Introduction
Viticulture is facing growing environmental challenges, particularly increased drought stress driven by climate change, alongside rising societal pressure to reduce pesticide use (Mwaka et al., 2024; van Leeuwen et al., 2024). Breeding cultivars is one of the most effective strategies to cope with these issues, typically requiring intraspecific or interspecific crossings and a long selection process (Gambetta et al., 2020; Pirrello et al., 2023). While marker-assisted selection can be routinely applied to develop new varieties with tolerance to major fungal diseases such as downy mildew (Plasmopara viticola) and powdery mildew (Erysiphe necator) (Pirrello et al., 2023), developing tolerant varieties to drought remains more difficult due to the multigenic nature of this trait (Gambetta et al., 2020). Consequently, selecting for such complex traits requires long, time-consuming phenotyping steps (Töpfer et al., 2011). Besides tolerance, berry aroma composition is another key character to take into consideration in the selection process as it contributes to wine quality and sensory perception (Zhu et al., 2016).
The concentrations in aroma compounds can be as low as the ng L
-1 or ng kg
-1 range, meaning that sample preparation and analyte preconcentration are often as critical as the analytical step itself (Marín-San Román et al., 2020). Regardless of the extraction method, analysis is mostly performed by gas chromatography (GC) or multidimensional gas chromatography (GC-GC) coupled with various detectors, including mass spectrometry (MS), flame ionisation detection (FID) or olfactometry (O) (Welke et al., 2022). These GC approaches share several limitations, including long analysis times (45 min to 2 h) and often complex method development (
Chin & Marriott, 2015). Chemical ionisation mass spectrometry (CIMS) techniques such as selected ion flow tube mass spectrometry (SIFT-MS) have been reported as alternatives to minimise sample preparation and enable gas-phase analysis within a few minutes (
Smith & Španěl, 2005). This instrument relies on soft ionisation using up to eight reagent ions in the most recent systems (H
3O
+, NO
+, O
2
+, NO
3
-, NO
2
-, O
-, O
2
-, and OH
-) and allows analysis of sample headspace with determination of ion abundances. It can be operated either in scan mode, covering the full mass range of expected ions, or in selected ion monitoring (SIM) mode, in which specific masses are targeted to achieve quantification at the parts-per-billion (ppb) level (
Smith & Španěl, 2005). The reagent ions are generated by a microwave plasma, and ion switching occurs within milliseconds.
The instrument has previously been applied in i) environmental studies, including air quality monitoring and the detection of hazardous chemicals in ambient air and soils, ii) healthcare for human breath analysis and the identification of disease-related biomarkers, and iii) food science for the detection of spoilage and for product quality assessment (Smith et al., 2025).
Despite its broad range of applications, the use of SIFT-MS to analyse grape berries, wine, and other alcoholic beverages remains very limited. In that field, SIFT-MS has recently been applied to analyse twenty-three grape varieties collected from a germplasm collection (Baerenzung dit Baron et al., 2022). This study demonstrated that the use of O
2
+ as a reagent ion allowed rapid discrimination of grape varieties based on their production levels of monoterpenols, C13-norisoprenoids, volatile phenols, small alcohols, and aldehydes. Interestingly, in most cases, varieties clustered according to their parentage relationships. Another field study conducted under warm and dry climatic conditions using the same model showed that the SIFT-MS volatilome remained stable from the second phase of maturation (Geffroy et al., 2022). This stability facilitates sample collection, which can therefore rely solely on phenological stage and elapsed time after mid-veraison. More recently, it has been shown that this instrument can be used to assess the aroma potential of berries treated with β-glucosidase, and that hybrids genetically distant from
Vitis vinifera (
e.g., Clinton,
Vitis riparia ×
Vitis labrusca) can be easily distinguished (Baerenzung dit Baron et al., 2026).
Such high-throughput technology could be an effective tool for phenotyping the large numbers of seedlings produced during breeding programs. During early growth stages, these individuals are typically grown in pots in a greenhouse, providing semi-controlled conditions. At later developmental stages, such conditions allow for the precise application of environmental stresses, particularly water stress, which is highly relevant in studies targeting adaptation strategies to climate change. To date, no research has examined the SIFT-MS volatilome of berries produced under these conditions or assessed the instrument's ability to discriminate between varieties.
