ConvCGP: A convolutional neural network to predict genetic values of agronomic traits from compressed genome-wide...

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genomic prediction faces computational challenges due to large genome-wide polymorphism datasets.
• Existing dimensionality reduction methods like PCA and autoencoders are often computationally intensive.
• Deep learning models struggle with high-dimensional genomic data due to training costs.

Purpose of the Study:

• To develop a computationally efficient and scalable deep learning framework for genomic prediction.
• To address the limitations of existing methods in handling large-scale genomic data.
• To integrate nonlinear data compression with predictive modeling for improved genomic prediction.

Main Methods:

• Proposed a novel deep learning framework, Compression-based Genomic Prediction using Convolutional Neural Networks (ConvCGP).
• Integrated autoencoder-based nonlinear compression with convolutional neural network (CNN) prediction in an end-to-end pipeline.
• Applied ConvCGP to high-dimensional rice and maize genomic datasets for trait prediction.

Main Results:

• ConvCGP achieved prediction accuracy comparable to models using uncompressed data, even at 98% data reduction.
• The framework significantly reduced storage needs and computational load.
• ConvCGP outperformed PCA, GBLUP, LASSO, and SVM methods in prediction accuracy and scalability.

Conclusions:

• ConvCGP is a powerful, efficient, and scalable solution for modern genomic prediction.
• The method effectively preserves predictive information under drastic dimensionality reduction.
• ConvCGP demonstrates robust performance on massive genomic datasets for animal and plant breeding.