Deep Learning and Experimental Validation of BRCA1-Associated Protein-1 Missense Variant Pathogenicity in Uveal...

To benchmark the pathogenicity predictions of AlphaMissense, a deep learning model, against high-throughput functional scores from saturation genome editing (SGE) for all experimentally tested BRCA1-associated protein-1 (BAP1) missense variants in uveal melanoma (UM).

Cross-sectional analytical study comparing computational predictions with experimentally derived functional classifications and clinical annotations.

ClusteredAll 4619 BAP1 single amino acid substitutions profiled in a published Regularly Interspaced Short Palindromic Repeats-Cas9 SGE viability assay. No separate control cohort was required. Clinical Variant Database (ClinVar)-annotated variants within this set served as an independent reference.

Saturation genome editing log₂-fitness scores were dichotomized at a validated depletion threshold. Prediction tools AlphaMissense, Rare Exome Variant Ensemble Learner (REVEL), Meta-predictor based on a Recurrent Neural Network (MetaRNN), and Combined Annotation Dependent Depletion (CADD) scores were aligned to SGE classifications. Diagnostic performance was assessed using receiver operating characteristic (ROC) and precision-recall (PR) analyses. Gene-specific threshold optimization for AlphaMissense was performed using Youden index. Structural mapping was performed using the AlphaFold2 BAP1 model.

(1) Concordance between in silico predictors and SGE labels (area under ROC and PR curves); (2) agreement with ClinVar classifications; (3) structural clustering of high-risk residues.

Of the 4619 BAP1 missense variants tested by SGE, 988 (21.4%) were experimentally classified as pathogenic. Using a default genome-wide threshold of 0.564, AlphaMissense predicted 2563/4618 variants (55.5%) as pathogenic. Receiver operating characteristic analysis showed good overall performance, with an area under the curve of 0.837, indicating that AlphaMissense pathogenic calls were the most strongly associated with experimentally defined pathogenicity when compared to REVEL, MetaRNN, and CADD. Receiver operating characteristic analysis informed a BAP1-specific optimized threshold of 0.952, improving sensitivity-specificity balance and identifying 1556/4618 variants (33.7%) as pathogenic. Of the 13 851 possible BAP1 missense variants, AlphaMissense predicted 8694 and 5816 as pathogenic by the genome-wide and optimized thresholds, respectively. Structural mapping highlighted clusters of high-risk substitutions within the catalytic core and protein-interaction motifs of the BAP1 protein. Finally, AlphaMissense predictions matched 11 of 12 ClinVar-classified variants (91.7%).

AlphaMissense aligns closely with the gold standard functional data for BAP1, providing a rapid, scalable, and interpretable approach to variant classification in UM. This study introduces an optimized, gene-specific threshold that further enhances its precision, supporting the integration of artificial intelligence-based tools into ocular oncology precision-medicine pipelines for risk stratification and clinical decision-making.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.