This study aimed to investigate the SIFT-MS volatilome, defined here as the set of
m/z signals detected in the berry headspace from a large set of
Vitis vinifera cultivars (n = 141) selected from a diversity panel and grown under greenhouse conditions, leveraging the high-throughput phenotyping capacity of SIFT-MS as an attractive alternative to conventional targeted approaches for large-scale screening purposes. Due to the limited availability of grape material, no more than two berries were analysed per sample.
Materials and methods
1. Plant material and sample collection
The five-year-old own-rooted plants used for this experiment were the same as those used and described in
Coindre et al. (2026), who studied leaf ecophysiological traits. One hundred and forty-one genotypes were selected for their berry availability from a panel of 279 designed to represent the worldwide genetic diversity of
Vitis vinifera (Nicolas et al., 2016). The list of these genotypes, their belonging to three previously described subpopulations (
e.g., Wine West, WW; Wine East, WE; and Table East, EE), their accession ID (Vassal), and their skin colour are shown in Table S1. It should be noted that such subpopulation differentiation, derived from microsatellite-based genetic diversity analyses, does not fully correspond to use patterns or geographic locations.
They were grown from May to August 2022 in a greenhouse in individual 9-L pots with a substrate composed of a mixture of loamy soil and organic compost on the PhenoArch high-throughput phenotyping platform at the Montpellier Plant Phenotyping Platforms (Cabrera‐Bosquet et al., 2016). Water was supplied at 1.05 g H
2O g
-1 dry soil to maintain a moderate water deficit, simulating early field drought conditions with moderate effects on leaf functioning. For seven cultivars, namely Alexandroouli, Bou Khanzir noir, Camaraou noir, Corbeau, Kolontar, Morrastel and Pagadebiti (Table S1), berries were sampled from two different plants over the same sampling period, each providing one pooled sample of two berries, to provide a preliminary assessment of the robustness of the SIFT-MS volatilome to inter-plant biological variability, not an exhaustive estimate of this variability for all studied genotypes. Such cultivars were selected on the sole basis of the availability of plant duplicates within the experimental setup.
Air temperature and relative humidity were continuously monitored using HC2-SH sensors (Rotronic, Bassersdorf, Switzerland), while photosynthetically active radiation (PAR) was measured using SKP 215 sensors (Skye Instruments, Llandrindod Well, United Kingdom), with instruments installed at ten locations throughout the greenhouse. Environmental conditions recorded during the June–July period, which corresponds to maturation, were 17 mol day⁻¹ for PAR, 1.58 kPa for average maximum daily vapour pressure deficit and 19.0 to 29.3 °C for mean daily air temperature. Exactly twenty-one days after the onset of veraison, as determined for each genotype, two berries per genotype were collected. This choice of two berries per sample was motivated by strong and deliberate experimental constraint given that the plant material came from single potted plants whose primary purpose was ecophysiological phenotyping (Coindre et al., 2026) rather than berry production in quantity. The two berries collected per genotype were pooled into a single sample prior to headspace SIFT-MS acquisition. Because of genetic variability in phenology, berry harvest spanned eight periods: P1 (12–14 June), P2 (21–23 June), P3 (28 June–1 July), P4 (4–8 July), P5 (11–15 July), P6 (19–22 July), P7 (26–29 July), and P8 (3 August). All berries were stored at –18 °C in 50 mL Falcon tubes prior to analysis.
2. Chemicals
Ultra-pure water from a Milli-Q system (Millipore, Burlington, Massachusetts, USA) was used in the study. N
2 was used as the SIFT-MS carrier gas and to compensate for the pressure drop induced by aspiration (Alphagaz, Air Liquide, 99.9999 %, Paris, France). Certified gas standards for SIFT-MS calibration were provided by Air Liquide America Speciality Gases LLC (Plumsteadville, PA, USA). Tartaric acid (≥ 99.0 %), sodium tartrate (≥ 99.0 %), sodium hydroxide (≥ 98 %), and sodium chloride (≥ 99.5 %) were purchased from Thermo Fisher Scientific (Illkirch, France).
