Per-Stage Controls and Failure Modes in a Co-Folding Virtual Screening Cascade for the Keap1–Nrf2 Interaction

Abstract

Virtual-screening cascades combine learned affinity prediction, co-folding, physics-based rescoring and molecular dynamics, but each score requires an independent control. We audited these stages for the Kelch-like ECH-associated protein 1 (Keap1)–nuclear factor erythroid 2-related factor 2 (Nrf2) interaction using a 130,793-compound make-on-demand library. Two affinity models showed molecular-weight coupling and ordered activity-cliff pairs above chance (about 68%), but neither showed a detectable association between predicted and measured potency gaps or above-chance enantiomer ordering. Co-folding confidence weakly discriminated valid from invalid poses, whereas training-free guidance raised the share of poses with preserved stereochemistry from 28% to 95–97%. For five co-crystal references, the best of five top-ranked poses lay within 2 Å; three structures were released after the model’s training cutoff. A seeded gate separated eleven seed compounds from model-selected decoys with a stratified area under the receiver operating characteristic curve (AUROC) of 0.98. In a size-stratified, anion-restricted test of 101 actives and 124 assay-defined inactives, molecular mechanics generalised Born surface area (MM-GBSA) rescoring reached an AUROC of 0.68 (95% confidence interval 0.60–0.76), was not resolved from the strongest descriptor baseline (0.64), and fell below both affinity models (0.82–0.86). Ensemble molecular dynamics of five candidates did not reproduce the single-structure ordering, and two solvation models disagreed on candidate separability. The results bound the supported use of each stage on this test system. The controls and their data are released as a reusable audit suite.