Chem2Gen-Bench
Benchmark for target-matched chemical-to-genetic translation in perturbation response space, testing whether virtual-cell/perturbation models align chemical (compound) responses with genetic (knockdown) responses at shared targets. Evaluates pairwise alignment, retrieval, protocol-covariate associations, feature spaces, and single-cell foundation-model embeddings.
Composite
63.3
Experimental validation
N/A — retrospective benchmark over public chemical and genetic perturbation profiles
Stages
Virtual CellTarget IDDisease modeling
Modalities
small molecule
Task types
retrievalclassificationregression
Size
chemical_perturbation_profiles: 260,084
genetic_perturbation_profiles: 1,099,045
splits: {'train': 0, 'val': 0, 'test': 1359129}
note: 260,084 chemical + 1,099,045 genetic perturbation profiles organized into cell-target contexts; target-matched K562 audit subset
genetic_perturbation_profiles: 1,099,045
splits: {'train': 0, 'val': 0, 'test': 1359129}
note: 260,084 chemical + 1,099,045 genetic perturbation profiles organized into cell-target contexts; target-matched K562 audit subset
License
Other — see arXiv 2606.21109 (derived from public perturbation datasets)
First release
2026-06-19
Last updated
2026-06-19
Official site
Leaderboard
→ leaderboard
Dataset
Code / GitHub
→ repository
HuggingFace
→ HF
Paper
Chem2Gen-Bench: Benchmarking Chemical-to-Genetic Translation in Perturbation Response Space · Yuxiang Lin, Ying Chen · 2026 · paper · doi:N/A — arXiv preprint 2606.21109 · 0 citations
Flags
none
Experts
Groups
Hosted by
—
Related benchmarks
Rubric (7-criterion)
rigor
4
coverage
4
maintenance
3
adoption
1
quality
4
accessibility
3
industry_relevance
3
Notes
Fills a real gap: chemical vs genetic perturbations are usually evaluated separately, so target-matched translation was under-tested (rigor 4, coverage 4 given >1.35M profiles across cell-target contexts). Honest negative result: evaluated foundation-model embeddings did not consistently beat simple gene-delta baselines in the K562 audit, and background adjustment lowered mean retrieval success — a useful reality check for virtual-cell claims (quality 4). Adoption 1 as a June 2026 preprint; accessibility 3 pending public code/leaderboard. Directly relevant to target ID and mechanism inference.