Computational & AI chemRxiv
Standard ML virtual screen of COX-2 yields candidates that dock no better than random compounds
- Random forest potency models for COX-2 and COX-1 were used to screen a large ZINC subset, and filtering left a small candidate list that was docked with Vina alongside a random control set from the same library.
- Docking was checked by redocking a reference inhibitor and by showing the protocol separates known actives from random compounds.
- The ML candidates docked no better than random, model predictions did not track docking ranks, and most candidates fell outside the applicability domain, despite good retrospective benchmark scores.
- The authors argue that applicability-domain coverage and same-library random baselines should be required reporting.
A 'same-library random baseline' is the right control, but Vina is a weak judge: its scores barely correlate with measured COX-2 affinity, so 'no better than random' may partly reflect docking noise. The sharper lesson is the domain gap: training sets of known NSAID-like chemotypes versus ZINC. Next step: confirm a handful of top picks in a biochemical assay.
Paper: chemRxiv ↗ A Standard Machine-Learning Virtual Screening Pipeline Produces Candidates Indistinguishable From Random Drug-Like Compounds: A Quantitative Audit Using COX-2 As Case Study