Comparison of different functional prediction scores using a gene-based permutation model for identifying cancer driver genes

Background: Identifying cancer driver genes (CDG) is a crucial step in cancer genomic toward the advancement of precision medicine. However, driver gene discovery is a very challenging task because we are not only dealing with huge amount of data; but we are also faced with the complexity of the disease including the heterogeneity of background somatic mutation rate in each cancer patient. It is generally accepted that CDG harbor variants conferring growth advantage in the malignant cell and they are positively selected, which are critical to cancer development; whereas, non-driver genes harbor random mutations with no functional consequence on cancer. Based on this fact, function prediction based approaches for identifying CDG have been proposed to interrogate the distribution of functional predictions among mutations in cancer genomes (eLS 1-16, 2016). Assuming most of the observed mutations are passenger mutations and given the quantitative predictions for the functional impact of the mutations, genes enriched of functional or deleterious mutations are more likely to be drivers. The promises of these methods have been continually refined and can therefore be applied to increase accuracy in detecting new candidate CDGs. However, current function prediction based approaches only focus on coding mutations and lack a systematic way to pick the best mutation deleteriousness prediction algorithms for usage. Results: In this study, we propose a new function prediction based approach to discover CDGs through a gene-based permutation approach. Our method not only covers both coding and non-coding regions of the genes; but it also accounts for the heterogeneous mutational context in cohort of cancer patients. The permutation model was implemented independently using seven popular deleteriousness prediction scores covering splicing regions (SPIDEX), coding regions (MetaLR, and VEST3) and pan-genome (CADD, DANN, Fathmm-MKL coding and Fathmm-MKL noncoding). We applied this new approach to somatic single nucleotide variants (SNVs) from whole-genome sequences of 119 breast and 24 lung cancer patients and compared the seven deleteriousness prediction scores for their performance in this study. Conclusion: The new function prediction based approach not only predicted known cancer genes listed in the Cancer Gene Census (CGC), but also new candidate CDGs that are worth further investigation. The results showed the advantage of utilizing pan-genome deleteriousness prediction scores in function prediction based methods. Although VEST3 score, a deleteriousness prediction score for missense mutations, has the best performance in breast cancer, it was topped by CADD and Fathmm-MKL coding, two pan-genome deleteriousness prediction scores, in lung cancer.