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

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 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

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

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

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

By Matt Lubin

Until the 20th century, there was no easy way to detect a pregnancy. One could wait for two missed menstrual cycles, watch for the first signs of a baby bump, or listen closely for a fetal heartbeat, but none of these methods work until several months into gestation. Ancient and medieval sources describe a variety of possible tests to determine whether a person was pregnant, but none were remotely reliable by modern standards.1

In the 19th and 20th centuries, researchers began looking for a solution in the emerging science of endocrinology, the systematic study of hormones. It was during this era that doctors and biologists began to shift from studying anatomy (observing the mechanics of visible organs) to exploring the invisible potency of “internal secretions” and “juices” that directed bodily phenomena. Early experiments by endocrinologists were crude, however, usually involving the injection of fluids from one animal (or human) into another and seeing what happened.

One pioneering endocrinologist who predicted the existence of hormone chemicals was Charles-Édouard Brown-Séquard. In 1889, he shared results from a “study” in which he injected himself with an elixir of “the three following parts: first, blood of the testicular veins; secondly, semen; and thirdly, juice extracted from a testicle, crushed immediately after it has been taken from dog or a guinea-pig.” Brown-Séquard found that these testes-semen injections restored to him all the vigor of his youth, though few others have been able to reproduce his results.2

In this spirit, two doctors working in Berlin, Selmar Aschheim and Bernhard Zondek,3 discovered an early pregnancy test in 1928 by injecting patients’ urine into laboratory mice. A few days after the injections, the mouse would be killed and dissected so that its ovaries could be examined; if the patient who provided the urine sample was pregnant, the mouse would develop “ovarian blood spots.”

In the U.S., Maurice Friedman adapted this approach for rabbits because a single rabbit could handle a larger volume of urine, their ovulation patterns are a bit more reliable, and clinical centers generally found them easier to house. His “rabbit test,” released in 1931, became the American standard for several years. Yet both the mouse and rabbit tests had significant drawbacks: they required waiting several days for results, and the animals had to be euthanized or at least operated upon to check for signs of ovulation.

It was around this same time that a British scientist living in South Africa, Lancelot Thomas Hogben, injected some ox pituitary extract into an African clawed frog, Xenopus laevis, as part of his efforts to understand how hormones influence ovulation. It did not take long for Hogben to realize he had struck gold (or rather, eggs) because, just hours after the injection, the frog laid many hundreds of them. Thrilled by this discovery, two former students of Hogben’s, Hillel Shapiro and Harry Zwarenstein, demonstrated that these African frogs would be much more suitable for pregnancy tests than mice or rabbits.4

Reliably, these frogs would ovulate within 18 hours of being injected with the urine sample of a pregnant person, laying easily visible eggs at any time of year. Unlike the rabbit test, this new “bioassay” was faster and left the animal perfectly unharmed for future use. Word spread quickly, and within a few years, biological institutions everywhere were inundated with ovulating Xenopus frogs.

Thanks to the discovery of the specific “pregnancy hormone” (chorionic gonadotropin, or hCG) and lateral flow devices that can detect its presence, modern-day pregnancy tests do not require rabbits, frogs, or indeed, any other animal. But their widespread use for pregnancy testing was what likely helped entrench Xenopus laevis (and a few of its cousins) in scientific research.

Xenopus frogs possess a strangely zen-like charisma. Their lidless eyes protrude from the tops of their flat heads, and when not feeding, swimming, or mating, they tend to laze in the water and stare at you. They lack tongues; instead, they open their maws and shovel in prey with their front feet as it passes. These same feet allow them to grip females during mating, while they use their webbed back feet, equipped with distinctive black claws on the last three toes, to swim. These appendages inspired their name, “Xenopus,” which means strange-footed.

The sudden abundance of Xenopus in laboratories during the 1930s and 40s may seem like a stroke of luck for biologists, but in truth, frogs were not newcomers to research labs. Well before Xenopus frogs were used as “obstetrical consultants” (as one scientist referred to such animals), many other frog species played a role in developmental biology.

The main reason for this was convenience: frogs are relatively easy to find, catch, and house, as anyone who grew up near a pond can attest. They also lay many eggs at a time, and their eggs are generally larger than most fish eggs but unencased by hard shells like bird eggs, making them excellent research subjects for understanding embryonic development.

The most prominent early frog experiments were performed way back in the 18th century by Italian scientist, Lazzaro Spallanzani (whose other accolades include the discovery that bats use echolocation and that digestion harnesses the chemistry of gastric juices). By brushing unfertilized eggs with frog sperm, Spallanzani proved that fertilization could occur outside the female body — a finding he later confirmed in mammals through the successful in vitro fertilization of a dog.

Perhaps his most memorable experiment, however, was his demonstration that it was specifically semen that caused fertilization. He proved this using a negative control: Spallanzani sent off male frogs to “seek the females with equal eagerness, and perform, as well as they can, the act of generation” while wearing miniature taffeta breeches that he had meticulously sewn for their tiny legs, preventing their semen from reaching any eggs. The result of this experiment, he wrote, “is such as may be expected: the eggs are never prolific [e.g., fertilized], for want of having been bedewed with semen, which sometimes may be seen in the breeches in the form of drops.”5

Over the next hundred years, European scientists continued to make crucial discoveries in the burgeoning field of embryology by manipulating amphibian eggs. However, these studies were usually done using locally available species, such as frogs from the genus Rana (which includes most common pond frogs of Europe, Asia, and North America). It was a century and a half after Spallanzani sewed his silk condom-pants that American and European biologists found themselves with an abundance of Xenopus eggs on their hands thanks to their use in fertility testing.

The same qualities that made African clawed frogs attractive for pregnancy testing also made them excellent laboratory models; they do fairly well in captivity. Xenopus are fully aquatic species6 with simple needs, at least by most amphibian standards. They eat almost anything, can survive in tap water,7 and can live for up to twenty years. And a single brood from these frogs yields hundreds to thousands of eggs at a time.

For scientists, these eggs are far more interesting than the adult frogs. Just one-tenth the size of a typical marble, but clearly visible to the naked eye, these two-toned eggs are dark on one side and light on the other, a pattern reminiscent of a Poké Ball. The contrast between these two colored sections (with a grey crescent in between) becomes more pronounced after fertilization, when the pigmentation concentrates at the site of sperm entry. This darker half of the egg is known as the “animal” pole, the half which will divide into smaller cells more quickly, and is thus more “animated” relative to its “vegetal” pole, the lighter-colored half characterized by dense yolk.

Today, one can find numerous videos showing how no special dyes or equipment (beyond, perhaps, a magnifying lens) are necessary to observe the early rounds of cellular division in Xenopus eggs; meticulous work by dozens of scientists has since determined which organs will arise for these early cell divisions.

Beyond their visual properties, frog eggs are also useful research subjects because their embryos are easily manipulated; eggs can be poked, prodded, and even completely torn apart while still maintaining the capability to develop into tadpoles. One of the most beautiful experiments from early embryology, conducted by Hans Spemann and his student Hilde Mangold8 in the 1920s, was originally done in newts, but can be easily replicated in frog eggs: if you remove a small piece of tissue off one amphibian embryo and attach it onto a second, this ripped-off chunk of cells will develop along with the “donor” egg such that the resulting tadpole grows to have two fully formed heads.

Spemann and Mangold’s extraordinary discovery sparked an international race, from Japan to the United States, to identify the chemical factors responsible for the activity of the “organizer” tissue, as Spemann and Mangold called it. The “miners” of the time, however, lacked the metaphorical tools needed to dig up the secrets of these organizer cells. Although they knew, generally, that all biological traits were controlled by chromosomes housed in the cell’s nucleus, they did not know how those chromosomes gave rise to genes.

That changed in 1953, when Watson and Crick published their celebrated papers describing the structure of the DNA molecule and how it suggested a mechanism of gene inheritance. The work over the next decade that would unveil the secrets of the genetic code was mostly done in simple models such as bacteria, viruses, or yeast cells. But it was, incidentally, a Xenopus experiment that provided the most dramatic demonstration that even adult animal cells retained the entire genetic library of information necessary to recreate an organism.

In 1968, following experiments by Robert Briggs and Thomas King showing that the nucleus of one cell can be successfully transferred into another, developmental biologist John Gurdon took the nucleus from an adult frog cell and injected it into an egg whose nucleus had been removed. The resulting egg, now carrying DNA from an adult “parent,” was capable of growing into a normal tadpole and then adult frog, with all the varied cell types of a normal animal. This new frog was a “clone,” from the Greek word for “twig,” being grown, as it were, from a clipping of an adult tree instead of from a seed.9 Dolly, a sheep cloned in 1996, gained notoriety as the first cloned mammal, but the first animal cloned from an adult cell was actually a Xenopus frog, nearly three decades earlier.

Today, if you overhear a molecular biologist talk about their “cloning work,” they are unlikely to be discussing attempts to make an identical twin out of a grown animal. Instead, they are probably speaking about the cloning of specific genes; synthesizing or copying the sequence of a gene (the specific combination of DNA chemical letters that spell it out) and amplifying it by taking advantage of the rapid proliferation (and gene copying capacity) of the bacterium E. coli. However, this still routes us back to the frog, as this basic technique of having bacteria express a gene from a eukaryote was first achieved with the successful cloning of a Xenopus gene in 1974.

Genetic cloning was made possible by three discoveries: restriction enzymes, molecular scissors that cut DNA at specific sequences, ligases that could rejoin those fragments into new arrangements, and the idea of including antibiotic resistance genes alongside the cloned gene so that successful constructs can be identified using antibiotics. Herb Boyer at UC San Francisco, who had purified the restriction enzyme EcoRI (and would go on to found Genentech, widely regarded as the world’s first biotechnology company, in 1976), conceived the idea alongside Stanford bacterial geneticist Stanley Cohen while the two were discussing their research over sandwiches at a Hawaiian deli in 1972.10 After transforming bacteria with their stitched-together plasmids, Boyer recounted checking the results: “I went to look at the gels in the darkroom, and there it was. It actually brought tears to my eyes, it was so exciting.”

Boyer and Cohen, along with other scientists, soon went further than introducing bacteria-derived genes back into other bacteria. In 1974, their team spliced the Xenopus ribosomal RNA gene into E. coli cells. This frog gene was selected because it had been thoroughly “characterized and can be isolated in [a large] quantity.” This extraordinary result demonstrated that bacteria could be coaxed into reading and copying genes from across a billion-year evolutionary divide, laying the foundation for the fields of genetic engineering, synthetic biology, and biomanufacturing.

For all its upsides, however, Xenopus laevis had its shortcomings. The frog, so favored for its use in fertility testing and embryology, turned out to be poorly suited to genetic research because, instead of having two copies of each of its genes (one from each parent), it has approximately four of each. This chromosomal messiness is not uncommon in nature, but having four similar — but not identical! — copies of their chromosomes makes sequencing laevis genomes significantly more difficult. A complete laevis genome assembly wasn’t released until 2016, trailing a decade or more behind other model organisms such as Mus musculus (mouse) and Rattus norvegicus (lab rat) or agricultural workhorses like Bos taurus (cow) and Gallus gallus (chicken).

Even without a solid understanding of Xenopus genetics, however, biologists in the 1980s found another remarkable way to put Xenopus eggs to good use. Eggs are essentially large cells; thus, a scientist with a large supply of eggs is also in possession of many easily manipulated cells. Just as young frog embryos remain intact after being prodded by needles, the cellular components of frog eggs remain biologically active even after those cells have been gently crushed and separated away from their yolks, chromosomes, and heavier components, thanks to the marvels of centrifugation. This cellular extract could be generated in large quantities, and the resulting soup of proteins and organelles would still carry out many of the most intricate operations of a cell, but in a more easily observable and controllable system.

Because these systems are not cells, but contain almost everything inside cells, they have been especially useful for understanding the amazingly well-coordinated mechanisms of cell division. Xenopus egg extracts were essential in the discovery of proteins that control the cell cycle, the dance of chromosomes as they arrange themselves for cell division, the self-assembly of structural proteins called microtubules as they direct and move those chromosomes, and the role of small messenger proteins, called Rans, in many aspects of cellular activity. What made all these discoveries possible is not only the fact that the components of these cell-free extracts are easily visualized, but also that they are exposed; there is no membranous barrier between the cellular components or any chemical that a scientist may wish to introduce into the system (or remove, such as by fishing out proteins using antibodies).

Throughout the 1980s and 1990s, scientists looking to conduct genetic research on Xenopus frogs without the genetic challenges of X. laevis did some chromosomal explorations of related species, and by the late 1990s, hit upon Xenopus tropicalis, a species that possesses many of the advantages of X. laevis but has only two sets of each chromosome instead of four. Some labs switched to using these frogs, but many groups that had well-established protocols for X. laevis were reluctant to do so, especially since Xenopus tropicalis requires different housing and handling conditions. Xenopus tropicalis frogs are also smaller than X. laevis, with smaller eggs and brood sizes. Thankfully, genetic engineering has advanced a great deal over the past two decades, and today both X. laevis and X. tropicalis are commonly used in biology research.

Even if much more is known now than in the 1930s, when Xenopus frogs were first hopping around research labs, they are still being used to solve pressing questions of basic biology, as vertebrate animal models for drug development and testing, and in experiments whose findings may only be fully understood in years to come. Thanks to their historic role in the science of fertility and embryology, four African clawed frogs were even lucky enough to be astronauts in 1992, when they boarded the Space Shuttle Endeavour to test whether or not reproduction and development could occur in zero gravity. These frogs were indeed capable of laying eggs and producing viable offspring while aboard the space shuttle. After all, nobody forced them into wearing pants.

Subscribe now

Matt Lubin is a PhD candidate studying microbiology at Johns Hopkins University.

Acknowledgements: The author thanks Dr. Ross Pederson and Dr. Christine Field, scientists who currently work with Xenopus, for their helpful comments and corrections, as well as the editorial team of Asimov Press.

Cite: Lubin, M. “A Brief History of Xenopus.” Asimov Press (2026). DOI: 10.62211/86wn-33ro

By Sam Clamons

The interior of a cell is densely packed with millions of molecules vibrating, jostling, and moving about. Sugar molecules fly through a cell at 250 miles per hour, ricocheting off of ribosomes, organelles, cytoskeletal fibers, and enzymes. Indeed, every protein in the cell is hit by about 10¹³ water molecules each second. This chaos makes biology seem hopelessly convoluted. With everything moving so quickly, how can we begin to understand biomolecules?

As with other hard-to-intuit quantities in science, one could look up biological rates using resources like PubMed or BioNumbers, to discover facts like “water flows through aquaporin at 100 million molecules per second,” or “yeast transcribes RNA at 0.12 molecules per minute.” But knowing a number doesn’t necessarily give one a feel for it. Are those rates… fast? How do they compare to protein folding? Or enzymatic activity? Or squeezing a muscle?

In short, how fast do things in a cell happen, from the perspective of the molecules it’s made from?

We can answer this question with a quantitative metaphor, by visualizing the most important goings-on of a typical cell slowed down to speeds that are still accurate relative to one another, but matched to what we experience in the everyday world. The slowdown factor we pick should make it easy to understand the molecular machines that run our cells — proteins. Ideally, we would scale the fastest functionally important protein event to match the shortest unit of human perception.

As a representative “fastest functionally important protein event,” I’ve picked the opening of a membrane-bound ion channel — specifically, a potassium-gated ion channel.1 This protein channel opens and closes to allow potassium ions into the cell. This conformational change must happen very quickly. So, for the sake of our metaphors, let’s imagine slowing down the opening of this channel 10,000x, so it takes as long as the blink of an eye. If this were the case, then…

Molecular Events

• The fastest motions in chemistry are the internal vibrations of molecules. Hydrogen atoms in a water molecule, for example, constantly stretch, bend, rock, and twist. Even slowed 10,000x, such molecular bonds would still wiggle billions of times each second — a million times faster than the vibration of a piano’s middle-C string.2

• Molecular velocity is a different story. At our 10,000x slowed-down scale, an average room-temperature water molecule moves at a zippy 6 centimeters (about a finger-length) per second (1,300 mph real time).3

• Finally, there is average displacement. Our zipping water molecules collide so frequently that they don’t appear to be traveling in one direction, but rather they mostly jiggle in place. They do diffuse around eventually, but that diffusion is much slower than 6 cm/s. Diffusion rate isn’t a speed,4 but we can see how long it would take (on average) to diffuse some important cell-scale distances. At 10,000x time scale, water in the cytoplasm diffuses:

  • The width of an E. coli (1 µm) in 8 seconds (~1 millisecond real time).

