A Chemometrics Approach for Nuclear Magnetic Resonance Data to Characterize the Partial Metabolome Banana Peels from Southern Brazil

2.1 Data Collection

2.2 Data Processing

The ¹H-NMR spectra were processed using the ACD/NMR processor software (advanced chemistry development, release 12.0) consisting of zero filling, Fourier transforming the 32 K data points, and automatically phased (Ph0 and Ph1). The baseline was manually corrected and all spectra referenced to the internal standard (TSP, ¹H – 0.00 ppm). The spectroscopy information of interest was exported as a .csv file containing a matrix with the chemical shifts (¹H pmm) and a peak intensity list. Typical resonance regions of the water and internal standard (TSP) signals removed from the dataset for further analysis. Further, each ¹H-NMR spectrum was processed using a routine implemented in the R language [23] through the package specmine [24]. Peak alignment grouped proximal peaks together according to their position using a moving window of 0.03 ppm. Peaks of the same group were aligned to their median positions across all samples. Also, missing value imputation was done filling with a constant value of 0.0005, and data pre-processing further contemplated log transformation and auto-scaling.

2.3 Chemometrics Analysis

The metadata considered in this study were the seasons of collection. However, as it was only possible to obtain one sample for the summer period, only three seasons were considered for data analysis. The season variable was therefore assigned as spring for samples from September and October (years 2010 and 2011) and November-2010; as summer/autumn for the samples from February, March, April and May-2011; and, finally, as winter for the June-2011, July-2010/2011 and August-2011 samples.

The chemometrics analysis was executed using the R specmine package. The analysis started with one-way analysis of variance (ANOVA) to test the difference in means between the metadata groups for each one of the seasons. Then, multivariate statistical analysis was performed, starting with hierarchical clustering, using an Euclidean distance between samples and complete linkage clustering, to evaluate how close samples are inside each season and between seasons. The former parameter was chosen because it tends to find compact clusters of approximately equal diameters and does not force clusters together due to single elements being close to each other, like in single linkage clustering. After that, K-means clustering was performed, by separating the samples in three or four groups, to observe if samples would be well grouped according to their seasons, either by the three seasons considered or by the four actual seasons. Finally, the analysis was followed by principal components analysis (PCA).

2.4 Metabolite Identification

To perform the identification of the metabolites present in the samples, firstly a code developed by the authors in R for the clustering of peaks, according to their correlation, was used based on the methods proposed in [25]. Each obtained cluster is considered a potential metabolite, as it can be assumed that peaks coming from the same molecule, showing similar behaviour across all samples and, therefore, correlate strongly with each other.

After this each cluster was compared with each reference metabolite, using the Jaccard index score. The Jaccard index is used to compare the similarity between sets, as it is defined by the division of the size of intersection by the size of the union of the sets. A peak tolerance of 0.03 ppm was used, as in [25] it showed to be the best threshold value to use with the reference metabolites from the database used, mentioned below.

To construct the library of reference metabolites for this study, the XML spectra files from the Human Metabolome Database (HMDB) [26], version 3.6, were downloaded and parsed. Only the metabolites whose spectra files were acquired using the ¹H-NMR technique, with a frequency of 500 MHz, were used as reference.

For each cluster, the top five reference metabolites with the best score were obtained.

Additionally, metabolite identification was performed by visual inspection of the resonances both in unidimensioal (¹H-NMR) and bidimensional (¹H/¹H, TOCSY and ¹H/¹³C, HSQC) NMR spectra.

2.5 Supplementary Material

The data used in the analysis, together with the reports generated using R Markdown are all given in supplementary material available in the URL: http://darwin.di.uminho.pt/pacbb2017/banana-nmr. This allows for the results to be understood in detail and fully reproducible.

3.1 Chemometrics Analysis

Starting our analysis of the results by observing the spectral profiles obtained for each sample collected, we could see that the samples have, approximately, the same peaks. When constructing the mean ¹H-NMR spectra for each group considered (Figure 1), we observed that the peaks did not vary much between seasons. Despite this, the intensity of some peaks varied. Thus, although the banana samples collected seems to have the same metabolites, their concentrations vary from season to season.

Figure 1:

¹H-NMR mean spectra plots for each season. Each plot was obtained from the ppm mean of the different samples for each season. (A) Spring season. (B) Summer/Autumn season. (C) Winter season.