3. Basic berry parameters
For each variety, the two berries were weighed after thawing, pooled and homogenised together into a single sample using an Ultra-Turrax T18 molecular grinder (IKA, Staufen, Germany). The decision to pool the two berries was made to obtain a satisfactory analytical signal given the limited amount of material available. The resulting must was analysed for soluble solids content (°Brix) using a digital MA885 Wine Refractometer (Milwaukee, Wisconsin, USA) and for pH using a PHM 210 MeterLab pH meter (Radiometer, Copenhagen, Denmark).
4. Sample preparation for SIFT-MS analysis
Following measurement of basic berry parameters, fifteen mL of tartrate buffer (pH 3.2) were added per 3 g of berries, and samples were subjected to a second grinding step to ensure complete tissue disruption. The tartrate buffer (pH 3.2) was prepared by dissolving tartaric acid in ultra-pure water and adjusting the pH with NaOH. Extracts were transferred to 100 mL Drechsel-style gas washing bottles (Servilab, Le Mans, France). Sodium chloride was then added (4.5 g per 3 g of berries, corresponding to 300 g·L⁻¹), and the mixture was sonicated for 15 min and heated at 70 °C for 40 min under continuous rotation (100 rpm). Under these conditions, without the use of exogenous β-glycosidase enzymes, SIFT-MS measures free volatile compounds rather than bound forms or precursors.
5. SIFT-MS analysis
Following this equilibrium period, the bottle containing the extract was connected on one side to a VOICE 200 Ultra instrument (Syft Technologies, Christchurch, New Zealand) and on the other side to a nitrogen source to enable dynamic analysis. The instrument was operated using nitrogen as the carrier gas in full scan mode (
m/z range: 15–250) with O
2⁺ as previously proposed (Baerenzung dit Baron et al., 2022). Samples were introduced through a calibrated capillary at 0.3 Torr·L·s⁻¹. Inlet and aspiration flow rates were balanced to minimise vacuum. The internal SIFT-MS flow rate was set at 2.0 Torr·L·
s-1 with the reaction between the analytes and the selected precursor occurring in the flow tube, maintained at 119 °C and 0.06 kPa. Instrumental repeatability and reproducibility over 3 consecutive days were estimated at 7 % and 10 %, respectively. For each sample, four full mass scans were recorded, of which the last three were retained for analysis.
Data acquisition and analysis were performed using LabSyft 1.6.2 software (Syft Technologies). A calibration procedure was performed before each instrument start to ensure linearity, range, accuracy, and precision by injecting a mix of certified gas standards.
In comparison with the previously published method (Baerenzung dit Baron et al., 2022), the modifications in container volume (100 mL vs 1 L), temperature (40 °C vs. 70 °C), and sampling mode (dynamic
vs. static) were implemented based on preliminary experiments to get sufficient volatiles in the headspace, detectable and repeatable signal from a limited amount of material (2 berries in most cases
vs. 50 g). Triplicate analysis of a thawed
Vitis
vinifera L. cv. Syrah grape sample using the two systems showed a significant correlation (
p-value < 0.001) for abundances measured in the
m/z range 100-250, with Pearson and Spearman correlation coefficients of 0.44 and 0.60, respectively. This indicates a strong monotonic relationship between the two methods, meaning that compounds with high abundance in one model tend to have high abundance in the other, even though the relationship is not strictly linear or proportional. The absence of a stronger relationship may reflect the higher concentrations at 70 °C in the headspace of heavier, less volatile compounds, such as sesquiterpenes (
Copolovici & Niinemets, 2015) (data not shown).
