AlphaRING: Missense Variant Classification via Protein Modelling and Residue Interaction Network Analysis

Abstract

BackgroundAccurate interpretation of missense variants remains a significant challenge hindering genomic diagnosis. While state-of-the-art machine learning and deep learning predictors offer high accuracy, they often lack the transparency required for clinical evidence weighting. To bridge this gap between accuracy and clinical utility, we developed AlphaRING-X. MethodsAlphaRING-X is an open-source framework that integrates local and global protein structural stability features using an efficient implementation of gradient boosting, XGBoost, to predict variant deleteriousness. Local stability is interrogated through a combination of AlphaFold-based modelling and residue interaction network analysis. Changes in global stability upon mutation ({Delta}{Delta}G) are estimated using FoldX. For each prediction, AlphaRING-X quantifies the magnitude and direction of each features contribution using Shapley additive explanations (SHAP), providing a unified yet interpretable score. We trained and evaluated the model using a gold standard ClinVar dataset of validated neutral and deleterious variants. ResultsAlphaRING-X achieved an area under the receiver operator curve of 0.94, significantly outperforming widely used predictors such as CADD. Crucially, it provided unambiguous classification for 95.5% of variants at 90% precision in both neutral and deleterious classes. We analysed each predictions SHAP values, identifying local connectivity and disorder as the strongest contributors. ConclusionsAlphaRING-X combines high-performance prediction with the interpretability necessary for clinical decision-making. Its flexible, open-source architecture makes it a powerful, transparent tool for enhancing genomic diagnostics and advancing precision medicine.