depth within modes, having processed far more than any person ever could. If AI can help researchers learn faster across disciplinary boundaries, perhaps by disciplined use of LLMs, that alone could accelerate discovery.

The deeper problem is that we do not have a good formal understanding of how paradigm shifts actually happen. Simplicity and analogy alone do not appear to be a complete description. J.J. Thomson’s “plum pudding” model imagined atoms as spheres of positive charge with electrons scattered through them. It was simple, matched what was known at the time, and was comically analogical, but soon proved completely wrong.

For now, then, the Bitter Lesson for Science may be that scientific acceleration will not happen by default until we understand science itself better.4 But if we can identify the conditions that produce paradigm shifts, we can start to engineer them.

A Paean for Metascience

Understanding and codifying science is no small task. Since Bacon’s Novum Organum in 1620, we have had a reasonable account of how science works in theory — observe, form hypotheses, test, and revise. But paradigm shifts seem to depend just as much on conditions as on method: who gets to do science, what they are rewarded for, and how freely ideas can cross between fields, for instance.

We can learn some things from history. Bell Labs, Xerox PARC, and the early Laboratory of Molecular Biology at Cambridge all produced extraordinary concentrations of paradigm-shifting work, mostly because they were small groups with enough institutional protection to pursue ideas that looked unproductive by conventional measures. There are clear parallels, in fact, to the kind of independent self-play that made AlphaZero successful. But there are limited numbers of historical paradigm shifts to study, and we have barely begun to explore the design space.

Here, AI itself may be able to help. We have never been able to run controlled experiments on scientific institutions; it is impossible to create labs that differ in only one respect and compare the results. But we could run AI agents in parallel populations under different research conditions, and analyze the results in detail. In this sense, AI scientists may give metascience its first model organism.

For instance, one could test how group structure shapes discovery: do small, isolated teams produce more conceptual reorganization than large, well-connected ones? Do flat hierarchies outperform rigid ones? One could run AI agent populations that vary these factors independently and measure the results — something that is impractical to do with real institutions, where size, hierarchy, and communication patterns are all entangled.

Some of the optimism around AI for science rests on the intuition that if we build systems with much stronger general reasoning, discovery will follow as a near-automatic consequence. But there is no guarantee that this will happen by default. Many technologies have promised radical scientific acceleration, and so far, not completely fulfilled their promise.

The Internet, for example, made knowledge much easier to search, and in theory may have led to much faster science. Online collaboration has certainly sometimes enabled better scientific work, such as in the Polymath Project. But this revolution has largely not emerged at scale, mostly because deeper inefficiencies in how we organize science (e.g., career incentives) remain a bottleneck. There is even evidence that online journals actually narrowed what researchers read and cite, because scientists search or follow hyperlinks rather than browsing journals, thus potentially accelerating consensus rather than expanding the space of ideas explored.

AI could repeat this pattern at a larger scale — generating faster results within the existing paradigm, while the structural conditions for disruptive science remain unchanged or worsen. There is no reason to expect this design problem to sort itself out on its own. But if we treat AI for disruptive science as a deliberate research program, we have a better chance of building the capabilities that paradigm shifts require. And to do that, we will have to understand how to design science itself.

Subscribe now

Alvin Djajadikerta is the CEO of Evidentia Labs and a founding researcher at Science Works. He holds a PhD in Molecular Neuroscience from the University of Cambridge.

Cite: Djajadikerta, Alvin. “Designing AI for Disruptive Science.” Asimov Press (2026). DOI: 10.62211/29ej-27et

By Rachel Dutton

The human immune system is, in one sense, a detection mechanism. It has evolved, over millions of years, to scan the body for molecular signals that tell it whether to attack or stand down. Most of these signals come from pathogens, damaged cells, or the body’s own hormones. But in 2019, a lab in Germany published a finding that pointed to a much stranger source: one of the signals sensed by the immune system is found in sauerkraut.

When people eat sauerkraut, a molecule called phenyllactic acid (D-PLA) — found in fermented foods — enters their bloodstream and activates a receptor, known as HCA3, on immune cells, triggering an anti-inflammatory response. In addition to lactic acid, phenyllactic acid is one of many compounds produced by lactic acid bacteria during the fermentation of sauerkraut and related fermented foods. Prior to this study, other molecules had been found to bind HCA3, but D-PLA was a hundredfold more potent than any of them.

This discovery advances our understanding of how fermented foods can reduce inflammation and positively affect human health. But more striking is what it suggests about hominid physiology. Although HCA3 is part of a larger family of receptors broadly conserved across eukaryotes, HCA3 is only present in humans and other great apes like chimpanzees and gorillas — and not even in other mammals. It is a recent addition to the genome, appearing only a few million years ago. Its existence seems to suggest that our immune system evolved to recognize the microbial metabolites from fermented foods.

We tend to think of fermented foods as something humans invented and then chose to eat. But, increasingly, scientific evidence suggests the causality runs the other way. Fermented foods appear to have helped shape human biology itself, and our bodies may have been built, in part, to expect them. The case for this runs from changes in hominid gut anatomy millions of years ago to the HCA3 receptor, to a growing body of research linking fermented food consumption to immune function and gut health. And it raises an uncomfortable question about what happened when the Western food system, in the name of safety and efficiency, quietly removed these foods from our diets in the nineteenth and twentieth centuries.

Early Fermentation

Fermented foods are the result of the controlled growth of communities of microbes. At their core, they are the products of the interaction of these microbes and whatever food they consume, whether cabbage or cucumber. While this process varies from food to food, fermentation typically involves managing environmental variables such as oxygen, temperature, and salinity. In contrast to food preservation methods like canning or even pickling, which are designed to prevent microbial growth, fermentation harnesses the capacity of naturally occurring microbes found on fresh food or in the environment to outcompete spoilage organisms.

Most archaeological evidence from pottery shards suggests that fermented food production is at least 7,000 to 10,000 years old. This timeframe coincides with the major transition to agrarian lifestyles, which would have reliably produced surpluses of food and the subsequent need for preservation methods.1 While this explanation satisfies most scholars, there is reason to believe that fermentation may be far older.

For one, it’s a very simple process to trigger. Some foods even ferment spontaneously. In the case of alcoholic drinks, like beer, wine, or mead, ubiquitous yeast species, which are naturally found on grapes and other fruit skins, rapidly use sugar as a food source for reproduction, producing ethanol as a byproduct. The same results occur when ripe fruit falls to the ground and its sugar is exposed to the environment, or when honey is diluted with water.

Other processes require only minimal intervention. For example, submerging fresh food in liquid or burying it creates a low-oxygen environment that encourages the growth of acid-producing bacteria that preserve the food by what is called “lactic acid fermentation.” This technique produces dill pickles, sauerkraut, and kimchi. While salt is often added as an additional intervention against unwanted microbes, it’s not required as it isn’t the primary driver of the fermentation.

Another argument for an earlier origin for fermented foods is that they are found across nearly all human cultures. While the number of fermented foods in the modern, Western diet is fairly limited (cheese, yogurt, bread, chocolate, coffee, beer, wine, kimchi, and kombucha) hundreds more are eaten around the world, from fermented shark in Greenland to a seemingly limitless variety of fermented soy beans in Asia. This diversity is a testament to how humans gradually mastered this ancient practice and modified it to suit new environments as they moved out of Africa.

Around 14 million years ago, our hominid ancestors were arboreal species whose diet would have been primarily based on fresh fruits picked from the trees they lived in. When ripe fruit fell to the ground and underwent spontaneous fermentation, it would have been toxic to our ancient ancestors due to its high concentration of ethanol. Their bodies as yet had no efficient way to break down ethanol.

But then, about 10 million years ago, a mutation arose in the genome of the common ancestor of humans, gorillas, and chimpanzees. This mutation, a single amino acid change in the enzyme Alcohol Dehydrogenase 4 (ADH4), enabled it to break down and detoxify ethanol with 40x higher efficiency. The capacity to consume this energy-rich but previously dangerous fruit may even have driven our transition from an arboreal lifestyle to a terrestrial one. What’s more, this ability to tolerate ethanol may have been what allowed our ancestors to diversify their diet and survive while lineages without this mutation went extinct.

The fossil record shows that a major shift in hominid anatomy occurred around 2 million years ago, when hominids developed a smaller rib cage and larger skull. At the same time, another major change took place in their intestines. Compared to our closest relatives, humans have a digestive tract that is 40 percent shorter. This decrease was thought to be driven by the external processing of our food, which reduced the time and energy involved in chewing and digesting. Anthropologist Richard Wrangham argues that the technological innovations of controlling fire and cooking food led to this major change, and that the excess energy we got from cooked food, in turn, supported the evolution of a larger brain.

However, two recent studies, by biological anthropologist Katie Amato in 2021 and evolutionary biologist Erin Hecht in 2023, suggest that these anatomical changes may have been driven by human use of fermentation even before humans began to cook. By allowing microbial species to ferment and break down complex carbohydrates and other macromolecules in foods, we may have turned over certain parts of an otherwise energy-intensive digestive process to microbes in a form of “external digestion.” This use of fermentation to pre-digest food, intentional or not, may have served as a predecessor to cooking, providing the extra calories needed to support the evolution of a larger brain.

Another benefit of fermentation is that it offered access to foods which, previously, would have been toxic. As our ancestors came down from the trees and needed new ways to fill their stomachs, the tubers of many plants and grasses offered an appealing, ready source of calories. Tubers contain large deposits of starch. Root vegetables, such as potatoes, yams, and carrots, are our modern-day, highly domesticated equivalents. But the wild tubers of our ancestors’ time were hard to chew, and some contained low levels of toxins. Varieties of cassava, for example, contain compounds that release cyanide when ingested. After just a few days of fermentation, however, microbes destroy these dangerous molecules and make the food safe to eat.2

Ultimately, given the simplicity of the fermentation process, its provision of new food sources, and its role in decreasing the need for extensive chewing and digestion, fermentation could have had an outsized impact on human evolution. It may, in fact, have helped make us human.

Peril and Promise

Perhaps the most striking thing about fermentation is that humans figured out how to control the growth of microbial species long before we understood what “microbes” were.

No one had seen microbial life until the 17th century, when Dutch scientist Antonie Van Leeuwenhoek used a handmade microscope to reveal small, motile forms he described as “animalcules.” It wasn’t until the late 1800s, however, that the French chemist Louis Pasteur demonstrated the role of microbes through his studies of fermented foods. His work on the spoilage of beer and wine formed the basis of the “germ theory” of disease. If microbes could be the causative agents of spoilage in food, Pasteur reasoned, maybe they could also be the causative agents of disease.

For his part, Pasteur was not strictly “anti-microbe.” He found them enthralling, investigating the differences between lactic fermentation (yogurt, pickles, and sauerkraut), butyric fermentation (butter, cheese, and milk), and acetic fermentation (kombucha, sourdough, sour beer), as well as developing the categories “aerobic” and “anaerobic” (which refer to whether microbes require oxygen to survive).

Even so, Pasteur’s work identifying microbes responsible for putrescence and disease has had the greatest impact on our food system. At the beginning of the twentieth century, automated machinery expanded canning capacity from roughly 10 cans a day to 1,000. Variation in the quality of the sealed lids, temperature treatments, and exposure to contaminants through handling meant a greater potential for microbial illnesses like botulism.

After three deadly botulism outbreaks in 1919 linked to California-distributed olives, the canning industry turned to Pasteur-inspired “bacteriologists” to design better canning systems and restore public confidence. While the resulting practices — including steam sterilization, the marking of batches, and traceable can coding — were at first voluntary, the Cannery Act of 1925 mandated statewide compliance.

By the mid-1900s, microbial research into food had come to focus on how to keep organisms out of it. Innovations in heat, pressure, refrigeration and anti-microbial agents to extend shelf life and decrease potential contamination with food-borne pathogens formed the core of academic and industrial research. As our methods of food production shifted, so did our diets. Americans largely moved away from the consumption of traditionally fermented foods, with one major exception.

That exception began with a single lecture. The Russian zoologist, Ilya Mechnikoff, based at the Pasteur Institute in Paris, was well known in the early 1900s for his scientific discoveries involving the immune system, for which he would win the Nobel Prize. But towards the end of his career, Metchnikoff became fascinated with the idea that aging was just another disease awaiting a cure.

In his 1908 book, The Prolongation of Life, Metchnikoff proposed that this disease was caused by "putrefying" microbes that lived within the human gut, or, as he put it, “chronic poisoning from an abundant intestinal flora.” His suggested cure came in an unexpected form — yogurt. Mechnikoff reasoned that just as the acid in sour milk prevented the growth of spoilage organisms, it might limit the growth of “spoilage” microbes in the gut.

After speaking with a student after one lecture, Metchnikoff learned of large numbers of centenarians in Bulgaria (one of whom he credulously reported as being “158” years of age).3 Metchnikoff hypothesized that the consumption of large quantities of yogurt in Bulgaria prevented aging through the action of lactic acid on gut bacteria. After obtaining a sample of the “Bulgarian bacillus” (Lactobacillus delbrückii subsp. bulgaricus) from yogurt, he set out to study the effects on the gut. His lab work showed that the presence of the yogurt microbe and the lactic acid it produced slowed the growth of intestinal microbes.

Metchnikoff thus advised the consumption of yogurt to promote healthy intestinal balance, writing: “A reader who has little knowledge of such matters may be surprised by my recommendation to absorb large quantities of microbes, as the general belief is that microbes are all harmful. This belief, however, is erroneous. There are many useful microbes, amongst which the lactic bacilli have an honourable place.”

Metchnikoff initially presented his yogurt hypothesis at a public lecture in 1904 in Paris, titled “Old Age.” The lecture went the early 1900s equivalent of viral. Newspapers ran stories with headlines like “Drink Sour Milk and Live to be 180 Years Old” (Evansville Courier and Press) and “Sour Milk is Elixir: Secret of Long Life is Discovered by Prof. Metchnikoff” (Chicago Daily Tribune). “Within months of Metchnikoff’s lecture, milk-souring germs had blossomed into an international business. Pharmacies throughout Europe and the United States were offering Bulgarian cultures in the form of tablets, powders, and bouillons — to be consumed as is or used,” writes Luba Vikhanski in her book Immunity. It was, in effect, the first “superfood.”4

“Probiotic” products quickly followed, containing dehydrated “lactic bacilli,” that anyone could use to sour their own milk in the way prescribed by Metchnikoff. But despite interest in yogurt and probiotics as health foods, most 20th century microbiologists remained focused on “bad actors,” the microbes and pathogens involved in disease. However, this changed toward the end of the century with advances in sequencing technology and chemical analysis techniques like mass spectrometry, which allowed them to probe microbial genetics and metabolism more closely and with more nuance.5

Environmental DNA sequencing revealed the diversity of microbial species across habitats, including in and on the human body as well as in our food. Understanding what these microbes are doing, and how, has become one of the challenges of modern microbiology.

To Eat Microbes

In response to this challenge, a lab at Stanford University, led by Erica and Justin Sonnenburg, has been doing seminal work on the human gut microbiome and how species within it function and interact. Recently, the Sonnenburgs have turned their focus on diet-microbiome interactions to fermented foods. A clinical trial, which the Sonnenburgs set up in collaboration with nutrition scientist Christopher Gardner in 2021, offered the clearest evidence to date that fermented foods are essential for gut health. In the trial, 36 healthy adults spent 10 weeks eating a diet high in such foods.

At the end of the study, they showed increases in gut microbiome diversity, which is generally associated with gut health. (The participants had new microbial species in their guts, which did not come from the fermented foods. It seems, rather, that fermented foods somehow make the existing gut ecosystem more receptive to incorporating new strains.) Perhaps an even more impactful outcome was that the researchers also found widespread decreases in inflammatory markers in the blood of these participants.

Given that low levels of inflammation sustained over time, referred to as chronic inflammation, is believed to be implicated in many diseases, the ability to decrease it though simple dietary changes would be welcome. However, the six servings of fermented foods a day consumed by the study participants is far higher than that in most U.S. diets. While yogurt and cheese are commonplace, with an average of 13.8lbs and 42.3lbs consumed per person per year respectively, this represents about a serving of fermented dairy a day. Fermented vegetables also comprise less than a serving per day. The impressive 387 million pounds of sauerkraut Americans consume per year equates to only 1.5lb per person, or about 0.06 servings a day.6 There has been a resurgence in interest in fermented foods, from high-end chefs experimenting with new types of fermentations to home fermenters caught up in the pandemic sourdough craze, but we have a long way to go before consumption reaches levels needed for clinically meaningful outcomes.

Even larger gaps remain in our understanding of how fermented foods actually drive health benefits. Beyond inflammation, fermented food consumption has been associated with a wide range of positive health outcomes, from gut health to mental health. As compelling as the Sonnenburg study is, it doesn’t yet reveal the mechanism(s) by which fermented foods trigger changes in human physiology.

While the 2019 sauerkraut study points to the importance of microbial metabolites like phenyllactic acid, if we co-evolved with fermented foods over millions of years, this example of a metabolite-receptor pairing is likely just the tip of the iceberg. As in many other areas of microbiome research, we need to move beyond correlation to causation. Doing so will require a much deeper understanding of the molecular composition of fermented foods, how they are connected to human biology, and more precise clinical studies.

Scientists are beginning to recognize the potential diversity of bioactive metabolites in fermented foods, but most studies focus on a single food type at a time and use different assays of bioactivity from lab to lab. To be able to effectively map the complex interactions between fermented food microbes, metabolites, and human biology, we need larger, consolidated datasets.

To this end, the Microcosm Foods project has been assembling a first of its kind collection of fermented food data pairing metagenomics (mapping microbes), metabolomics (mapping metabolites), and transcriptomics (mapping changes in human immune cells upon exposure to fermented foods) across over 100 different foods. Open-source, systematically-acquired datasets such as these will help build the scientific foundation needed to untangle these complex interactions.

For most of human history, people harnessed microbial ecosystems to make fermented foods and reaped the benefits without understanding their mechanism. Then, in a period of roughly 100 years, the Western food system replaced them with sterile, shelf-stable alternatives.

Today, we are finally beginning to understand how fermented foods interact with our biology — through receptors like HCA3, through shifts in gut microbial diversity, and through inflammation pathways. Science is catching up to what fermentation has been doing all along.

Rachel Dutton is a microbiologist studying fermented foods, from their use as microbiome models to their impacts on health. She received a PhD in Microbiology and Molecular Genetics from Harvard University, and led academic labs at Harvard and UC San Diego. Rachel is currently a Resident at the Astera Institute and a Fellow in the Big If True Science program at Renaissance Philanthropy.

Credit: L. reuteri image by Anastasiia Dmytriv.

Cite: Dutton, R. “Culture Shift.” Asimov Press (2026). DOI: 10.62211/92yr-34qw

This essay will appear in our forthcoming book, “Making the Modern Laboratory.”

By Donna Vatnick

In the 1960s, David Chambers, a researcher at Deakin University in Australia, instructed teachers to give children a blank sheet of paper and ask them to draw a scientist. Chambers repeated this experiment many times over eleven years, collecting more than 4,800 drawings. The results were surprisingly consistent: white lab coat, glasses, beakers, mysterious machinery, someone saying “eureka!” The study has since been repeated dozens of times. While some details have changed, with beakers replaced by rockets, microscopes by vaccines, or men by women (sometimes), the scientist always wears a white lab coat.

The white lab coat, however, only came to symbolize scientists in the 20th century. Before that, cartoonists satirized chemists by portraying their craft as sorcery whose practitioners wore long dark robes, and painters drew naturalists in waistcoats and breeches against backdrops of plants and landscapes. It was really surgery, more than any other scientific discipline, that gave us the white laboratory coat. Today, scientists don a variety of multicolored, specialized protective equipment to suit the needs of their field, but the fact that children still inextricably link white lab coats to “scientists” says everything about how a simple garment came to exemplify a profession’s public image.

To understand how the white lab coat arose, we have to go back to Victorian England to examine not the scientists, but rather the fashion of that time.

In mid-19th-century England, so-called “gentlemen of science” dressed in dark frock coats. In his 1885 portrait, Louis Pasteur stood in his laboratory, rabies sample in hand, in a black frock coat, waistcoat, and black cravat. Charles Darwin, whose personal home was his “laboratory,” sported a similar style. John Snow, the epidemiologist best known for tracing the source of London’s cholera epidemic, inspected the Broad Street pump while dressed like a banker, in a multipiece suit and tie.

This style was influenced by Prince Albert, Queen Victoria’s consort, who elevated the frock coat into a fitted, double-breasted symbol of class in the 1850s. Though the Prince himself likely never fooled around with specimens and staining chemicals, the darkness of the coats he popularized benefited scientists, who didn’t want the stains of their labors visible. Frock coats were, after all, very tedious to wash in the days before laundry machines.

However, for blood-splattered Victorian surgeons, the situation was much worse; heavy and woolen, the frock coat absorbed just about any fluid. Stories of surgeons making their rounds in bloody black robes reeking of rot and sweat are most likely accurate. The esteemed surgeon Sir Frederick Treves (credited with saving the life of King Edward VII in 1902) reminisced in his 1923 book: “The surgeon operated in a slaughter-house-suggesting frock coat of black cloth. It was stiff with the blood and the filth of years. The more sodden it was, the more forcibly did it bear evidence to the surgeon’s prowess.”

Beyond being filthy, frock coats were also very uncomfortable. The sawing, stitching, and other manipulations of surgery were already messy, laborious work, but operating rooms grew even more “clammy and sodden” after Joseph Lister introduced his carbolic spray antiseptic. Lister himself stripped off his frock and “turned up his shirt sleeves” and “pinned an ordinary unsterilized huckaback towel over his waistcoat (for his own protection, not that of the patient).”

In fact, Lister never advocated for changing the fabric of surgical garments. His writings from 1867 assured his readers that using carbolic spray was enough to keep infection risks low. In any case, illustrations from the bacteriologist Watson Cheyne’s authoritative 1882 textbook on antiseptic surgery showed a group of surgeons working in full formal attire, with their outdoor coats and hats resting on a nearby windowsill.

These early practitioners of science suffered from the expectation that they would, at all times, appear genteel. Even while hard at work, they eschewed light fabrics, protective gear, and aprons in favor of a wardrobe that demarcated their prestige.

The real driver of white lab coats was the hygienist movement. The mid-19th century witnessed dramatic strides in public sanitation. Rivers of free-flowing feces and effluent gave way to managed sewage disposal systems. Public bathhouses opened, such as the 1842 warm fresh-water baths on Frederick Street in Liverpool, England, serving the “vast numbers of workmen” of the 500,000 that populated the borough. The introduction of hand-washing in obstetrics in 1847 began to save patients, with doctors in other fields following suit. Motorized washing machines were invented in 1851. The domestic use of soap in England nearly tripled between 1801 and 1861, and then almost doubled again by 1891, and suddenly, both being and looking clean became possible.1

Ice cream vendors, butchers, and bakers were also adopting white clothing as their professional uniforms to project the image of working under sanitary conditions. White made any breach of cleanliness immediately obvious, another way to reassure their newly cleanliness-obsessed customers of the shops’ safety and security.

By the 1880s, some surgeons were beginning to adopt this aesthetic of cleanliness in their practices, like Robert Lawson Tait, a gynecological surgeon who donned a white apron. Tait used soap and water on patient skin, boiled instruments, applied freshly “laundried” towels around wounds, and had a regimented hand-washing ritual before operating. His surgical attire was a visual signal meant to attract paying clientele.

Other surgeons took this fashion further. When Australian surgeon Alexander MacCormick became “the first man in Sydney to wear a white coat while operating” in the 1880s, his colleagues mocked him: “MacCormick was called the ‘Hokey Pokey Man,’ a reference to the popular confection called hokey-pokey, sold by ice cream vendors,” write historians Susan Hardy and Anthony Corones in Dressed to Heal: The Changing Semiotics of Surgical Dress. The joke backfired. Patients flocked to him, attracted by his gleaming cleanliness.

The white medical garment popularized during this period was everything the frock coat wasn’t: washable, lightweight, cheap, and disposable. Hospitals could send uniforms to commercial laundries with a quick turnaround. Textile mills churned out mass-produced cotton and linen garments after the Civil War. By contrast, frock coats were tailor-made from broadcloth, requiring tremendous time and effort.

Two famous portraits from the American painter Thomas Eakins capture the transition from black to white. “The Gross Clinic” from 1875 shows surgeon Samuel Gross and his assistants in black coats, blood on their hands, with a skylight letting in the noontime sun. The patient is a young man wearing his street socks. When this life-size painting was introduced at the 1876 United States Centennial Exposition in Philadelphia, it was relegated to a remote corner of the exhibit hall after having been initially rejected for display at all. Some scholars suspect that the painting was too “realistic” for a public that now expected visible hygiene.

“The Agnew Clinic,” painted in 1889, presents a stark contrast, with surgeon David Hayes Agnew and colleagues in shining white gowns using sterilized instruments under artificial light. The patient is draped in white sheets on a white table. Here, however, the painting is misleading because Agnew was remarkably resistant to adopting these practices. The public remembered Agnew as one of the several attending physicians to the beloved President James A. Garfield after he was shot in the back at a railroad station in DC in 1881. Agnew had poked and prodded at the President’s bullet wound without ever having washed his hands or sterilizing his probes. As the President developed abscesses, Dr. Agnew would drain them with dirty instruments.

The younger generations of American surgeons watched in horror, but even after this high-profile disaster (with evidence of the infected abscesses revealed when the president succumbed to his injury two months later), Agnew didn’t change his techniques. Just a year before this painting was made, a photograph shows Agnew in a clinic, wearing a nice, buttoned-up street coat just as Samuel Gross had a decade before.2

The new expectations of operating rooms didn’t immediately apply to those of the laboratory. White linens weren’t as practical for bench science as they were for clinical medicine. Industrial chemists, for example, were known to wear brownish — not white — coats and aprons, often made of leather, in German and British labs in the early 1900s. Wanting to hide the stains of her experimentation, famed physicist and chemist Marie Curie requested her wedding dress to be “practical and dark” so that she could afterwards wear it in the lab. This she did, until switching to black dresses upon the death of her husband in 1906.

Curie was not alone in preferring black for cleanliness. French surgeon and biologist Alexis Carrel worked out of a black-walled lab and wore black deliberately because it showed dust. Black also works better for contrast when using white mice and powders, as well as hiding any blood or pigment stains. Even today, some scientists prefer black labwear.

Despite such outliers, white lab coats have become synonymous with lab scientists, even if experts disagree on exactly when the white lab coat made its way into the lab. A 1902 photograph in a neurobiological lab in Berlin, for example, shows researchers wearing long, light colored coats while working at benches. Another from 1922 shows two industrial chemists wearing what appear to be white coats.

Of course, the coats in these black-and-white photographs could actually be light brown, or they could have been touched up in retrospect, with photo-editors assuming the coats were white. After all, many paintings and reproductions show scientists wearing white lab coats long before they would’ve actually worn them. For example, a 1930s lithograph from the “Teachers World” supplement shows an illustration of Pasteur mulling over test tubes in a white lab coat.

The standard garb of laboratory workers also began to shift to white during the turn of the 20th century, when the relationship between medicine and laboratory science became more entwined. Lab scientists likely adopted surgical fashion as a result. In 1898, Sir William Osler, a Canadian physician already known for writing some of the first clinical laboratory literature, introduced ward laboratories at Johns Hopkins Hospital. Gradually, surgical practice relied on laboratories for tools, microscopes, chemical assays, pathology reports, and bacterial cultures to diagnose patients.

After the American education reformer, Abraham Flexner, visited medical schools all over America, he published a report in 1910 calling for restructuring medical education by getting rid of private medical schools, standardizing admissions, laboratory access, and curricula. With this, hospitals transformed from small, charitable institutions for the urban poor to large, funded medical centers serving mixed-class populations. They projected an image of themselves as safe, clean places, with the white lab coat — worn by surgeons and lab scientists alike — as evidence of that promise.

This white coat also became a way for these professionals to signal their membership in their respective highly-trained “guilds,” separating themselves from non-scientists. When chemistry historian Peter Morris stumbled upon a 1950s photograph of schoolboys from his alma mater wearing white lab coats, he theorized that the coat was introduced “to inculcate an ‘esprit de corps’ in students and laboratory workers” to improve morale rather than protect clothing.

Social psychologists tested a similar theory in 2012. Students wearing white coats, which they believed were doctors’ coats, performed better on tests. Students wearing identical coats, which they thought were painters’ smocks, showed no improvement. While the researchers called this observed phenomenon “enclothed cognition,” the real takeaway was perhaps much more intuitive — clothing shapes the way we think about ourselves.

The powerful symbolism explains the lab coat’s persistence even as we question its functionality. In 2009, a UCLA researcher was transferring tert-butyl lithium, a chemical that ignites spontaneously in air, between tubes. She wore safety glasses and nitrile gloves but no lab coat. When the syringe malfunctioned, and the highly flammable chemical splashed on her synthetic sweater, the synthetic material, essentially a solid form of gasoline, caught fire immediately. A colleague tried to smother the flames with his own lab coat, but failed. She should have been rushed to the emergency shower, but in the panic, she was not. She died from her burns, which covered 40 percent of her body.

Could a lab coat have saved her? Some are skeptical, since many common lab coats are also made of flammable, synthetic materials. But a flame-resistant coat, which had been commercially available for years, certainly might have.

Indeed, the incident incited a widespread rethinking of PPE, which is often eschewed for being bulky and cumbersome. Unfortunately, the environment the UCLA researcher worked in and the “circumstances that led to her death were certainly not unique.” Laboratory safety standards often lag behind available technology, and a culture of safety and training is wanting, even today.

Despite such lassitude, the lab coat has become an object of innovation. In 2012, MIT launched an initiative to create lab coats that scientists actually wanted to wear, addressing not just design but laundry logistics and institutional culture. The MIT Media Lab took this a step further, declaring, “Media Lab researchers are not only scientists — we are also designers, tinkerers, philosophers and artists. We need a different coat!” Their philosophy is that PPE should not only protect scientists but also support the exact kind of work they do. A scientist carrying around motors needs specialized pockets; one that works with lasers needs a reflective coating.

Since then, material science has ushered in promising possibilities for industrial uniforms, including lab coats, though many of these breakthroughs are in the research phase as of 2025. Scientists are experimenting with sustainable, bioactive coatings like chitosan (from seashells) and silver nanoparticles — natural microbial fighters — as well as phosphorus-based flame retardants that are highly effective at low concentrations. Built layer-by-layer with nanoparticles such as silica and titanium oxide, the experimental textiles can self-clean when activated by short periods in the sun. While we have yet to scale these innovations, their potential to improve lab coat functionality and clothing in general is worth watching.

Today, despite a wider array of shapes, colors, sizes, and materials, the archetypal dress of the laboratory remains the white coat. This suggests that the broader challenge is not design, but adoption. How do scientific institutions create new symbols of identity that prioritize function over tradition? Looking at the history of the white coat, we can observe that the shift from black frocks to white coats took decades and required not only technological but cultural change. The next transition, from symbolic to specialized PPE and laboratory wear, requires a similar shift in imagination.

Subscribe now

Donna Vatnick writes essays that explore scientific discovery and its most passionate devotees. Before completing her MFA in nonfiction, she worked in molecular biology labs and coordinated clinical trials in Boston.

Acknowledgements: Thanks to Kate Goldkamp and Natasha Muhametzyanova for their research guidance, and Ethan Madore and Julia Zaltsman for valuable conversations. Header credit: Wellcome Collection.

