Genomic prediction in plants: opportunities for ensemble machine learning based approaches

Area of Science:

• Plant genetics
• Bioinformatics
• Computational biology

Background:

• Machine learning (ML) methods are increasingly used for genomic prediction (GP) in plants.
• A clear rationale for selecting ML over traditional parametric methods is often missing.
• GP model performance depends on factors like sample size, marker density, and genetic architecture.

Purpose of the Study:

• To identify dataset and problem characteristics associated with successful ML-based genomic prediction.
• To compare the predictive performance of ensemble ML methods against established parametric approaches.

Main Methods:

• Compared Random Forest and Extreme Gradient Boosting (ensemble ML) with GBLUP, RKHS, BayesA, and BayesB (parametric methods).
• Utilized simulated and real plant trait data with varying genetic complexity (QTLs, heritability, population structure, linkage disequilibrium).

Main Results:

• Ensemble ML methods excel with nonlinear phenotypes and are competitive with Bayesian methods for linear phenotypes with large-effect Quantitative Trait Nucleotides (QTNs).
• ML methods are vulnerable to population structure confounding but less affected by low linkage disequilibrium than linear parametric methods.

Conclusions:

• Provides insights into the application and performance of ML in genomic prediction.
• Offers practical guidelines for selecting appropriate GP methods based on trait characteristics and data.