Towards a scientific interpretation of the terroir concept: plasticity of the grape berry metabolome

The fully-mature berry metabolome is principally affected by vintage

Corvina clone 48 berries were harvested at three time points corresponding to the
beginning of vèraison (that is the term used by viticulturist to indicate the onset
of ripening), mid-ripening and full maturity in seven vineyards located in the three
most important macrozones for wine production surrounding Verona (Soave, Valpolicella
and Lake Garda; Additional file 1: Table S1) during the 2006, 2007 and 2008 growing seasons. Parameters reflecting
the uniform degree of ripeness among different vineyards and growing seasons have
been reported in Additional file 1: Table S1 and, only for some of the vineyards/vintages, also in Dal Santo et al.,
2013 14].

HPLC-ESI-MS was used to characterize the non-volatile metabolites. Among 551 signals,
73 were assigned to molecules, 131 to aglycones, fragments and molecular adducts,
and the others remained unidentified. The identified metabolites included 18 anthocyanins,
13 flavan-3-ols and procyanidins, 14 flavonols and flavanols, 18 stilbenes and viniferins,
6 hydroxycinnamic acids, and a small number of sugars, amino acids and non-aromatic
organic acids. Structural characterization by MS/MS and database searching revealed
eight new molecules that were not identified in the previously-reported Corvina metabolome
15], 16]; Additional file 5: Table S3).

GC-MS was used to investigate the volatile molecules, revealing 48 identifiable molecules
in the ripe berry metabolome (Additional file 6: Table S4). Many of these molecules were sesquiterpenes (representing 40.8 % of all
the compounds identified by GC-MS). The other identifiable volatile compounds were
aldehydes (14.3 %), carboxylic acids (12.2 %), monoterpenes (8.2 %), alcohols (8.2 %),
hydrocarbons (6.1 %), esters (4.1 %), norisoprenoids (4.1 %) and other sesquiterpenoids
(2 %).

The analysis of variance (ANOVA) based on Split-plot design was preliminarly used
to retrieve all those metabolites that significantly varied through the different
vintages and producers (Additional file 7: Table S5). Considering only the identified metabolites, most of them varied according
to the vintage and the producers. Going into details, among the non volatile metabolites,
67 % of them varied according to the vintage and the 69 % according to the producers.
These variables belonged to all the main classes of metabolites. Among the volatile
metabolites, 39 % of them varied according to the vintage and 67 % of them according
to the producers. Interestingly, among the volatile metabolites the sesquiterpenes
showed the strongest modulation according to the producers. Then, the effects of vintage
and producer on the metabolite profile results to be complex to investigate. For this
reason we performed our strategy for data modeling based on orthogonal constrained
PLS-DA that allowed us to exclude the effects of vintage on the metabolite profile.

The entire HPLC-ESI-MS data set was explored by PCA. The score scatter plot shows
that PC1, explaining 31 % of the total variance, could mainly distinguish the developmental
stage, separating véraison stage from mid ripening and fully mature stages (Fig. 1a), whereas PC2 and PC3, explaining 20 % of the total variance, separated the samples
according to vintage (Fig. 1b).

Fig. 1. PCA score scatter plot of the model obtained for the metabolites detected by HPLC-ESI-MS.
Samples, corresponding to the seven vineyards (sampled in vintages 2006, 2007 and
2008 at three time points) are roughly separated according to developmental stage
(a; explained variance equal to 44 %). Stage 1: beginning of véraison; stage 2: pre-ripening;
stage 3: full maturity. PCA score scatter plot of the same data set used in (a) colored according to vintage (b; explained variance equal to 20 %). Blue: 2006; green: 2007; red: 2008. PCA score
scatter plot of fully-ripe grapes (c; explained variance equal to 35 %). Blue: 2006; green: 2007; red: 2008. Vineyards:
??=?AM; ??=?BA; ??=?BM; ??=?CS; ??=?FA; ??=?MN; ??=?PM

By applying a supervised PLS-DA approach, we obtained a reliable model (two latent
components, R
²
?=?0.55, Q
²_{6-fold CV}
?=?0.51, Q
²_{7-fold CV}
?=?0.49, Q
²_{8-fold CV}
?=?0.51) showing as expected that the fully mature berry was mainly characterized
by higher levels of anthocyanins and stilbenes, and by lower levels of hydroxycinnamic
acids and procyanindis, compared to the véraison phase (Additional file 8: Figure S1A, B).

