Trends in Food Science & Technology 72 (2018) 83–90

Contents lists available at ScienceDirect

Trends in Food Science & Technology

journal homepage: www.elsevier.com/locate/tifs

Use of principal component analysis (PCA) and hierarchical cluster analysis T

(HCA) for multivariate association between bioactive compounds and
functional properties in foods: A critical perspective
Daniel Granatoa,∗, Jânio S. Santosa, Graziela B. Eschera, Bruno L. Ferreirab, Rubén M. Maggioc,∗∗
a
Graduation Program in Food Science and Technology, State University of Ponta Grossa, Av. Carlos Cavalcanti, 4748, 84030-900 Ponta Grossa, Brazil
b
Graduation Program in Food Science, Federal University of Santa Catarina, Rodovia Admar Gonzaga, 1346, Itacorubi, Florianópolis, Brazil
c
Área Análisis de Medicamentos, Facultad de Ciencias Bioquímicas y Farmacéuticas, Universidad Nacional de Rosario, Instituto de Química de Rosario (IQUIR-
CONICET), Suipacha 531, S2002LRK Rosario, Argentina

A R T I C L E I N F O A B S T R A C T

Keywords: Background: The development of statistical software has enabled food scientists to perform a wide variety of
Chemometrics mathematical/statistical analyses and solve problems. Therefore, not only sophisticated analytical methods but
Principal component analysis also the application of multivariate statistical methods have increased considerably. Herein, principal compo-
Cluster analysis nent analysis (PCA) and hierarchical cluster analysis (HCA) are the most widely used tools to explore similarities
Correlation analysis
and hidden patterns among samples where relationship on data and grouping are until unclear. Usually, larger
Bioactive compounds
chemical data sets, bioactive compounds and functional properties are the target of these methodologies.
Functional properties
Scope and approach: In this article, we criticize these methods when correlation analysis should be calculated
and results analyzed.
Key findings and conclusions: The use of PCA and HCA in food chemistry studies has increased because the results
are easy to interpret and discuss. However, their indiscriminate use to assess the association between bioactive
compounds and in vitro functional properties is criticized as they provide a qualitative view of the data. When
appropriate, one should bear in mind that the correlation between the content of chemical compounds and
bioactivity could be duly discussed using correlation coefficients.

1. Introduction 2014; Munck, Nørgaard, Engelsen, Bro, & Andersson, 1998; Qannari,
2017). Conversely, the application of chemometrics for assessing the
As well stressed by Ropodi, Panagou, and Nychas (2016), in the 21st adulteration and geographical origin of foods based on chemical mar-
century, governmental, industrial, and academic problems need to be kers is well established in food science (Granato, Koot, Schnitzler, & van
addressed by using sophisticated analytical tools with proper data Ruth, 2015; Granato, Margraf, Brotzakis, Capuano, & van Ruth, 2015;
collection, analysis and interpretation. In this sense, data mining and Paneque, Morales, Burgos, Ponce, & Callejón, 2017; Giannetti, Mariani,
data analysis are two interrelated approaches developed rapidly to Mannino, & Marini, 2017; Opatić et al., 2018). For example, Garrido-
address problems related to engineering and technology, as well as Delgado, Muñoz-Pérez, and Arce (2018) used ion mobility spectrometry
medicine, economics, biology, and food science (Brown, 2017). (IMS) to determine the origin of the olive oil, quality and adulteration
Chemometrics is an interfacial discipline that extracts useful in- with low-cost vegetable oils. Using different statistical tools, authors
formation from large chemical and biochemical data sets using different were able to predict the level of contaminating oil in olive oil. There-
mathematical and statistical methods (Brown, 2017; Nunes, Alvarenga, fore, there is no doubt that chemometric tools is of fundamental im-
Sant'Ana, Santos, & Granato, 2015). In applied chemistry, the use of portance to solve real life problems.
chemometrics has been spread and well recognized since 1960 Granato, Nunes, and Barba (2017) stated that the use of design of
(Brereton, 2014), but in food sciences and technology the applications experiments together with appropriate statistical data analysis is of
of chemometrics and sensometrics (multivariate methods applied to pivotal importance to assess the association between nutrition, biology,
sensory data and studies consumers) are somewhat new (Aquino et al., pharmacology, functional properties and the chemical components of