6. Data treatment & statistics
Averaged SIFT-MS data across the three technical replicates were pre-processed by removing masses with
m/z below 100 and abundances below the noise threshold, following the methodology described by Baerenzung dit
Baron et al. (2022). Signal intensities for 128 different
m/z ratios, collectively referred to hereafter as the berry volatilome or volatile fingerprint, were standardised by berry weight to enable comparison between cultivars. These
m/z values correspond to raw, unannotated SIFT-MS signals rather than to formally identified volatile compounds or characterised ion fragments. For each cultivar with plant duplicates, Spearman correlation coefficients and relative differences (median and standard deviation) between replicates were calculated to assess biological variability in the SIFT-MS data following recommended practices for mass spectrometry-based metabolomics data (Ghosh et al., 2021). For variability in basic berry parameters, absolute differences were calculated for sugar concentration and pH, while relative differences were used for berry weight.
The data, including those averaged across replicates for the variety concerned, were then subjected to Principal Component Analysis (PCA), followed by Hierarchical Clustering Analysis (HCA) using Euclidean distance and Ward’s method, a common approach for multivariate analysis and clustering (
Hair, 2011). The optimal number of clusters was chosen to maximise the silhouette and Calinski–Harabasz indices. One-way ANOVA and chi-square test followed by a Marascuilo post hoc test were performed to detect whether abundances, basic berry parameters, and the proportion of genotypes belonging to the three genetic subpopulations collected over the eight sampling periods significantly differed between the clusters identified by HCA. For ANOVA performed on abundances, p-values were adjusted for multiple testing using the false discovery rate (FDR) method. A heatmap was also generated from the dataset, an approach previously implemented for SIFT-MS data for overall visualisation purposes (Dryahina et al., 2016), with hierarchical clustering applied to both samples and variables.
All data analyses were performed using XLSTAT 3.1 software (Addinsoft, Paris), except for PCA and the heatmap, which were created using ClustVis and DataMap (
Ge, 2025;
Metsalu & Vilo, 2015).
Results and discussion
1. Basic berry parameters and SIFT-MS data for the 141 genotypes and their variability across biological replicates
Average basic berry parameters determined on the sampled berries for the 141 genotypes provide an overview of grape maturity at the time of harvest and enable contextualisation of SIFT-MS data. They were 14.1 °Brix ± 3.2 for sugar concentration, 3.64 ± 0.33 for pH, and 3.60 g ± 1.78 for berry weight (Table S1). To our knowledge, no other study has been conducted under a comparable experimental context and sampling timing. However, the sugar concentrations obtained are consistent with those reported in a field trial, where berries sampled 30 days after veraison showed sugar concentrations between 17 and 18.5 °Brix (Geffroy et al., 2014). These results highlight an overall low level of maturation, which might be the consequence of the early sample collection that occurred 21 days after the onset of veraison. The decision to sample individual berries within this post-veraison delay was supported by an earlier field study demonstrating that such a period was adequate for sampling single berries within their active period of sugar accumulation, when sugar concentration rapidly increases from 200 to 1000 mM (
i.e., from 3.5 to 18.5 °Brix) (Tavernier et al., 2025), as confirmed here. The discrepancies in sugar concentration across cultivars may be largely explained by the unknown delay between the veraison of the first berry and the sampled one. However, this sampling procedure warrants that samples did not reach the berry shrivelling period after the cessation of phloem import (Savoi et al., 2021), which is marked by abrupt transcriptomic and metabolic changes (Tavernier et al., 2025). It should be noted that, although this sampling time point was not specifically investigated for VOC dynamics, it reflects a trade-off between physiological relevance and experimental constraints, as the berries were also used for additional phenotypic measurements.
The relatively high average berry weight, a trait influenced not only by genotype but also by environmental factors such as water availability (Dai et al., 2011), with maximum values reaching 12.2 g for Angoor Kalan, suggests that water was not limiting under the conditions of this experiment. It should be noted that pH values may have been biased and overestimated, because berries were stored at –18 °C prior to analysis (Santesteban et al., 2013). However, such a process is expected to remove proportionally more K⁺ than tartaric acid, leading to a decrease in pH. The relatively high pH values observed may therefore result from intense tissue disruption during homogenisation and/or elevated potassium levels, potentially linked to fertilisation conditions.