Focusing specifically on fully-mature berries, PCA revealed that the vintage effect
was so strong that it prevented any obvious clustering according to vineyards, each
representing a specific terroir (Fig. 1c). The behavior of the 2006 vintage was intermediate between the 2007 and 2008 vintages,
as previously reported for the full transcriptomic data set based on the same biological
material 14].

PLS-DA generated a model with two components (R
²
?=?0.93, Q
²_{6-fold CV}
?=?0.92, Q
²_{7-fold CV}
?=?0.92, Q
²_{8-fold CV}
?=?0.91) that could distinguish the vintage. Analysis of the loading structure showed
that the 2008 vintage promoted the accumulation of secondary metabolites, particularly
anthocyanins and stilbenes (Additional file 8: Figure S1C, D).

The GC-MS data set for fully-mature berries was also investigated by PCA, and showed
a rough clustering based on vintage. A clearer separation was obtained by PLS-DA (three
components, R
²
?=?0.61, Q
²_{6-fold CV}
?=?0.45, Q
²_{7-fold CV}
?=?0.51, Q
²_{8-fold CV}
?=?0.41) but no metabolites were correlated strongly with a specific vintage (Additional
file 8: Figure S1E, F).

Some metabolome components show enhanced plasticity

The vintage-specific effects on the metabolite content of our berry samples masked
the other environmental effects (Fig. 1b, c). We therefore used a constrained technique to model the data, by generating latent
variables orthogonal to the vintage by oCPLS2-DA. We initially analyzed the data according
to geographical origin (the three macrozones) and then by the different vineyards
within each macrozone.

The geographical oCPLS2-DA model for non-volatile metabolites showed four components
(R
²
?=?0.79, Q
²_{6-fold CV}
?=?0.71, Q
²_{7-fold CV}
?=?0.73, Q
²_{8-fold CV}
?=?0.71). The score scatter plot in Fig. 2a shows a clear separation of the samples from each of the three macrozones. The correlation
loading plot (Fig. 2b) revealed the presence of groups of metabolites characterizing each macrozone. Specifically,
stilbenes clearly characterized vineyards located in the Lake Garda macrozone, some
flavonoids characterized Soave and Valpolicella vineyards, and the different vineyards
and macrozones were also characterized by different anthocyanins (Additional file
9: Table S6). These differences were investigated in more detail by characterizing
the putative markers of fully-mature berries listed in Additional file 9: Table S6 and assigning them to a particular chemical class (Additional file 10: Table S7). The results are shown for each of the seven vineyards in Fig. 3.

Fig. 2. oCPLS2-DA score scatter plot (a) and correlation loading plot (b) of the model for the metabolites detected by HPLC-ESI-MS. Samples, corresponding
to seven grape vineyards at three developmental stages are separated according to
the geographical macrozones, regardless of the vintage. Groups of metabolites are
depicted in different colors. Vineyards: ??=?AM; ??=?BA; ??=?BM; ??=?CS; ??=?FA; ??=?MN;
??=?PM. aa?=?amino acid; ac?=?anthocyanin; flav?=?flavonoid; hb?=?hydroxybenzoic acid;
hc?=?hydroxycinnamic acid; oa?=?organic acid; pr?=?procyanidin; s?=?sugar; st?=?stilbene
and viniferin; ui?=?unidentified

Fig. 3. Distribution of macrozone metabolic markers, determined by HPLC-MS analysis, among
the individual vineyards and in all three vintages. The markers are listed in Additional
file 9: Table S6 and are assigned to a chemical class and classified according to macrozone
relevance, as shown in Additional file 10: Table S7. Blue bars?=?2006 vintage; green bars?=?2007 vintage; red bars?=?2008 vintage.
Yellow rectangle: Lake Garda macrozone; sky blue: Soave macrozone; fuchsia: Valpolicella
macrozone. a.u. = arbitrary units

Among the stilbenes that were markers of the Lake Garda macrozone, resveratrol dimers,
trimers and tetramers (ST oligomers) were particularly associated with vineyard BA.
In contrast, the ST monomers resveratrol, resveratrol glucoside (piceide) and piceatannol
glucoside (astringin) were not identified as general markers of the Lake Garda macrozone
and were not associated with vineyard BA, but they were positively correlated with
the other Lake Garda vineyard, CS.

