Abstract
PURPOSE: AlphaMissense is a new artificial intelligence-based approach to predicting the pathogenicity of missense variants. However, whether its predictions can be directly applied to clinical decision making remains unclear. This study aimed to evaluate the accuracy of AlphaMissense predictions for DNA damage repair (DDR) genes using genomic and clinical characteristics. METHODS: Sequencing data from 56,965 patients with cancer who underwent Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT) testing between April 2015 and March 2023 and data from The Cancer Genome Atlas (TCGA)/Pan-Cancer Analysis of Whole Genomes (PCAWG) were analyzed. AlphaMissense pathogenicity was evaluated in six commonly mutated DDR genes. Missense mutations were classified into three categories on the basis of AlphaMissense and OncoKB: known pathogenic, newly identified pathogenic by AlphaMissense, or benign. RESULTS: In the MSK-IMPACT cohort, 1,182 (17.5%) of 6,743 unique DDR gene missense mutations were newly identified as pathogenic. In breast, ovarian, pancreatic, and prostate cancers, homologous recombination (HR)-deficiency signatures were more common in tumors with new pathogenic missense mutations in BRCA1/2, PALB2, and RAD51C than in benign missense mutations (66.7% v 35.2%, P = .021). In the TCGA/PCAWG data set, HR-deficiency signatures were also more frequent in tumors with new pathogenic mutations than in wild-type (46.2% v 19.2%, P = .036). For tumors with POLE missense mutations, there were no significant differences in tumor mutation burden and POLE-associated signatures between new pathogenic and benign mutations. Patients with new pathogenic ATM missense mutations had fewer TP53 mutations (30.5% v 54.6%, P < .001) and showed improved irradiated tumor control (hazard ratio, 0.58 [95% CI, 0.35 to 0.95]; P = .03) compared with those with benign missense mutations. CONCLUSION: Our findings suggest that AlphaMissense can help identify previously unknown pathogenic DDR gene mutations, but its accuracy is gene-dependent. AlphaMissense prediction still requires additional confirmation with clinical and functional validation.