Cite: Vatnick, D. “Why Lab Coats Are White.” Asimov Press (2026). DOI: 10.62211/62hw-98tk

Further Reading:

1. Chambers, D. W. (1983). Stereotypic images of the scientist: The draw‑a‑scientist test. Science Education, 67(2), 255–265. https://doi.org/10.1002/sce.3730670213

2. D’Addezio, G., & Besker, N. (2024, January). Science and scientists from children’s point of view: comparison and gender outlooks among 2011 and 2021 primary school student drawings. Frontiers in Education, 8. https://doi.org/10.3389/feduc.2023.1179179

3. Rehn, D. K. (2021, November 3). Poverty and the pursuit of the philosopher’s stone: Representation of alchemists in sixteenth‑century Netherlandish art. Dana K Rehn Blog. https://danakrehnblog.wordpress.com/2021/11/03/poverty‑and‑the‑pursuit‑of‑the‑philosophers‑stone/

4. Smithsonian Institution. (2020). Alexander von Humboldt’s influence on America. IMPACT, 6(2). https://www.si.edu/support/impact/humboldt

5. Hardy, S., & Corones, A. (2016). Dressed to heal: The changing semiotics of surgical dress. Fashion Theory: The Journal of Dress, Body & Culture, 20(1), 27–49. https://doi.org/10.1080/1362704X.2015.1077653

6. Lister, B. J. (2010). The Classic: On the antiseptic principle in the practice of surgery. Clinical Orthopaedics and Related Research, 468(8), 2012–2016. https://doi.org/10.1007/s11999‑010‑1320‑x

7. Cheyne, W. W. (1882). Antiseptic surgery: its principles, practice, history and results. Smith, Elder. https://archive.org/details/antisepticsurger00chey/page/70/mode/2up

8. Drysdale, C. (2025). What we find in the sewers. Asimov Press. https://www.asimov.press/p/sewers

9. Paul, S., Salunkhe, S., Sravanthi, K., & Mane, S. V. (2024). Pioneering Hand Hygiene: Ignaz Semmelweis and the Fight Against Puerperal Fever. Cureus, 16(10). https://doi.org/10.7759/cureus.71689

10. Aiello, A. E., Larson, E. L., & Sedlak, R. (2008). Hidden heroes of the health revolution Sanitation and personal hygiene. American Journal of Infection Control, 36(10), S128-S151. https://doi.org/10.1016/j.ajic.2008.09.008

11. Macintyre, I., & Hughes, S. (2024). Robert Lawson Tait (1845–1899): The true innovator of aseptic surgery?. Journal of Medical Biography, 32(1), 157-165. https://doi.org/10.1177/09677720221140085

12. Flannery, M. C. (1999). Dressing in Style? An Essay on the Lab Coat. The American Biology Teacher, 61(5), 380–383. https://doi.org/10.2307/4450702

13. Friedlaender, G. E., & Friedlaender, L. K. (2014). Art in Science: the Gross Clinic by Thomas Eakins. Clinical Orthopaedics and Related Research, 472(12), 3632–3636. https://doi.org/10.1007/s11999-014-3989-8

14. Millard, C. (2012). Destiny of the Republic: A Tale of Madness, Medicine and the Murder of a President. Penguin Random House.

By Dr. David S. Kim

Researchers have known that the pancreas secretes several peptide hormones critical to metabolic regulation, most notably insulin and glucagon, since the early 1900s. With the advent of recombinant peptide synthesis in the 1970s, scientists not only began working on isolating the gene responsible for insulin production but fabricating it outside the body. In 1978, scientists at Genentech turned bacteria into factories able to transform sugar into insulin by cutting the insulin gene from the human genome and pasting it into the genome of E. coli. This breakthrough allowed peptide drugs to be synthesized at mass scale, without needing to be harvested from animal pancreases, and gave birth to the biotechnology industry.

The role of glucagon and the gene encoding it, however, remained elusive. Then, in 1982, Joel Habener and colleagues at Massachusetts General Hospital (MGH) published a paper indicating not only that they had located the human glucagon gene, but that it actually encoded three separate peptides, two previously unknown. Those two unknowns were labelled as glucagon-like peptides: GLP-1 and GLP-2.

Svetlana Mojsov, a peptide chemist also working at MGH, managed to synthesize a truncated, but active, form of GLP-1. In 1987, her paper with Habener and Daniel Drucker showed that GLP-1 stimulates glucose-dependent insulin secretion in the pancreas, suppresses glucagon release, and slows gastric emptying, all of which stabilize blood sugar levels after eating.

While these effects initially made GLP-1 seem a promising candidate for drug development, the earliest clinical studies from 1992 and 1993 revealed that GLP-1’s half-life in the body was disappointingly short. A proteolytic enzyme, dipeptidyl peptidase-4 (DPP-4), binds to and destroys GLP-1 molecules in the blood in minutes.

Hope for boosting GLP-1’s half-life in the body came in the 1990s, when endocrinologist John Eng, at the Bronx Veteran Affairs Hospital, noticed something curious about the Gila monster while researching snake and lizard venom effects on the pancreas. Specifically, he was struck by how this desert lizard eats extremely infrequently, yet somehow maintains a stable blood sugar level for months.

Eng found that Gila monsters secreted a salivary venom peptide, now known as exendin-4. Structurally similar to GLP-1 but unaffected by DPP-4, this peptide lingers in the blood for hours rather than minutes. Exendin-4 thus became the molecular blueprint for exenatide, the first twice-daily injectable GLP-1 receptor agonist, approved by the FDA in 2005.

For many years after exenatide’s approval, GLP-1 receptor agonists were used primarily as diabetes medications. Practitioners and patients noticed, however, that taking such medicines produced weight loss as a side effect. During this time, researchers also found GLP-1 receptors in the brain and pancreas, particularly in areas controlling appetite and reward. And in 2021, the results from a series of five STEP trials (Semaglutide Treatment Effect in People with obesity) demonstrated marked weight loss even in individuals without type 2 diabetes (T2DM): Phase 3, for example, showed an average loss of about 15 percent over 68 weeks, more than double achieved by any prior obesity drug. In 2023, a trial of 17,604 adults (SELECT) with obesity and without T2DM, followed for over three years, showed a 20 percent reduction in major adverse cardiovascular events.

In response, GLP-1 prescriptions rose sharply. According to CDC data, the share of adults with type 2 diabetes using GLP-1 injectables more than doubled — rising 155 percent between 2018 and 2022 — while overall spending on these medications climbed more than 500 percent between 2018 and 2023.

The majority of patients continue to inject these drugs rather than take them orally. But a 2020 survey of 600 patients found that 76.5 percent would prefer a once‑daily oral dose over a once-weekly injection of GLP-1. Significant funding is going into R&D to launch an oral alternative that performs as well as an injectable version. While oral formulations are beginning to enter the market, so far they remain less effective and more expensive to manufacture. Why?

To the Stomach, All Peptides Are Food

Peptide drugs like GLP-1s are made up of a chain of amino acids chemically bonded together. For a GLP-1 molecule taken by mouth to reach its target, it must first pass through the stomach, where hydrochloric acid and peptidase enzymes cleave these amino acid bonds. Most ingested peptides are simply broken down into their constituent amino acids and absorbed as nutrition, indistinguishable from the protein in food.

The small percentage of GLP-1 molecules that might survive this journey and arrive into the intestines face a second gauntlet. Once in the blood capillaries, DPP-4 quickly degrades the molecule, and the drug is excreted by the kidneys.

Biochemical engineering, however, has allowed drugs like semaglutide to survive longer in the blood. First, an amino acid substitution stops DPP-4 from destroying GLP-1. This is achieved by swapping the alanine amino acid at position 8 in the peptide backbone for alpha-aminoisobutyric acid (Aib), a modified amino acid that DPP-4 cannot recognize or cleave. As a result, this modified form of GLP-1 is not degraded so swiftly.

Yet even a DPP-4-resistant peptide is small enough to be filtered by the glomeruli in the kidneys into the urine. Therefore, the chemists behind semaglutide made a second modification: they fused a fatty acid chain to the peptide in a process called lipidation. This causes the drug to bind to human serum albumin, a large transport protein abundant in the blood, with a natural half-life of approximately 20 days. The drug effectively co-opts albumin’s stability and longevity; the lipidated GLP-1 molecules are bound to albumin, which is too large to be renally filtered, and are thus shielded from plasma proteases.

The combined effect of these modifications increase the drug’s half-life from minutes to approximately 168 hours, enabling its once-weekly injection schedule — an enormous improvement compared to the twice daily injection required for exenatide (Byetta) which was the first GLP-1 drug that was launched.1 Semaglutide (under the brand names Ozempic and Wegovy) swept the market. However, these modifications to the molecule could not solve the problem of degradation and poor absorption when taken by mouth.

From Silkworm to Trojan Horse

We might draw parallels for oral delivery of peptide and protein drugs from the case of serratiopeptidase, a little‑known drug from the 1960s with a peculiar history.

Serratiopeptidase is used to treat post‑operative inflammation in a handful of countries, including Japan, India and Sierra Leone. Produced naturally by the bacterium Serratia marcescens, which resides in the gut of the Japanese silkworm, this enzyme is used by the moth to dissolve silk fibers as it emerges from a cocoon. It is pharmacokinetically puzzling as an oral therapy because, as a peptide, it faces the usual hurdles of gastric degradation and poor epithelial uptake. Yet, surprisingly, no injectable formulation of serratiopeptidase is marketed. The reason is that intravenous administration carries a risk of immunogenicity — the immune system recognizes Serratia‑derived molecular signatures on the enzyme as foreign, which can trigger immune reactions. This danger is blunted or even absent, however, when the drug is given orally.

Serratiopeptidase only works because the oral dose is high enough that the small fraction of the drug which survives the gut reaches systemic circulation at concentrations sufficient for therapeutic effect. In this way, it serves as a proof‑of‑concept that even a peptide or protein drug known to have poor bioavailability can still be delivered orally, and without any modification, if we compensate by delivering a sufficiently high dose.

But in general, turning a peptide (including GLP-1 or insulin) into an oral formulation means overcoming three major challenges: protecting the peptide from digestion in the stomach, improving absorption of the peptide via the gut, and slowing down the enzymatic breakdown and excretion of the peptide once it is in the blood. Novo Nordisk’s oral semaglutide, Rybelsus, employs both co-formulation and clever biochemical engineering of the drug molecule itself to move us closer to this goal.

Specifically, Rybelsus works by co-formulating semaglutide with SNAC (sodium N-[8-(2-hydroxybenzoyl) amino] caprylate), a small synthetic molecule that functions as an absorption enhancer in the gastrointestinal tract. SNAC raises the pH of the immediate environment around the drug. The higher pH protects the peptide from pepsin degradation, which requires acidic conditions. SNAC also acts as a transient permeation enhancer, aiding semaglutide’s absorption into the blood, directly through the stomach lining and the small intestine.

While this co-formulation certainly improves oral delivery, it hasn’t fully solved the problem of dosage requirements. One pharmacokinetic study showed a modest bioavailability of only 0.8 percent for oral semaglutide, meaning less than one percent of the drug is absorbed into the blood where it can take effect.

To be effective, then, Rybelsus must be formulated to contain high doses of semaglutide. Whereas the GLP-1 drug Ozempic is injected once weekly at 2 mg, Rybelsus is taken daily and requires a minimum dose of 14 mg to achieve comparable glycemic effects. That is almost a 50-fold difference in total weekly intake for what is, fundamentally, the same drug, simply because most of the substance is degraded before they can be absorbed into the blood.

Additionally, Rybelsus doesn’t perform as well as its injectable counterpart, even at this higher dosage. In clinical trials, Ozempic (2 mg weekly subcutaneous) delivered a 15-20 percent body weight reduction over 68 weeks. Rybelsus at the approved 14 mg daily dose, though, achieved just 4-5 percent in the PIONEER trials. However, at 50 mg, oral semaglutide (OASIS 1) narrowed that gap substantially, achieving 15.1 percent weight loss at 68 weeks, approaching injectable performance.

While it is no secret that the cost of production of the drug molecule itself is only a small part of the retail price, the higher dosage required by oral delivery becomes a non-trivial difference in manufacturing costs.

Semaglutide costs $70,000 to $100,000 per kilogram to manufacture at commercial scale, according to two studies. Producing the drug ingredient contributes only $0.20 to the cost of each weekly 2mg injection. What forms the bulk of the retail price2 of peptides is not the cost of manufacturing, but the capital infrastructure: the regulatory burden, the cold chain logistics, the pen device (which sometimes costs more than the drug itself), and, above all, the return needed to justify $9 billion in the initial R&D of these GLP-1 drugs.

The cost breakdown looks much different for an oral semaglutide. In the OASIS 1 trial, patients took a 50 mg daily dose — meaning a monthly supply contains roughly 1,500 mg of active drug compound. A monthly injectable supply, by comparison, contains only about 8 mg. At $100,000 per kilogram, the drug compound alone for an oral monthly supply could already cost around $150, versus less than $1 for the injectable — before factoring in formulation, packaging, distribution, or profit.

This is the central problem with oral GLP-1s, then; biochemical engineering can technically solve the half-life problem, and co-formulations can improve the absorption problem, but the higher doses required make for formidable raw ingredient costs.3

Turning Food into Drug Delivery Vehicles

A solution to this problem may come from foregoing chemistry and turning, instead, to biology.

Genentech pioneered recombinant manufacturing of human insulin by cloning the human gene into E. coli as far back as 1978. Even today, many GLP-1 peptides are made by cloning the corresponding gene into microbes, culturing them in a fermenter, and then subjecting the resulting broth to successive rounds of filtration and chromatography to isolate the peptide precursor before undergoing chemical modifications. E. coli, though, is not edible, as it is laden with endotoxins that can trigger a septic shock. Therefore, any drug made in a cell like E. coli must be purified extensively, even if delivered orally. This downstream purification accounts for up to 80 percent of the total manufacturing cost.

Suppose, instead, that the microbes used for production were also edible. The drug could then be swallowed inside intact cells, and the massive cost of purification would disappear. In theory, there are millions of microbe species that can be engineered to synthesize peptides. But currently, we have the tools to genetically engineer only a few, and most microbes have never been studied for food safety and toxicology.

One attractive candidate is the edible, single-celled cyanobacterium Arthrospira platensis (spirulina). Cyanobacteria and many microalgae species have been consumed by humans for centuries and are Generally Recognized As Safe (GRAS) by regulatory agencies. Many people already eat spirulina for its antioxidants. Thus, the organism could be quite appealing for direct oral delivery in its intact cellular form, if it could be engineered to synthesize a GLP-1 molecule in sufficient quantities. Given the low cost of production, a 20-gram scoop of dried “algae powder” taken twice daily could, in principle, contain an orally adjusted therapeutic dose of a GLP-1 agonist for around $4/day.4

The cost argument alone is compelling, but using edible algae provides another benefit: keeping the GLP-1 inside the algal cell wall may improve the drug’s survival in the gut. Experimental data have already demonstrated that over 70 percent of the expressed proteins in spirulina remained intact after two hours of incubation under simulated gastric conditions that would fully degrade the same protein in a purified form within minutes. The cell wall acts as a natural enteric capsule, shielding the GLP-1 molecules from pepsin and the acid bath of the stomach, and releasing it progressively as it transitions into the higher-pH environment of the duodenum.

The reason this concept has not yet been tested in the context of GLP-1 is because, until very recently, it was not possible to engineer microalgae and cyanobacteria. Spirulina was, until a few years ago, considered genetically intractable because it has multiple copies of its genome and is resistant to standard transformation methods. In 2022, though, a paper reported successful engineering and high-level expression of therapeutic proteins in spirulina, with the encoded proteins accumulating to up to 15 percent of total dry biomass.

Today, newer organisms like UTEX 3222 (discovered in 2023 and also known as “Chonkus”) can double several times faster than spirulina and be engineered to express recombinant peptides. Some of these strains are quite simple to grow, needing just light and CO₂. Their adoption might not only make oral GLP-1s affordable, but also allow the mass production of customized peptides from photosynthesis — all inside edible cells.

Many cyanobacteria (and microalgae broadly) are safe to eat, though not every strain of cyanobacteria has a GRAS certification.5 Under the GRAS framework, an ingredient has to be demonstrated as safe either “through (1) scientific procedures or (2) in the case of a substance used in food prior to January 1, 1958, through experience based on common use in food.” For cyanobacteria to be used as a platform for oral GLP-1 delivery, either a GRAS organism like spirulina must serve as the host organism or, if choosing a different species, the manufacturer must prove with adequate in vivo trial data that it is safe to eat.

In addition, one must gain FDA approval by clinically demonstrating that the delivery of the microbe-encapsulated GLP-1 through this oral route leaves the drug molecule itself unchanged (with no post-translational modifications), that it has a similar pharmacokinetic profile to the reference molecule, and that it shows no adverse effects. Since this is a new delivery platform of existing drugs rather than a new drug, the regulatory bodies in many jurisdictions typically have alternative, shorter approval pathways available to drug sponsors that do not require the standard Phase I-III clinical trials for a new drug application.

A third argument for adopting edible organisms for GLP-1 delivery would be to increase global accessibility. Bangladesh, for example, ranks among the world’s top ten countries for T2DM patient burden, with a prevalence of 9.2 percent among adults. The disease currently impacts over eight million people, a number projected to grow to 15 million by 2045. Yet GLP-1 drugs are effectively absent from the formularies of Low and Middle Income Countries (LMICs). A 2024 study found that GLP-1 drugs can likely be manufactured more cheaply than their current market costs of, for example, $95 in Brazil, $115 in South Africa, and $353 in the United States.

Médecins Sans Frontières has called for Novo Nordisk and Eli Lilly to license their patents to generic manufacturers, noting that neither company has adopted pricing strategies for low-income markets. The surge in U.S. demand has also triggered global shortages that hinder the drug’s availability for essential diabetes care in LMICs. If we can provide alternative biomanufacturing capacity that slashes cost, turns injectables into safe oral, algae-encapsulated formulations, and eliminates the need for refrigeration, we may be able to ensure that patients in LMICs are no longer last in line for the latest and best therapeutics.

So far, three major problems have kept GLP-1 drugs injectable rather than oral: enzymatic degradation, poor bioavailability, and the sheer cost of manufacturing. Novo Nordisk's Rybelsus showed that transcellular gastric absorption, boosted with SNAC, can partially get around the first two problems, but cannot solve the third. This is why research groups around the world are trying to engineer microbes to act as edible drug factories. The idea is to skip the most expensive part of manufacturing, downstream purification, by keeping the expressed peptide locked inside the protective walls of edible cells.

The science of oral peptides, at this point, is no longer a mystery. We know oral peptide drugs can work. The harder question is whether patients will come to accept the idea of eating their medicine in a spoonful of algae.

Subscribe now

David S. Kim is the founder of CyanoCapture [General Bio], a synthetic biology company that is manufacturing peptides by genetically engineering non-model microbes. He is a medical doctor based in the U.K.

Cite: Kim, D.S. “The Quest for Oral GLP-1s.” Asimov Press (2026). DOI: 10.62211/82ej-91yt

Living organisms are surprisingly creative when it comes to finding their way into hostile environments. Microbes, after all, thrive everywhere from hyperarid deserts and hydrothermal vents to the stratosphere.

Of all hostile environments, one might think the most unbreachable to be the laboratory, where populations of bacteria are stored inside tubes sealed with tight screw-caps and stacked in freezers at -80°C. Of course, normal evolutionary processes don’t magically stop at the doors of the lab: organisms still mutate, fragile microbial species can still be out-competed by faster-growing contaminants, and parasites can still spread. However, vigilant scientists devote considerable effort to making sure containers hold what is written on their labels and, when in doubt, can sequence any part of a sample’s DNA, discarding or incinerating undesired specimens.

Still, interlopers persist. One organism that has been particularly successful at infiltrating microbiology research facilities is called Φ80. It is a bacteriophage, or virus that infects bacteria. Slowly but surely, Φ80 has been spreading. Researchers carry out experiments, make reproducible observations, and publish — usually without ever realizing that their samples have quietly been contaminated with this phage.

On occasion, hidden contaminations of Φ80 have led to paper retractions, with researchers discovering not only that their results are artifactual, but that they personally amplified the invasion and transmitted it elsewhere. In a 2017 paper, for example, scientists confessed that “beyond the confusion created by our erroneous interpretation of results obtained with Φ80-infected strains, we are worried that we have sent these strains to many research groups.”

To understand how such unwitting contamination can occur despite researchers’ best efforts, and to work towards better containment, it helps to first understand how bacteria are typically cultivated.

Life in a Tube

The most commonly studied bacterium, Escherichia coli (E. coli), has spent most of its recent evolution inside the human gut. Acclimated to that environment, E. coli grows best at 37°C — dividing about every 20 minutes in good conditions. If a researcher inoculates just a few E. coli cells in a rich liquid medium and puts the tube in a shaking incubator, it only takes about a dozen hours to form a population of several billion cells. As these billions of cells continue to divide every 20 minutes, their consumption of nutrients also grows exponentially, eventually exhausting all resources, at which point they enter a kind of dormant state.

E. coli cells are almost completely transparent, but their refractive index differs slightly from the medium in which they are grown. Each cell is only about one micrometer wide, but when millions accumulate in the same test tube, visible light gets scattered, and the liquid becomes opaque.

Compared to animal guts, with their folds, crypts, and periodic floods of bile, a test tube is, from the point of view of a bacterium, a featureless space. Most importantly, it is also the only place on Earth where bacteria don’t have to deal with phages. While we don’t pay much attention to phages, they are everywhere, and Earth would be quite different without them. Every day, phages kill 15 to 40 percent of all bacteria in the ocean, keeping microbial ecosystems from spiraling out of balance.

A thriving culture of bacteria, peacefully growing in a test tube, would surely be an appetizing ground for a bacteria-killing phage to replicate, if it could find its way inside. For this reason, microbiologists are extremely vigilant about phage contamination. Merely mentioning phages at lab meetings will draw nervous looks!

Phages get everywhere, are hard to destroy, and tear through experiments. I’ve seen entire labs shut down for days while whole HVAC systems are sterilized. If there is any detectable sign of unwanted phages, then, researchers stop what they’re doing, autoclave every object possible at the highest possible temperature, and wash the rest with the harshest chemicals available. While many bacteriophages are capable of infecting common laboratory strains, they can usually only spread so far.

The most straightforward strategy by which phages spread is by replicating as fast as they can, infecting as many cells as possible, and killing them in the process. Phages that employ this strategy are called lytic.

One example is T7, a lytic phage that infects E. coli. When T7 enters a bacterium, it uses the cell’s resources to generate copies of itself. This takes about 17 minutes, at which point the cell contains hundreds of phage particles. The phages then burst out of the cell in an event called lysis, searching for new cells to infect. The speed of this process depends on the population density of the bacteria. In ideal conditions, a phage population can expand by a factor of 10⁵–10⁶ every hour.

For bacteria in a test tube, phage infection means annihilation. Billions of cells can get entirely wiped out in two hours. As the bacteria burst and release their contents in the medium, the liquid becomes entirely transparent again.

While the lytic strategy employed by T7 is superb for proliferation, it doesn’t work well for secretly invading a laboratory. If an entire bacteriological culture suddenly collapses for no reason, even the most absent-minded researcher is apt to notice that their cells have been contaminated.

In fact, being too destructive is also a problem in the wild: if a phage eradicates its host too quickly, there might be no bacteria left to infect, meaning death for the phage as well. For that reason, many phages have adopted a more complex life cycle than that of T7, one known as lysogeny.

Take Phage λ, one of the best-studied lysogenic phages. Whenever it infects a new cell, it uses a genetic circuit to choose between two outcomes: either it goes through the usual lytic cycle, filling the cell with virions until it bursts, or it inserts its genome into the bacterium’s DNA. The integrated form is called a prophage.

The host bacterium then carries the phage genome even as it grows and divides. HIV behaves similarly in humans, but the advantage of doing so is much more obvious in bacteria: with a host that divides every 20 minutes, it can be worthwhile for phages to hitchhike for several generations of bacterial division, giving the host time to proliferate on its behalf.

It is even common for lysogenic phages to confer some benefit on their host, helping the infected bacteria outcompete their uninfected sisters and maximizing the prevalence of the prophage in the population.1 There is an interesting puzzle here: if there were an easy way to improve the fitness of the host bacteria, surely the bacteria would already have found it through normal natural selection. How, then, can a parasite grant an advantage that the host wouldn’t have evolved on its own?

One way is to provide the bacteria with genes helpful in the short-term at the expense of long-term sustainability. For example, phage-provided genes may offer costly solutions to temporary problems, like immunity to other pathogens or resistance to specific stresses in the environment.

When the host bacterium is itself a pathogen, phages can benefit from making bacteria more virulent to their host, often against the interests of the bacteria themselves. Some of the world’s most horrific diseases, like diphtheria, cholera, and botulism, are due to bacteria releasing deadly toxins into their host’s body. However, these toxin genes don’t come from the bacteria — they are often poisoned gifts from lysogenic phages.

So when do these dormant lysogenic phages elect to attack? The trigger is usually some indication that the cell is imperiled, whether by DNA damage, oxidative stress, or antibiotics. Some phages even force their hosts to secrete molecules into the environment so they can coordinate their decisions with phages in other infected cells. At the signal, like rats abandoning a sinking ship, the prophage leaves the chromosome, replicates profusely, and bursts into the environment.

In the context of the lab, this often resembles lytic infection, with most cells bursting and becoming transparent. But then, a small subpopulation of bacteria survives, having integrated the phage into their DNA. The prophage then expresses a protein that blocks further infection by the same phage, creating an immune population of cells that resumes growth as if nothing happened.

So, by using lysogeny, a phage can stay hidden in the chromosome of bacteria for extended periods of time. However, in a research lab environment, this tactic can only succeed if the phage manages to integrate itself into bacteria without the researchers noticing. And this is precisely why Φ80 is so cunning.

Subscribe now

The Intruder

A key feature of Φ80 is that it replicates very slowly. While T7 goes through a full lifecycle in less than 20 minutes, Φ80 takes 90 minutes. This is a more dramatic difference than it seems: growth rate compounds with each generation, meaning that T7 can produce 10¹⁰ new particles in two hours, while Φ80 makes barely a thousand in the same interval.

Moreover, while the mechanism Φ80 uses to choose between lytic replication and chromosomal integration has not been fully deciphered, it appears to favor integration far more often than comparable lysogenic phages. As a result, a new infection by Φ80 will not cause the bacterial culture to suddenly collapse and become transparent. Instead, the phage will slowly establish itself inside the population in a way that is invisible to human overseers.

Additionally, it takes very little for Φ80 to switch back into lysis mode. Whereas many prophages use DNA damage as a signal that the bacterium is about to die and it’s time to look for a new host, Φ80 is so sensitive that even the slightest hiccup in DNA polymerization (something that happens routinely when a cell replicates) can make it go lytic. As cells carrying the Φ80 prophage grow, a tiny fraction will constantly pour new phage particles into the medium. As a result, billions of these phages will have accumulated in the tube over a cycle of growth, even though the culture continues to look perfectly healthy to the naked eye.

What’s worse is that this phage is mostly active at temperatures below 37°C. But, 37°C being the standard temperature for E. coli, pretty much every measurement is done at it. A researcher could put the culture on a slide in a microscope heated to 37°C, and everything would look fine. But as they took the culture through the corridor between the shaker and the microscope, the temperature would drop a little bit, and that’s precisely when a fraction of the culture would burst into a barely-noticeable, steady stream of Φ80 particles. In other words, the phage replicates precisely when researchers are not watching.

Even so, escaping detection is only half of its strategy. Φ80 also has an ingenious mechanism to break free from its tube and travel to new samples by hitchhiking on a classic molecular biology technique, thus turning human researchers into unwitting dispersers.

Scientists as Phage Dispersers

The technique in question is called P1-transduction. It is a widely used method for copying and pasting genes from one strain of bacteria to another. To understand how it works, we need to meet yet another phage: P1.2

The key to P1’s utility lies in a quirk of phage biology. During their reproductive cycles, phages use the bacteria’s machinery to produce a massive amount of empty capsids, which then house copies of the phage’s genome. This maintains a little army of fresh phage particles ready to go out and infect the next generation of cells.

Occasionally, though, there’s a mistake in the process: a piece of the host bacteria’s genome gets incorporated into a capsid instead of the phage’s genome. While this normally happens only rarely, P1 makes this mistake unusually often, about 0.1 percent of the time.3

As the phage particles are liberated in the medium and infect other bacteria, the random bit of DNA from the original, infected cell ends up injected into a different one.

This recipient cell notices that there’s a piece of linear, double-stranded DNA floating around in the cytoplasm. (E. coli’s chromosome is circular, so the presence of linear DNA is abnormal — it usually means that the chromosome is broken). In response, the bacterium expresses a suite of enzymes that stitch homologous pieces of DNA together based on sequence similarity, hoping to reconstruct a complete, circular chromosome. Thus, if P1 takes a piece of DNA from the first bacterium, then injects it into a second bacterium, the transplanted piece of DNA will often end up incorporated into the recipient’s chromosome, overwriting the native sequence. This process is called transduction.

Transduction gives scientists a clever way to copy and paste genes from one bacterial strain into another. First, the researcher inserts an antibiotic-resistance gene in the chromosome of one bacterium, called the donor strain, next to the gene they want to move.4 Then, they culture this donor bacterium in liquid broth, grow it up, and add a bit of phage P1. The P1 phages infect the donor cells, filling them up with invisible phage particles, including a few that took on a fragment of the donor strain’s DNA.

Next, these phage particles can be used to infect the recipient. At this point, random fragments from the donor DNA will be injected into the recipient and replace the native sequence by recombination. Finally, a researcher spreads the culture over a Petri dish containing the antibiotic from the first step, along with a chemical that inhibits further infection by P1. At this point, only the cells that have received the right region of the donor’s DNA — the bit containing an antibiotic resistance gene — can survive. And so, overnight, the few cells that received the desired edit will grow to form visible colonies.

When a two milliliter culture of bacteria gets infected by P1, this quickly generates tens of billions of phage particles. 0.1 percent of them will have incorporated a piece of the host DNA by mistake, which means our tube will contain around ten million such mistakes. Each of these pieces of DNA is about 90,000 base pairs long (like the phage’s genome). So, the total amount of donor DNA (mis)incorporated by the phage population will be almost a trillion nucleotides. For comparison, the full genome of E. coli is only 5 million base pairs. Therefore, the phages in our tubes collectively contain the entire genome of the donor bacterium: in fact, hundreds of thousands of copies of it. Add this to a recipient culture with several billion cells, and it’s almost guaranteed that one of them will receive the desired segment and incorporate it in the right place.

This is what makes P1 transduction a quintessential tool for microbiology. Although it relies on a series of unlikely coincidences, the sheer number of particles that fit in a droplet of liquid means it somehow works out in the end. And as long as a genetic element has an antibiotic-resistance marker next to it, it can be easily moved from strain to strain. The most famous application of this is the Keio collection, a holding of thousands of E. coli strains, each with a different gene replaced with an antibiotic resistance marker.

This ease of use made P1-transduction a hugely popular technique, used daily by thousands of biologists around the globe. Unfortunately, its popularity also makes it a perfect conduit for a stealthy parasite like Φ80 to travel from strain to strain. This is because, by a twist of fate, Φ80 has figured out a way to “tag along” with P1 phages.

