ALGI: Sparse Convolutional Denoising Autoencoder Utilizing Local Genomic Information for Genotype Imputation

Abstract

Genotype imputation (GI) is widely used to predict missing genetic information, which is essential for genomic studies and breeding programs. Recent deep learning approaches have shown promising results without relying on reference panels, but they often overlook important local genomic patterns that could improve prediction accuracy. In this study, we developed a novel method called ALGI (sparse convolutional denoising autoencoder with local genomic information), which integrates local genomic information into a deep learning framework. By grouping samples based on local genomic windows and applying a sparse convolutional denoising autoencoder, the model effectively captures complex genetic structures and improves the accuracy and stability of genotype imputation. This approach provides a more reliable and efficient solution for genomic data analysis, supporting advances in precision breeding and genetic research. Keywords: genotype imputation, K-means, autoencoder, deep learning, local genomic, convolution, genotype inference, Beagle