Multi-trait, Multi-environment Deep Learning Modeling for Genomic-Enabled Prediction of Plant Traits

Area of Science:

• Animal and Plant Breeding
• Statistical Genetics
• Machine Learning in Agriculture

Background:

• Genomic selection (GS) utilizes multi-trait and multi-environment data for improved breeding outcomes.
• Existing multi-trait models enhance prediction accuracy by leveraging trait correlations but often demand significant computational resources.
• There is a need for advanced statistical models that efficiently exploit trait correlations in GS.

Purpose of the Study:

• To evaluate the prediction accuracy of multi-trait deep learning (MTDL) models in genomic selection.
• To compare the performance of MTDL models against a Bayesian multi-trait multi-environment (BMTME) model.
• To assess the impact of genotype×environment interaction on model performance.

Main Methods:

• Comparison of MTDL models with the Bayesian multi-trait and multi-environment (BMTME) model.
• Evaluation of models with and without genotype×environment interaction terms.
• Cross-validation using Pearson's correlation to assess prediction performance.

Main Results:

• MTDL models demonstrated competitive prediction accuracy, often similar to BMTME models across datasets.
• MTDL models outperformed BMTME models when genotype×environment interaction was excluded.
• BMTME models showed superior performance when genotype×environment interaction was included.

Conclusions:

• MTDL models are a highly competitive alternative for predictions in GS, offering comparable accuracy to BMTME models.
• MTDL models present a practical advantage due to their lower computational resource requirements.
• The choice between MTDL and BMTME models may depend on the inclusion of genotype×environment interactions.