EpiRank: partner-agnostic surface-patch retrieval for epitope prediction of protein binding interfaces

Structure-based interface predictors are normally reported as a single pooled ROC AUC over a residue test set. For a protein designer the operative question is different and much harder: given one target chain, will the top-ranked surface patch be a real binding site? We built EpiRank, a partner-agnostic epitope predictor, on HumanPPI 29,246 predicted human dimers, 12,296 target chains, 4.18 M residues using only single-chain geometry and physicochemistry (170 features, no partner, no sequence search, no structural alphabet), and evaluated it as a retrieval problem with homology- and partner-graph-controlled splits, on the 3,529 chains that survive an explicit partner-contamination filter. The model places a real epitope in its top-ranked patch 64 % of the time. That number is far less informative than it looks. Compact patches placed at random hit a real epitope 41 % of the time under the same rule, so the model's lift is 1.57 x, not 64 %; 54 % of all predicted patches land on nothing; and when the same chains are scored by models trained on a 2,000-chain and an 11,669-chain superset, a substantial fraction change answer while the aggregate barely moves. We document four specific ways this task invites overstatement partner-graph percolation across splits, single-linkage homology percolation at the conventional 30 % identity threshold, a count-only hit rule that rewards spatial smoothing rather than signal, and hotspot precision conditioned on success, which ranks a random baseline above the trained model. Each was hit during development and each inverted a conclusion we had already drawn. We argue that interface-prediction papers should report lift over a geometry-matched random baseline, patch precision, and per-target stability alongside AUC, and we release the evaluation harness that computes them.