BiU-Net: A Biologically Informed U-Net for Genotype Imputation.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Missing genotypes in genomic data reduce statistical power for genome-wide association studies (GWAS).
• Reference-based imputation methods face challenges in complex genomic regions and with population mismatches.
• Current reference-free deep learning models struggle with imputing rare variants, especially in smaller datasets.

Purpose of the Study:

• To develop a novel deep learning model, BiU-Net, for accurate genotype imputation.
• To enhance the imputation of rare variants and preserve genomic context.
• To evaluate BiU-Net's performance against existing state-of-the-art methods.

Main Methods:

• BiU-Net, a biologically informed U-Net architecture, was developed to segment genotype data.
• The model encodes positional information to maintain genomic context during imputation.
• Performance was evaluated on three diverse datasets: 1000 Genomes Project, Louisiana Osteoporosis Study, and Simons Genome Diversity Project.

Main Results:

• BiU-Net demonstrated superior performance compared to Beagle and a sparse convolutional denoising autoencoder.
• The model achieved better overall imputation metrics across all evaluated datasets.
• BiU-Net showed significant improvements in imputing rare variants, particularly when stratified by minor allele frequency.

Conclusions:

• BiU-Net offers a robust and biologically informed approach to genotype imputation.
• The model effectively addresses limitations of existing methods, especially for rare variant imputation.
• BiU-Net shows promise for improving the accuracy and power of genome-wide association studies.