Multi-modal deep learning improves grain yield prediction in wheat breeding by fusing genomics and phenomics

Area of Science:

• Plant breeding
• Agricultural science
• Machine learning

Background:

• Developing superior crop varieties is crucial for global food security.
• Current plant breeding programs face limitations due to long field cycles and selection processes.
• Existing yield prediction models using genotype or phenotype data require improvement.

Purpose of the Study:

• To develop an integrated machine learning model for enhanced crop yield prediction.
• To leverage both genetic (genotype) and multi-source phenotypic data, including unmanned aerial systems (UAS) imagery.
• To improve the interpretability of yield prediction models.

Main Methods:

• A deep multiple instance learning framework with an attention mechanism was employed.
• The model fused genetic variants with UAS-collected data.
• Model performance was evaluated for yield prediction in similar and unseen environments.

Main Results:

• The model achieved a 0.754 Pearson correlation coefficient for yield prediction in similar environments, a 34.8% improvement over a genotype-only baseline.
• In unseen environments, the model improved genotype-only yield prediction accuracy by 13.5% (0.386 correlation coefficient).
• The multi-modal architecture effectively integrated plant health and environmental data, enhancing prediction accuracy and interpretability.

Conclusions:

• The proposed machine learning model significantly enhances crop yield prediction accuracy by integrating diverse data sources.
• This approach accelerates the breeding of improved crop varieties, contributing to sustainable food security.
• The developed model and associated data are publicly available to support further research.