Partial Least Squares Enhances Genomic Prediction of New Environments

Abstract

In plant breeding, the need to improve the prediction of future seasons or new locations and/or environments, also denoted as “leave one environment out,” is of paramount importance to increase the genetic gain in breeding programs and contribute to food and nutrition security worldwide. Genomic selection (GS) has the potential to increase the accuracy of future seasons or new locations because it is a predictive methodology. However, most statistical machine learning methods used for the task of predicting a new environment or season struggle to produce moderate or high prediction accuracies. For this reason, in this study we explore the use of the partial least squares (PLS) regression methodology for this specific task, and we benchmark its performance with the Bayesian Genomic Best Linear Unbiased Predictor (GBLUP) method. The benchmarking process was done with 14 real datasets. We found that in all datasets the PLS method outperformed the popular GBLUP method by margins between 0% (in the Indica data) and 228.28% (in the Disease data) across traits, environments, and types of predictors. Our results show great empirical evidence of the power of the PLS methodology for the prediction of future seasons or new environments.

Multi-trait genome prediction of new environments with partial least squares

“…For example, Vargas et al (1998) , Vargas et al (1999) and Crossa et al (1999) used this methodology for interpreting genotype by environment interaction in maize and wheat. For prediction in GS, this methodology has been used by Monteverde et al (2019) , Colombani et al (2012) and by Montesinos-López et al (2022a) for UT predictions.…”

“…Multi-trait models under artificial deep neural networks have even been explored in genomic selection ( Montesinos-López et al, 2018 ; 2019c ). Recently, Montesinos-López et al (2022a) explored the use of the partial least square (PLS) regression methodology for the prediction of one full environment of a single trait. The authors’ benchmarked the performance of the PLS for predicting a UT with the Bayesian Genomic Best Linear Unbiased Predictor (GBLUP) method, and in all data sets the UT-PLS method outperformed the UT-GBLUP method by margins between 0% and 228.28% across traits, environments and types of predictors.…”

“…In the current research, we evaluate the prediction performance of two multi-trait methods, MT-PLS and MT-GBLUP, in the context of leave-one-location-out cross validation. The proposed MT-PLS model is an extension of the UT-PLS recently proposed by Montesinos-López et al (2022a) . The goals of this study are three-fold: 1) benchmark the genomic-enabled prediction power of MT-PLS and UT-GBLUP, 2) compare the prediction performance of these two multi-trait methods (MT-PLS and MT-GBLUP) with their corresponding UT versions (UT-PLS and UT-GBLUP) and 3) evaluate the prediction power of these methods within three different predictor models (E + G, E + G + GE, and G + GE where E = Environment; G = genomic, and GE is genomic × environment interaction) under the leave one location out cross validation.…”

A Comparison of Three Machine Learning Methods for Multivariate Genomic Prediction Using the Sparse Kernels Method (SKM) Library

“…In a recent study, Montesinos-López et al [ 35 ] investigated the partial least square (PLS) regression methodology for the prediction of one full environment of a single trait (ST) and compared its prediction performance with that of the GBLUP method. In all datasets, the ST-PLS method outperformed the ST-GBLUP method by margins between 0% and 228.28% across traits, environments, and types of predictors.…”

“…However, because we only reported the prediction performance across traits, other details could not be fully observed in the provided tables and figures. Furthermore, under the implemented sevenfold CV, we observed that in general, the MT-PLS model achieved slightly worse performance than MT-GBLUP and MT random forest models; however, this finding cannot be extrapolated to all types of cross-validation strategies because evidence suggests that when the goal is to predict a complete environment (or year), the ST-PLS and MT-PLS models outperform the ST-GBLUP and MT-GBLUP models by considerable margins [ 34 , 35 ], which can be attributed, in part, to the fact that the ST and MT-PLS models first involve a variable selection process during which a considerable amount noise is discarded, and at the end, the training process applied with a latent variable of inputs.…”