Benchmarking AlphaMissense pathogenicity predictions against APP, PSEN1, and PSEN2 variants of unknown significance

Alzheimer's disease (AD) is a progressive neurodegenerative disease characterized by progressive cognitive decline. Over 200 pathogenic mutations in amyloid-β precursor protein (APP), presenilin-1 (PSEN1), and presenilin-2 (PSEN2), have been implicated in AD. Yet, many rare and common variants have not been completely classified as protective or benign, risk-modifiers, or pathogenic, which is important for research on the disease mechanisms and discovery of treatment methods. The majority of these variants are missense mutations, and there is an active need for computational approaches to accurately predict their molecular consequences. AlphaMissense (AM) is a novel technology that uses population frequency data along with structural and sequential contexts from AlphaFold to predict the pathogenicity of missense mutations. Herein, we sought to evaluate the capabilities of AM on 114 variants of unknown significance (VUS), including 56 missense variants of PSEN1, 25 of APP, and 33 of PSEN2 by benchmarking its prediction against their respective Aβ isoform levels in vitro, respectively. We found that the AM scores correlated moderately well with the critical Aβ42/Aβ40 biomarker and Aβ40 levels in the transmembrane proteins compared to weaker correlations in traditional approaches, including Combined Annotation Dependent Depletion (CADD) v1.7, evolutionary model of variant effect (EVE), and Evolutionary Scale Modeling-1b (ESM-1B). Yet, there were non-significant correlations identified with Aβ42 levels in all models. Furthermore, we found that AM does not rely completely on structural contexts from AlphaFold2, as it accurately predicted the effects of known variants on residues with a low predicted local distance difference test (pLDDT) score. Additionally, based on the receiver operating characteristic-area under the curve analysis (ROC-AUC), we found that AM retained a high performance on 263 validated variants of these amyloidogenic genes, and performed the greatest compared to other models for the 114 VUS. We believe this is the first study to provide comprehensive characterization and validation of AM in comparison to the widely utilized pathogenicity scoring models for VUS involved in proteins implicated in AD.