Using Local Convolutional Neural Networks for Genomic Prediction.

Area of Science:

• Genetics and breeding
• Bioinformatics
• Machine learning in agriculture

Background:

• Genomic prediction is crucial for livestock and crop breeding, aiming to predict breeding values and phenotypes.
• Artificial neural networks (ANNs) offer potential for genomic prediction, with local convolutional neural networks (LCNNs) showing promise due to their region-specific filtering capabilities.

Purpose of the Study:

• To evaluate the performance of LCNNs for genomic prediction compared to traditional ANNs and state-of-the-art methods.
• To assess the impact of LCNN architecture on predictive ability across different datasets and trait heritabilities.

Main Methods:

• Utilized simulated maize and real Arabidopsis datasets for performance evaluation.
• Compared LCNNs against multi-layer perceptrons, convolutional neural networks, genomic best linear unbiased prediction (GBLUP), and Bayesian models.
• Investigated the effect of LCNN kernel size, stride, and fully connected layer configurations.

Main Results:

• Baseline LCNN outperformed other ANNs for most traits across both datasets.
• LCNNs showed superior predictive ability (up to 24% increase) over GBLUP and Bayesian models for high heritability traits in large populations.
• State-of-the-art methods outperformed ANNs in small training populations, though LCNNs still showed a ~10% advantage over other ANNs.

Conclusions:

• LCNNs represent a promising advancement in genomic prediction, particularly for large-scale datasets and complex genetic architectures.
• While ANNs offer flexibility for incorporating diverse data inputs, challenges remain regarding heritability estimation and the additive nature of breeding values.
• Continued advancements in phenotyping and genotyping may position ANNs as a viable alternative for genomic prediction in breeding programs.