SeverYX: An ensemble framework for robust variant severity classification

Abstract

Accurate interpretation of genetic variants in clinical genomics depends on reconciling multiple computational predictors that often disagree or provide incomplete evidence. We present SeverYX, a transparent, rule-based ensemble framework that consolidates conservation, splicing, and functional-impact predictors into a three-tier severity classification ('High', 'Medium', 'Low') through an effect-aware decision architecture. We benchmarked SeverYX against a curated ClinVar dataset of 270,208 MANE-transcript variants (86,161 pathogenic/likely pathogenic, 184,047 benign/likely benign), stratified by molecular consequence, review status, and haploinsufficiency. The baseline configuration achieved 97.6% concordance with ClinVar; at a fixed High cut, 90.0% of pathogenic and 2.0% of benign variants were assigned High. Concordance was highest for loss-of-function variants and lowest for non-coding and unclassified pathogenic variants, reflecting known gaps in current predictive tools. Systematic threshold optimization, combined with the addition of AlphaMissense and REVEL as missense predictors, raised overall concordance to 98.7%, reduced the benign High rate to 0.4% (at a cost of pathogenic High sensitivity falling to 88.3%), and improved the Matthews correlation coefficient from 0.895 to 0.908. These results show that SeverYX offers a transparent, reproducible and extensible approach for consolidating heterogeneous variant-effect evidence into clinically interpretable severity calls, and a scalable path for incorporating future predictive models into genomic interpretation workflows.