Breadth Over Depth: Predictor-Agnostic Selection for Improved Antibody-Antigen Co-Folding

Protein structure prediction workflows commonly generate many structures from one predictor and select among them using that model’s confidence score. We ask whether a fixed inference budget is better spent sampling broadly across complementary co-folding predictors, and whether the re- sulting mixed pools can be ranked without predictor-specific information. On 103 cluster-disjoint, post-2021 antibody–antigen complexes, native confi- dence is only weakly associated with pose quality within targets (Spearman ρ = 0.08–0.20). We introduce a predictor-agnostic structural selector that combines cross-pose agreement, within-source reproducibility and inter- face geometry, using no predictor identity, no native confidence and no predictor-specific calibration. Sampling only two structures from each of five predictors yields a ten-structure pool reaching 55.4% medium-or-better Success@1 (DockQ ≥ 0.49), above every individual predictor sampled 100 times (26–52%): +23.4 to +29.2 percentage points (pp) against the three weaker predictors and +6.9 and +3.9 pp against the two strongest, at 1.8- fold less GPU time. A selector trained on four predictors loses at most 2.5 pp when applied to pools containing the unseen fifth, and adding that predictor at fixed budget improves Success@1 by 0.6–12.7 pp. Generating full MSA-conditioned ESMFold2 and adding it as a sixth predictor without retraining gives +1.1 pp. Experiments show the gain requires genuinely different predictors rather than extra structures or artificial source labeling.