Prediction of Maize Phenotypic Traits With Genomic and Environmental Predictors Using Gradient Boosting Frameworks

Area of Science:

• Agricultural Science
• Genetics
• Data Science

Background:

• Climate change poses challenges for stable crop variety development.
• Genomic prediction models can be enhanced by field-level environmental data to capture genotype-by-environment interactions.
• Machine learning (ML) offers advanced approaches for handling complex data and nonlinear relationships.

Purpose of the Study:

• To evaluate the predictive ability of ML models for maize phenotypic traits using environmental data.
• To compare ML models against traditional linear models for genomic prediction.
• To identify key environmental factors influencing crop performance.

Main Methods:

• Utilized multi-environment trials (METs) data from the Maize Genomes to Fields (G2F) Initiative (2014-2017).
• Incorporated genotypic data with soil and weather variables into prediction models.
• Compared linear random effects models with elastic net, XGBoost, and LightGBM (ML methods).
• Evaluated models across four prediction scenarios involving new genotypes and environments.

Main Results:

• Gradient boosting ML methods improved grain yield prediction accuracy by up to 20% for new genotypes in new years when environmental predictors were included.
• ML methods and detailed environmental data did not enhance predictive ability for plant height.
• Temperature during flowering, water availability during vegetative/grain filling, and soil organic matter were identified as key predictors for grain yield.

Conclusions:

• Machine learning, particularly gradient boosting, combined with environmental data, significantly enhances genomic prediction for maize grain yield.
• Predictive gains for plant height were not observed, suggesting trait-specific responses to environmental data integration.
• Environmental variables like temperature, water, and soil organic matter are critical for accurate yield forecasting in maize breeding.