Clinical trials are engines for scientific discovery. Better drugs require not just more trials, but also improved data collection, to create therapeutic feedback loops.
Part of the problem, really a large part is lack of funding. I moved our entire genetics lab to Denmark because they offered support and funding. We will turn out new CAR-T studies and applications in the future, just not in the US.
Great article! Interesting that you think that phase II is the highest impact phase for this kind of work in your field. I’ve done something similar for infectious disease clinical trials (https://www.nature.com/articles/s41467-019-10092-5), using pathogen genome sequencing. We didn’t find any “hits” for the clinical trial outcome, but is a relatively low cost high value extension of the clinical trial IMHO. But need the numbers of a phase III trial to do that kind of work probably.
According to IQVIA, trials are increasingly complex (profusion of endpoints, inclusion/exclusion criteria etc) and therefore costly. As you’ve pointed out there is significant informational value “even” in failed trials, as these can form the basis of subsequent success. There is clearly some trade-off between the need to both expedite clinical trials and to run information rich experiments and I don’t know where the optimum setting lies.
My own experience is that often companies run trials using almost the exact same criteria as their competitors in an attempt to derisk the asset. Not an optimal solution but an understandable one.
One thing this essay made me think about is a gap that feels more sociological than technical.
Outside formal biomedical research, there’s a huge community of patients, biohackers, and amateur researchers who are deeply invested in medical progress. They generate observations, ideas, and informal experiments at scale. A lot of it is messy, anecdotal, and unregulated—but there’s often real signal and urgency there that never seems to make it into formal research pipelines.
Inside the system, professional investigators and trialists are just as motivated, but they work in an environment shaped by regulation, institutional risk, and slow incentives. Over time, that seems to produce a kind of learned immobility, where even obviously suboptimal processes feel hard to change.
What’s striking is how little structured interaction there is between these two worlds. The outside community has energy and numbers but lacks legitimacy; the inside community has rigor and legitimacy but limited flexibility. Bridging that gap—without compromising safety—feels like it could create both practical momentum and broader support for the kind of clinical abundance you’re arguing for.
I’m curious whether you see any concrete ways to do this. Are there institutional or technical tools that could make patient or enthusiast input more usable for professional researchers—and make it easier for investigators to take that energy seriously within credible frameworks?
I also wonder if other fields have solved versions of this problem. In cooking, radio, art, or software, amateur communities and professionals often coexist through clear interfaces—public work, sandboxed experimentation, reputational filtering—without lowering standards. Biomedicine feels unusually closed by comparison. Do you see realistic ways to build similar interfaces here, so that patient and enthusiast energy becomes an input to research rather than something that lives entirely outside the system?
So, I strongly agree with this article's thesis that trials aren't just a binary pass/fail pipeline, but instead are central to figuring out how a disease works.
However, the example in the article only demonstrates one type of loop: how to program T cells. This is very "narrow" learning in that it's "just" refinement of a product until it works. But there are other types of lessons that can be drawn from trials!
For example, HDL trials kept failing in improving cardiovascular outcomes, even though they were massively raising HDL ('good cholesterol'). However, researchers eventually found a minor improved outcomes and when they looked into the data, this was entirely explained by the 'side effect' of lowering LDL. So now LDL-lowering drugs are about to enter the market.
Another example is SGLT2 inhibitors: meant to reduce excess glucose in diabetes patients, but showed benefits in heart health and kidney health, teaching us about the link between heart, kidneys, and metabolism. The literature is full of examples like these where a drug had unintended but positive side effects.
In sepsis, many failed trials that tried to reduce inflammation created enough data (of high quality, as patients were intensively monitored in the ICUs) to figure out that there were contradictory subtypes. Some patients needed immunosuppression, but some needed immune stimulation!
And then there's the first-in-class learning. These are risky projects for pharma because of a high risk of failure + the first drug often gets quickly overtaken by a more optimized fast follower. But once 1 drug shows that the concept works, many follow. Examples are Hep C antivirals, oncogene addiction drugs, GLP1 antagonists for weight loss.
For all of these, faster, cheaper trials allow more experiments and serendipity. And in the LDL and sepsis cases, many trials were needed to eventually get enough data to spot a signal. I'm also thinking that first-in-class drugs should perhaps be rewarded more?
Lastly, I wonder how many insights remain hidden in proprietary datasets from "failed" trials, and how we can incentivize the sharing of at least utilisation of them 🤔
---
Note: I'm not an expert and relied on AI for most of these examples.
Wonderful article, thanks for writing this. The notion that real-world problem-and-technology-rich use environments can be "engines" for learning and scientific discovery is super under appreciated. I like to call it "reverse translation" and "true" Pasteur's quadrant research (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5600892). I'm more familiar with the physical sciences, so know how important it is there. It makes sense that it is just as important in the biomedical sciences. Also thanks to Smrithi Sunil for calling attention to this nice article!