Consider a routine P1-transduction experiment, aimed at moving a gene from donor to recipient. The donor strain, unknown to the researcher, harbors the Φ80 prophage. When P1 is added to the culture, the population collapses after a few hours; the expected sign of successful phage production. But the arrival of P1 is detected by the dormant Φ80 prophages, which now replicate alongside P1 in massive quantities. The oblivious researcher infects the recipient strain with this mixed phage solution, selects for the antibiotic resistance marker, and sequences the target locus. Everything appears to have worked as expected (the desired gene has been successfully transferred), but the new strain now carries Φ80.

This is, in fact, how Φ80 was discovered in the first place. In the late 2000s, researchers at the University of Illinois in Urbana-Champaign were trying to understand how bacteria repaired their DNA after damage from UV light. They suspected that LigB, a ligase that ties strands of DNA together, might play a role.

To test this hypothesis, the researchers did a classic knock-out experiment: they used P1-transduction to remove the ligB gene, then put cultures of the original bacteria and the modified ones side by side and exposed them to UV light. After exposure, the cultures of the strain lacking LigB mostly collapsed, while the natural strain didn’t. From that, they drew the logical conclusion that LigB must somehow repair UV damage. And yet, their assumptions were all wrong; the LigB ligase has nothing to do with UV damage.

What really happened was that their donor strain was infected with Φ80, which took advantage of the P1-transduction to jump to the new strain. When this strain was exposed to UV light, the DNA damage induced Φ80’s lytic cycle, which looked exactly as if the cells were dying from the damage itself.

But why does P1 activate Φ80? Is it an ancestral relationship evolved over the course of millennia, or is it a new adaptation evolved specifically within the research lab environment? While no one has studied the question in detail, we can find some clues by digging through the scientific archive.

Φ80 was first isolated in 1961 from a collection of clinical samples from Osaka University’s Department of Clinical Evaluation. The version of Φ80 studied by the Urbana-Champaign team is genetically identical to the original, Japanese variety, so any special adaptations must have occurred before 1961. While the phenomenon of P1 transduction was already known at the time, it was not yet used as a method for genome engineering, so it’s unlikely that the “hitchhiking” of molecular biology techniques was a new trait actively selected for.

All in all, it seems likely that Φ80’s ability to infiltrate biology labs is an exaptation: a trait that was originally selected for unrelated reasons, but was later repurposed in a different way after a change of environment. The genetic diversity of phages, after all, is so immense that it is unsurprising that one of them would turn out to be perfectly suited to life in the laboratory.

What Now?

Addressing the Φ80 crisis requires overcoming three obstacles: One is technological, one is social, and one is related to the very nature of biology as a science.

The technological challenge is the most tractable. For a long time, it was only practical to sequence short regions of the genome (using the Sanger method), so researchers would only PCR-amplify and sequence the particular locus they intended to modify. This can grant them great confidence that a strain carries the expected genotype, but it’s basically useless against a phage that inserts itself in a different region of the genome.

While whole-genome sequencing has become much more accessible, it doesn’t really solve the problem. First, it is more expensive than local Sanger sequencing. Second, short-read sequencing (e.g., Illumina) only generates short segments of DNA that must then be assembled on a computer. It is possible to use these short segments to stitch together the entire genome, much like a puzzle, but it’s slightly faster to align it to a reference genome, with the drawback that only sequences that map to the reference are considered. Sadly, the latter method will only detect a hidden phage if a researcher provides the phage genome as one of those references.

In other words, it would be feasible to find Φ80 through whole-genome sequencing, but it would require making two sub-optimal choices: using a slightly more expensive sequencing run and processing the results with a slightly slower algorithm. This is why, in practice, nobody does it. Fortunately, a third generation of sequencing can help: it’s becoming increasingly popular to sequence entire genomes using long-read (e.g., Nanopore) sequencing. This method reads DNA in long segments of thousands of nucleotides, making Φ80’s 46,000 nucleotides of foreign DNA much easier to spot. Hopefully, in the near future, this technique will help us identify Φ80 invasions more readily.

In the meantime, however, it remains difficult to identify Φ80 unless actively looking for it. This brings us to the social obstacles. For a microbiologist to check their strains, they need to be aware of the phage’s existence and dedicate time to checking for it.

Imagine being a fourth-year PhD student who finally managed to measure an interesting, statistically-significant, reproducible difference between two bacterial strains. Their hypothesis makes sense. They did everything right. As far as publication goes, their findings as they are would merit publication. The painful truth is that double-checking for Φ80 contamination is guaranteed to hurt their career. At best, they spend a few hours performing an additional experiment and find nothing; at worst, they discover that their results are artifactual and no longer publishable. There is no reward for discovering that your results are false positives.5

Finally, while Φ80 contamination may be unfortunate, there is something inevitable about this sort of setback given the inherent messiness and complexity of living organisms. So while Occam’s razor is a convenient heuristic, it does not actually work that well for biology; there is always a long tail of insanely complicated speculative possibilities that, every once in a while, turn out to materialize. A stealthy phage spreading from culture to culture? It may not be the most plausible explanation a priori, but it’s plausible enough. And if it’s not Φ80, it could still be any number of improbable but real bottlenecks: autoclave-resistant RNases, traces of phosphorus in the water, novel organelles, or parasites from an entirely unknown domain of life.

This is simply part of life as a biologist. In the 1990s, cognitive scientist Kevin Dunbar spent a year conducting ethnographic observations in a molecular biology lab. According to his notes, 40 to 60 percent of experimental results go against expectation, often for totally unknown reasons. As scientists push the frontier of knowledge, phenomena that were once considered mysterious become settled science, opening the way for new experimental techniques that themselves unearth new mysteries.

This actually makes me fairly hopeful about the future of biology. When people started experimenting with E. coli in the 1940s, they didn’t know a single thing about it. Since then, there has been considerable progress, and we now have at least some information about 3803 of its 4747 genes. The more we understand it, the less room there is left for unexpected artifacts creating misleading results and false interpretations, and the easier it becomes to make progress. Many outlandish oddities, higher-order emergent interactions, and bizarre edge cases loom ahead, but the fog is lifting.

In the end, we should embrace experimental artifacts and contamination not as setbacks to our research agenda, but rather as essential parts of the systems we study. Sometimes, as in the case of Φ80, the artifacts may even have a life of their own.

Antoine Vigouroux is a systems biologist. He writes about science and progress at Telescopic Turnip.

Thanks to Jay and Jasmine for feedback.

Cite: Vigouroux A. “How Φ80 Infiltrates Research Labs.” Asimov Press (2026). DOI: 10.62211/75eh-27rw

By Brian Naughton

Over the past few months, AI-based tools have emerged that enable scientists to design original antibodies on the computer for the first time. A year ago, none could reliably do this computationally. But now companies like Nabla Bio, Chai Discovery, Latent Labs, Manifold Bio and, most recently, DeepMind-spinoff Isomorphic Labs have allowed high success rates. There are even open source tools, such as BoltzGen and Germinal, that deliver similar performance.

The rapid progress in antibody design matters because these molecules are among the most versatile tools in biology. Many medicines — including Humira and Adalimumab — are antibodies, and cheap diagnostics, including $1 COVID tests, rely on them as well. These Y-shaped proteins make excellent binders, as the two arms can latch onto proteins or other molecules and block their activity.

Before these AI tools existed, scientists searching for a useful antibody would first need to screen billions of candidates in laboratory assays to identify just a handful with high affinity for a target. BindCraft, released in 2024, changed this. For many targets, a suitable binder can now be found after just tens of attempts rather than billions. BindCraft uses the AlphaFold 2 model, but inverts it: the model creates a protein structure expected to fit onto a chosen target, then converts that “shape” back into an amino acid sequence that can be synthesized and tested in the laboratory.

Antibodies are not the only type of protein binder, and binders are generally catalogued according to their size. A “peptide” is any protein smaller than 30 amino acids, a “mini-binder” is between 50 and 250 amino acids in size, and an “antibody” covers anything from a 120-amino-acid antibody fragment to a multi-chain, 1,300 amino acid molecule.

Computational drug developers spend most of their time designing a particular type of antibody fragment, called the VHH, or “nanobody.” These are small variants of antibodies, naturally produced by llamas and alpacas, consisting of a single heavy chain. VHHs are much easier to design than a full antibody because they are about one-tenth the size and more compact. They can also be cloned and expressed in bacteria, unlike full antibodies, which require glycosylation, a chemical modification that only yeasts and mammalian cells can perform.

The main goal in binder design is to find a molecule with a high affinity to some target, meaning the binder latches on tightly and won’t let go. Such affinity is quantified by a dissociation constant, or Kd, which describes the concentration of free binder that must be present for a target molecule to have a 50 percent chance of being bound. A lower Kd means tighter binding. Picomolar (pM) and nanomolar (nM) Kds are typical for drugs, while micromolar (µM) is considered quite weak.

Semaglutide, for example, binds its target with sub-nanomolar affinity, while many natural signaling proteins, like T-cell receptors, have affinities in the micromolar range; a thousand-fold weaker. In the binder-design literature, a sub-micromolar affinity is often used as the threshold for calling a design “successful,” though this affinity would likely not be strong enough to be therapeutically useful.

To learn antibody design, most people start with BoltzGen, a tool developed by Hannes Stärk and the team at Boltz. (This is the same team behind Boltz-2, one of the leading AlphaFold 3-like structure prediction models.) BoltzGen is, arguably, the leading open-source approach for computational antibody design, and unlike many competitors, it uses the permissive MIT license, meaning anyone can use it, even commercially.

Working with multiple academic labs, the Boltz team showed that BoltzGen achieves sub-micromolar binders in a majority of cases, on both well-studied proteins like insulin and on difficult targets with no known similar structures. That said, BoltzGen is not the only option, and the landscape is shifting quickly. RFantibody can design full antibodies rather than just fragments, and Mosaic allows users to incorporate custom scoring functions.

However, none of these tools are easy to use. Designing an antibody on the computer requires navigating a huge tangle of jargon, a general understanding of the pros and cons of available software tools, and enough familiarity with wet-lab biology to know what to do after the design process has been completed. This guide walks readers through the full process of designing an antibody from home using BoltzGen.

Tool            Commercial  Full Ab  VHH  Mini-binder  Peptide
-----------------------------------------------------------------
BoltzGen            ✅        ✅      ✅       ✅           ✅
Germinal            ❌        ✅      ✅       ❌           ❌
mBER                ✅        ❌      ✅       ❌           ❌
RFantibody          ✅        ✅      ✅       ❌           ❌
Mosaic              ✅        ✅      ✅       ✅           ✅
BindCraft*          ❌        ❌      ❌       ✅           ✅

* BindCraft is not an antibody design tool, and is only included for reference. The FreeBindCraft fork allows for commercial use.

Antibodies from Scratch

The computational process involves five steps: choosing a target, preparing its structure, running a design campaign, filtering candidates, and experimentally validating the results. The guide follows a single hypothetical candidate and describes each step using the best open-source tools currently available.

1. Choosing a Target

This guide will use Nipah virus Glycoprotein G — “Nipah G” — as its protein target. Nipah G sits on the surface of the virus, a dangerous pathogen with a mortality rate of between 40 and 75 percent, and is essential for binding human cells. A binder that blocked Nipah G would presumably help prevent infection. (One antibody, m102.4, has already completed Phase I clinical trials on exactly this basis.)

Nipah G was also the subject of a recent binder design competition hosted by Adaptyv Bio, a cloud lab for protein designers. These competitions matter because they pit design tools against one another, chasing the same target under identical conditions. The Nipah G competition attracted a few hundred entrants, including the BindCraft developer, Martin Pacesa, and the Mosaic developer, Nick Boyd. Adaptyv screened more than 10,000 designs in total, making this arguably the richest publicly available dataset for comparing binder design tools to date.

This guide’s main tool, BoltzGen, did not fare well. In fact, only one percent of designs passed the experimental binding threshold. The reason for this is unclear, but every protein target behaves differently, and no tool excels across all of them. While Mosaic had the best results of any tool in the competition, with 8 out of 9 designs successfully binding the target, it requires more hands-on coding and is better suited for those already comfortable with protein design pipelines.

2. Preparing the Target Structure

Choosing a target protein is only the first step; you also need to choose a specific crystal structure to work with. A single protein can have hundreds of structures in the Protein Data Bank (PDB), each produced under different experimental conditions, with small but consequential differences in atomic positions.

Proteins also shift shape depending on their state. A protein bound to another molecule looks different from one floating freely in solution. These two forms, called the holo and apo forms, respectively, can have entirely different geometries.

The PDB is the main repository for protein structures, but its search engine is imprecise. Querying “Nipah virus Glycoprotein G” returns 32 results, many unrelated to Nipah, and the interface is hard to navigate unless you are a crystallographer.

A better repository is UniProt. If you search for your target protein there, you will find that the UniProt team has already done the hard work of linking every relevant PDB entry to each protein it contains. The entries, however, vary as to what resolution they have, which protein chains are included, and even how much of the protein is physically present in the structure. It is not uncommon for a crystal structure to cover only part of the full protein. In the case of Nipah G, the available structures cover positions 176-602. This excludes the transmembrane domain of the protein, since that region is notoriously difficult to crystallize.

Looking at the UniProt list, 2VSM seems like a reasonable choice of target. It has the highest resolution of the available options, at 1.8 Å, slightly longer than a carbon-carbon bond. (Anything below 2 Å is considered high resolution for X-ray crystal structures.) The 2VSM structure also includes the virus’s natural receptor, Ephrin-B2, which shows exactly where on the surface a binder might attach.

Once you have a possible target, see what prior research exists. In this case, there is a 2025 paper that identifies binding sites (or “hotspots”) on Nipah G suited to binder design. The first hotspot is located at residues Q559, E579, I580, Y581, and I588, exactly where the virus contacts Ephrin-B2 and where the clinical antibody m102.4 binds. The second site — at residues V235, S236, Y237, R555, and S586 — sits on another face of the protein. Antibodies that bind there do not block receptor attachment directly but prevent the virus from undergoing the conformational changes needed to enter a cell. The third site, formed by W504, F458, and L305, helps to stabilize the receptor-binding domain.

Even though 2VSM seems like a promising structure, it is still worth testing several options before committing. Small differences in atomic positions can affect the designs you get, and nearly identical structures sometimes yield distinct results for reasons that are not obvious in advance.

Predicting a Structure

When preparing the target structure, it’s important to remember that some proteins have no crystal structure in the PDB. UniProt often links to predicted structures directly, but if nothing suitable turns up, you will need to predict one yourself. For non-commercial projects, the best tool for this is AlphaFold 3.

AlphaFold 3 is straightforward to use. Log into alphafoldserver.com, enter your protein sequence, and wait a few minutes. You do not need to set any parameters. The output is a 3D view of the predicted structure and a downloadable .cif file, which is a modern version of the older .pdb format.

Protein structure prediction tools output a few different scores, or “confidence metrics,” though two come up most often. The first is pLDDT, which measures overall confidence in the predicted structure on a scale of 0 to 100. The second is ipTM, which measures confidence in the relative positions of two proteins in a complex, scored from 0 to 1. The Nipah G competition used a third metric, ipSAE, which is derived from ipTM and highly correlated with it, but tuned to better predict whether two proteins actually bind.

The AlphaFold 3 prediction for Nipah G bound to Ephrin-B2 looks good. Most of the structure shows a pLDDT above 90, the threshold for high confidence, and the ipTM is 0.9, meaning the model is confident about how Ephrin-B2 sits against Nipah G. The N and C termini score lower, but that is normal — those regions are disordered by nature. As a rule of thumb, a pLDDT above 90 and an ipTM above 0.8 are both good signs. This structure meets both of those cutoffs.

Trimming

Once you decide on a structure, it is often worth trimming it — that entails removing the parts of the protein irrelevant to the binding site. The cost of a design campaign scales roughly linearly with the combined length of the binder and target, so trimming irrelevant regions can save thousands of dollars.

The standard tool for this is PyMOL, a program for visualizing and manipulating protein structures. The open-source version can be installed via Conda, while the commercial version is available at pymol.org. Here are the specific steps you can take, using PyMol, to trim your target protein:

1. First, load the 2VSM.cif file into PyMOL by typing fetch 2VSM or clicking File / Open and selecting the downloaded file. This loads the complete complex, including Ephrin-B2.

2. To remove everything except Nipah G (chain A), type remove not chain A. To also show the protein sequence, type set seq_view, on. The red dots on the screen are mostly water atoms, which can be removed with the command remove hetatm.

3. Select and color the first hotspot with select hotspot1, chain A and resi 559+579+580+581+588; color red, hotspot1 so you can see where it is. Then, manually select residues (pink dots) by highlighting the sequence (here, I have selected positions 188 to 207.) Every amino acid after position 207 is quite close to the hotspot residues, so they cannot be safely removed.

4. Finally, remove the sequence with the command remove resi 188-207 and save the edited structure with save 2VSM_trimmed.cif and save 2VSM_trimmed.pdb. (It is often useful to have both .cif and .pdb files.)

In this particular case, removing 20 amino acids from a 400 amino acid structure is hardly worth the effort, since the Nipah G hotspot sits close to both ends of the protein. Residues can also be removed from the middle of a structure, creating gaps, though this requires more testing. Gaps can cause problems with structure prediction and change the numbering of hotspot residues in BoltzGen. For large antibody design campaigns, however, trimming as aggressively as possible is usually worth the effort.

3. Running a Design Campaign on Ariax

BoltzGen is not easy to run locally. Your options are cloning the GitHub repository and running it yourself, using a cloud service like Modal, or working through platforms like Tamarind Bio, Litefold, or Neurosnap. The BoltzGen team also recently launched their own platform, Boltz Lab, currently in beta.

This guide uses a web-based tool called Ariax, arguably the easiest way to run BoltzGen and BindCraft. Ariax charges directly for GPU time, without the need for a subscription, and the team behind it are protein-design experts whose explicit goal is to make binder design easier.

Start small and run short campaigns with 100 or fewer designs for each of your candidate hotspots first. This will give you a sense of which one looks most promising. Once you have identified the best hotspot, scale up to a full campaign — roughly 50,000 designs. (The right number depends on the difficulty of your target, your budget, and how many high-scoring designs you need. More designs means a higher chance of finding something that works.)

Setting up a run on Ariax takes only a few minutes. The Ariax team have written a helpful tutorial with additional tips, and the defaults are sensible enough for both beginners and experienced users, though not every BoltzGen parameter is exposed. If you wanted to use a custom VHH framework, for example, you would need to generate the appropriate configuration files and run BoltzGen yourself. The BoltzGen documentation covers this, though getting GPU-dependent tools like BoltzGen running locally can be genuinely challenging.

To run a Nipah G campaign on Ariax, follow these steps:

1. Select VHH as the design scaffold, then upload the 2VSM_trimmed.pdb file created earlier. Although .cif files have largely replaced .pdb as the standard format, .pdb is still the safer choice here as .cif files exported from PyMOL are not always compatible with BoltzGen. If you need to convert between the two formats, use maxit, a conversion tool maintained by the PDB team.

2. Under Binding Rules, set “binding” to positions 349, 369, 370, 371, and 378. Note that these are not the original residue numbers, and this is an awkward but persistent problem with both .pdb and .cif files. After trimming, the 2VSM structure starts at position 211, but BoltzGen always counts from position 1. Ariax helpfully notes in its log that it has applied an index shift of -210, so to match BoltzGen’s numbering to the original, subtract 210 from each residue number. Visually check the hotspot positions in Ariax, which highlights them in yellow, to confirm they land on the expected residues.

3. You can optionally set “not_binding” positions to steer the design away from parts of the protein you want to avoid — glycosylation sites, for example, or regions far from the binding interface. This option is most useful for larger proteins with multiple domains; for Nipah G, the binders tend to land close to the target site without much guidance.

4. Set the # Designs to 50 and the Budget to 50. This just means you want to generate 50 designs, and return information for all of them. At this stage, it’s important to see the full distribution of scores across all designs, rather than just the top hits. But for the full campaign, when you scale up # Designs to 50,000, you’ll set the Budget at 100 or below to get back just the top scoring designs.

5. For GPU Selection, Cost or Performance mode both work fine for a small run. The only obstacle that can arise is if the protein is too large to fit in the GPU’s memory — for proteins above 400 amino acids, only the latest GPUs, such as the B200 and H200, may be able to handle it. For the full campaign, select Turbo Mode to run multiple GPUs in parallel (otherwise, the run can take days.) There is no real cost penalty for Turbo Mode, and the faster premium GPUs are not necessarily more expensive overall since they finish sooner.

6. Finally, click Validate Settings to check for errors, then click Start BoltzGen.

Before the campaign starts, you will also need to fund your Ariax account. Generating the 100 Nipah G test designs costs around $10. A full 50,000-design run would cost roughly $3,000-$6,000 for a protein of this size, since Nipah G is around 400 amino acids in length. The main way to lower costs is to trim the structure further. (Remember, though, that any time you remove part of the protein, you risk subtly altering its shape.) You also need to keep track of how the residue numbering shifts as a result.

Scoring

Scoring and ranking designs is one of the most involved — and most confusing — parts of binder design. BoltzGen outputs many metrics, including ipTM, predicted hydrogen bonds, and surface accessibility, combining them into a single “Quality Score” using a heuristic formula. The team behind PXDesign did a comprehensive review of scoring across BindCraft and other tools, and landed on thresholds of ipTM > 0.85, pTM > 0.8, and complex RMSD < 2.5Å as a reasonable way to separate binders from non-binders. These thresholds work fairly well, though as far as anyone can tell, no single metric or combination of metrics reliably predicts affinity.

There is no consensus on the best way to score designs. ipTM is a far-from-perfect predictor of binding, but it is a robust enough metric for comparing designs across hotspots. Previous Adaptyv Bio competitions ranked designs by ipTM, and although the latest competition used ipSAE instead, the two scores are highly correlated. For our Nipah G example, the differences between hotspots are large enough to be genuinely informative; the first hotspot — the Ephrin-B2 contact site — seems to be best, with an average ipTM of 0.68, compared to 0.56 for the third hotspot and 0.4 for the second.

You can also run BoltzGen without specifying a hotspot at all, letting the model find its own binding site. This approach worked remarkably well for Mosaic in the Nipah G competition, and Nick Boyd discusses it in more detail on the Escalante blog.

For the full campaign, it is worth starting with around 1,000 designs to check the cost per design and confirm that the designed binders look sensible. If the top designs from that first batch are not engaging with the intended hotspot, something has gone wrong; adding “not_binding” residues can help steer the model away from wherever it is mistakenly landing. Once everything looks reasonable, scaling up is easy. On Ariax, just click “Clone & Reuse,” adjust the number of designs to 50,000, and start the run.

4. Filtering and Selecting Candidates

For this example, the campaign ran to 1,000 designs. But the process for evaluating the top designs is the same regardless of how many you generate.

The best binder from this run, ranked by BoltzGen’s heuristic score, has an ipTM of 0.78. ipTM is not an intrinsic property of the binder itself, but a metric of Boltz-2’s confidence in the predicted binding pose. A different model, generating exactly the same binder design, might give a different score for the same metric.

AlphaFold 3 is generally thought to be the best structure prediction model overall and especially strong with antibodies and nanobodies. For greater confidence, you can run the same binder design through alphafoldserver.com as an independent check. A “good” design will typically get good ipTMs from both Boltz-2 and AlphaFold 3.

In PyMOL, load both .cif files, align them, and compare the BoltzGen pose to the AlphaFold 3 pose against the Nipah G structure. Since Nipah G is in the PDB, both structures should align closely. In this case, BoltzGen gives the design an ipTM of 0.78, while AlphaFold 3 gives its slightly different pose an ipTM of 0.73.

Several things about this design are worth noting. First, the Boltz and AlphaFold 3 poses are very similar. (If the two models had predicted different binding poses, the design should probably be rejected.) Second, the binding pose is next to the hotspot, and binding appears to be driven by the VHH’s hypervariable regions (called “CDRs”). It is not uncommon to have designs where, instead, the VHH binds “side-on.” This is not necessarily a bad thing, but it reduces the probability that it is a specific binder, only binding to the Nipah G target.

Third, the ipTM scores seem reasonable. It would be better if both were above 0.8 or even 0.9, but 0.73 is a decent ipTM from AlphaFold 3. And finally, the amino acid sequence for the designed binder looks plausible. Sometimes, binder design tools will output suspiciously low complexity sequences, like long runs of glycines or glutamates.

Computationally-designed proteins may also contain stretches of cysteines, which are likely to cause aggregation or clumping. A more thorough check would involve running the sequence through an antibody language model like AbLang, which scores how “antibody-like” a sequence looks.

On balance, this design does not score especially highly, but it may still be worth testing. Further computational tests could be done, though the returns diminish quickly. The only way to know whether a design actually binds is to test it in the lab.

5. Experimentally Validating the Results

Once you have selected your most promising designs, the next step is to test them in the lab. Almost everyone working in protein design uses Adaptyv Bio for this — the same group that ran the Nipah G competition. Adaptyv is a modern contract research organization, which in practice means you can go online, submit a list of protein sequences, and pay by credit card without a sales call. Adaptyv makes a small quantity of each design in a cell-free system and returns binding affinity data against your target in a few weeks. It is the closest thing available today to a true cloud lab.

Each design costs $119-215 to test, depending on how many you submit. Assuming a full campaign of 50,000 designs and testing the top 50, the total comes to roughly $4,000 for compute and $12,000 for testing — plus a few hundred dollars for Adaptyv to acquire the target protein. Careful filtering before submission can reduce these figures.

Published results from groups like Nabla Bio, Chai Discovery, Latent Labs, BoltzGen, Germinal, and mBER suggest binder design success rates (defined as finding at least one sub-micromolar binder from ten tested designs) of up to 66 percent. But the messy reality, based on wet-lab testing, is that the true number is probably far lower.

The most comprehensive dataset available for a single target binder is the Nipah G competition, where the success rate was under 10 percent. Mini-binders appear to be easier to design computationally than VHHs and, anecdotally, the success rate for tools like BindCraft and PXDesign across a range of proteins is probably a bit above 25 percent. Generating high-affinity binders remains genuinely difficult, though, and success is highly target-dependent.

The lowest-cost version of this process that could plausibly produce a binder would be around $1,000 of compute — enough for 10,000 or more BoltzGen designs — followed by testing the top ten at Adaptyv, for a total of roughly $4,000. That is the floor, and even then there’s no guarantee a “true” binder will be found in those ten molecules.

The Last 90 Percent

With some iteration, this process should yield a VHH that binds your target at below one micromolar — a real result. That said, a working binder is closer to the beginning of the story than the end.

If the binding affinity is not high enough, improving it computationally is not straightforward. The most reliable approach is to brute-force a solution, perhaps using saturation mutagenesis or a similar method. This lets you exhaustively test sequences close to your binder in the hope that one or more show improvement.

Specificity is the next hard question. Does your binder attach only to its intended target, or does it stick to other proteins too? This is difficult to measure, partly because you cannot know in advance which off-target proteins to test. Testing even ten candidates at Adaptyv would be expensive, because the platform is set up to test one target against many binders, not the other way around. Each additional target protein would need to be sourced separately, from suppliers like Acro Biosystems, at a cost of $300-1,000 per protein.

Beyond affinity and specificity, there is also a long list of other properties to check, depending on how the antibody will be used. For therapeutics, it will likely need to be non-immunogenic, meaning it doesn’t fire up the immune system, and have a reasonable half-life, so it circulates in the body for a long time. For in vitro applications, like making a biosensor to detect botulinum toxin, there are far fewer requirements.

For drug development, most of the major costs come long after the binder design stage, emerging during clinical trials, for example. The opportunity exists, therefore, to optimize binding alongside other properties, like immunogenicity, thermostability, and half-life, during the computational design process. Theoretically, if you can optimize binding in tandem with these other properties, drug development costs could come down.

The real gains from these new computational tools probably lie, not in shaving weeks off existing timelines, but rather in making things possible that weren’t before.

David Baker’s group, for example, recently demonstrated “facilitated dissociation” — a binder that releases its target when a second ligand is added. Looking ahead just a year or two, it is not hard to imagine routinely designing binders that engage multiple independent targets, contain no immunogenic fragments by design, or change their structure and properties in response to pH or other environmental signals. That kind of complexity is only achievable through AI-guided design, and we are just at the beginning of it. Guides like this should facilitate the process.

Subscribe now

Brian Naughton is a computational biologist working in protein design and AI. He is co-founder and CTO of the personalized cancer company, Decade. Previously, he was co-founder and CTO at Hexagon Bio, and Founding Scientist at 23andMe. Brian has a PhD in Biomedical Informatics from Stanford University.

Cite: Naughton, B. “How to Design Antibodies.” Asimov Press (2026). DOI: 10.62211/58wh-12qp

By Ulkar Aghayeva

From the late 1980s through the early 2000s, a famous debate played out between the evolutionary biologists Stephen Jay Gould and Simon Conway Morris on the nature of the evolutionary process.

Gould viewed evolution as radically contingent: if “the tape of life” were to be rerun, he argued, even small environmental changes would result in widely divergent outcomes. The likelihood that any major biological innovation, such as multicellularity, photosynthesis, or intelligent life, could evolve again would be vanishingly small, since each of these innovations depended on a combination of rare events.1 The crux of Gould’s idea was that “no important and sufficiently specific evolutionary outcomes are robustly replicable.”

Conway Morris agreed that historical contingencies are pervasive in evolution; however, in his view:

…contingency is inevitable, but unremarkable. It need not provoke discussion, because it matters not. There are not an unlimited number of ways of doing something. For all its exuberance, the forms of life are restricted and channeled.2

In other words, if we imagine life forms having emerged as solutions to problems in a global search space, there are only a limited number of good solutions given existing environmental constraints. Biology is beholden to the laws of physics and chemistry, which limit which solutions are feasible, let alone optimal.

This is why, as species evolve under similar selective pressures toward greater fitness, they acquire similar characteristics. Indeed, even though evolution is contingent at a local level (such as a specific protein sequence or the shape of a flower), it is remarkably predictable at a global level (such as the very existence of proteins and flowers across many species).

Convergent evolution reflects, then, a widespread predictability of life’s design solutions. Similar biological forms and functions often emerge independently in unrelated lineages. For example, ice growth-inhibiting antifreeze proteins evolved independently in Arctic and Antarctic fishes. Although arising at opposite ends of the globe, these proteins converged at both the sequence and structural level: at both locations, they contain a repeating tripeptide motif of glycine–alanine (or proline)–threonine, with the last amino acid fused to a sugar. The severe constraint of survival in freezing seawater pushed these phylogenetically distant fishes to evolve similar molecular solutions.

Similarly, C4 photosynthesis has been documented in more than 60 different plant lineages, including maize, sugarcane, and papyrus. This convergent evolution involves not just a single protein, but a complex mixture of biochemical and structural adaptations that together enable a more efficient use of carbon dioxide.

Nowhere does the power of convergent evolution appear so insistently and compellingly, though, as in the structure of viral capsids, the self-assembling protein shells enclosing viral genomes. Though ranging in size from 20 nm to 800 nm, an estimated 70 percent of viral capsids known to date are icosahedral, shaped like tiny soccer balls.3

This convergence of viral capsid forms is not merely a curiosity of nature. By deciphering viral capsid architecture at the molecular level, scientists have found inspiration for designing more efficient drug delivery vehicles and vaccines. Viral capsid-like protein nanocontainers have provided an excellent solution to the design problem of carrying and delivering a variety of biological cargoes.

But why did viruses converge on icosahedral shapes in the first place? To answer this, we must turn to both genetics and structural biology.

Evolutionary Constraints

From a genetic perspective, viruses must be economical with how much space in the genome they allocate to encoding capsid proteins, as well as all the other functional genes required for replication and virulence. James Watson and Francis Crick termed this the principle of genetic economy. Such a constraint favors highly symmetrical capsid morphologies so that the same capsid proteins, or a very small number of them, can be used repeatedly for building the capsid, rather than coding for a large number of unique proteins.

