Cancer is a heterogeneous disease, largely driven by the accumulation of somatic mutations arising from numerous mutational processes. While the majority of mutations are neutral passengers, a small subset, called driver mutations, contribute to cancer development. Elucidating these mutational processes and identifying driver mutations are crucial for understanding the complex mechanisms of cancer and for developing diagnostic and therapeutic strategies. Although large-scale sequencing projects have yielded extensive catalogs of somatic mutations, distinguishing driver mutations from abundant passenger mutations remains difficult. This challenge is further compounded by the lack of an objective gold standard for defining driver and passenger mutations, which limits the development and validation of novel computational methods. This dissertation aims to address these challenges through two complementary studies. First, I present a new computational approach to predict driver mutations in DNA polymerase ε (POLE). This approach is based on the observation that driver POLE mutations produce specific mutational fingerprints throughout genomes of cancer patients. By implementing this method, we have found that the presence of POLE driver mutations is strongly associated with the presence of specific sequence motifs in the whole-exome sequencing data — namely, G>T transversions in polypurine sequence contexts. As a result, we have developed a machine learning classifier to identify tumors that may carry unknown POLE driver mutations. We have discovered novel POLE drivers and confirmed their functional significance by the experimental analyses of mutator effects in yeast. Second, this dissertation contributes to constructing the most comprehensive benchmark of experimentally validated driver and passenger mutations. Using this benchmark, I present a study evaluating next-generation pathogenicity predictors, which are commonly used to identify disease-related mutations and are applied to detect driver mutations in cancer patients. The findings indicate that pathogenicity predictors demonstrate commendable performance even though they generally underperform compared to cancer-specific methods.