The MarkerLynx^TM application manager of the MassLynx^TM software (Waters Corporation, Manchester, UK) was used for raw UHPLC-MS pre-processing (matrix creation) with the software parameters set as: 3–13 min retention time (Rt) range of the chromatogram, mass range 100–900 Da, mass tolerance 0.05 Da, mass window 0.05 Da and a Rt window of 0.20 min. The data matrices obtained were exported to SIMCA (Soft Independent Modelling of Class Analogy) software, version 13.0.2 (Umetrics, Umea, Sweden) for multivariate data analysis to generate models such as principal component analyses (PCA) derived score plots and hierarchical clustering analysis (HCA). Orthogonal projection to latent structures-discriminant analysis (OPLS-DA) derived score plots and S-plots were used to identify signatory biomarkers perturbed by or associated with the treatments (Tugizimana et al., 2012). Pareto- scaled data was used to generate all models. The quality of PCA models was determined by (i) cumulative modelled variation in the X matrix, R²X (cum) also known as the goodness-of-fit parameter and (ii) the predictive ability parameter Q² (cum). Values close to 1 indicate a robust model (Supplementary Table S1). The OPLS-DA models were statistically reliable with CV-ANOVA p-value of ≤0.001 (Supplementary Table S2).

The LC-MS data were further analyzed also using XCMS (various forms of chromatography-mass spectrometry) software package¹ based on the R statistical language (Benton et al., 2008; Gowda et al., 2014). The parameters for XCMS parameters were as follows: (i) Feature detection was performed with m/z deviation of 15 ppm, maximum peak width of 20 s, (ii) Rt correction was achieved using Orbiwarp method, (iii) Alignment was performed using minimum fraction of samples of 0.5, Rt deviation of 5 s, m/z window of 0.015, (iv) statistical test was performed using the unpaired parametric t-test, at a p-value threshold of 0.05 and a fold change threshold of 1.5, (v) annotation was performed with an m/z absolute error of 0.002.

Principal component analyses is a non-supervised model that reduces the multi-dimensionality of complex data from analytical platforms. Thus, it depicts a global overview of all the similarity and/or dissimilarities between different treated sample groups (Trygg et al., 2007; Tugizimana et al., 2013; Eriksson et al., 2014). As such, the models facilitated descriptive visual evaluation of the effect of each treatment condition on C. asiatica cells. On the other hand, OPLS-DA is a supervised model that allowed the extraction of the significant biomarkers responsible for the separation of the different treated sample groups (Yamamoto et al., 2009; Tugizimana et al., 2013; Saccenti et al., 2014).

Based on the chromatograms (Figure 1), the PCA score plot (Figure 3A) reveals that there is a time-dependent response of C. asiatica cells to ASM-treatment as reflected in the clustering at different coordinates of samples harvested at different time points, whereas the loadings plot shows each variable (m/z and Rt) representing each ion (Figure 3B). The metabolome profile of the cells is already changed at 6 h post-treatment. In addition, the close clustering of the sample groups of the cells harvested 6- and 12 h post-treatment implies ongoing changes, indicative of an early response which progresses to result in a distinctive cluster at 24 h post treatment (OPLS-DA scores-plot, Figure 3C and S-plot, Figure 3D).

Multivariate data analyses of changes occurring in C. asiatica cells in response to increasing concentrations of ASM are shown in Figure 4. The PCA scores plot (Figure 4A) reflect concentration-dependent metabolic response of C. asiatica cells to ASM-treatment as seen on the chromatograms (Figure 2), whereas the loadings plot shows each variable (m/z and Rt) representing each ion (Figure 4B).

Based on the same data sets, the OPLS-DA model of control vs. 300 μM ASM treatment (Figure 4C) allows for the identification of the features (m/z ions at specific Rts) or biomarkers associated with the response of the cells towards ASM (Figure 4D).

