In the (Confi)Dock: model-agnostic DockQ prediction for antibody-antigen co-folding

We present ConfiDock, a lightweight, model-agnostic graph neural network that predicts DockQ directly from co-folded antibody-antigen structures, producing scores comparable across different models without per-backbone calibration or metric selection. Co-folding models are increasingly central to structure-based drug discovery, and inference-time scaling (many candidate structures sampled and the best is selected) can push accuracy well beyond any single prediction. Realizing this potential is particularly challenging because native confidence scores correlate poorly with structural accuracy, and when candidate structures from multiple co-folding models are compared together, each model's scores are calibrated to its own internal representation, making direct comparison unreliable. Trained on decoys from three co-folding models (Boltz-2, ESMFold2, and Protenix-v2) across $\sim$$3{,}200$ antibody-antigen complexes from the SAbDab database and evaluated on two held-out out-of-distribution splits, ConfiDock reaches per-complex Spearman $\rho=0.59$/$0.56$ (Easy/Hard) against ground-truth DockQ, compared to $0.34$/$0.45$ for pooled native confidence scores. A per-backbone rank-transform, which removes scale differences between models while preserving their relative ordering, closely matches ConfiDock's performance, confirming that the core contribution is a unified scoring function that enables direct comparison of confidence scores across co-folding models without per-model calibration/labeled data.