PRP: pathogenic risk prediction for rare nonsynonymous single nucleotide variants.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate prediction of pathogenic variants is essential for personalized medicine and genomic diagnosis.
• Rare nonsynonymous single nucleotide variants (nsSNVs) pose a challenge due to their low frequency and potential impact on protein function.

Purpose of the Study:

• To develop and validate PRP, a robust and interpretable prediction tool for rare nsSNVs.
• To compare PRP's performance against existing state-of-the-art prediction tools.

Main Methods:

• PRP utilizes 34 features across frequency, conservation, substitution, and gene intolerance metrics.
• Machine learning algorithms were optimized using Optuna, with feature importance analyzed via SHAP.
• The model was trained on ClinVar data and validated on three independent test sets.

Main Results:

• PRP consistently outperformed 20 other prediction tools across eight key performance metrics (AUC, AUPRC, Accuracy, F1-score, MCC, Precision, Recall, Specificity).
• The tool demonstrated high sensitivity and specificity without overestimating pathogenic variants.
• PRP showed robustness in predicting rare genetic variants.

Conclusions:

• PRP offers a reliable and interpretable solution for predicting pathogenic rare variants.
• This tool has the potential to significantly advance genomic medicine and personalized diagnosis.
• The datasets, code, and pre-computed scores for PRP are publicly available for research use.