Incorporation of clinical and molecular variant properties improves the performance of in silico pathogenicity prediction tools.

Abstract

In silico pathogenicity prediction tools performance can vary depending on molecular and clinical contexts. This study aims to assess the performance of commonly used tools under different conditions. Additionally, the study aims to recalibrate score thresholds to better reflect evidence of pathogenicity. ClinVar variants were stratified by allele frequency, conservation, mode of inheritance, and disease category. For each subset, Bayesian methods were used to recalibrate thresholds corresponding to the levels of evidence defined by the American College of Medical Genetics. Tools exhibited reduced accuracy for variants with higher allele frequencies and for variants located in regions with high conservation. Variants affecting autosomal recessive and X-linked genes were more accurately classified compared with those affecting autosomal dominant genes. Recalibrated thresholds consistently showed higher odds of correctly estimating pathogenicity (OR=3.78 [1.74, 8.55]; P < .001) and produced significantly higher scores for known pathogenic variants and lower scores for benign/likely benign. This improved discriminatory performance was particularly notable in variants found in low-conservation regions and autosomal recessive genes. Pathogenicity prediction tools should be evaluated using various variant subsets during development. Score threshold recalibration extends the range of evidence and improves overall pathogenicity probability estimation and classification.