Cancer predisposition genes (CPGs) are defined as those inherited variants that increase the risk of tumorigenesis. Despite their functional and clinical role in tumorigenesis, the number of identified CPGs remains limited. Due to the rarity of pathogenic germline variants and the limitations of traditional case-control approaches, the systematic approach to identify novel cancer predisposition genes is demanding. We developed a machine learning-based predictive model integrating 13 various biological features, including germline genomic mutation status, gene expression, somatic second hit mutation, and general characteristics of human genes, to identify novel candidates. While single features possessed moderate predictive performance, our integrative model outperformed all single-feature models (AUC =0.74 vs. 0.53-0.58), representing the value of multi-feature integration. We further applied the model across individual cancer types with sufficient sample sizes. Overall, we identified the top 8 high-confidence CPG candidates at both cancer-type-specific and pan-cancer levels. Several of the candidates were validated using an independent cancer cohort. This approach provides a generalizable and scalable framework for uncovering novel CPGs and offers a new direction for understanding the genetic basis of cancer susceptibility.

Co-authors: A-Reum Nam, Solip Park