Comparison of the S-plots generated to illustrate the separation at other time points (0 vs. 6 h and 0 vs. 12 h, data not shown) and concentration points (0 vs. 100 μM and 0 vs. 200 μM, data not shown) projected a similar pattern; i.e., showed the same significant biomarkers. However, as mentioned, the conditions that resulted in a greater response was at the highest concentration (300 μM) and the longest incubation period (24 h). The ions that are positively correlated to ASM-treatment on the OPLS-DA derived S-plot (Figure 3D and 4D) are similar to those reported in the SA-treated C. asiatica cells (Ncube et al., 2016), and the same set of ions are defined as signatory biomarkers for both time- and concentration studies. However, contrary to the results obtained from the SA study (Ncube et al., 2016), the ASM-treatment resulted in the up-regulation/ increased biosynthesis of the CGA- derivatives represented by those ions.

The UHPLC-MS generated data was further analyzed using XCMS online to compliment the OPLS-DA derived S-plots (Figures 3D and 4D) as it also allows for the extraction of statistical significant biomarkers. Similarly, the four (1–4) upregulated biomarkers with m/z = 367.0994, 515.1201, 601.1159, and 529.1414 of Rts = 7.56, 8.99, 9.56, and 10.46 min respectively, as per the S-plot, were seen on the Cloud plots as well (Figure 5). It is also interesting to note that these biomarkers appear to be responsive to the time - and the concentration of the treatment. These results also confirm that incubating the cells with 300 μM ASM for 24 h results in a change in the metabolite profile of C. asiatica accompanied by enhancement of CGA derivatives. Thus, these metabolites (Table 1) were tentatively annotated and relatively quantified in Sections “Metabolite Annotation” and “Relative Concentration of CGA Biomarkers,” respectively.

Hierarchical cluster analysis (HCA) provides a clearer overview of the relationship between groups (Madala et al., 2014a). Figures 6A,B indicates that extracts from the ASM treated cells and the extracts from cells treated with ShA alone were similar, and clustered close to the control (non-treated) cells. Overall, the treatment of the cells with ASM relatively has the least effect on the metabolite profile of the cells as these samples appear closely related to the control (non-treated cells). However, the combined treatments of [ASM + QA] and [ASM + ShA] exhibited clear differences compared to the control cells, indicating that the precursor feeding was significant in contributing to the differential clustering.

The differences between the QA and ShA treatments are noteworthy: the addition of ShA as a precursor seemingly resulted in less pronounced effects on the metabolic response of the cells. On the other hand, QA seems to result in a more pronounced effect on the metabolite profile of the cells as both the precursor molecule and the combination with ASM [ASM + QA], cluster the furthest from the control samples. Moreover, the [ASM + QA] combination appears to result in a similar response to just the QA alone, but the [ASM + ShA] combination treatment is significantly different from the individual ASM and ShA treatments of the cells (Figure 6B).

Accordingly, these findings show that QA and ShA have differential effects on C. asiatica cells, although they are both primary intermediates contributing to the phenylpropanoid pathways. As such, a study by Sonnante et al. (2010) reported that hydroxycinnamoyl-CoA quinate hydroxycinnamoyl transferase (HQT) isolated from Artichoke shows higher affinity for QA over ShA. However, these enzymatic dynamics have not been studied for C. asiatica. A separate study elsewhere confirmed that although QA and ShA are interconvertible, albeit the reation equilibration is highly dependent on a number of contributing environmental factors (Knop et al., 2001). Therefore, the clustering observed in (Figure 6B) indicates that the underlying metabolic pathways are complex and further studies are needed in order to unravel them.

In order to extract the significant biomarkers associated with each of these groups, OPLS-DA models comparing non-treated and each precursor treatment were generated (Supplementary Figure S2).

