Genomic prediction accuracy for low hydrogen cyanide selection in fresh cassava roots. Comparative model analysis

Abstract

Cassava (Manihot esculenta Crunz) is a staple food crop for millions of people in Africa. The crop’s tolerance to drought, ability to survive in marginal soils and amenability to a variety of uses endears it to farmers. However, utilization of the crop’s full potential for food security is limited by the presence of cyanogenic glucosides (HCN) in the roots. These not only make cassava bitter to the taste but can also be toxic to lethal levels. Breeding for low HCN cassava is necessary to mitigate against high dietary HCN consumption. However, efforts at variety development have been slow, in part due to limited use of molecular tools. This study assessed genomic prediction accuracy for fresh cassava root HCN content using the GBLUP model and compared its performance to 5 Bayesian models (Bayes A, Bayes B, Bayes C, Bayesian Ridge Regression (BRR), Bayesian Lasso (BL) and the Reproducing Kernel Hilbert Spaces (RKHS) model. We used a genomic selection (GS) cycle two population consisting of 434 clones which were genotyped with 24,040 SNP markers using the DArTseq platform. Data were filtered to remove markers with minor allele frequency (MAF) less than 0.05, call rate less than 80%, clones with more than 20% missing calls and clones with no matching phenotypic data. Phenotypic data on fresh cassava root HCN content was collected on roots at 12 months after planting. The accuracy of GBLUP model was low, ranging from -0.41 to 0.68 and averaging 0.22 across folds. On the other hand, the accuracy of RKHS and Bayesian models was moderate, with BL and RKHS being most accurate (r = 0.52), followed by Bayes A (r = 0.49) and Bayes B and Bayes C (r = 0.48). The BRR model was least accurate (r = 0.18). Collectively, these results provide a foundation for implementing GS to accelerate efforts to deliver desired genetic gains for low HCN cassava in farmers’ fields.