Predicting drug responses of unseen cell types through transfer learning with foundation models.

Abstract

Drug repurposing through single-cell perturbation response prediction provides a cost-effective approach for drug development, but accurately predicting responses in unseen cell types that emerge during disease progression remains challenging. Existing methods struggle to achieve generalizable cell-type-specific predictions. To address these limitations, we introduce the cell-type-specific drug perturbatIon responses predictor (CRISP), a framework for predicting perturbation responses in previously unseen cell types at single-cell resolution. CRISP leverages foundation models and cell-type-specific learning strategies to enable effective transfer of information from control to perturbed states even with limited empirical data. Through systematic evaluation across increasingly challenging scenarios, from unseen cell types to cross-platform predictions, CRISP shows generalizability and performance improvements. We demonstrate CRISP's drug repurposing potential through zero-shot prediction from solid tumor data to sorafenib's therapeutic effects in chronic myeloid leukemia. The predicted anti-tumor mechanisms, including CXCR4 pathway inhibition, are supported by independent studies as an effective therapeutic strategy in chronic myeloid leukemia, aligning with past studies and clinical trials.