Beyond Genome-Wide Predictors: A Domain-Specific Model for ABCA4 Variant Interpretation

Background Inherited retinal degenerations (IRDs) are genetically heterogeneous, with over 270 implicated genes, roughly one-third of which encode membrane-bound proteins. Missense variant interpretation is particularly difficult in this subset, and ABCA4 exemplifies the challenge: nearly half of its missense variants remain classified as variants of uncertain significance (VUS), limiting diagnostic clarity and delaying eligibility for emerging gene-targeted therapies. Under ACMG/AMP guidelines, computational evidence (PP3/BP4) plays a key role in variant classification, but genome-wide predictors are trained broadly across genes and may not adequately capture the structurally complex, partially disordered regulatory domains (RD1 and RD2) of ABCA4 . To address this gap, we benchmarked eight genome-wide pathogenicity predictors on a curated dataset of membrane-associated retinal degeneration genes and then developed a domain-specific machine learning model, that uses a random forest classifier, focused on the regulatory domains of ABCA4 , with performance evaluated using cross-validation and permutation testing. Results During benchmarking, MetaRNN showed the strongest overall performance among the genome-wide tools evaluated. We then assessed our ABCA4-loc alized model, which was trained on a small, labeled dataset (n = 18), and achieved complete discrimination between pathogenic and benign variants, with robust performance under permutation testing. This indicates that it captured meaningful biological signals rather than memorizing the training data. Limited training data currently prevent the model from independently supporting PP3 or BP4 evidence under ACMG/AMP guidelines; however, it effectively prioritizes high-impact variants for follow-up when combined with genome-wide tools and protein structural analysis. Conclusion Genome-wide and localized predictors serve complementary roles in variant interpretation. Genome-wide tools provide broad, well-powered assessments, whereas localized models capture gene- and domain-specific biological features. Together, these findings support a new in-silico pipeline that integrates genome-wide screening with domain-specific modeling and structural analysis, offering a practical framework for prioritizing and interpreting missense variants in ABCA4 and other disease-associated genes.