Metaxvp: An Interpretable Machine Learning Framework for Deep Insight Into Variant Pathogenicity and Vus Classification Publisher Pubmed

Summary: Can AI clarify uncertain genetic variants? MetaXVP combined prediction features to classify missense variants and released scores for about 83 million possible changes. #Genomics #MachineLearning

Dehghan Tezerjani M ; Sehhati M ; Tabatabaiefar M A
Source: Scientific Reports Published:2026

Abstract

The interpretation and classification of nonsynonymous single nucleotide variants (nsSNVs) remains a significant challenge in clinical genomics particularly for variants of uncertain significance (VUS). Although many computational methods have been developed for predicting variant pathogenicity, their performance remains limited and often lacks interpretability. To address these limitations, we developed MetaXVP (Meta eXplainable XGBoost Variant Predictor),designed to combine high predictive performance with effective feature explanation using rigorous methodology. MetaXVP was built using the XGBoost algorithm and trained on high-confidence pathogenic and benign nsSNVs from ClinVar (submissions up to April 2024). Feature annotations were obtained from dbNSFP v4.7 (last updated March 2024) where 32 features spanning conservation metrics, functional scores, ensemble predictors, and allele frequencies were selected through correlation-based feature selection. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) methods were also used to make the model interpretable. Final evaluation was conducted on four independent ClinVar test sets (isolated from training data), a TP53 functional dataset, and a cancer somatic driver dataset using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC) and evaluation metrics based on Bayesian threshold calculations. In addition, the performance was compared with 32 individual features and 6 benchmark tools. MetaXVP achieved high classification performance across ClinVar test sets (AUROC = 0.991 for pathogenic/benign and 0.986 for VUS reclassified to pathogenic/benign), and somatic cancer test set (AUROC = 0.924). In addition, the stability of the Bayesian decision threshold was evaluated using bootstrap resampling, demonstrating robustness to sampling variability. SHAP analysis also revealed that ensemble predictors such as ClinPred and gMVP; and population allele frequency (gnomAD) were among the most influential features in the model based on mean absolute SHAP values. Pre-computed MetaXVP scores for all potential human nsSNVs (~ 83 million) are provided at https://github.com/Masouddehghant/MetaXVP for community use, which is particularly helpful for classification of VUS. © The Author(s) 2026.