PerturbBench

Benchmark for generalization of perturbation foundation models to unseen genetic perturbations / cell contexts.

Composite

71.4

Experimental validation

Retrospective

Stages

Virtual Cell

Modalities

cross-modality

Task types

perturbation-prediction

Size

cells: 400,000
perturbations: 200

License

Apache-2.0

First release

2024-12

Last updated

2025-06

Official site

→ project page

Leaderboard

→ leaderboard

Dataset

→ dataset

Code / GitHub

→ repository

HuggingFace

→ HF

Paper

PerturbBench: Benchmarking Single-Cell Perturbation Foundation Models · Wu Y, Barry T, Wang K, et al. · 2024 · paper · doi:10.48550/arXiv.2412.10091 · 35 citations

Flags

none

Experts

Aviv Regev

Groups

Genentech gRED (ML)

Hosted by

—

Related benchmarks

Open Problems: Perturbation Prediction, scPerturb

Rubric (7-criterion)

rigor

coverage

maintenance

adoption

quality

accessibility

industry_relevance

Notes

Pharma-led (Genentech); well-specified eval.

← Back to all benchmarks

Compare:

Open comparison →