ShallowBench
Strictly curated benchmark of 5,780 shallow-pocket protein targets extracted from CrossDocked2020 for stress-testing structure-based generative drug design models on low-pocketability / historically 'undruggable' interfaces (e.g. KRAS, MYC).
Composite
48.5
Experimental validation
No — purely in silico evaluation of predicted binding affinity on curated shallow pockets
Stages
Hit ID
Modalities
small molecule
Task types
generationdocking
Size
targets: 5,780
splits: {'train': 0, 'val': 0, 'test': 5780}
note: 5,780 shallow-pocket targets isolated from CrossDocked2020 via Alpha Shape lid-vs-voxel concavity filtering; used as an evaluation set
splits: {'train': 0, 'val': 0, 'test': 5780}
note: 5,780 shallow-pocket targets isolated from CrossDocked2020 via Alpha Shape lid-vs-voxel concavity filtering; used as an evaluation set
License
Other — derived from CrossDocked2020 (CC-BY-4.0 upstream)
First release
2026-06-04
Last updated
2026-06-04
Official site
Leaderboard
→ leaderboard
Dataset
Code / GitHub
→ repository
HuggingFace
→ HF
Paper
ShallowBench: Benchmarking Generative Drug Design Models on Shallow-Pocket Targets · Saket Reddy, Shiwei Liu · 2026 · paper · doi:N/A — arXiv preprint 2606.06717 · 0 citations
Flags
none
Experts
Groups
Hosted by
—
Related benchmarks
Rubric (7-criterion)
rigor
3
coverage
3
maintenance
2
adoption
1
quality
4
accessibility
3
industry_relevance
3
Notes
Fills a real gap: most structure-based generative benchmarks reward deep, druggable pockets, so models look strong without ever facing the hard, flat oncology interfaces (KRAS, MYC) that matter clinically. Curation logic (Alpha Shape lid volume minus protein atom voxel volume) is principled and the 5,780-target scale is solid (quality 4). Scored conservatively on rigor (preprint, no peer review or reproduction yet), maintenance (single release, no leaderboard infrastructure), and adoption (0 citations, June 2026). Industry relevance is real given the undruggable-target focus but unproven in pharma pipelines.