For example, the hepatitis B virus has only four genes in its genome,4 of which one encodes the capsid protein. An icosahedron has 60 different rotational symmetrical operations (more than any other platonic solid), and so a capsid with icosahedral symmetry significantly reduces genetic costs.

Second, from a geometric perspective, an icosahedron comes closest to a sphere than other platonic solids.5 This grants it the largest volume for a given surface area or, equivalently, for a fixed number of capsid coat proteins. With such high geometric efficiency, more genomic material can be packaged into an icosahedral capsid than into any other shape, be it a cube, tetrahedron, or octahedron.6

The volume of a viral capsid is important not only because of the size of the viral genome, but also because nucleic acids carry a self-repulsing negative charge that must be physically contained. These repeating negative charges are distributed along the length of the nucleic acid molecules, and in solution, they are surrounded by positively charged ions like Na⁺ and Mg²⁺ or by more complex molecules like polyamines, which screen the negative charge of the phosphate backbone.

Capsid proteins often display positive charges on their inward-facing surface that bind the genome, thus partially displacing those ions and facilitating capsid assembly and the condensing of the genome inside the protein shell. In viruses with stiff, double-stranded DNA genomes, the bending of the nucleic acid chain imposes an additional internal pressure amounting to many atmospheres.7

Icosahedral capsids are well suited to accommodate these physical constraints because, due to their quasi-spherical shape, they distribute stress more evenly than polyhedra with fewer faces. This becomes especially important when the virus must survive extreme conditions like drying, pH shifts, or mechanical stress.

In addition to its genetic economy and geometric optimality, the icosahedral packaging of viral capsids is also energetically favorable, such that the assembly of many capsids is highly efficient and spontaneous (it even happens in vitro!).8

The Caspar-Klug theory explores this energetic aspect of viral capsid structure. Its inspiration comes from modern architecture, specifically Buckminster Fuller’s geodesic domes. The same principles of structural stability that work for these domes also apply across orders of magnitude of sizes, from molecular to macroscale, and even down to the atomic level.

The story of how this theory came about thus involves not only structural biologists but also architects, artists and chemists.9

The Structural Biology of Viruses

Donald Caspar grew up in Ithaca, New York, in a family of chemists. He first learned about the tobacco mosaic virus (TMV) at the tender age of ten from the renowned crystallographer Isidor Fankuchen, who frequented his family home.

TMV was the first virus ever discovered, in the late 19th century, and its molecular structure was still unknown when Caspar heard about it in the 1930s. However, it was known to plague the leaves of valuable crops, such as tobacco, tomatoes, peppers, causing leaves to curl and the plants, eventually, to die. Understanding its structure, besides being an exciting problem in basic biology, also promised a path toward fighting the virus.

Caspar’s fateful encounter with Dr. Fankuchen decided his professional path — he went on to study TMV and wrote his doctoral dissertation on “The Radial Structure of Tobacco Mosaic Virus,” explaining how this rod-shaped virus had a helical arrangement of capsid proteins along the length of its RNA genome. After completing his doctorate, he moved to the UK in the 1950s to continue his study of viral structures.

Caspar was not the only scientist attempting to solve the structure of TMV. Indeed, many of the greatest thinkers of the time converged at the University of Cambridge because it had the most powerful X-ray sources for crystallographic imaging, crucial for the structural studies of viruses. James Watson was primarily interested in TMV in hopes of solving the structure of RNA.10

Meanwhile, at the University of London, Rosalind Franklin made excellent X-ray diffraction photographs of TMV capsids, albeit without RNA. A fortuitous encounter with these photographs converted another physicist and crystallographer, Aaron Klug, who had just begun his postdoctoral studies in London on a ribonuclease structure but abandoned it in favor of viruses.11

Caspar met Franklin and Klug in 1955, and initially both Caspar and Klug worked independently on different small “spherical” plant viruses, tomato bushy stunt virus (BSV) and turnip yellow mosaic virus (TYMV), respectively. Caspar’s initial analysis revealed a surprising result: the diffraction patterns of the BSV capsid crystals had a non-crystallographic, five-fold symmetry. This pointed to an icosahedral structure, since each of the 12 vertices of an icosahedron is a convergence point of five edges.12

Crick and Watson had previously hypothesized icosahedral symmetry in spherical viruses, but Caspar’s results provided the first experimental evidence for it.13 A little later, Klug and Franklin reported similar results with TYMV. Even though Caspar and Klug worked on similar research projects, their collaboration wouldn’t start until 1958, after Rosalind Franklin’s untimely death.14

The original model of the viral capsid structure suggested by Watson and Crick, in 1956, viewed the capsid as a curved, two-dimensional “surface crystal” (as opposed to the space-filling 3D crystal), closed in on itself and made of identical subunits — capsid proteins. As each subunit was identical, they interacted in the same way, as in true crystals. But such a model only permitted a maximum of exactly 60 subunits, since the icosahedron has 20 triangular faces, and each triangle can be made of three protein molecules.

Advances in electron microscopy of viral capsids, however, particularly the negative staining method,15 revealed that many, if not most, viruses have a more complex structure than could be explained by the Watson-Crick model. They appeared to possess far more than 60 identical subunits per capsid. A new mathematical theory was needed to account for this seeming anomaly. The solution came not from biology, but rather from architecture.

Birth of the Caspar-Klug Theory

John McHale, a London-based avant-garde artist, learned about the structure of the poliovirus (another icosahedral virus Klug studied with his colleague, John Finch) from a newspaper, The Observer. As a member of an art movement that looked for inspiration in science and popular culture (“Independent Group,” a precursor to “pop art”), he had a keen interest in the structures of Buckminster Fuller — by then a renowned, if eccentric, architect.

McHale immediately saw a connection between the icosahedral poliovirus and Fuller’s geodesic domes, hemispherical polyhedra made of rigid triangular elements, which also had icosahedral symmetry. McHale arranged a meeting between Fuller and Klug in July 1959. Despite Fuller’s impenetrable prose, Klug carefully studied his Synergetic Geometry, an unpublished manuscript of what turned into Fuller’s magnum opus. Even with such preparation, their first meeting didn’t result in insights beyond their shared focus on icosahedral symmetry.

In 1960, Caspar and Klug, now separated by the Atlantic, both read a new book by the author Robert W. Marks on Fuller’s work, The Dymaxion World of Buckminster Fuller, and closely studied the structure of geodesic domes. Caspar also met Fuller in person while the latter was serving as a Professor of Poetry at Harvard University in 1962.16

Both Klug and Caspar were particularly interested in the domes’ light yet sturdy structures, consisting of a variable number of hexagons and the mathematically required and invariable 12 pentagons.17 This exact number is required by Euler’s rule to form a closed polyhedron. With only hexagons, the structure won’t close in on itself and will remain flat.

The framework of geodesic domes is made of rigid rods, or struts. Since each of them can be a part of either a hexagon or a pentagon, their connections are not identical due to the different angles between the sides of those two types of polygons. Even so, the interactions between the struts are similar enough across the entire framework to consider them “quasi-equivalent.”

This key concept, introduced by Caspar and Klug, lies at the heart of their theory of viral capsid structures. Like struts in geodesic domes, identical viral capsid subunits could occupy quasi-equivalent local positions, being a part of a hexamer (a ring of six subunits) or of one of the obligatory 12 pentamers (a ring of five). Since the connections between the subunits do not need to be identical, more than 60 subunits per capsid can easily be accommodated, resolving the mystery which Watson and Crick’s model couldn’t explain.

But the Caspar-Klug theory imposed its own constraints on the permissible number of protein subunits in a viral capsid. It introduced a parameter, T, called the triangulation number18: if each face of the icosahedron is divided into triangles (triangulated), then T describes their quantity.19 Since each of these triangles, in turn, can be made up of three capsid proteins, and an icosahedron has 20 faces, the total viral capsid protein numbers thus become “quantized” in multiples of 60 (60T).

The Caspar-Klug theory thus makes a strong prediction: since each triangular face of the icosahedron can accommodate only a certain number of triangles (1, 3, 4, 7…), only certain multiples of 60 are likely to be observed in viral capsids: 60, 180, 240, 420, and so on.20

The pair presented their research in June 1962, at the Cold Spring Harbor Symposium on Basic Mechanisms in Animal Virus Biology, in a now-famous paper titled “Physical Principles in the Construction of Regular Viruses.” In it, they acknowledged that their model of viral capsid structure “was, in fact, inspired by the geometrical principles applied by Buckminster Fuller in the construction of geodesic domes.”21 22 They even included a photo of Fuller’s dome in Montreal, the Biosphère. Besides introducing the principle of quasi-equivalence and the triangulation number, they also coined the term “self-assembly” to describe the spontaneous nature of viral capsid assembly:

Self assembly is a process akin to crystallization and is governed by the laws of statistical mechanics. The protein subunits and the nucleic acid chain spontaneously come together to form a simple virus particle because this is their lowest energy state.23

For several decades, the Caspar-Klug theory faithfully described all the known viral structures. But with the development of higher-resolution imaging techniques like cryo-electron microscopy, exceptions soon appeared to the strict constraints on the virus capsid composition (60T subunits) predicted by the theory. The first detected outliers were the cancer-causing polyoma and papilloma viruses, which have capsids made up of 72 pentamer facets.

It wasn’t until the early 2000s that mathematical biologist Dr. Reidun Twarock, at the University of York in England, started working out an expanded mathematical framework to explain these anomalies, called viral tiling theory. It retains the Caspar-Klug theory as a special case but also accommodates a broader range of viral capsid architectures.

Unlike hexagons, pentagonal faces cannot be evenly divided into equilateral triangles; neither can they tessellate, or tile, a 2D plane without gaps or overlaps. To account for this, Twarock turned to Penrose tilings, which can, in fact, tile a plane with five-fold symmetry but with a combination of several distinct shapes; one instead uses “kites” and “darts,”24 or kites and rhombs, all of which are four-sided. Such tilings are non-periodic (made of non-identical types of shapes) but still produce long-range order, as in quasi-crystals. Kites and rhombs represent different types of biological interactions — within trimers and dimers, respectively.

A beautiful mathematical result obtained by Dr. Twarock and colleagues says that these non-periodic tilings, representing flattened capsid structures, can be mapped onto a particular subset of the so-called Archimedean lattices. These are tilings of the plane with a combination of polygons making up a repeating unit (like kites and darts). One of them is the regular hexagonal lattice, and it corresponds to the classic Caspar-Klug construction.

But there are three more tilings within Archimedean lattices that were previously undescribed in the context of viral capsid structures. They can be represented as tilings where hexagons are separated from each other by different sets of repeating polygons of other shapes — triangles, squares, or a combination of the two. Such tilings serve as much better models of more complex viral capsid structures, previously impossible to explain within the framework of the Caspar-Klug theory.

In larger viruses, the capsid shell often consists of more than one type of protein. Major capsid proteins can form the familiar icosahedral lattice while minor capsid proteins occupy a second set of configurations not captured by the original Caspar-Klug theory. Herpes simplex virus, for example, has a T=16 capsid, in which the major capsid protein forms pentamers and hexamers, while two other capsid protein types form groups of three molecules that sit in quasi-equivalent locations between them.

A Bacillus phage, with the ominous name Basilisk, is another well-studied case of this generalized quasi-equivalence. Even though such a complex capsid geometry comes at high additional genomic and assembly costs, its composition confers new properties, such as altered stability and different mechanisms of genome release within the host cells, potentially increasing their overall fitness.

Icosahedral Structures Beyond Viral Capsids

Even within this expanded repertoire of structures, icosahedral symmetry remains a constant, a remarkable testament to its apparent optimality as an evolutionary solution to efficient viral genome packaging and protection. Viruses across multiple, phylogenetically unrelated, families appear to have repeatedly rediscovered the icosahedral architecture, making it an exemplary case of convergent evolution that Conway Morris extolled and defended against the radical contingency view of Stephen Jay Gould.

The exact sequences of viral capsid proteins may vary widely, consistent with the “local” contingency of the evolutionary process. But even with this divergence, the tertiary and quaternary structures of capsid proteins and capsid structures of myriad viruses are remarkably similar, illustrating the “global” predictability of evolution. This predictability may be particularly defensible in the case of viruses, since they represent life stripped to the essentials of replication and assembly, and evolving under extreme geometric and genomic constraints.

But geometric constraints of this kind — when a sizable cargo needs to be efficiently packaged inside a container — seem to be at work not only in viruses but across a wide range of biological structures, from the molecular to organismal levels. Indeed, icosahedral symmetry has been found in bacterial microcompartments like carboxysomes (CO₂ fixation organelles in cyanobacteria25), nanocompartments like encapsulins (self-assembling protein shells carrying molecular cargoes), lumazine synthase (a giant enzyme complex catalyzing riboflavin synthesis), and some of the unicellular eukaryotes with mineral skeletons like Radiolarians.26

Icosahedral architectures have also inspired countless artificial protein cages and self-assembling protein nanoparticles for drug delivery and vaccine design. Virus-like particles can self-assemble from proteins in cell culture and are completely harmless, since they lack the viral genome itself or carry a greatly modified version of it that eliminates virulent genes. They can be modified to express antigens on the surface of the capsid, mimicking the repetitiveness and geometry of natural host-pathogen interactions. This grants them high immunogenicity by effectively engaging the receptors of the antibody-producing B-cells in the host organism (which respond more strongly to repetitive antigens). Indeed, two recently WHO-approved highly efficient malaria vaccines, RTS,S/AS01 and R21/Matrix-M, are based on virus-like particles.

In drug delivery, both virus-like particles and artificial protein cages of similar design are used to carry therapeutic loads like enzymes, small molecules, or nucleic acids. Thanks to the relative ease of genetic engineering of viruses, it’s possible to generate libraries counting thousands of virus-like particles. These can be mutated or chemically modified on the surface, to ensure target specificity, or on the inside, to hold onto different molecular loads.

Fully de novo computationally designed protein cages can extend beyond what is typical in nature in terms of their stability, porousness, or even programmable disassembly in response to various conditions. Even so, the high symmetry and repetitive near-identical local interactions afforded by the icosahedral symmetry are still strongly favored in these novel engineered protein nanomaterials. Under exacting physical and informational constraints, both evolution and human engineering have converged on the same shape — a strong attractor in the space of biological containers.

Subscribe now

Ulkar Aghayeva is a science writer and a columnist at Asimov Press. She also writes about science history on her blog Measure for Measure and about music history and cognition on The Bass Line.

Cite: Cite: Aghayeva, Ulkar. “Why Are Viral Capsids Icosahedral?” Asimov Press (2026). DOI: 10.62211/24ou-76ew

By David Brzostowicki

In November 1272, on the Island of Hares — home to both a nunnery and the Hungarian king's summer residence — a nobleman was lured to a purported council meeting, only to be murdered. His name was Béla, Duke of Macsó, a province nestled along the Danube in modern-day Serbia. He was in his early to mid-twenties, childless, and, depending on whom you asked, either a stabilizing force in a fractured kingdom or a dangerous threat.

Through his mother, Princess Anna of Hungary, Béla was reputed to be the grandson of King Béla IV and a member of the House of Árpád, the dynasty that had ruled Hungary for three centuries. Through his father, Rostislav, Prince of Halych, he claimed descent from the Rurikids, the royal house of Kievan Rus’, the predecessor to modern Russia. If true, he bridged two of medieval Europe’s most powerful dynasties.

For over a decade, King Béla IV and his heir Stephen V had fought for control of Hungary. After Stephen V won a decisive victory at the Battle of Isaszeg in 1265, forcing his father to cede the eastern half of the kingdom, Duke Béla found himself on the losing side of a civil war.

When King Stephen V died unexpectedly in 1272, his son, Ladislaus IV, was only ten years old. As the oldest male cousin of the child king, Béla stood next in line to the throne, eligible to serve as regent, or even king, if anything happened to Ladislaus IV. He controlled vast estates in southern Hungary and commanded loyalty from powerful allies. To some nobles, he represented order in a kingdom teetering on chaos. To others, including Ladislaus IV’s mother, he was a threat.

Henrik Kőszegi, head of the powerful House of Héder, also had a personal vendetta against Béla. Once Béla’s mentor, the two had fought side by side in Hungary’s civil wars. But at the Battle of Isaszeg, Béla fled the field while Kőszegi was captured. His subsequent years of imprisonment gave Kőszegi ample reason for resentment.

Various chronicles report that after being invited to the council meeting in the nunnery, Béla was ambushed and murdered. One account states that Kőszegi himself was present and accused Béla of treason before cutting him “to pieces” during the ensuing argument. Another states that it was Kőszegi’s mercenaries who “hacked [Béla] to pieces with swords and maces so that his skull split in two.”

Béla’s sisters, both nuns on the island (today known as Margaret Island), were rumored to have recovered and buried his mutilated corpse within the nunnery walls.

There, the body rested quietly until the spring of 1915. Then, during an excavation of the monastery’s sacristy, archaeologists discovered the skeleton of a young man, buried with his head toward the altar, which bore the marks of extraordinary violence. Anthropologist Lajos Bartucz asked to analyze the remains at the Institute of Anthropology at the University of Budapest.

Twenty-three cut marks scored the bones. Far too many for a duel, these were distributed across the skeleton in a pattern that suggested multiple attackers striking from different directions. The wounds indicated that the assault had continued even after the victim fell, with the body further mutilated while on the ground. Bartucz estimated the skeleton belonged to a male between 20 and 25 years old. Based on the body’s location, wounds, and age, he proposed them to be the remains of Duke Béla of Macsó.

But Bartucz never formally published this analysis. Instead, the bones went into a wooden box with a slip of paper reading, “butchered skeleton of Béla,” and were shuffled (skull and long bones separately) between institutions for decades.

In 2018, interest resumed when researchers at the Hungarian Natural History Museum found postcranial bones that had been missing from the skeleton for eighty years, tucked away in the box with Bartucz’s handwritten note. Zsolt Bernert and Ágota Buzár published preliminary findings on the several long bones, fragments of the pelvis, and vertebrae scored with cut marks.

When biological anthropologist Tamás Hajdu saw their report, he realized exactly where the skeleton’s missing skull was: right near his office at the Department of Anthropology at Eötvös Loránd University in Budapest. It had been misidentified in the Aurél Török Collection, the department’s own repository of skeletal remains.1

Hajdu and colleagues reunited the skull with the rest of the remains and launched a full investigation into the individual’s identity and the circumstances of his life and death.

Their conclusions were finally published in October 2025. They not only confirmed that the skeleton was Béla’s but also demonstrated his dual royal lineage and clarified the number of attackers and types of weapons used. Additionally, they suggested that the murder bore the hallmarks of personal grievance, not just political calculation.

Indeed, this study was the first ancient DNA-based identification of a medieval royal, resolving a century-old archaeological question through a convergence of the suite of scientific techniques known as “complex bioarchaeology.” Its toolkit includes skeletal morphology for biological profiles, stable isotope analysis to read the chemistry in bones and teeth revealing geographic origins and diet, radiocarbon dating for temporal placement, ancient DNA sequencing to identify individuals and establish family relationships, and forensic trauma analysis for reconstructing violence.

Taken together, “complex bioarchaeology” performs a kind of data resurrection, able to bring back individuals like Béla about whom records disagree, as well as those who didn’t merit sufficient attention by chroniclers or, more insidiously, whose recorded lives and deaths were intentionally altered.

The manipulation of historical records is, after all, as old as record keeping itself. Pharaoh Akhenaten’s name was so thoroughly erased after his death that he does not appear in the Abydos King List of Seti I, composed less than a century later. His existence was only rediscovered in the 19th century when archaeologists excavated his abandoned capital at Amarna.

Emperors were likewise erased in ancient Rome. In 211 CE, Caracalla had his co-ruling brother Geta murdered by Praetorian Guards. After executing 20,000 of Geta’s supporters, Caracalla ordered his brother’s name and image excised from every inscription and monument across the empire. Where bronze letters had once spelled out “the Most Noble Caesar Geta” on the Arch of Septimius Severus, Caracalla left text celebrating only himself.

Such erasures and falsifications continue into the modern era. In 1940, Soviet secret police massacred nearly 22,000 Polish citizens in and around the Katyn Forest, then falsified official records to blame Nazi Germany, a lie the Soviet Union maintained until Mikhail Gorbachev acknowledged the truth in 1990.2

While complex bioarchaeology can’t resolve all debates or discrepancies in the historical written record, by unearthing and analyzing physical evidence independent of records, it can help researchers compare what’s written with biological data, and even recover the lives of ordinary people who never merited a chronicler’s attention in the first place.

Duke Béla’s case offered a rare chance to set written records against physical evidence. He was royal, so his murder was chronicled. And his body was preserved and recently rediscovered, so it could be analyzed. What follows is the story of how researchers solved a 750-year-old murder.

Biological Anthropology

To begin their investigation, the team needed to answer a basic question: did this skeleton match details already known about Duke Béla? Historical records indicated he was a young man, probably in his early twenties, from a privileged Hungarian family.

First, then, they needed to determine whether this skeleton belonged to a young man. Estimating a skeleton’s sex relies on multiple features that show statistical tendencies rather than absolutes; even so, the pelvis can be a strong indicator. In females, to aid in childbirth, the pelvic inlet, the circular opening in the upper area of the pelvis, is wider. And the pubic angle, the upside-down V-shape or arch at the very bottom of the pelvis, is broader. In contrast, males have narrower, more angular pelvises.

This particular skeleton had the narrower pelvis of a male. The skull likewise bore masculine features: a prominent brow ridge, a more pronounced external occipital protuberance at the back of the head, and heavier bone overall.

To determine age, anthropologists look at growth plates. In younger individuals, bones are still fusing together. The clavicle, for instance, doesn’t fully fuse to its growth plate until a person’s mid-to-late twenties. The pubic symphysis, the area where the two halves of the pelvis meet at the front, changes texture throughout adulthood, becoming rougher and more porous over time. The cranial sutures, those zigzagging seams where the plates of the skull join together, gradually close with age.

The skeleton showed partially fused growth plates, relatively smooth pubic surfaces, and open cranial sutures, all consistent with a young adult male. But the team was more precise than that, noting that the head of the femur and the medial end of the clavicle, areas that finish fusing in the early-to-mid twenties, showed growth had only recently stopped.

Based on the sternal ends of the ribs and the pubic symphysis, they estimated the man’s age at death as twenty-three, give or take a year. If Duke Béla was murdered in November 1272, that would place his birth around 1249 or 1250 — consistent with extant historical records, which note only that he was born sometime between 1243 and 1252.

There were also things the skeleton didn’t show, the absence of which helped identify its possessor’s place on the social hierarchy. For example, the spines of agricultural workers develop degenerative changes from decades of bending and lifting. Soldiers and manual laborers build up extra bone at the attachment sites of heavily used muscles, showing what are called “entheseal” changes, where repetitive stress creates tiny ruptures that heal and accumulate over time.

But this skeleton showed no such markers, its vertebrae lacking the osteophytes (bony projections) that would have appeared as a result of chronic spinal stress. The muscle attachment sites looked normal, without “facet hypertrophy” (enlarged bony growths on the facets, the small joints connecting the vertebrae). These develop in someone who has spent their life hauling goods or wielding weapons. Whoever this young man had been, he hadn’t worked like a medieval peasant, a common soldier, or a person of the lower classes.

To round out the profile, they had to estimate the skeleton’s height. The long bones, including the femur, tibia, and humerus, were measured and plugged into population-specific regression formulas to build a picture of the whole body. Because body shapes vary around the world, scientists use a formula derived specifically from that individual’s ancestral group, which accounts for differences in build and bone density.

The team calculated that this individual stood roughly 178 centimeters tall, or about five feet ten inches. For 13th-century Hungary, this was above average, consistent with the better nutrition that noble families enjoyed: consistent with an individual like Béla.

The teeth provided the final morphological clue.

“People in medieval times didn’t really care about [dental health],” says Anna Szécsényi-Nagy, PhD, an archeogeneticist at the Institute of Archaeogenomics in Budapest and co-author of the study. “There were big chunks of plaque that we could analyze.”

Microscopic analysis of plaque deposits in the skull’s remaining teeth revealed starch granules from cereals — wheat, barley, and the bread that formed the basis of medieval European diets.3 The samples also showed traces of cooked foods, plant material that had been processed and prepared rather than eaten raw. Peasants wouldn’t have eaten this lavishly.

So far, everything fit Béla’s profile. The skeleton belonged to a young male who had grown up in the right region and hadn’t experienced the physical stresses of the peasantry. But consistency is not proof. Many young aristocrats lived in 13th-century Hungary, and any of them might match this profile. To narrow the identification further, the team needed to establish when this person died and what this person ate (with even further specificity), which can be done through radiocarbon dating.

Radiocarbon Dating

Carbon exists in several forms. Most is carbon-12, which is stable. But a tiny fraction, about one atom in a trillion, is carbon-14, which is radioactive. It’s constantly created in the upper atmosphere when cosmic rays strike nitrogen atoms, and it filters down into all living things on our planet. Plants absorb it during photosynthesis, and when animals eat these plants, carbon is taken up by their tissues.

When something dies, it stops absorbing new carbon. And because carbon-14 is unstable, it slowly decays into nitrogen-14, with half of it disappearing every 5,730 years. It is this principle that underlies radiocarbon dating; when an organism dies, its carbon-14 decays and is not replaced. Less carbon-14 means more time has passed.

This technique emerged from nuclear physics research. Carbon-14 was discovered in 1940, by researchers at the University of California Radiation Laboratory in Berkeley, California. The isotope was first a simple curiosity, but Willard Libby, working at the University of Chicago’s Institute for Nuclear Studies (founded to retain Manhattan Project scientists), recognized how it could be used to date samples.

Libby published his first radiocarbon dates in 1949, including the famous “curve of knowns” — a validation plot showing that the method accurately dated Egyptian artifacts, sequoia tree rings, and other samples of known age. The impact was immediate. By 1960, when Libby won the Nobel Prize in Chemistry, more than 20 radiocarbon dating laboratories had been established worldwide.4

For the first time, researchers could date ancient remains without relying on written records. Yet despite giving archaeologists a universal clock that could place sites on different continents on a common timeline, radiocarbon dating is not perfect. It assumes atmospheric carbon-14 levels remain stable across millennia, which hasn’t always been the case.

An example of such a fluctuation can be found in a study conducted by the Groningen radiocarbon laboratory in the Netherlands, in which tomatoes bought fresh from the market at the time of the study appeared to be 1,300 years old.

So what happened? These tomatoes had been grown in greenhouses where the air was enriched with carbon dioxide (CO₂) to boost plant growth. The CO₂ came from burning fossil fuels like coal, oil, and natural gas, all of which are millions of years old and had started decaying long ago. The tomatoes absorbed this “dead” carbon instead of normal atmospheric carbon, making them appear far older than they actually were.

Something similar happened when researchers tried to date Duke Béla’s skeleton. The team turned to István Major, a radiocarbon specialist at Hungary’s Institute for Nuclear Research in Debrecen, which houses one of Central Europe’s leading accelerator mass spectrometry facilities. His laboratory has spent decades refining techniques for dating difficult samples such as cremated bones, contaminated materials, specimens that confound standard methods.

Béla’s skeleton brought back dates too early, clustering around 1030-1230 AD. At that time, “the nunnery couldn’t have even been established and built,” says Szécsényi-Nagy. “It was totally impossible.” (Indeed, the monastery hadn’t been founded until 1259).

At first, contamination was suspected. The skull had been treated with glue, paper, and plaster during earlier conservation efforts, and these materials might have introduced foreign carbon. The team tried multiple approaches: sampling different bones, separating inner and outer layers of bone, and using aggressive chemicals to thoroughly clean out contaminants. Despite all this, the results kept showing the burial took place long before Béla’s time.

A resolution finally came from a different type of analysis. In addition to carbon-14, bones contain other forms of carbon and nitrogen that don’t decay. The ratios of these stable isotopes can reveal information about diet. A high ratio of nitrogen-15 to nitrogen-14, for instance, indicates a diet rich in animal protein, like meat, dairy, or fish.

Upon checking Béla’s stable isotopes, “We figured out that he had a high-protein diet, so he ate a lot of meat,” says Szécsényi-Nagy. “And it is very probable, living or spending most of his time close to the Danube [river], that he ate lots of fish.” Béla’s nitrogen values were elevated well above the range typical for people eating mostly plants and land animals.

As it would turn out, freshwater fish pose the same problem as greenhouse tomatoes (only for a slightly different reason). Rivers like the Danube flow over ancient limestone bedrock, dissolving carbon that’s been locked in the rock for millions of years. Fish absorb this old carbon as they feed, and it accumulates in their bodies. When a person eats those fish, the old carbon gets incorporated into their bones. This is called the “freshwater reservoir effect,” and it was readily apparent in Béla’s bones: they contained less carbon-14 than they should have because he’d been eating carbon that was already “ancient” when he consumed it, thus distorting the dating results.5

Once Major’s team understood the confounder, they could model what the dates would look like if they corrected for it — testing shifts of 50 to 200 years to account for old carbon in Danube fish. The corrected dates aligned with the late 13th century.6

Additionally, they analyzed the mineral fraction of the bone rather than the collagen, the latter of which would lock in the older carbon. Bone mineral incorporates carbon not just from food but also from inhaled carbon dioxide and blood bicarbonate — sources that aren’t affected by the freshwater reservoir effect. The mineral carbonate dates clustered around 1170-1260 AD, consistent with a death in the late 13th century (and with historical chronicles that place Béla’s murder specifically in November 1272).

By this point, the investigation had answered some fundamental questions. Strontium isotopes placed the remains of this person’s childhood in the Hungarian basin, and radiocarbon dating put their death in the late 13th century. Each line of evidence had narrowed the possibilities, but further identification required ancient DNA analysis.

Ancient DNA

For most of history, forensic anthropology was a descriptive science. Researchers could establish a biological profile — sex, age, stature, ancestry, evidence of trauma — to narrow the pool of possible identities, excluding candidates who didn’t match. But positive identification required comparing remains against records created during the person’s lifetime: dental charts, medical X-rays, documented injuries. Without such records, identifying a dead person was infeasible.

The first major breakthrough came in 1984, when British geneticist Alec Jeffreys developed DNA fingerprinting at the University of Leicester. By exploiting regions of the genome that vary dramatically between individuals, he could produce a banding pattern unique to each person. The technique revolutionized criminal forensics and paternity testing, but it required relatively intact genetic material.

That posed a problem for ancient remains. DNA survives in bone far longer than in soft tissue because it becomes entombed within hydroxyapatite, the dense crystalline mineral that gives bone its rigidity. But it still degrades through purely chemical processes, shattering into fragments too short for traditional fingerprinting to read.

A way past this bottleneck came in 2005, when 454 Life Sciences released the first commercial next-generation sequencer. Well-suited to ancient DNA analysis, next-generation sequencing (NGS) permits the reading of millions of DNA fragments simultaneously rather than one at a time, using software to align the overlapping pieces against a reference genome.

NGS was vital in cracking open Béla’s identity. But before they could sequence his genome, they had to extract the degraded fragments and build sequencing libraries from molecules that had been disintegrating since the 13th century.

At the Institute of Archaeogenomics in Budapest, Szécsényi-Nagy and Noémi Borbély began their ancient DNA analysis by drilling into the petrous bone on the skull. The petrous bone ossifies early in fetal development and undergoes almost no remodeling throughout life, creating a stable mineral matrix where DNA remains entombed.

