Do Perturbation Predictors Learn Gene Interactions? Diagnosing Extrapolation in Agentic Screening
The study tests if machine learning models learn gene interactions when predicting effects of unseen gene combinations. It introduces Additive Disagreement (AD), which measures how much a model's prediction deviates from the simple sum of individual gene effects. AD proves more reliable than standard confidence signals for spotting extrapolation errors in screening workflows.