Predicting Cellular Perturbations: Biological Benchmarks and AI-Assisted Model Discovery

Large-scale genetic perturbation screens have enabled researchers to probe gene function and regulatory networks, but measuring changes in gene expression for every perturbation in a new cellular context remains experimentally costly, if not infeasible. This limitation has motivated substantial interest in machine learning (ML) and artificial intelligence (AI) models that predict transcriptional responses to unmeasured perturbations and cellular contexts.

In this talk, I will give an overview of a simplified framing of the problem and present some recent work addressing these questions through statistical modeling, biological benchmarking, and AI-assisted model development. We begin with a prediction benchmarking study that compares popular deep learning methods versus simplified statistical models built on top of pretrained embeddings (vectors summarizing biological information) for classifying whether perturbation–gene pairs lead to up, down, or no changes in differential expression (DE). We find that models with strong expression-level accuracy do not necessarily identify the genes that respond to a perturbation. In particular, discrepancies between predicted and observed expression distributions can produce substantial disagreement in differential expression calls.

I will also discuss our preliminary experiments using AI agents to develop perturbation-prediction methods. Some agent-developed models outperform established baselines on expression prediction, and inspection of their analysis workflows reveals recurring uses of biological knowledge and statistical adjustments that are conceptually simple, yet elaborate and informative. These results highlight opportunities for agents to help discover useful modeling strategies, while raising questions about reproducibility, evaluation, and the translation and generalization of predictive gains to other related problems. Our work explores how biological knowledge, statistical structure, and increasingly automated model development can improve the prediction and interpretation of cellular perturbations.