Looking at the hierarchical clustering performed, shown in Figure 2, the samples were grouped in the considered seasons quite well. The samples from winter grouped very closely, except for the sample from July 2011, that was closer to the samples from May 2011 (summer/autumn) and September 2011 (spring). Adding this, the other three winter samples seem to be closer to the samples from April 2011, from summer/autumn and October 2011, spring, than these samples are to their groups. Furthermore, the spring samples were also well grouped, apart from the already mentioned ones from September 2011 and October 2011. On the contrary, the summer/autumn group revealed to be the most difficult one to group, not just regarding the summer sample in comparison to the autumn samples, but also between the autumn samples analysed.

Figure 2:

Dendrogram plot of the hierarchical clustering, with Euclidean distance between samples. Spring samples are in black, summer/autumn samples in red and winter samples in green.

Another clustering method performed was K-means clustering. Two executions of this method were performed, one grouping the samples into three groups and the other into four. The results from both these approaches are present in Table 1.

When performing K-means clustering, all winter samples were grouped together, along with the sample from October 2011 (spring). This was an expected result, after analysing the previous results from the hierarchical clustering. The samples from May 2011 and September 2011 were also grouped together in one of the clusters. This similarity was also previously observed in the hierarchical clustering.

When performing K-means clustering to separate samples in three different groups, the remaining summer/autumn and spring samples were grouped together. This is a result that could be expected, as autumn and spring are two seasons that normally might show similar climate conditions in southern Brazil. Furthermore, the difference for the K-means clustering for four different groups lies on the separation of this last cluster into two clusters: February 2011 and March 2011 form a separate cluster from the other samples.

Some of these inconsistencies in separating the samples in the right groups may be due to abnormal climatic conditions from the months in question. In fact, according to CPTEC/INPE (Center for Weather Forecasting and Climate Research/National Institute of Space Research) (http://clima1.cptec.inpe.br/monitoramentobrasil/pt) in October 2011 were recorded abnormally lower precipitation, while the months of June and July showed abnormally higher precipitation, revealing values very similar, which are not typical. Therefore, this might have been one of the factors to why October 2011, a spring sample, was grouped closer to the winter samples in both clustering methods.

Looking at the results from ANOVA, few were the peaks that showed a low corrected p-value, i.e. few peaks showed to have intensity means significantly different across samples from different seasons. The p-values were corrected by using the false discovery rate (FDR) method. After this correction, three peaks showed a corrected p-value below 0.1, as it can be observed in Table 2.

Furthermore, all peaks have means significantly different between the pairs of groups spring and winter, and summer/autumn and winter. Only the peak at 1.89 ppm showed a mean significantly different between summer/autumn and spring. This seems to corroborate some observations taken from the previous analysis performed i.e. the winter and spring samples are the ones that are better distinguished, and the summer/autumn samples are not so well distinguished from the spring samples. The three peaks occur in the aliphatic (1.89 ppm) and anomeric (4.01 and 4.05 ppm) regions of the ¹H-NMR spectrum.

On the other hand, the PCA analysis showed that the first three components generated led to a cumulative explanation of more than 50 % of the data variability, as it can be seen in Figure 3. Only the first component is already able to explain more than 20 % of the data variability.

Figure 3:

Screen plot of the PCA, showing the percentage of explained data variability for each principal component obtained. The blue line corresponds to the individual percentage and the red one to the cumulative percentage.

By further observing the pairs plot generated by these first three components (Figure 4), we can realize that PC1 allows the distinction of the winter and spring groups from summer/autumn, while PC2 leads to the distinction of the groups spring and summer/autumn from winter, and the PC3 discriminates the winter and summer/autumn seasons from spring.

Figure 4:

PCA pairs plot of the first three components. The variables in pink correspond to the spring group, the green ones to the summer/autumn group and the blue to the winter group.

3.2 Identified Metabolites and their Relevance

Metabolite identification was performed through a code developed in R language by the authors. According to their correlations, peaks were clustered and each set of peaks grouped has been considered a metabolite, as they come from the same molecule. Further, by comparing the clustered peaks with resonances of reference metabolites using the Jaccard index score, twenty-three metabolites have been detected as shown in Table 3.

