scRADAR: Dissecting intratumoral drug response heterogeneity at single-cell resolution via mechanism-guided prototype routing.

Precision oncology requires resolving intratumoral heterogeneity to identify drug-resistant cell states associated with treatment failure and relapse. Although single-cell RNA sequencing enables characterization of heterogeneous resistance-associated states, single-cell drug-response phenotype prediction remains challenging because of sparsity, noise, class imbalance, and limited mechanistic interpretability. Here, we present scRADAR (Response Analysis via Drug-Aware Routing), a mechanism-guided prototype routing framework for predicting and interpreting drug-response phenotypes at single-cell resolution. Rather than relying on cell-line-anchored transfer learning, scRADAR learns directly from labeled single-cell cohorts. The framework integrates metabolic and signaling pathway activities to form a dual-view cellular representation, conditions pathway embeddings on drug mechanisms through feature-wise linear modulation, and uses sparse prototype routing to decompose predictions into interpretable response archetypes. Across nine independent cohorts, scRADAR showed strong predictive performance and consistent cross-cohort behavior, particularly under imbalanced settings. Post hoc attribution analyses highlighted candidate TGF-β-associated epithelial-to-mesenchymal transition signatures in Erlotinib-associated Resistant-labeled states and cytoskeletal/metabolic response-associated signatures in BET-inhibitor-associated Resistant-labeled states. These results suggest that scRADAR provides an interpretable framework for single-cell drug-response phenotype prediction and for generating hypotheses about resistance-associated programs from heterogeneous tumor transcriptomes.