Gene-based calibration of high-throughput functional assays for clinical variant classification

High-throughput functional assays measure the effects of variants on macromolecular function and can aid in reclassifying the rapidly growing number of variants of uncertain significance. Under the current clinical variant classification guidelines, using functional data as a line of evidence to assert pathogenicity relies on determining assay score thresholds that define variants as functionally normal or functionally abnormal. These thresholds are designed to maximize the separation of variants with known clinical effects (benign, pathogenic) and often incorporate expert opinion. However, this approach lacks the rigor of calibration, in which a variant's posterior probability of pathogenicity must be estimated from the raw experimental score and mapped to discrete evidence strengths. To build upon the existing guidelines, we introduce and evaluate a method for calibrating continuous high-throughput functional data as a line of evidence in clinical variant classification. Assay score distributions of synonymous variants and variants appearing in gnomAD for a given functional scoreset are jointly modeled with score distributions of known pathogenic and benign variants using a multi-sample skew normal mixture of distributions. This model is learned using a constrained expectation-maximization algorithm that provably preserves the monotonicity of pathogenicity posteriors and is subsequently used to calculate variant-specific evidence strengths for use in the clinic. Using 24 datasets from 14 genes, we first assess the model's ability to capture assay score distributions. We then demonstrate its potential impact on reclassifying variants by comparing the evidence strengths assigned at the variant-level with those assigned uniformly to all functionally normal and abnormal variants under the existing ClinGen guidelines. An improved classification of variants will directly improve the accuracy of genetic diagnosis and subsequent medical management for individuals affected by Mendelian disorders.

https://github.com/dzeiberg/mave_calibration.