The annotation of the signatory biomarkers (1–4) extracted from the OPLS-DA derived S-plot (Figures 3D and 4D) and XCMS cloud plots (Figure 5) was validated in a similar manner as previously described (Ncube et al., 2016). Therefore, the tentative annotation (Table 1) was carried out following MSI level 2 (i.e., putatively identified, comparison of physicochemical properties and/or spectral similarity with public or commercial spectral libraries without authentic chemical standard) (Sumner et al., 2007) using an optimized UHPLC-QTOF-MS ISCID approach (Ncube et al., 2014) that has proved to effectively overcome the analytical challenges posed by the geometric and positional isomerism of these compounds (Jaiswal et al., 2014). The annotation followed the hierarchical scheme for the identification of CGAs by (Clifford et al., 2003). Here, the generation of fragmentation patterns using tandem MS (MS²), facilitated the monitoring of molecules containing a cinnamic acid moiety. For instance, metabolites with a quinic acid moiety were characterized by their fragmentation patterns with ion peaks representing Q1 [quinic acid-H] at m/z 191 and Q2 [quinic acid-H₂O] at m/z 173 wherein they were also considered for CGA derivatives along with other ion peaks representing C1 [caffeic acid-H] at m/z 179 and C2 [caffeic acid-CO₂] at m/z 135 respectively. Where available, authentic standards (particularly 3- and 5-CQA, 3,5- and 3,4-di-CQA for the current study) were also used to confirm the annotation of the signatory biomarkers. Thus, a total of four biomarkers (from the S-plots and XCMS cloud plots, Figures 3D, 4D, and 5, respectively) were annotated as described below.

As mentioned, the annotated CGA derivatives (Table 1) were also reported as significant biomarkers in the SA study (Ncube et al., 2016), and their detailed characterization is supplied in the Supplementary Figure S3.

Similarly, the annotation of CGAs was performed following the hierarchical classification scheme (Clifford et al., 2003). In addition, other significant biomarkers were either annotated based on accurate mass searches of various databases or using tandem MS spectral fragmentation patterns.

A total of 16 metabolites (13 putatively annotated) including CGA-derivatives, hydroxycinnamic acid (HCA) derivatives, terpenoids and phenolics were found to be positively correlated with the precursor/ precursor combination treatments (Table 2). As this is the first study to report on the use of precursor feeding in C. asiatica cells, a Venn diagram-based comparison was carried out in order to investigate which biosynthetic pathways are affected by each of the treatments based on the up-regulated metabolites (Figure 7).

Evaluation of the Venn diagram revealed that the HCA-derivatives and terpenoids were up-regulated by all the elicitor/ precursor combination treatments, although the former group of metabolites appear to be well represented among the different treatments. Focusing on the CGA-derivatives, only trans-5-FQA and 3-C, 5-FQA were upregulated by some treatments (i.e., [ASM + QA], [ASM + ShA] and ShA, respectively) (Table 3).

However, the peak intensities of identified biomarker metabolites in extracts from the different treated cell samples appeared to be at varying levels (reflecting different concentrations).

In general, ASM appears to have a positive effect on the annotated CGA biomarkers (Table 1). For some of the CGA-derivatives, the combined treatments (pre-induction followed by precursor feeding) appear to decrease their relative concentrations (Figure 8A). However, all the precursor/ inducer treatments appear to not induce the biosynthesis of new CGA-derivatives. According to earlier reports, the effects of exogenous application of ASM may vary depending on different pathosystems (i.e., particular host–pathogen interactions) and concentration of ASM applied to the plants (Oostendorp et al., 2001; Hong et al., 2011; Walters et al., 2013).

Based on our previous findings (Ncube et al., 2016), it is evident that the phenylpropanoid pathway in C. asiatica is more responsive to ASM than to SA. However, these inducers seemingly have different effects on the concentration of CGA-derivatives in C. asiatica cells. Therefore, the enhancement of the already existing phenylpropanoid pathway (where SA failed) is in agreement with the notion that ASM is a more effective inducer of the pathway compared to SA (Du et al., 2012; Sillero et al., 2012). In addition, the presence of only these biomarkers [1–4] in SA- and ASM-treated C. asiatica cells could be an indication that the enzymatic machinery of this plant predominately catalyze the biosynthesis of only these CGA-derivatives. Similarly to previous reports, C. asiatica cells appear to constitutively biosynthesize a relatively large pool of 3,5 diCQA [2] (Satake et al., 2007) that may be the substrate for conversion to 3,5-O-dicaffeoyl-4-O-malonylquinic acid [3] (irbic acid). Irbic acid is a species-specific compound (Antognoni et al., 2011). It is a known phenomenon that some plant species only contain a fraction of possible CGA isomers. Case in point, Mormodica species have been reported to synthesize only the 4-acyl CGAs (Madala et al., 2014b). On the other hand, some plants such as tobacco (Ncube et al., 2014) and coffee (Mahesh et al., 2007; Mehari et al., 2016) have been reported to biosynthesize diverse CGA isomers.

