drGT: Interpretable Drug Response Prediction with Attention-Guided Gene Attribution on a Drug-Cell-Gene Heterogeneous...

Area of Science:

• Computational Biology
• Pharmacogenomics
• Machine Learning

Background:

• Accurate drug response prediction is crucial for translational medicine.
• Biological plausibility of predictive features is essential for understanding drug mechanisms.
• Existing models often lack interpretability, hindering biological discovery.

Purpose of the Study:

• To develop drGT, a heterogeneous graph deep learning model for drug response prediction.
• To couple prediction accuracy with mechanism-oriented interpretability using attention coefficients (ACs).
• To assess the model's predictive generalization and biological plausibility across diverse datasets.

Main Methods:

• Utilized a heterogeneous graph deep learning approach integrating drugs, genes, and cell lines.
• Employed attention coefficients (ACs) for mechanism-oriented interpretability.
• Evaluated on GDSC, NCI60, and CTRP datasets using various cross-validation strategies (random, unseen-drug, unseen-cell, zero-shot).
• Assessed biological plausibility using text-mined PubMed gene-drug co-mentions and a structure-based drug-target interaction (DTI) predictor.

Main Results:

• drGT achieved top regression performance and competitive classification accuracy for drug sensitivity across benchmarks.
• Outperformed baselines in regression under random 5-fold cross-validation (up to R²=0.690).
• Demonstrated strong performance in leave-one-out tests for unseen cell lines (AUROC=0.706) and drugs (AUROC=0.844, R²=0.022).
• Achieved highest scores in zero-shot prediction (AUROC=0.786, R²=0.334).
• AC-derived drug-gene links showed high concordance with known DTIs (36.9%) and literature support (63.7%).
• Enrichment analyses of AC-prioritized genes revealed drug-perturbed biological processes.

Conclusions:

• drGT advances predictive generalization and mechanism-centered interpretability in drug response prediction.
• Offers state-of-the-art regression accuracy and generates literature-supported biological hypotheses.
• Demonstrates the utility of graph learning from heterogeneous data for biological discovery.