Part of the problem, really a large part is lack of funding. I moved our entire genetics lab to Denmark because they offered support and funding. We will turn out new CAR-T studies and applications in the future, just not in the US.
Great article! Interesting that you think that phase II is the highest impact phase for this kind of work in your field. I’ve done something similar for infectious disease clinical trials (https://www.nature.com/articles/s41467-019-10092-5), using pathogen genome sequencing. We didn’t find any “hits” for the clinical trial outcome, but is a relatively low cost high value extension of the clinical trial IMHO. But need the numbers of a phase III trial to do that kind of work probably.
According to IQVIA, trials are increasingly complex (profusion of endpoints, inclusion/exclusion criteria etc) and therefore costly. As you’ve pointed out there is significant informational value “even” in failed trials, as these can form the basis of subsequent success. There is clearly some trade-off between the need to both expedite clinical trials and to run information rich experiments and I don’t know where the optimum setting lies.
My own experience is that often companies run trials using almost the exact same criteria as their competitors in an attempt to derisk the asset. Not an optimal solution but an understandable one.
One thing this essay made me think about is a gap that feels more sociological than technical.
Outside formal biomedical research, there’s a huge community of patients, biohackers, and amateur researchers who are deeply invested in medical progress. They generate observations, ideas, and informal experiments at scale. A lot of it is messy, anecdotal, and unregulated—but there’s often real signal and urgency there that never seems to make it into formal research pipelines.
Inside the system, professional investigators and trialists are just as motivated, but they work in an environment shaped by regulation, institutional risk, and slow incentives. Over time, that seems to produce a kind of learned immobility, where even obviously suboptimal processes feel hard to change.
What’s striking is how little structured interaction there is between these two worlds. The outside community has energy and numbers but lacks legitimacy; the inside community has rigor and legitimacy but limited flexibility. Bridging that gap—without compromising safety—feels like it could create both practical momentum and broader support for the kind of clinical abundance you’re arguing for.
I’m curious whether you see any concrete ways to do this. Are there institutional or technical tools that could make patient or enthusiast input more usable for professional researchers—and make it easier for investigators to take that energy seriously within credible frameworks?
I also wonder if other fields have solved versions of this problem. In cooking, radio, art, or software, amateur communities and professionals often coexist through clear interfaces—public work, sandboxed experimentation, reputational filtering—without lowering standards. Biomedicine feels unusually closed by comparison. Do you see realistic ways to build similar interfaces here, so that patient and enthusiast energy becomes an input to research rather than something that lives entirely outside the system?
So, I strongly agree with this article's thesis that trials aren't just a binary pass/fail pipeline, but instead are central to figuring out how a disease works.
However, the example in the article only demonstrates one type of loop: how to program T cells. This is very "narrow" learning in that it's "just" refinement of a product until it works. But there are other types of lessons that can be drawn from trials!
For example, HDL trials kept failing in improving cardiovascular outcomes, even though they were massively raising HDL ('good cholesterol'). However, researchers eventually found a minor improved outcomes and when they looked into the data, this was entirely explained by the 'side effect' of lowering LDL. So now LDL-lowering drugs are about to enter the market.
Another example is SGLT2 inhibitors: meant to reduce excess glucose in diabetes patients, but showed benefits in heart health and kidney health, teaching us about the link between heart, kidneys, and metabolism. The literature is full of examples like these where a drug had unintended but positive side effects.
In sepsis, many failed trials that tried to reduce inflammation created enough data (of high quality, as patients were intensively monitored in the ICUs) to figure out that there were contradictory subtypes. Some patients needed immunosuppression, but some needed immune stimulation!
And then there's the first-in-class learning. These are risky projects for pharma because of a high risk of failure + the first drug often gets quickly overtaken by a more optimized fast follower. But once 1 drug shows that the concept works, many follow. Examples are Hep C antivirals, oncogene addiction drugs, GLP1 antagonists for weight loss.
For all of these, faster, cheaper trials allow more experiments and serendipity. And in the LDL and sepsis cases, many trials were needed to eventually get enough data to spot a signal. I'm also thinking that first-in-class drugs should perhaps be rewarded more?
Lastly, I wonder how many insights remain hidden in proprietary datasets from "failed" trials, and how we can incentivize the sharing of at least utilisation of them 🤔
---
Note: I'm not an expert and relied on AI for most of these examples.
Wonderful article, thanks for writing this. The notion that real-world problem-and-technology-rich use environments can be "engines" for learning and scientific discovery is super under appreciated. I like to call it "reverse translation" and "true" Pasteur's quadrant research (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5600892). I'm more familiar with the physical sciences, so know how important it is there. It makes sense that it is just as important in the biomedical sciences. Also thanks to Smrithi Sunil for calling attention to this nice article!