Multi-trait genomic prediction improves selection accuracy for yield traits in cassava breeding panel

A preprint posted to bioRxiv describes an evaluation of multi-trait genomic prediction (MT-GBLUP) against single-trait (ST-GBLUP) models in a Brazilian cassava breeding panel comprising 1,078 clones genotyped with 25,923 single-nucleotide polymorphisms (SNPs) and phenotyped for six agronomic traits. The authors used stage-wise mixed models to generate best linear unbiased estimates (BLUEs) as inputs to both model types, then applied five cross-validation schemes designed to reflect realistic breeding scenarios.

The cross-validation strategies tested include a baseline single-trait scheme (CV1), a naive multi-trait prediction for unphenotyped candidates (CV2), and additional scenarios that exploit auxiliary trait information to varying degrees. According to the preprint's lede, the multi-trait framework consistently improved predictive accuracy compared with single-trait modelling, particularly for traits that are expensive or time-consuming to phenotype directly in the field.

Cassava is a major staple crop in tropical regions, and improving genomic selection efficiency has direct relevance for food security breeding programmes. For researchers in plant genetics, quantitative genetics, and genomic prediction methodology, the work offers a comparison of model architectures under conditions that reflect real breeding-pipeline constraints. The use of a clonally propagated species also presents particular modelling considerations, as standard assumptions about random mating do not apply. This is a preprint and has not yet been peer-reviewed.