PNNGS, a multi-convolutional parallel neural network for genomic selection.

Area of Science:

• Plant breeding
• Genomics
• Machine learning

Background:

• Genomic selection (GS) accelerates crop and livestock improvement over traditional phenotypic selection.
• Enhancing prediction accuracy is crucial for the widespread adoption and effectiveness of GS.
• Deep learning models offer potential for improving GS but require optimization for complex genomic data.

Purpose of the Study:

• To introduce a novel deep learning model, the parallel neural network for genomic selection (PNNGS), to enhance GS prediction accuracy and stability.
• To investigate the impact of parallel convolutional paths and different loss functions on GS performance.
• To address challenges posed by unbalanced cluster data in genomic prediction.

Main Methods:

• Developed PNNGS, incorporating parallel convolutions with varying kernel sizes and residual connections.
• Trained PNNGS using four L_p loss functions and determined optimal parallel path numbers for different species (rice, sunflower, wheat, maize).
• Compared PNNGS against RRBLUP, RF, SVR, and DNNGP using 24 prediction cases, and employed PCA and K-means for data clustering and stratified sampling.

Main Results:

• PNNGS and serial DNNGP outperformed RRBLUP, RF, and SVR in phenotype prediction accuracy.
• PNNGS demonstrated an average prediction accuracy 0.031 higher than DNNGP, confirming the benefit of parallelism.
• Stratified sampling improved PNNGS prediction stability and accuracy, with significant accuracy drops when small clusters had reduced sample sizes.

Conclusions:

• The parallel neural network for genomic selection (PNNGS) effectively enhances prediction accuracy and stability in genomic selection.
• Parallelism in deep learning models is beneficial for improving genomic prediction performance.
• Addressing data imbalance through stratified sampling and increasing sample sizes in smaller clusters is critical for robust genomic selection.