MDAGCN: Predicting Mutation-Drug Associations Through Signed Graph Convolutional Networks via Graph Sampling

The surge in accessible high-throughput molecular data presents computational challenges for the precision medicine in cancer. Genetic mutations have the potential to act as reliable biomarkers indicating responses to targeted drugs. Accurate prediction of mutation-drug associations is critically important for drug development and cancer treatment. We propose a novel graph convolutional network method, MDAGCN, to predict the mutation-drug associations with specific types (sensitive/resistant) in cancer. To enhance both the efficiency and accuracy of training, we begin by constructing a feature and topological graph using the k-Nearest Neighbors algorithm, incorporating the structural relationship and feature data associated with mutation-drug interactions. Experimental results show that MDAGCN outperforms state-of-the-art methods in different experimental settings. Moreover, we show the effectiveness of graph sampling technique for training signed graphs. MDAGCN is a comprehensive end-to-end framework that could be broadly applicable to cancer pharmacogenomics. This framework has the potential to facilitate the mapping from the discovering novel mutation-drug associations to in-depth analysis of drug sensitivity and resistance.