Hacker News new | past | comments | ask | show | jobs | submit
steveBK123 | on: Anthropic IPO
My understanding of drug dev is that finding stuff that might work is not the challenge, but verifying how well and with what side effects is.

That verification might start in some simulated computer environment but needs to move into physical wet lab, animal testing and then human testing.

So medicine is quite a bit different than code/math where you have a perfectly verifiable problem space that can be iterated within by AI infinitely.

A good example might be Alzheimer's where the brain plaque was thought to be a cause, lots of research on various ways to prevent or reduce the plaque and then it turns out that doesn't matter at all, or is more a byproduct than a cause of the disease.

edit: In general we have many diseases we don't understand and drugs that work via mechanisms we don't understand. So it seems very hard to have a computer simulation in which AI can test medicine. The simulator work seems like the actual problem to solve, not the invention portion.

HDThoreaun | parent
I think this is a bit misleading. Less than 10% of phase 1 trials end up passing phase 3 and going to market. You can claim that means the challenge is verifying how well they work since you need to run a bunch of phase 3 trials but I think the real challenge is finding a phase 1 target that will end up passing phase 3. Doing a better job of quickly eliminating bad targets is the real challenge. If AI turns that 10% number into 20% then you can get twice as many marketable drugs for the same number of trials. I guess another way of putting it is that finding stuff that has a small chance of working is not a challenge, but finding stuff that has a good chance of working is a huge challenge.