Session # 18, Abstract Program #82

 Gene-specific allele frequency thresholds for benign evidence to empower variant interpretation

•Allele frequency is typically used as evidence for variant pathogenicity, but current assessments are based on fixed or prevalence-adjusted thresholds which do not adequately discriminate benign and pathogenic missense variants for all genes.

•We developed a gene-specific allele frequency threshold approach that enabled us to assign benign evidence to more missense variants at a lower false classification rate than exiting methods.

 

Allele frequency is often used as evidence of whether a variant is likely to be causative for a rare disease. However, current assessments of allele frequency for variant classification rely on either fixed or prevalence-adjusted thresholds and make no use of gene-specific information on variant pathogenicity. The publicly available Genome Aggregation Database provides an unprecedented spectrum of population-based human genetic variation that can be leveraged to examine the relationship between allele frequency and variant pathogenicity in classified variants on a gene-by-gene basis. Using a cohort of 176,280 patients who underwent genetic testing at a single diagnostic laboratory in 2012-2016, we assembled a training dataset of 1,388 classified missense variants in 10 genes (BRCA1, BRCA2, CDH1, PALB2, PTEN, TP53, MLH1, MSH2, MSH6 and PMS2) included on hereditary cancer panels. The number of variants per gene ranged 44 to 438. We developed a constrained distribution fitting (CDF) approach to quantify gene-specific allele frequency thresholds (AFT) using data mining techniques of bounded constraints, monotonic distribution fitting and bootstrap sampling, each targeted to conservative estimates in case of uncertainty. Across variants in all 10 genes, positive predictive values (PPVs) for benign evidence were 0.40 ± 0.33, 0.43 ± 0.33 and 0.47 ± 0.33 using AFTs designated by a fixed 1% cutoff, prevalence-adjusted method, and CDF method, respectively. Negative predictive values (NPVs) were 1.00 ± 0.00 for fixed 1% cutoff and CDF methods, and 0.88 ± 0.31 for prevalence-adjusted method. Thus, the AFTs estimated by CDF showed 9% to 17% higher PPV than the other two methods and 12% higher NPV than the prevalence-adjusted method. Notably, differences between gene-specific AFTs estimated by CDF vs. other methods were striking for several genes. For example, using the CDF method, the AFT for benign evidence was as low as 0.0019% for CDH1, due to extremely rare pathogenic/likely pathogenic variants and as high as 0.43% for PMS2 due to pathogenic/likely pathogenic variants having a wide range of frequencies. In contrast, AFTs estimated by the prevalence-adjusted method were nearly identical from 0.060% to 0.063% for both genes. Our results underscore the tremendous need for and practical usefulness of gene-specific allele frequency thresholds for benign evidence to empower variant interpretation.

• Speakers: Dajun Qian
• Conference: ASHG 2017
• Date: Wednesday, Oct 18, 2017 12:15pm - 12:30pm