Germline Pathogenic Variant Prediction Model for Tumor-Only Sequencing Based on Japanese Clinicogenomic Database

Germline pathogenic variants (GPV) are frequently identified as secondary findings in cancer gene panel testing. Due to limited data on germline conversion rates (GCR) in the Japanese population, clinical decisions have relied on European Society for Medical Oncology (ESMO) criteria. We aimed to develop a variant-level GCR prediction algorithm using a Japanese tumor-normal matched panel database and compare its utility with existing standards.

We analyzed 7,078 Japanese cases from the NCC Oncopanel dataset, focusing on 32 hereditary cancer genes. Clinical features, sample information, sequence results, and minor allele frequency (MAF) in healthy populations were incorporated into a machine learning model and nomogram. Clinical utility was assessed via decision curve analysis and validated using the GenMineTOP dataset.

Among 3,372 cases (mean age 61; 51% male), 4,905 pathogenic variants were identified, including 491 GPV (GCR: 10%). High disease-specific GCR were observed in BAP1 (11% in ocular tumors), BRCA1 and/or BRCA2 (13%-16% in ovarian/peritoneal cancers), and NF1 (16% in peripheral nerve tumors). Genes with >50% GCR included RAD51C, BRCA1, PALB2, CHEK2, RET, BRCA2, and PMS2. Significant predictors included age <30, multiple cancers, gene type, cancer type, MAF, relative variant allele frequency to tumor purity, and tumor allele ratio (TAR). The model achieved a c-index of 0.96 to 0.97, outperforming ESMO (0.88), with a 1.2% net benefit at a 5% threshold. The results were confirmed using the GenMineTOP dataset.

Variant-level prediction models for Japanese patients with cancer incorporating TAR and MAF offer improved GPV prediction over gene-level approaches and support clinical decision-making and personalized medicine.