  • The width of a typical skin cell (20 µm) in an hour (⅓ seconds real time).

  • The width of a human hair (100 µm) in a day (10 seconds real time).

Protein Events

The fundamental mechanical unit of biology is the protein — machines built out of chains of amino acids that do most of the work of life at the molecular scale. Before a protein can be used by a cell, though, it must be built. The sequence of the protein must first be read from DNA to messenger RNA (transcription) and then from the mRNA into a chain of amino acids (translation). Assembled amino acid sequences also need to fold into their proper shape, and sometimes require further processing before they’re fully functional.5 At our 10,000x slowdown….

• During transcription, RNA polymerases add ribonucleic acids to growing mRNA at a rate of about one base every 3.5 minutes (50 bp/s real time). This means that a typical bacterial gene will get transcribed in about 2.5 days (20 seconds real time).

  • Human genes contain massive non-coding introns, so they take more like two months (10 minutes real time). These mRNAs will need to be processed by further downstream steps before they can be passed to a ribosome.

• During translation, a ribosome adds an amino acid to the end of a growing protein about once every half hour (5/second real time). For a typical protein (bacterial or eukaryotic) of a few hundred amino acids, this means that translation takes about 6 days (1 minute real time).

• A chain of amino acids needs to fold into its functional form to be useful. The time this takes is roughly exponential with the square root of the length of the protein, which adds a lot of variability to protein construction rates:

  • The fastest-folding proteins, which are little more than a single helix or similar structure, fold in about 20 milliseconds, which is a fifth of an eyeblink (2 microseconds real time).

  • Proteins of more typical size (a few hundred amino acids) take more like an hour or a day to fold (0.3-8 seconds real time). Slow, but still faster than translation — these proteins likely fold as they’re being built, and may be ready to work shortly after they’re released from the ribosome.

  • Proteins with complex or slow maturation processes (some of which require covalent modification, not just folding) can take much, much longer to mature than they take to be synthesized by a ribosome! GFP (green fluorescent protein) takes 5.5 days to translate but almost a year to become functional (1 minute real time to translate, 1 hour real time to mature)!

What kind of time-scales do proteins act on, in our metaphor? Proteins do many different things over many different time-scales, so we’ll have to select some illustrative examples. At the 10,000x slowdown…

• When a voltage-gated potassium channel opens, the outside-the-membrane half of the gate first uncoils gradually over the course of about 1 second (100 microseconds real time). After a delay averaging another second, the inner half of the channel will snap open as fast as an eyeblink, after which potassium will start flowing in a rapid stream (10 microseconds real time).

• ATP synthase, in the process of producing ATP, rotates about once every 50 seconds (200 rotations/second real time). Each rotation assembles three ATP molecules.

• A distant cousin of ATP synthase, the swimming flagella of an E. coli bacteria, is a bit faster, rotating every 30 seconds (30 rotations/second real time). A similar flagellum used by Vibrio alginolyticus rotates five times faster, once every six seconds (2,000 rotations/second real time).

• Superoxide dismutase, which converts superoxide free radicals into less-toxic hydrogen peroxide, is one of the fastest enzymes, working at a maximum rate of ten reactions per second (100,000 reactions per second real time!) — about as fast as the action of a voltage-gated ion channel.6

• An average metabolic enzyme under saturating conditions, on the other hand, processes about one molecule every 16 minutes (10 reactions per second real time).

• A kinesin motor takes one (8 nanometer) step in about 30 seconds (3 milliseconds real time).

How long do proteins persist within a cell? Again, answers vary substantially.

• Most proteins in growing yeasts have metaphor-scale half-lives between 50 days and 5 years, with a median of about 300 days (8 minutes to 4 hours real time, median 45 minutes).

  • The shortest-lived yeast proteins last about 3 weeks (3 minutes real time). The longest-lived have half-lives of over 200 years (8 days real time, or about 130 yeast-generations).

• Most proteins in a non-growing human fibroblast cell7 now have half-lives between 20 and 80 years, with a median of about 65 years (1-3 days real time, median 2.5 days).

• Ornithine decarboxylase, among the most ephemeral human proteins, has a metaphorical half-life of about 75 days (10 minutes real time).

• On the other end of the spectrum, the longest-lived known human protein is collagen, which provides strength and elasticity to skin. In physiological conditions, the average collagen protein sticks around for about 1.2 million years (120 years real time) — longer than the human in whose skin it is embedded!

Cell and Organism Replication

• DNA can be copied by a reasonably fast DNA polymerase, like that of E. coli, at 1 base pair every 17 seconds (600 bases/second real time). In E. coli, that polymerase sticks to the DNA for an average of 10 days before falling off (1.5 minutes real time), having replicated about 1 percent of the genome.

• A fast-growing E. coli strain growing in ideal conditions in a laboratory divides once every 7 months (30 minutes real time).

• A human fibroblast in growing tissue (in a well-kept flask of media) will divide every 22 years (20 hours real time). Each cell cycle can be divided into four broad phases:

  • G1: 7.5 years (7 hours real time). This is when the cell produces RNA and proteins required for DNA replication.

  • S: 10 years (9 hours real time). This is when the chromosomes are replicated.

  • G2: 3.5 years (3 hours real time). This is additional protein production, growth, and preparation for mitosis.

  • M: 1 year (50 minutes real time). Mitosis, during which time the cell actually divides.

• A fruit fly in the lab lives about 820 years (1 month real time). Since nobody has a sense for what 820 years feels like, this is probably farther than we should stretch this particular quantitative metaphor — we’ll need a different one to directly relate to animal lifespans.

Brains and Neurons

To get a sense of the timescales of brain operation, let’s trace out one stimulus-response cycle (specifically, how a stimulus triggers a nerve impulse, which then leads to a response) at metaphorical speed. Consider a typical (non-athlete) college student being given the ruler-drop reflex test, where they are asked to wait until a ruler is dropped, then catch it as quickly as possible. In our metaphorical, 10,000x slowed-down time scale:

• Without intervention, a 1-foot ruler will fully pass through the student’s hand in 41.7 minutes (250 milliseconds real time).

• Light takes about 10 microseconds (3 nanoseconds real time) to travel from the ruler to the student’s eye. This is the physical limit on how quickly any physical process could react to the drop. As we’ll see, it’s a trivial duration compared to the time it will actually take a brain to react.

• Photoreceptors start sending measurable voltage signals within 50 seconds (5 milliseconds real time) of being struck by light. This process is known as phototransduction.

• Neurons in the retina start working in response to signals from photoreceptors within 1.5 minutes (9 milliseconds real time).

• After passing through a handful of layers of neurons in the retina, signals travel through the optic nerve to the visual processing system of the brain at somewhere between ⅓ and 1 millimeter per second (3-10 meters per second real time). The optic nerve is about 8 cm long, so that translates to a 1–4 minute transit time (8-24 milliseconds real time) from eye to brain.8

• The visual cortex collates raw contrast information into lines and textures, lines and textures into simple shapes, and simple shapes into complex shapes and moving objects. About 15 minutes (90 milliseconds real time) after the first instant of the drop, the brain has processed the first photons coming off the ruler and can, at least in theory, start issuing a command to grab it.

• The brain acts on its awareness of the drop and decides to catch the ruler in about 6 minutes (36 milliseconds real time).9

• Once the brain decides it’s time to catch the ruler, it needs to get a signal from the skull to the fingers. The peripheral neurons that carry this kind of information typically move signals at 4-5 mm/s (40-50 meters per second real time), which adds a delay of about 4 minutes (24 milliseconds real time) between the decision and the hand’s response.

• Finally, it takes a bit of time for the finger’s muscles to actually move to catch the ruler. Fingertip muscles can accelerate at about 1.1 mm/s², which translates to a travel time of 5 minutes over the 5mm or so required to pinch the ruler (30 milliseconds real time).

• From direct measurements, we know that this whole process for a typical young adult takes about 30 minutes (200 milliseconds real time). A mixed martial arts fighter can catch the ruler in about ¾ the time.

In short, at our metaphorical time scale, a speed test that we normally perceive at about the speed of an eyeblink is instead a half-hour affair, of which about half is waiting for neural processing and a quarter is waiting for signals to move along nerve fibers.

Human-Made Things

Now that we’ve calibrated a timescale for molecular biology, we can apply our metaphor to human-created devices, like electronics. What does human technology look like from the perspective of biochemistry?

• Accessing one byte of memory from RAM takes around 10 milliseconds — ten times faster than opening/closing a voltage-gated ion channel (1 microsecond real time).

• Accessing one byte of memory from a solid-state SATA hard drive takes about 1 second (100 microseconds real time).

• A typical consumer car engine at cruising speed cycles through one combustion cycle every 5 minutes (30 microseconds real time, 2,000 RPM). The jet engine of an F-22 Raptor is moderately faster, at about 40 seconds per turbine rotation (4 microseconds real time, 15,000 RPM).

• An LED flat screen TV refreshes once every 1.4 or 2.8 minutes, depending on the refresh rate (8.5 or 17 milliseconds real time). Similarly, one frame of a 30FPS video game lasts about 5.5 minutes (33 milliseconds real time), and one frame of a Hollywood-standard movie lasts about 7 minutes (42 milliseconds real time).

• A one-second clock tick by an analog wrist watch takes about 3 hours (1 second real time).

Human technology is pretty speedy, even compared to biochemistry. Processors act on timescales that make voltage-gated ion channels look positively sluggish, and, amazingly, modern jet engines can rotate a little bit faster than ATP synthase, even though they are tens of millions of times larger and more than 24 orders of magnitude more massive!10

Humans and the Everyday

As could be guessed from the last few biological examples, our time metaphor loses much of its usefulness when applied to entire organisms. If a metaphor describes events in terms of epochs longer than the subjects of most history books, it’s hard to argue that the metaphor helps ground intuitions in the everyday.

This is a good place to step back and take stock of the temporal ground we’ve covered. From the vibrations of intramolecular bonds to the lifespans of simple invertebrates, we’ve spanned 24 orders of temporal magnitude — one order of magnitude more than the span covered in our spatial metaphor article.

The shortest ten orders of magnitude of duration are the realm of molecular motion, with molecules vibrating, spinning, and crashing about without discernible biological function. The next nine orders of magnitude — from tenths of a second to months — cover most of the biological processes we think about in molecular biology, from enzymatic processing of small molecules to sophisticated neural signal processing up through division of a fast-growing bacteria. Going up another four orders of magnitude allows us to see eukaryotic cell division, turnover of the longest-lived proteins, and lifespans of long-lived animals like humans.

Yet we still haven’t covered the time scales of ecology and evolution! Even for our metaphor, evolutionary time scales are geologically slow compared to chemical ones. To build intuitions about rates of evolution, then, we’ll need to use a different time metaphor entirely, as we’ll explain in the next essay in this series.

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.

Cite: Clamons, S. “Metaphors for Biology: Time.” Asimov Press (2025). DOI: 10.62211/27qi-28hq

Header image by Ella Watkins-Dulaney.

Correction: This article has been updated to include real rates, in parentheses.

Science’s most studied organism is, without contest, E. coli. It has prompted at least half a million academic papers, spanning nearly a century of work. And yet a quarter of its genes still do not have an experimentally-determined function, according to a 2019 study.1

If we don’t yet understand the inner workings of a bacterium that scientists have obsessed over for so long, consider how little we are likely to know about everything else.

Scientists have estimated that about one trillion microbial species inhabit the Earth, of which 99.999 percent remain undiscovered.2 Of those that have been cataloged, perhaps a few thousand have been grown in a laboratory. In other words, a vast chasm remains between the microbes one can see under a microscope or locate in the dirt and those we can grow on a petri dish.3 Rounding to the nearest whole number, we have deeply studied zero percent of lifeforms on this planet.

Even if such rounding is hyperbolic, it points to the vast number of possibly useful discoveries still on the table. CRISPR was first found in a salt-loving microbe in Spain. The thermostable DNA polymerase used in modern PCR was discovered in a Yellowstone geyser. Biologists find useful things nearly everywhere they look.

So why don’t we grow and study more organisms in the laboratory?

A few reasons. First, the genetic tools for manipulating more “established” organisms, like E. coli, yeast, and HeLa cells, are quite reliable. Why would a PhD student spend the formative years of her career mixing up liquids to grow an obscure microbe that nobody else has ever heard of? It’s much easier to take a labmate’s vial of cells and get to work straight away on a well-defined research problem.

Another reason is that the combinatorial complexity of biology is vast and difficult to navigate. A typical protein in E. coli is 300 amino acids long, with 20 possible amino acids at each position. The number of possible sequences for this hypothetical protein is thus 20³⁰⁰, more options than there are atoms in the Universe. And that’s just for proteins!4

Similarly, finding the “perfect” ingredients to grow and engineer a new organism requires an equally vast search through combinatorial space. Some microbes need special carbon or nitrogen sources, vitamins, pH levels, and temperatures. Others die if they are exposed to even a hint of oxygen. If our hypothetical PhD student makes a growth media, dips cells into it, and nothing grows, then that failed experiment also returns little useful information; was it the pH or oxygen to blame, or something else?

Finally, even if one manages to grow cells, that is only the beginning for most biology research. We also need to be able to transform, or manipulate DNA inside of cells. After all, scientists figure out how genes work by subtracting them from the genome and then watching what happens to the cell (usually, it dies). Or, to confirm a gene’s role in a phenotype or disease, scientists replace a defective copy with an unmutated version and see if the addition renders a fix.

This substitution process only works if scientists can grow and manipulate cells using genetic tools. Both steps are required, and the latter bit — manipulation — usually begins by coaxing cells to take up “foreign” DNA with a technique called electroporation, in which pulses of electricity are used to punch physical holes in a cell through which DNA can sneak inside. But much like finding ideal growth conditions, electroporation has dozens of different parameters spanning many millions of possible combinations, and it’s not often clear which settings will lead to success.

Now it should be clearer why making headway in the biology lab is so difficult. As it so happens, this is exactly the problem that Cultivarium, a nonprofit research organization, is trying to solve. For a recent study, they created a robot that finds working conditions to both grow and engineer non-model microbes. They are giving these growth protocols away for free, in the hopes that others will use them and begin studying a wider range of lifeforms on Earth.

Search Space

In the 1950s, physiologists Alan Hodgkin and Andrew Huxley were studying giant squid axons, which can extend nearly a meter in length and are visible to the naked eye, to understand how electrical signals travel through them. They discovered that electrical signals, rippling along this axon, somehow triggered alternating tides of sodium and potassium moving in or out of neurons. Twenty years later, two scientists at the Weizmann Institute, Eberhard Neumann and Kurt Rosenheck, returned to the Hodgkin-Huxley experiment, wondering if the electrical signals that coax small molecules to move about naturally in the brain could also be used to “program” a cell to release other molecules instead.

Neumann and Rosenheck took lipid vesicles and filled them with proteins and ATP, a type of energy-storing molecule. These vesicles were then wedged between platinum electrodes and zapped. The duo measured which molecules were released from the vesicles using a spectrophotometer, and found that the shocked vesicles dumped ATP while larger proteins remained trapped inside. Their results suggested that membranes became permeable — but were not entirely destroyed — when electrocuted.

In 1982, Neumann probed further, investigating whether he could use electricity to get molecules, such as DNA, into cells. At the time, recombinant DNA was becoming more popular, and dozens of scientists were trying to clone things like the human insulin gene into bacteria. The main methods for doing this were dipping cells in calcium-phosphate (which carries DNA through the membrane via electrostatic charges), microinjection, and viruses. However, none worked reliably across different cell types.

So Neumann, ever the tinkerer, found a tube of frozen mouse fibroblast cells deficient in thymidine kinase, an enzyme required for DNA synthesis. These cells die in HAT medium, a selective liquid that only allows cells with a functional form of this enzyme to survive and grow. First, Neumann created a plasmid carrying the thymidine kinase gene. Next, he performed several trials, trying many different parameters to push this thymidine kinase gene into these deficient cells. He tuned the electrical field strength, pulse durations, temperature, and so on. Finally, he found settings that worked: out of one million cells in the shocked vial, 95 took up the plasmid and grew. The cells that took up the DNA regained the thymidine kinase gene, made the enzyme, and thus survived in HAT. Neumann called his method “electroporation.”5

Today, electroporation remains the most common way to get DNA into cells. The electrical pulses punch tiny, temporary holes in the lipid bilayer, thus allowing charged molecules (like DNA) to slip inside. The method is general, meaning it can work (in theory) on any organism. But it isn’t foolproof, and when it fails entirely, “you often don’t know what went wrong,” says Nili Ostrov, CSO of Cultivarium. “If you get no colonies at all, it could be the plasmid, it could be the conditions, it could be the machine, it could be the buffer, it could be the voltage, and you just don’t know.”

