InteractBind
Large-scale protein-ligand dataset and benchmark that probes whether models genuinely localize binding sites and recover interaction types, rather than merely predicting binding likelihood, using ligand-similarity-controlled splits.
Composite
47.5
Experimental validation
Retrospective
Stages
Hit ID
Modalities
small moleculebiologic
Task types
classificationretrievaldocking
Size
complexes: 0
tasks: 3
splits: {'train': 0, 'val': 0, 'test': 0}
note: Curated from PDB-derived structures; exact counts pending camera-ready release
tasks: 3
splits: {'train': 0, 'val': 0, 'test': 0}
note: Curated from PDB-derived structures; exact counts pending camera-ready release
License
CC-BY
First release
2026-05-21
Last updated
2026-05-21
Official site
Leaderboard
→ leaderboard
Dataset
→ dataset
Code / GitHub
→ repository
HuggingFace
→ HF
Paper
A Large-Scale Dataset and Benchmark: Do Protein-Ligand Models Learn Binding Sites or Just Binding Likelihood? · Zhaohan Meng, Zhen Bai, Ke Yuan, Iadh Ounis, Zaiqiao Meng, Hao Xu, Joseph Loscalzo · 2026 · paper · doi:10.48550/arXiv.2605.24045 · 0 citations
Flags
none
Experts
—
Groups
—
Hosted by
—
Related benchmarks
Rubric (7-criterion)
rigor
4
coverage
3
maintenance
3
adoption
1
quality
4
accessibility
2
industry_relevance
3
Notes
arXiv preprint (21 May 2026). Valuable diagnostic framing: separates true binding-site localization from affinity/likelihood shortcuts, with ligand-similarity-controlled splits to test generalization to novel proteins. Brand new (zero citations), code not yet public, and dataset size not yet finalized — accessibility and adoption scored conservatively. Strong scientific motivation co-authored by a clinical systems-biology group (Loscalzo, Harvard).