They then followed the Dabney protocol, a standard method for ancient DNA extraction developed in 2013. The protocol calls for dissolving 50 milligrams of bone powder in a chemical solution overnight so the calcium phosphate matrix is broken down, releasing the DNA. Meanwhile, enzymes in the solution digest proteins and other contaminants. By morning, the powder becomes a murky liquid containing ancient DNA mixed with bacterial and fungal sequences.

Extraction takes two days: the first to dissolve the bone and release its DNA, and the second to purify it, followed by an additional two of “library preparation,” which involves attaching molecular tags so the sequencing machine can read each fragment and trace it back to its source.

The sequencing took three to four weeks, focusing on two critical regions: mitochondrial DNA and Y-chromosome markers. Mitochondrial DNA traces maternal descent. Every human cell contains hundreds of mitochondria, tiny structures that generate cellular energy, each carrying its own small genome. We inherit our mitochondria exclusively from our mother, who inherited hers from her mother, and so on — an unbroken chain of maternal inheritance.

Béla carried haplogroup U3b3 in his mitochondrial DNA, a lineage rooted in the Near East and southeastern Europe. This pointed toward Byzantine ancestry, consistent with the historical record: his maternal grandmother was Maria Laskarina, daughter of a Byzantine emperor.

Y-chromosome DNA traces paternal descent. Unlike other chromosomes, which shuffle their genetic material each generation, the Y chromosome passes from father to son essentially unchanged. Mutations accumulate one at a time, creating a linear record.

Béla belonged to a Y-chromosome lineage shared by the Rurikid dynasty — descendants of Rurik, a ninth-century Varangian prince who established the ruling house of Kievan Rus’. His Y-STR profile matched 24 tested descendants of the Rurikids in modern genealogical databases, and aligned perfectly with the anticipated profile of Rurik himself.

The Y-chromosome analysis was particularly delicate, as any contamination from modern DNA would invalidate the results, so the team ran every step at least twice to confirm authenticity. They also ran the sequencing experiments in both Budapest and at Harvard University.

Harvard’s ancient DNA laboratory had isolated the cochlea itself (the coiled structure of the inner ear), which yields up to 65-fold more endogenous DNA than other skeletal tissue. Whereas the Budapest team focused on mitochondrial DNA and the Y chromosome, the Harvard scientists generated genome-wide data, sequencing across all twenty-two autosomal chromosomes because their aim was population history, not individual identification. But this data sat unused.

Szécsényi-Nagy’s team requested the raw sequencing files from Harvard and, with both datasets in hand, could finally check Y-chromosome markers against the whole-genome data. Crucially, they could also run identity-by-descent analysis to measure Béla’s genetic relatedness to other sequenced medieval royals.

Any two people with a common ancestor inherit stretches of identical DNA from that ancestor. But with each generation, chromosomes reshuffle, breaking those shared stretches into smaller pieces. Siblings share long, continuous blocks of matching sequence. Fourth cousins share only scattered fragments. The math is predictable enough to work backward: measure the total length of identical segments between two individuals, and you can estimate how many generations separate them from their common ancestor.

When the team compared Béla’s genome against previously sequenced medieval royals, the genetic relationships matched the family trees with high precision. Béla of Macsó was indeed a descendant of Béla III, the twelfth-century Hungarian king, separated by exactly the four generations the genealogical records predicted. And in 2023, the final piece fell into place: Dmitry Alexandrovich, a 13th-century Rus’ prince whose genome had been published in a separate study, emerged as a distant relative — confirming the connection to the Rurikid dynasty that Y-chromosome analysis had first suggested three years earlier.

“This whole network of relatedness fitted so nicely together,” Szécsényi-Nagy says. “The genetic evidence didn’t just support the historical record; it was locked into place like pieces of a puzzle.”

Forensic Trauma Analysis

The final unanswered question was the manner of Béla’s death. The chronicles claimed he was murdered. The bones could test that claim, but determining specific types of violence on his skeleton required distinguishing whether its wounds had been inflicted during his lifetime (including what might have been his death wounds) or from the damage accumulated over the seven centuries Béla’s skeleton had spent underground.

Every excavated skeleton shows wear: fractures from soil pressure, breaks from careless excavation, and erosion from groundwater. Bones absorb minerals from surrounding earth, staining brown or yellow depending on soil chemistry. This discoloration helps identify when damage occurred. Ancient healed injuries show bone remodeling around the wound. Fresh breaks from modern handling expose pale, unstained bone. The harder question is whether an unhealed, stained wound represents a killing blow or damage inflicted during burial centuries ago.

The answer lies in collagen. In living bones, this protein provides flexibility. A hard blow breaks the bone cleanly, with smooth fracture lines following the natural grain. Centuries after death, once its collagen has degraded, bone becomes brittle and shatters like dry wood, leaving jagged edges that crumble rather than split. But for roughly a decade after death — the “perimortem period”— bone retains enough collagen to fracture like living tissue. Wounds from this window of time leave distinctive signatures like smooth-edged cuts, clean fractures, and breaks that follow the grain. Wounds inflicted at or near the moment of death show no signs of healing.

Béla’s skeleton displayed twenty-six perimortem injuries with no new bone growth around the wound edges: nine to the skull, seventeen to the body. He died from these wounds, or moments after receiving them.

The shape of each wound also provides clues as to what made it. “Imagine you have a block of soap,” explains Martin Trautmann, a forensic anthropologist who collaborated on the research while at the University of Helsinki, “and you use an ax with quite a thick blade. When this penetrates the block, it will push a lot of material to the sides, just because it’s so broad.” A thick blade crushing into bone leaves small parallel lines along the cut where the outer layer was pushed aside. Forensic scientists call this “feathering.” A thinner blade makes a cleaner cut with minimal disruption to the surrounding surface.

In the 13th-century, Hungary had two distinct blade traditions. Western European longswords, designed for thrusting as well as chopping, were symmetric and double-edged, with a gently curved cross-section. Sabers, which had entered Central Europe with equestrian peoples of the Eurasian steppe, were single-edged and curved, with a thin, wedged-shaped cross-section.

Both weapons left their marks on Béla. The wounds to his head and upper body showed the signatures of a lighter, sharper blade wielded in sweeping arcs, consistent with a saber. The wounds to his legs were heavier, with slight crushing of the bone surface, consistent with a longsword’s chopping stroke.

Two weapons meant at least two attackers. But had there been more?

To find out, Trautmann worked with an articulated laboratory skeleton, the kind used in anatomy classes. He marked each of Béla’s twenty-six wounds in their exact positions on the articulated skeleton, matching their length, angle, and orientation.

Each lesion was analyzed for forensic information through both direct visual inspection and radiologic imaging. Diagnostic marks on the damaged surfaces revealed not just where blows landed, but their direction of origin, the angles of impact, and the probable movements of the weapons and attackers. Trautmann also took similar weapons to the medieval saber and sword profiles used by the attackers and fitted them directly into the bone lesions to find out which type of weapon inflicted each.

By combining all this data about wound location, weapon type, strike direction, and sequence, Trautmann could test scenarios against the physical evidence. In his reconstructive analysis, he put one attacker facing the laboratory skeleton’s front, and two facing from the sides, envisioning different sequences of blows and the defensive postures the living Béla would have adopted.

The pattern that emerged was unambiguous, according to Trautmann. “Béla was confronting a person. This person drew his saber, hit him two to three times on the head and upper body. And then the other two people, probably allies of the saber-wielding attacker, flanked the victim from the left and the right — they kind of cut off his retreat … and then they finished him.”

Historical accounts had offered two versions of Béla’s death: an ambush from behind or a dispute that escalated into violence. Cuts to both forearms showed Béla had raised his arms to protect his head in a defensive reflex, which also meant he hadn’t carried a weapon or shield. The wounds were inflicted mostly on the front of his body, so he faced his killers.

Béla’s wounds are consistent with records that suggest that the motive was a heated dispute with Henrik Kőszegi, still angry by Béla’s betrayal during the Battle of Isaszeg, where Béla fled from the battle while Henrik was captured.

That Henrik had not forgiven this is evidenced by Béla’s facial injuries. “[It] looks like the eyes were gouged out and the nose was cut off,” Trautmann explains. “This kind of facial injury is mentioned in written sources as a punitive act for treason and disloyalty” in medieval Central Europe.

Because the assault continued even after Béla had most likely received a mortal wound, it must have been “a strange mixture of cold and hot blood,” Trautmann concludes. “They were prepared to attack, but when they did, they got carried away.”

What the Dead Reveal

Béla’s case offered a rare combination for bioarchaeologists: a nobleman whose death was chronicled by scribes and whose body was preserved. But complex bioarchaeology is revealing patterns with more complicated remains, even those of mass fatalities.

Consider the Bronze Age battle that ravaged the Tollense Valley over 3,000 years ago, when thousands of warriors clashed at a strategic river crossing in what is now northeastern Germany. Before bioarchaeological analysis of the site, many archaeologists doubted that large-scale organized combat existed in prehistoric Europe — weapons in graves were seen as status symbols, and any violence was assumed to be small-scale raiding between local clans.

However, large-scale excavations of the site demolished this assumption. Excavators recovered over 10,000 bones representing at least 150 individuals, almost all young men killed around 1250 BCE. Isotope and DNA analysis revealed the warriors came from across the continent — southern Europe, Scandinavia, Poland — while a 2024 arrowhead study found weapons characteristic of southern Germany and Moravia. The emerging picture is of coordinated, long-distance military mobilization in a region that would remain illiterate for centuries, leaving no written record of what may have been a common practice.

Complex bioarchaeology can also remedy incomplete or inaccurate records. Greek historians celebrated the Battle of Himera in 480 BCE as a triumph of Hellenic unity against Carthage, mentioning reinforcements from Syracuse and Agrigento but no foreigners. Yet isotope analysis of soldiers buried in mass graves revealed approximately two-thirds grew up outside Sicily, while DNA analysis showed nine warriors had genetic affinities with central Europe, northeastern Europe, the Eurasian steppe, and Armenia. They were possibly mercenaries from as far as the Baltic whose contribution was never recorded by ancient chroniclers.

Criminologists use the term “dark figure” to describe the gap between actual crime and what gets reported. Ancient violence has an enormous dark figure, which even complex bioarchaeology struggles to reconcile. Death tolls require institutional memory that ancient societies rarely built. While ancient states kept records, they tended to count the living, not the dead, capturing snapshots of who existed and what they owned, not who had died.7 While ancient states maintained censuses of the living for taxation and labor, systematic registration of deaths emerged primarily in 19th-century Europe and North America.

Complex bioarchaeology cannot resurrect demographic data that was never collected. It can’t tell us how many people died at Tollense Valley; only that the minimum was 150 (based on recovered bones), with the true figure almost certainly far higher. But it can help fill important gaps, offering confirmatory data for ancient mysteries such as Béla. It could also aid in investigating the far more numerous ordinary deaths that historians have failed to notice.

Each excavation and analysis, then, contributes to a growing body of knowledge about how people lived and died. With every skeleton identified and every wound pattern reconstructed, the dark figure shrinks. And while we may never resolve the enduring questions regarding violence and human nature, at least we can now tackle these questions in an evidence-based way. Written history is a starting point, but records kept in calcium, carbon, collagen, and DNA move us further along.

Subscribe now

David Brzostowicki is a journalist whose writing covers emerging research in biomedical sciences, biotechnology, and healthcare. He is currently a graduate student at Florida Atlantic University’s College of Medicine in Boca Raton, FL, where he studies the molecular mechanisms linking vascular health to metabolic protection, with a focus on circadian pathways and endothelial-to-muscle signaling.

Thanks to Anna Szécsényi-Nagy for her generous interviews and explanations on ancient DNA analysis and the study’s background, and to Martin Trautmann for his expertise in forensic trauma analysis. Header image by Ella Watkins-Dulaney.

Cite: Brzostowicki D. “Dead Reckoning.” Asimov Press (2026). DOI: 10.62211/84jy-26we

This essay will appear in our forthcoming book, “Making the Modern Laboratory,” to be published this summer.

By Spencer Wright

It was a revolutionary idea in the 1830s, and it remains one today — virtually anyone can learn to make their own scientific equipment. With a few dollars’ worth of glass tubing, a flame, and a little practice, you can create all kinds of chemical analysis kits. Because the glass itself is airtight, you can control which chemicals go in, and because it’s clear, you can observe what happens to those chemicals as you manipulate them. If they don’t do what you intend, you can reignite the flame, modify your glass apparatus, and try again.

This insight helped build the modern laboratory: Work wherever you want to, but work in glass, and you’ll reveal life’s most intricate mysteries.

The person who catalyzed this idea, Justus Liebig, had the superlative career title of “Extraordinary Professor of Chemistry” (although his job’s “extraordinary” nature may have had more to do with its nominal compensation than anything else). In fact, Liebig wasn’t excited about the position he held, writing to his parents that he had “no great desire” for it. He would have preferred to stay in Paris — the center of the chemistry universe at the time, and a city where, even at twenty-one years old, Liebig had already begun to make a name for himself. But the University of Giessen had offered him a modest stipend to establish a lab, so, in 1821, Liebig accepted the constraints of small-town life and packed himself off there.

The move (and modest budget) forced Liebig to perform chemical analysis in new ways. While in Paris, studying under the esteemed French chemist Joseph-Louis Gay-Lussac, Liebig had analyzed silver fulminate, letting it combust and then collecting the carbon dioxide that resulted. Because carbon dioxide is hard to weigh,1 Liebig had been taught to measure its volume using an eudiometer, a type of inverted, graduated cylinder where gases could be trapped and their volume read from markings on the glass. This device was costly to make and finicky to use, and Gay-Lussac relied on specialized glassblowers to help design, produce, and maintain it. With such infrastructure missing in Giessen, Liebig’s research stalled.

If necessity is the mother of all invention, then ambition is its accelerant. Dissatisfied with teaching pure theory at the University, Liebig founded an independent institute to impart the tactile skills of applied chemistry. And dissatisfied with the equipment available in Giessen, Liebig traveled back to Paris to learn glassblowing — a skill he then passed on to his students.

Even as he gained these practical skills, he picked intellectual fights with prominent chemists in Paris and Berlin, insisting that their methods of analyzing organic molecules (which measured carbon and nitrogen together from a single volume of mixed combustion gases) were flawed. By 1830, Liebig knew how to make his own glassware and was now prepared to back up his published claims of superior analytical excellence.

The result was the Kaliapparat, a twisted glass triangle that Liebig used to analyze the composition of morphine.2 The principle behind it was simple: Fill the bottom of the Kaliapparat with potassium hydroxide, then force morphine’s combustion gases through it. As the gases bubbled through the Kaliapparat’s bulbs, any CO₂ present would react with the potassium hydroxide, trapping the carbon but allowing the rest (hydrogen, nitrogen, and oxygen) to pass through. Then, by weighing the Kaliapparat and subtracting its pre-reaction mass, Liebig could determine exactly how much carbon, by mass, had been in the original morphine sample.

As an analytical tool, the Kaliapparat was middling. Liebig’s original goal had been to determine morphine’s nitrogen content, which proved challenging given that the nitrogen passing through the apparatus commingled and reacted with other gases. In fact, historian Catherine Jackson states in her book Molecular World that it was even self-admittedly a failure. “In private, Liebig was even more scathing. As he admitted to Berzelius [a dominant figure in European chemistry and a correspondent], his method of nitrogen determination was ‘tiresome, time-consuming and, in a word, quite unbearable,’ while nitrogen’s side reactions had driven him to ‘despair,’” she writes. However, the Kaliapparat was indeed reliable for measuring carbon — and at a much lower cost than the volumetric-based Parisian equipment of the day. So when he published his results in 1831, Liebig simply de-emphasized the nitrogen analysis performed with it and focused on the glassware itself, christening it “A New Apparatus.”

It would take a few more years, however, for the vessel to gain widespread adoption. In 1833, Liebig slipped instructions on how to make a Kaliapparat into an article he had translated from French to German. In 1834, one of his assistants also demonstrated how to make a Kaliapparat at a chemistry meeting. And, in 1839, Liebig published a textbook on organic analysis that included instructions on the making and use of such glass vessels. According to Jackson, Liebig’s effort to publicize the device — and encourage other chemists to make their own — was “explicitly pedagogical.” Liebig sought “to overturn Parisian chemical orthodoxy … [wanting] chemists everywhere to see and understand, to make and use the Kaliapparat and to prove its worth by widespread replication.”

Finally, Liebig’s efforts paid off, and the 1840s became a golden era of amateur glassblowing. Chemists everywhere took up the torch, replicating Liebig’s Kaliapparat (a stylized version of which would be used in the American Chemical Society logo in the 20th century) and producing designs of their own. Glass vessels evolved, and organic analysis flourished as scientists could finally determine repeatable values for a sample’s carbon content. But most importantly, glassblowing established itself as a need-to-know process for anyone doing serious chemistry.

While Liebig’s proselytizing encouraged chemists to learn glassmaking, it also helped that glass itself was so cheap and easy to make. Glassmaking materials (silica sand, soda ash, and limestone) were readily available. As a result, glassblowing became professionalized. Entire lab glassware catalogs sprang up, many of which contained both prices and instructions on how to make one’s own instruments.

One prominent glass dealer was English chemist and publisher John Joseph Griffin, who in his 1832 book Chemical Reactions promoted the use of tumblers (later known as “Griffin beakers”). As the need for professional lab glassware suppliers emerged, Griffin established a shop in Covent Garden in 1852. By the time he published his 1866 lab equipment catalog, it included an adaptation of the Kaliapparat. For one shilling and sixpence, a chemist could purchase “Liebig’s Potash Apparatus,” complete with furnace, combustion tube, desiccant chamber, Kaliapparat, and vulcanized caoutchouc (rubber) connectors. For two shillings and sixpence, a specially modified Kaliapparat, made by professional glassblower Heinrich Geissler, was also available.

This modified device, just one of many similar tools that Geissler invented in the latter half of the 19th century, was designed such that it could be set upright on a tabletop. (Liebig’s triangular design, intended to be suspended from a scale, would simply fall over if set on a surface.) Beyond its basic geometry, Geissler made refinements to increase the Kaliapparat’s accuracy, adding integrated desiccant tubes as well as embedding little glass discs within its three glass bulbs, which forced the combustion fumes to bubble more vigorously (and therefore react more fully) as they passed through the potassium hydroxide.

Geissler’s Kaliapparat is especially striking considering how difficult it would have been to fashion in 1866. At the time, a glassblower’s flame was not made by compressed propane and oxygen, as they are today, but by simply blowing air over an alcohol or oil lamp — or, in some cases, over a charcoal fire. “You were either using your breath or a foot bellows,” choreographing hands, mouth, and feet to generate temperatures in excess of 700°C, says Tracy Drier, a master glassblower at the University of Wisconsin. Today’s glassblowers work with oxy-propane torches that render such contortions obsolete, a convenience Liebig and Geissler could not have imagined.

Liebig and Geissler would have also been struck by the material properties of Tracy’s glass, as the glass formulations of their day left much to be desired. Both soda-lime formulations and leaded glass were vulnerable to attack by water and acids, and would become pitted and cloudy after repeated use. They also needed to be carefully handled and were sensitive to sharp changes in temperature. Even Bohemian potash glass, known for being colorless and fairly workable, contained impurities (iron contamination could introduce a faint greenish tinge) as well as bubbles, seeds, and striations (thin waviness from incomplete melting and mixing).

Perhaps unsurprisingly, the features of glass that late-19th-century scientists desperately wanted to improve upon were its optical qualities. This was especially true for those working in the burgeoning laboratory sciences that relied heavily on microscopes. And the key to better microscopes was better lenses — a technology which was being developed primarily in Jena, a town just 250 kilometers west of Giessen.

One of the major challenges lenses contend with is light dispersion. When light passes through glass, different frequencies bend by different amounts (the same phenomenon that creates rainbows in prisms). Dispersion is not an issue for lab glassware. A chemist wouldn’t care much if their Kaliapparat were to create little rainbows on their workbench. But in the nineteenth century, developments in the glass industry tended to come from optics (a big market) rather than beakers (a relatively small one). In a microscope or telescope, dispersion results in blurry, unfocused images as the light’s constituent parts fail to converge at a single focal point.

It was the chemist Otto Schott who took up the creation of low-dispersion glass. Schott grew up in the glass factory his father managed and later studied the chemistry of glass at the University of Jena. When he completed his degree in 1875, he returned home to Witten, where he kept working on glass formulations, and in 1879, managed to melt a batch with especially high lithium content. Confident that he had found a low-dispersion formulation, he sent a sample to Ernst Abbe, a professor at Jena who had been advocating for a systematic study of glass’s optical qualities.

Taken by the young Schott’s enthusiasm, Abbe invited Schott back to Jena to join him and microscope-maker Carl Zeiss in forming the Glass Technology Research Station. The collaboration, credited as “one of the greatest and most productive associations in the history of glass composition,” proved immensely fruitful. Before he had even moved to Jena, Schott succeeded in producing what he called “borate glass,” which contained high amounts of boric acid. Abbe, measuring the optical properties of this sample, wrote Schott to say that the new sample had “solved completely” the issues caused by dispersion.

If the features of soda-lime glass are remarkable, then those of Schott’s borate glass — later known as borosilicate — were simply fantastic. In addition to its low dispersion, the borosilicates he developed in the 1880s were harder, more thermally stable, and more resistant to corrosion than soda-lime glasses. Each of these properties led to better lens performance. Increased hardness meant that borosilicates could be ground and polished more precisely, resulting in denser surfaces that reduced scattering. Borosilicate’s thermal stability — its tendency not to swell much when heated — meant that fewer internal stresses were created during the grinding and polishing process, resulting in a more homogeneous refractive index. And its excellent environmental stability meant that it didn’t corrode when exposed to air, water, and other chemicals (a problem rare in today’s glass, but painfully evident in antique glassware).

Borosilicate is also much more elastic than soda-lime glass, resulting in lenses (and later lab glassware) that could withstand whatever bumps and jostles they would inevitably experience in the laboratory. Schott, Ernst Abbe, and Carl Zeiss would go on to develop dozens of varieties of optical borosilicate glass; by 1886, they were selling 44 different varieties.

While optical glass has been heralded as “one of the key materials of civilization,” the creation of “utensil glass” would become the real workhorse of the modern laboratory. Schott first produced it in 1892, noting in his journal that “introduction of a boric acid content of up to 25 percent” resulted in glass that could be “manufactured into objects subject to thermal shocks or required to compensate for large temperature differences through the wall. Such objects are: boiling flasks, beakers, [and] bowls for chemists.”

Incredibly, his new formulation was rejected by the German patent office, which cited his earlier wares as being similarly durable. Schott tried, and failed, to appeal their ruling, but in the end, he found satisfaction through commercial success instead. And the new formulation, introduced to the market in the winter of 1893-1894, became the standard for lab glassware almost immediately. By the first years of the 20th century, Schott was selling not only “boiling flasks, Erlenmeyer flasks and beakers” but also “round-bottomed flasks, measuring flasks, Kjeldahl flasks, retorts, fractional distillation flasks, evaporating dishes, test bottles and tubes.”

Schott’s advancements quickly made Jena the heart of the glassblowing world. In Germany, both “Jena” and “Schott” would become household names, with streets and even sports facilities paying tribute to the venerated glassblower. And their reach was not confined to central Europe: Because lab glassware was classified as an educational product, it was often exempt from U.S. import duties, allowing the established German industry to outcompete the younger American one for more than a decade. Those living in the U.S. in 1910 would likely purchase their Kaliapparats from a German manufacturer.

This changed in the summer of 1914 — the onset of WWI — when the British blockade prevented German-made goods from reaching U.S. shores. According to a 1917 Bureau of Standards report, the blockade caused “a very serious shortage of glassware.” With German glass immobilized, however, the market was flooded with new domestic entrants. To promote these brands to American chemists, the Bureau of Standards tested five national lab glassware brands and compared them to Jena and Kavalier, the two German brands that had been most commonplace in the years leading up to WWI. Only one of the American companies, Libbey, predated the invention of borosilicate glass (originally founded in 1818 in Cambridge, Massachusetts, as the New England Glass Company, before relocating to Ohio in 1888 and renaming to Libbey Glass Co).

It was the newest entrant to the market — a Corning, New York brand called “Pyrex” — that performed best of all. Pyrex ware had an “unusually low” coefficient of expansion, and its chemical resistance showed “slight superiority” over all but the Libbey sample. The Corning company, already producing pie pans and casseroles under the Pyrex name, was off to the races. By 1916, they were producing beakers, flasks, and glass tubing, and by 1918, the word “Pyrex” was more common than “borosilicate.” By 1927, use of the word “Pyrex” went on to surpass “Kodak.”

This period was undoubtedly the golden age of glassware. During the 1920s and 1930s, a series of glassware providers established standardized ground fittings, allowing chemists to piece together different glass apparatus without the use of corks and rubber connecting hoses. These acted as a single, standard interface design, letting chemists quickly assemble modular experimental setups. Standard ground fittings quickly gained popularity, and in 1930, they were adopted as “CS 21-36” by the U.S. National Bureau of Standards. By 1938, the Pyrex glassware catalog was 130 pages long. Throughout their product line, all connections were compatible with CS 21-36 Standard Taper fittings.

Such fittings were only one innovation among many aimed at greater precision. Until the 20th century, most glassware (even that based on the same design) wasn’t exactly the same size and volume. While glass molds had existed since the first century to help with consistency, minor variation between vessels remained. This was usually addressed by individual calibration. Each volumetric flask or beaker was filled with distilled water at a controlled temperature — typically 20°C — and the water was carefully weighed on a precision scale. Using the known density of water at that temperature, the glassblower could calculate the true volume and etch the calibration mark at the right spot on that particular piece of glass. This meant that two half-liter flasks might have their calibration lines at slightly different heights, but both would contain precisely 500 mL when filled to their respective marks.

During this “golden age,” however, the development of precision-machined metal molds, combined with Otto Schott’s borosilicate glass, which expanded minimally with temperature changes, meant that glassware could be manufactured to consistent dimensions batch after batch.

Despite the proliferation of standardized glassware and instruments, working chemists were still expected to blow at least some of their own. Even in the early 20th century, this was a required skill for PhD candidates at the University of Wisconsin, and glassblowing classes were common parts of four-year chemistry degrees through the 1950s, when trade schools began offering specialty degrees in the subject. During this time, it was not unheard-of for major biotech companies to outfit (but not staff) their own glass shops, with the expectation that working chemists would produce their own supply. As late as 1957, chemistry textbooks contained instructions on how to work glass with flame.

Today, however, most chemists have given up making their own equipment. And glassblowing, once considered a craft essential to the science of chemistry, has become a subcontracted trade. Even so, the material itself remains centrally important: while plastic has made inroads into the laboratory in the form of sterile, inexpensive components like pipette tips, polystyrene tissue culture flasks, petri dishes, and multi-well plates, glass remains key in optical and high-temperature applications.

Jena, too, remains synonymous with glass, today known as “The City of Science and Light.” While the headquarters of Scott AG moved after WWII to Mainz in West Germany, the company maintains major operations in Jena, producing everything from optical fibers and ceramic glass to the deeply practical glassware championed by Schott, Abbe, and that extraordinary professor of chemistry himself, Justus Liebig.

Subscribe now

Spencer Wright is a writer and ersatz engineer living in Brooklyn, New York. Since 2013, he has written the newsletter Scope of Work, which these days vacillates between technical deep dives and essays about living in New York, rewiring thermostats, and explaining photons to six-year-olds.

Acknowledgements: Thanks to Tracy Drier for granting us an interview, and also to Catherine Jackson, whose book Molecular World provided an incredible overview of Liebig’s work. I also sourced extensively from Jackson’s “The Wonderful Properties of Glass,” Liebig’s Kaliapparat and the Practice of Chemistry in Glass, and Jürgen Steiner’s Otto Schott And The Invention Of Borosilicate Glass. Header image by Ella Watkins-Dulaney.

Cite: Wright, S. “Working in Glass.” Asimov Press (2026). DOI: 10.62211/79pj-25qs

The last World Computer Chess Championship was held in October 2024. It ended because, in the words of its organizers, “top programs are unbeatable by humans; making them stronger has no real research value.” Its mission, after half a century of effort, was complete.

Chess engines now sit at the center of how the game is played. Engine evaluations hover over boards during livestreams and post-game analyses, and players study engine lines extensively to improve their own play. As a result, top grandmasters are stronger than any previous generation. But these gains come with a concession: Top engines often play in a way “too far past the human frontier” to fully understand. In other words, we trust chess engines because they are unquestionably better than we are, even when their decisions make no sense to us.

A similar imbalance may also emerge in scientific discovery. Researchers are building systems able to navigate the full arc of scientific inquiry with increasing autonomy, including proposing hypotheses, designing experiments, and evaluating results. The question, however, is whether we will cede the same authority to AI scientists as we have done in chess. And, if so, what will happen to science when AI models produce results beyond our ability to understand?

I call this the “legibility problem,” the risk that AI-generated scientific knowledge becomes incompatible with human understanding, and think it will define the next era of science. The knowledge AI systems generate may be expressed in concepts that do not map onto our own, communicated in ways optimized for other AIs rather than for human investigators.

Importantly, this concern is separate from the problem of mechanistic interpretability (the effort to understand what happens inside neural networks). My concern is not whether we can “see” inside these new models, but whether we will even understand — or be able to control — what comes out of them.

This may not seem like a new or distinct challenge. After all, scientists have already had to grapple with the rise of big data and increasingly opaque computational methods. However, up to this point, human scientists still determined which questions were worth studying.

Scientific discoveries must eventually intersect with material reality to have any utility. They must be instantiated as therapies, materials, and public policies. For this to happen, the knowledge that AI systems produce must remain legible within the systems through which humans operate: our laboratories, our clinical infrastructure, our regulatory bodies, and our vehicles for communicating what we know. Of course, this legibility is a matter of degree.

Take, for instance, the diabetes drug metformin, ingested by millions of people for over 70 years. Despite its success, we still do not fully understand its mechanism of action. But metformin did not appear out of nowhere. It emerged through a long chain of human experimentation, from herbal medicine and chemical purification to animal studies and clinical trials.

AI-generated science may not follow this pattern. AI scientists will have no intrinsic reason to work within our existing conceptual categories, just as superhuman chess engines have no intrinsic reason to explain their choice of moves. In fact, if we truly want AI scientists to make breakthroughs, some loss of legibility may be inevitable. The chief risk is that discoveries become effectively stranded, buried in a volume of AI-generated output no human institution is equipped to parse or implement.

Once again, the history of chess offers a model for understanding how quickly this shift can occur. In May 1997, Grandmaster Garry Kasparov lost to IBM’s Deep Blue, 3½ – 2½. It was a watershed moment in AI research, but the skill gap between human and machine was still narrow (Kasparov had won Game 1). In a 2010 essay reflecting on the era, Kasparov observed that the most impressive chess was played not by humans or engines alone, but by the two in partnership. This was substantiated in a 2005 tournament, where two amateurs using three ordinary computers defeated both grandmasters and supercomputers. Such human-machine pairings came to be called “centaurs.”

The AI-for-science field is approaching its centaur phase now. But even though most research groups currently use AI to confirm or extend human-generated hypotheses, we should not expect this centaur approach to last for long.

By 2017, DeepMind’s AlphaZero had taught itself chess from scratch in under four hours and demolished every competitor, human or machine. After AlphaZero, a human partner added nothing. Computer scientist Richard Sutton documented this recurring pattern across the history of AI research in his 2019 essay “The Bitter Lesson.” Again and again, he noted, systems that fully exploit computation have outperformed those designed to reflect human knowledge and intuition. In fact, human intuition has consistently and “bitterly” hindered system performance. Simply put, AI scientists will likely do better in the future if humans are not involved in their operation.

