FoldBench published in Nature Communications

Our paper Benchmarking all-atom biomolecular structure prediction with FoldBench has been published in Nature Communications (volume 17, Article 442, 2026).

FoldBench is a comprehensive benchmark dataset consisting of 1522 biological assemblies categorized into nine distinct prediction tasks across proteins, nucleic acids, ligands, and diverse interfaces. We systematically evaluate five recent all-atom predictors (AlphaFold 3, Boltz-1, Chai-1, HelixFold 3, and Protenix), quantify performance dependence on training set and ligand similarity, and map strengths and failure modes across tasks.

Xu, S., Feng, Q., Qiao, L. et al. Benchmarking all-atom biomolecular structure prediction with FoldBench. Nat Commun 17, 442 (2026).