FoldBench
Comprehensive benchmark of 1,522 biological assemblies across 9 prediction tasks evaluating all-atom biomolecular structure prediction models including proteins, nucleic acids, ligands, and ions.
Composite
55.8
Experimental validation
Retrospective
Stages
Hit ID
Modalities
small moleculebiologic
Task types
dockingstructure prediction
Size
assemblies: 1,522
tasks: 9
splits: {'train': 0, 'val': 0, 'test': 1522}
tasks: 9
splits: {'train': 0, 'val': 0, 'test': 1522}
License
CC-BY
First release
2025-05-27
Last updated
2026-03-15
Official site
Leaderboard
→ leaderboard
Dataset
→ dataset
Code / GitHub
HuggingFace
→ HF
Paper
Benchmarking all-atom biomolecular structure prediction with FoldBench · Shitong Xu, Qian Feng, Liang Qiao · 2026 · paper · doi:10.1038/s41467-026-00442-x · 12 citations
Flags
none
Experts
—
Groups
—
Hosted by
—
Related benchmarks
Rubric (7-criterion)
rigor
4
coverage
5
maintenance
4
adoption
2
quality
4
accessibility
3
industry_relevance
4
Notes
Published Nat Comms 2026. Covers 9 task types (monomer, multimer, nucleic acid, ligand, ion, antibody-antigen, etc). Revealed that ligand docking accuracy decreases with training set dissimilarity and antibody-antigen predictions fail >50%. Strong structural benchmark but still gaining community traction.