These low sugar levels may have influenced the SIFT-MS data collected, which have previously been shown to remain relatively stable during the later stages of maturation, particularly from 28 days after mid-veraison under field conditions, although under specific sample handling conditions (Geffroy et al., 2022). However, direct comparisons of timing are difficult, as the earlier study assessed mixed berry populations, whereas our work focused on two berries. While a formal study under greenhouse conditions would be needed to evaluate the effect of post-veraison timing fully, preliminary insights can be drawn from the metrics used to assess biological variability, as presented in Table 1.
Table 1. Within-cultivar variability of basic berry parameters and SIFT-MS data between biological replicates (n = 2). Metrics are expressed in absolute difference for sugar concentration and pH, and in relative difference for berry weight. RD stands for relative difference.
|
Full cultivar name
|
Basic berry parameters
|
|
SIFT-MS abundance data
|
|
Sugar content
(°Brix)
|
pH
(pH unit)
|
Berry weight
(%)
|
|
R
Spearman
|
RD (%) Median
|
RD (%) Standard deviation
|
|
Alexandroouli
|
0.6
|
0.61
|
12.5
|
|
0.9882
|
18.6
|
17.7
|
|
Bou Khanzir noir
|
2.0
|
0.03
|
47.9
|
|
0.9763
|
27.4
|
20.9
|
|
Camaraou noir
|
4.9
|
0.13
|
26.0
|
|
0.9849
|
44.0
|
25.0
|
|
Corbeau
|
0.0
|
0.01
|
33.1
|
|
0.9787
|
22.8
|
25.0
|
|
Kolontar
|
3.7
|
0.13
|
33.9
|
|
0.9632
|
41.9
|
32.9
|
|
Morrastel
|
1.7
|
0.15
|
11.5
|
|
0.9656
|
30.6
|
26.2
|
|
Pagadebiti
|
4.9
|
0.23
|
8.6
|
|
0.9916
|
16.0
|
15.9
|
The seven replicate plants enabled assessment of the biological variability for basic berry parameters and SIFT-MS data (Table 1). Sugar concentration showed greater variability for Camaraou and Pagadebiti, with an absolute difference of 4.9 °Brix between replicates. However, despite this variability in sugar content, Pagadebiti exhibited low variability across most other metrics, particularly those related to SIFT-MS signal intensities. This apparent decoupling between sugar concentration and SIFT-MS measurements suggests that, within a given genotype, the berry volatilome is only weakly influenced by sugar concentration within the studied range.
It must be pointed out that Alexandroouli also displayed low overall variability, except for pH. The higher variability observed for this parameter suggests that pH values should be interpreted with caution, as discussed previously.
Interestingly, the variability among plants was low for the SIFT-MS data, as reflected by high Spearman correlation coefficients ranging from 0.9632 for Kolontar to 0.9916 for Pagadebiti, along with low relative differences, with median values varying between 16.0 % for Pagadebiti and 44.0 % for Camaraou. Although the moderate inter-plant variability observed for the seven duplicated cultivars suggests reasonable reproducibility of the SIFT-MS signal despite the limited number of berries analysed per sample, results obtained at the individual cultivar level should be interpreted with caution, as two berries cannot fully capture the biological variability inherent to a given genotype. A higher number of biological replicates would be required to support any fine-scale varietal characterisation. Overall, this reproducibility across biological replicates complements the analytical reliability of each individual measurement, which itself results from averaging three consecutive technical mass scans per sample after discarding the first scan for signal stabilisation, with an estimated instrumental repeatability of 7 %
2. PCA and HAC clustering results
Principal Component Analysis (PCA) was performed on the full set of 128
m/z ratios retained after filtering, in order to explore in an unsupervised manner whether genotypes could be structured into distinct groups based on their overall SIFT-MS volatilome. Hierarchical Clustering Analysis (HCA), applied to the resulting PCA coordinates, identified two genotype clusters (C1 and C2), presented in Figures S1 and S2 and summarised in Table S1. A heatmap was additionally generated directly from the full dataset, independently of the PCA/HCA workflow, providing a complementary visualisation of
