Predicting phenotypes from genetic, environment, management, and historical data using CNNs.

Area of Science:

• Genomics
• Machine Learning
• Agricultural Science

Background:

• Predicting phenotypes from genetic, environmental, and management factors is crucial for agriculture, medicine, and conservation.
• Traditional methods often oversimplify complex datasets through subjective feature engineering.
• Deep learning offers a data-driven approach to identify important predictive factors.

Purpose of the Study:

• To develop and evaluate Convolutional Neural Networks (CNNs) for predicting agronomic yield.
• To compare CNN performance against standard genomic prediction methods.
• To investigate the impact of pre-training on prediction accuracy.

Main Methods:

• Developed CNN models utilizing replicated trials and historical yield data.
• Compared CNN predictions with standard methods using held-out genetic, environmental, and management data.
• Employed saliency map analysis to interpret CNN decision-making.

Main Results:

• CNNs achieved higher accuracy than standard methods when predicting with genetic, environmental, and management data (r=0.50 vs. 0.43).
• CNNs performed slightly less accurately than standard methods when only genetic data was held out (r=0.74 vs. 0.80).
• Pre-training CNNs on historical data significantly improved prediction accuracy.

Conclusions:

• CNNs offer a powerful alternative to standard genomic prediction, especially with rich datasets.
• CNNs can autonomously identify and prioritize agriculturally relevant predictive factors.
• The approach holds potential for enhancing prediction accuracy in diverse biological and agricultural applications.