Interest of phenomic prediction as an alternative to genomic prediction in grapevine

Abstract

Background: Phenomic prediction has been defined as an alternative to genomic prediction by using spectra instead of molecular markers. A reflectance spectrum provides information on the biochemical composition within a tissue, itself being under genetic determinism. Thus, a relationship matrix built from spectra could potentially capture genetic signal. This new methodology has been mainly applied in several annual crop species but little is known so far about its interest in perennial species. Besides, phenomic prediction has only been tested for a restricted set of traits, mainly related to yield or phenology. This study aims at applying phenomic prediction for the first time in grapevine, using spectra collected on two tissues and over two consecutive years, on two populations and for 15 traits, related to berry composition, phenology, morphological and vigour. A major novelty of this study was to collect spectra and phenotypes several years apart from each other. First, we characterized the genetic signal in spectra and under which condition it could be maximized, then phenomic predictive ability was compared to genomic predictive ability.

Results: For the first time, we showed that the similarity between spectra and genomic relationship matrices was stable across tissues or years, but variable across populations, with co-inertia around 0.3 and 0.6 for diversity panel and half-diallel populations, respectively. Applying a mixed model on spectra data increased phenomic predictive ability, while using spectra collected on wood or leaves from one year or another had less impact. Differences between populations were also observed for predictive ability of phenomic prediction, with an average of 0.27 for the diversity panel and 0.35 for the half-diallel. For both populations, a significant positive correlation was found across traits between predictive ability of genomic and phenomic predictions.

Conclusion: NIRS is a new low-cost alternative to genotyping for predicting complex traits in perennial species such as grapevine. Having spectra and phenotypes from different years allowed us to exclude genotype-by-environment interactions and confirms that phenomic prediction can rely only on genetics.

Keywords: Genomic prediction; Grapevine; Phenomic prediction; Phenomics; Plant breeding; Spectroscopy.

Fig. 1

Variance components from the mixed models applied to NIRS after der1 pre-process. A in the diversity panel population, B in the half-diallel population. x and y correspond to field plot coordinates

Fig. 2

Rho-vector (RV) coefficient between the genomic relationship matrix (“snp”) and the relationship matrices derived from wood and leaf NIRS BLUPs of genotype + cross or subpopulation effects with both years included in the mixed model (“wood.2y”, “leaves.2y”, respectively). A in the diversity panel, B in the half-diallel. RV : rho-vector

Fig. 3

Predictive ability of phenomic prediction for 15 traits, in two populations with two methods, using base spectra (in golden) or spectra BLUPs (in dark blue) after der1 pre-processing. For each trait, PA distribution was over the 6 models retained for years and tissues (and also over the 10 crosses in the half-diallel)

Fig. 4

Predictive ability of phenomic prediction with a single vs. both years and tissues, over the 15 traits in both populations and the 10 crosses in the half-diallel. Prediction models were fitted with glmnet in the diversity panel (except for wood+leaves configuration) and with lme4GS in the half-diallel. In both populations, models were carried out after der1 pre-processing. The white cross indicates the average PA for each configuration

Fig. 5

Predictive ability in two settings for 15 traits with lme4GS: SNPs: GP; wood+leaves: PP with two variance-covariance matrices, for wood and leaf NIRS, for “2 years” NIRS BLUPs derived after der1 pre-process. Prediction models were fitted with lme4GS. Error bars correspond to the 95% confidence interval around the mean, based on CV repetitions

Fig. 6

Predictive ability of phenomic against genomic prediction. A in the diversity panel, B in the half-diallel. The correlation coeficient and associated p-value was computed over 15 observations (traits). In the half-diallel, PA was averaged across the ten crosses, hence standard error around each point is displayed. The red line is the regression line and the gray dashed line is the identity line