Graph Convolutional Networks Improve the Prediction of Cancer Driver Genes

Despite the vast increase of high-throughput molecular data, the prediction of important disease genes and the underlying molecular mechanisms of multi-factorial diseases remains a challenging task. In this work we use a powerful deep learning classifier, based on Graph Convolutional Networks (GCNs) to tackle the task of cancer gene prediction across different cancer types. Compared to previous cancer gene prediction methods, our GCN-based model is able to combine several heterogeneous omics data types with a graph representation of the data into a single predictive model and learn abstract features from both data types. The graph formalizes relations between genes which work together in regulatory cellular pathways. GCNs outperform other state-of-the-art methods, such as network propagation algorithms and graph attention networks in the prediction of cancer genes. Furthermore, they demonstrate that including the interaction network topology greatly helps to characterize novel cancer genes, as well as entire disease modules. In this work, we go one step forward and enable the interpretation of our deep learning model to answer the following question: what is the molecular cause underlying the prediction of a disease genes and are there differences across samples?