Learning Cancer Genes through Mutation Impact

Compiling the catalogue of genes actively involved in tumorigenesis (known as cancer drivers) is an ongoing endeavor, with profound implications to the understanding of tumorigenesis and treatment of the disease. An abundance of computational methods have been developed to screening the genome for candidate driver genes based on genomic data of somatic mutations in tumors. Most methods rely on detecting genes displaying excessive mutation rates compared to some background model. This approach is susceptible to false discoveries, due to its sensitivity to the assumptions of the background model, such as the need to account for hyper-mutated samples, cancer types and genomic loci. We present a fundamentally different approach. Instead of focusing on the number of mutations, we examine their content, and their expected effects on the functions of genes. We use a machine-learning model to predict functional effect scores of somatic mutations. For each gene, we compare the distribution of observed effect scores with the distribution expected at random, and report genes showing significant bias. By applying our framework on the ~20k protein-coding human genes, we detected 593 genes showing significant bias towards harmful mutations in the context of cancer. In contrast, we found only 6 significant genes biased in the opposite direction. The list of 593 genes, constructed without any prior knowledge of their role in cancer, shows an overwhelming overlap with known cancer driver genes, but also highlights many overlooked genes. The genes identified by our framework, including the overlooked ones, are strongly predictive of patient survival. These overlooked genes are promising candidates for novel cancer drivers. Our model is generic and is not restricted to the context of cancer. Applying the same framework to data of human-population genetic variation reveals the opposite trend. Unlike cancer, which is dominated by a bias towards harmful mutations, long-term evolution in healthy individuals results a bias towards less harmful mutations. The underlying assumptions of our framework are minimal, making it ideal for analyzing genetic data in search of genes subjected to positive or negative selection. It is fully open sourced and available for installation and use. Our framework presents a substantial development towards the application of state-of-the-art machine-learning algorithms in genetic studies. This work is accessible in bioRxiv (doi: https://doi.org/10.1101/242354).