Approximate Bayesian neural networks in genomic prediction.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genome-wide marker data are crucial for phenotypic genome-wide association studies (GWAS) and genome-wide prediction (GWP).
• High-dimensional SNP data from numerous individuals are typically analyzed.
• Neural networks (NN) and deep learning are underutilized in genomic studies.

Purpose of the Study:

• To present a novel neural network (NN) model for analyzing genomic SNP data.
• To evaluate the performance of the proposed NN model in genomic prediction and association studies.

Main Methods:

• Development of an approximate Bayesian neural network (ABNN) model using weight decay and dropout for regularization.
• Implementation of the ABNN model in mxnet.
• Comparison of ABNN with genomic best linear unbiased prediction (GBLUP) and Bayesian LASSO.

Main Results:

• The ABNN model demonstrated superior prediction accuracy compared to GBLUP and Bayesian LASSO.
• Mean squared error was reduced by at least 6.5% in simulated data and 1% in real pig data.
• A shallow NN model with specific configurations (one layer, one neuron, one-hot encoding, linear activation) outperformed more complex models.

Conclusions:

• The ABNN model offers a computationally efficient approach with strong prediction performance for GWP and GWAS.
• SNP importance can be inferred from ABNN model weights.
• The ABNN model is suitable for both genome-wide prediction and genome-wide association studies.