The impact of machine learning models in reducing variants of uncertain significance for individuals from underrepresented populations who are undergoing testing for inborn errors of immunity

Abstract

The classification of missense variants is challenging due to limited evidence. Consequently, many remain categorized as variants of uncertain significance (VUS). VUS are at the core of healthcare disparities, as individuals from race, ethnicity and ancestry (REA) populations underrepresented in large genomic databases and medical literature receive VUS more often. To generate definitive genetic results more equitably across REA groups, we sought to develop gene-specific machine learning (ML) models and evaluate their utility in genetic testing for inborn errors of immunity (IEI).From 1/1/2022 to 11/15/2022, gene-specific ML algorithms were validated and integrated at Invitae. We incorporated existing models such as SpliceAI, and developed our own by leveraging large datasets including gnomAD, AlphaFold protein structures, and others, to model variant effects. Evidence from these ML models were incorporated into Sherloc, a semi-quantitative variant interpretation framework based on ACMG/AMP guidelines. Only evidence that met a negative or positive predictive value >80% was incorporated during variant interpretation. At least 1 validated model was available for 232 genes during the study period. VUS reduction was calculated, stratified by REA groups.Of 19 784 US based individuals who underwent multi-gene panel testing for IEI, ~16 900 (85%) had ML evidence applied to at least one variant, and ~5700 (29%) individuals were from an underrepresented REA group. Models contributed to the classification of at least 1 benign/likely benign variant in ~10 300 (52%) individuals and at least 1 pathogenic/likely pathogenic variant in ~210 (1%) individuals. A higher percentage of Black (67%), Asian (56%) and Hispanic (60%) individuals had a definitive classification (P/LP or B/LB) that was dependent on ML evidence compared to White (51%) individuals by one-tailed, two-sample proportion test (p <0.03). The average amount of benign or pathogenic ML evidence applied per reclassified variant by Sherloc was similar across populations. Among individuals who had at least 1 variant with ML evidence applied, over 60% would have received a VUS variant without this evidence. ML modeling has demonstrated utility during clinical variant interpretation for IEI, particularly in underrepresented populations.