Sub-sampling graph neural networks for genomic prediction of quantitative phenotypes

Area of Science:

• Genomics and Bioinformatics
• Machine Learning in Biology
• Quantitative Genetics

Background:

• Deep learning (DL) is increasingly used in genomics to analyze complex biological data.
• Traditional genome-wide prediction (GWP) methods may not fully capture intricate, multi-generational relationships.
• Existing DL approaches for GWP often overlook the non-Euclidean graph structure inherent in genomic relationships.

Purpose of the Study:

• To propose novel deep learning architectures, a global convolutional neural network (GCN) and a GCN with random sub-sampling (GCN-RS), for genomic prediction.
• To address the limitations of existing DL methods by incorporating the non-Euclidean graph structure of individual relationships.
• To evaluate the performance of GCN and GCN-RS against traditional GWP methods.

Main Methods:

• Developed a GCN tailored for non-Euclidean spaces using graph convolutional layers.
• Introduced the GCN-RS architecture, which sub-samples the graph to reduce data dimensionality and improve efficiency.
• Constructed relationship graphs using an iterative nearest neighbor approach.

Main Results:

• The GCN-RS model demonstrated superior performance compared to the Genomic Best Linear Unbiased Predictor (GBLUP) method.
• Predictive accuracy improvements ranged from 4.4% to 49.4% in test mean squared error across one simulated and three real datasets (wheat, mice, pig).
• GCN-RS proved to be computationally efficient, suitable for large-scale genomic prediction applications.

Conclusions:

• The proposed GCN-RS architecture is a powerful and efficient tool for genomic prediction in both plants and animals.
• Incorporating graph convolutional layers effectively models complex genomic relationships for improved predictive accuracy.
• GCN-RS offers a promising advancement over traditional GWP methods, particularly for large and complex genomic datasets.