Among the anthocyanin markers, some Valpolicella and Soave vineyards were characterized
by acylated anthocyanins (AC1), whereas Lake Garda and Valpolicella vineyards were
characterized by some non-acylated anthocyanins (AC2), and other Valpolicella vineyards
were strongly characterized by other non-acylated anthocyanins (AC3, especially the
more decorated molecules delphinidin and petunidin). Among the flavonoid markers,
some quercetin derivatives characterized the Valpolicella and Soave vineyards (FLAV1),
one taxifolin derivative mainly characterized the Lake Garda vineyards (FLAV2), and
another putative flavanone characterized the Valpolicella vineyards (FLAV3).

Other common flavonoids, such as myricetin glycosides and various flavanones (dihydrokaempferol
and naringenin glycosides) did not strongly characterize any of the vineyards under
investigation. Furthermore, the flavan-3-ols, procyanidins and phenolic acid derivatives
did not strongly correlate with any of the samples under investigation, with the exception
of a hydroxytyrosol derivative that negatively correlated with the Lake Garda vineyards.
This indicated substantial differences between distinct classes of secondary metabolites
in terms of their ability to respond to terroir-specific environmental stimuli.

In the second data analysis step, oCPLS2-DA was applied in each of the three geographical
regions and the models showed that the producers were clearly separated from each
other (Fig. 4 and Additional file 11: Table S8). The resulting model for Lake Garda had two components (R
²
?=?0.97, Q
²_{6-fold CV}
?=?0.89, Q
²_{7-fold CV}
?=?0.92, Q
²_{8-fold CV}
?=?0.91), the model for Valpolicella had three components (R
²
?=?0.95, Q
²_{6-fold CV}
?=?0.93, Q
²_{7-fold CV}
?=?0.92, Q
²_{8-fold CV}
?=?0.92) and the model for Soave had two components (R
²
?=?0.95, Q
²_{6-fold CV}
?=?0.91, Q
²_{7-fold CV}
?=?0.91, Q
²_{8-fold CV}
?=?0.92).

Fig. 4. oCPLS2-DA models using the metabolites detected by HPLC-ESI-MS applied within each
of the three geographical regions to distinguish the vineyards. For each model, the
score scatter plot (a, c, e) and correlation loading plot (b, d, f) are provided. Samples, corresponding to seven vineyards at three developmental stages
are separated regardless of the vintage. Vineyards: ??=?AM; ??=?BA; ??=?BM; ??=?CS;
??=?FA; ??=?MN; ??=?PM. Yellow (a, b): Lake Garda macrozone; sky blue (c, d): Soave macrozone; fuchsia (e, f): Valpolicella macrozone. Groups of metabolites are shown in different colors. aa?=?amino
acid; ac?=?anthocyanin; flav?=?flavonoid; hb?=?hydroxybenzoic acid; hc?=?hydroxycinnamic
acid; oa?=?organic acid; pr?=?procyanidin; s?=?sugar; st?=?stilbene and viniferin;
ui?=?unidentified

The two vineyards in the Lake Garda macrozone were characterized by the abundance
of stilbenes (BA) and some anthocyanins and flavonoids (CS). Within the Soave macrozone,
vineyard AM was characterized by certain stilbenes, anthocyanins and flavonoids, whereas
vineyard PM was characterized predominantly by unidentified metabolites. The three
Valpolicella vineyards could be distinguished based on the content of flavan-3-ols
and procyanidins (BM), coumarated malvidin (FA) and certain stilbenes (MN).

The same oCPLS2-DA strategy was applied to the volatile metabolites detected by GC-MS.
Once again, we were able to distinguish the three macrozones and each of the vineyards
within each macrozone. The oCPLS2-DA model for geographical origin revealed five components
(R
²
?=?0.68, Q
²_{6-fold CV}
?=?0.46, Q
²_{7-fold CV}
?=?0.49, Q
²_{8-fold CV}
?=?0.45) whereas the model for the Lake Garda producers had one component (R
²
?=?0.92, Q
²_{6-fold CV}
?=?0.89, Q
²_{7-fold CV}
?=?0.91, Q
²_{8-fold CV}
?=?0.90), the model for the Valpolicella producers had five components (R
²
?=?0.93, Q
²_{6-fold CV}
?=?0.76, Q
²_{7-fold CV}
?=?0.80, Q
²_{8-fold CV}
?=?0.79) and the model for the Soave producers had three components (R
²
?=?0.95, Q
²_{6-fold CV}
?=?0.80, Q
²_{7-fold CV}
?=?0.80, Q
²_{8-fold CV}
?=?0.82) as shown in Figs. 5a and 6. The Lake Garda vineyards were best characterized by this approach, on the basis
of benzene derivatives, esters, sesquiterpenes and monoterpenes (Fig. 5b). Vineyard BA was mainly characterized by sesquiterpenes and C13 norisoprenoids,
whereas vineyard CS was characterized by certain sesquiterpenes (Fig. 6a, b and Additional file 12: Table S9). In the Soave macrozone, vineyard AM was characterized by benzene derivatives,
esters and several sesquiterpenes (Fig. 6c, d). Finally, in the Valpolicella macrozone, vineyard MN was characterized by C6 aldehydes
and C13-norisoprenoids, whereas vineyard FA was characterized by low levels of benzene
derivatives and some sesquiterpenes (Fig. 6e, f).

