points by elcritch 1 day ago | 0 comments
> No one knows how a molecule will interact with human biology. It can't be simulated computationally to generate data (too slow),
Prior to AlphaFold that was the state of predicting protein structures. That model seemingly figured out the "language" used by biology for making protein structures from DNA. It's not unreasonable to believe there's a good chance something similar exists for protein interactions.
Even if it was coarse grained but could eliminate drugs with high likelihood of unfavorable reactions it'd still be valuable. Still not sure LLMs would be ideal for that work.
[–] danielmarkbruce 1 day ago
[–] elcritch 14 hours ago
However the majority of a protein’s structure is functionally inactive. Only a relatively small region tends to be active and affect rate kinetics of various chemical reactions. Alternatively similar interactions affect the mechanical properties of other proteins.
We also know many of the pathways for most key biological processes. Of course we’re no where near simulating an entire human, however we could feasibly filter drug candidate drugs for strong affects on key biological pathways.
That’s not too different in a broad since than using surrogates like mice to test for toxicity across highly conserved pathways in metabolism, reproduction, liver function, etc. It’s a much more tractable problem as well.
We have a lot of experimental data for the effect of many proteins on important chemical pathways. Presumably drug companies have more information on how many drug candidates react with or interfere with key pathways.
My conjecture would be that we wouldn’t need fine grained interactions but that even coarse grained filtering could be valuable. As in sure drug XYZ is great for binding to serotonin but it also has high affinity to key stomach proteins. That could avoid expensive studies just to find out a promising anti-depression drug causes severe gastrointestinal issues in a lot of patients.
Even better if we could use genome variations to protein structure variations which could help differentiate between subsets of patients with or without severe side effects.
[–] danielmarkbruce 5 hours ago
Honestly once you go through all the details... drug discovery feels like the one of the overhyped and unlikely to work areas for AI. Other areas of medicine (like diagnosis) are already seeing pretty good results, and anecdotally seems to work well and tends to fit within the realm of text analysis where most of the AI gains have happened.
[–] elcritch 3 hours ago
Another bigger impact might come by making regular researchers, lab workers, etc all more efficient by making data analysis more efficient or creating better tools. Possibly with less errors. That'd of course require big pharma to adopt them. We'll see on that point.