New AI model predicts how cells respond to drugs and genetic perturbations across diverse contexts

Scientists at the Arc Institute have developed State, a machine learning model that predicts how individual cells will respond to drugs, genetic edits, and other perturbations—even in cell types or tissues where no such experiments have been performed. The work, reported today in Cell, is led by senior author Yusuf H. Roohani and first authors Abhinav K. Adduri, Dhruv Gautam, and Beatrice Bevilacqua.

State addresses a central challenge in cell biology: current perturbation screens are expensive and cannot be run across every cell type, tissue, or patient. To fill these gaps, the model learns from 167 million observational and interventional cells and over 100 million perturbed cells, capturing both biological heterogeneity within a population and technical noise across experiments. When tested on a large drug screen covering 50 cancer cell lines, State achieved an absolute improvement in perturbation discrimination of 66% and outperformed the next-best method by 91% in correlating predicted and actual gene expression changes.

The model also distinguishes cell-type-specific responses. For example, it correctly predicted that the drug trametinib triggers a strong, specific response in C32 melanoma cells—a cell line it had never seen during training. In another test, State predicted cell viability from simulated gene expression profiles with a correlation of 0.52, far outperforming simple mean baselines.

A key innovation is the Cell-Eval framework, introduced alongside State, which provides a standardized way to evaluate perturbation prediction models across multiple biologically meaningful metrics. The authors note that while State performs well on zero-shot prediction of overall effect sizes, its accuracy on individual differentially expressed genes is more sensitive to dataset size and quality—a limitation that may improve as larger perturbation atlases become available.

Overall, State demonstrates that machine learning can generalize perturbation effects across cellular contexts with practical accuracy, opening the door to in silico experimentation for drug repurposing, patient-specific treatment planning, and hypothesis generation in basic biology.

Reference: Adduri AK, Gautam D, Bevilacqua B, et al. Predicting cellular responses to perturbation across diverse contexts with State. Cell. 2026. DOI: 10.1016/j.cell.2026.08.031