MGTDR: A Multi-modal Graph Transformer Network for Cancer Drug Response Prediction

Drug response prediction in cancer cell lines can guide researchers to design personalized treatments for different patients. However, accurately predicting drug response remains a challenging task. This study proposes MGTDR, a multi-modal graph transformer framework for drug response prediction. First, using an auto-encoder, MGTDR learns the latent features of cancer cell lines. Secondly, It employs graph convolutional neural networks (GCN) and multi-layer perceptrons (MLP) to understand features of drugs from the simplified molecular input line entry specification (SMILES) and molecular fingerprints of drugs. Thirdly, it utilizes miRNA expression, DNA methy-lation, and drug physicochemical properties to calculate cell line similarity and drug similarity. Subsequently, it constructs a heterogeneous network by combining cell line similarity and drug similarity. The cell line features and drug features calculated earlier are then employed as the features of the nodes in the network. Finally, it applies graph transformer networks and MLP to predict drug sensitivity. Extensive experiments on publicly available datasets demonstrated the effectiveness and efficiency of the proposed method in predicting drug response and its potential value in guiding personalized therapy.