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

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genotype imputation (GI) is crucial for genomic studies, but current methods struggle with local genomic patterns.
• Reference-free deep learning models show promise but often overlook local genomic context, limiting accuracy and stability.

Purpose of the Study:

• To develop ALGI, a novel method for genotype imputation that integrates local genomic window information with group-specific feature learning.
• To improve prediction accuracy, stability, and robustness in genotype imputation.

Main Methods:

• ALGI utilizes a sparse convolutional denoising autoencoder.
• It employs K-means clustering to group samples based on local genomic windows.
• Learns hidden genotype configurations specific to each group to capture fine-scale local patterns and haplotype structures.

Main Results:

• ALGI demonstrated consistent improvements over conventional methods (Beagle) and deep learning approaches (AE, SCDA).
• The method showed enhanced accuracy, stability, and robustness across yeast, human, and pig MHC regions.
• Performance was evaluated under various scenarios, including different window sizes, missing rates, sample sizes, and variant numbers.

Conclusions:

• ALGI offers a user-friendly and publicly available tool for genotype imputation.
• Its strong performance suggests broad applicability in genomic selection and complex trait genetics, although further genome-wide validation is recommended.