Another interesting point to note is that of all the phenylpropanoid molecules, only CGA-derivatives were found as significant biomarkers in both SA and ASM studies. This confirms that C. asiatica cells can be stimulated to increase the existing pool of not only centelloids but CGA-derivatives as well (Antognoni et al., 2011; Long et al., 2012; Gray et al., 2014; Maulidiani et al., 2014).

Relevant factors to consider in precursor feeding studies is the timing and the concentration of the precursor molecule to be added to a cell culture, as well as whether the particular precursor added would feed into a certain or a particular network of pathways (Jackson and Attalla, 2010; Murthy et al., 2014). Increased metabolic pools of QA and ShA might lead to corresponding increases in CGA concentrations. However, if the flux through the pathway is under strong metabolic control, negative feed-back inhibition of enzymes such as hydroxycinnamoyl-CoA shikimate/quinate hydroxycinnamoyl transferase (HCT) and hydroxycinnamoyl-CoA quinate hydroxycinnamoyl transferase (HQT) involved in CGA biosynthesis (Hoffmann et al., 2003; Niggeweg et al., 2004; Li et al., 2016), could have resulted. This could, therefore, explain the down-regulation of some of the biomarkers upon treatment with some inducer and precursor combinations (Figure 8). Further optimisation of the concentration of the inducer and the incubation period may result in increased biosynthesis of new CGA-derivatives. However, the treatment of the cells with just the ShA as a precursor or in combination with ASM can also increase the concentration of trans-5-FQA [1] and 3-C, 5-FQA [4] as shown in Figure 8.

This current study presents a biotechnological attempt to increase CGA accumulation (using chemical elicitation with ASM and precursors), but work with a similar objective has been carried out elsewhere using recombinant DNA approaches. Previous work of the latter resulting from the overexpression of the HQT gene in tomato only increased CGAs by 85% (1.85 fold) (Niggeweg et al., 2004). Constitutive expression of the artichoke HQT gene in tobacco resulted in a 3 fold increase in CGAs (Sonnante et al., 2010) and constitutive expression of the AtMYB12 transcription factor gene from A. thaliana in potato increased CGA concentrations by 3.35-fold on average (Li et al., 2016). Thus, by either recombinant DNA manipulation or chemical elicitation, the metabolic control over CGA accumulation seems to be very tightly controlled, limiting the accumulation of CGAs.

Evidently, the exogenous application of precursors resulted in a metabolic perturbation of C. asiatica cells as the treatments resulted in the up-regulation of several biomarkers (Table 2). Furthermore, the finding of rishitin in precursor-treated cells is also of particular interest as this sesquiterpene is reportedly exclusive to the plants in the Solanaceae family (Tugizimana et al., 2012). In our previous report (James et al., 2013), we reported the up-regulation of rishitin in MeJA-treated C. asiatica cells. Therefore, this finding further verifies the role of this sesquiterpene as a phytoalexin in stressed C. asiatica cells. Although the aim of this study was to increase the biosynthesis of CGAs, it is interesting to note the positive effect of all the treatments on the HCA category in particular (Table 2). These findings further clarify the role of ASM as a highly potent elicitor of phenylpropanoids as well as the prospective use of QA and ShA in precursor feeding approaches to increase the biosynthesis of phenylpropanoids.

Moreover, these results also indicate that ASM has an effect on the terpenoid pathway of the plant as rishitin along with its derivative were found to be up-regulated by this treatment. Interestingly, in our previous work, the exogenous application of MeJA to C. asiatica cells resulted in the biosynthesis of both terpenoids and phenylpropanoids (Tugizimana et al., 2015). Similarly to this cross-talk effect of MeJA in C. asiatica, the various treatments applied here resulted in the up-regulation of both terpenoids and phenylpropanoids as well.

The co-occurrence of sesquiterpenoids, pentacyclic triterpenoids, hydroxycinnamic acid derivatives and chlorogenic acids as bio-active metabolites, thus allow for this plant to cover a wider pharmacophore space.