5 ms·
AlphaFold Excitement
- dekhn 5y agoI like reading Derek Lowe because he's one of the few prominent scientist authors who actually just comes out and calls the modern narrative about drug discovery a fiction. SV has this weird addiction to rapid success and biotech drug discovery is sort of the opposite of that. So there are a lot of cases where a data scientist thinks they applied some magic tech and solved a hard bottleneck that will suddenly transform drug discovery into an efficient scientific enterprise. That's not going to happen. There is no magic loss function that solves the hard problem and even if you had a computer that could compute NP-complete problems quickly, it wouldn't change the process at all. Instead, the one step that would be the most effective is to make a model that could predict if a late stage molecule could get through FDA approval using only the data that is submitted to the FDA (lots of boring paperwork about trials). With prediction accuracy (both false positive and negative below 1%) you could easily save an enormous amount of time and money.
- axg11 5y agoDerek Lowe is rightfully pointing out that protein structures (AlphaFold's main purpose) is not the rate limiting step in drug discovery today. Other articles overhype AlphaFold's achievements and this one pours water on that, a natural balance in the commentary from the scientific community. AlphaFold is still a great achievement. In the long term, the biggest impact won't be the model itself, but the demonstration that hard biological problems can be tackled, with high accuracy, with deep learning. Toxicity, specificity, and candidate generation are all super-problems that are the limiting steps today. These will be broken down into sub-problems and eventually tackled by machine learning models once we have the data and understanding to do so.