A Simulative Deep Learning Model of SNP Interactions on Chromosome 19 for Predicting Alzheimer's Disease Risk and Rates of Disease Progression

Abstract

Background: Understanding Alzheimer’s disease (AD) genetic dynamics is key to unraveling its pathophysiology and advancing precision medicine. Current genetic studies, however, fall short in analyzing epistatic interactions between single nucleotide polymorphisms (SNPs). Here, we introduce a novel capsule network–based deep learning framework designed to model and quantify these complex SNP–SNP interactions on AD risk. Methods: In Chapter 1, we developed a novel deep learning model that can examine epistatic interactions of SNPs. Chromosome 19 genetic data from ADNI and ImaGene were used. Their epistatic impacts on AD development were quantified and the top 35 AD-risk SNPs were identified. In Chapter 2, we explored the clinical utility of the top 35 AD-risk SNPs. We performed computational gene-editing simulations, i.e., substituting each risk allele with its reference counterpart, to estimate how these edits would alter an individual’s Alzheimer’s risk. Further, we correlated each SNP’s quantified impact with the rate of cognitive declines and cerebrospinal fluid proteins changes using regression analysis. Results: The model was successfully trained and mapped genetic dynamics of AD in chromosome 19. Rs561311966 (APOC1) and rs2229918 (ERCC1) emerged as the strongest AD-risk SNPs. Computational gene-editing simulation with rs56131196 reduced the likelihood of AD by 7.9%, converting 36% of predicted AD participants to cognitive unimpaired individuals. Regression analyses using the top 35 SNPs yielded significant associations (p < 0.05) with disease progression, with the strongest correlations observed for executive function decline (adjusted r² = 0.433) and the ratio of amyloid beta over total tau change. (adjusted r² = 0.973). Discussion: Our model provided a comprehensive view of SNP interactions on chromosome 19 underlying Alzheimer’s development in a fully hypothesis free manner. The top 35 risk variants formed clusters in six genes: APOC1, TOMM40, ZNF473, VRK3, ERCC1 and APOC2. Multiple biological and clinical studies demonstrate that variants in these genes contribute to Alzheimer’s pathology, particularly through oxidative stress related mechanisms. We quantified the individual impact of these risk variants, enabling in-silico gene editing simulations and prediction of disease progression. This work has the potential to transform preventive precision medicine.