OpenLB Open Library of Bioscience

RefRGim: an intelligent reference panel reconstruction method for genotype imputation with convolutional neural networks.

Shuo Shi, Qiheng Qian, Shuhuan Yu, Qi Wang, Jinyue Wang, Jingyao Zeng, Zhenglin Du, Jingfa Xiao
Author Information
  1. Shuo Shi: National Genomics Data Center of Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing, China.
  2. Qiheng Qian: National Genomics Data Center of Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing, China.
  3. Shuhuan Yu: National Genomics Data Center of Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing, China.
  4. Qi Wang: Qujiang culture finance holding (Group) Co., Ltd, Xian, China.
  5. Jinyue Wang: National Genomics Data Center of Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing, China.
  6. Jingyao Zeng: National Genomics Data Center of Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing, China.
  7. Zhenglin Du: National Genomics Data Center of Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing, China.
  8. Jingfa Xiao: National Genomics Data Center of Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing, China. ORCID

Abstract

Genotype imputation is a statistical method for estimating missing genotypes from a denser haplotype reference panel. Existing methods usually performed well on common variants, but they may not be ideal for low-frequency and rare variants. Previous studies showed that the population similarity between study and reference panels is one of the key factors influencing the imputation accuracy. Here, we developed an imputation reference panel reconstruction method (RefRGim) using convolutional neural networks (CNNs), which can generate a study-specified reference panel for each input data based on the genetic similarity of individuals from current study and references. The CNNs were pretrained with single nucleotide polymorphism data from the 1000 Genomes Project. Our evaluations showed that genotype imputation with RefRGim can achieve higher accuracies than original reference panel, especially for low-frequency and rare variants. RefRGim will serve as an efficient reference panel reconstruction method for genotype imputation. RefRGim is freely available via GitHub: https://github.com/shishuo16/RefRGim.

Keywords

MeSH Term

Algorithms
Computational Biology
Databases, Genetic
Deep Learning
Genetics, Population
Genome-Wide Association Study
Genotype
Genotyping Techniques
Humans
Neural Networks, Computer
Reproducibility of Results
Software
Web Browser