m/z abundance patterns across samples. Minor differences in sample groupings observed between the two approaches likely reflect the heatmap's greater sensitivity to a small number of dominant
m/z values, whereas PCA/HCA integrates variance across the full set of variables. Surprisingly, given the genetic variability of the studied varieties and the expected distinct composition in aroma compounds arising from it (Robinson et al., 2014), only two homogeneous clusters, hereafter referred to as Cluster 1 (C1, n = 91) and Cluster 2 (C2, n = 50), were identified by PCA followed by HAC. In previous fieldwork conducted on 23 varieties sampled from a germplasm collection, the same method identified four distinct clusters, highlighting more pronounced differences within the dataset (Baerenzung dit Baron et al., 2022). Rather than indicating a lack of discrimination, the observed binary structure may reflect the existence of a biological continuum within the studied dataset. Such a continuum is consistent with the genetic design of the panel specifically selected to represent a wide range of
Vitis vinifera diversity (Nicolas et al., 2016). For genetically diverse populations, metabolomic traits often vary along gradients rather than forming discrete classes (
Hall, 2006). This continuum can also be observed in the PCA plot with confidence ellipses for the two clusters slightly overlapping (Figure 1A). In addition, the relatively homogeneous environmental conditions (temperature, light, water and nitrogen supply) under which plants were grown may have revealed a limited level of genetic variability in the berry volatilome, as environmental effects known to influence aroma biosynthesis (van Leeuwen et al., 2020) were minimised. Furthermore, extraction under buffered conditions when preparing samples for SIFT-MS analysis may have contributed to reducing additional sources of variability.
It can be pointed out that the obtained PCA provides a reliable representation of the sample distribution, with PC1 and PC2 explaining 69.5 % of the total variance. PC1 showed strong positive loadings for most
m/z variables, suggesting that it captures overall signal intensity variation across samples, while PC2 differentiated groups of
m/z values with contrasting contributions, reflecting compositional differences in the volatilome.
Figure 1. Factor scores with 95 % confidence ellipse for a principal component analysis (PCA) performed on the SIFT-MS abundance data using O
<sub>2</sub>
<sup>+</sup> (n = 141). Ellipses represent groups by A) cluster identified through PCA followed by HAC and B) subgroup panel.
![]()
3. SIFT-MS spectral results for the two clusters identified
The mean SIFT-MS spectrum obtained for the two clusters (Figure 2) enabled the visualisation of differences in
m/z abundance between the two clusters. Overall, the two spectra show comparable general patterns, with abundances showing the highest values at low
m/z (100–150), gradually decreasing toward higher
m/z, with
m/z above 200 showing the lowest signal. Overall, this pattern is consistent with previous studies (Baerenzung dit Baron et al., 2022; Geffroy et al., 2022), which demonstrated the ability of SIFT-MS to discriminate between low and high producers of volatile compounds, although the peaks observed around
m/z 196-201 have not been previously described.
Figure 2. Mean SIFT-MS spectrum obtained with O
<sub>2</sub>
<sup>+</sup> reagent ion for all the sampled varieties combined in C1 (n = 91) and C2 (n = 50). Error bars are 95 % confidence intervals. The y-axis is displayed on a logarithmic scale.
Une image contenant texte, capture d’écran, conception Le contenu généré par l’IA peut être incorrect.
SIFT-MS signals observed with O
2
+ can be related to molecular ions (M⁺) formed by charge transfer, fragments or water adducts (M⁺·(H
2O)
n) (
Smith & Španěl, 2005). The peaks around
m/z 196–201 could correspond to molecular ions of sesquiterpenes, oxygenated terpenols (
e.g., geranyl acetate or linalyl acetate), or C13-norisoprenoids, some free compounds found in grapes (Robinson et al., 2014) with nominal masses in this
m/z range whose relatively low volatility due to higher molecular mass may have been increased by change in methodology notably in equilibrium temperature (70 °C
vs. 40 °C).
