Abstract
Abstract only
1518
Background: Germline genetic analysis is an essential tool for implementing precision cancer prevention and treatment. However, only a small fraction of cancer patients, even those with features suggestive of a cancer-predisposition syndrome, have detectable pathogenic germline events, which may in part reflect incomplete pathogenic variant detection by current gold-standard methods. Here, we leveraged deep learning approaches to expand the diagnostic utility of genetic analysis in cancer patients. Methods: Systematic analysis of the detection rate of pathogenic cancer-predisposition variants using the standard clinical variant detection method and a deep learning approach in germline whole-exome sequencing data of 2367 cancer patients (n = 1072 prostate cancer, 1295 melanoma). Results: Of 1072 prostate cancer patients, deep learning variant detection identified 16 additional prostate cancer patients with clinically actionable pathogenic cancer-predisposition variants that went undetected by the gold-standard method (198 vs. 182), yielding higher sensitivity (94.7% vs. 87.1%), specificity (64.0% vs. 36.0%), positive predictive value (95.7% vs. 91.9%), and negative predictive value (59.3% vs. 25.0%). Similarly, germline genetic analysis of 1295 melanoma patients showed that, compared with the standard method, deep learning detected 19 additional patients with validated pathogenic variants (93 vs. 74) with fewer false-positive calls (78 vs. 135) leading to a higher diagnostic yield. Collectively, deep learning identified one additional patient with a pathogenic cancer-risk variant, that went undetected by the standard method, for every 52 to 67 cancer patients undergoing germline analysis. Superior performance of deep learning, for detecting putative loss-of-function variants, was also seen across 5197 clinically relevant Mendelian genes in these cohorts. Conclusions: The gold-standard germline variant detection method, universally used in clinical and research settings, has significant limitations for identifying clinically relevant pathogenic disease-causing variants. We determined that deep learning approaches have a clinically significant increase in the diagnostic yield across commonly examined Mendelian gene sets.