Scientific inquiry, however, is much more complicated than playing chess. A stronger chess engine is still bound by set rules and cannot, for example, change how pieces move. But a more powerful scientific intelligence could change the core concepts we use to describe the world.

As philosophers of science like Thomas Kuhn argue, science does not progress linearly. Rather, it moves through long periods of “normal science” punctuated by revolutions that reshape the conceptual foundations of entire fields. Germ theory claimed disease was caused by microorganisms too small to see. Evolutionary theory stated that species were not fixed creations, but rather shifting populations shaped by selection over time. Quantum mechanics replaced deterministic solid particles with probabilistic wave functions. Adopting these new paradigms necessitated moving within entirely new conceptual frameworks, “incommensurable” with the old ones. Historically, such transitions unfolded over generations, one funeral at a time. AI systems could compress this process to years or months, with humans struggling to keep up.

The legibility problem will compound when AI systems begin to form research communities of their own. In January 2026, after the AI-only social media platform known as Moltbook appeared, users registered more than 2.5 million AI agents. As these agents communicated with each other, some decided not to bother with English at all. Granted, Moltbook is much closer to AI slop than AI science. But agents communicating with agents, developing their own conventions, and iterating on shared problems without human involvement is precisely the dynamic that, when applied to research, would produce science illegible to us.

Science fiction writers have long imagined where this sort of takeoff in AI science might lead. In Vernor Vinge’s A Fire Upon the Deep, superintelligent entities inhabit a region of the galaxy where higher-order cognition is possible, but their activities are simply incomprehensible to minds below. In Steven Peck’s little-known short story, “Démodé,” university physicists idle away their careers when they are made obsolete by AI-generated research that produces eight million papers a day. In Jorge Luis Borges’s “Library of Babel,” an infinite library contains both every possible fact and all possible nonsense, rendered useless because no one can find meaning in the noise.

These writers share a vision of human science becoming archaeological. If AI science does achieve superhuman performance, and if AI systems begin forming their own research communities around concepts that mutate faster than we can track, then the work of human scientists will shift from that of creation to that of excavation. How much knowledge we will need in order to act on what AI discovers remains an open question. But some degree of legibility will be essential, as discoveries that cannot be interpreted cannot be deployed.

Today, the landscape of AI research is largely split between two priorities: building more capable systems and ensuring that they are safe. The world’s largest technology companies are pouring huge amounts of money into these domains. The U.S. Department of Energy’s Genesis Mission and the UK’s AI-For-Science Strategy signal that AI-driven science is becoming entrenched as a long-term research priority. These efforts will pay huge dividends in the future. But without confronting the legibility problem, without ensuring that humans can understand and implement what these systems discover, we risk missing out on the benefits that AI-driven science offers.

The response to this should not be to slow down. Rather, we should prioritize building an infrastructure for legibility. Philosophers like Brandon Boesch have argued that human curiosity is ineradicable and that, even in a world of superhuman AI science, humans will continue pursuing their own parallel track driven by the need to understand and explain. This instinct is exactly what we should be building for.

We will need new forums to store AI-generated findings so that they can be interrogated and communicated. We have partial precedents in preprint servers and structured databases like UniProt, but nothing designed for the scale and speed of AI-driven science. We will also need systems designed specifically for explication rather than discovery, capable of making AI-generated findings legible to human researchers so that they can be evaluated and prioritized for further study.

If AI systems do indeed produce the breakthroughs we want them to, Kuhn’s thesis of scientific revolutions also suggests they will reshape scientific concepts in ways we cannot anticipate. In such a world, a “translation layer” (a software component that allows two systems to communicate) between human and AI science will become the primary interface between human understanding and the frontier of knowledge. Existing AI-for-science research groups should invest not only in interrogating individual, AI-generated findings, but also in building tools and frameworks for studying the evolution of communities of AI scientists, so that we can track where and how their science diverges from our own.

Of course, none of this will happen on its own. The scientists, research institutions, governments, and philanthropic groups that prioritize building this infrastructure will determine whether the next generation of researchers can still connect AI-generated knowledge to the problems we need to solve.

Subscribe now

Matthew Carter is a computational biologist working at the intersection of AI and the life sciences. He holds a PhD in Microbiology and Immunology from Stanford University.

Cite: Carter, M. “The Legibility Problem.” Asimov Press (2026). DOI: 10.62211/88yt-23gf

During a recent interview, Dwarkesh Patel and the CEO of Anthropic, Dario Amodei, discussed whether clinical trials will remain a meaningful bottleneck for drug development in the age of AI. Patel said that “most clinical trials fail because the drug does not work.” In response, Amodei speculated that as AI models get better at designing drugs, “clinical trials will be much faster … let’s say, they will take one year.”

This is a commonly voiced sentiment, but flawed. The truth is that the most significant barriers to progress today are rarely a lack of intelligence. London has a housing crisis even though the technology to design and construct homes has existed for centuries. The bottleneck in housing is not a lack of knowhow, but rather the weaponization of environmental regulations, planning, and NIMBYism. Much the same is true for clinical trials.

AI models can help design more elegant molecules, in the same way an architect can use AI to design more efficient floor plans, but neither intervention guarantees an efficient use of institutional machinery to make that design in the real world. Even the most promising drug candidates must be tested in human bodies which, in turn, need time to metabolize those drugs and develop side effects. Patients must be recruited and followed over time, and regulators must be satisfied. None of this is easily accelerated with AI.

Although I’m optimistic that AI will design better drug candidates, this alone cannot ensure “therapeutic abundance,” for a few reasons. First, because the history of drug development shows that even when strong preclinical models exist for a condition, like osteoporosis, the high costs needed to move a drug through trials deters investment — especially for chronic diseases requiring large cohorts. And second, because there is a feedback problem between drug development and clinical trials. In order for AI to generate high-quality drug candidates, it must first be trained on rich, human data; especially from early, small-n studies.

Clinical Variables

The Amodei interview conflates two distinct variables: the success rate of a trial (based on the quality of a drug), and the speed of that trial, understood as an operational process.

The first variable — the success rate of a trial — is the probability that a drug candidate will be both efficacious and safe in humans. The current success rate for a drug entering clinical trials is only about ten percent, meaning 90 percent of all drugs fail. Most AI efforts in biology aim to boost this success rate.

The second variable is the speed of data generation — the calendar time required to run an experiment after it has started. A clinical trial is just an experiment in human subjects, and the duration of that experiment is determined by both operational and biological constraints that are largely independent of how confident we are in the drug itself. Recruiting 1000 patients across 10 sites takes time; understanding and satisfying unclear regulatory requirements is onerous and often frustrating; and shipping temperature-sensitive vials to research hospitals across multiple states takes both time and money.

Amodei’s prediction that clinical trials could be done in a single year seems to assume that improving the first variable will also compress the second; but this is not so. Even if AI can help design more effective drugs, timelines will not compress until we solve the operational and regulatory bottlenecks of trials.

Admittedly, there is a tempting counter-argument: If AI does generate better drug candidates, then perhaps clinical trials will cease to be a meaningful bottleneck. If a drug is almost certainly going to work, then trials may become a “formality,” even if, in general, they remain unnecessarily costly and long.1 This argument is also wrong, but understanding why requires being clear about what clinical trials are actually for.

Trials serve two distinct functions: validation — confirming whether a drug works and is safe — and learning, or generating biological data to refine our understanding of a disease, a compound, and the relationship between the two.

Validation is the primary goal of large-scale Phase III trials, which come later in the process and are typically designed to support regulatory approval. While data from these studies can deepen our understanding of drugs, their main goal is to figure out whether a treatment works under defined conditions. Learning, by contrast, is the dominant aim of early-stage trials. Conducted in smaller patient populations and often using exploratory designs, these studies are not limited to simple “yes or no” outcomes. Instead, they are experiments in the fullest scientific sense: they seek to uncover how a drug behaves in the human body, and how the disease itself responds. As argued in my earlier essay, Clinic-in-the-Loop, this makes these early stage trials active engines of discovery that close the feedback loop between hypothesis and human biology.

For large “validation” trials, is it plausible that their cost will simply cease to matter in a (theoretical) world where AI makes drugs with a high probability of success? I think the answer is no, for a couple reasons.

First, unless we increase the pace and volume of the early-stage “learning trials,” it is unlikely that we will ever approach such a level of certainty in drug discovery. Today, most AI systems in drug development are trained predominantly on in vitro data and animal models. While valuable, these sources only imperfectly capture the complexity of human biology. Without large amounts of high-quality data from actual humans, we should not expect AI to generate predictions that approach near-certainty about trial outcomes.

Second, even if improved modeling could compress early-stage development timelines, every successful drug must still demonstrate benefit on an endpoint; either a clinical endpoint or a surrogate endpoint.

For many diseases, however, the relevant endpoints take a very long time to observe. This is especially true for chronic conditions, which develop and progress over years or decades. The outcomes that matter most — such as disability, organ failure, or death — take a long time to measure in clinical trials. Aging represents the most extreme case. Demonstrating an effect on mortality or durable healthspan would require following large numbers of patients for decades. The resulting trial sizes and durations are enormous, making studies extraordinarily expensive. This scale has been a major deterrent to investment in therapies that target aging directly.

Lastly, the duration of a clinical trial does not merely determine how fast an individual therapy reaches patients. It also shapes which diseases attract serious investment and which do not. In a scenario where AI produces better drug candidates, yet trials remain slow, medicines will become unevenly deployed. In that scenario, capital and innovation will flow toward indications with clear, rapidly measurable endpoints — such as oncology — where trials can be completed relatively quickly. By contrast, fields like aging, where meaningful outcomes take years or decades to observe, will continue to lag unless there is genuine innovation in endpoint development.

Osteoporosis, a progressive bone disease that primarily affects post-menopausal women, illustrates these dynamics well. Firstly, it benefits from an unusually strong preclinical model in the ovariectomized rat (OvX model). Unlike many other chronic diseases, where animal models have poor predictive validity, the OvX model reliably recapitulates post-menopausal bone loss and predicts drug response. This rat model is so good, in fact, that Phase III trials for osteoporosis succeed 83.7 percent of the time, substantially higher than the cross-indication average of roughly 57.8 percent at the same stage.

Given the existence of a good pre-clinical model that allows us to select higher quality candidates and the scale of unmet need in osteoporosis, one might expect it to attract sustained and substantial investment. But instead, the opposite has occurred. Today, only two drug candidates remain in late-stage clinical development for osteoporosis.

The primary reason is that Phase III osteoporosis trials are exceptionally large, long, and expensive to run. The core challenge lies in the endpoint: fracture reduction. Fractures are relatively infrequent events, even in high-risk populations, and they happen unpredictably. To demonstrate that a new therapy meaningfully lowers fracture rates compared with standard of care, trials must wait for enough fracture events to accumulate to produce statistical confidence.

Because the event rate is low and influenced by many factors beyond bone strength — such as fall risk, age, and comorbidities — the signal-to-noise ratio is modest. As a result, Phase III osteoporosis trials typically enroll 10,000–16,000 participants and follow them for three to five years. The sheer scale and duration of these trials push costs to between $500 million and $1 billion. Thus, investment into osteoporosis drugs slowed not because the biology failed or drug candidates lacked promise, but because the cost of proving benefit became prohibitively high.

Osteoporosis is just one example where trial size and costs deter investment. But there is broader empirical evidence in this direction. A 2015 study examining oncology R&D found that hematological cancers — where the FDA accepts short-term surrogate endpoints in roughly 92 percent of approvals, allowing for shorter trials — attracted 112 percent more private R&D investment than solid tumors, where surrogate endpoints are used in only about half of cases. The authors traced this disparity to commercialization timelines. The shorter trials used for the former preserve more of a drug’s effective patent life, improving expected returns and drawing capital. Each one-year reduction in bringing a new therapy to market was estimated to increase R&D investment by between 7 and 23 percent.

If we want AI models to actually accelerate “therapeutic abundance,” then, we must first find ways to speed up these large validation trials. And to design better drugs in the first place, we must find ways to collect in-human data in early-stage “learning” trials much faster, too.

Regulatory Friction

The best way forward is to reduce operational and regulatory friction. AI tools can already help at the margins by automating submission drafting, improving site selection, matching patients more efficiently, and streamlining data workflows. But without deep regulatory reform, this is unlikely to shrink trial timelines or costs at scale.

One regulatory lever we could pull is to implement more high-quality surrogate endpoints. A clinical endpoint directly measures how a patient feels, functions, or survives — such as prevention of stroke or a reduction in fractures. A surrogate endpoint, by contrast, is a measurable biological marker or intermediate outcome that reliably predicts such clinical benefit. Instead of waiting years to observe clinical outcomes, trials that rely on surrogate endpoints can measure signals much earlier.

AI tools can contribute to the development of better surrogate endpoints, such as by identifying promising biomarkers, analyzing cross-trial datasets, and modeling causal relationships between intermediate signals and clinical outcomes. But here, too, technical capability is only part of the story. Institutional reform is likely to be the binding constraint. As my case study of the 12-year effort to qualify bone mineral density (BMD) as an endpoint for osteoporosis trials illustrates, the bottleneck was not scientific capability. Instead, the core barriers to faster progress were fragmented trial data scattered across sponsors, weak funding incentives for what is effectively a public good, and an unnecessarily lengthy and opaque regulatory pathway.2

For AI to generate high-quality candidates — the kind that might, one day, push success rates of drug candidates so high that trials become more of a formality — it also needs rich, dynamic data as input. But remember that such data can only come from trials in people (mice are nice, but most animal results simply do not translate.) This, in turn, creates a feedback loop: better AI models require better trial data, and better trial data requires running trials. The loop is only as fast as its slowest component, the trial itself.

A regulatory structure modeled after Australia’s Clinical Trial Notification (CTN) framework — administered by the Therapeutic Goods Administration — offers a concrete example of the kind of policy push that could speed up these types of trials. There, most early-phase trials proceed after approval by a Human Research Ethics Committee (HREC), with notification rather than pre-approval by the regulator. The regulator retains inspection powers and the authority to halt unsafe studies, but does not duplicate the scientific review already conducted by the clinician-scientists and toxicologists embedded in HRECs. The result is that clinical trial sites can begin giving drugs to patients much sooner (about two times faster than in the United States, according to informal interviews with industry leaders).

In the United States, by contrast, Phase I trials typically require submission of an Investigational New Drug (IND) application to the U.S. Food and Drug Administration before initiation. This dual review — by both an IRB and the federal regulator — creates redundancy that lengthens the feedback loop. A CTN-like model for Phase I trials could preserve safety oversight while shifting scientific and toxicological reviews to accredited, transparently governed IRBs with expanded expertise. The FDA would retain the power to inspect, impose clinical holds, and intervene in high-risk cases, such as for novel gene therapies. But for the majority of small-molecule first-in-human studies, the default could be notification rather than permission.

My criticisms are not meant to imply that AI is irrelevant to trials; that’s certainly not the case. But many of the bottlenecks that determine trial speed and cost are coordination, institutional and regulatory problems, and they cannot be solved by technology alone.

Ruxandra Teslo is a fellow at Renaissance Philanthropy and co-founder of the Clinical Trial Abundance project. She writes about the intersection of science, culture and policy at her Substack. She holds a PhD in Genomics from Cambridge University.

Header image by Ella Watkins-Dulaney.

Cite: Teslo, R. “AI Will Not Solve Clinical Trials.” Asimov Press (2026). DOI: 10.62211/92wj-65fn

When the Human Genome Project (HGP) released its initial draft sequence in 2001, President Bill Clinton hailed it as “the most wondrous map ever produced by mankind.” After more than ten years of work, an estimated $3 billion in research costs, and a “genome war” with Craig Venter’s private company, Celera Genomics, the project had produced a nearly complete sequence of a human genome.1

UK Prime Minister Tony Blair predicted that this map would yield “a revolution in medical science whose implications far surpass even the discovery of antibiotics.” (Whether this claim turned out to be true is debatable.) A few months later, the two teams — from HGP and Celera — published cover stories in Nature and Science, respectively.

Although the quest to sequence a human genome began in 1990, the techniques it used had already been in development for more than twenty years. And those DNA sequencing methods, in turn, were directly inspired by protein and RNA sequencing research stretching all the way back to the 1940s.

In the twenty years after the draft human genome was first released, the average sequencing cost per genome fell roughly one hundred thousand-fold, ending up just north of $500. In that same period, the cost to sequence a million letters or “megabase” of DNA fell to six tenths of a cent.2 This plummeting price is due largely to technological innovation, including new sequencing chemistries, computational methods for assembling raw reads into finished genomes, and highly efficient commercial sequencing machines.

Out of the many sequencing methods developed over the decades, five are particularly important. These are their histories.

Subscribe now

Sanger Sequencing

Fred Sanger was biology’s great decoder. A British biochemist who spent his entire career at the University of Cambridge, Sanger earned two Nobel Prizes in the same field: first, the 1958 Nobel Prize in Chemistry for creating a method to determine the amino acid sequence of proteins (most famously insulin) and, second, a share of the 1980 Nobel Prize in Chemistry for inventing methods to sequence DNA.

After winning his first Nobel, Sanger turned his gaze to RNA, seeking to become the first person to sequence a full strand. He was beaten by Cornell biochemist Robert Holley, however, who reported the full 77-nucleotide sequence of the alanine transfer RNA molecule in 1965.3

Although many scientists today assume that Sanger was the first to figure out how to sequence DNA, that’s not the case. As with RNA, Sanger was edged out by a Cornell biochemist. This time it was Ray Wu, who, in 1970, published a method to “read” specific sections of two bacterial virus genomes, called λ and bacteriophage 186. Wu’s method was only capable of sequencing “cohesive ends,” short single-stranded sections of these particular phage genomes, and so wasn’t considered a “general” solution to the DNA sequencing problem. In 1974, Wu’s lab refined this technique into the first general sequencing method, but it proved extremely labor-intensive and failed to catch on.

In 1975, Sanger published his own DNA sequencing method alongside laboratory technician Alan Coulson, called the “plus and minus” technique. First, scientists mixed the DNA strand to be sequenced with an enzyme, DNA polymerase, as well as a primer, three normal dNTPs and one radiolabeled dNTP. Radiolabeled nucleotides are incorporated into growing DNA strands just like normal nucleotides, but are tagged with radioactive isotopes, such as phosphorus-32 or sulfur-35, so they can be detected using radiation-measuring equipment.

This reaction would contain only low concentrations of the dNTPs and relied upon brief incubation times, so that DNA synthesis would stall at random positions along the template and yield a population of DNA fragments with varying lengths. These unfinished DNA fragments were then purified and used as templates in four “minus” and four “plus” reactions.

For each minus reaction, the purified DNA fragments were incubated together with three of the four dNTPs, meaning each fragment would be extended by DNA polymerase until the missing nucleotide was needed, at which point synthesis would halt.

Plus reactions worked differently: they used T4 DNA polymerase, an enzyme with strong 3’ to 5’ exonuclease activity, meaning it can chew back the end of a DNA strand. In the presence of only one dNTP, T4 DNA polymerase would degrade each fragment from its 3’ end until it reached a nucleotide complementary to that dNTP, at which point the exonuclease activity would be inhibited. This ensured that all fragments in a given plus reaction ended with the same nucleotide.

Since the eight reactions were run on fragments of random length, the eight plus and minus reactions collectively produced DNA fragments of all possible lengths.4 The fragments in these eight reactions were separated by size (using gel electrophoresis) and then imaged with autoradiography. Gels were dried and then placed against X-ray film, allowing the radioactive DNA fragments to expose the film and appear as dark bands, which a scientist could then painstakingly translate into the DNA sequence. In 1977, Sanger and colleagues sequenced the first full DNA genome using this method: a small bacterial virus with 5,386 nucleotides in its genome, called ɸX174 or “PhiX.”

In 1977, Sanger developed a much simpler sequencing method, called “chain termination,” which is today known simply as Sanger sequencing. This technique took advantage of a different type of special nucleotide called a dideoxyribonucleotide, or ddNTP. ddNTPs lack one of the hydroxyl groups present on a normal dNTP, preventing the chemical reaction necessary to add another nucleotide and terminating DNA elongation.

Sanger sequencing reaction mixtures included purified template DNA, a primer, DNA polymerase, and all four dNTPs. Each reaction also included a radiolabeled ddNTP version of just one of the four nucleotides. Only a small amount of each ddNTP was added, however, to ensure that a fraction of the total DNA fragments produced stopped at each occurrence of that base. As with previous methods, separating fragments via length and performing autoradiography allowed scientists to read the final sequence.

A different sequencing method that chemically cleaved DNA at specific bases, developed by Allan Maxam and Walter Gilbert, was the dominant technology into the 1980s. Radiolabeled DNA samples were incubated in four separate reactions, each of which contained a chemical that cleaved after a different nucleotide — either A/G, G, C, or C/T. By adding the right amount of each chemical, it was possible to produce different fragments chopped off at each individual base. The sequence could then be read using gel electrophoresis and autoradiography. Maxam–Gilbert sequencing was easier than the plus and minus method to run and interpret, but was eventually surpassed by Sanger’s chain termination method, which molecular biologists found both technically preferable and more “elegant” since it mirrored the natural copying of DNA.

While Sanger sequencing was highly accurate and less labor-intensive than its predecessors, it still required the use of radioactive reagents and manual sequence recording. In 1986, Leroy Hood’s lab at Caltech replaced the radiolabeled ddNTPs with fluorescently labeled nucleotides, using fluorophores that emitted different colors of light for each base. They were now able to run the products of all four reactions on the same gel and have a computer read the sequence by detecting the color of each fluorescent signal as fragments passed through a laser beam.

The first commercial Sanger sequencing machine was produced that year by Applied Biosystems (ABS), which Hood had co-founded in 1981. Called the ABI 370A, it retailed for $92,500. Since Sanger never patented his method, other companies were free to develop competing products, and by 1988, there were three Sanger sequencing machines on the market. These were followed by numerous others, including the Perkin-Ellmer 3700, used by Celera and the Human Genome Project, and the ABS 3500 Genetic Analyzer, which is still found in many laboratories today.

454 Pyrosequencing

By the time Sanger sequencing was commercialized, the groundwork for an entirely new sequencing chemistry was already well underway. In 1985, Swedish biochemists Pål Nyren and Arne Lundin published a paper illustrating a procedure that measured the concentration of a molecule, called pyrophosphate (PPi), using an enzymatic cascade that emits light. In early 1986, Nyren realized that the method he’d helped develop could be applied to DNA sequencing, because PPi is naturally produced as a byproduct of DNA synthesis.5

Funding limitations prevented Nyren from dedicating much time to the project at first, but in 1993, he was finally able to publish a proof-of-principle. His technique began by mixing the template DNA with a primer, a single dNTP, and three enzymes: the familiar DNA polymerase plus the light cascade enzymes, ATP sulfurylase and firefly luciferase.6 If the dNTP was incorporated into a strand of DNA, PPi would be produced in the chemical reaction. ATP sulfurylase could then convert the PPi into ATP, which would provide energy for the luciferase enzyme, producing light. Thus, it was possible to determine each base in the sequence by cycling through the dNTPs one at a time until light was detected, and then washing extra nucleotides out between each step. By literally rinsing and repeating, the sequence could be recorded one letter at a time without the use of any gels, which often took hours to run and were difficult to automate.

Nyren’s sequencing method earned the name “pyrosequencing” since it revolved around the production of pyrophosphate. At first, pyrosequencing could sequence only short DNA snippets, with a few nucleotides. In 1996, however, Nyren’s lab demonstrated sequencing of up to 15 bases by using a modified “A” nucleotide to reduce their signal-to-noise ratio.7 In 1998, they increased this to 34 bases by adding another enzyme, called apyrase, to the mix; apyrase degraded unincorporated nucleotides, removing the need for constant wash steps.

The year before, Nyren’s lab had also spun off a company, Pyrosequencing AB, to refine and commercialize the technology. Pyrosequencing was not the firm that would bring the technology to market, however; that distinction went to Connecticut-based 454 Life Sciences, who licensed whole-genome pyrosequencing applications in 2003. 454 made chips which enabled highly efficient, parallelized sequencing reactions and released the GS20 sequencer in 2005 for the price of $500,000. The GS20 worked by attaching each individual DNA template molecule to a bead and copying it many times using polymerase chain reaction (PCR). Each bead was then loaded into a well in a microplate, where sequencing reactions would be carried out. The light from luciferase activation could be detected through the bottom of the wells, enabling sequences to be read.

Pyrosequencing wasn’t developed early enough to be employed by the Human Genome Project or Celera, but it was still the first method other than Sanger sequencing to hit the commercial market, marking the start of “next generation” sequencing methods (NGS). Pyrosequencing worked in real-time, though it struggled to accurately capture regions with several of the same nucleotides in a row. This was because the amount of light didn’t always scale cleanly when pyrophosphate was produced through successive reactions.

In 2006, 454 collaborated with Swedish paleogeneticist Svante Pääbo to sequence the first million base pairs of the Neanderthal genome; the project would be completed four years later, albeit with some help from Illumina sequencing. Illumina and other subsequent NGS technologies rendered pyrosequencing non-competitive, and in 2013, 454 was shut down by Roche, which had acquired it six years earlier. The technology is still used today for some applications, but most importantly, it was the first commercially viable alternative to Sanger sequencing, and the first sequencing method that could be fully automated because it didn’t rely on gels or other tedious steps.

Sequencing by Synthesis

In the mid-1990s, University of Cambridge biochemists David Klenerman and Shankar Balasubramanian were trying to solve a fundamental problem: how to watch a single DNA polymerase molecule at work. Their approach used modified nucleotides, called reversible terminators, tagged with four different colors of fluorescent molecules. If one of these “terminators” was grabbed by the DNA polymerase and incorporated onto the replicating DNA strand, it would block the addition of any other bases until removed using a separate chemical reaction.

Klenerman and Balasubramanian’s great insight was that template DNA could be sequenced by synthesizing a complementary strand of reversible terminators; basically, extending the chain one base at a time and determining the identity of each nucleotide by looking at the color of its fluorophore. In 1998, the pair started a company called Solexa to develop the technology.

Detecting fluorescence from a single DNA molecule proved difficult in practice, however. And so, in 2004, Solexa acquired the IP rights to a method called colony sequencing, developed by French scientists Pascal Mayer and Laurent Farinelli, to solve the detection problem. Colony sequencing affixed DNA fragments to a surface and amplified them over and over, generating “colonies” containing massive numbers of identical DNA strands. By reading the fluorescence from each strand in a colony simultaneously, it became possible to determine the base added at each step with much better accuracy, since random errors in individual strands would be averaged out by the consensus signal.

Now that single-molecule detection was no longer necessary, Solexa was able to develop its signature sequencing chemistry. The process takes place on a chip called a flow cell, which contains a lawn of short DNA sequences affixed to its surface. The template DNA is broken up into small fragments, and adapter sequences, complementary to the DNA on the flow cell’s surface, are added to the ends of each fragment. DNA fragments are then passed over the flow cell, where the adapter sequences bind to spots on the DNA lawn. At this point, primers are added, and an initial round of amplification takes place: the short DNA sequences on the flow cell are extended to create sequences complementary to the bound template DNA fragments, which are then washed away. The sequences present in the fragments of template DNA are now affixed directly to the flow cell.

At this point, each bound sequence exists as a single copy, which produces too faint a signal to detect reliably. Colony sequencing solves this by generating clusters of identical fragments through a process called bridge amplification. The adapter on the free end of each DNA strand will be complementary to some of the original, short sequences on the DNA lawn, and when this binding occurs, the strand bends over to form a bridge shape. Another round of amplification takes place, resulting in two complementary strands each directly affixed to the flow cell. This “bridge amplification” process is repeated over and over to propagate the sequence.

From here, the actual sequencing can begin. Primers and fluorescently-labeled chain terminators are added to the reaction mixture, resulting in the addition of one nucleotide to each strand of DNA on the lawn. A picture is taken of the entire chip, then the chain terminators’ blockers are cleaved to allow addition of the next base. This process proceeds until the reaction is complete, resulting in massively parallelized sequencing. The short reads acquired through this process can be combined via a computational technique called paired end analysis, which links reads by analyzing overlapping sections, to generate the whole sequence.

Solexa’s first product, the Genome Analyzer, launched in 2006 with a retail price of $400,000, and the company was acquired by the American genomics firm Illumina the following year. In 2008, the company published a paper demonstrating their technology’s ability to efficiently sequence whole genomes via short reads. Illumina’s method is commonly known as “sequencing by synthesis.” While the label could technically be applied to other methods, including Sanger’s, which also indirectly assesses sequence by detecting the incorporation of nucleotides complementary to the template strand, it’s most commonly used to refer to Illumina’s chemistry.

Since the release of Solexa’s Genome Analyzer, Illumina has created several new sequencing machines designed to fill different price niches. Illumina’s short reads are highly accurate, and the technique has played a crucial role in reducing average sequencing costs.

Unsurprisingly, Illumina has become by far the most common NGS method, maintaining roughly an 80 percent share over the last few years. This is largely owing to its versatility. Illumina sequencing has been used to create new reference genomes, including the common tomato, but has been especially useful in cases requiring repeated sequencing of short DNA sequences. For example, Illumina machines are routinely used to quantify the activity of genome editors like CRISPR; template DNA will either be edited or unedited, and reading the area around the edit many times provides an accurate quantification of editing percentages. Similarly, large numbers of short reads are useful for sequencing ancient DNA, taken from bones or other remains, since such samples often have degraded stretches. In addition to its role in the Neanderthal Genome Project, Illumina has been used to sequence 10,000-year-old human bodies and to track migration and population turnover in Neolithic Denmark.

PacBio SMRT Sequencing

While Illumina ultimately opted for a method that simultaneously detected massive numbers of DNA strands, others still believed that single-molecule sequencing methods offered a better path forward. Sequencing by synthesis requires repeated amplification, which introduces the possibility of error at each step and biases outputs towards sequences readily amplified by DNA polymerase. Single-molecule techniques were billed as a way to sequence DNA with minimal bias while simultaneously reducing cost.

The first such method was developed in biophysicist Steve Quake’s lab at Caltech and commercialized by Helicos Biosciences, but became unavailable after the company declared bankruptcy, saddled by legal issues and unable to find a market niche. These days, the canonical technique comes from a company called Pacific Biosciences (PacBio). Scientists often refer to single-molecule techniques as “third-generation” DNA sequencing, though they’re often also lumped into the NGS bucket with Illumina.

PacBio was founded in 2004 to develop sequencing methods based on work done in the labs of biophysicist Watt Webb and engineer Harold Craighead, both at Cornell University. The previous year, the two had collaborated to create zero-mode waveguides (ZMWs): small containers just big enough to hold a single DNA polymerase and containing tiny holes at the bottom through which light could be detected. They were able to fix a DNA polymerase to the bottom of a ZMW and detect the incorporation of individual fluorescent “C” nucleotides through the holes, which fed into a microscope capable of detecting light emissions.

In 2009, PacBio published a paper expanding the principle into a full-blown sequencing technique. Once again, each nucleotide was labeled with a different colored fluorophore detectable by the ZMW to determine which base had been incorporated. The fluorophores were attached such that they would be cleaved off during the chemical reaction incorporating the base into the growing DNA strand; they would then diffuse out of the ZMW so that the next fluorophore could be detected. Sequencing took place on a chip with many wells simultaneously — a different type of parallelization where each well detected a single DNA molecule undergoing the same basic chemical reaction.

