UshEffect-3D: Structure-informed Classification of USH2A Missense Variants for Inherited Retinal Disease

PurposeVariants of uncertain significance (VUS) in USH2A represent a critical interpretive challenge in inherited retinal disease, with over 70% of ClinVar submissions for this gene currently unresolved. We aimed to develop a gene-specific, structure-informed machine learning framework to improve the clinical classification of USH2A missense variant and provide a tractable tool to aid the diagnosis of Usher Syndrome II. MethodsA dataset of 545 curated USH2A missense variants with established clinical classifications was assembled from ClinVar and LOVD. AlphaFold2-predicted domain structures were used to generate local structural descriptors and biochemical features combined with sequence-based evolutionary conservation scores, yielding 153 candidate features reduced to nine via sequential feature selection. Eleven machine learning classifiers were trained using a 10-fold cross-validation strategy, then independently assessed on a blind test set and validated against 78 ACMG-classified pathogenic variants. Model predictions were benchmarked against five general-purpose variant effect predictors and applied to 2639 USH2A VUS from ClinVar. Feature contributions were analysed using SHAP analysis and ablation studies. ResultsThe Random Forest classifier achieved the highest performance on the blind test set, with an MCC of 0.87 and AUC of 0.97. On independent ACMG validation, sensitivity reached 0.73 with perfect precision. UshEffect-3D substantially outperformed all general-purpose predictors, including PolyPhen-2 (MCC = 0.61), AlphaMissense (MCC = 0.42), and ESM-1b (MCC = 0.32). SHAP analysis identified evolutionary conservation as a dominant predictor, with structural stability providing an independent but complementary signal. Applied to 2639 ClinVar VUS, the model prioritised 888 variants (33.6%) as likely pathogenic, particularly enriched within the Laminin N-terminal and Laminin G-like domains. ConclusionsUshEffect-3D demonstrates that gene-specific, structure-informed machine learning substantially outperforms general-purpose variant effect predictors for USH2A missense variant interpretation. This framework provides a high-confidence prioritization resource for the large unresolved VUS burden in this gene to facilitate earlier molecular resolution of USH2A-associated disease. As genedirected therapies for USH2A-associated retinal disease advance toward clinical application, accurate and interpretable variant classification will be essential for equitable patient selection. UshEffect-3D is freely accessible via an interactive web server.