A Hybrid Graph and Transformer-Based Deep Learning Model for Multi-Omics Drug Response Prediction

Background:Predicting drug response is critical for precision medicine, enabling personalized therapies and minimizing adverse effects. However, multi-omics data pose challenges such as high dimensionality, missing values, noise, and class imbalance.Methods:This study proposes a hybrid graph-aware framework, GraphTrans-Omics, integrating multi-stage preprocessing and learning. Data are normalized using log₂ transformation and z-score scaling, followed by imputation via Singular Value Thresholding (SVT). Feature selection combines Mutual Information, Recursive Feature Elimination, and LASSO. To address class imbalance, a graph-based augmentation method (GSMOTE-GMC) is introduced. Random Forest and Multi-Layer Perceptron models are trained on the processed features with graph-based enhancement.Results:The framework achieves competitive performance with accuracy of 0.42–0.46, precision of 0.44, recall of 0.43, AUC of 0.64, and RMSE of 0.69. Improvements are observed over baseline models in terms of stability and minority-class representation.Conclusion:The proposed approach provides a robust and interpretable solution for multi-omics drug response prediction. It demonstrates potential for pharmacogenomics applications and offers a foundation for future graph-based deep learning models.