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

Missing genotypes reduce statistical power and hinder genome-wide association studies. While reference-based methods are popular, they struggle in complex regions and under population mismatch. Existing reference-free deep learning models show promise in addressing this issue but often fail to impute rare variants in small datasets. We propose BiU-Net, a biologically informed U-Net model that segments genotype data and encodes positional information to preserve the genomic context. Evaluated on the 1000 Genomes Project, Louisiana Osteoporosis Study, and Simons Genome Diversity Project datasets, BiU-Net outperformed Beagle and sparse convolutional denoising autoencoder in overall metrics and in metrics stratified by minor allele frequency.