Identification of SNP interactions in chromosome 19 for Alzheimer’s Disease through Capsule Network

Abstract Background Alzheimer’s disease (AD) is highly heritable. Genome‐wide association studies and whole genome sequencing have uncovered many AD‐risk variants but are still falling short from providing an explanation of the full heritability. To address this issue, we devised a novel deep learning approach that can quantify the polygenic risk score (PRS) of genetic variants and their interactive effects at the individual level. Method We applied a novel deep learning model to the 266,161 SNP chromosome 19 data of 313 AD and 457 cognitively unimpaired (CU) ADNI patients. Through occlusion maps, we quantified the predictive value of 70,643 possible AD‐determining SNPs and identified the top 0.05%, i.e., 35 SNPs, without imposing any additional pre‐assumptions. Epistatic effects increasing or decreasing the risk of AD were further investigated by using the co‐occlusion method. The top 35 AD‐determining SNPs were validated in an external dataset. Result The model achieved 68.18% in predicting AD vs. CU. The top 35 AD SNPs showed dominant impact in AD prediction as single as well as an interactive factors. 574 interactions that increased AD risk were found for AD‐predicted patients while 358 interactions predictive of cognitively unimpaired performance were found for CU‐predicted patients. Rs56131196 (APOC1) was recognized as the most significant AD‐determining SNP. It showed interactions with at least 8 other SNPs that further influenced the risk of AD. Rs2229918 (EERC1) was recognized as the most powerful SNP influencing CU diagnostic accuracy. Conclusion Our deep learning model provides a novel approach for assessing the genetic risk for AD at the individual chromosomal level. The results indicate that APOC1 and ERCC1 are strong genetic risk factors for the development of AD. These findings are consistent to the previous medical research. Our model’s capability in deciphering the impact of SNPs on AD development could prove critical for early presymptomatic risk assessment and personalized therapeutic strategy approaches.