The total space of parameters for a single electroporation experiment is vast. One commercially available electroporator machine, called the BTX Gemini X2, has more than 4.2 million possible settings for voltage, resistance, and capacitance. And this doesn’t even include biological variables, like the buffer in which cells are washed, pH levels, or temperature.

In practice, most scientists today do electroporation using small cuvettes. They wash their cells with some buffer, mix in the DNA, and then put the cuvette into a machine that zaps them. The cells are then transferred into liquid media to “recover” (as they get stressed by the process) after which they might be plated with some selection media, like an antibiotic. If the transformation works, little colonies will appear on the plate after a few hours. If nothing grows, then the scientist has to adjust and try again.

This time-consuming procedure inspired Cultivarium to automate the process. As opposed to a human painstakingly choosing parameters and filling the cuvettes, they built a custom electroporator robot that clamps down onto a 96-well plate and can deliver different electrical pulses to each well. Every well can also have different buffers and DNA concentrations. The machine is controlled through an API, and the team can test 24 different conditions simultaneously in a single run. (The device isn’t yet for sale and Cultivarium has not released the blueprints. They also declined a request to share a photograph of the robot, citing biosecurity concerns.)

Cultivarium’s first experiments were relatively simple. They took different microbes and seeded them across 24 wells of a microplate. In each well, they varied the wash buffer (sucrose, sorbitol, water, or glycerol), voltage (1, 2, or 3 kV), and waveform (exponential decay or square wave). They shocked all 24 conditions at once, coaxing the cells to take in a plasmid containing an antibiotic-resistance gene. Then, a second robot transferred each well of cells into liquid growth media with the antibiotic. These cells recovered overnight. The next day, a laser measured the optical density, or cloudiness, of each well. Cloudy wells meant bacteria grew; clear wells meant failure.

Cultivarium quickly found electroporation settings that work for many non-model microbes: organisms that nobody had ever transformed before, including Halomonas elongata (used by some companies to make bioplastics) and Piscinibacter sakaiensis (a microbe famous for eating plastic).

Some of the “optimal” parameters were quite surprising. For example, the wash buffer, or liquid used to prepare cells for transformation, had a huge effect on how well the electroporation actually worked. For each of the microbes, one buffer outperformed the others by at least 100-fold. Voltage was also important; while one kilovolt was usually enough to transform cells, increasing this to two or three kilovolts could, in some cases, lead to 1,000-fold improvements. (Higher voltages can also kill the cells, though, so one must tread carefully.)

But electroporation can fail for another reason: even if DNA enters the cell successfully, the plasmid won't survive unless the bacterium recognizes its origin of replication, or the genetic element that instructs the cell to copy the plasmid. To disentangle these two problems, Cultivarium thus combined its electroporation robot with a library of plasmids — each with a different origin of replication — to see which ones worked in each microbe.

They shocked all these plasmids into the microbes at once, grew the survivors, and then sequenced whatever grew to identify the working plasmids. Using this approach, they were able to transform two species of ocean-dwelling microbes, called Shewanella, that nobody had managed to work with before.

Finally, the team developed an active learning framework, in which they let a Bayesian optimizer choose which settings to try in each experiment. To begin, Cultivarium’s robot runs an initial round of experiments, testing 176 different conditions spread across the parameter space. The results — like which conditions yielded colonies, and how many colonies — are fed into the algorithm, which then suggests the next conditions.

“We take these results and then ask the model to suggest which ones we should optimize around,” says Stephanie Brumwell, lead author on the paper. “The goal is to let the model suggest what we cannot rationalize.”

They used this Bayesian optimizer on Cupriavidus necator, a bacterium that converts carbon dioxide from the air into a biodegradable plastic, called PHA. Over three cycles of active learning, the model found transformation efficiencies 8.6-fold higher than anything previously described in the literature.

Untapped Nature

I’ve profiled Cultivarium before, and my doing so again is a first. But their work feels deeply important because it is by tinkering with non-model organisms that biotechnologists have, historically, found the most useful tools.

Taq polymerase and CRISPR were both discovered in unusual organisms, as I said before. Green fluorescent protein was first isolated from jellyfish caught off the coast of Friday Harbor, Washington. Luciferase came from the North American firefly. Rapamycin, an immunosuppressant used in organ transplants, came from Easter Island soil microbes. Many more useful discoveries are surely awaiting, if only we could grow the many lifeforms that nobody has yet studied in any meaningful way.

Consider a scenario in which scientists discover a microbe that eats plastics, like Piscinibacter sakaiensis. They’d like to study this microbe in the laboratory. They begin by sequencing the cells’ genome, quickly spotting the stretch of DNA encoding the plastic-degrading enzymes.

From here, the scientists could try and synthesize the plastic-eating genes and put them into another microbe, like E. coli, that is easier to work with. This is basically the logic of metagenomics: sequence environmental samples, identify interesting genes computationally, and express them in a more cooperative host. This approach was recently used to discover Bridge recombinases, a new type of genome-editing tool.

But this doesn't always work. Many fascinating biochemical mechanisms, especially those involving multiple genes, can only be understood in context; they stop working entirely when transplanted into another cell.

To deeply study a biological mechanism, then, the only option is often to work with the native host, complete with its evolutionary baggage. So the scientists take some broth off the shelf and dip the plastic-eating cells in, yet nothing grows. So they try a different medium, and another, again and again. They tweak the temperature, pH, and oxygen levels and then — at last! — the cells begin to grow.

Now they want to actually use this organism to do things. Perhaps they want to delete a gene to prove its link to plastic degradation. Or they want to overexpress an enzyme to see if they can speed up its plastic-munching. For these experiments, they need to get foreign DNA into the cell. They need electroporation — or something like it — to work.

All of this tinkering takes time.

And this, at last, is why I think Cultivarium’s robot is such a big deal. It’s not only about scaling up plastic-eating microbes, but about scaling anything else, too! This work makes it feasible for scientists to work much more rapidly with a vastly larger number of organisms, which I would argue is one of the most fundamental bottlenecks in biology and which, when solved, will fling scientific discovery into unimaginable directions.

Niko McCarty is a founding editor of Asimov Press.

Cite: McCarty, N. “Solving the Electroporation Bottleneck.” Asimov Press (2026). DOI: 10.62211/28hw-48hq

In biology, genetic variation often results in things that look more or less the same. Despite the panoply of genetic sequences in our cells, for example, we end up with a limited number of tissue types. This is known as canalization, the idea that, despite genetic variation, environmental forces, and randomness, lots of genotypes yield the same phenotype. This is why many different diseases cause similar symptoms. Or why body shapes are more limited than one might expect.

As a lapsed biologist, a favored pastime of mine is taking ideas from biology and applying them to other domains; parachuting into a field with a simple model shouldn’t just be the province of physicists! So I am tempted to ask where we find canalization in the world of metascience.

One obvious place to look is institutional forms, or “phenotypes” if you will. While there’s a high-dimensional space of possible institutional forms, we have traditionally only explored a small subset of it: universities, corporate research labs, startups, and a handful of others. Gratifyingly, over the past few years, there has been an explosion of new research organizations; researchers are trying new things and traversing this space, and I’ve been collecting them in my Overedge Catalog.

Yet, it seems that the new research organizations that have cropped up are not all that different from previous iterations. Many are cool and interesting, and I’m thrilled they exist, but, whether in the Overedge Catalog or not, many appear almost interchangeable in their structure. There has been a canalization of organizational forms.

If you go the route of a for-profit research lab, for example, your institution might end up looking like a startup. Or you might try to construct a strange research institution but end up with something that looks like an independent version of a university department, with colleagues that look like faculty and perform faculty-like tasks. While canalization is potentially valuable in biology in that it provides a kind of developmental robustness, it is probably not ideal for institutional innovation, as it constrains the form, and thus the function, of these organizations.

But why does this canalization occur in science?

Well, with organizations that end up looking university-like, it comes down to risk, or more precisely, risk aversion. Imagine you build a weird institution devoted to some odd interdisciplinary topic. Then, you hire people to work in it. These are often academic researchers with varying levels of exposure to academia’s incentives and requirements. Each hire then begins to think the following: “What if this organization crashes and burns? Or what if I’m simply not a good fit for it? I need to make sure I can get another job afterward.” And for the most part, “job afterward” means another academic job.

Thus, due to what metascience commentators Michael Nielsen and Kanjun Qiu term the “shadow of the future,” there is a certain amount of risk aversion: these employees still want to publish in traditional journals or do things that seem reasonable upon their return to the academic community. As Nielsen and Qiu note:

Scientists considering working for Jazzy Not-for-Profit (or for-Profit) Startup Institute must ponder: do they really want to give up publication in high impact journals? Or to work on risky projects that may not pan out? Or to do anything else which violates the norms of their scientific community? If they ever decide to leave their Jazzy Startup Institute job, won’t they then have a tough time finding another good job? After all, other potential employers haven’t changed their standards, just because Jazzy Startup Institute has. The shadow of the future looms strongly for such ventures, causing a kind of regression to the institutional mean … The perennial question within such organizations is: if I hew to local aspirations, will that damage my chances of getting a job anywhere else?

And thus, many of these organizations are funneled by their own staff, not necessarily even consciously, into similar kinds of forms. The people within the organization are driven to do academic-like activities, making the organization as a whole more academic. This “regression to the institutional mean” is a kind of canalization, yielding a smaller subset of forms than we might expect (or desire).

If, on the other hand, you go the route of trying to build a venture-backed R&D lab, you are faced with another canalizing force: investors and their money. While there are no doubt exceptions, such as investors with a great deal of patience and willingness for researchers to try lots of things, or companies like Genentech that made fundamental research breakthroughs, too often venture investment means a startup must do startup things. Ben Reinhardt, who runs Speculative Technologies (I’m an advisor), has written a wonderful little essay with the provocative title “When should an idea that smells like research be a startup?” Reinhardt writes:

A big reason that market uncertainty can kill researchy startups is less about investment and more about the fact that there is a tradeoff between an organizational culture that is good at addressing market uncertainty and one that is good at addressing technological uncertainty. At some point a researchy startup needs to do a dramatic gear shift into growth and product-market fit mode. This transition often either prematurely kills the research potential or the company dies because it’s being run by people with a research mindset.

To be clear, if venture investors push for this, that’s a reasonable and responsible part of their job (I work for a venture capital firm, for goodness sake!). A startup needs to scale to succeed within the venture capital model. Trying to bend this kind of structure and make it weird and capacious for research can yield valuable science, but it is unlikely to yield a fundamentally different and novel kind of institutional form.

Even when you have a single investor who seems mission-aligned, you can still end up with a reversion to the mean. Paul Allen, the cofounder of Microsoft, created Interval Research in the 1990s with lots of open-ended research, but within a few years, Allen asked for “less R and more D.” As one researcher there noted, “We’re moving from the ‘Let a thousand flowers bloom’ stage to being more tightly focused on commercializing our technology.” By 2000, Interval Research had shuttered. So it goes.

Yet another canalizing force is seen implicit in the first line of the excerpt from Nielsen and Qiu: the tax code. Due to my involvement in the space of non-traditional research organizations, I speak with many people who are thinking about building new institutions. A common question that I get asked is whether to go non-profit or for-profit. This decision will impact the kind of people or organizations they approach for fundraising, the regulations they will need to adhere to, and so forth, and these things should not be taken lightly. Yet, is it not odd that our tax codes have reduced the complexity in how researchers think about science? We imagine the vast and high-dimensional space of outlier research institutions, and then are forced to collapse it into these two categories because of tax implications.

Given these pressures, canalization might seem our destiny. However, I don’t think it is. There are clearly counterexamples to this narrative, where organizational structures exist that are indeed novel.

For example, focused research organizations seem qualitatively different and are something worth fostering: they blend aspects of startups and research organizations within a time-bound structure to construct something new. There was also a for-profit machine learning research lab called Fast Forward Labs that was entirely bootstrapped — taking no venture backing — that made money by producing quarterly reports about machine learning advances (it was eventually acquired by Cloudera). This was definitely different! And there’s Ink & Switch, which thinks about research work in terms of the Hollywood movie model, where teams assemble for specific projects, and then members might move on to something else.

But while it’s possible to push against canalization, it’s difficult. You have to try hard to avoid traditional incentive structures and do something far less pigeon-holeable. So what mechanisms might allow greater success?

One option may be to hire more weirdos or misfits (or whatever it makes sense to call those people less concerned with the shadow of the future). If these individuals aren’t thinking about returning to academic positions, this could allow an organization to avoid being channeled into academia-lite.

Or if everyone is obsessed with non-profits and for-profits, we may need to create new legal structures. There is the designation of the benefit corporation and even the B Corp certification, so perhaps, as a law professor friend of mine once suggested, we need an R corp? What would a Research Corporation look like, and what legal structure could allow different kinds of science to flourish within it? I don’t have the answer, but I think it’s worth exploring.

I think we might also just need more of these organizations, with funders willing to create the space and time for experiments to play out, rather than prematurely directing them into specific categories. This means that we need more rich people to do some really weird philanthropy, supporting crazy ideas in science and scientific organization. Hey, tech multimillionaires, be risky! Worry less about fitting the philanthropic mold and, instead, find people who want to make an institution that is truly different. The key for philanthropists aiming to start a successful research organization is simply to give money to interesting teams and then walk away. Minimize the oversight, lock the money up, and let smart people do their thing. Let the research take its time.

Or perhaps we need an institution that is really just a group of loosely affiliated independent researchers. That organization would have one person who fundraises for the researchers — The front man? The hype man? — and that’s it. Might this end up doing something truly different and weird? Let’s find out.

The National Science Foundation also recently announced its Tech Labs initiative to fund teams with large amounts of money for long stretches of time for research outside the traditional academic setting. This experiment is still in its early stages, but there are some exciting possibilities here.

Suppose, however, that your organization ends up succumbing to the forces of canalization. Even that’s not all bad. These are tried and true institutional forms, and they can still do great things. OpenAI, for example, has crept closer to the structure of a traditional tech startup over time, but it has certainly had an outsized impact on the world of AI. If, however, you want to explore the high-dimensional space of institutional structures, keep pushing for something new.

Samuel Arbesman is Scientist in Residence at Lux Capital and the author of The Magic of Code, Overcomplicated, and The Half-Life of Facts. In addition, he is an adjunct professor at Case Western Reserve University’s Weatherhead School of Management and a research fellow at the Long Now Foundation.

Cite: Arbesman, S. “Why Do Research Institutes Often Look the Same?” Asimov Press (2026). DOI: 10.62211/28jw-57eh

For the last several years, I have been trying to understand why biomedical progress, especially in therapeutics, has become less productive despite staggering advances in basic science.

I am not the only person vexed by this. In 2012, biotechnologist Jack Scannell formally described the dwindling returns on therapeutic investments, coining the term Eroom’s Law (Moore’s Law in reverse). Eroom’s law states that the inflation-adjusted cost to bring a new drug to market roughly doubles every nine years: a trend that has held since the 1950s. With the goal of upending Eroom’s law, I have spent the last year studying the structural bottlenecks that shape how new medicines are tested and the FDA’s role in such decisions.

Much of my time has been focused on clinical trials, which, despite their central role in the creation of pharmaceuticals, receive remarkably little systematic attention. This led me to launch the Clinical Trial Abundance Project, a framework aimed at increasing not only the number of clinical trials, but also their speed and how much we learn from them. Recently, I co-authored an essay with Scannell, arguing that making trials more efficient and informative is essential to breaking Eroom’s Law.

Critics of our essay, however, argued that making clinical trials more efficient risks treating biotechnology like a casino. In their view, making it easier to run clinical trials would risk allowing more potentially harmful drugs to be tested in patients and, instead, biotechnologists should focus on making better drugs that are more likely to gain approval. These critics see Clinical Trial Abundance as accepting the status quo of drug development rather than challenging it.

But this is a misunderstanding.

In fact, Clinical Trial Abundance and better hypotheses for drugs are not merely compatible, but self-reinforcing. Faster testing in the clinic creates a feedback loop: ideas become trials, trials generate rich data (including both successes and failures), these data improve models, and better models inform the next generation of ideas. In this view, the clinic is not an endpoint of discovery but a central component of it.

