Background: Accurate drug response prediction alone is insufficient for translational impact; computational models must also generate biologically plausible hypotheses. We present drGT, a heterogeneous graph neural network (GNN) model over drugs, genes, and cell lines that couples prediction with mechanism-oriented interpretability via attention coefficients (ACs). Results: drGT encodes a drug-gene-cell line graph and uses ACs to describe drug-gene associations. We assess both predictive generalization (random, unseen-drug, unseen-cell, and zero-shot splits) and biological credibility (text-mined PubMed co-mentions and comparison to a structure-based DTI predictor) on GDSC1, GDSC2, NCI60, and CTRP datasets. Here, we evaluate both regression (IC50 values) and binary classification (cell-line sensitivity) settings to reflect typical pharmacogenomic experiments; we report R² for regression and AUROC for classification tasks. Across benchmarks, drGT consistently delivers top regression performance while maintaining competitive classification accuracy. Under random 5-fold cross-validation, where 20% of samples are randomly masked for testing in each fold, drGT attains an R² of up to 0.690 (1st overall against similar methods) and an AUROC of up to 0.945 (3rd overall). In leave-one-out generalization tests, where either all samples from one cell line or all samples of one drug are excluded during training, drGT achieves R² values of 0.692 (1st) and 0.022 (1st) and AUROCs of 0.706 (3rd) and 0.844 (2nd), respectively. In the zero-shot prediction test, trained on GDSC1 and evaluated on GDSC2, drGT achieves an R² of 0.334 (1st) and an AUROC of 0.786 (1st), both of which represent the highest scores among all models. For interpretability, AC-derived drug-gene links recover known biology: among 976 drugs from the NCI60 dataset with known DTIs, 36.9% of predicted links match established DTIs, and 63.7% are supported by either PubMed abstracts or a state-of-the-art (SOTA) structure-based DTI prediction model. Moreover, an over-representation analysis of AC-ranked genes identifies mechanism of action (MoA)-consistent pathways, such as KRAS signaling, for relevant kinase inhibitors, providing pathway-level explanations for the model predictions. Conclusions: drGT advances predictive generalization and mechanism-centered interpretability with AC-derived drug-gene links, offering SOTA regression accuracy and literature-supported biological hypotheses, demonstrating how interpretable graph learning can bridge AI prediction and biological discovery. Code: https://github.com/sciluna/drGT