Penalized regression techniques for modeling relationships between metabolites and tomato taste attributes

Abstract

The search for models which link tomato taste attributes to their metabolic profiling, is a main challenge within the breeding programs that aim to enhance tomato flavor. In this paper, we compared such models calculated by the traditional statistical approach, stepwise regression, with models obtained by the new generation of regression techniques, known as penalized regression or regularization methods. In addition, for penalized regression, different scenarios and various model selection criteria were discussed to conclude that classical crossvalidation, selects models with many superfluous variables whereas model selection criteria such as Bayesian information criterion, seem to be more suitable, when the goal is to find parsimonious models, to explain tomato taste attributes based on metabolic information. An exhaustive comparison of the discussed methodology was done for six sensory traits, showing that the most important covariates were identified by the stepwise regression as well as by some of the penalized regression methods, despite the general disagreement on the size of the regression coefficients between them. In particular, for stepwise regression the coefficients are inflated due to their high variance which is not the case with penalized regression, showing that this new methodology, can be an alternative to obtain more accurate models.

The genetic and functional analysis of flavor in commercial tomato: the FLORAL4 gene underlies a QTL for floral aroma volatiles in tomato fruit

“…To assess chemical, sensorial and genetic diversity in cultivated tomato, a diversity panel (DP) of 94 cultivars was established by asking several leading breeding companies for cultivars of different genetic backgrounds and diverse fruit flavor characteristics. A comprehensive analysis of volatile compounds in fruits of this collection was performed and has been reported previously (Tikunov et al ., 2005; Ursem et al ., 2008; Menéndez et al ., 2012). To study the association between the genetic, chemical and sensorial variation further, a half diallel was developed by crossing the parents of four cultivars representative of the variation across the panel.…”

NON-SMOKY GLYCOSYLTRANSFERASE1 Prevents the Release of Smoky Aroma from Tomato Fruit

“…In order to find a candidate GT, an untargeted gene expression analysis using next-generation transcript sequencing called digital gene expression (DGE) was performed. Two groups, consisting of 25 'smoky' and 25 'nonsmoky' cultivars, were selected from a set of commercial cultivars characterized previously (Tikunov et al, 2005(Tikunov et al, , 2010Menéndez et al, 2012). Fruit material of each of the groups was pooled at two ripening stages: mature green (MG) and turning (T).…”

“…The marker scores corresponded perfectly with PhP-V release (see Supplemental Figure 5 online). In addition, we genotyped a diverse set of commercial tomato cultivars (F1 hybrids), which had previously been profiled for taste and aroma and for volatile and nonvolatile metabolites (Tikunov et al, 2005(Tikunov et al, , 2010Menéndez et al, 2012). The two markers revealed 100% cosegregation with the different types of PhP-V glycoconjugates and the resulting difference in the release of the corresponding volatiles (see Supplemental Figure 6 online).…”

“…Therefore, the ripe fruits of these cultivars retained the ability to release the corresponding volatiles upon fruit tissue damage. This release of PhP-V has been associated with the smoky aroma of tomato as determined using a trained fruit tasting panel (Menéndez et al, 2012).…”

Comparison of Regularized Regression Methods for ~Omics Data

“…In the context of regression, Kiers and Smilde [9] did a comparison of various multiple regression methods on simulated data with collinear variables, but their study was mainly focused on prediction and comparison of the regression coefficients when predictor variables are collinear. Menendez et al [23] reported comparison of stepwise linear regression, LASSO, EN and RR, but did not cover other penalization methods, such as SPLS, PLS, PCR, RF and SVM. We compare these methods (RR, EN, LASSO, SPLS, RF, SVM, PLS, PCR), in terms of mean square error of prediction, goodness of fit, variable selection and the ranking of the variables.…”