To understand why clinical abundance is important, we must step outside the prevailing view of clinical testing as a mere “validation step” for scientific ideas. The familiar funnel metaphor of drug discovery, depicting a linear progression from basic science to regulatory approval, reinforces the flawed notion of clinical testing as a passive filter designed to screen pre-existing ideas. While this model is narrowly correct in a regulatory sense, it obscures the clinic’s role as an active engine of discovery.

The reality is that clinical trials rarely just deliver a “yes/no” verdict on a drug’s efficacy. Instead, the history of drug development shows that many successful therapies emerged only after initial versions failed in specific, informative ways. When a trial fails, it provides a unique physiological stress test that reveals exactly where a drug’s design fell short. By collecting data from “failed” trials, we can transform negative results into experimental corrections for the next iteration.

Subscribe now

Consider CAR-T cell therapies. Once thought implausible or risky, CAR-T therapies now deliver long-term, treatment-free remissions in cancers where relapse had been almost certain.1 In pediatric B-cell acute lymphoblastic leukemia (B-ALL) and aggressive B-cell lymphomas, for example, CAR-T has cured patients who, previously, had been given only months to live.

CAR-T therapy works by turning a patient’s own immune cells into living drugs. Doctors collect T cells from the blood, genetically reprogram them to recognize a protein on cancer cells, and reinfuse the modified T-cells into the patient. These engineered cells expand inside the body, move to tumor sites, and destroy malignant cells.

In 2017, the FDA approved Kymriah, the first CAR-T therapy, for children and young adults with relapsed or refractory B-ALL, a cancer of arrested development in which immature blood cells, specifically B cells, multiply out of control while failing to mature into viable immune cells. Relapsed B-ALL is the most severe form of the disease, because it means the cancer has returned after prior therapy. Even with aggressive care, only 10-20 percent of patients with relapsed or refractory B-ALL survived beyond five years.

Against this backdrop, Kymriah received accelerated approval from the FDA based on results from the Phase II ELIANA trial, a global, multicenter study sponsored by Novartis. In ELIANA, 82 percent of treated patients achieved complete remission, and subsequent follow-up analyses revealed that five-year survival rose to approximately 55-60 percent.

ELIANA was not a sudden breakthrough, though. It was, rather, the culmination of nearly two decades of clinical studies. During this period, CAR-T therapies evolved through repeated failure in the clinic, as careful studies of underwhelming results spurred new ideas to correct them. The ELIANA trial was led by investigators at the University of Pennsylvania, a group that had spent years studying CAR-T cells directly in patients well before regulatory approval.

In the mid-2000s, the earliest CAR-T therapies first entered human testing. And they emerged from a fundamental question: is it possible to engineer and redirect the T cell’s innate killing power against malignant cells?

Two well-established biological concepts made this seem plausible. First, T cells are extraordinarily cytotoxic.2 However, their natural activation is governed by a “layered permission” system, meaning they cannot recognize targets directly, but must wait for other cells to process and present protein fragments in a precise molecular context. While this evolutionary safeguard keeps us from being attacked by our own immune system, it also provides cancer with many opportunities to evade detection by suppressing these signaling pathways.

To bypass these safeguards, researchers relied on a second insight: the ability of antibodies to bind directly and precisely to proteins on the surface of cells, called antigens. By equipping T cells with a synthetic Chimeric Antigen Receptor (CAR), bioengineers created a functional shortcut that bypassed the need for permission systems. This receptor uses antibody-style recognition to lock onto a cancer cell and is wired directly to CD3ζ, a signaling molecule that triggers the T cell’s internal “kill switch.” The moment the receptor engages its target, it flips the internal switch, activating the cell’s killing program.

In laboratory experiments, these first-generation CAR-Ts were formidable, displaying antigen recognition and potent killing power against tumor cell lines. Yet, this in vitro prowess vanished in patients and did not yield durable clinical responses. Understanding why this happened, though, was not simple. The failure could have been caused by a breakdown in in vivo antigen recognition, poor signaling strength, or other defects that only emerged after the cells were injected into the body.

Progress in understanding why early CAR-T therapies did not live up to their promise came from treating first-in-human trials not simply as therapeutic attempts, but as opportunities to learn. These information-dense studies were conducted throughout the mid- and late-2000s and were relatively small (usually enrolling fewer than ten patients). However, they were designed to be maximally revealing. Researchers used many tools to monitor CAR-T persistence and activity in the body, turning information from a small number of patients into a mechanistic understanding.

One such tool was quantitative PCR (qPCR), a lab method that detects and counts specific DNA sequences, which allowed researchers to measure how many CAR-T cells were in patients' blood. This showed that CAR-T cells successfully entered the body and were easy to detect after infusion. But the signal quickly faded, suggesting that the cells died off quickly. Other experiments shed light on the problem: CAR-T cells could recognize and eliminate cancer cells in patients — meaning antigen recognition was working — but their functional activity fell over time, suggesting that something in the blood was blocking them.

At this point, the diagnosis of why first-generation CAR-T therapies were failing matched long-standing insights from basic immunology. Decades of research had shown that T cells are not governed by a single on-off switch: signaling through CD3ζ provides only the first activation signal. To keep working, T cells need additional “costimulatory” signals,3 delivered through receptors such as CD28 or 4-1BB. First-generation CARs had been designed to deliver signal one without signal two, which explained their poor performance.

This hypothesis guided the next wave of clinical experiments, which investigated whether adding a costimulatory domain would make CAR-T cells more effective at clearing tumors in vivo.

The field-defining result came from Carl June’s group at the University of Pennsylvania. June and colleagues explored a costimulatory domain called 4-1BB. In a first-in-human study published in 2011, they treated a patient with chronic lymphocytic leukemia using CAR-T cells containing both CD3ζ and 4-1BB. They also administered a dose that was remarkably small by cell therapy standards at that time: just 1.5 × 10⁵ CAR-T cells per kilogram of body weight. (A first-generation CAR-T trial targeting renal cell carcinoma, published in 2013, used a dose more than 100-times higher.)

What followed was unprecedented. The CAR-T cells multiplied more than a thousandfold in patients, at their peak comprising a large fraction of immune cells in the blood. The CAR-T cells also persisted for months. Now, at last, CAR-T cells were a long-lived and self-maintaining immune population. Many cancer patients treated with these second-generation CAR-T therapies have achieved complete remission.

The next step, though, was asking whether the same CAR-T behavior would work in a faster, more aggressive blood cancer, such as B-ALL. In 2013, Carl June’s lab reported striking results in two children with relapsed B-ALL, again showing that the engineered T cells could multiply, persist, and drive cancer into remission.

All of these lessons were built into ELIANA, the study that ultimately supported Kymriah’s approval. Led by Stephan Grupp, who had treated the earliest pediatric patients and worked closely with June, ELIANA translated the early insights into standardized practice. This trial codified chemotherapy given before CAR-T infusion, scaled up cell manufacturing, and measured success using tools like qPCR.

Viewed through this lens, clinical trials are not an alternative to basic science, but rather a mechanism within it that closes a feedback loop. Foundational immunology, antibody engineering, and molecular biology made first-generation CAR-T cells possible in the first place, but early human trials quickly revealed that these designs were incomplete and suggested ways to fix them.

Yet theory alone did not prove this would work; the expansion and persistence observed with 4-1BB–based CAR-T cells came as a genuine surprise even to the therapy's designers. “It was unexpected,” they reported, “that the very low dose of chimeric antigen receptor T cells that we infused would result in a clinically evident antitumor response.”

This shows why the “casino biotech” critique is flawed. It assumes that experimentation simply reveals a fixed probability of success. But trials can change those probabilities. When clinical testing is understood as part of a continuous feedback system, optimizing trial efficiency is not about accepting failure but about learning fast enough to make success more likely.

The most discovery-rich experiments are often not massive Phase III trials, either, but small, academic, investigator-initiated studies that sit close to the design loop.4 These are also the trials most burdened by regulatory, institutional, and manufacturing bottlenecks.

For example, early CAR-T studies are often required to use full “good manufacturing practice” (GMP) facilities — the same industrial standards used to mass-produce approved therapies — which can drive costs into the millions even for studies with just a handful of patients. Many researchers argue that such custom, tightly monitored studies do not need full commercial-scale GMP, and Australia offers a real-world example: Its regulatory system allows small-scale trials to use more flexible manufacturing under strict oversight, enabling safe human studies at far lower cost than in the U.S. or Europe. More worrying still, the researchers I interviewed who run these trials consistently report that institutional bureaucracy has become harder to overcome in the last several years.

If we want biomedical progress to accelerate, we should stop treating clinical trials as an afterthought or as a binary “yes/no” part of drug development. Instead, we should ask how to make the design-build-test loop better in humans. Some answers will be policy-driven, like lowering barriers to ethical investigator-initiated trials, enabling adaptive development, and making better use of clinical data. Others will be technological, including richer in-human measurement and better monitoring.

The aim of the Clinical Trial Abundance project is thus broader than merely proposing policy solutions for accelerating trials. It seeks to remind people of the importance of in-human data as a tool for scientific discovery.

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. “Clinic-in-the-Loop.” Asimov Press (2026). DOI: 10.62211/38jw-33ht

The Harz Mountains, whose low ridges sprawl across northern Germany, have been said to be teeming with witches. The poet Goethe wrote in Faust: “Da drängen sich Hexen zu tausend zuhauf” (“Witches throng together by the thousand”). But while such a coven has never actually been found in the Harz, something even more storied has — Arabidopsis thaliana, the premier model for plant biology.

Arabidopsis thaliana was first described by Johannes Thal, a German doctor and botanist born in 1542, who stumbled across it while on an alpine walk.1 From a young age, Thal had been transfixed by the nature of the Harz, collecting and cataloging grasses, herbs, and various resinous plants that might prove medically useful.2 No species was too unremarkable for Thal, fortunate given that Arabidopsis thaliana resembles one of those spindly flowers that sprout between cracks on sidewalks or in the window wells of abandoned cars.

While it may look like a weed, Arabidopsis belongs to the Brassicaceae, or mustard family. Native throughout Africa and Eurasia, it typically grows in rocky, sandy, and chalky soils. Sometimes called “thale cress” or “mouse-ear” cress, Arabidopsis does indeed have flowers whose pale petals resemble the soft, rounded tip of a mouse’s ear. Atop a 20-30 centimeter leaf-studded shoot grow several pale flowers, each with four petals arranged in a whorl.

Although Thal was the first person to pay any mind to this unprepossessing plant, another German botanist, Friedrich Laibach, took the first steps to bring Arabidopsis from the herbarium into the modern research laboratory. As part of his work as a Ph.D. candidate in Bonn in the early 1900s, Laibach spent his time analyzing the number of chromosomes in the various plants that he had collected around the city and his hometown of Limburg. He did this by staining the plant tissue with a special dye that binds to chromatin, the mixture of DNA and proteins that form chromosomes. By examining cells during meiosis — when chromosomes are most condensed and distinct — he could directly count the number of chromosomes possessed by each species. When Laibach turned his attention to Arabidopsis, he found that the plant carried only five pairs, one of the smallest numbers known at the time.

This genetic simplicity captivated Laibach, who began to collect seeds from every Arabidopsis type he could get his hands on, imploring colleagues to look out for new species while traveling. Between 1930 and 1950, he collected seeds from over 150 different ecotypes (later called “accessions”), which he maintained at the Botanical Institute at Goethe University in Frankfurt.

In 1943, Laibach published a paper in which he made a case for the use of the plant as a genetic model. A. thaliana, he asserted, was not only easily cultivated year-round but also self-fertilizing, meaning pure lines could be easily maintained. It was also able to be cross-pollinated by hand, which made it well-suited to experimental breeding. Finally, its small genome made it ideal for cytological and genetic studies, since its hereditary factors could be more easily traced.

The plant’s reception, however, was lukewarm. In the 1940s, plant biology was dominated by economically important crops like maize, wheat, and tobacco, which already had established research communities and clear agricultural relevance. Few agreed with Laibach’s evaluation of the small weedy plant.

One exception was Hungarian plant biologist György Rédei, who read Laibach’s article in 1955 and recognized Arabidopsis’s potential for genetic studies. He obtained four natural accessions (Graz, Limburg, Estland, and Landsberg) from Laibach and brought them to the University of Missouri in Columbia when he left Europe. For the next 20 years, Rédei remained the only researcher working on Arabidopsis in the United States.

While working in a lab he inherited from Nobel Prize-winning cytogeneticist Barbara McClintock, Rédei used radiation to create Arabidopsis mutants with stunted growth. He hoped that by studying the inheritance patterns of this trait, he could map the genes responsible for developmental processes such as stem elongation, hormone regulation, or cell division.

As he worked with these mutants, Rédei discovered that the Landsberg population he had received from Laibach contained genetic variation accumulated in the wild. To establish stable reference lines for future work, Rédei selected two different plants from this heterogeneous population. The first was from the collection of Landsberg seeds that he had irradiated. From these, Rédei selected a line dubbed “Landsberg erecta” after the upright inflorescences (arrangement of the components in the flower head) of the originating mutant. A few years later, he established another line, Columbia (Col-0), which would become the other major reference strain and the one still used to this day.

Enthusiasm for Arabidopsis research remained tepid in the United States during the 60s and sparked only slightly more attention in Germany. In 1964, the eminent plant scientist Gerhard Röbbelen published the first Arabidopsis Information Service (AIS) newsletter to connect the small, but fervent Arabidopsis community.3 The newsletter also acted as a seedstock, including Laibach’s original accessions from decades earlier. A year later, in 1965, Röbbelen organized the first international Arabidopsis conference in Göttingen — attracting just 25 participants.

This modest attendance wasn’t so much a marker of exclusivity as a reflection of the ambivalence still surrounding the value of this plant model. During this period, plant biologists were not only preoccupied with agriculturally important plants, but also believed that tissue culture — the ability to grow and manipulate plant cells in laboratory dishes — would be essential for the future of plant biology. Plants like petunia and tobacco were favored because researchers could readily culture their cells, manipulate them, and then regenerate whole plants from those cultured cells. Attempts to do this with Arabidopsis, however, were frustrating; members of the Brassicaceae family do not grow well in dishes, often developing with deformed or fused leaves.

In the late 1960s and 70s, the world of plant research didn’t have any interest to spare for Arabidopsis. Researchers were busy puzzling over the bacterial pathogen Agrobacterium tumefaciens, which caused tumor-like growths on many plants. In 1969, Australian plant pathologist Allen Kerr recognized that bacteria could pass the tumor-causing trait to each other, writing that “this seems to be the first unequivocal evidence for transfer of virulence from a plant pathogenic to a saprophytic species of bacteria … ”

Shortly after, studies by groups including Jeff Schell and Marc van Montagu in Belgium, and Mary-Dell Chilton in Eugene Nester’s laboratory in the United States, demonstrated the relationship between crown gall disease and the tumor-inducing plasmid (Ti plasmid). Taken together, these discoveries pointed toward the radical hypothesis that A. tumefaciens transfers its own genetic material to plants, and that it might even be possible to exploit this natural process to transfer genes into plants.

A paper on the first transgenic Arabidopsis plant was published in May of 1986, and another followed just five months after from a team at Monsanto, quickly sparking excitement over its genetic makeup. Its authors noted that “The low incidence of repetitive DNA and small genome of A. thaliana suggest several potential uses for the more than 100 independently derived transgenic plants obtained with this procedure,” and concluded that “Demonstrations of the feasibility of these approaches to the identification, isolation, and analysis of specific plant genes will guarantee the position of A. thaliana as ‘the Escherichia coli of the plant kingdom.’”

The breakthrough that really vaulted Arabidopsis closer to this lofty status came in 1989, when researchers first demonstrated that Agrobacterium’s T-DNA could systematically knock out genes. By inserting T-DNA randomly into the Arabidopsis genome, they could disrupt individual genes and create mutant plants lacking specific gene functions. This “reverse genetics approach” became the standard method for characterizing plant gene function and led directly to the creation of comprehensive mutant collections. The Salk Institute’s T-DNA collection and similar global resources now provide researchers with ready-made knockout mutants for nearly every Arabidopsis gene, making it possible to study gene function without generating new mutants from scratch.

By the early 90s, it appeared the transformation of Arabidopsis had occurred at two different levels — that of its DNA (now manipulable thanks to plant transformation) and that of its larger role within the research community. No longer an innocuous mustard varietal, Arabidopsis seemed a promising model organism. This was also the “Genome Era,” with the Human Genome Project well underway by 1990. Plant biology found itself swept up in the same current of excitement, enticed by the possibility of mapping a plant’s genetic blueprint.

