Cross-modal transfer enables zero-shot prediction of drug responses across cell types

A recent study introduces GRAFT, a framework that leverages genetic perturbation responses to predict drug responses across different cell types without the need for direct drug measurements in the target cell context. By combining genetic and chemical perturbations, GRAFT takes advantage of shared regulatory networks, allowing for the transfer of response information between perturbation modalities.

This innovative approach was tested against existing methods and demonstrated improved correlation, agreement, and accuracy in predicting drug-induced transcriptional responses, particularly for drugs with context-dependent effects. To further enhance prediction capabilities, GRAFT-Adapt refines predictions of unmeasured responses using a limited number of drug measurements from the target context.

The GRAFT-Atlas, a comprehensive resource, spans 13 cellular contexts and 7,856 drug catalog entries, resulting in 102,128 drug-context predictions. The atlas reveals recognizable pharmacological neighborhoods, shared functional programs, context-dependent transcriptional changes, and structural-functional associations supported by molecular docking.

In a validation study involving ten iPSC lines from different donors, GRAFT successfully identified candidate inter-donor differences, the functional relevance of which was confirmed by measured genetic perturbation responses. In summary, GRAFT enables perturbation-informed prediction and comparison of drug responses across diverse cellular contexts, providing valuable insights for drug discovery and development.