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

Abstract

Virtual-screening cascades assembled from co-folding and learned-affinity models are now routine, and each stage is usually trusted on the strength of its own score. We measured, stage by stage, whether those scores carry the information a practitioner would attribute to them, using the Keap1–Nrf2 protein–protein interaction and a 130,793-compound make-on-demand library as the test bed. Most measurements are negative, and none of the failure modes is specific to one of the two models tested. Two independently developed affinity models order activity-cliff pairs above chance (about 68% of 181 pairs) but carry no information about the size of the potency gap, and both couple to molecular weight; the built-in molecular-weight correction rescales without de-biasing. Neither model resolves enantiomers, including the model that generates an explicit high-confidence three-dimensional pose. Pose-model confidence is only weakly informative about geometric validity, so an explicit validity check is required rather than optional. Training-free guidance is required for stereochemical fidelity in pose generation, and diffusion steps can be halved without detectable loss. A seeded matched-decoy gate validates the pose-and-energy stage where no single physicochemical descriptor does, and the pose model reproduces all five co-crystal references below 2 Å, three of them released after its training cutoff. Single-structure MM-GBSA ordering is not validated by ensemble refinement, and the two solvation models disagree both on the ordering and on which candidates are separable at all. The control suite is released as a reusable artifact.