Multi-trait/environment sparse genomic prediction using the SFSI R-package

Sparse selection indices (SSIs) can be used to predict the genetic merit of selection candidates using high-dimensional phenotypes (e.g., crop imaging) measured on each of the candidates of selection. Unlike traditional selection indices, SSIs can perform variable selection, thus enabling borrowing of information from a subset of the measured phenotypes. Likewise, sparse genomic prediction (SGP) can be used to predict genetic merit by borrowing information from a subset of the training dataset. In this study, we introduce a framework for multi-trait/environment SGP (MT-SGP) that combines the features of SSI and SGP into a single model. For candidates of selection, an MT-SGP produces prediction equations that use subsets of the training data, borrowing information from correlated traits expressed in training genotypes that are genetically close to the candidates of selection. Along with the methodology, we present an R-package (sparse family and selection index) that provides functions to solve SSIs, SGP, and MT-SGP problems. After presenting simplified examples that illustrate the use of the functions included in the package, we provide extensive benchmarks (using three data sets covering three crops and 30 traits/environments). Our results suggest that MT-SGP either outperforms (with up to 15% gains in prediction accuracy) or performs similarly to MT-genomic best linear unbiased prediction. The benchmarks provide insight regarding the conditions (sample size, genetic correlation among traits, and trait heritability) under which the use of MT-SGP can lead to gains in prediction accuracy.