Learning the mutational landscape of the cancer genome

An ongoing challenge to better understand and treat cancer is to distinguish neutral mutations, which do not affect tumor fitness, from those that provide a proliferative advantage. However, the variability of mutation rates has limited our ability to model patterns of neutral mutations and therefore identify cancer driver mutations. Here, we predict cancer-specific mutation rates genome-wide by leveraging deep neural networks to learn mutation rates within kilobase-scale regions and then refining these estimates to test for evidence of selection on combinations of mutations by comparing observed to expected mutation counts. We mapped mutation rates for 37 cancer types and used these maps to identify new putative drivers in understudied regions of the genome including cryptic alternative-splice sites, 5 prime untranslated regions and infrequently mutated genes. These results, available for exploration via web interface, indicate the potential for high-resolution neutral mutation models to empower further driver discovery as cancer sequencing cohorts grow.