FCMEDriver: Identifying Cancer Driver Gene by Combining Mutual Exclusivity of Embedded Features and Optimized Mutation Frequency Score

Efficiently identifying cancer driver genes is critical to drug design, cancer diagnosis and treatment. Current unsupervised cancer driver gene prediction approaches mainly exploit mutual exclusivity of mutated driver genes and integrate multi-omics data with gene function networks. Some of them identify driver genes based on the gene features learned by network embedding algorithms. However, these methods are limited to using the mutual exclusivity from original data without considering the mutual exclusivity implanted in the learned features. Additionally, they simply assume that all driver genes have high mutation frequencies. Thus, we propose a novel unsupervised framework FCMEDriver, which utilizes the mutual exclusivity from the learned features and mutation frequency to predict driver genes. In FCMEDriver, a feature clustering algorithm is designed to obtain modules. Based on the modules, our extensive experiments show that the Euclidean distances between learned features are highly related with the mutual exclusivity defined on the original data, and they can reveal more information compared to mutual exclusivity. Thus, we apply the Euclidean distances of learned gene features for each module to calculate a module importance score for each gene. Since the fact that most of driver genes have intermediate mutation frequencies, we design a mutation frequency scoring function for each gene to optimize the existing mutation frequency scoring strategy in which genes with intermediate mutation frequencies are more inclined to obtain similar high scores as those genes with high mutation frequencies. The weighted sum of the module importance score and the mutation frequency score is used to prioritize the genes. The experiment results show that FCMEDriver outperforms other four state-of-the-art methods for cancer driver identification.