Performance Evaluation of in Silico Predictors for the Classification of ClinVar Variants

The purpose of this paper is to evaluate the performance obtained by the software tools (in silico) used by researchers and clinicians to determine the pathogenicity for the genetic variants. In order to evaluate their performance, variants that were reported pathogenic or benign were extracted from the ClinVar dataset. The variants were filtered by certain criteria and then they were annotated using specific tools (ANNOVAR and dbNSFP). In order to determine the performance, a series of statistical measures were extracted. The list of performance measures consisted of: specificity, sensitivity, precision, the harmonic mean (F
₁
score) and the regular mean between sensitivity and specificity. We found that the three predictors who had the highest sensitivity scores (CADD 96.2%, DANN 98.1%, M-CAP 96.8%) for determining the pathogenic variants, had mediocre specificity scores. The best overall performance was attributed to another three predictors (REVEL 91.5%, MetaSVM 85.5%, PolyPhen-2 HVAR 85%), which had the highest regular mean score. As future use, a combination of these tools is suggested.