Structure-Informed Scoring Bridges the Generation–Selection Gap in Antibody–Antigen Structure Prediction

Protein co-folding models often sample accurate antibody--antigen complex structures, yet their confidence scores fail to reliably identify them, exposing a gap between structure generation and candidate selection. Across extensive sampling, confidence ranking outperforms random selection mainly by identifying better seeds, while selection among samples from the same seed remains near random. These observations are consistent with a seed-shared representation bottleneck: confidence is strongly influenced by seed-level representations but is comparatively insensitive to candidate-specific structural differences. Motivated by this observation, we introduce ReCon, which reconstructs a candidate-specific representation from each full-complex structure while preserving its interchain geometry before confidence evaluation. ReCon improves both overall ranking and within-seed discrimination. Representation analyses show that reconstruction strengthens interface signals more for acceptable structures than for incorrect structures, while uncertainty in the original pair representations identifies targets more likely to benefit from rescoring. Together, these results are consistent with seed-shared representations constraining within-seed discrimination; candidate-specific reconstruction partially restores structural sensitivity but remains limited in distinguishing candidates of similar quality.