Assessing the performance of 28 pathogenicity prediction methods on rare single nucleotide variants in coding regions.

backgroundAccurate pathogenicity prediction of rare variants in coding regions is crucial for prioritizing candidate variants in human diseases and advancing personalized precision medicine. Although many prediction methods have been developed, it remains unclear how they perform specifically on rare variants.

resultsIn this study, the performance of 28 pathogenicity prediction methods was assessed using the latest ClinVar dataset, with a focus on rare variants and various allele frequency (AF) ranges. Ten evaluation metrics were employed to comprehensively assess the predictive performance of each method. The methods were selected based on their training approaches, including whether the training dataset was filtered by AF and whether AF was incorporated as a feature. Most methods focused on missense and start-lost variants, covering only a subset of nonsynonymous SNVs. The average missing rate of approximately 10% was observed in these variants, indicating that prediction scores were unavailable for them. MetaRNN and ClinPred, which incorporated conservation, other prediction scores, and AFs as features, demonstrated the highest predictive power on rare variants. For most methods, specificity was lower than sensitivity. Across various AF ranges, most performance metrics tended to decline as AF decreased, with specificity showing a particularly large decline.

conclusionsThese results provide insights into the strengths and limitations of each method in predicting the pathogenicity of rare variants, which may guide future improvements in predictive models. Furthermore, while AF and existing prediction scores offer valuable information for prediction methods, the identification of novel biological features is essential to overcome current limitations and further improve predictive performance.