Functional fine-mapping of noncoding risk variants in amyotrophic lateral sclerosis utilizing convolutional neural network

Ali Yousefian-Jazi,Jung Kyoon Choi,Taeyeop Lee,Yoon-Ho Hong,Jinwook Choi,Min Kyung Sung

doi:10.1038/s41598-020-69790-6

Abstract

Recent large-scale genome-wide association studies have identified common genetic variations that may contribute to the risk of amyotrophic lateral sclerosis (ALS). However, pinpointing the risk variants in noncoding regions and underlying biological mechanisms remains a major challenge. Here, we constructed a convolutional neural network model with a large-scale GWAS meta-analysis dataset to unravel functional noncoding variants associated with ALS based on their epigenetic features. After filtering and prioritizing of candidates, we fine-mapped two new risk variants, rs2370964 and rs3093720, on chromosome 3 and 17, respectively. Further analysis revealed that these polymorphisms are associated with the expression level of CX3CR1 and TNFAIP1, and affect the transcription factor binding sites for CTCF, NFATc1 and NR3C1. Our results may provide new insights for ALS pathogenesis, and the proposed research methodology can be applied for other complex diseases as well.

Highlights

Recent large-scale genome-wide association studies have identified common genetic variations that may contribute to the risk of amyotrophic lateral sclerosis (ALS)
We used the genetic associations from a large-scale genome-wide association studies (GWAS) metaanalysis including 8,697,640 single-nucleotide polymorphisms (SNPs) genotyped in 14,791 ALS patients and 26,898 healthy controls from 41 cohorts organized in 27 platform- and country-defined strata[3]
We reached to 274 association blocks carrying the lead SNPs and their nonoverlapped neighboring SNPs. (2) Annotate each SNP with functional features from four different categories (“Methods” section), DHS mapping data, histone modifications, target gene functions, and transcription factor binding sites (TFBS). (3) Train the convolutional neural network (CNN) model with uncertain labels (“Methods” section) on the extracted epigenetic feature map using a large number of hyperparameters and an autoencoder for pre-training

Summary

Introduction

Recent large-scale genome-wide association studies have identified common genetic variations that may contribute to the risk of amyotrophic lateral sclerosis (ALS). We constructed a convolutional neural network model with a large-scale GWAS meta-analysis dataset to unravel functional noncoding variants associated with ALS based on their epigenetic features. Genome-wide association studies (GWAS) have identified the common genetic variations that may contribute to the risk of ALS. We proposed a post-GWAS analysis method using a convolutional neural network (CNN) trained on epigenetic features to find functional rare noncoding risk variants[9]. The CNN model was constructed with uncertain class labels on the epigenetic feature map extracted from the largest available GWAS d ata[3] to predict functional noncoding variants associated with ALS

Methods

Results

Conclusion