4 ms·

I'm a noob on this topic, but I think drug discovery is more amenable to this structurally than other problems. Simulating biology is what we were doing with pr

by mindwok 2mo ago

I'm a noob on this topic, but I think drug discovery is more amenable to this structurally than other problems. Simulating biology is what we were doing with protein folding before Alphafold, and the search space was far too large to find stuff in reasonable timeframes. Alphafold showed that you could take a physical process and make a neural net clever enough to learn just enough structure that it starts finding things we might care about, and still physically accurate, much faster. Drug discovery is similar AFAIK. The space of possibilities is even larger than protein folding, but it's structurally similar enough that I think AI will help to make progress on the discovery side. Actually getting the drug tested and approved is another matter though for sure.

tsimionescu 2mo ago
There are two relevant problems here, I think. For one, the question for Anthropic is whether LLMs, specifically, not AI techniques more generally, can help significantly with cancer research. And here, all experience so far is that LLMs only really work when they can easily automatically verify their own outputs and self correct - such as in math (using automatic proof verifiers) or programming (using compilers and unit tests). The second problem is that biological research speed is highly dependent on slow biological processes, such as cultures and long term studies. In programming, if an LLM could provide excellent insights and research suggestions 100x faster than a human, it would speed up the work roughly 100x. But in biology, it would only speed up the total work by a small amount - as any insight, even if absolutely brilliant and spot on, would still require months and years of actual experimentation.
- epihelix 2mo ago
  For drug discovery, you can simulate interactions in silico, driven in an agentic loop via LLM. Mostly stil using non-LLM tools, of course, but it saves time. If I was Dario, and I was after a cynical sales grab to go with the "actually curing cancer!" spiel, I'd probably aim for a drug repurposing strategy. I'd use the above approach to find existing drugs that might target known pathways that drive incurable cancers. I'd only screen compounds with extensive safety data and easy delivery mechanisms, and from the in silico hits, I'd throw a tonne of money at rapidly experimentally screening all those candidates in parallel, and then rapidly push those that worked into clinical trials. I'd assume that, with a small but non-negligible prior, and the money to push through hundreds and hundreds of candidate compounds through at once, I'd have a reasonable chance of getting one drug through to a "Claude Cured Cancer!" show-stopper headline. But I think even then, with all of Anthropic's wealth, you'd need 2 years minimum to move from initial screening targets to a Phase 3 trial. And it would be an almost criminal waste of research funding to get there -- the dollars for discoveries ratio would be appallingly bad. (If I was Dario, and I actually wanted to cure some forms of cancer? I'd just use my obscene profits to fund actual cancer research, step back, and let the researchers get on with it.)
  - disgruntledphd2 2mo ago
    This plan would still take 3-5 years, even in the best case scenario. Clinical trials are really hard.
  - chasd00 2mo ago
    The headline “AI discovered cure for cancer headed to trials!” is all he needs.
disgruntledphd2 2mo ago
> Drug discovery is similar AFAIK. The space of possibilities is even larger than protein folding, but it's structurally similar enough that I think AI will help to make progress on the discovery side. I do agree that this kind of targeted approach makes sense. However, discovery is not really the issue here. Running the clinical trials (1/2/3) is much much more difficult, and consumes basically all of the time in drug development, so even if LLMs perfectly automate this, the speedup will not be particularly large.
fizzbarnull 2mo ago
> Simulating biology is what we were doing with protein folding before Alphafold This is not correct, Alphafold predicts crystal structures of proteins. It is a known issue in the field that these folding models do not generate ensembles of protein conformations like the ensembles generated by DE Shaw Research (who built a super computer to simulate proteins). And these folding models fail when we don't have crystal structures, so they are certainly not generalizing. A great paper on the subject: https://www.biorxiv.org/content/10.1101/2025.02.03.636309v1 https://www.biorxiv.org/content/10.1101/2025.02.03.636309v1 Drug discovery is significantly more complicated than protein folding. Small molecules with similar chemistries can adopt novel binding poses, can end up binding to off targets (PXR, hERG, etc) or simply never make it to the protein due to solubility or permeability. I will always defer to Pat Walters who seems like the most careful and sane person in this field: https://patwalters.github.io/ https://patwalters.github.io/