For the 128
m/z ratios studied, with the exception of
m/z 114 for which no significant difference was observed, abundances were significantly higher (
adjusted p < 0.001, FDR) for C2 with an average 1.91-fold increase (± 0.20) compared to C1. Among the most cluster-specific signals,
m/z 140, 149, 105, 191, and 159 showed the strongest differential abundances, potentially reflecting variety-specific compounds that may be tentatively associated with fragments or molecular ions of terpenoids or norisoprenoids, as reported in previous SIFT-MS studies (Baerenzung dit Baron et al., 2022). In contrast,
m/z 114, 214, and 172 were the most stable abundances, suggesting signals related to core metabolites common to both clusters.
The molecular identification of these SIFT-MS markers remains challenging. Each observed signal may arise from a molecular ion, a fragment, or a rearrangement product, and compounds of the same chemical family with similar masses can produce identical fragment ions (Baerenzung dit Baron et al., 2022). Even standards provide limited guidance, as they cannot capture all fragmentation pathways or overlapping signals. Moreover, SIFTMS signals can appear ± 1–3 m/z from the nominal value due to multiple product ions and isotopologue contributions, and unit mass resolution can shift peaks relative to theoretical masses (
Smith & Španěl, 2005). Further work using GC-MS would be valuable, although such an approach would be challenging to implement at such a large scale. In the same way, the assessment of genetic associations between these cluster-specific signals and molecular markers characterising the diversity panel would be useful to further investigate these relationships.
4. Sub-population analysis and further discussion
The proportion of samples by skin colour, subgroup panel, and sampling period (P1 to P8) for the two clusters (Table 2) enabled the discussion of the observed clustering. Interestingly, C2 included a larger proportion of varieties from the WW subgroup and a lower proportion from the TE subgroup (Table 2), suggesting that wine-oriented varieties may exhibit higher overall concentrations of aroma compounds, although these subpopulation assignments do not fully correspond to use categories. However, this pattern is not clearly visible in the heatmap (Figure S2); the PCA plot, where samples tend to be grouped by subgroup (Figure 1B), is consistent with this assumption and reveals a similar pattern for the WE subgroup. A one-way ANOVA using panel subpopulation as a qualitative factor revealed that 39
m/z ratios were significantly different (
adjusted p < 0.05, FDR), with TE generally showing lower abundances, WW higher values, and WE intermediate levels (results not shown).
A first explanation could be that centuries of wine grape selection favoured genotypes with higher concentrations of aroma compounds, whereas breeding of table grapes primarily targeted other traits. A second hypothesis relates to differences in berry maturity and higher average sugar concentration observed in C2 varieties (15.1 °Brix ± 3.0
vs 13.6 °Brix ± 3.3 for C1,
p < 0.01), consistent with higher pH values (3.77 ± 0.27
vs 3.57 ± 0.34 for C1,
p < 0.01), which could have enhanced the richness in aroma compounds. However, this explanation appears unlikely, as it has been demonstrated that most
m/z signals decreased during maturation, possibly due to volatile emission or the accumulation of bound, non-detectable forms (Geffroy et al., 2022). Moreover, this interpretation is inconsistent with our previous discussion based on variability metrics, which suggested only a limited influence of sugar concentration on SIFT-MS volatile fingerprints. A third and more plausible explanation lies in the significant difference in average berry weight between clusters (2.52 g ± 0.78 for C2
vs 4.18 g ± 1.90 for C1,
p < 0.001). Indeed, the TE subpopulation has already been shown to exhibit higher berry weights than WW (Nicolas et al., 2016), a pattern that was confirmed in our study, where berry weight decreased significantly (
p < 0.001) from TE (5.46 ± 2.64 g) to WE (3.60 ± 1.35 g) and WW (2.81 ± 0.86 g). As SIFT-MS data were standardised per berry weight, smaller berries are likely to contain a larger proportion of skin per unit mass, which may have induced higher signals. Indeed, the berry skin is known to be particularly rich in phenolic and aroma compounds (Urcan et al., 2016). This latter hypothesis does not fully contradict the first one, as historical selection for smaller berries in wine grapes may have favoured higher concentrations of these compounds.
