Sparse Convolutional Neural Networks for Genome-Wide Prediction.

Area of Science:

• Genomics
• Artificial Intelligence
• Animal Breeding

Background:

• Genome-wide prediction (GWP) is crucial for artificial selection, facing challenges with large datasets.
• Machine learning, particularly neural networks (NN) and deep learning (DL), shows potential but is underutilized in GWP.
• Computational efficiency is a key consideration for GWP methods.

Purpose of the Study:

• To introduce a novel, powerful, and extensible neural network method for GWP.
• To leverage deep learning for improved prediction accuracy and computational efficiency in genomic selection.
• To evaluate the performance of the proposed method against existing GWP approaches.

Main Methods:

• A one-dimensional convolutional neural network (CNN) was developed, named CNNGWP, to utilize ordinal marker information.
• The CNNGWP incorporates pooling and ℓ1-norm regularization for sparsity and computational efficiency.
• Hyper-parameters were optimized using Bayesian optimization, and model-averaged ensemble predictions were employed.

Main Results:

• CNNGWP demonstrated a significant reduction in prediction error, exceeding 25% on simulated data.
• The method achieved approximately 3% improvement in prediction error on real pig data compared to GBLUP and LASSO.
• The CNNGWP approach proved to be computationally efficient for large-scale genomic datasets.

Conclusions:

• CNNGWP presents a promising and powerful deep learning approach for genome-wide prediction.
• The method offers improved accuracy and computational efficiency, particularly for complex genetic architectures.
• Further research may explore its application across diverse species and genetic scenarios, considering heritability impacts.