Systematic Evaluation of Functional and Computational Evidence for Large-Scale Resolution of Variants of Uncertain Significance

Abstract

Interpreting missense variants remains a major challenge in clinical genomics, with approximately 90% of observed missense variants classified as variants of uncertain significance (VUS). These uncertain classifications limit the clinical utility of genetic testing and hinder medical management for individuals. In this thesis I show that computational evidence is gene dependent: variant-effect predictors (VEPs) vary substantially across genes in their ability to distinguish benign and pathogenic variation. This gene-level heterogeneity demonstrates that gene-agnostic evidence thresholds can overestimate or underestimate evidence strength for clinical variant classification, motivating the development of gene-specific calibration frameworks within the IGVF Consortium. A central contribution of this thesis addresses the growing challenge of VUS, where I integrate calibrated large-scale functional measurements from multiplexed assays of variant effect (MAVEs) with gene-specific calibrated variant-effect predictions to build a scalable framework for resolving VUS. Applied across 40 clinically relevant genes, this framework resolves ~75% of 16,115 ClinVar VUS while maintaining an error rate of <1% among benign and pathogenic controls. In addition, applying this framework to >90,000 unobserved variants enable preemptive classification of 62% of variants before they are observed in clinical databases, further reducing the future burden of VUS. Together, these results show that production-scale functional data, combined with computational evidence, can substantially reduce VUS and help realize the promise of precision medicine.