A Deep Autoencoder Compression-Based Genomic Prediction Method for Whole-Genome Sequencing Data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genomic prediction using whole-genome sequencing (WGS) data faces challenges due to high dimensionality (p) and limited sample size (n), where n ≪ p.
• The computational burden of high-dimensional WGS data restricts its practical application in genomic selection.
• Existing methods struggle to efficiently handle the vast amount of genetic information in WGS data.

Purpose of the Study:

• To introduce DAGP, a novel deep autoencoder compression method for reducing WGS data dimensionality.
• To evaluate the computational efficiency and prediction accuracy of DAGP in genomic prediction.
• To assess the performance of DAGP when integrated with various genomic prediction models.

Main Methods:

• Developed DAGP, a deep autoencoder model to compress WGS data, reducing dimensionality by over 99% while retaining essential genetic information.
• Integrated DAGP with Genomic Best Linear Unbiased Prediction (GBLUP), Bayesian, and machine learning models for genomic prediction.
• Validated the method on sturgeon (50K markers) and maize (20K markers) datasets, comparing prediction accuracy with traditional WGS-based GBLUP.

Main Results:

• DAGP achieved over 99% dimensionality reduction of WGS data, significantly enhancing computational efficiency.
• Genomic prediction accuracy using DAGP combined with GBLUP was comparable to using full WGS data, even at reduced marker densities.
• Integration of DAGP with Bayesian and machine learning models resulted in improved genomic prediction accuracy, with average gains of 6.05% and 5.35% over WGS-based GBLUP, respectively.

Conclusions:

• DAGP offers an efficient and scalable solution for genomic prediction with high-dimensional WGS data.
• The method preserves crucial genetic information during compression, enabling accurate genomic predictions.
• DAGP facilitates the practical application of WGS data in genomic selection across various species, improving both computational feasibility and predictive performance.