I have spent a lot of time recently thinking about why biology is hard. The goal of biomedical research is nominally to cure disease. However, most people don’t seem to be very serious about it. When a biotech startup says it’s out to cure a disease, it usually means it wants to find an intervention that will increase lifespan by 3 months. Success usually means an IPO, regardless of whether the drug eventually succeeds or fails. When the Chan Zuckerberg Initiative says they want to cure all disease by 2100, people mostly chuckle at the naïveté. By contrast, when OpenAI said they want to make AGI, people mostly took them seriously, even before the recent developments in LLMs, and the question was whether it would happen in 2030 or 2050, not whether it would happen by 2100. What gives?
If we want to have AI-style success in curing disease, we need three things: speed, knowledge, and talent.
Speed
I always tell my students that the primary factor that determines whether their project will succeed or fail is the rate at which they can do experiments. If you can do experiments once a day, you will learn things 5 times faster than if you can do experiments once a week. The goal of biomedical research is to cure human disease. In the biomedical industry at the moment, it takes about 10 or 20 years to conduct an experiment in which you try to cure human disease. We are still trying to support or disprove the amyloid hypothesis for Alzheimer’s disease, which dates from the 1990s.
The issue is essentially regulatory. Just compare how easy it is to cure a disease in mice to our ability to come up for cures for the same diseases in humans. You can just try things in mice and see what happens — most of your hypotheses don’t work, but sometimes you get lucky and stumble upon a cure. For good reasons, however, you can’t “just try things” in human beings. You need to know that whatever you’re putting into the human is safe, and you need to have some reason to believe they will work. But the result is that it takes 100 times longer to try a new idea in humans than it does to try the same idea in mice. And, for most diseases, there’s really no substitute for trying an idea in a human being. Mice don’t get Alzheimer’s disease, per se — any headline you see about a cure for Alzheimer’s disease in mice is just hype. Serious drug developers have long since learned not to trust animal models when it comes to predicting the efficacy of a treatment for most diseases. (This should NOT be construed as an argument that animal research is unimportant in general; animal models are essential for many aspects of biomedical research, just not usually for drug efficacy.)
Our progress towards curing disease will continue to proceed at a snail’s pace unless we can figure out an ethical way to do experiments faster. Even if we had a super AI that could predict drug efficacy, it would still take us 20 years to pull its predictions through clinical trials, and it would probably need to go through several iterations. I am also skeptical of the ability of even an AI trained on the entire scientific literature to predict drug efficacy for diseases for which have no effective drugs and no understanding of how they work. ML may help us design better experiments and interpret those experiments better, but those experiments will still need to be done, and that will take time.
Solving this problem will require both technical and regulatory work. The best idea I can see at the moment is to use machine learning to predict drug safety, and then work with regulators to enable drugs that are predicted with high confidence to be safe, bioavailable, and to have good PK/PD etc. to be tested directly in human patients without other preclinical trials.
Knowledge
There is a phenomenon that all biologists will be aware of, where after working on a new idea for 2 years, you one day come across a paper from 2008 and say, “oh my god, if only I had known this two years ago.” If we want biology to move fast, we need to figure out how to eliminate this phenomenon.
The biomedical literature is vast and suffers from three problems: it does not lend itself to summarization in textbooks; it is unreliable by commission; and it is unreliable by omission. The first problem is simple: biology is too diverse. Every disease, every gene, every organism, and every cell type is its own grand challenge. The second problem is trickier — some things in the literature are simply wrong, made up by trainees or professors who were desperate to publish rather than perish. But it is the third problem that is really pernicious: many things in the literature are uninterpretable or misleading due to the omission of key details by the authors, intentional or otherwise. Authors may report a new, general strategy for targeting nanoparticles to cells expressing specific receptor proteins and show that it works for HER2 and EGFR, while declining to mention that it does not work for any one of the 20 other receptors they tried. Other times, a lab may decline to mention that their method only works above 50% relative humidity, because they never realized that was an issue. The unreliability of he literature by commission means that fundamentally, if you want to know that an experiment really works, you have to try it yourself. The unreliability by omission means that if you want to understand the real limitations of a technique, you usually have to work on it for 6 months to figure out how it really works and why.
How much progress we can make in biology in the next hundred years will depend on the extent to which the language models are able to solve these problems. The primary questions here will be: what fraction of knowledge in the world can be generalized from knowledge already in the literature, and how valuable is literature synthesis when a large portion of it is incorrect? The problem of summarization will be solved by language models soon. Within a few years at most (and maybe in a few months), every lab will have immediate access to the world’s expert in all of biology, which will happily educate them about the state of the art and the relevant subtleties in whatever field they choose. Simply tell the language model exactly what you want to do, and it will summarize for you everything relevant that is known, thereby avoiding the “if only I had known this” problem. It seems to me that the unreliability of the literature, however, will mean that for the foreseeable future (essentially until we have fully parameterized lab automation), AIs will be better at suggesting experiments and interpreting results than they will be at drawing conclusions from existing literature.
A note on lab automation: chemistry may be all heat, weighing, and mixing, but keep in mind that we recently discovered after 6 months of investigation that one of the projects in my lab was failing because a chemical that was leaching out of a rubber gasket in one of our cell culture wells was interfering with an enzymatic reaction. Full parameterization means, the robot needs to be able to pick up the Petri dish and notice that there is some oily residue floating on the surface that isn’t supposed to be there. Until then, humans will probably stay significantly in the loop.
Talent
When I talk to people in AI/ML, it seems like the smartest people usually want to go work at DeepMind, OpenAI, FAIR etc. where they can continue to do research themselves on the biggest and most exciting problems in the field. In biology, until recently, it seemed like everyone wanted to be a professor or start a company, i.e., that the only high status thing you could do after your PhD was to become a manager. (Including for people who were really better as researchers than as managers.) If we want to make progress in biology, we need a high-status way for top biology researchers to keep doing research, in the same way that DeepMind, OpenAI, FAIR, etc. have created a new high-status way for AI researchers to keep doing research. We need to create a new career path.
My guess is that this new career path will require some new kind of institution. Academia is not built to retain top talent; on the contrary, it is explicitly built to train people and then kick them out once they know what they’re doing. Academia also has the wrong approach to teamwork: academia works on a basis of name recognition, and it’s harder to accumulate name recognition if you have to share paper authorship with many people, so the incentive is almost always to keep team sizes small even if the work suffers as a result. I think people usually don’t want to stay in environments that discourage them from working with their colleagues. At the same time, though, the for-profit biotech universe is ultra-competitive, way more competitive than the FAANG universe, and doesn’t lend itself to industrial labs like OpenAI, DeepMind, or FAIR. My guess is that we need to see more institutions in the space that Adam Marblestone and I have been exploring with the FRO model: dedicated non-academic non-profit research organizations. More thoughts on this to come.
I suspect that the Knowledge and Talent problems are pretty tractable. I also suspect that the Regulatory problem may actually be more tractable than you might think, but that it’s fundamentally technical at the core: can we come up with a way of making drugs that is guaranteed to be safe? Then we might get somewhere.