TMEDRP: Decoding Tumor-Intrinsic and Microenvironmental Signatures for Clinical Drug Response Prediction.

MOTIVATION: Clinical drug response prediction is constrained by the paucity of patient data, forcing a reliance on in vitro cell line models. Current transfer learning-based methods typically focus on learning domain-invariant representations, but often overlook the critical role of tumor microenvironment (TME) during cross-domain translation. Given that TME discrepancies between in vitro and in vivo settings are critical factors of therapeutic resistance, explicitly modeling the TME is essential to bridge the gap between preclinical models and patient-specific responses. RESULTS: We propose TMEDRP, a novel two-stage learning framework that integrates tumor-intrinsic signatures and TME-specific components for enhanced clinical drug response prediction. TMEDRP introduces TME influences via a pathway-informed disentanglement strategy, a neural co-expression module, and an uncertainty-driven adaptation mechanism. Extensive benchmarks demonstrate that TMEDRP outperforms existing state-of-the-art methods. Importantly, it exhibits promising zero-shot generalization across unseen cancer lineages and novel compounds, suggesting its robustness across diverse scenarios. Furthermore, quantitative analysis of the model's latent embeddings uncovers key tumor-intrinsic pathways and extrinsic TME factors that distinguish drug-sensitive from resistant cohorts. The identified shared and drug-specific molecular markers align with established clinical mechanisms, supporting the model's biological interpretability and clinical utility. AVAILABILITY AND IMPLEMENTATION: The code and results are available at https://github.com/kybinn/TMEDRP.