Table 2. Proportion (expressed as %) of samples by skin colour, subgroup panel, and sampling period (from P1 to P8) for the two clusters identified by PCA followed by HCA, and significance (
<italic>P</italic>) calculated from the chi-square test.
|
Cluster
|
White
|
Red
|
TE
|
WE
|
WW
|
P1
|
P2
|
P3
|
P4
|
P5
|
P6
|
P7
|
P8
|
|
C1
|
52
|
43
|
24
|
42
|
34
|
2
|
4
|
5
|
15
|
35
|
27
|
9
|
1
|
|
C2
|
56
|
44
|
8
|
30
|
62
|
0
|
2
|
10
|
16
|
12
|
40
|
16
|
4
|
|
P
|
0.620
|
0.896
|
0.018
|
0.168
|
<0.01
|
0.291
|
0.462
|
0.319
|
0.903
|
<0.01
|
0.127
|
0.197
|
0.253
|
To obtain preliminary insights, although the number of seeds was not recorded in our experiment despite being a key parameter known to influence berry weight (Dai et al., 2011), berry volume (V) was estimated for each variety based on sugar concentration and corresponding density values reported in OIV tables (
OIV, 2021). Berry diameter (d) was then calculated using the formula V = πd
3/6. From this estimation, berry skin surface area (S) was derived using S = πd
2, allowing the determination of skin surface per unit mass. This variable significantly discriminated between clusters (
p < 0.001), with values of 3.02 cm
2 g⁻
1 ± 0.42 for C1 and 3.50 cm
2 g⁻
1 ± 0.37 for C2. Overall, C2 exhibited values approximately 1.16-fold higher than C1. However, this difference is markedly smaller than the 1.91-fold increase observed in SIFT-MS abundances across all
m/z. This suggests that skin surface area alone cannot fully account for the differences in abundances between clusters. Together, these results support the hypothesis of enhanced overall volatile compound production in varieties belonging to the C2 cluster, which includes a greater proportion of WW varieties.
It is also worth noting that cluster C1 included a significantly higher proportion of varieties sampled at P5, whereas the greatest proportion of C2 were harvested later at P6. This pattern suggests that C1 varieties reached maturity earlier in the season, which is consistent with previous observations for table grape varieties (Nicolas et al., 2016) that are more frequently represented in C1. Therefore, C1 varieties may have also been exposed to cooler conditions during maturation. The absence of significant differences between clusters in the distribution of varieties across the other sampling periods, together with the lack of clear clustering in the heatmap (Figure S2), further supports the hypothesis that environmental conditions had a limited impact on the dataset, as expected under semi-controlled conditions. Consistent with this, a one-way ANOVA using sampling period as a qualitative factor revealed a significant but limited effect on sugar concentration (
p < 0.05), driven by higher values at P6 (15.4 ± 3.0 °Brix) and P7 (15.4 ± 3.7 °Brix) compared to P3 (11.8 ± 3.2 °Brix), with no significant differences among the remaining periods. This moderate increase likely reflects the progressive rise in greenhouse temperature over the June-August period. pH also varied significantly across periods, with the highest values at P1 (4.10 ± 0.00, n = 2) and the lowest, as for sugar concentration, at P3 (3.36 ± 0.46). However, unlike sugar concentration, pH did not follow a comparable temperature-related trend across the season, suggesting that this period effect likely reflects differences in the specific set of genotypes sampled within each period. Such effects are unlikely to have affected the SIFT-MS-based clustering, given the weak coupling between sugar concentration and volatilome already noted among replicate plants (Table 1).
Finally, it should be noted that the thermal treatment applied (70 °C, 40 min) to obtain a detectable and reproducible headspace signal from a limited amount of berry material may have altered the native volatilome of fresh berries, notably through partial thermal hydrolysis of esters, degradation of heat-sensitive sulphur compounds, and the possible formation of newly generated compounds via Strecker/Maillard-type reactions. As a result, the SIFT-MS volatile fingerprints obtained here are not strictly representative of fresh berry aroma. However, since all samples underwent a rigorously identical thermal treatment, the relative comparisons between cultivars and clusters which form the core of the present results remain valid.