Evaluation of Multivariate and Random Regression Models in High-Throughput Phenotyping Traits to Improve Genomic Prediction for Yield in Wheat

High-throughput phenotyping (HTP) traits, such as, canopy temperature (CT) and normalized difference vegetation index (NDVI), can be applied for indirect selection for yield because of their correlations. The objectives are: 1) Identify statistical models that could efficiently utilize the information of HTP longitudinal data by comparing multivariate (MV) with random regression (RR) models; 2) Test the ability of HTP traits for improving genomic prediction accuracy for yield based on the proposed models. In this study, HTP traits, yield, and genetic data were collected for 557 wheat lines. HTP traits at different time points were considered as multiple response variables in a multivariate (MV) model. The random regression (RR) model involves growth trajectory for HTP traits collected over time and the high correlation between HTP observations can be reduced by a random regression function in the RR model. We found that genomic prediction accuracy for yield was increased in both MV and RR models for HTP traits in both training and testing populations by comparing with HTP traits only in the training population. Heritability for HTP traits was higher in the RR model than the MV model, however genetic correlation between HTP traits and yield were lower in the RR model compared to the MV model. In addition, although the RR model has advantages of fitting random regression coefficients individually, it was not superior to the MV model based on current genomic prediction results. In order to take advantage of longitudinal information efficiently using the RR model, more work is necessary to determine the growth curve between HTP traits and time and variation of each individual genotype around the curve.