Multiparadigm Benchmark of Molecular Docking: From Physics to Co-Folding and Hybrid Models

Generalizability of molecular docking predictions across novel protein targets and distinct small-molecule chemotypes is key to the successful application of deep learning co-folding approaches in drug discovery, as highlighted by recent benchmarks, including Runs-n-Poses [Škrinjar, P.; et al.Nat. Struct. Mol. Biol.2026, 33, 782-794.]. To compare generalizability across physics-based, hybrid rescoring models, and co-folding approaches, we systematically benchmarked several molecular docking tools representing these paradigms. Our results support previous observations that co-folding approaches show superior performance on targets resembling their training datasets. However, they deteriorate sharply to unacceptably low 20-40% success rates on novel dissimilar systems, consistent with memorization. In contrast, physics-based and hybrid methods exhibit greater out-of-distribution robustness, maintaining more than 60% success rates for docking protein-ligand complexes with minimal similarity to previously known complexes. Our results highlight the complementarity of AI-based approaches and methods based on physical sampling in all-atom models, in terms of their applicability range, and argue for the benefits of tighter integration.