Because Arabidopsis had one of the smallest and simplest plant genomes (~135 million bases across five chromosomes), researchers saw it as the ideal candidate for the first plant to be fully sequenced. While doubtlessly easier to sequence than a human genome, it still took roughly a decade to complete. The complete sequence was published in Nature in 2000 (four years after Baker’s yeast, the first eukaryotic genome, and one year before the human genome).

In the years that have elapsed, the value of Arabidopsis has only grown. For one, it has become much easier to work with. The year before the full genome sequence was published, researchers discovered the “floral dip method” in which flowering Arabidopsis plants are simply dipped into a solution containing Agrobacterium, allowing the bacteria to naturally transform developing seeds without any need for tissue culture. This breakthrough expanded the Arabidopsis research community dramatically. “Until other plants with fully sequenced genomes can be transformed with comparable ease, Arabidopsis is unlikely to be displaced as the plant of first choice for experimental molecular geneticists,” write Arabidopsis pioneers Chris Somerville and Maarten Koornneef in Nature.

A. thaliana’s primacy as a plant model, however, extends beyond molecular genetics; it’s been employed to decipher the genetic control of flowering, leaf and root formation, hormone signaling, circadian rhythms, and plant–pathogen interactions. And because many of its genes and pathways are conserved across plant species, insights gained from Arabidopsis routinely guide crop improvement and environmental research, making it a central reference species. It has even grown in space in an orbital laboratory, proving its resilience is not gravity-dependent.

In the end, Arabidopsis’s origin in Harz, a region renowned for its witchcraft, seems fitting. This thale cress is indeed a shapeshifter; once the passion project of a few dozen researchers, it has transformed itself into the laboratory’s most iconic plant.

Subscribe now

Xander Balwit is editor-in-chief of Asimov Press and a Brassicaceae superfan.

Huge thanks to Evan Groover for providing feedback and clarifications on earlier versions of this draft. Header image by Ella Watkins-Dulaney, with credit to Marie-Lan Nguyen.

Thank you also to Dr. Marc Somssich, whose piece, Short History of Arabidopsis thaliana, clarified a lot of the plant’s early history and pioneers.

Cite: Balwit, X. “A Most Important Mustard.” Asimov Press (2025). https://doi.org/10.62211/83gr-98jh

In 1674, a Dutch cloth merchant in Delft, Antoni van Leeuwenhoek, spent his free time tinkering with lenses. One day, he pressed a drop of rainwater beneath his homemade microscope and observed what he called “animalcules” darting about. Leeuwenhoek described these creatures in two letters he sent to the Royal Society in London for publication, each replete with charming descriptions:

[T]he motion of most of these animalcules in the water was so swift, and so various upwards, downwards and round about that ‘twas wonderful to see: and I judged that some of these little creatures were above a thousand times smaller than the smallest ones I have ever yet seen upon the rind of cheese.

Leeuwenhoek may have been the first person to see microbes in motion, but his microscopes weren’t powerful enough to see the actual machinery responsible. (Leeuwenhoek mused that his animalcules might be using “little paws” to move.) It wasn’t until the 1830s that a German naturalist, Christian Gottfried Ehrenberg, saw whisker-like appendages, later named flagella, protruding from microbes. Still, the mechanism by which they worked remained unknown until the latter half of the 20th century, when electron microscopes finally homed in on the thousands of proteins that make a flagellar motor, able to convert flowing protons into mechanical motion.

Even more recently, a surge of research has revealed how evolution has finetuned the flagellum to operate in vastly different ways based on a cell’s niche. Whereas an E. coli flagellum spins around nearly 20,000 times per minute, the flagellum in a microbe called Vibrio alginolyticus spins about five times faster, or slightly more than 100,000 times per minute. (For context, a Boeing 737 rotor has a maximum speed of 14,000 rpm.) This extra rotary speed is because Vibrio cells must traverse the ocean, where ion gradients (used to drive the motors) are large and nutrients more spread out.

A microbe found in the human digestive system, Campylobacter jejuni, also has a flagellar motor that generates much more torque than the one in E. coli — about 3,600 piconewton-nanometers. It uses this higher torque to propel itself through the viscous environment of the human gut.

Earlier this year, a detailed structure of the high-torque C. jejuni motor was resolved for the first time. The structure revealed that the bacterium has extra motors located further away than normal from its central driveshaft. These motors push and spin the tail with greater leverage, and thus higher torque, than any other flagellum yet discovered.

This recent paper is just one contribution toward a structural-biology revolution that is finally shedding light on the molecular mechanisms underlying Leeuwenhoek’s 17th-century speculation about microbes’ “little paws.” It shows we’re still tackling a puzzle first uncovered more than 350 years ago.

Cryo Structures

A flagellum looks deceptively simple from the outside (“just a long tail!”) but contains hundreds of interlocking proteins of many different types. Flagella have evolved independently at least three times — in bacteria, archaea, and eukaryotes — because they solve a fundamental problem for cells: namely, locomotion.1

Despite a high variance in structures, bacterial flagella are composed of four parts:

• the basal body, which makes up both the motor and driveshaft;

• stators, which convert the energy of flowing protons into torque;

• a flexible hook that connects the motor to the final part;

• a tail, which spins around like a corkscrew to push the cell forward.

The most amazing part of the flagellum is not only its composition of interlocking proteins or nature’s construction of literal motors, but rather that this mechanism self-assembles, piece by piece. Every protein in the flagellum fits perfectly to its neighbors. From within the cell, each protein finds its way to the growing structure, guided by transport systems that thread new parts through the base and out to the tip, where they click into place to complete the machine.

The working parts of the flagellum lie embedded inside the cell, spanning the inner membrane, periplasm (space between membranes), and outer membrane. When a bacterial flagellum assembles, it begins with the basal body, which is itself made from more than a hundred interlocking proteins and spans all three of these layers.

In 2021, Oxford researchers used cryo-electron microscopy to resolve the molecular details of the basal body from a microbe called Salmonella Typhimurium. The first part of the basal body, the C-ring, is located inside the cytoplasm just beneath the inner membrane and is built from three proteins — FliG, FliM, and FliN — repeated 34 times around the base. This structure converts torque from the stators into mechanical rotation on the rod.

Above it lies the MS-ring, made entirely of a protein called FliF. About 34 FliF subunits assemble into a double-layered ring embedded in the inner membrane, acting together like a chassis to connect the C-ring in the cytoplasm to the drive shaft, or rod.

The rod itself is made from a rigid column of five proteins called FliE, FlgB, FlgC, FlgF, and FlgG. These proteins stack end to end, transmitting torque from the rotating MS-ring to the external hook and filament. Around the rod, two more rings, called the P-ring and L-ring, act as friction-reducing bearings, keeping the rod aligned as it spins thousands of times per minute.

The next piece of the flagellum is the hook, a short, curved segment that functions like a universal joint. It’s made from about 120 copies of a single protein, called FlgE, that hold the tail as it spins.

And then there is the tail itself, a helical propeller that spins like a corkscrew, pushing on the fluid in its viscous environment to drive its cell forward. A bacterial flagellum is typically up to ten micrometers in length, which, in E. coli, is significantly longer than the bacterium itself. Proportionally, this is akin to the tails of grass lizards or widowbirds, trailing three or four body lengths behind them.

Unlike animal tails, however, the flagellum is not a bundle of muscle or bone but rather a hollow cylinder of protein that builds itself, one subunit at a time. In fact, the tail is made almost entirely of repeating subunits of one protein, called flagellin. After the hook has self-assembled, the cell begins pumping unfolded flagellin proteins through it and into the hollow tail outside the cell. Each protein diffuses through this tail tube, reaches the growing tip, folds, and locks into place. As writer James Somers once quipped, “it’s as if the flagellum were built by vomiting forth parts of itself.”

Because the movement of each flagellin protein is diffusion-limited (meaning each protein drifts randomly through the narrow channel, rather than being actively pushed), each new protein needs to move further than the protein preceding it. And the further each flagellin protein must travel, the longer it takes the tail to grow. Longer flagella therefore elongate more slowly than shorter ones.

Finally, there are the stators, the actual motors that convert flowing protons into the mechanical energy that spins the tail.

It was long believed that the flagellum spins by burning ATP, an energy-storage molecule abundant in cells. But in 1977, researchers in Colorado did a clever experiment that disproved this. The researchers poisoned cells, collapsing the proton gradient without disturbing ATP levels. As soon as the poison was added, the cells stopped swimming. This showed that the flagellum is powered not by ATP, but rather by the flow of protons across the cell membrane.2

These protons flow from outside to inside a cell through small channels formed by two proteins embedded in the membrane, called MotA and MotB. Every time a proton passes through this channel, MotA changes its shape ever so slightly, nudging a protein embedded in the rotor, called FliG. This happens around a million times each second, collectively spinning the tail tens of thousands of times per minute.

These stators are incredibly energy efficient, too. It takes about 50 protons to power one revolution of a stator, with more than 90 percent of available energy being converted into mechanical work. To put that into context, a typical combustion engine converts only 20-30 percent of the chemical energy stored in gasoline into mechanical work at the crankshaft. The rest is lost to heat and friction. The cell as a whole swims far less efficiently, though. Only about 0.2 percent of energy is transferred into forward motion, because much is lost to the drag of water at microscopic scales.

But despite their elegant design, not all flagella are created equal. Different bacteria have adapted their flagellar motors for the worlds they inhabit, with some optimized for speed and others for force. One of the most striking variants, in my eyes, belongs to the microbe, Campylobacter jejuni, which uses its flagellum to drill through the viscous mucus of our intestines.

Evolving Torque

A C. jejuni cell looks a bit like a microscopic eel. Its name derives from the Greek kampylos and baktron, meaning “curved rod.” It is a pathogen, and one of the most common causes of food poisoning in the U.S. and Europe. This microbe thrives in the human gut; an environment far more viscous than water. To swim through it, C. jejuni has evolved a flagellum that has a torque far greater than the one in E. coli.

In July of this year, researchers at Imperial College London imaged the C. jejuni flagellum using cryo-electron microscopy, which works by rapidly freezing biological molecules into a thin layer of glass-like ice, trapping them in a nearly-native state. When an electron beam is shot through the frozen sample, particles strike the sample and scatter into a detector to form a noisy, two-dimensional image. By collecting hundreds or thousands of such images and aligning them computationally, researchers can calculate 3D maps of proteins.

It wasn’t possible to solve the C. jejuni flagellum directly because it sits deeply embedded in the bacterial envelope, making the cells too thick for electrons to pass through. To surmount this problem, then, the researchers first coaxed the microbes to make “minicells,” basically tiny, round daughter cells carrying only a small amount of cytoplasm.

They did so by deleting a gene called flhG, which controls where the cell division machinery assembles, and other genes encoding flagellins, the proteins used to build the flagellum’s long tail. The result was a population of uniform, 200-nanometre-wide minicells, each carrying an intact, tailless flagellar motor near the surface. These minicells were thin enough for electrons to penetrate, thus allowing the researchers to image the motor at subnanometer resolution.3

The solved structure revealed that C. jejuni’s core motor isn’t that different, in principle, from E. coli’s, with a rotor in the middle and stators around it that act like little engines, each powered by flowing protons. But the C. jejuni flagellum has 17 stators compared to the 11 in E. coli. These stators are also positioned further away from the central driveshaft, or axle, by a large, newly-evolved scaffold structure, which enables them to exert more leverage.

Extra protein disks in their periplasm act as a cage to fix their stators in place, holding them at just the right distance from the rotor. In E. coli, the stators are loosely tethered and can drift in and out of position, so the motor runs below capacity. In C. jejuni, however, they’re locked down and permanently switched on. The end result of more engines, pushing from farther out, is three times the torque.

This paper on flagella is remarkable because it reveals how Darwin’s admiration of evolution’s “endless forms most beautiful” also applies at the molecular scale. The flagellum is not a single design but a spectrum of mechanical solutions to the same problem: how to best move a cell. What began in the 1600s as an investigation into the “little paws” that Van Leeuwenhoek could only imagine has evolved today into a high-resolution portrait of the proteins that drive their mechanisms.

Subscribe now

Niko McCarty is a founding editor of Asimov Press.

Thanks to Morgan Beeby, associate professor of structural biology at Imperial College London, for helpful comments on this essay. Header image by Ella Watkins-Dulaney. Mistakes are my own.

Cite: McCarty, N. “Animalcules and Their Motors.” Asimov Press (2025). DOI: https://doi.org/10.62211/36qa-94vx

In 1712, Thomas Newcomen, a Baptist preacher and ironmonger from Dartmouth, Devon, constructed his newly designed “atmospheric engine” for Coneygree Coalworks. Each stroke of the 20-ton machine would raise about 37 liters of water from the flooded mines below. It worked continuously and without tiring, unlike the dozen horses it replaced.

Like those horses, however, Newcomen’s engine needed to be fed, burning copious amounts of the very coal that it was helping to unearth. Although these engines were cost-effective compared to horses or humans, they were still expensive to operate, and improvements to their efficiency were especially sought after.

In 1763, while working as a mathematical instrument maker at the University of Glasgow, James Watt was tasked with repairing the university’s scale model of the Newcomen engine. As he worked, Watt envisioned ways to improve the efficiency of the design. In 1776, he unveiled an engine with seemingly extraordinary modifications: it consumed 75-80 percent less fuel than Newcomen’s. Tasks that would have burned 100 kilograms of coal could now be done with a mere 20.

Although a triumph of engineering, Watt’s engine was nowhere near its optimal performance. The difference between Newcomen’s and Watt’s engines, in fact, is between 0.5 percent and 2.5 percent efficiency. But these inventors could not have known this because the concept of a theoretical limit — asking how efficient an engine design could be, in principle — had not yet been imagined.

What was missing was a notion we might call “Limit Thinking.” This abstract approach forces one to focus solely on the features of a system essential for its performance, so that one can make predictions or evaluations, regardless of the specifics of how each individual system is built. It grounds problems in mathematics — as one cannot calculate limits without being precise about what is being measured and in what units. Once such calculations have been determined, they often drive rapid progress, signaling not only when we have reached diminishing returns, but also just how far we can aspire.

To understand the power of this approach, let’s begin with engines, returning to the work of Nicolas Léonard Sadi Carnot.

Carnot, named after the Persian poet Saadi, was born in Paris in 1796 and attended the École Polytechnique before serving as an officer in the engineering arm of the French army. In 1819, at the age of 22, he took a part-time job in the army at half pay as a travailleur scientifique (scientific worker) and attended lectures at the Sorbonne, the Collège de France, and the Conservatoire des Arts et Métiers in his spare time. It was during this period that Carnot’s interest turned to heat engines.

In 1824, when Carnot was just 28 years old, he published Reflections on the Motive Power of Fire and on Machines Fitted to Develop this Power, in which he aimed to determine the fundamental limits of how heat could be converted into mechanical work. This 118-page booklet, of which he printed 600 copies at his own expense, represents essentially the whole of his scientific output (regrettably, Carnot died at 36 years old from a combination of scarlet fever and cholera).

Reflections was largely ignored until 1834, when the French engineer and physicist Émile Clapeyron highlighted its importance in his own work, Memoir on the Motive Power of Heat.1

Carnot’s work showed that what mattered for the efficiency of a heat engine was the temperature differential between the hot and cold reservoirs, rather than any particular design feature. He says of the motive power of heat: “Its quantity is fixed solely by the temperatures of the bodies between which it is effected.” In retrospect, this simple fact explained why the Watt engine was superior to the Newcomen engine; the separate condenser allowed for a larger difference between the reservoirs.

Carnot’s work would go on to inspire not only Clapeyron but also a generation of prominent physicists such as Rudolph Clausius, James Joule, and Lord Kelvin himself. It laid the groundwork for the modern theory of thermodynamics and even presented an outline of what would later become its second law. Otto Diesel directly refers to the theory of Carnot2 as the rationale for his design of the engine that now bears his name (a design which achieved an efficiency of an astounding 26 percent when first tested in 1897). While the 65 years between Newcomen and Watt had seen efficiency improve roughly two percent, in the 65 years after Carnot, it leaped by about 23 percent and has climbed steadily higher even since.

125 years after Reflections, another engineer, Claude Shannon, was working at Bell Telephone Laboratories on a different practical problem: how to make telephone conversations clearer. Long-distance calls were plagued by distortion and static; amplifiers added noise, cables weakened signals, and cross-talk muddled speech. Engineers could tinker with hardware, but they lacked any way to quantify “clarity” itself — a measure of how much of a message survived its transfer from sender to receiver. Shannon reframed this into a Limit problem: not how to send more information, but how much information could ever be sent through any noisy channel.

