Benchmarking confidence estimation and rescoring for cyclic peptide–protein complex predictions

Abstract

Cyclic peptides are a rapidly expanding class of therapeutics, but the reliability of deep-learning structure prediction for cyclic peptide–protein complexes has not been systematically evaluated. We assembled a curated benchmark of 111 non-redundant complexes spanning five cyclization chemistries and assessed two co-folding models, Boltz and Protenix, each generating 100 poses per target (22,200 total). Stratifying all poses by complex attributes, we found that disulfide-cyclized peptides and small protein targets (200 or fewer target residues) were predicted significantly worse by both tools, with target size the largest and most consistent effect; overall accuracy nevertheless remained high (median top-pose DockQ of about 0.89, 96–98% of targets Acceptable or better), indicating that pose generation is rarely the bottleneck. Conversely, native model ranking scores correlated only moderately with pose quality (Spearman rank correlations of 0.53–0.66): approximately 12% of poses showed high model ranking score/confidence despite poor pose DockQ quality, and the highest-quality pose was not ranked first for nearly every target. We therefore augmented the native score with externally computed interface descriptors normalized by chain length, principally the per-residue density of inter-chain hydrogen bonds, in a gradient-boosted rescoring model evaluated under target-grouped cross-validation that prevents leakage, improving out-of-fold ROC-AUC for both tools, significantly so for Protenix. Together, these findings identify pose ranking, rather than pose generation, as the major limitation of current cyclic peptide–protein complex prediction and demonstrate that complementary structural features can improve confidence-based pose selection.