Hermann Gregor Dallinger, Franziska Löschenberger, Herbert Bistrich, Christian Ametz,  Herbert Hetzendorfer, Laura Morales, Sebastian Michel & Hermann Buerstmayr

Theoretical and Applied Genetics; November 2023; vol. 136 (235)

Key message

NIRS of wheat grains as phenomic predictors for grain yield show inflated prediction ability and are biased toward grain protein content.

Abstract

Estimating the breeding value of individuals using genome-wide marker data (genomic prediction) is currently one of the most important drivers of breeding progress in major crops. Recently, phenomic technologies, including remote sensing and aerial hyperspectral imaging of plant canopies, have made it feasible to predict the breeding value of individuals in the absence of genetic marker data. This is commonly referred to as phenomic prediction. Hyperspectral measurements in the form of near-infrared spectroscopy have been used since the 1980 s to predict compositional parameters of harvest products. Moreover, in recent studies NIRS from grains was used to predict grain yield. The same studies showed that phenomic prediction can outperform genomic prediction for grain yield. The genome is static and not environment dependent, thereby limiting genomic prediction ability. Gene expression is tissue specific and differs under environmental influences, leading to a tissue- and environment-specific phenome, potentially explaining the higher predictive ability of phenomic prediction. Here, we compare genomic prediction and phenomic prediction from hyperspectral measurements of wheat grains for the prediction of a variety of traits including grain yield. We show that phenomic predictions outperform genomic prediction for some traits. However, phenomic predictions are biased toward the information present in the predictor. Future studies on this topic should investigate whether population parameters are retained in phenomic prediction as they are in genomic prediction. Furthermore, we find that unbiased phenomic prediction abilities are considerably lower than previously reported and recommend a method to circumvent this issue.

See https://link.springer.com/article/10.1007/s00122-023-04479-8

Figure 1: Scatter plots and phenotypic correlations of traits with GY and PC in three environments. Scatter plots for grain yield (GY), protein content (PC), grain yield deviation (GYD), protein yield (PY), equal weight protein yield (EWPY), test weight (TW), thousand kernel/grain weight (TKW) including linear regression line and Pearson’s correlation coefficient R

Fig. 1 t-SNE plots representing the relationships among the G2P-Sol pepper accessions used in the study (trimmed dataset—see “Methods”) and their classifcations under diferent schemes, as represented with t-SNE plots (calculated from IBS distances, perplexity=30). Markers labelled with Roman numerals are referred to in the main text. For k-nearest-neighbours, van Bemmelen van der Plaat (2021) recommend k=3 in general. We use k = 6, which we judged to give the best result possible for this dataset (see Supplementary Materials for a range of values). At k = 7, C. eximium (marker ix) are reassigned as C. chacoense. The unassignment cutof r is set to 4/6. Parameters for the kernelbased method are 휆 = 320, d = 0.2, r = 0.66, k = ∞