In answering this, Shannon defined information mathematically, in terms of entropy, and calculated the maximum rate at which a noisy channel can transmit data with negligible error — the “Shannon limit.” After nearly a decade of work, his 1948 paper, “A Mathematical Theory of Communication,” established the discipline of information theory and laid the groundwork for digital communication.

The limits of thermodynamics and information theory have proven robust enough to be codified into laws and theorems that continue to influence their respective fields. But in biology, such limits have been harder to find. Living systems are messy, adaptive, and often seem to violate our expectations. Even if it is challenging to apply Limit Thinking to biology, it’s worth attempting, as it helps biologists home in on the units, models, and general principles that point us toward new ways of thinking.

Two examples help illustrate this.

For decades, enzymologists believed that the speed of a chemical reaction was limited by how fast substrate molecules could find each other by diffusion. For a one-molar solution of enzymes and substrates (an extremely concentrated amount), this diffusion limit was thought to be about 100 million to 1 billion collisions per second.

But in the early 1970s, when biochemists began directly measuring how transcription factors bind to their target sequences on DNA, some appeared to exceed this supposed speed limit. The paradox was that the diffusion limit had been derived for motion in three dimensions, yet DNA confines this search to fewer dimensions and so the limit is actually much higher. It was thus proposed that molecules could restrict the dimensionality of their search and thus achieve effective rates beyond what theory predicted. This discrepancy alerted biochemists to the need to update their models.

And, indeed, they did: By the late 1980s, free-diffusion of a transcription factor to its genomic target was replaced with a combination of one-dimensional sliding along DNA, and occasional hopping between strands. The behaviors predicted by this model were observed directly in 2007 using a fluorescently labelled lac repressor.

Another example that reveals the utility of Limit Thinking for biology is error rates.

In 1972, after a stint at Bell Labs, physicist John J. Hopfield was thinking about the limits of molecular recognition; specifically, how molecules find the correct partner amid a sea of similar ones. It was well known at the time that biology used macromolecules to assemble basic building blocks into complex polymer proteins. This relied on the “genetic code,” which translates a sequence of nucleic acids into a sequence of amino acids. Hopfield recognized that to be a useful code, the assembly process must use physics to preferentially associate the correct amino acid with its correct triplet of nucleic acid bases. His major insight was to ask: “How accurate can such a process possibly be?”

According to classical thermodynamics, the error rate should depend only on the difference in binding energy between correct and incorrect pairs. But when Hopfield did the math, biology turned out to be about a thousand times more accurate than any equilibrium model allowed. Something fundamental was missing.

The only way to beat the equilibrium limit, Hopfield realized, was to drive the system out of equilibrium — to spend energy to enforce correctness. Hopfield proposed a mechanism in which enzymes use chemical energy (from ATP hydrolysis, for example) to make certain steps irreversible, effectively giving molecules a second chance to “proofread” before committing to an error. He called this process “kinetic proofreading.”

This idea marked a turning point. It showed that biological accuracy arises not from perfect, efficient binding but, counter-intuitively, from active, energy-consuming processes that operate out of equilibrium.

Like the heat engine and the information channel, kinetic proofreading is an abstract model. And because it is agnostic to the details of exactly how it is implemented, the same underlying theory has been used to explain not only error correction in protein synthesis, but also proofreading in nucleic acid copying, antigen discrimination in the immune system, and even the dynamics of how microtubules can rapidly reorganize the cytoplasm when necessary.

Kinetic proofreading has also inspired new iterations of Limit Thinking. When it was noticed that the highest accuracy limits are only achieved when the process becomes exceedingly slow and energetically expensive, researchers began to map the limits of the trade-offs between speed, accuracy, and energy use.

Hopfield is well known for his work on proofreading, but is perhaps even better known for his work on artificial neural networks, often called Hopfield networks, for which he was awarded a Nobel prize in 2024. In his Nobel lecture, Hopfield acknowledges the role his work on proofreading played in his thinking about neural networks.

This 1974 paper was important in my approach to biological problems, for it led me to think about the function of the structure of reaction networks in biology, rather than the function of the structure of the molecules themselves. Six years later, I was generalizing this view in thinking about networks of neurons rather than the properties of a single neuron.

The Hopfield neural network, like the biomolecular networks that influenced it, also lent itself to Limit Thinking. Once formalized, it was clear to Hopfield that his associative networks could not store memories densely enough. The theory put a limit on how many memories a given number of neurons could store, but a real biological network could apparently store much more than that. Despite their utility, Hopfield’s model still lacked something essential, and this recognition, based on Limit Thinking, has led to Hopfield’s recent work on dense associative memories.

Ultimately, Limit Thinking is about how to frame problems well. Not every abstraction is useful, but if an abstraction allows you to say something quantitative about the fundamental limits of a system, you’re probably on the right track.

Carnot understood this better than anyone. “The motive power of heat,” he wrote, “is independent of the agents employed to realize it,” and “the quantity of caloric absorbed or relinquished is always the same, whatever may be the nature of the gas chosen.” Those statements freed generations of engineers to tinker with any mechanism they liked, as long as it increased the temperature differential between reservoirs.

It’s easy to forget that, in Carnot’s time, it wasn’t obvious whether such a limit even existed. Carnot notes:

The question has often been raised whether the motive power of heat is unbounded, whether the possible improvements in steam-engines have an assignable limit — a limit which the nature of things will not allow to be passed by any means whatever; or whether, on the contrary, these improvements may be carried on indefinitely.

To define a limit is not to set an arbitrary boundary but to understand the possible bounds of any given system. Carnot’s limit turned heat into a science; Shannon’s turned communication into information; Hopfield’s turned chemistry into computation. Each began with a simple question about what nature would allow and ended with a new theory that progressed technology. Apparently, searching for limits is precisely what removes them from scientific advancement as a whole.

Subscribe now

David Jordan is a researcher and founder of the Living Physics Lab, based in Cambridge, UK, and funded by the UK’s Funded by the Advanced Research + Invention Agency (ARIA) through project code NACB-SE01-P03.

Thanks to Isabelle Zane, Somsubhro Bagchi, and the ARIA Innovator Circles Programme. Thanks to Xander Balwit and Niko McCarty for editing this essay, and Ella Watkins-Dulaney for many fact-checking questions and for making the header image.

Cite: Jordan, David. “The Power of Limit Thinking.” Asimov Press (2025). DOI: https://doi.org/10.62211/63ok-78ht

Further Reading:

1. Rolt, L. T. C., and Allen, John Scott. The Steam Engine of Thomas Newcomen. United Kingdom, Moorland Publishing Company, 1977.

2. How Claude Shannon Invented the Future, Quanta Magazine

Biology can be hard to intuit, in part because it operates across vastly different scales, from single atoms all the way up to entire ecosystems. Students of biology therefore often first meet its agents and mechanisms through metaphors: molecules are charged balls connected by sticks! evolution designs organisms to maximize their fitness! mitochondria are the powerhouses of the cell! While metaphors give us qualitative handles to grasp, they often oversimplify complex ideas.

This is because most metaphors fail to address specifics — especially regarding numbers. Consider another common bio-metaphor: DNA is the blueprint of the cell. That’s useful for conceptual understanding, but how big is this blueprint? Is it as big as a novel or an encyclopedia? How much space does it take up? It’s possible to look up or calculate the answers to these questions; the human genome is 6.2 billion base pairs, which takes up about 10 cubic microns.1 But how big is that compared to the total volume of a cell? Is it most of it or just a tiny fraction?

To answer questions like these, you need more quantitative metaphors. Whereas a standard metaphor says that mitochondria are the powerhouses of the cell, a quantitative metaphor says how big a battery a mitochondrion would be, say, if a cell were a toaster. If qualitative metaphors are like containers, then quantitative metaphors are closer to yardsticks.

And these yardsticks are exceptionally useful down at the scales where biology operates. We all know cells are small. But so are proteins, nucleic acids, and water molecules. We often think of everything “small” as equally small, but that is not the case. Proteins are titanic, hulking machines compared to water molecules, and the mRNA that encodes a protein is orders of magnitude larger still!

To make the sizes and shapes of various biomolecules concrete, let’s imagine that each water molecule within a cell has been blown up to the size of a grain of sand.2 If this were the case, then …

Subscribe now

Proteins

• Proteins come in a variety of shapes and sizes, but a typical protein would now be a knobby ball the size of a blueberry.3

• The largest human protein, titin, would be a floppy spring 3-4 mm in diameter. Coiled loosely, it would be about as long as a golden retriever end-to-end but could stretch out to the height of a double-garage door.

• An antibody would consist of three blobby arms, each roughly the dimension of a grain of basmati rice, connected to a common center by a short chain.

DNA

• A nucleotide would be a paper-thin oblong disk about the width of a poppy seed (1mm, also the width of three grains of sand).

• To make double-stranded DNA, glue two stacks of nucleotides together side-to-side, then twist them so that they make a complete turn every ten bases. You should end up with a 2-stranded braid about as thick as a #2 pencil lead. Real nucleotides are extremely thin but have some space between them; for simplicity, imagine them as tightly-packed stacks of card stock. Short chains are stiff, but they start to get floppy around 5cm (2 inches).

• A typical human chromosome would be a thread of double-stranded DNA about 100,000,000 bases long, which is just long enough to span the English Channel (34 km).4 In its natural context, it’s usually a snarled hairball roughly the size of a bottlenose dolphin. During mitosis, it tightens up into a cow-length cylinder about one foot wide.5

• A typical bacterial chromosome is a thread or circle of double-stranded DNA a couple million bases long, which would have a length comparable to the height of the tallest buildings. It is normally loosely coiled into lots of twisty bundles, like a headphone cable left loose in a bag; in this form, it would be a meter or two across.6

Single-Celled Organisms

• Human viruses come in a variety of shapes and sizes, but most are balls of protein covered in a lipid membrane studded in protruding proteins, between ping-pong ball and soccer ball in size.

• E. coli (a fairly typical bacteria) is a capped cylinder one meter across and two meters long — about the size of a cow and weighing about two metric tons.

• Single-celled eukaryotes would come in a variety of shapes, and range between cow and large town in size, with most between whale-size and city-block-size. Yeasts would tend toward the smaller end, with Saccharomyces cerevisiae being a sphere the size of a medium-weight moving van. The largest amoebae, on the other hand, would be several kilometers across.

Human Cells and Organelles

• A “typical” human cell is a fibroblast, which is the cell that makes up most of skin and connective tissue. This cell likes to attach itself to a surface and spread itself into an irregular shape, which would be about the dimensions of an 5x5 square block of king-bed, standard-sized hotel rooms.

• The nucleus of such a fibroblast, which contains its DNA and transcriptional machinery, would be an oblong disk roughly the size of two hotel rooms. And remember those dolphin-sized human chromosomes? We need to pack 46 of these into the nucleus, or 92 if it’s about to divide (although the cell nucleus expands before cell division). The chromosomes take up only a few percent of the nucleus’s total volume, loosely packed most of the time and, therefore, tending to expand to fill the space.

• The mitochondria are coral-like webs of branching tubes reaching throughout the cell, with widths between that of a human and a cow.

• Various organelles are attached to a dense spiderweb of fibers called the cytoskeleton. These come in a variety of thicknesses, lengths, and consistencies, but most are either:

  • actin fibers, which are tough threads that would be about as thick as a fat nail, or

  • microtubules — stiff, hollow, finger-width tubes.

• The cell contains hundreds of lysosomes (organelles specialized for breaking down food, waste products, and damaged organelles), which would be blobby bags the size of a rat at the smaller end or house cat at the larger one.

• The cell is encased by a lipid membrane about the thickness of a thin porcelain tile (~5mm). It is rough with proteins, and probably7 looks patchy, with many temporary “rafts” of different densities of lipids and proteins. A 2x2cm square of membrane (about the dimensions of a key from a desktop keyboard) contains around 2,000 lipids and 10-15 proteins.

Organs and Body Parts

• A human hair is a towering cylinder of tightly-packed dead cells. Shape and thickness will vary by hair color and texture, but a fairly typical hair would be about as thick as a soccer pitch is long. At shoulder-length, the hair would extend about 250 kilometers, or long enough to cross California (or England) east to west.

• Human blood vessels vary massively in scale. The smallest capillaries would be as thin as five meters (smaller in cross-section than a fibroblast!), just large enough to pass one or two inflatable-pool-sized red blood cells at a time. In contrast, the aorta, where the heart pumps into, would almost extend across Washington D.C., with surfaces a couple of kilometers thick.

• A human eye would be an orb about 25 kilometers (15.5 miles) across, with a wall 0.5-1 kilometers thick. The cornea (the clear part at the front of the eye encompassing the pupil and iris) would be some 11 kilometers across. The retina (the “camera sensor” at the back of the eye) covers about three-quarters of the interior surface of the eye and would be made up of ten distinct layers of cells, each about 7-8 building stories tall.

• A human brain, removed from the body and placed on the ground, would span half the surface area of Belgium and almost reach space.

Animals

• Perhaps the smallest commonly-studied animal is C. elegans, a near-microscopic nematode worm consisting of exactly 959 cells.8 According to our metaphoric yardstick, this nematode would be a tube a kilometer long (about as long as the Golden Gate Bridge) and 40 meters (about 12 stories) high and wide.

• A fruit fly would be a behemoth some 2-3 kilometers from mouthparts to abdomen, and 1-2 kilometers across.9 Each eye would be a rough hemisphere made up of thousands of whale-sized facets, and every other vertex between facets would sport a giant guard-spike some twenty meters long (as long as a semi truck).

• The average human would be about 1,700 kilometers tall. The International Space Station would orbit somewhere around the knee.10 This human could cover the U.S. west coast with their stretched-out arms.

Manmade Things

Just to hammer home how compact biological entities are, let’s look at some human-made artifacts at the same scale:

• A US penny laying flat on the ground would now be a disk a kilometer and a half tall (a bit shy of twice the height of the Burj Khalifa) and nineteen kilometers across (a reasonable day’s walk for an experienced hiker).

• The ball from a fine-tipped ballpoint pen would form a sphere 400 meters tall — as tall as the Empire State Building. This 400-meter height is precise to within 250 millimeters (about 10 inches). It is not a perfect sphere, but it is surprisingly close — its manufacturing imperfections have an average height (or depth) of less than three centimeters (about an inch).

• A transistor would form a rectangular block about two centimeters (¾ inches) across.11 An iPhone processor has 15-20 billion of these blocks, and is roughly 10 kilometers on a side.

Room Scales and Larger

Unfortunately, the sand-as-water-molecule metaphor begins to lose its usefulness when we get to the scale of the spaces in which humans and other animals live. Arguably, it has already lost concreteness by the time you get to whole humans (do you really grasp the length of the U.S. west coast or the orbital height of the International Space Station?).

If you wanted to visualize, say, a hotel room at sand-as-water-molecule scale, you’d be talking about carpet fibers as tall as mountains and areas comparable to continents. And real organisms can operate over much larger distances than the width of a hotel room! Scaled up, the migrations of the monarch butterfly or the arctic tern would be literally astronomical, best measured in units of solar system radii. Traditionally, astronomical distance is where we start using quantitative metaphors to scale down.

This shift in metaphoric type is also useful — the difference between the scales of a cell and the scales of the everyday are like the difference between everyday earthly lengths and the smallest “astronomical” ones. In our next essay on metaphors, we’ll tackle the other dimension of biology: time.

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.

Cite: Clamons, S. “Metaphors for Biology: Sizes.” Asimov Press (2025). https://doi.org/10.62211/27he-92qw

Header image by Ella Watkins-Dulaney.

Correction: This article was updated on 5 November 2025 to address discrepancies in the sizes of chromosomes and nuclei. The original estimates were based on entries reported in BioNumbers, which appear to be flawed. Also, a human cell packs 92 chromosomes into its nucleus prior to cell division, rather than 46.

Growing up, my family’s llamas never did anything useful. There was no need for Andean pack animals on their small acreage farm in Oregon wine country, so the creatures simply idled on my grandparents’ property, spitting at us and occasionally getting stuck in the pond. Looking back, it is hard to imagine that these beasts bear any resemblance to the llamas at the center of a recent study on making better antivenom — yet they do.