∗
Corresponding author.
∗∗
Corresponding author.
E-mail addresses: [email protected] (D. Granato), [email protected] (R.M. Maggio).

https://doi.org/10.1016/j.tifs.2017.12.006
Received 19 October 2017; Received in revised form 10 December 2017; Accepted 13 December 2017
Available online 15 December 2017
0924-2244/ © 2017 Elsevier Ltd. All rights reserved.
D. Granato et al. Trends in Food Science & Technology 72 (2018) 83–90

Abbreviations NMR nuclear magnetic resonance

OPLS-DA orthogonal partial least squared discriminant analysis
ABTS 2,2′-azino-bis(3-ethylbenzothiazoline-6-sulphonic acid) ORAC oxygen radical absorbance capacity
ANN artificial neural networks PCA principal component analysis
CAIMAN classification and influence matrix analysis PLS-DA partial least squared discriminant analysis
DD-SIMCA data-driven soft independent modeling of class analogy PRIMA pattern recognition by independent multi-category ana-
DPPH 2,2-diphenyl-1-picrylhydrazyl lysis
FRAP ferric reducing antioxidant power QDA quadratic discriminant analysis
FuRES fuzzy rule-building expert system RF random forests
HCA hierarchical cluster analysis SIMCA soft independent modeling of class analogy
HPLC high performance liquid chromatography sPLS-DA super partial least squared discriminant analysis
IMS ion mobility spectrometry SVM support vector machine
k-NN k-nearest neighbors UHPLC-MS ultra-high performance liquid chromatography – mass
LDA linear discriminant analysis spectrometry

foods and their extracts. In this sense, chemometric tools and other orthogonal partial least squared discriminant (OPLS-DA). After ana-
statistical methodologies may be of interest when different food ex- lyzing the experimental data, authors were able to identify 83 com-
tracts and bioactivities need to be evaluated (Granato, de Araújo pounds, in which 39 were metabolites, in the biological samples. In
Calado, & Jarvis, 2014). addition, the metabolic pathway (glucoronidation) by which these
In real life applications, chemometrics may be employed in food metabolites formed after oral administration of the decoction was
science and technology studies either to assess similarities/differences identified by using OPLS-DA. This research is an example on how
between multiple objects (samples) or to project the objects in a two/ chemometric tools are important aids in not only in the food chemistry
three-dimensional factor-plane based on various characteristics. field but also in the experimental nutrition studies.
Therefore, clusterings can be observed and the reasons for the grouping According to Brereton (2015), chemometrics users tend to ‘follow
can be pinpointed (Erasmus, Muller, Butler, & Hoffman, 2018; Jandrić the crowd’ and use indiscriminately the available software without
& Cannavan, 2017; Lund, Brown, & Shipley, 2017). Additionally, knowing the principles and fundamentals of each method applied in
multivariate techniques have been widely used to authenticate/trace their research data analysis. In food chemistry studies, Principal Com-
the geographical origin of foods, to verify the farming system employed ponents Analysis (PCA) and Hierarchical Cluster Analysis (HCA) are
by a company and check whether it complies to the information de- widely (and, sometimes, improperly) applied as “unsupervised classifi-
clared on the label, and to check for adulterations (intentional or not) of cation” methods to assess the association between bioactive compounds
foods and raw materials (Granato, Koot, &, van Ruth, 2015; Chiesa and in vitro functional properties (i.e., antioxidant and inhibition of
et al., 2016; Müller-Maatsch, Schweiggert, & Carle, 2016; Tavares et al., enzymes). Herein, a critical perspective on these display techniques
2016; Zhu, Wang, & Chen, 2017; Karabagias et al., 2017; Chung et al., (PCA and HCA) is made together with some comments on their use in
2017; Giannetti et al., 2017; Acierno, Alewijn, Zomer, & van Ruth, the field of bioactive compounds.
2018).
For example, Luo, Shi, and Feng (2017) aimed to characterize the
metabolites of Zhi-Zi-Hou-Po decoction, a traditional Chinese medicine, 2. Study of bioactive compounds and in vitro potential functional
in rat bile, urine and feces after oral administration, using untargeted properties with the use of chemometrics
liquid chromatography time of flight mass spectrometry combined with
Chemometrics may be used for both qualitative and quantitative

Fig. 1. Summary of selected multivariate statis-

tical methods applied in food research.

