Benchmarking antibody-antigen co-folding on human monomeric antigens

Co-folding methods have revolutionized protein complex prediction, yet antibody-antigen interactions continue to pose a challenge due to their flexible complementarity determining region (CDR) loops and the lack of co-evolutionary signals guiding prediction. Efforts are being made to improve co-folding models, increase sampling, and incorporate experimental data like epitope constraints.

The HuMonoAg-Bench, a benchmark of 412 experimentally determined antibody complexes with human monomeric antigens, was created to independently assess ten co-folding protocols. The latest methods significantly surpass older ones, generating medium-or-better top-ranked models (DockQ [≥] 0.49) for roughly half of post-cutoff Fv complexes without templates or experimental restraints, and delivering similar results on both pre- and post-cutoff antigens.

Improved CDRH3 modeling was primarily responsible for these gains, while antigen structures and the remaining CDR loops were modeled comparably well across methods. Providing true epitope residues as constraints enhanced the performance of earlier methods by around 20-30 percentage points, aligning their performance with the top unconstrained methods.

Failures were primarily due to the inability to sample the correct binding mode rather than ranking it. Despite improvements from employing multiple methods, the overall success rate remained modest, as the remaining unsolved complexes are structurally diverse, lacking a single structural property that accounts for the current limitations.

In summary, while recent progress has been substantial, many antibody-antigen complexes still exceed the capabilities of current co-folding methods, with CDRH3 modeling and accurate binding mode sampling being the main challenges.