Today, researchers reported in Nature how a llama and an alpaca were injected with venoms from 18 of the most deadly snakes on the African continent to make a broad-spectrum antivenom, the product of a vast amount of experimental work, spanning years of effort.1 By isolating antibodies from these animals, expressing more than 3,000 recombinantly in engineered E. coli cells, and combining eight, the resultant antivenom prevented venom-induced death in mice injected with 17 of Africa’s most lethal elapid snake venoms. And while this antivenom has yet to move into phase I trials, it offers greater protection to mice than Inoserp PAN-Africa, a commercial antivenom approved by the WHO that was specifically developed for snakes found in sub-Saharan Africa.2

Although a recent computational approach for a designed antivenom by the Baker group at the Institute for Protein Design has garnered significant attention in the past year, this paper underscores that there is still much value that can be extracted from traditional antibody development techniques. By injecting animals with the full slate of diverse toxins present within venom and then screening thousands of antibodies to identify which ones bind to and neutralize a wider range of damaging toxin subfamilies — not only the three-finger toxins (3FTxs) at the center of Baker’s work but also phospholipase A₂ (PLA₂) and Kunitz-type serine protease inhibitors (KUNs) — it was possible to create a potent antivenom that protects against multiple toxin types.

The paper thus offers a promising solution to what has been dubbed “the antivenom crisis;” a truly broad, manufacturable antivenom cocktail that outperforms the best remedy currently on the market (at least in mice).

The Problem

If ever there were a truly menacing snake family, it would be the elapids, characterized not only by a set of permanently erect frontal fangs but also a threat display involving rearing up and fanning out a neck flap. While elapids are found worldwide in tropical and subtropical regions, they pose an enormous problem in sub-Saharan Africa, where their bites cause 7,000 deaths and 10,000 amputations each year (though these are likely underestimates, because many deaths go unreported).

When biting, these snakes inject their venom through hollow or grooved fangs that work like hypodermic needles. Once inside the body, the venom spreads toxins via the lymphatic system and bloodstream.3 As they circulate, the various foreign proteins present in the venom target specific systems: neurotoxins attack the nervous system, causing paralysis and respiratory failure, hemotoxins destroy blood cells and damage blood vessels, leading to internal bleeding and tissue death, and cytotoxins break down cells at the bite site, causing severe local tissue damage and necrosis. While cowboy Westerns often depict the application of tobacco poultices or vigorous sucking and spitting to draw out the poison, antivenom is the only way to halt this cascade of cellular damage.

Antivenom is typically produced by immunizing horses or sheep with small amounts of snake venom and then harvesting and purifying their antibody-rich serum. When administered intravenously, this serum circulates rapidly throughout the bloodstream, while the antibodies within bind to venom molecules through highly specific protein-protein interactions, forming immune complexes that neutralize the toxins before they can rampage unchecked throughout the body.

The effectiveness of an antivenom depends critically on timing.4 It works best when administered within hours of the bite, before venom molecules have bound to their targets and caused irreversible damage. The problem, however, is that getting to a hospital quickly is often impossible. In a piece on the antivenom crisis from Works in Progress, author Mathias Kirk Bonde contrasts a suburban Australian seeking attention at a nearby clinic with what transpires in poor, rural regions in India:

For over 34 percent of Indian snakebite victims, it takes more than six hours to receive treatment. If the clinic lacks a cold chain — a coordinated system of temperature-controlled environments — it is limited to using antivenom that can survive room-temperature storage. Without equipment to diagnose the species of snake the patient was bitten by, doctors must use polyvalent antivenom … which is typically more expensive and less effective than species-specific monovalent antivenoms.

This excerpt underscores two major challenges facing antivenom: first, the problem of transport (both of the patient to the clinic and the antivenom to temperature-controlled storage), and second, the difficulties surrounding polyvalent antivenom (those effective against multiple species), which are not only more expensive but also less targeted. Because these polyvalent antivenoms contain antibodies against multiple snake venoms, patients must receive a much larger dose of foreign animal proteins than they would with monospecific antivenoms, significantly increasing the risk of adverse immune reactions.

Looming over all of this is an even more essential challenge: venom is a staggeringly complex biological substance. “Snake venom is not just one thing, but rather an umbrella term for a cocktail of proteins,” writes Abhishaike Mahajan in a blog post. “The therapeutic issue here is that the composition of the toxin can vary.” This variation can occur by snake families (elapids versus vipers), snake species (Black mamba versus Cape cobra), snakes of the same species (dictated by things like geographic location), and even within the same species over time (influenced by whether the animal is a juvenile or adult at the time of the bite). Each snake’s venom, in other words, includes a unique blend of harmful proteins.

Such complexity also helps explain why timing is not the only critical factor after someone has been bitten by a snake. Because of species-specificity, it also matters that they receive an antivenom that matches their snake’s particular venom composition. For example, while taipan antivenom (developed against the coastal taipan) effectively neutralizes the toxins of both coastal and inland taipans, it’s less effective against inland taipan venom. Therefore, snake-specific antivenoms (even between subspecies) are essential.

This bevy of challenges, however, has spurred researchers to explore a variety of ways to address the antivenom crisis.

Back in January, for example, David Baker and his team used a computational tool called RFDiffusion to create proteins that bind to, and neutralize, three-finger toxins, small toxic proteins whose name comes from their characteristic molecular structure resembling three fingers extending from a central core. Three-finger toxins bind to the receptors where motor neurons communicate with muscle fibers, thus blocking the electrical signals that allow muscles to move and causing a progressive paralysis that, if untreated, results in respiratory failure; the primary cause of death from elapid bites. When tested in mice, pre- or post-injection of the computer-designed binders completely protected animals from lethal neurotoxin doses, with full survival and no paralysis up to 24 hours after exposure.

This computational paper marked an important step toward solving some of the biggest challenges facing antivenoms. For one, the de novo proteins were smaller and simpler than antibodies, making them easier to manufacture using recombinant DNA technology (and avoiding the batch-to-batch variation that often limits the effectiveness of traditional antivenoms, made from serums.) Their small size also helps them penetrate tissue deeply to more quickly neutralize toxins. And finally, these proteins were explicitly designed to have a high thermal stability, meaning they can withstand higher temperatures without degrading or unfolding. This helps with antivenom transport and storage; they can be kept in warmer environments that may lack an adequate cold chain.

The limitation of Baker lab’s approach, however, is that, while it might help in effectively neutralizing three-finger toxins, it does little to fight the various other components at work within venom. It neglects a particularly sinister family of toxins known as the phospholipase A₂ (PLA₂) family, a group of enzymes that hydrolyze phospholipids in cell membranes, breaking apart the fat molecules that form the protective outer barrier of cells. In many elapid venoms (especially those found in spitting cobras and rinkhals), these PLA₂ toxins are major contributors to local tissue destruction, “dermonecrosis.” When venom doesn’t kill, but leaves the bitten individual maimed or suffering life-altering wounds, PLA₂ is often to blame.

For another study, published in Cell this past June, researchers extended their antivenom neutralization effort to include PLA₂. Here, a group led by Jacob Glanville and Peter Kwong identified two human antibodies recovered from a hyperimmune human donor who had self-immunized with hundreds of venoms that could neutralize toxins shared across many snake species. Structural studies revealed that these antibodies (dubbed “LNX-D09” and “SNX-B03”) block the same receptor-binding surfaces that the computationally-designed antivenom, made by Baker’s group, was targeting. When combined with varespladib, a small-molecule inhibitor that disables PLA₂, the resulting cocktail protected mice from lethal doses of venom from 19 medically important Elapid snakes (though not from the same geographic region).

Although this study provides yet another approach toward the creation of “universal antivenoms,” it, too, had limitations. For one, varespladib has a short half-life (about 5.5 hours in humans). So while it initially neutralized PLA₂, it didn’t remain in the bloodstream long enough to confer sufficient protection. For that, the researchers had to re-dose varespladib every eight hours,5 which is impractical for emergency treatment, especially in low-resource settings where continuous medical supervision isn’t feasible.

Desperately needed, then, is a broad antivenom able to work continuously across all of the deleterious toxic subfamilies and offer long-lasting protection in a medically realistic scenario.

Going Broader

The approach in the Nature study out today aims to provide just that. The researchers began with a process quite similar to traditional antivenom production, milking venom from 18 species of African elapids and injecting each of them into one llama and one alpaca. From here, they increased the dose of these venoms over 60 weeks, administering the animals with larger and larger doses of snake venoms over the span of more than a year. Once inside the camelids’ bodies, these venoms produced an immune response, and the animals began to make antibodies that could be withdrawn and interrogated for the particular traits these researchers sought.

Next, the researchers used a technique called phage display to screen billions of antibody fragments and identify which ones bound best to various toxins within the venoms. Based on the results from the phage display, the researchers took 3,000 different antibodies and expressed each of them recombinantly in E. coli cells. The engineered microbes pumped out each of these antibodies, which were then isolated, one by one, and screened for binding against various toxins in the venoms. They found that 60 percent of the antibodies bound to their target toxin and more than 50 percent bound to multiple toxins from the same toxin subfamily. In other words, the llama and alpaca antibodies proved to be naturally promiscuous; a protein that neutralizes a three-finger toxin is likely to also neutralize other three-finger toxins in that same family.

After running some in vivo tests (involving injecting mice with antibodies that seem to have high cross-reactivity, and also that seem to bind their target toxins exceptionally tightly), they winnowed their antibody candidates down and combined just eight of them into an antivenom cocktail.

From here, the researchers conducted two sets of experiments. First, they did a pre-incubation experiment in which the antivenom cocktail was mixed with various snake venoms, and then the full mixture was injected into mice. Each animal was monitored for 24 hours to see whether they survived. This was not a “medically realistic” experiment, because the antivenom is given with the venoms. But the goal is merely to help establish baseline potency and answer the question of whether the antivenom will neutralize the toxins if it contacts them directly.

The researchers observed no signs of envenoming for 13 out of 18 of the snakes whose venom was injected during the pre-incubation work. For two of the snakes, pre-incubation with the antivenom still led to “minor signs of envenoming, including closed eyes and periods of lethargy interspersed with periods of excessive grooming,” but the animals survived. Two other species had similar late-onset symptoms, but likewise recovered. Overall, minor signs of envenoming were observed in mice injected with venom from 4 out of 18 snakes. The only snake that the cocktail failed to protect against was D. Angusticeps, the Eastern green mamba, a thin neon snake harrowingly nicknamed the “people-chaser.”6 However, even the mice injected with this cocktail survived twice as long as they would if injected with the venom alone (from three hours to six hours of survival).

The next experiments were the “rescue experiments,” so named because the antivenom is administered after exposure to the venom (more realistically what would occur after envenoming). Each mouse was injected with three-times the lethal dose of a given snake venom, and then treated with antivenom five minutes later. Once again, the mice were monitored for 24 hours.

The researchers benchmarked both experiments against the commercially available Inoserp PAN-AFRICA, with the manufacturer’s recommended dose similarly scaled up to contend with three times the median lethal dose of venom. The recombinant antivenom completely prevented lethality for 6 of the snake venoms in the rescue experiments, with no signs of envenoming. The Inoserp PAN-AFRICA antivenom, by contrast, “showed only partial neutralization of all the tested venoms and an extension of time of death for the venom from N. melanoleuca.” In other words, the recombinant antivenom “performed better than the commercial antivenom on all included venoms at the tested doses.

Toward Synthetic, Broad Antivenoms

There are several reasons to think that this new approach might help in addressing the antivenom crisis.

The first is that it may be more easily scalable, at least compared to the conventional production of antivenoms from horses and sheep. This has to do with this antivenom’s use of camelid antibodies. Historically, one of the reasons antivenoms were produced in horses was because they produce conventional antibodies (IgGs) that have a standard Y-shaped structure with two heavy chains and two light chains, resulting in large, complex molecules (approximately 150 kDa) that more closely mirror our own.

Camelids (camels, llamas, alpacas), in contrast, uniquely produce both conventional antibodies and a special type that lack light chains entirely, called “nanobodies.” These nanobodies are only about 15 kDa in size, or roughly one-tenth as large as conventional antibodies. Their small size allows them to reach venom molecules in hard-to-access areas of the body, and it was these nanobodies which were used to make the antivenom cocktail reported in this paper.

Nanobodies are also more stable, soluble, and less likely to trigger immune reactions in patients because of their simpler structure. And, finally, they are easier to manufacture than conventional antibodies because they are made from a single polypeptide chain. This structural simplicity means that they can be manufactured using microbes, rather than mammalian cells, which “might facilitate a lower cost of goods for treatment,” according to the paper.

While this approach already shows great promise, there are ways to optimize it further. This is especially true (and worth restating) because this biochemical approach is complementary to the work being done on AI-designed proteins.

In fact, Baker’s 2025 paper acknowledges that “RFdiffusion could be used to generate designs that neutralize other medically relevant toxins, expediting the formulation of antivenoms with broader species coverage.” When this work is expanded, researchers may be able to readily mix and match, manufacturing each nanobody in E. coli and combining them in desired ratios to confer even broader protection against snake venoms. An eight-protein cocktail, after all, is not the end-all be-all of broad antivenoms; these cocktails could expand to include AI-generated versions that cover that 18th snake, or that resolve some of the other envenoming symptoms observed in the mice in this study.

There is also the question of whether this approach can work elsewhere. While the paper demonstrates protection against African elapid venoms, more work is needed to see whether, for example, this approach could work in India, where 50,000 to 60,000 people a year die from snakebites.

In India, a group of snakes known as “The Big Four” — the Indian cobra, common krait, Russell’s viper, and saw-scaled viper — cause most of the deaths. The big four are split between elapids and vipers, resulting in a different venom composition. Whereas elapids contain the aforementioned toxin families (3FTx, PLA₂, and KUNs), viper venoms are predominantly composed of enzymatic proteins, such as snake venom metalloproteinases (SVMPs), snake venom serine proteases (SVSPs), and L-amino acid oxidases (LAAOs). These proteins attack blood vessels, prevent clotting, and destroy tissue. Protecting against these symptoms would require a different approach.

Finally, there is the longstanding question of who will pay for the development and manufacture of antivenoms, a question that has seen lots of recent handwringing as people consider what it would take to meet the WHO’s plan to halve the number of deaths and disabilities due to snake bites in South-East Asia by 2030. To revisit Bonde:

The field remains poorly funded. Investment into venomics research makes up just two percent of the snakebite budget in the WHO’s 2030 plan. Relying on venture capital instead, start-ups in the field are urged to focus on richer markets with high profit margins rather than where their products would do the most good. This is not to point blame at venture capitalists who are providing much-needed funding…but to emphasize that nonprofit funding in the space could shift priorities for the better.

To return for a minute, to my family’s llamas. When one was stuck in the pond, there was little one could do to encourage it to exit on its own. Instead, we had to wade in and haul it out. The same seems true of better antivenoms. Basic biochemistry or computational modeling might make them broad, but without our working to manufacture them at scale and haul them into the field, they will not arrive there on their own. As venomic researchers issue plaintive calls to philanthropists and discuss funding mechanisms like advanced market commitments, it would be remiss not to emphasize that innovation cannot carry a technology to widespread adoption. Scaling requires both buy-in and investment.

Xander Balwit is editor-in-chief of Asimov Press and recently changed her estimation of llamas.

Cite: Balwit, X. “An Antivenom Cocktail, Made by a Llama.” Asimov Press (2025). https://doi.org/10.62211/29kf-44ty

A huge thanks to Niko McCarty for directing me to this paper, talking me through it, and sending various reading materials in support. Thanks also to Ella Watkins-Dulaney for additional feedback and for creating the header image.

Cells are crowded and frenzied places. But their interiors still contain more order than their surroundings.

Inside a cell, molecular machines convert raw inputs into highly ordered structures; DNA coils into chromosomes and proteins fold into precise three-dimensional shapes. Outside the cell, however, ions drift without gradients, while DNA and proteins exist mostly as scattered building blocks. Only a thin membrane separates the managed chaos inside from the high entropy outside.

If a cell’s membrane is punctured, the fluid within — made mostly of water — drains away, and the cell swiftly dies. And yet, the membrane must let molecules in (food, nutrients, ions) and out (waste, toxins, messages) for the cell to survive. The membrane does so by means of hyperspecific proteins that allow only particular molecules to pass. It is thanks to these atomic-scale filters that life can exist at all.

Subscribe now

What makes some of these filters remarkable is that they work without any “added” energy. Separating water from protons, or potassium from sodium, seems like it ought to require a great deal of resources. And yet, the only energy a cell requires to do so is that used for building the proteins themselves. In other words, evolution has figured out how to perform (nearly) energy-free filtration at a scale that keeps cells alive in an environment that would otherwise overwhelm them.

