Poster #88 - Zachary Davis

Machine Learning Approaches to Prioritize Variants of Uncertain Significance in the Regulatory Domains of ABCA4

Davis, Z. University of Delaware, Center for Bioinformatics and Computational Biology Biswas, S. University of Delaware, Department of Medical and Molecular Sciences, College of Health Sciences Biswas-Fiss, E. University of Delaware, Department of Medical and Molecular Sciences, College of Health Sciences

This study leverages supervised machine learning to classify and prioritize variants of uncertain significance (VUS) in ABCA4, a key gene implicated in inherited retinal degenerations. By integrating labeled variant datasets with ensemble predictors and leave-one-out cross-validation, this framework identifies high-priority variants for further investigation and potential clinical interpretation. Purpose: Missense variants in ABCA4 are a major cause of inherited retinal degenerations such as Stargardt disease, yet nearly half of reported variants are classified as VUS or have conflicting interpretations of pathogenicity. This uncertainty complicates clinical decision-making and genetic counseling. Accurate computational classification is critical for prioritizing variants for experimental validation and improving patient care. Approach: A labeled dataset of 15 ABCA4 regulatory domain variants (11 Pathogenic/Likely Pathogenic, 4 Benign/Likely Benign) was used to train random forests with Leave-one-out cross-validation (LOOCV) implemented to ensure robust evaluation with limited data. The best-performing models were applied to predict pathogenicity for 65 VUS and 14 variants with conflicting interpretations. Previous studies noted challenges in predicting variants within intrinsically disordered regions, and these limitations were assessed here. Results: The random forest model achieved perfect classification under leave-one-out cross-validation, outperforming commonly used genome-wide pathogenicity predictors. When applied to unresolved variants, the model was able to classify VUS as high-priority, providing a focused set of candidates for downstream study. Structural visualization in PyMOL supported these predictions, with likely pathogenic variants showing steric clashes or destabilizing effects, while variants trending toward benign showed no such disruptions, which reinforces the validity of this approach. Conclusions: This work demonstrates the potential of machine learning in resolving uncertainty around ABCA4 variants. By prioritizing VUS and conflicting variants and benchmarking against functional data, this approach provides a pathway for improved variant interpretation and supports future experimental validation.