High-throughput perturbation screens increasingly rely on machine-learning models to predict the effects of gene combinations that were never measured. A new arXiv preprint focuses on a particularly tricky scenario: when both genes in a test pair have been observed individually, but their combined effect has not.

The paper, titled "Do Perturbation Predictors Learn Gene Interactions? Diagnosing Extrapolation in Agentic Screening," asks whether such models genuinely learn gene interactions or simply extrapolate from single-gene measurements. The abstract highlights this as a difficult case for prediction, and the title suggests the authors are concerned about overconfident extrapolation.

The study introduces a diagnostic approach for what it calls agentic screening, presumably referring to automated or iterative screening pipelines. As a preprint, the work has not yet undergone peer review, and the abstract does not disclose specific findings or conclusions.