Beyond AlphaFold 3: a standardised benchmark of open-source co-folding models for antibody–antigen complexes, from docking accuracy to interaction specificity

Hs) from SAbDab2-Bench, a newly curated dataset, and a reconstructed antibody-antigen subset of FoldBench, Ab-FoldBench. Using standardized inputs, sampling, and interface metrics that isolate the antibody-antigen interface, we compare docking accuracy, inference-time scaling, confidence-based candidate structure selection and sensitivity to interaction perturbations. OpenDDE achieved the highest median DockQ and docking pass rate, followed by Protenix-v2, although their difference was not statistically significant. Both models outperformed AlphaFold 3 on these measures, alongside offering fewer licensing restrictions on model and prediction use. We further find that model confidence was generally effective at identifying accurate predictions but showed model-dependent variations. To probe interaction specificity, which is essential for antibody design, we generated decoys through non-conservative interface mutations, and non-cognate antigen shuffling. Confidence decreased for both perturbations, but more consistently for non-cognate antigens than for interface mutations, with OpenDDE retaining comparatively high confidence for many decoys. Overall, Protenix-v2 combined high docking accuracy with strong rejection of perturbed interfaces, OpenDDE achieved the highest docking accuracy, albeit at the longest runtime, while ESMFold2-fast offered the fastest inference for large-scale triage. These results provide an up-to-date guidance framework for selecting and configuring co-folding models in antibody discovery, design, and engineering.