New Machine Learning Model Predicts Drug Responses from Basic Statistical Data

Rhaister is a perturbation-response prediction framework that operates on screen-level summary statistics rather than raw molecular data, allowing it to generalize drug response predictions to new biological contexts after measuring only a small number of perturbations. The tool was developed alongside Emerald Bay, a purpose-built dataset derived from the Tahoe platform that integrates multi-day cancer drug perturbation experiments, pooled Mosaic tumor contexts, and paired transcriptomic readouts. Rhaister can work with both fine-grained molecular outputs, such as transcriptional profiles from large perturbation screens like Tahoe-100M, and higher-level phenotypic endpoints. A key extension, Rhaister-O, enables zero-shot prediction of drug responses in entirely new contexts using only baseline gene expression data, which the authors describe as the first model of its kind for this task. Benchmarking results indicate that Rhaister matches or surpasses substantially more resource-intensive virtual-cell models, often reaching ceiling values on standard evaluation metrics, while completing training in seconds and inference in milliseconds. The authors position Rhaister as a fast, interpretable alternative to deep learning-based virtual cell approaches, with practical applications in prioritizing experiments and identifying context-dependent drug mechanisms.