None of the three peaks that revealed the lowest corrected p-value in the ANOVA analysis were found in a cluster. Because the code is developed so that only clusters with two or more related peaks are taken into consideration, this could mean that each one of these peaks belongs to metabolites whose spectra is only composed by one peak. In fact, after submitting each of these peaks to HMDB, the peak 1.89 matched, with a ppm tolerance of 0.03, with the peak of the acetic acid spectra (1.91 ppm). For the other two peaks, 4.05 and 4.01, the best match found was β-D-fructose (4.08 ppm, m, 3-CH; 4.05 ppm, m, 5-CH) as previously described [27]. However, the correlation between these two peaks was not considered enough, as the optimum correlation value calculated by the code (value that leads to the greater number of clusters) that two peaks have to have, at least, is 0.9, and their correlation was lower. Thus, those resonances have been tentatively assigned do that monosaccharide.

Taking into account a biochemical approach, the polar fraction of the banana peel metabolome recovered by the aqueous extracts revealed to be a chemically complex matrix, containing a series of distinct primary and secondary metabolites, e.g. carbohydrates and derivatives (sugar alcohols, e.g.), amino acids, nitrogenous bases, organic and phenolic acids, vitamin B6 (pyridoxine), and polyamines. Since the number of identified compounds allow one discussing several biochemical relevant issues to, e.g. nutritional, agronomic and pharmacological areas, in line with previous studies of our research group [1], [22], focus will be put on the metabolites with claimed effects on the wound healing process of cutaneous tissues. Thus, banana peels collected over the seasons in southern Brazil contain important bioactive compounds of interest to human health. In fact, in the Brazilian folk medicine banana peel has a history of utility to promote wound healing when used topically [1], [3].

An important class of biologically active molecules in banana peels are the phenolic compounds. In the present study, NMR spectroscopy coupled to bioinformatics tools have been able to detect caffeic acid (3, 4-dihydroxycinnamic acid), 4-hydroxybenzoic acid and syringic acid (3, 5-dimethoxybenzoic acid) in the aqueous extracts investigated. The former is originated from the mevolanate-shikimate biosynthesis pathways as the benzoic acids are produced via the loss of a two-carbon moiety from phenylpropanoids. These phenolic compounds possess antioxidant activity and anti-inflammatory properties, as well as are cytotoxic toward certain tumor cell lines [28]. Caffeic acid, for instance, is recognized as a potent antioxidant due to the delocalization of an unpaired electron caused by the extended conjugated side chain. Besides, its o-dihydroxyl group forms a hydrogen bond, which creates a more stable configuration after breaking the O-H bond [29].

Another class of secondary metabolites found in bananas are the natural amines. At least fifteen bioactive amines and derivatives haven been identified in bananas so far. Among these, the polyamines spermidine, spermine and putrescine are commonly found, being spermidine as herein shown the prevalent one, followed by its precursor putrescine [30]. These are secondary metabolites with well-known metabolic and physiological functions on the growth and development of plants, as well as presenting neuroactive, psychoactive and vasoactive effects in mammals. Importantly, polyamines are required in higher contents during periods of wound healing and post-surgery recovery [31] what could explain in certain extension the positive effects found by our research group in using standardized aqueous extracts of banana peels in the wound healing of mechanically damage-cutaneous tissue of isogenic Balb/C mice – Mus muscullus [22]. In fact, arginine, a dibasic amino acid and its metabolites nitric oxide, proline and polyamines (spermidine, spermine and putrescine, e.g.) are quite important in wound healing, affecting all phases of the process. Through the arginase pathway, L-ornithine, proline and polyamines are produced, the latter directly stimulating cell proliferation [31] and favouring the regeneration of damaged tissues. Besides, more recently, it has been shown that expression levels of heparan sulphate in human skin and wounded tissues of mouse are well correlated with polyamine contents [32].

Finally, vitamin B6 is a long known to play a critical role in protein metabolism. All the three forms of that water-soluble vitamin, i.e. pyridoxine, pyridoxal and pyridoxamine occur in the human body, being stable under acidic conditions [33]. Regarding the effect of pyridoxine on the skin wound healing process, studies have shown in rats that upon deficiency a decrease in the content, synthesis and maturation of the skin collagen is found, with negative effects on the healing of excision and incision wounds [34].