Fusarium
head blight
(FHB), also known as head scab, is an economically and environmentally
concerning disease of wheat (Triticum aestivum L). A two-pronged
approach of marker assisted selection coupled with genomic selection has been
suggested when breeding for FHB resistance. In the present study, an historical
dataset comprised of entries in the Southern Uniform Winter Wheat Nursery
(SUWWSN) ranging from 2011-2021 was partitioned and used for cross validation
within the dataset from the years 2011-2019 and forward validation on the 2020-2021
environments in the dataset. Two traits were curated: percent Fusarium damaged
kernels (FDK) and Deoxynivalenol (DON) content in parts per million. Two
methods of data partitioning were entertained to establish a training
population for each trait: 1) sub-setting environments by genomic heritability
and 2) sub-setting environments based on like performance of checks. Heritability
was estimated for each trait-by-environment combination and data were
partitioned on environments with greater heritability than 0.10, 0.25, 0.50,
0.75, and 0.90. For each trait assessed, a consistent set of check lines were
drawn from each environment in the SUWWSN, and K-means clustering was performed
to establish clusters with like-check performance; possible number of clusters
from 2 to 15 were entertained and the optimal clustering was selected by
multiple clustering criterion. By majority rule among the clustering criterion,
two clusters were identified for FDK and three for DON. When using the combined
data from 2011-2019 for training, forward validation for FDK on the SUWWSN 2020
and 2021 indicated a predictive accuracy R=0.58 and R=0.53, respectively. In comparison, forward validation using only
the environments in the first cluster of like-check-performance environments
for FDK indicated a predictive accuracy of R= 0.65 and R=0.60. Forward validation for DON using the total
data available indicated a predictive accuracy of R=0.57 and R=0.45. Likewise, forward validation using only the like-check
performance environments in cluster one for DON indicated a predictive accuracy
of R=0.67 and R=0.60. These results indicate that selecting
environments based on like-check performance may produce substantially higher
forward prediction accuracies than using the total available data. This work
may be used as a model to create a public resource for genomic prediction of
FHB resistance traits across public wheat breeding programs.