TABLE 4.

Highest performing genomic selection (GS) model for each trait/cycle combination across five breeding cycles representing two different breeding programs.

Prediction site	Year	Free threshing	Plant height	Seed mass	Shattering	Spikelets per inflorescence	Spike length	Spike yield
UMN-C1	2012	MN†	LOO	TLI-PCA		LOO	KS	LOO
	2013	LOO	LOO	LOO		TLI-PCA	LOO	LOO
UMN-C2	2014	MN	MN		LOO		KS	LOO
	2015	LOO	LOO		LOO		LOO	MN
TLI-C6	2016	LOO	KS	LOO	UMN-PCA	LOO	LOO	MN
	2017	LOO	LOO	LOO	LOO	LOO	LOO	MN
TLI-C7	2018	LOO	LOO	TLI-PCA	LOO	LOO	LOO	TLI-PCA
	2019	LOO	LOO	KS	TLI-PCA	LOO		KS
TLI-C8	2019	UMN-PCA	LOO	LOO	UMN-PCA	UMN-PCA		UMN-PCA

Predictive ability was assessed as the correlation between the GS predicted value and the phenotypic best linear unbiased predictor (BLUP). Models differed with respect to the training population used to develop the model. The leave one-out, LOO, model was used as the reference model and only if a model exceeded the 95% confidence interval of the LOO model was it considered superior. ^†Models are: LOO, leave-one-out, prediction cycle is left out of the training set, and all other cycles are used to train the model. MN and KS are breeding-program specific where only genets from Minnesota (or Kansas) are used to predict each cycle. For TLI-C6 2016 plant height, KS training population would consist of TLI-C7 and TLI-C8, with TLI-C6 as the prediction population. UMN-PCA and TLI-PCA are where the training population is made from PCA analysis of the marker matrix, with UMN-PCA encompassing most UMN lines and some of TLI that were more similar to UMN material than the TLI subset.