DeepCGP: A Deep Learning Method to Compress Genome-Wide Polymorphisms for Predicting Phenotype of Rice

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genomic selection (GS) accelerates breeding but faces challenges with large genome-wide data storage and computational demands.
• Existing data compression methods often compromise data quality, while prediction models can be computationally intensive and require original data.

Purpose of the Study:

• To develop a novel Deep Learning Compression-based Genomic Prediction (DeepCGP) model to address limitations in genomic data compression and prediction.
• To integrate deep learning for efficient genome-wide data compression and subsequent phenotype prediction.

Main Methods:

• Proposed the DeepCGP model, comprising a deep neural network autoencoder for data compression and regression models (RF, GBLUP, BayesB) for phenotype prediction from compressed data.
• Applied the DeepCGP model to two rice datasets containing genome-wide marker genotypes and target trait phenotypes.

Main Results:

• The DeepCGP model achieved up to 99% prediction accuracy with 98% data compression for a target trait.
• Compared to RF, GBLUP, and BayesB, DeepCGP demonstrated superior performance in both data compression and prediction accuracy.
• BayesB, while accurate, was computationally intensive and only functional with compressed data.

Conclusions:

• DeepCGP offers a powerful solution for managing large genomic datasets, enhancing prediction accuracy, and accelerating breeding applications.
• The developed model outperforms existing state-of-the-art methods, providing a significant advancement in genomic prediction efficiency.