Averaging in gene perturbation benchmarks hides perturbation-strength variation: Consequences for model evaluation
Yash Sharma, Madhav Subramaniyam, Pio Passariello, Dylan Jay, David Render, Márió Szegedy
The paper evaluates how averaging perturbation responses in gene perturbation benchmarks affects the ability to assess perturbation-strength variation, finding that current models fail to reliably recover perturbation identity.
bioRxiv · Oct 23, 2026 · PreprintBioinformatics