84
D. Granato et al. Trends in Food Science & Technology 72 (2018) 83–90

networks (ANN), show several applications in food science and tech-

nology, especially in the classification and authentication problems
(Aloglu, Harrington, Sahin, Demir, & Gunes, 2017; Mehretie, Al Riza,
Yoshito, & Kondo, 2018; Tian et al., 2017; Torkashvand, Ahmadi, &
Nikravesh, 2017).
Unsupervised methods, also named clustering or displays methods, are
used to study the data structure, look for similarities between multiple
objects, and check for outliers in the data set (Liu, Koot, Hettinga, de
Jong, & van Ruth, 2018). Mixture models, self-organizing maps, k-
means, HCA and PCA are representatives of unsupervised methods.
However, PCA and HCA are the most used in food and chemistry field,
representing both sub-classes visualization and agglomerative algo-
rithms, respectively (Wang, Zeng, Contreras, & Wang, 2017).
The goal of multivariate unsupervised methods is to evaluate whe-
ther clustering exists in a dataset without using class membership in-
formation in the calculations (Beebe, Pell, & Seasholtz, 1998, p. 348).
Natural clustering of samples/objects is the result of understanding the
measurement system used to characterize the samples and this union
between statistical analysis and analytical methods aids in elucidating
the physical reasons for the presence/absence of clustering in the data.
For further information on these methods, the reader is referred to
existing literature (Oliveri & Downey, 2012; de Oliveira, de Araújo
Calado, Ares, & Granato, 2015).
Here we show some recent applications of unsupervised multi-
variate techniques in the field of bioactivity of food components. When
it comes to studies relating bioactive compounds, almost all reports aim
to associate the level of certain chemical compounds, i.e., phenolic
compounds and carotenoids, with antioxidant activity and other func-
tionalities. Additionally, a critical perspective on the use of display
techniques (PCA and HCA) is made together with some comments on
their use in the field of bioactive compounds.

2.1. Principal component analysis

The term PCA is statistical test that belongs to a group of factor

analysis. PCA is a mathematical tool that aims to represent the variation
present in the dataset (i.e., responses used to characterize the samples)
Fig. 2. PCA of juice samples based on chemical composition and antioxidant activity: A – using a small number of factors. For visual analysis, usually two-di-
represents the number of PCs e the explained variance. B- represents the projection of mensional or three-dimensional projection of samples is constructed
samples on the factor-plane. For illustration purposes, red starts represent orange juice, having the axes (principal components, PC) as the factors. Each PC is a
green stars represent lemon juices, and violet stars represent grape juices. (For inter- linear combination of the original responses (that retain some correla-
pretation of the references to colour in this figure legend, the reader is referred to the Web tion among) and PCs are orthogonal to each other. PCs iteratively
version of this article.)
calculated hold as much variation from original data set as possible, in a
way that PC1 explains more the data variation than PC2, and PC2 ex-
analysis of experimental data (Martínez et al., 2017; Szymanska et al., plain more data variation than PC3 and so on. That is why a few PCs
2015). Determining whether a rice sample comes from European explain the variation of a large number of original responses. One
countries or elsewhere based on the NMR spectra or the presence or possible way to determine the number of PC is based on the Kaiser
absence of a chemical compound in a HPLC chromatogram are two criterion (Kaiser, 1960): eigenvalues higher than 1 are considered as
typical examples of qualitative data. On the other hand, assessing the “significant” in the PCA analysis. In addition, the use of Bartlett's test of
correlation between the content of chlorogenic acid derivatives and sphericity is of interest to check for correlation between responses. This
antioxidant activity of coffee brews represents a quantitative approach. test indicates that the responses are (un)related and therefore (un)sui-
A summary of selected multivariate statistical methods is shown in table for structure detection.
Fig. 1. Fig. 2 contains an example of PCA of fruit juices (i.e., orange, lemon
Overall, chemometrics may be divided into calibration, classification and grape) based on chemical composition and antioxidant activity: the
and exploratory methods. According to Oliveri and Simonetti (2016), responses used to generate the 2D-scatter plot are based on correlation
chemometrics may be divided into supervised and unsupervised methods. analysis of each response with the first three PCs. As first step an ex-
The first class encompasses a varied number of methods/algorithms, ploration of cumulative variance explained should be carried out and
including both qualitative and quantitative approaches. Among quali- the Kaiser criterion (eigenvalues higher than 1) may be used to define
tative methods, k-nearest neighbors (k-NN), partial least squares dis- the number of significant PC. Usually this decision is taken according to
criminant analysis (PLS-DA), super PLS-DA (sPLS-DA), fuzzy rule- pre-established level of variance (90, 95, 99, or 99.9%) or based on
building expert system (FuRES), soft independent modeling of class experimental error.
analogy (SIMCA) and linear or quadratic discriminant analysis (LDA or Using a factor loadings analysis (Table 1), PC1 retained about 50%
QDA) are the most used techniques. However, some methods, such as of data variation and differentiate the juice samples according to the
classification and influence matrix analysis (CAIMAN), pattern re- contents of caffeic acid, (−)-epicatechin, (+)-catechin, quercetin, lu-
cognition by independent multi-category analysis (PRIMA), support teolin and antioxidant activity (2,2-diphenyl-1-picrylhydrazyl – DPPH,
vector machine (SVM), random forests (RF), and artificial neural 2,2′-azino-bis (3-ethylbenzothiazoline-6-sulfonic acid – ABTS), and

