MetaXVP: an interpretable machine learning framework for deep insight into variant pathogenicity and VUS...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Classifying nonsynonymous single nucleotide variants (nsSNVs) is crucial for clinical genomics but challenging, especially for variants of uncertain significance (VUS).
• Existing computational methods for variant pathogenicity prediction often lack sufficient performance and interpretability.

Purpose of the Study:

• To develop MetaXVP (Meta eXplainable XGBoost Variant Predictor), a novel computational tool combining high predictive accuracy with feature interpretability for nsSNV classification.
• To provide a robust and explainable model for classifying VUS and aiding clinical decision-making.

Main Methods:

• Developed MetaXVP using the XGBoost algorithm, trained on high-confidence pathogenic and benign nsSNVs from ClinVar.
• Integrated 32 features from dbNSFP, selected via correlation-based feature selection.
• Employed SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) for model interpretability.
• Evaluated performance on independent ClinVar test sets, a TP53 functional dataset, and a cancer somatic driver dataset using AUROC and AUPRC metrics.

Main Results:

• MetaXVP achieved high classification performance across various test sets, with AUROC values of 0.991 for pathogenic/benign classification and 0.986 for VUS reclassification.
• Demonstrated strong performance on a somatic cancer driver dataset (AUROC = 0.924).
• SHAP analysis identified ensemble predictors (ClinPred, gMVP) and population allele frequency (gnomAD) as key influential features.

Conclusions:

• MetaXVP offers a high-performing and interpretable solution for nsSNV classification, significantly advancing the interpretation of variants of uncertain significance.
• The model's robustness and feature explanations provide valuable insights for clinical genomics applications.
• Pre-computed scores for millions of nsSNVs are available to the research community.