by mindwok 2mo ago

tsimionescu 2mo ago

There are two relevant problems here, I think. For one, the question for Anthropic is whether LLMs, specifically, not AI techniques more generally, can help significantly with cancer research. And here, all experience so far is that LLMs only really work when they can easily automatically verify their own outputs and self correct - such as in math (using automatic proof verifiers) or programming (using compilers and unit tests). The second problem is that biological research speed is highly dependent on slow biological processes, such as cultures and long term studies. In programming, if an LLM could provide excellent insights and research suggestions 100x faster than a human, it would speed up the work roughly 100x. But in biology, it would only speed up the total work by a small amount - as any insight, even if absolutely brilliant and spot on, would still require months and years of actual experimentation.

epihelix 2mo ago
For drug discovery, you can simulate interactions in silico, driven in an agentic loop via LLM. Mostly stil using non-LLM tools, of course, but it saves time. If I was Dario, and I was after a cynical sales grab to go with the "actually curing cancer!" spiel, I'd probably aim for a drug repurposing strategy. I'd use the above approach to find existing drugs that might target known pathways that drive incurable cancers. I'd only screen compounds with extensive safety data and easy delivery mechanisms, and from the in silico hits, I'd throw a tonne of money at rapidly experimentally screening all those candidates in parallel, and then rapidly push those that worked into clinical trials. I'd assume that, with a small but non-negligible prior, and the money to push through hundreds and hundreds of candidate compounds through at once, I'd have a reasonable chance of getting one drug through to a "Claude Cured Cancer!" show-stopper headline. But I think even then, with all of Anthropic's wealth, you'd need 2 years minimum to move from initial screening targets to a Phase 3 trial. And it would be an almost criminal waste of research funding to get there -- the dollars for discoveries ratio would be appallingly bad. (If I was Dario, and I actually wanted to cure some forms of cancer? I'd just use my obscene profits to fund actual cancer research, step back, and let the researchers get on with it.)
- disgruntledphd2 2mo ago
  This plan would still take 3-5 years, even in the best case scenario. Clinical trials are really hard.
- chasd00 2mo ago
  The headline “AI discovered cure for cancer headed to trials!” is all he needs.

disgruntledphd2 2mo ago

> Drug discovery is similar AFAIK. The space of possibilities is even larger than protein folding, but it's structurally similar enough that I think AI will help to make progress on the discovery side. I do agree that this kind of targeted approach makes sense. However, discovery is not really the issue here. Running the clinical trials (1/2/3) is much much more difficult, and consumes basically all of the time in drug development, so even if LLMs perfectly automate this, the speedup will not be particularly large.

fizzbarnull 2mo ago

> Simulating biology is what we were doing with protein folding before Alphafold This is not correct, Alphafold predicts crystal structures of proteins. It is a known issue in the field that these folding models do not generate ensembles of protein conformations like the ensembles generated by DE Shaw Research (who built a super computer to simulate proteins). And these folding models fail when we don't have crystal structures, so they are certainly not generalizing. A great paper on the subject: https://www.biorxiv.org/content/10.1101/2025.02.03.636309v1 https://www.biorxiv.org/content/10.1101/2025.02.03.636309v1 Drug discovery is significantly more complicated than protein folding. Small molecules with similar chemistries can adopt novel binding poses, can end up binding to off targets (PXR, hERG, etc) or simply never make it to the protein due to solubility or permeability. I will always defer to Pat Walters who seems like the most careful and sane person in this field: https://patwalters.github.io/ https://patwalters.github.io/