DeepTNR: predicting tumor drug sensitivity from spatial transcriptomics with deep graph contrastive and transfer learning

Heterogeneity within the tumor microenvironment remains a major obstacle to precision oncology. Although existing drug-sensitivity models can capture transcriptional heterogeneity, they generally do not account for the spatial organization of predicted drug responses. Here, we developed DeepTNR, a spatially resolved framework that integrates tumor-stroma interface features, spatial autocorrelation, and distance-dependent associations between local cellular composition and predicted drug sensitivity. Across nine tumor types, with colorectal cancer (CRC) as a representative example, DeepTNR identified irinotecan as the most sensitive drug and revealed significant spatial autocorrelation in its predicted response patterns. Spatial proximity effect analysis further showed that local enrichment of immune and proliferative programs positively was associated with increased irinotecan sensitivity. Together, these findings demonstrate the potential of DeepTNR to resolve intratumoral therapeutic heterogeneity and identify niche-specific vulnerabilities.