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

Area of Science:

• Genomics and Precision Medicine
• Cancer Genetics
• Machine Learning in Healthcare

Background:

• Germline pathogenic variants (GPVs) are often found incidentally during cancer genetic testing.
• Limited data on germline conversion rates (GCRs) in Japan necessitates better predictive tools.
• Current clinical decisions for GPVs in Japan often rely on European Society for Medical Oncology (ESMO) criteria.

Purpose of the Study:

• To develop a variant-level algorithm predicting GCRs in the Japanese population.
• To utilize a Japanese tumor-normal matched panel database for algorithm development.
• To compare the new algorithm's clinical utility against existing standards like ESMO criteria.

Main Methods:

• Analysis of 7,078 Japanese cases from the NCC Oncopanel dataset, focusing on 32 hereditary cancer genes.
• Incorporation of clinical features, sample data, sequence results, and minor allele frequency (MAF) into a machine learning model and nomogram.
• Assessment of clinical utility via decision curve analysis and validation using the GenMineTOP dataset.

Main Results:

• The developed model achieved a high predictive accuracy (c-index 0.96-0.97), significantly outperforming ESMO criteria (0.88).
• High disease-specific GCRs were noted in BAP1, BRCA1/2, and NF1 genes, with several genes showing >50% GCRs.
• Key predictors included age, multiple cancers, gene type, cancer type, MAF, and tumor allele ratio (TAR).

Conclusions:

• Variant-level prediction models incorporating TAR and MAF enhance GPV prediction accuracy in Japanese cancer patients.
• The developed algorithm offers superior clinical utility compared to gene-level approaches and ESMO criteria.
• This predictive model supports improved clinical decision-making and advances personalized medicine for hereditary cancers.