AblateCell agent automates ablation studies on virtual cell repos
AblateCell is a reproduce-then-ablate agent that automatically reproduces baselines and runs closed-loop ablation studies on AI Virtual Cell repositories, achieving 88.9% end-to-end workflow success and 93.3% accuracy in identifying critical components.
Score breakdown
AI/coding practitioners building or evaluating biological ML pipelines can use AblateCell to automate the otherwise manual, error-prone process of reproducing baselines and identifying which model components actually drive performance gains.
- 01AblateCell is a reproduce-then-ablate agent targeting AI Virtual Cell repositories.
- 02Systematic ablations are rarely performed in this domain due to under-standardized, tightly coupled biological repositories.
- 03The agent auto-configures environments, resolves dependency and data issues, and reruns official evaluations to reproduce baselines.
AblateCell addresses a persistent gap in AI Virtual Cell research: systematic ablation studies — which attribute performance gains to specific components — are rarely conducted because biological code repositories are under-standardized and tightly coupled to domain-specific data and formats. Existing coding agents can translate ideas into implementations but lack a verifier capable of reproducing strong baselines and rigorously testing which components truly matter. AblateCell fills this role with a two-phase reproduce-then-ablate pipeline.
In the first phase, the agent auto-configures environments, resolves dependency and data issues, and reruns official evaluations while emitting verifiable artifacts.
In the first phase, the agent auto-configures environments, resolves dependency and data issues, and reruns official evaluations while emitting verifiable artifacts. In the second phase, it conducts closed-loop ablation by constructing a graph of isolated repository mutations and adaptively selecting experiments using a reward function that trades off performance impact against execution cost. Evaluated on three single-cell perturbation prediction repositories — CPA, GEARS, and BioLORD — AblateCell achieves 88.9% end-to-end workflow success, a +29.9% improvement over a human expert baseline, and 93.3% accuracy in recovering ground-truth critical components, a +53.3% improvement over a heuristic approach. These results demonstrate scalable, repository-grounded verification and attribution directly on biological codebases.
Key facts
- 01AblateCell is a reproduce-then-ablate agent targeting AI Virtual Cell repositories.
- 02Systematic ablations are rarely performed in this domain due to under-standardized, tightly coupled biological repositories.
- 03The agent auto-configures environments, resolves dependency and data issues, and reruns official evaluations to reproduce baselines.
- 04Closed-loop ablation uses a graph of isolated repository mutations with adaptive experiment selection under a performance-vs-cost reward.
- 05Evaluated on three single-cell perturbation prediction repositories: CPA, GEARS, and BioLORD.
- 06Achieved 88.9% end-to-end workflow success, +29.9% over a human expert baseline.
- 07Achieved 93.3% accuracy in recovering ground-truth critical components, +53.3% over a heuristic baseline.