Two types of protein filters, aquaporin and the potassium ion channel, offer particularly elegant examples. Aquaporins let billions of water molecules pass in and out each second in a single file line, while blocking protons. Potassium channels conduct about 100 million K⁺ ions per second yet decisively reject Na⁺, a slightly smaller ion with the same charge. Both rely on geometry alone (namely, the precise placement of amino acids) to achieve their selectivity. And both demonstrate how seemingly complex problems are often solved in biology simply by positioning the correct atom, with the appropriate charge, at the perfect place.

Aquaporins were first theorized to exist in 1970 by two physiologists at the University of California, Berkeley. These scientists noticed that, when human red blood cells were treated with chemicals that latch onto cysteine, an amino acid found in many membrane-embedded proteins, the blood cells’ permeability dropped by about tenfold. Their finding came at a time when scientists believed that water passively moved into or out of cells through the cell membrane itself. They were the first to suggest, instead, that proteins were, by some as yet unknown mechanism, involved with the movement of water.

The protein responsible was not discovered until 1992 by a team of scientists led by Peter Agre at Johns Hopkins University. While purifying Rh antigens1 from red blood cells, an additional, unknown protein kept turning up. Agre’s team called it CHIP28.2

To figure out its function, the researchers injected frog eggs with RNA molecules encoding the CHIP28 protein. Doing so caused the eggs to swell up and burst, hinting that the mystery protein was some kind of water channel. Follow-up studies, from inhibitor assays in frog eggs to reconstitution in artificial membranes, later confirmed that CHIP28 was a dedicated water channel. In 2000, a high-resolution structure of the protein, renamed aquaporin, was published in Nature.

“The atomic model provides a possible molecular explanation to a longstanding puzzle in physiology,” the authors wrote. Namely, “how membranes can be freely permeable to water but impermeable to protons.”

Early structural imaging revealed that each aquaporin protein is made from six alpha-helices, or looping rods that run back and forth through the cell membrane. These helices form an hourglass shape, such that the channel is cone-shaped at both ends and narrow in the middle. The two ends of aquaporin mirror each other, with their point of symmetry occurring at the narrowest part of the hourglass.3

As water molecules flow from outside to inside the cell, the first obstacle they encounter is the ar/R filter. This is a cluster of four bulky or positively-charged amino acids, including arginine, situated inside the aquaporin channel. The positively charged amino acid forces away sodium ions and loose protons, while the bulky amino acids help to narrow the protein’s center to the width, roughly, of a single water molecule.

Once a water molecule moves past this first filter, it quickly encounters a second, called the NPA motif (short for asparagine–proline–alanine). There are two NPA filters pointing at each other inside the aquaporin.

When a water molecule passes the ar/R filter and moves deeper into the channel, it latches onto the asparagine amino acid in the first NPA motif. But then — almost instantaneously — the local electric field at the channel’s midpoint forces the water molecule to rotate 180° and latch onto the asparagine in the second NPA motif. In other words, these duelling NPA filters force each water molecule to flip around inside the aquaporin, thus breaking them away from any hitchhiking ions. Incredibly, this happens billions of times each second.

There are at least 13 different types of aquaporins in human cells, and nine different aquaporins are found in kidney cells alone. Some aquaporins transport not only water, but also small molecules like glycerol. It is even thought that some aquaporins may move gases — such as CO₂, NH₃, NO, O₂ — across the membrane.

A second kind of atomic-scale filter is the potassium ion channel. These channels, too, span the cell membrane. But instead of transporting water, they admit potassium (K+) ions. Potassium channels are especially important in the brain. After a neuron triggers an action potential, it is these filters that open up and allow potassium to rush into the cell, thus restoring the neuron to its resting potential in preparation for its next firing.

The first clues about potassium channels came, yet again, not from biochemistry but from physiology. In the 1940s and early 1950s, Alan Hodgkin and Andrew Huxley used the squid giant axon as a model system to study how electrical signals travel along nerves. The squid giant axon stretches several centimeters in length, making it large enough to handle the direct insertion of electrodes. With this setup, and their later invention of the voltage-clamp, Hodgkin and Huxley could hold the membrane at a fixed voltage and watch the underlying currents run across the axon.

What they found, over the course of many years of research, was that action potentials in the axon were caused by a fast, transient inward current (sodium ions), followed by a slower outward current (potassium ions). Hodgkin and Huxley’s discovery of the action potential, and its mechanism, earned them a Nobel Prize in 1963.

The two men never figured out the proteins responsible for those potentials, however. That breakthrough came only in the 1980s, when researchers studying fruit flies with a mutant “Shaker” phenotype — flies whose muscles shook under ether anesthesia — traced the defect to a single gene. Cloning and sequencing the Shaker gene revealed that it encoded a voltage-gated potassium channel, thus proving Hodgkin and Huxley’s predictions at the molecular level. The protein structure was finally solved by Roderick MacKinnon’s team at Rockefeller University in 1998.

A single potassium channel funnels in 100 million ions per second while being 10,000-times more permeable to potassium than sodium. This selectivity is remarkable given that these two ions both carry a positive charge and are of similar size.4

The mechanism lies in the potassium channel’s interior, which is lined with oxygen atoms. Normally, potassium floats surrounded by a shell of water molecules, and stripping away those water molecules is energetically costly. But the potassium channel positions its oxygens (via a motif called P-loops) so that K+ ions shed their water and grab onto the protein instead. Na+, being slightly smaller, can’t make contact with all the oxygen molecules encircling the channel at once and so gets filtered out by geometry alone.

This geometry explains the potassium channel’s selectivity, but what explains its speed?

Once again, the answer came from structural biology. Researchers crystallized the potassium channel and then soaked the crystals in rubidium and cesium; two ions that behave similarly to potassium but scatter X-rays more strongly. These resolved structures showed that the inside of the potassium channel, which is 12 Å long, holds as many as four potassium ions at the same time, separated by about 3 Å of space.5 This distance is just close enough for the ions — held in space by oxygens within the channel — to repel one another.

The channel harnesses this repulsive force as an atomic propulsion system. When a new potassium ion enters from above, it nudges the others forward and pushes out the ion at the bottom. This “knock-on” mechanism helps explain how the channel can be both extremely selective and also incredibly fast.

What makes these filters so compelling is that they show how small details add up to enormous precision. In aquaporins and potassium channels, a single amino acid placed just so decides whether molecules pass or not. (This idea feels reminiscent of a J.S. Bach quote regarding the piano: “There’s nothing remarkable about it. All one has to do is hit the right keys at the right time and the instrument plays itself.”) Simply put, selectivity in biology rests on geometry and charge, arranged at the scale of angstroms.6

These filters also show that often, in biology, solutions to complicated problems (like how to separate out individual ions when many others carry the same charge) are often strikingly simple. Cells are bags of vibrating molecules, and they only work if the right molecules reach the appropriate places at the correct times. Much of cell biology lies downstream from this basic idea.

Niko McCarty is a founding editor of Asimov Press.

Illustrations by Ella Watkins-Dulaney.

Cite: McCarty, Niko. “Atomic-Scale Protein Filters.” Asimov Press (2025). https://doi.org/10.62211/82gy-21uq

Nature has evolved a stunning array of biosensors for detecting the physical world.

A single E. coli cell, for example, can precisely sense chemical gradients and “swim” toward or away from them. Some bird species, including robins and warblers, can see magnetic fields using cryptochrome proteins embedded in their eyes to guide them during their annual migration. Bogong moths use photons from distant stars as a compass while soaring 1,000 kilometers across southeast Australia. In other words, organisms can sense not only tastes and smells, but also individual molecules, magnetic fields, and infrared or ultraviolet light.

Humans have long used other creatures’ senses to aid and extend our own, too. As far back as 1,000 BCE, humans employed pigeons to carry messages across cities and kingdoms, taking advantage of their remarkable homing instinct. Dogs’ superior sense of smell is often used to sniff out disease, truffles, contraband, and explosives. And today, the city of Poznań, in Poland, uses just eight mussels to monitor their water quality.1

But increasingly, over the last quarter century, scientists have not only used entire organisms to sense the natural world, but have also taken particular genes from those organisms and adapted them into molecular biosensors. Just as a smoke detector has a sensor that detects particles in the air and a buzzer that then alerts us, all human-made biosensors have two basic components.

The first is the sensor itself — an enzyme, antibody, or engineered cell — that physically recognizes a target, whether a pollutant, virus, or rise in temperature. The second is the transducer, which converts that recognition event into a signal we can perceive, such as a glowing light.

Although bioengineers have adapted hundreds of biosensors from nature, they have been less successful in making better transducers.2 Nearly every biosensor today still relies on a narrow set of outputs (aka “reporters”), such as green fluorescent protein (GFP), luciferase, or colorful pigments. Most transducers can only be seen from close up with a direct line of sight, usually using a microscope. And almost all man-made reporters fail to work inside the body or at a distance. This is because visible light does not penetrate solid materials, such as human skin, and easily “blends in” with other photons in the environment.3

Recently, however, bioengineers have developed transducers that transcend such limitations. To make biosensors that work inside the body, scientists have discovered genetically encoded transducers that can be measured using ultrasound or even MRI machines. And for a recent paper in Nature Biotechnology, scientists have reported — for the first time — a new type of transducer that can even be seen from up to 90 meters away using “hyperspectral” cameras mounted to drones. This new technology makes it feasible to monitor individual molecules, as sensed by engineered bacteria, across entire ecosystems.

Hyperspectral Photos

The first hyperspectral cameras were developed in the early 1980s by NASA scientists, who wanted to capture information about Earth, including mineral deposits and ocean algal blooms, from the air. Unlike conventional cameras, which record just three bands of light (red, green, blue), hyperspectral cameras split incoming light into hundreds of narrow spectral bands, including ultraviolet and near-infrared wavelengths.

Because each type of molecule absorbs and reflects light in a distinct way, the camera can be mounted onto satellites and used to record a full spectrum for every pixel on the ground. In plants, for example, these cameras can quantify shifts in chlorophyll levels because those molecules strongly absorb light in the blue and red regions. For soils, the spectra contain characteristic dips and peaks that correspond to moisture levels.

But the idea that these same cameras could be used to detect bacteria required a leap of imagination. It first came to Chris Voigt, professor of biological engineering at MIT, while touring a military facility, where soldiers explained how hyperspectral drones were being used to spot plastic objects from the sky. Foreign militaries sometimes hide explosives or sensors inside plastic casings and disguise them as rocks, but because real rocks reflect light differently than plastic dupes, hyperspectral cameras can distinguish between them.

If the military can distinguish plastic from rock, Voigt wondered, why not microbes from soil?

The work to answer this question fell to Yonatan Chemla and Itai Levin, a postdoctoral fellow and graduate student in Voigt’s laboratory. Their first challenge was to find molecules that cells make that could produce a distinctive hyperspectral fingerprint visible from a distance. So the duo began by searching through hundreds of thousands of metabolites listed in scientific databases, finding that only about 100 have any recorded absorption spectra.

Upon realizing that we don’t understand how the overwhelming majority of biomolecules reflect light, Chemla and Levin decided to investigate themselves. They bought a hyperspectral camera and a large number of purified molecules from online chemical suppliers — such as indigo and porphyrins — and started testing them in the laboratory. They sprayed these molecules onto soils or rocks, took pictures, and then tried to work out which ones produced a clear signal against background noise.

The duo also used computational tools to identify candidate molecules that might act as hyperspectral reporters. Together with collaborators at MIT, they ran quantum chemistry simulations on a selection of 20,000 metabolites to predict how each one would respond to light. These simulations calculated which wavelengths of light each chemical would absorb, and how strong those peaks would be.4 After running these computational tests, Chemla and Levin filtered this list down to a few hundred with unusual peaks or that absorbed light in parts of the spectrum where biology is usually quiet, especially near-infrared wavelengths.

Finally, they considered which of these molecules would be easiest and most efficient for a microbe to make, favoring ones that could be made by slightly altering natural pathways or requiring the addition of only a few recombinant genes. Since microbes can have very different metabolisms, they also weighed which hosts would be the best for each possible molecule. After this winnowing process, they ended up with just two: biliverdin IXα made by Pseudomonas putida, and bacteriochlorophyll a made by Rhodocyclus gelatinosus.

Biliverdin IXα is a green pigment that naturally forms when heme, the molecule carrying oxygen in red blood cells, is broken down and recycled. To make it in P. putida, the team only needed to add two enzymes. Bacteriochlorophyll a, on the other hand, is a photosynthetic pigment found in purple bacteria.5 R. gelatinosus is itself a purple bacterium, meaning that all the team needed to do was amend its existing genome to produce much larger quantities of bacteriochlorophyll a.

With these two engineered microbial strains in hand, the researchers traveled to Fort Devens in Massachusetts — alongside two undergraduate students, Anna Johnson and Yueyang Fan — and sprayed the cells onto little patches of soil. They flew a hyperspectral drone overhead and took pictures of one acre, or about 4,000 square meters, across the entire military facility. Using a hyperspectral detection algorithm that separated the molecular signal from background “noise” of soil and dirt, Chemla and Levin could clearly identify the engineered microbes from up to 90 meters away.6

Alas, the cells were layered on top of sand, in direct line of sight to the camera. But in many cases, the things we want to sense — like explosives or pathogens invading plant roots — are hidden underground. Chemla is now searching for volatile molecules that diffuse upward through the soil and into the air, creating a spectral signature that a camera can detect from high above (possibly even from outer space.)

Environmental Release

Despite this scientific breakthrough, it will be difficult to move these biosensors into the real world. Researchers have been testing engineered microbes in field trials for the last four decades, but few have been commercialized.

Field trials for genetically-engineered microbes peaked in the early 1990s but have fallen off since then, mainly due to increased regulations and mixed field trial results. In the late 1980s, engineered Agrobacterium radiobacter K1026 was approved in both Australia and the U.S. to fight crown gall disease in trees. (The microbe outcompetes disease-causing bacteria, killing them.)

But getting approval to release a microbe into the wild, without containment, can be incredibly arduous. The regulatory pathway is divided across the EPA, USDA, and FDA. Each agency has jurisdiction depending on the intended use; pesticides fall to the EPA, other agricultural products go to the USDA, and ingestible microbes fall under the province of the FDA. Anything that does not easily fit into these categories, including environmental biosensors, is lumped under the EPA’s Toxic Substances Control Act, or TSCA.

The TSCA regulates genetically engineered microbes based on their method of engineering, rather than the product itself. This practice is outdated and should be revised, Chemla says. Any microbe containing DNA from another genus — say, moving a gene from Escherichia coli into Pseudomonas putida — is flagged by the TSCA and unlikely to get approval, even if researchers can prove that the product is safe. More than 200 TSCA submissions were filed between 1987 and 2018, but none of those submissions have led to a commercialized product.

There are ways to skirt these regulations, though. Pivot Bio sells genetically-engineered microbes that colonize plant roots and convert atmospheric nitrogen (N₂) into ammonia (NH₃), a chemical form that plants can use. This reduces the amount of fertilizer needed for a field, thus decreasing the leaching of fertilizer byproducts into water.7

Pivot Bio sidestepped some regulatory hurdles by avoiding the transfer of genes from one species to another; they simply remodeled their organism’s existing genome. The company still must get USDA approval to ship its product across state lines, but that is a simpler and less insurmountable regulatory hurdle.

In the case of hyperspectral reporters, there may be similar ways to circumvent the most onerous regulations. Even in this study, the R. gelatinosus strain engineered to make bacteriochlorophyll a did not have any DNA from foreign microbes. It could, in principle, sidestep the TSCA regulations. A startup called Fieldstone Bio has spun out from the Voigt laboratory with the goal of commercializing this hyperspectral technology.

Regardless, the barrier to commercializing these biosensors is not scientific feasibility but rather a patchwork of rules written long before anyone imagined microbes capable of broadcasting messages into space.

Still, it’s promising to see that synthetic biology is moving past its reliance on visible light toward a broader range of transducers that let us measure biology in places once thought inaccessible, from the molecules inside a tumor to antibiotic resistance genes hidden in soil. The challenge ahead is not discovering what cells can sense, but engineering more reliable ways for them to communicate those impressions back to us.

Niko McCarty is a founding editor of Asimov Press.

Thanks to Xander Balwit and Ella Watkins-Dulaney for reading drafts of this.

Cite: McCarty, Niko. “Seeing Microbes from the Sky.” Asimov Press (2025). https://doi.org/10.62211/23jr-64kt