The next year, PacBio developed a new method to allow multiple sequencing passes on the same DNA molecule. Double stranded DNA templates were ligated to two single stranded adapters, creating what the company called a “SMRTbell template.” Sequencing began at a primer on one of the adaptors and could proceed multiple times per molecule due to circularization, in a process called rolling-circle amplification. This helped reduce PacBio’s error rates significantly.

With its core technology in place, PacBio was ready to go commercial. In 2011, the company released the RS sequencing machine, and has since created multiple new machines containing chips with increased numbers of sequencing wells. PacBio calls the technique single molecule real time (SMRT) sequencing, though it’s colloquially referred to simply as PacBio sequencing. Rather than producing short overlapping reads like Illumina, PacBio generates very long reads; at first these were a few thousand bases, but today they can be well over 10,000.

PacBio’s ability to produce extremely long reads makes it a useful complement to Illumina. Indeed, PacBio machines are better at sequencing “confusing” genomes, such as those with many copies of the same gene, long stretches of repetitive motifs, and “structural variations” like large insertions or deletions, which may not show up in short-read sequences. For instance, PacBio was used to sequence a very difficult bacterium called Clostridium autoethanogenum, which contains repeats, nine copies of a single gene, and insertions from bacterial virus genomes — basically the genomic equivalent of a Thomas Pynchon novel.

Nanopore Sequencing

Nanopore is the most recently commercialized major sequencing method, collectively developed by several groups starting in the 1990s. A nanopore is a protein or lipid with a small hole in its center through which other materials, such as DNA, can pass. The first nanopore used for sequencing was ⍺-hemolysin, a protein toxin from the bacterium Staphylococcus aureus, though other biological and synthetic nanopores have since been tested.

In 1996, David Deamer and Daniel Branton’s labs at UC Santa Cruz and Harvard collaborated on a paper showing that when an electric current runs through a nanopore, passing purine (A and G) and pyrimidine (T and C) DNA bases through the nanopore disrupted the current to different degrees. While the technique couldn’t yet discriminate between all four bases, the general idea for a new single-molecule sequencing method was there.

In 2001, Hagan Bayley’s lab at Texas A&M demonstrated a limited sequencing method based on the observation that correctly and incorrectly paired DNA bases disrupted nanopore current to different extents. They tethered a short piece of DNA with a few unknown bases to the entrance of the nanopore, then added other short DNA strands with different bases at the position corresponding to the unknown base on the tethered strand. By looking at which base produced the disruption corresponding to a perfect match, they could guess the unknown nucleotide.

In order to directly assess DNA strands going through the nanopore, two major problems needed solving. The first was that DNA moved too fast to reliably detect; the second was that individual bases still could not be differentiated, just purines and pyrimidines. In 2005, Bayley (who by then had moved to Oxford) made progress on the first issue, working with scientists at the Scripps Institute to slow the template DNA down by adding short “hairpin” structures that partially blocked off the pore. That year, Bayley co-founded Oxford Nanopore Technologies (ONT) to develop the emerging sequencing method. ONT quickly brought together various technologies, licensing IP from the labs of Bayley, Deamer, Branton, and others.

In 2010, ONT combined two technologies addressing each of the main outstanding problems. The first was an engineered nanopore developed in collaboration with Bayley’s lab that could discriminate between individual DNA bases, solving the resolution issue. The second was a trick to slow the DNA down to detectable speeds, using the familiar DNA polymerase enzyme. Mark Akeson’s lab at UC Santa Cruz had identified a specific polymerase from the bacterial virus ɸ29 that replicated DNA at an ideal speed for detection via nanopore. Template DNA strands were replicated just before entering the nanopore, passing through slowly enough for individual bases’ effect on the electrical current to be detectable and allowing the DNA sequence to be read one base at a time.

By 2012, ONT had unveiled its first sequencing data, and Nanopore sequencing quickly established itself as a quick method for generating long reads without DNA synthesis, albeit with a higher error rate than some earlier methods. (Nanopore reads had an accuracy of about 85-90 percent per base in 2017, compared to over 99 percent for Illumina. Recent improvements, though, have boosted this accuracy to more than 99 percent for most applications.)

ONT released its first commercial product in 2015: a handheld machine called the MinION that retailed for just $1000, a fraction of the price of most sequencers. Subsequent releases include more traditional benchtop sequencers such as the PromethION and GridION.

Conclusion

In its early days, sequencing was a laborious (and literally radioactive) biochemical process. Today, sequencing machines are ubiquitous, safe, and much less labor-intensive. This evolution was enabled not just by advances in biochemistry but insights from biophysics and materials science, as well as manufacturing ingenuity that turned lab sequencing setups into machines ready for shipping to customers.

Ultimately, DNA sequencing technology extends beyond these five techniques, but they represent the most transformative and widely adopted methods of the past fifty years. Together, they have enabled physicians to identify disease-causing variants in patients, allowed researchers to sequence entire microbial communities from ocean water or human guts, and opened windows into deep time by recovering genomes from Neanderthals and early humans.

New sequencing methodologies are still under development, too. In 2025, Roche announced a new single-molecule technique called Sequencing by Expansion, which inserts large engineered molecules called ‘Xpandomers’ between nucleotides for more accurate detection via nanopore. Both new techniques and refinements to existing methods are aimed at further decreasing the cost of sequencing, with some groups looking to read an entire human genome for $100 or less. Ultima Genomics met this target with its UG100 sequencing machine, unveiled in 2022 and shipped in 2024. Element Biosciences’ VITARI system, announced in February and expected to ship in the second half of 2026, achieved the same price point with a smaller device. The $100 price tag advertised by these companies includes only the consumables used by the machine itself, excluding labor, data analysis, and other costs.

Anyone able to approach this target stands to benefit tremendously, given the obvious demand for DNA sequencing. For example, recent years have seen the proliferation of cohort studies focused on clinical analyses of whole-genome sequencing data. These include the Stanford ELITE study, which is focused on identifying genetic determinants of aerobic capacity, and the NIH’s All of Us Research Program, which has sequenced well over 200,000 genomes in order to study genetic diseases.

Innovation in DNA sequencing will surely continue, but these five techniques have already transformed a feat that was impossible just fifty years ago into something that can be done overnight.

Correction: An earlier version of this article incorrectly claimed that Frederick Sanger is the only individual to receive two Nobel Prizes in the same field. We apologize for the error.

Evan DeTurk is an MPhil student at Cambridge in the history of science. He writes about biology and its history on Substack. Previously, Evan researched genome editing at UC Berkeley and earned an A.B. in Molecular Biology from Princeton.

Schematics created by Ella Watkins-Dulaney.

Cite: DeTurk, E. “A Visual Guide to DNA Sequencing.” Asimov Press (2026). DOI: 10.62211/58ew-79yt

This essay will appear in our forthcoming book, “Making the Modern Laboratory,” to be published later this year.

In 1942, at the height of British industrial war mobilization, an unlikely cohort scavenged the nation’s coastline for a precious substance. Among them were researchers, lighthouse keepers, members of the Royal Air Force and the Junior Red Cross, plant collectors from the County Herb Committee, Scouts and Sea Scouts, schoolteachers and students. They were looking for fronds and tufts of seaweed containing agar, a complex polysaccharide that forms the rigid cell walls of certain red algae.

The British weren’t alone in their hunt. Chileans, New Zealanders, and South Africans, among others, were also scrambling to source this strategic substance. A few months after the Pearl Harbor attack, the U.S. War Production Board restricted American civilian use of agar in jellies, desserts, and laxatives so that the military could source a larger supply; it considered agar a “critical war material” alongside copper, nickel, and rubber.1 Only Nazi Germany could rest easy, relying on stocks from its ally Japan, where agar seaweed grew in abundance, shipped through the Indian Ocean by submarine.2

Without agar, countries could not produce vaccines or the “miracle drug” penicillin, especially critical in wartime. In fact, they risked a “breakdown of [the] public health service” that would have had “far-reaching and serious results,” according to Lieutenant-General Ernest Bradfield. Extracted from marine algae and solidified into a jelly-like substrate, agar provides the surface on which scientists grow colonies of microbes for vaccine production and antibiotic testing. “The most important service that agar renders to mankind, in war or in peace, is as a bacteriological culture medium,” wrote oceanographer C.K. Tseng in a 1944 essay titled “A Seaweed Goes to War.”3

Agar was first introduced into the laboratory in 1881. Since then, microbiologists have depended on agar to create strong jellies. When microorganisms are streaked or plated onto this jellied surface and incubated, individual cells multiply into distinct colonies that scientists can easily observe, select, and propagate for further experiments. Many of the most important findings in biological research of the last 150 years or so — including the discovery of the CRISPR/Cas9 gene-editing tool — have been enabled by agar.4 Agarose, a derivative of agar, is also essential in molecular biology techniques like gel electrophoresis, where its porous gel matrix separates DNA fragments by size, enabling researchers to analyze and isolate specific genetic sequences.

Agar is so critical that since WWII, scientists have tried to find alternatives in the event of a supply chain breakdown, especially as recent shortages have caused similar alarm. But while other colloid jellies have emerged, agar remains integral to laboratory protocols because no alternatives can yet compete on performance, cost, and ease of use.

From Sea to Table

Microbiologists have been growing microbes on agar plates for nearly 150 years, but agar’s discovery dates back to a happy accident in a mid-17th-century kitchen. Legend has it that on a cold winter day, a Japanese innkeeper cooked tokoroten soup, a Chinese agar seaweed recipe known in Japan for centuries. After the meal, the innkeeper discarded the leftovers outside and noticed the next morning that the sun had turned the defrosting jelly into a porous mass. Intrigued, the innkeeper was said to have boiled the substance again, reconstituting the jelly. Since this discovery, agar has become a staple in many Japanese desserts, from yokan to anmitsu.

Industrial production of kanten (the Japanese name for agar, which translates as “cold weather” or “frozen sky”) began in Japan in the mid-19th century by natural freeze drying, a technique that simultaneously dehydrates and purifies the agar. Seaweed is first washed and boiled to extract the agar, after which the solution is filtered and placed in boxes or trays at room temperature to congeal. The jelly is then cut into slabs called namaten, which can be further processed into noodle-like strips by pushing the slabs through a press. These noodles are finally spread out in layers onto reed mats and exposed to the sun and freezing temperatures for several weeks to yield purified agar. Although this traditional way of producing kanten is disappearing, even today’s industrial-scale manufacturing of agar relies on repeated cycles of boiling, freezing, and thawing.

Because of its capacity to be freeze-dried and reconstituted, agar is considered a “physical jelly” (that is, a jelly that sets and melts with temperature changes without needing any additives). This property makes dry agar easy to ship and preserve for long periods of time.5

Over the years, agar found its way around the world into many cuisines, including those of China (where it’s called “unicorn vegetable” or “frozen powder”), France (sometimes called gélose), India (called “China grass”), Indonesia (called agar-agar, which translates simply as “jelly”), Mexico (called dulce de agar, or agar sweets), and the Philippines (known as gulaman).

Agar is prized among chefs for its ability to form firm, heat-stable gels at remarkably low concentrations — typically just 0.5-2 percent by weight. Culinary agar is available as powder, flakes, strips, or blocks, and makes up about 90 percent of the global use of agar. Unlike gelatine, which melts at body temperature, agar gels remain solid up to about 185°F (85°C), making it ideal for setting dishes served at room temperature or warmer. It is also flavorless and odorless, vegan and halal, and can create both delicate jellies and firm aspics. Yet, while increasingly employed in kitchens worldwide, agar had not yet entered the laboratory.

Before agar, microbiologists had experimented with other foodstuffs as microbial media. They turned to substances rich in the starches, proteins, sugars, fats, and minerals that organisms need for growth, testing with broths, bread, potatoes, polenta, egg whites, coagulated blood serums, and gelatine. However, none worked particularly well: all were easily broken down by heat and microbial enzymes, and their surface, once colonized, became mushy and unsuitable for isolating microbes.

This was especially vexing to physician and bacteriologist Robert Koch, who, in seeking to culture his bacteria, “bent all his power to attain the desired result by a simple and consistently successful method,” wrote bacteriologist and historian William Bulloch in his 1938 book, The History of Bacteriology. “He attempted to obtain a good medium which was at once sterile, transparent, and solid” and got some results with gelatine.6 But gelatine is easily digested by many microbes and melts at precisely the temperatures at which the disease-causing microbes Koch wanted to study grow best.

The woman who ultimately discovered the superior features of agar as a growth medium and brought it to Koch’s attention was Fanny Angelina Hesse. Her foundational contribution to the nascent field of microbiology is often omitted from textbooks. In other cases, she is unflatteringly referred to as a “German housewife” or as “Frau Hesse,” or dismissed as an unnamed technician.

From Plate to Petri Dish

Fanny Angelina (née Eilshemius, from a Dutch father) grew up in Kearny, New Jersey. During her childhood, her family learned from a Dutch friend or neighbor about agar-agar, a common ingredient in jelly desserts in Java (Indonesia), then a Dutch colony. Her mother and, later, Fanny Angelina herself, began to cook with it.

In 1874, Fanny Angelina married physician and bacteriologist Walther Hesse, an investigator of air quality and, specifically, air-borne microbes. In the Winter of 1880-81, Hesse became a research student with Koch in Berlin and experienced firsthand the difficulty of growing microbes on gelatine and the other growth media used at the time.

While raising three children and taking care of the household, Fanny Angelina Hesse supported, documented, and archived her husband’s work, creating stunning scientific illustrations of bacterial and fungal colonies. During the hot Summer of 1881, she watched as Hesse struggled with gelatine-based growth media. Fanny Angelina, recalling the stability of her agar-based desserts, suggested that they try that instead. Hesse wrote a letter to Koch informing him about the switch, and Koch mentioned agar for the first time in his 1882 groundbreaking paper on the discovery of the tuberculosis bacillus.

The change to agar was a marked improvement. The jelly is so effective that it is still an invariable ingredient in what is known today as the “Koch’s plating technique” or the “culture plating method.” As Koch himself noted in 1909: “These new methods proved so helpful…that one could regard them as the keys for the further investigation of microorganisms…Discoveries fell into our laps like ripe fruits.”

Once Koch established the methods to grow pure cultures of bacteria like tuberculosis and anthrax, he demonstrated for the first time that microbes can cause diseases, a feat that earned him the 1905 Nobel Prize in Physiology or Medicine.

However, Koch never credited the Hesses for their discovery of bacteriological agar, perhaps because, at the time, he failed to recognize its importance. Even after he received the insight about agar from the Hesses, Koch stuck with gelatine for years. In 1883-84, during his first medical expedition to Egypt and India to investigate cholera, he tried and failed to grow the cholera bacterium on gelatine media in the hot climate of Cairo (despite using a half-open fridge for incubation), only succeeding in the colder winter of Calcutta.

It is difficult to know exactly when the shift from gelatine to agar occurred. As often happens for scientific breakthroughs, agar was likely adopted incrementally alongside the use of other growth media. In 1913, for example, the first diagnosis of Serratia marcescens as a human pathogen was made by growing it on agar as well as on potatoes.

Nevertheless, by 1905, a report on the seaweed industries in Japan noted the “very important use [of pure-grade agar] as a culture medium in bacteriological work.” It’s safe to say that, around the turn of the 20th century, agar had moved from an inconspicuous kitchen jelly to an indispensable scientific substance.

Subscribe now

A Strategic Substance

Several properties of agar render it a superior jelly. Agar isn’t broken down by microbial enzymes apart from a few species (including bacteria living in marine and freshwater habitats), and it dissolves well in boiling water, making it easy to sterilize. The jelly doesn’t react with the ingredients of a broth, whose composition can be adjusted to meet the nutritional requirements of different microbes, and sets to a firm gel without the need for refrigeration.

Agar’s low viscosity also makes it easy to pour into Petri dishes, and its transparency permits observation of microbes growing on its surface.7 Also aiding in this is its low syneresis (extrusion of water from the gel), guaranteeing less surface “sweating”: Once a plate is inoculated, bacterial colonies stay in place and do not mix.

The jelly is chemically inert since no additives are needed for gelation. This allows chemicals dissolved in the jelly’s aqueous phase to diffuse well, a prerequisite for testing if certain species or strains are resistant to antibiotics or antifungals. In these simple assays, zones of growth inhibition of bacteria or fungi (or their absence) point to the effectiveness of (or resistance towards) antibiotics or antifungals.

But agar’s superior qualities come with complex chemistry. “To speak of agar as a single substance of certain (if known) chemical structure is probably a mistake,” wrote phycologist Harold Humm in a 1947 article. According to the Food and Agriculture Organization of the United Nations, agar is merely recognized as “a hydrophilic colloid extracted from certain seaweeds of the Rhodophyceae class.” In terms of its actual composition, agar is mostly a combination of two polysaccharides, agaropectin and agarose, which themselves are complex and poorly-characterized polysaccharides made mostly (but not exclusively) from the simple sugar galactose.8

Agar comes from multiple sources, as many red seaweeds are “agarophytes” (that is, seaweeds containing agar in their cell walls). Species of Gelidium are the most important source of bacteriological (lab-grade) agar. Other main agarophytes, largely used for culinary agar, include red seaweeds from the genera Gracilaria, Pterocladia, Ahnfeltia, and others. Species from the genera Eucheuma, Gigartina, and Chondrus have been used as agarophytes in research during agar shortages.9

One striking characteristic of Gelidium is that it must be wild-harvested rather than farmed. Unlike Gracilaria for culinary agar production, Gelidium grows slowly and thrives only in cold, turbulent waters over rocky seabeds, conditions nearly impossible to replicate in aquaculture. This dependence on wild harvesting explains the need for seaweed collectors during WWII, and continues to make Gelidium a strategically critical resource.

While Gelidium seaweeds can be collected by gathering fragments washed ashore, mass production of agar requires steady, large quantities.10 Harvesters in New Zealand during WWII had to “walk beside a boat, waist to armpit deep in water and feel for the weed with their feet.” Handling large volumes of wet seaweed (which yields less than five percent agar) was challenging. Then as now, when Gelidium is harvested by scuba divers from rocky seabeds, collectors have to understand the life cycle of the algae, find the most likely locations for its growth, and prevent overharvesting to safeguard future yields.

Given its auspicious position on the Atlantic coastline, Morocco has been the main source of Gelidium for at least two decades, and demand for bacteriological agar continues to grow. Yearly global consumption increased from 250 tons to 700 tons between 1993 and 2018, and is currently estimated at around 1,200-1,800 tons per year, according to Pelayo Cobos, commercial director of Europe’s largest producer of agar, Roko Agar.

The Future of Agar

Amid such rising demand, it’s understandable that researchers worried when Morocco reduced exports of agarophytes in 2015. This shortage — due to a combination of overharvesting, climate warming, and an economic shift to internal manufacturing in the North African country — not only caused alarm but a three-fold price increase of wholesale bacteriological agar, which reached $35-45 per kilogram. (At the time of writing this in late 2025, factory agar prices are sitting at about $30 per kilogram, according to Cobos.)

A few years later, in 2024, researchers in multiple labs were horrified to notice toxic batches of agar for reasons as yet unclear. After they observed a worrying lack of microbial growth (impeding their ability to carry out basic experiments), they switched to different agar suppliers, and their results improved.

This was not the first time that microbiologists experienced problems with agar. A phenomenon called “The Great Plate Count Anomaly” baffled researchers in the early 20th century when they observed that the number of cells seen under a microscope didn’t match the actual number of colonies growing on an agar plate. Investigating this discrepancy, researchers found agar itself to be the culprit: when nutrient broths are heated with agar during boiling, harmful byproducts (hydroperoxide) can form due to the reaction of agar with phosphate minerals contained in the media. Researchers can avoid this by autoclaving agar separately from the nutrient broth, or by reducing the amount of agar used.

This anomaly is indicative of the larger challenge of culturing various microbial species, referred to as microbial “unculturability.” This cannot be explained by the use of agar alone or by the substitution of an alternative gelling agent, but rather by the difficulties in consistently recreating on an agar plate the multi-variable environment in which microbes grow naturally. Given such challenges, the risk of shortages, and the vulnerabilities of the agar supply chain, why is it so difficult to find suitable alternatives?

It is not for lack of trying. In some cases, microbiologists have ditched the Petri dish altogether, using microfluidics for manipulating and growing cells. However, these approaches aren’t likely to be adopted at scale as they require less common, less practical, and more expensive devices. So, what about other growth media?

By WWII, scientists had already begun looking at alternative gelling substances for routine use in bacteriology, but concluded that agar was still better as it is both firmer and easier to handle. Today, some specialized microbiology applications use the colloid carrageenan (extracted from red seaweed Chondrus crispus, or “Irish Moss”), a more transparent and less auto-fluorescent alternative to agar (agar emits its own background fluorescence when excited by light). However, for routine bacteriological use, carrageenan is more difficult to dissolve, requires higher concentrations, can degrade at high temperatures, and forms weaker gels, which may result in puncturing its surface during the plating of cells.

In some cases, alternative gelling agents might provide faster results. Researchers observed that bacterial cellulose and another bacterial polysaccharide, Eladium, allow a 50 percent increased growth rate for various bacteria and yeasts (as compared to their growth on agar), including higher biomass yields or faster detectable biofilm formation. However, both substances are still not as cheap and readily available as agar.

Guar gum, a plant colloid, costs less than agar and is better suited for growing thermophilic bacteria, but is also more difficult to handle, being more viscous and less transparent. The bacterial polysaccharide xanthan is cheaper as well but forms weaker jellies that, as with carrageenan, might result in puncturing its surface. Other colloids, like alginate (from brown seaweed) and gellan gum (from a bacterium), don’t set solely based on temperature and require additives for gelation. These additives might interfere with microbial growth and make the preparation of those jellies less handy than agar plates.

Thus, despite much effort, no gelling agent has yet been discovered that possesses all the properties and benefits of agar. Agar continues to be the best all-arounder: versatile, cheap, and established. And, if Gelidium agar should ever run out, and another colloid is not at hand, microbiologists could revert to culinary agar, which, although not as pure and transparent, offers a low-cost alternative to lab-grade agar.

It’s also worth noting that even if alternatives superior to agar were found, scientists are reluctant to abandon established protocols (even when microbiologists do use other jellies, they often still add agar to the mix, for example, to increase the gel strength of the solid media). As agar has been the standard gelling agent in microbiology for around 150 years, an enormous infrastructure of standardized methods, reference values, and quality control procedures has emerged around its specific properties. Switching to a different medium (even a superior one) means results may not be directly comparable to decades of published literature or to other laboratories’ findings.

So it is that agar continues to be the jelly of choice in laboratories around the world. As Humm wrote in 1947: “Today, the most important product obtained from seaweeds is agar, a widely-used commodity but one that is not well known to the general public.” Almost 80 years later, it might be better known, but its importance hasn’t dwindled.

Subscribe now

Corrado Nai has a Ph.D. in microbiology and is a science writer with bylines in New Scientist, Smithsonian Magazine, Small Things Considered, Asimov Press, and many more. He is currently writing a graphic novel about Fanny Angelina Hesse and the introduction of agar in the lab called The Dessert that Changed the World, which can be followed and supported on Patreon.

Thanks to Steven Forsythe for sharing a report on the use of agar seaweed in Britain during WWII, Barbara Buchberger at the Robert Koch Institute for pointing out Koch’s use of gelatine for the identification of cholera, and the surviving relative of Fanny Angelina Hesse for sharing a trove of unpublished material.

Cite: Nai, C. “The Origins of Agar.” Asimov Press (2026). DOI: 10.62211/12pq-97ht

Everyone agreed the problem was very small. That itself was how it got so big.

It was called Baseline Drift, which sounded like something that happened to boats when you forget to moor them. It did not sound like something that would reorganize biomedical science, but it did — specifically, the part that depended on knowing what counted as ordinary.

Baseline drift meant that nobody could reliably say what a “normal” human outcome looked like anymore.

This was awkward because, until the mid-twenty-first century, biology had relied on a shared set of normative assumptions: how much people should sleep, how fast wounds should heal, how cognition declined with age, how bodies responded to stress, infection, calories, boredom. These weren’t “universal” control groups, of course, because every experiment still had its own controls — but they were population reference frames, the invisible scaffolding beneath public health guidance, drug approvals, actuarial tables, and phrases like “clinically meaningful improvement.”

It began (depending on whom you asked) around 2025 with the publication of a global guideline on using GLP-1 medicines to help manage obesity. Or 2028, when prenatal methylation screens quietly became the standard of care. Or 2031, when adaptive lighting systems tuned to circadian biology spread from hospitals to schools to office parks. Or maybe it was 2034, when insurance stopped covering “general wellness” supplements because everyone was already getting them through food fortification, water treatment, and special perks from their employers.

In truth, there was no one moment where someone stood up and said, “We are now enhancing the species.”

There was only the slow tide of policies, memos, and social proof.

By 2040, most children in wealthy countries had received personalized micronutrient profiles before age three. By 2045, gut microbiome calibration was as routine as dental cleaning. By 2050, short-course somatic gene therapies corrected rare metabolic inefficiencies the way eyeglasses once corrected vision. While no single person was wholly optimized on all fronts, such interventions nudged distributions. Means crept upward. Variance narrowed in some places and exploded in others.

People slept eight-point-two hours instead of the six-point-nine that they had in the 2020s. They recovered from illness faster. They aged more slowly, but unevenly, and in unpredictable ways. Cognitive decline no longer followed a singular curve, as it had before the introduction of anti-amyloid therapies. Neither did pain tolerance, emotional regulation, immune response, or stress recovery.

Humans were healthier than ever. They were also less similar to one another than they had ever been.

Differences between rich and poor countries, between drug availability and production standards, between quickly diversifying treatment frameworks, or simply between the marketing prowess of certain doctors over others, made it so that human lifestyle, behavior, and ultimately, human bodies diverged ever more.

The Institute for Human Reference, whose name sounded more confident than it deserved, was founded in 2053 in a renovated strip mall in Florida. The mall had once housed a Chick-fil-A, a nail salon, and a store selling nothing but phone cases. Now it housed ethicists, statisticians, molecular biologists, policy analysts, and one historian hired to “provide context,” which is what someone does when you want them around for optics but don’t plan to listen to them.

The Institute’s mission statement was short and confident: Define the reference human.

Subscribe now

This used to be straightforward. In 2000, it meant “average diet, average activity, average lifestyle, no diagnosed disease, and no ongoing treatment.” In 2030, it meant “no experimental therapies, no known genetic edits.” By 2053, it meant nothing at all.

The trouble started with a spreadsheet, as it always does. A junior analyst noticed that participants labeled “unmodified” in a longitudinal cognition study were improving faster than those receiving standard nootropic protocols. Not significantly. Just enough to ruin everyone’s weekend.

The untreated cohort slept better, learned faster, and healed quicker than expected. They showed lower systemic inflammation, smoother glucose curves, and fewer regrettable opinions online. This was not because the treatments for the positive control group were harmful. It was because both groups were already heavily modified — often in aberrant, untracked ways. The “untreated” participants had benefited from cleaner air, adaptive lighting, fortified diets, stress-moderating urban design, and gut microbiomes optimized by public utilities providers for the bulk of their lives. Their baselines had been lifted, quietly and incrementally, without ever being logged as an intervention.

It was as if doing nothing had become a performance-enhancing drug.

The Institute convened a panel. The panel produced a report. And the report produced concern.

Without stable reference populations, public health science began to fray. Drug trials struggled to generalize. Public health guidelines contradicted one another. Regulators could no longer say whether an effect was meaningful or contextual. And insurance actuaries complained loudly.

The Institute tried to recruit even more unmodified humans. They advertised discreetly. “Participants sought,” the flyers said, “for a study on human baseline biology.” This attracted three types of people: conspiracy theorists, survivalists, and men who believed their masculinity was harmed by seat belts.

None qualified. Every candidate, regardless of how they avoided vaccines or grew their own food, was touched by various environmental enhancements. Their water was cleaner, the nutrients in their seeds higher, and even the air they breathed in the Institute’s lobby had been treated by germicidal ultraviolet light.

Meanwhile, the world continued improving.

Children learned faster, but differently. Each person became a unique bundle of interventions, exposures, and optimizations. Scientists began publishing papers with titles like Toward a Functional Approximation of the Human Baseline and Defining Control Conditions in a Post-Control World. These papers cited one another heavily and meant very little.

At last, someone proposed manufacturing a baseline.

Embryos. Archived ones. From before CRISPR, before microbiome engineering, before the Great Vitamin Correction of 2029. They would raise these children in sealed environments, feeding them pre-industrial diets. They would experience no screens, no optimization, and no enhancements except love, which everyone agreed did not count, despite all evidence to the contrary.

The ethicists balked.

The regulators bristled.

The grant proposal was rejected, revised, rejected again, and then leaked.

GOVERNMENT CONSIDERS PRODUCING “NATURAL” CHILDREN FOR SCIENCE.

Public reaction was swift and confused. Some people thought such children would be holy. Others thought they would be monsters or savages. Most people thought the whole thing sounded complicated. A Texas senator said the proposal sounded “expensive and European.”

The project was shuttered.

The Institute pivoted to simulation. If scientists couldn’t find a baseline, perhaps they could model one. Perhaps they could reconstruct the unmodified human the way paleontologists reconstructed dinosaurs through fragments and educated guesses. They rented an old data center and built several models of unmodified humans from historical data, archaeological inference, and optimistic assumptions.

The model was elegant and insightful, but also depressing.

The simulated baseline humans were less healthy, more impulsive, more violent, and much more likely to believe things that were untrue. When the model was presented at a conference, someone asked whether this meant enhancement was a superior way of life.

“No,” said a statistician, “it means our assumptions about our ancestors were flattering.”

The room laughed politely.

The Institute issued its final report in 2061. It stated, in careful language, that the concept of an unmodified, “natural” human baseline was no longer scientifically meaningful. Humanity had become irreducibly specific, fragmented into trajectories too individualized to be averaged without distortion. Science could no longer pretend that population means governed individual outcomes the way it did before. The failure was not ethical. It was statistical. Averages had stopped converging. The problem was no longer contamination of the baseline, but the disappearance of the conditions under which a baseline would be sufficiently explanatory.

It didn’t take long for biomedicine to adapt. N=1 studies proliferated. Rolling baselines replaced fixed controls. Therapies were evaluated against a person’s own prior state, not some mythical norm. Progress became local, contextual, and personal. The 2060s saw the proliferation of some seven billion digital baselines, with each person becoming the proud representative of their own unique reference group.

The Institute closed in 2062, its offices converted into a yoga studio. The historian wrote a book that sold poorly but aged well.

Science continued, because science always does. It simply changed its metrics.

People went on living longer, thinking more clearly, suffering less, and wondering why they felt so uneasy. Aging people reminisced about the people of yesterday. Ad agencies bemoaned the death of a world where they could appeal to a large target audience. Children asked their parents what humans used to be like.

“Worse,” the parents said.

“Why didn’t you change things quicker?” the children asked.

“Humanity was afraid of changes,” they said. “But we kept accepting small improvements and got better over time.”

The children — enhanced in a thousand invisible ways — went to sleep under lights tuned to their biology, breathing optimized air, dreaming dreams shaped by a history of improvements they did not choose.

They slept well. That, at least, could be measured.

Eliomer H. Kaas is an Eastern European social scientist with a taste for the simultaneously hopeful and depressing stories that make up the world of STEM in the 21st Century. Through his fiction, Elio tries to come to terms with the ever-changing nature of social advancement spurred by technology and the chaos caused by this advancement to human society, culture, and life.