Fig. 5. oCPLS2-DA score plot (a) and correlation loading plot (b) using the volatile metabolites as X variables. Samples, corresponding to seven vineyards at three developmental stages
are separated according to the geographical macrozones, regardless of the vintage.
Groups of metabolites are shown in different colors. ui?=?unidentified. Vineyards:
??=?AM; ??=?BA; ??=?BM; ??=?CS; ??=?FA; ??=?MN; ??=?PM

Fig. 6. oCPLS2-DA models using the volatile metabolites applied within each of the three geographical
regions to distinguish the vineyards. For each model, the score scatter plot (a, c, e) and the correlation loading plot (b, d, f) are provided. Samples, corresponding to seven vineyards at three developmental stages
are separated regardless of the vintage. Vineyards: ??=?AM; ??=?BA; ??=?BM; ??=?CS;
??=?FA; ??=?MN; ??=?PM. Yellow: Lake Garda macrozone (a, b); sky blue: Soave macrozone (c, d); fuchsia: Valpolicella macrozone (e, f). Groups of metabolites are shown in different colors. ui?=?unidentified

Berry transcriptome analysis supports environment-dependent metabolome plasticity

In order to investigate the environment-dependent plasticity of some components of
the Corvina metabolome, we retrieved berry transcriptomic data from the seven wine
vineyards sampled in the 2008 growing season (BA, CS, BM, MN, FA, AM and PM) from
our previous work 14] in which we reported the general plasticity of the entire grapevine berry transcriptome
using the same biological material described herein. First, we inspected the expression
profiles of the Vitis vinifera stilbene synthase gene family 19] throughout our experimental design. Stilbene synthases are key enzymes catalyzing
the final step in the phenylalanine/polymalonate branch of the phenylpropanoid pathway
that eventually produces stilbenes. The heat map shows the clear upregulation of most
of the family starting from the mid-ripening stage in berries from vineyards BA and
CS (Lake Garda) and pronounced upregulation in fully-mature berries from vineyards
BM and MN, in line with the metabolomic data (Fig. 3). We then analyzed the expression profiles of the laccase gene family (Additional
file 13: Figure S2), one member of which (transparent testa 10, tt10) is involved in the oxidative polymerization of phenolic compounds in the Arabidopsis thaliana phenylpropanoid pathway 20]. Analysis using LacSubPred software 21] showed that the laccases expressed after véraison were mainly class 8 enzymes like
tt10, and were expressed differentially in berries from vineyards BA and CS, which are
characterized by stilbenes with different degrees of polymerization (Figs. 2 and 3).

