Interpretable Transfer Learning for Cancer Drug Resistance: Candidate Target Identification.

Tumor drug resistance exhibits substantial heterogeneity across cancer types, reflecting distinct molecular mechanisms in each malignancy. To characterize this complexity, we developed a pan-cancer transfer learning framework that integrates bulk RNA-seq data with a residual variational autoencoder (Res VAE) backbone. Five models were trained on the Genomics of Drug Sensitivity in Cancer (GDSC) dataset, which includes drug response profiles for 72 chemotherapeutic agents. Among them, three models are specially designed by incorporating variational autoencoders and large pretrained models (LLMs): the LLM large VAE (VAE_LL), the LLM small VAE (VAE_LS), and the LLM distillation VAE (VAE_LD). Random Forest (RF) and eXtreme Gradient Boosting (XGB) were included as ensemble learning baselines. After internal cross-validation, the top four models (VAE_LL, VAE_LD, XGB, and RF) were applied to five representative TCGA cohorts comprising 1,836 patients. For each cancer type, resistance to nine clinically relevant first-line drugs was modeled, resulting in 180 drug-cancer prediction tasks. Among all models, VAE_LD achieved the best overall performance, with a mean AUC of 0.81 and an F1 score of 0.92 on the GDSC benchmark, and maintained strong predictive power in the clinical validation phase. Interpretation analyses identified tumor-specific resistance biomarkers with clinical significance. In lung adenocarcinoma, elevated expression of