An autoencoder-based deep learning method for genotype imputation.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Genotype imputation is crucial for genome-wide association studies (GWAS), aiding in statistical power, locus discovery, and variant prioritization.
• Deep learning (DL) methods, like sparse convolutional denoising autoencoders (SCDA), have emerged for genotype imputation, but optimizing their accuracy remains challenging.

Purpose of the Study:

• To develop and evaluate a novel convolutional autoencoder (AE) model for genotype imputation with improved accuracy.
• To address the challenge of optimizing DL-based genotype imputation learning processes.

Main Methods:

• Developed a convolutional autoencoder (AE) model for genotype imputation.
• Implemented a customized training loop using a single batch loss instead of average batch loss.
• Evaluated the model on yeast, 1,000 Genomes Project (1KGP) human leukocyte antigen (HLA) data, and Louisiana Osteoporosis Study (LOS) genotype data.

Main Results:

• The modified AE imputation model demonstrated comparable or superior performance to the SCDA model across all tested datasets.
• Achieved high evaluation metrics, including concordance rate (CR), Hellinger score, scaled Euclidean norm (SEN) score, and imputation quality score (IQS).
• For HLA data, the AE model achieved an average CR of 0.9468 and Hellinger score of 0.9765 at 10% missingness.

Conclusions:

• The proposed AE-based genotype imputation method shows significant potential for enhancing GWAS statistical power.
• The customized training approach improves imputation accuracy, offering a valuable tool for post-GWAS analyses.