Multi-locus random marker effects GWAS, ensemble models and variable selection improve genomic prediction for pro-vitamin A carotenoids in cassava

Abstract

Cassava biofortification is an effective strategy for addressing malnutrition challenges in communities that depend on the crop as a staple. The lack of reliable genomic selection models constrains the use of genomic techniques for rapid crop biofortification. Therefore, this study aimed to determine effective strategies for optimizing parametric (Bayesian Ridge Regression [BRR], Bayesian Lasso [BL], BayesA, BayesB, and BayesC) and non-parametric (random forest [RF], XGBTree, k-nearest neighbors [KNN], SVMRadial, NNET, and Ensemble) genomic prediction (GP) models for predicting pro-vitamin A carotenoid (pVAC) levels in cassava. Parametric and non-parametric GP models were developed using 93 pVAC cassava genotypes that were genotyped for SNP markers and phenotyped for beta-carotene content using HPLC. The addition of significant SNPs from fixed-marker-effect and random-marker-effect GWAS to parametric GP models greatly improved the prediction performance of these models. The initial prediction abilities of the non-parametric GP models ranged between r = 0.04 and r = 0.32, with an ensemble result of r = 0.44. However, the use of variable selection improved the prediction performance of the models, which ranged from r = 0.34 to r = 0.79. The ensemble of these models resulted in a prediction performance similar to that of a high-performing single model (r = 0.79). In conclusion, using significant SNPs from multi-locus random marker effects GWAS, variable selection, and an ensemble of models enhanced the prediction power of GP models. This study contributes to ongoing efforts to accelerate cassava biofortification in Kenya using genomic prediction.