High throughput phenotyping and genomics assisted breeding for low hydrogen cyanide in cassava

Abstract

Cassava has diverse uses for food and feed. However, the utility of the crop is restricted by the presence of hydrogen cyanide (HCN) in the roots. The HCN not only makes roots bitter to the taste, but can also be poisonous to fatal levels. Limiting societal exposure to high HCN cassava by breeding for low HCN varieties is the most sustainable way of limiting HCN dietary toxicity. However, progress has been slow, partly due to the difficulty in HCN phenotyping. This study aimed to contribute to accelerated breeding of low HCN cassava varieties by developing faster, accurate and reproducible HCN prediction methods to inform selection decisions. Multi-location field trials were established; leaf samples picked three months after planting and genotyped with kompetitive allele specific polymerase chain reaction (KASP) markers and also genotyped by sequencing (GBS). Roots were harvested 12 months after planting for HCN quantification and collection of near infrared spectroscopy (NIRS) spectra. Highly significant differences were observed among clones (p < 0.001) and locations (p < 0.001) with significant clone by environment interactions ranging from (p < 0.05) to (p < 0.001). The HCN phenotype range 1 to 9 (< 50 ppm to 800 ppm. Three KASP markers snpME00404, snpME00405 and snpME00406 showed significant co-segregation of genotypes with HCN phenotype, explaining 27%, 17% and 14% of the HCN phenotypic variation respectively. The NIRS prediction with modified partial least squares (MPLS) algorithm had low accuracy (R2p = 0.38). However, binary classification with machine learning algorithms achieved up to 99% accuracy with partial least squares discriminant analysis (PLS-DA), 75% with support vector machine (SVM) and 74% with logistic regression (LR) using the full spectral range (400 – 2500 nm). The wavelengths 961, 1165, 1403–1505, 1913–1981, and 2491 nm were influential in discrimination of low and high HCN accessions. Using these selected wavelengths, LR achieved 100% classification accuracy. The accuracy of genomic prediction models was low for Bayesian Ridge Regression (BRR) (r = 0.18) and Genomic Best Linear Unbiased Prediction (GBLUP) (r = 0.22) but moderate for Bayes B and C (r = 0.48), and Bayes A (r = 0.49). Reproducing Kernel Hibert Spaces (RKHS) and Bayesian Lasso (BL) models had the highest prediction accuracy (r = 0.52). Collectively, these results offer frameworks for low HCN cassava breeding in Uganda. The NIRS and KASP markers can be used in routine phenotyping and selection in early generation trials while the genomic selection models RKHS or BL can be deployed in cyclic population improvement for low HCN.