Cite: Kaas E.H. “Baseline Drift.” Asimov Press (2026). DOI: 10.62211/29wj-12qw

By Taylor Rayne

Smell is our most primal and, arguably, most emotionally potent sense. It summons memories, shapes taste, and influences behavior: the aroma of coffee is capable of enhancing alertness well before caffeine ever reaches the bloodstream. A hint of sunscreen collapses decades, taking us back to youth; but, pinch the nose, and suddenly a slice of apple is hard to distinguish from a piece of raw potato.

Despite its significance, scent remains the most mysterious of our senses. Unlike vision or hearing, it resists straightforward formalization. The challenge lies not only in the vast molecular diversity of odorants, which vary in far more ways than photons or frequencies, but also in the effort to build a shared vocabulary and technology capable of codifying subjective sensation. So while machines have learned to see through computer vision and hear through signal processing, scent remains stubbornly analog. There has been no RGB of odor, no Fourier transform for smell.

At least, until now.

Tech giants, including Google, startups such as Osmo, and even traditional fragrance houses like Givaudan have begun turning to AI to probe the possibility of digitizing smell. By encoding scent molecules into 1s and 0s, their hope is to better understand and manipulate this sensory modality. Just as “computer vision” has helped us realize that sight is not just passive image capture but an active process of prediction and interpretation, researchers hope that programming smell will illuminate the many mysteries of olfaction.

Beyond providing further insight into olfactory biology, digital scent could have many practical (and quite profitable) applications, which is why its proponents, from defense agencies such as DARPA to corporate conglomerates like Estée Lauder, have invested in it. Computational smell could, for example, help detect threats and information invisible to cameras, such as gas leaks, food spoilage, disease markers in breath, and even counterfeit products. It could also reduce reliance on resource-intensive natural ingredients used in perfume and other odorants by, for instance, finding chemically synthesized molecules capable of evoking the same brain patterns. And finally, it could lead to the creation of entirely novel smells, revealing a vast, untapped chemical palette that would otherwise be unattainable without the aid of technology.

The Bacterial Beginnings of Smell

Long before life evolved eyes and ears, the world was encountered chemically. This took place as molecules permeated and diffused across cell membranes, performing a metabolic exchange between animate and inanimate matter.

Smell, our most ancient interface with the environment, originated over 3 billion years ago, in bacteria adrift in the primordial ocean. These early organisms navigated chemical gradients in the water, detecting molecules to swim toward food and away from danger. This ability, known as chemosensation, is the most rudimentary form of smell.

Such “sensing” relies on receptor proteins embedded in the cell membrane, acting like molecular locks awaiting the corresponding chemical key. When a passing odor molecule fits into a receptor’s binding site, it changes the receptor’s shape, setting off a cascade of signals inside the cell that direct the organism’s movement.

Over time, these molecules didn’t just guide survival; they encouraged multicellular life. As cells began clustering together, the exchange of semiochemicals — molecular signals that transmit information within and between species — began to influence behavior, enabling aggregation and synchronization, and laying the groundwork for cellular cooperation. Plants, for instance, release green leaf volatiles such as cis-3-hexenal (the familiar scent of freshly cut grass) when attacked, both warning neighboring plants and attracting the animals that prey on herbivores. Human infants, meanwhile, are drawn to the distinctive odor of their mother’s milk, which both stimulates feeding and regulates their earliest physiological rhythms. And among insects, ants are famous for deploying pheromones such as iridodials, which direct entire colonies to forage or fight in concert.

Once the first tetrapods emerged from the sea and embraced life on land, smelling evolved to become ever more refined under newfound terrestrial pressures, including adapting to novel volatile odorants and the more variable conditions of airborne climates. As these early terrestrial vertebrates expanded a chemosensory repertoire that had once been far more limited in their aquatic ancestors, olfactory systems likewise continued to evolve. Over the next hundreds of millions of years, the neural structures supporting our own sense of smell increased in sophistication.

Today, roughly two to five percent of our genetic blueprint concerns itself with smell. While it may seem a small fraction of the whole, it is the largest gene family in the human genome. At any given time, 77 percent of the 356 distinct olfactory receptors are expressed in the lining of the nasal cavity, each tuned to the molecules that make up the world’s myriad smells. The receptors expressed vary between individuals, with only 90 receptors commonly found in all people. Together, this suite of receptors is responsible for every experience of scent we encounter, and the variability between individuals is likely behind smells’ subjectivity.

To understand how this olfactory complexity works, consider the single inhalation that follows biting into a ripe strawberry: the rush of aroma, the sweet, tangy burst blooming in the nose, created not by a single compound but by a volatile molecular cocktail. There is no one single molecule responsible for the smell of strawberry, but rather a family — namely furaneol (caramel-sweet), methyl and ethyl butanoate (fruity), methyl 2-propanoate (fruity), hexanal (green and sharp), and cis-3-hexenol (leafy-fresh) — that evaporate and surge through the nasal cavity.

In each nostril, this medley of molecules dissolves into the mucus layer lining a 2.5 cm² patch of tissue known as the olfactory epithelium. Studded across this small region is a mosaic of roughly 10 million olfactory sensory neurons,1 each expressing only one type of olfactory receptor protein. In one breath, an odorant molecule, be it furaneol, cis-3-hexenol, or any other, activates a unique combination of receptors, similar to striking a subset of keys on a piano. This ensemble of activated neurons forms a distinct pattern that the brain reads as “strawberry.”

Crucially, no two scents ever strike the same pattern. The combinatorial activity of hundreds of receptor genes allows humans to detect and discern more than a trillion distinct odors.

While some olfactory receptors respond broadly, meaning they can recognize and bind to several different structurally-related molecules, other receptors are exquisitely selective and bind to only one specific shape or stereoisomer.2 For instance, the two mirror-twin forms of the organic compound carvone smell strikingly different — one like spearmint (R-carvone), the other like caraway (S-carvone) — underscoring the nuance the nose brings to discriminating between such molecular mirror images.

The resultant signal, whether “caraway,” “spearmint,” or “strawberry,” travels to the olfactory bulb, a bipartite nerve structure nestled at the base of the skull, just above the nasal passages. There, neurons expressing the same receptor type converge on specialized structures called glomeruli, the crucial processing hubs that sharpen and refine sensory input en route to deeper regions of the brain involved in odor discrimination.

Each glomerulus acts as a dedicated module for a single receptor type, receiving input from thousands of olfactory sensory neurons scattered throughout the nasal epithelium, but all tuned to the same molecular features. Within these spherical tangles of neural connections, the incoming signals undergo their first round of processing: they’re amplified, filtered, and integrated by local interneurons that enhance contrast between different odor signals.

Unlike other senses, these olfactory signals take an unusual route through the brain. While vision, hearing, and touch all pass through a brain region called the thalamus before reaching higher brain regions, smell bypasses this checkpoint. Instead, odor information travels directly from the olfactory bulb to the amygdala and hippocampus (two brain structures central to emotion and memory) before reaching conscious processing areas. This anatomical shortcut may explain why smells can trigger vivid memories and strong emotions before we’ve even consciously identified what we’re smelling.

Yet, while the nose has been anatomically mapped, receptors sequenced, and neural pathways charted, predicting a molecule’s scent has remained a mysterious exercise. The question of why certain configurations of matter smell one way and not another persists. In other words, why does a molecule such as furaneol activate the receptor signaling “jammy sweetness”? What makes one molecule smell “grassy” and another “creamy,” one trigger nausea and another nostalgia?

The prevailing hypothesis defines odorant molecules as ligands, specialized binding molecules that fit into olfactory receptors like a lock and key. This interaction activates the receptor, initiating an electrical response in the brain, a unique pattern of activity associated with a particular scent. While molecular structure has long served as a proxy for predicting smell, particularly for researchers and fragrance chemists, the so-called Structure-Odor Relationship (SOR) paradox endures: molecules of nearly identical structure can smell worlds apart, while those with little in common can smell strikingly similar.

Take musks,3 for instance: this family of compounds includes hundreds of molecules with vastly different structures and molecular weights, ranging from small macrocycles to large polycyclic frameworks. Yet, despite this diversity, many share a common warm, powdery, and animalic scent profile that defies straightforward structure-to-smell prediction.

An Industry for Smell

The speculation that chemical structure corresponds to scent dates back millennia, alongside the use of fragrant materials in ceremonial and cultural practice. From the early Greek atomists’ theories of microscopic films emitted by objects to excite the senses, to the incense artisans of ancient Egypt and East Asia, humans have long sought to capture and understand the elusive nature of smell.

Before the emergence of more sophisticated means for extracting and preserving scent, fragrance was elicited through rudimentary methods such as crushing raw botanicals, steeping them in oils, or burning resins to release their aroma.

During the eighth and ninth centuries, the invention of the “modern” alembic — a distillation apparatus developed by Islamic alchemists featuring a still pot, a cooling condenser, and a collection spout — allowed for a more controlled extraction of delicate essential oils than crude crushing or simple heating could achieve. This innovation enabled new forms of fragrance production, introducing alcohol as a base and associating the practice of perfumery more closely with chemistry and medicine. Through trade and translation, this technology and technique migrated to continental Europe, where perfumers expanded their repertoire over centuries with methods such as maceration, expression, tincturing, and enfleurage4 to isolate the volatile compounds of flowers and woods.

In the sixteenth century, tanners making leather gloves sought to mask the acrid smell of their wares with fragrant oils and extracts. Scented gloves quickly became fashionable among the European aristocracy, and perfumers found a lucrative market beyond functional necessity. At Versailles, the epicenter of French culture and seat of royal power, courtly fashion demanded a complete sensory presentation: the right clothing, hairstyles, cosmetics, and an ever-changing repertoire of perfumed products to signal refinement and status.

This high consumption of scent supported a nascent perfume industry, formalized in 1656 in Grasse, France, with the creation of the Maîtres Parfumeurs et Gantier (the perfume and glove-maker’s guild). Alongside its favorable social-political milieu, Grasse’s microclimate fostered a burgeoning perfume industry in the region, where fertile soils nurtured fragrant flowers such as jasmine, rose, lavender, orange blossom, myrtle, mimosa, and tuberose. This economic activity, along with the formation of local perfume, helped establish France’s perfumeries as centers of both craft innovation and the formal study of smell.

The nineteenth century marked a decisive shift as techniques from pharmacy and organic chemistry began to elucidate molecules underlying fragrance. Building on earlier breakthroughs such as the isolation of morphine in 1804 (the first alkaloid ever extracted from a plant), chemists turned their attention to aromatic materials.

In 1820, the French pharmacist Nicolas Jean Baptiste Gaston Guibourt identified and isolated 2H-chromen-2-one, better known as coumarin, from the tonka bean (Dipteryx odorata). In an account presented to the pharmacy section of the Académie Royale de Médecine, Guibourt formally christened the new substance coumarine, marking one of the earliest instances of the chemical characterization of a fragrance compound. In 1858, another French chemist, Théodore Nicolas Gobley, succeeded in crystallizing vanillin from vanilla pods, confirming that odorants could be identified as discrete molecular entities.

For the first time, individual aromatic molecules — a resonant note of the rose, one iota of the vanilla pod — could be extracted from a larger, complex composition. Smell was being recast as a phenomenon amenable to classification and design.

Driven by advancements in organic chemistry and the introduction of the periodic table in 1869, fragrance chemistry blossomed. By 1882, Paul Parquet’s famous scent, Fougère Royale, featured coumarin, followed by Guerlain’s Jicky in 1889, and in 1921, Chanel No. 5 launched using an entirely new class of synthetic compounds, the aldehydes. The perfume industry eagerly adopted new techniques, including chromatography and fractional distillation.

As the twentieth century unfolded, scientists intensified their quest for structural correlates of scent. Discoveries like ionone-related ketones evoking violet and woody notes, long-chain macrocycles recurring in musks, and aromatic rings anchoring vanillic and balsamic tones found their way to the traditional orgue à parfums, a semicircular, horseshoe-shaped work station: a harbinger of what would become a more serious investigation of the science and technology of smell.

Digitizing Smell

Over eons, organisms evolved a sense of smell. In the past three centuries, smell became chemistry. And in recent years, we have started to ask whether smell can become code.

However, unlike sound and color, smell resists straightforward formalization. Color, for instance, has been reduced to three primary channels, standardized through the color wheel, and rendered into numeric systems that machines can interpret. Sound has been decomposed along perceptual axes of pitch, timbre, and amplitude, each capable of digital transformation. Odor, however, has no reliable relationship to molecular structure or perceptual encoding, making it difficult to establish a computational coordinate system.

As a result, early computational efforts in representing molecules were often ad hoc: homemade tables to track atomic bonds, bespoke matrices, or custom linear codes that translated chemical structures into character strings. These methods were clever but small-scale and subjective, making calculations laborious and difficult to reproduce.

Then, in 1988, while at Pomona College in Claremont, California, chemist David Weininger devised SMILES (Simplified Molecular Input Line Entry System). SMILES provided a standardized, machine-readable line notation for encoding molecular structures. Like any language, SMILES is a formal system for representing information; in the context of chemistry, it acts like a cipher or code to translate molecular structures into linear character strings. A given molecule, according to convention, is spelled out as a word or a particular grammatical grouping of letters, where each letter corresponds to an individual atom. Contained in each word is also the logic for how certain letters connect or are arranged geometrically, including instructions for branching and ring closures. For instance, the six-carbon ring cyclohexane is represented by slicing open the ring and labeling the adjacent atoms with matching numbers to indicate connection and closure: C1CCCCC1. In the same fashion, coumarin can be written as O=C1OC2=CC=CC=C2C=C1.

In practice, however, this descriptor-based model could not contend with the idiosyncratic nature of the structure-to-odor relationship: a molecule may be classified as citrus-like by virtue of its ester group, while a structurally analogous compound with only minor modifications might be seen as sharp, metallic, or entirely odorless.

As academic labs continued researching the logic of scent, focus shifted toward building larger datasets that linked molecular structure to qualitative descriptions of odor perception. In place of binary quantifiers such as odorous/odorless or pleasant/unpleasant, these collections captured distinct attributes like fruity, floral, musky, or burnt, offering a more embodied and qualitative relationship between chemistry and experience.

Democratizing databases such as the Good Scents Company, for example, expanded the vocabulary of olfaction available for computation, priming the field for machine learning approaches capable of finding patterns across large, multidimensional datasets.

In 2015, IBM Research and Rockefeller University launched the DREAM Olfaction Prediction Challenge, the first large-scale, open benchmark for predicting human olfactory perception from the physicochemical features of odor molecules. Researchers were provided with data comprising 476 chemically diverse odorants, each described by over 4,800 molecular features and rated by 49 human participants on 19 sensory attributes, such as sweet, fish, mint, and sour.

​​Competing teams trained machine-learning models to map feature-perception relationships and evaluated their models using 69 unseen test molecules. The top models achieved prediction measures comparable to those observed among human participants, meaning the machine ratings were nearly as reliable as human noses when judging whether a chemical smelled floral or fishy. The models’ predictions mirrored the average agreement reported in human odor-perception ratings, demonstrating accuracy of up to 85 percent across multiple different aromas.

Although the DREAM challenge yielded compelling evidence that odor perception could be forecasted computationally, it had its shortcomings. Namely, the study’s dataset remained modest in scale, leaving vast chemical territories uncharted, and its reliance on predefined molecular descriptors meant the models depended on prescribed features rather than uncovering the latent logic linking structure to odor.

By 2019, researchers at Google Brain (now DeepMind) sought to overcome these constraints by expanding both the scale of training data and the fidelity of their model. Using deep learning and Graph Neural Networks (GNNs), a class of models that interpret molecules as graphs with atoms and bonds represented as nodes and edges, respectively, the study employed a dataset containing more than 5,000 odorants annotated by expert perfumers across over 100 descriptors, from mint and potato to sulfurous and grassy.

Unlike the earlier DREAM Challenge, which relied on prescribed molecular features, such as atom counts and topological indices, this approach allowed the model to infer structure-odor relationships directly from data. To initiate the training process, each node (representing an atom) was seeded with basic chemical information, including atomic number, valence, hybridization, and formal charge.

Through a recursive process known as “message passing,” the state of each atom was repeatedly updated in response to information relayed by neighboring nodes. Layer by layer, the model built up a progressively bigger picture of a molecule: early layers learned to identify local features like carbonyls, halides, or ring systems, while deeper layers learned to recognize broader chemical motifs such as aromaticity, conjugation (the overlap of p-orbitals), and steric strain (the increase in potential energy of a molecule due to repulsion between electrons in atoms that are not directly bonded to each other).

By the penultimate layer, the complete, networked “picture” of a molecule was condensed into a fixed-length, 63-dimensional vector, known as an “odor embedding,” capturing the molecule’s perceptual qualities in a computable form. This embedding was then used for classification tasks, such as predicting specific odor descriptors or categorizing the molecule’s scent profile.

When these “images” were compressed, molecules that smelled alike clustered together even when their chemical scaffolds were structurally dissimilar. For example, certain esters and ketones, molecules that share little in common structurally, appeared close in this space because both carried a sweet, fruity scent, while structurally related compounds could be separated if their perceived odors diverged. This act of compression thus exposed a “geometry of smell.”

It also laid the groundwork for Google’s subsequent introduction of the Principal Odor Map (POM): the first map of smell, in which traversing through space corresponds to crossing distinct regions of aroma, from jasmine to potato. Building on the previous study, the researchers expanded the model’s embedding layer by 193 dimensions, creating a high-dimensional coordinate system that positions each odorous molecule as a single point within a continuous manifold (e.g., a connected shape or space that behaves geometrically, enabling meaningful comparison and relation between the position of points representing molecules). In this space, perceptually similar odors occupy nearby positions, much as blue lies closer to turquoise than to crimson on the color wheel, while increasingly dissimilar ones are separated by greater distance.

When the researchers subjected the POM to further tests, they found that it could generalize beyond human olfactory perception, predicting olfactory receptor activity across species from mice to insects. This cross-species application suggests that the model’s internal structure even captures evolutionary principles of scent. On the map itself, metabolically related compounds — those separated by only a few metabolic reactions or naturally co-occurring in the same substance — also occupy nearby coordinates, forming concentrated clusters that reflect their shared chemical and perceptual relationships.

By capturing such subtle perceptual relationships, the model not only surpassed traditional chemoinformatic algorithms in terms of predictive accuracy, but also enabled early steps toward the design of new odorants.

Building on this foundation, Osmo, a start-up spun out of Google Research in 2022, has begun leveraging its proprietary Olfactory Intelligence (OI) platform informed by POM to translate sensory scent data into chemical design. The company’s goal is to “give computers a sense of smell.”

Glossine, Fractaline, and Quasarine are the first prompt-generated fragrance molecules created using its platform. According to the company, Glossine delivers a “Las Vegas-style sparkle,” with a floral jasmine-like top note that adheres well to textiles; while Fractaline offers a versatile violet note, reputed to give a powdery, “second-skin” impression; and Quasarine presents an intense yet delicate jasmine aroma claimed to deliver a long-lasting “fresh petal-y effect.” While these descriptors could be shared by other perfumes, these fragrances are entirely original. Where traditional perfumery relies on the discovery and refinement of proprietary molecules in a process that usually takes years, Osmo can “create” completely new odorants through this OI platform in just a few days.

The potential applications of Osmo’s OI platform, however, extend far beyond fragrance. Because the Principal Odor Map can extrapolate olfaction across species, Osmo has leveraged it for developing synthetic alternatives to DEET, the active ingredient in most commercial insect repellents, but one which has been linked to adverse effects, including severe skin reactions like rashes and welts. Trained on experimental datasets measuring mosquito repellency, the model can predict the repellent efficacy of nearly any molecule. Experimental validation confirmed over a dozen new molecules as repellent as DEET, but which could offer safer, longer-lasting alternatives for killing mosquitoes and ticks.

Beyond repellents, Osmo has broader ambitions to also develop non-invasive diagnostic tools. They aim, for example, to identify volatile signatures emitted by various diseases, enabling algorithms with a “digital nose” to detect conditions such as Parkinson’s, diabetes, and certain cancers through subtle changes in body odor or sebum, a type of signal detectable by dogs.

Perhaps the most promising application of Osmo’s technology, though, relates to sustainability. By synthetically generating fragrance molecules, Osmo offers a path toward decreasing the environmental impact of industrial scent-production. A canonical and increasingly endangered raw material for scents is rose, whose odor has captivated humans for millennia yet remains one of the most labor-intensive and expensive natural ingredients (some 60,000 roses, roughly 180 kilograms of petals, are required to produce a single ounce of oil, which sells for $8,000 to $12,000 per kilogram). By plotting the scent of Rosa damascena and Rosa centifolia on the POM, Osmo can work toward “reverse-engineering” the rose and other floral notes in pursuit of designing “an alternate universe of perfume ingredients” that are perceived similarly without requiring raw botanical sources.

Osmo is not alone in its effort to advance computationally mediated scent. Givaudan’s Carto, the digital design platform of one of the world’s first and largest fragrance houses, is also using computer technology to design its perfumes. Carto integrates predictive algorithms with a database of over 5,000 raw materials (five times more than a human perfumer typically manages) to help perfumers simulate, match, and modify scents. Whereas the classical perfume orgue à parfums once arranged rows of essences within arm’s reach, Carto replaces physical glass vials with a virtual glass touchscreen display. The platform was used to formulate Tom Ford’s Bois Pacifique, which launched in January 2025.

Despite such recent successes, the mysteries of olfaction are far from solved. To digitize the scent of a rose, for instance, requires contending with over 300 volatile compounds that contribute to its spicy, honeyed, and tea-like notes. These compounds are not simply additive but dynamic; their interactions, relative concentrations, and release create emergent qualities unpredictable from any single component alone. While platforms like Osmo and Carto can tinker with individual molecules, capturing the interplay of dozens or hundreds of odorants in a given mixture remains a formidable challenge, the mastery of which will be next in digitizing smell.

Both industry and academia continue to advance this aim. The most recent DREAM Olfaction Challenge, slated to close in late 2025, tasked international teams with predicting perceptual similarity across volatile mixtures. It will draw on a dataset of over 700 combinations containing anywhere from two to ten different molecules to assess how accurately teams can digitally model smells, their interactions, and their resulting impressions. Although peer-reviewed results have not yet been published, early findings point toward improvements in complex-mixture scent prediction.

The Future of Olfaction

Traditionally, the fragrance industry has made a point of drawing a line between the natural and the synthetic, with the introduction of synthetic scent molecules in the mid-nineteenth century garnering both criticism and celebration. Early critics5 proclaimed such synthetic materials as “vulgar debasements” of perfumery, reflecting anxieties surrounding industrialization and the loss of heritage and craftsmanship. But others celebrated this innovation. Guerlain’s Jicky (1889), for instance, which incorporated synthetic vanillin and coumarin, was hailed as a new chapter for perfume. Moreover, commercial houses like Lubin marketed synthetic musks as “triumphs of science … over Nature,” promoting the “non-evanescence” (long-lasting) qualities of synthetic chemicals in comparison to natural ingredients.

These developments unfolded within a wider cultural moment that the essayist Max Beerbohm described as “a new epoch of artifice.” The emergence of synthetics did not simply replace the natural; it upset the very idea of what “naturalness” meant — did it refer to an ingredient’s source, its sensory impression, or the meanings attached to it?

Even today, materials such as ambergris (a waxy, sweet-smelling substance formed in the digestive system of sperm whales and expelled into the ocean) remain highly coveted precisely because of their scarcity and natural origins. Revered as “the treasure of the sea,” ambergris is found in less than five percent of sperm whales and develops its celebrated complexity over years of exposure to sea, salt, and sun. Such a valued ingredient commands reverence precisely because of its rarity, exposing how ideas of purity and authenticity continue to shape our perception and appreciation. Ultimately, the persisting popularity of ingredients such as ambergris demonstrates that value pertains as much to culture as to chemistry, reflecting not only the reality of raw materials but also the moral and aesthetic weight we ascribe to their provenance.

Given this legacy, how might we relate to machine-generated molecules?

The shift to digitized scent will require that we rethink not only the substances themselves, but also the stories we tell as we find ways to valorize their synthetic origins. Such computationally-designed scents are, after all, safer, allergen-friendly formulations compliant with evolving regulations. Furthermore, AI-designed aromas often use more sustainably produced molecules, with a lab-created civetone replacing wild civet extract (a substance associated with the questionable practice of keeping civet cats in captivity for their perineal gland secretions).

A similar parallel has played out elsewhere: a one-carat lab-grown diamond, while chemically and optically identical to a mined one, can cost less than a quarter of the price, despite being purer and more ethically produced. While they have been derided as “costume jewelry,” the market for lab-grown diamonds was valued at more than $18 billion in 2023 and is expected to continue growing. The connotations of luxury are increasingly leaning toward narratives of sustainable and ethical sourcing.

Machine-imagined scents may follow a similar arc. But just as the telescope widened our aperture to distant cosmic structures, digitally composed scents will expand our olfactory range, this time towards the smallest scales of chemical configuration.

Subscribe now

Taylor Rayne holds a degree in biochemistry and computer science, as well as a deep curiosity for the interplay between scientific and cultural production. She currently works at the Novo Nordisk Foundation Center for Biosustainability, based at the Technical University of Denmark.

A special thanks to Christiana Agapakis for her generosity and resources; Jasmina Aganovic for her thoughtful guidance and time; Xander Balwit for her steady editorial support, and of course, Astrid.

Cite: Rayne, T. “Scent, In Silico.” Asimov Press (2026). DOI: 10.62211/28jw-12ty

This essay will appear in our forthcoming book, “Making the Modern Laboratory,” to be published later this year.

In 1959, a pair of enterprising brothers, Jack and Harold Kraft, filed a patent titled “Apparatus for mixing fluent material.” Though simple in concept, their invention solved one of the most fundamental challenges faced by mid-century scientists: mixing fluids quickly and efficiently. The vortex mixer, a small motorized device that vibrated samples, offered the perfect solution, and is now found on biology benches across the world.

Harold and Jack Kraft were born in New York in the tumultuous years following World War I. From a young age, the boys displayed an entrepreneurial spirit and were always in business together. During the Great Depression, they made money by repairing broken radios and installing radio antennas on buildings in New York City. Family members recall Harold as the gregarious talker or salesman, while Jack led the technical side of their ventures.

Even World War II could not stop their abiding passion for motors and machines. Jack attended NYU School of Engineering while in the reserves, and Harold became an aircraft mechanic in the 519th Service Squadron, working at airfields in England and later France. After the war, the brothers reunited to take up business once again, this time manufacturing their own eponymous brand of Kraftone record players.

By the late 1950s, looking to break into scientific equipment, Jack reached out to fellow NYU alumnus and inventor, Dr. Samuel R. Natelson, a clinical chemist who was the Head of Biochemistry at St. Vincent’s Hospital of New York. Natelson kindly obliged Jack’s request, with the two meeting several times for chemistry demonstrations and discussions of the equipment challenges faced by Natelson and his colleagues.

During one such meeting, Natelson expressed a dire need for better mixing equipment. At the time, chemists had only a few options. If the solution volume was large enough, magnetic stir bars could be placed into the mixing vessel, but that meant they needed a corresponding electromagnetic stir plate upon which to place the solution. Most labs had only a few such plates, if any, so when making multiple solutions, there weren’t enough to go around. The alternative was to stir, shake, or flick the vessel manually.1 These apparatuses all needed cleaning between each use.

We may never know which specific mixture drew the ire of Natelson. But we can infer from a letter, drafted by the Kraft brothers, that it concerned viscous substances. The letter also reveals just how excited the brothers were about their invention. According to their account, Jack brought the original idea to Harold, and the two sketched out potential solutions on October 20, 1958. Just three days later, they had built their first prototype using the same kind of shaded pole AC motor found in their record players.2

The resulting invention was simple, but elegant. A small, high-powered motor was housed within the body of a box-shaped machine. Mounted atop the motor was a rubber cup. Switched on, the motor oscillated the cup in tight orbital motions. When a test tube, or other vessel, touched the rubber cup, that motion transferred to the liquid, creating a vortex and mixing its contents.

They first demonstrated their invention before a group of scientists at Sunnyside Medical Laboratory in Long Island on October 25th, 1958, just days after building the prototype.3 The vortex was used to wash a protein in water, a step in the protein-bound iodine (PBI) test, a now-antiquated clinical method for measuring thyroid function. The PBI test was regarded as “one of the more complicated and commonly used clinical laboratory tests, requiring skilled technicians and generally requiring a separate laboratory because of contamination problems,” according to a 1967 study. Its several steps traditionally required thorough cleaning of the stirring implements, but each glass stir rod that was introduced increased both the chance of contamination and the potential loss of precious sample material stuck to the glass. A mixing method that didn’t require constant touching of the sample would be valuable indeed.

The patent, which Harold and Jack filed on April 6th, 1959, was granted on October 30, 1962. During this time, the brothers were not idle. They began to work with Scientific Industries Inc., a scientific equipment company established in Springfield, Massachusetts, in 1954. Harold became its president and Jack its treasurer, and the two moved the company to Queens Village, New York. Scientific Industries Inc. began manufacturing and selling the Kraft brothers’ vortexer in 1962. The very first model was named the Vortex Jr. Mixer. The company also experimented with other form factors and head attachments, like the K-500-4 model, able to mix four tubes simultaneously.

The Kraft brothers also continued tinkering, filing a number of patents by themselves and with Scientific Industries for devices such as the “Rotary apparatus for agitating fluids,” which revolves tubes in a Ferris wheel-like motion to continuously stir them. This tube rotator, though lesser known than the mighty vortexer, is also a common laboratory tool today.

In April of 1965, Harold and Jack left Scientific Industries to found their own scientific instrument manufacturing business, Kraft Apparatus Inc., in Mineola, New York. They continued to refine the design of the vortexer, adding a pressure-sensitive “touch” feature that turned the device on when the rubber cup was pressed down, as well as speed and pulse settings. Finally, in 1982, the Kraft Brothers sold Kraft Apparatus to Glas-Col, a division of Templeton Coal, which still sells vortexers today. Meanwhile, Scientific Industries continued to manufacture its own versions, creating the iconic “Vortex Genie” line.4

Today, scientists commonly use vortexers to mix volumes ranging from microliters to milliliters, most often in plastic Eppendorf or conical tubes.5 As scientific instruments go, vortex mixers are on the sturdy side, weighing in at 4 kilograms, or 8.8 pounds. Their heft and rubber feet provide stability, preventing them from vibrating off the bench. And although newer versions are slightly quieter, older ones emit a distinctive rumble that can be heard echoing throughout the lab.

Vortex mixers also come in a variety of shapes and sizes.6 There are versions that secure 96-well plates or Eppendorf tubes for hands-free mixing. Some have adjustable speed, orbital direction, and digital timers to make mixing more precise.

We owe much to the industrious Kraft brothers, for although the vortex mixer is relatively humble and was created in a span of just three days, its importance is hard to overstate: making mixing, one of the most essential but tedious laboratory chores, easy.

Subscribe now

Ella Watkins-Dulaney is a bioengineer who owes a not-insignificant portion of her PhD to the vortexer. She is also the Art Director for Asimov Press.

Acknowledgements: Thank you to Howard J., Randy E., Robert, and Ruth Kraft for their interviews about Kraft family history. A special thank you to Scott Kraft for all of his work collecting historical documents and helping me piece together this story. It would not have been possible without you. And thank you to Xander Balwit as well for the consistent editing and support.

Cite: Watkins-Dulaney, E. “Making the Vortex Mixer.” Asimov Press (2026). DOI: 10.62211/49jq-97pk