The statistical approach described above was used to retrieve transcripts associated
with the geographical area regardless of the vintage. This was achieved by creating
a data set containing berry transcriptomic data representing all three developmental
stages of each vintage, sourced from three vineyards, one representing each macrozone
(CS from Lake Garda, MN from Valpolicella and AM from Soave). The data set included
292 selected genes involved in non-volatile secondary metabolism (Additional file
14: Table S10). Based on PCA results (Fig. 7b), we applied oCPLS2-DA to both the mid ripening and fully mature berries, because
the accumulation of a metabolite in fully mature fruits is often triggered by an earlier
transcriptional change (Fig. 7c, d). The score scatter plot and the correlation loading plot of the obtained model (four
components, R
²
?=?0.83, Q
²_{6-fold CV}
?=?0.70, Q
²_{7-fold CV}
?=?0.77, Q
²_{8-fold CV}
?=?0.73) are reported in Fig. 7c and d, respectively. Vineyard MN, which is associated with the positive metabolomic markers
AC1, AC3, FLAV1 and FLAV3 (Fig. 3), was also found to be associated with transcripts for the three transcription factors
VvMybA1, VvMybA2 and VvMybA3 (VIT_02s0033g00410, VIT_02s0033g00380 and VIT_02s0033g00450,
respectively), a flavonoid 3?,5?-hydroxylase (VIT_06s0009g02910) and a 4-coumarate-CoA
ligase (VIT_17s0000g01790) (Additional file 14: Table S10), all of which are active in the berry anthocyanin biosynthesis pathway
22]. This vineyard was also associated with transcripts for two flavonol synthases (VIT_13s0047g00210,
VIT_07s0031g00100) and the transcription factor VvMybF1 (VIT_07s0005g01210), which
are involved in berry flavonol synthesis 23], again supporting the metabolomic data. Similarly, vineyard CS, which is characterized
at the metabolomic level by the abundance of stilbenes (Fig. 3), was found to be associated with transcripts for the R2-R3 MYB transcription factor
VvMYB14 (VIT_07s0005g03340) (Additional file 14: Table S10) which regulates berry stilbene biosynthesis 24]. Interestingly, Soave vineyard AM lacked strongly positive transcriptomic markers,
but was associated with several negative transcriptomic markers linked to the low
level of AC2 anthocyanins (Fig. 3), including anthocyanin O-methyltransferase VvAOMT1 (VIT_01s0010g03510), MATE efflux
family protein VvAnthoMATE2 (VIT_16s0050g00910), UDP glucose:flavonoid 3-o-glucosyltransferase
VvUFGT (VIT_16s0039g02230) and anthocyanin membrane protein 1 (Anm1, VIT_08s0007G03560).
We also observed a correlation between the low level of FLAV1 molecules in berries
from vineyard CS and the presence among its negative markers of VvMyb5a (VIT_08s0007g07230),
a transcription factor involved in the general grapevine flavonoid pathway 25], 26]. When the same statistical approach was applied to a data set of selected volatile-related
transcripts, we found no correlation among the transcripts and volatile metabolites
(data not shown).

Fig. 7. Grapevine berry transcriptome analysis. Heat map of the stilbene synthase gene family
(VvSTSs) showing transcriptional profiles (a). The heat map was generated with TMeV v4.8.1 using the average expression level
of the three replicates. Data were normalized based on the mean center genes/rows
adjustment, and Pearson’s correlation was chosen as the statistical metric. PCA score
scatter plot obtained using transcripts related to secondary metabolism (b; explained variance equal to 69 %). Stage 1: beginning of véraison; stage 2: pre-ripening;
stage 3: full maturity. oCPLS2-DA score scatter plot (c) and correlation loading plot (d). Samples are separated according to the geographical macrozones, regardless of the
vintage. Vineyards: ??=?AM; ??=?CS; ??=?MN. In (d) the circles represent the macrozone, while the ? symbols represent the transcripts

Correlation between secondary metabolites and specific terroir features

We investigated the specific responses of Corvina berries to terroir-specific environmental
components by generating several classifications for each of the components described
in Additional file 1: Table S1; only some of these classifications resulted in O2PLS-DA models, reported
in Table 1, correlating the berry metabolome with the vineyard features (see Additional file
1: Table S1). We considered as reliable only those models with a Q
²_{7-fold CV}
value greater than 0.5 that also passed the permutation test on the response (400
random permutations). Table 1 shows a complete list of the models we tested and their relative Q
²_{7-fold CV}
values, with the reliable models highlighted in bold.

Table 1. List of the OPLS-DA models that were validated using a cross-validation test with
200 permutations, showing the classes that were used

The loading plot showed that no individual metabolites correlated strongly (pq(corr)
0.75) with any of the terroir features. Even in the best models for metabolites detected by LC-MS (macrozone, Lake
Garda vs. others, soil total lime, soil clay, soil exchangeable potassium) and by
GC-MS (macrozone, Lake Garda vs. others, training system and soil active lime) there
was only a low correlation between individual metabolites and specific terroir features (Additional file 15). The existence of reliable O2PLS-DA models lacking strong characteristic metabolites
suggests that the observed correlations between specific terroir features and the berry metabolome probably reflect many small metabolomic changes
rather than a small number of major metabolic shifts. In the context of these slight
correlations, we found once again that flavonoids and stilbenes were assigned to the
more plastic fraction of the metabolome, given that some flavonoids correlated with
the Valpolicella and Soave macrozones, low soil clay, total lime and exchangeable
potassium, whereas stilbenes correlated with the Lake Garda macrozone, low soil clay
and an average amount of soil exchangeable potassium.