Automated Machine Learning: A Case Study of Genomic "Image-Based" Prediction in Maize Hybrids

Area of Science:

• Genomics and Bioinformatics
• Plant Breeding and Genetics
• Machine Learning Applications

Background:

• Machine learning, including multilayer perceptrons (MLP) and Convolutional Neural Networks (CNN), is increasingly recognized for its potential in genomic prediction (GP).
• Evaluating these methods against established techniques like Genomic Best Linear Unbiased Prediction (GBLUP) is crucial for advancing breeding strategies.

Purpose of the Study:

• To assess the performance of MLP and CNN for genomic prediction in maize hybrids.
• To compare MLP and CNN against GBLUP in regression tasks and against each other in classification tasks.
• To explore the utility of automated hyperparameter tuning for machine learning models in GP.

Main Methods:

• Genomic data from maize hybrids were utilized, with MLP receiving a relationship matrix and CNN receiving 'genomic images'.
• Regression and classification tasks were performed, with traits discretized for classification to create balanced and unbalanced datasets.
• An automated hyperparameter search was conducted for MLP and CNN, followed by performance evaluation using various metrics and a validation scheme.

Main Results:

• Both MLP and CNN demonstrated competitive predictive performance compared to GBLUP in regression tasks.
• MLP and CNN showed comparable results in classification tasks, even with varying dataset balances representing different selection intensities.
• The study provides insights into automated machine learning for optimizing genomic prediction models.

Conclusions:

• MLP and CNN are viable and competitive alternatives to GBLUP for genomic prediction in plant breeding.
• Automated machine learning approaches can enhance the efficiency and effectiveness of genomic prediction.
• These findings have significant implications for accelerating genetic gain and improving crop traits through advanced breeding technologies.