85
D. Granato et al. Trends in Food Science & Technology 72 (2018) 83–90

Table 1 guava fruit pulps (Psidium guajava L.) by HPLC, including (+)-catechin,
Factor loadings for illustrating the interpretation of Fig. 2. gallic, ferulic, trans-cinnamic, chlorogenic, caffeic, p-coumaric, syringic,
vanillic, and ellagic acid, rutin, quercetin, and kaempferol. The ex-
Factor PC1 PC2 PC3 PC4
traction procedure was optimized using different concentrations of
DPPH 0.69 −0.47 0.16 −0.42 ethyl alcohol and methyl alcohol for 15–90 min using a sample to
ABTS 0.68 0.06 −0.40 0.44 solvent ratio between 1:30 and 1:100 w/v. The extracts were also
FRAP 0.63 −0.65 −0.12 −0.02
analyzed for total phenolic content, ascorbic acid, and flavonoids, to-
Gallic acid 0.50 −0.66 0.09 −0.30
Caffeic acid 0.81 −0.23 0.22 0.15 gether with the antioxidant activity toward DPPH and ABTS radicals.
5-O-caffeoylquinic acid 0.04 −0.70 −0.50 0.27 PCA was able to explain only 60% of data variability with 2 PC, but a
(+)-Epicatechin −0.75 −0.54 −0.30 −0.09 clear separation between ripe and green guava fruits was observed from
(+)-Catechin −0.90 −0.07 0.08 −0.17 the scatter plot. The main responses that separated the groups were
Quercetin −0.90 −0.19 0.03 −0.08
syringic acid, (+)-catechin, p-coumaric acid, caffeic acid, ellagic acid,
Quercetrin −0.52 −0.26 0.70 −0.36
Luteolin −0.78 −0.05 −0.65 0.09 trans-cinnamic acid and rutin for the green guava, while for ripe and
Ellagic acid 0.13 0.48 −0.46 −0.73 white guava, the better markers were gallic acid and chlorogenic acid.
Eigenvalue 5.39 3.78 0.56 0.23 As rational subsequent step, authors applied ANN (a supervised algo-
Explained variance (%) 50.35 30.56 8.05 3.18
rithm) on same data set to obtain a reliable methodology to classify
Note: bold numbers are factor loadings higher than 0.60.
their samples. ANN showed a suitable separation between not only
green and white variety but also ripe and unripe guava fruits. It should
ferric reducing antioxidant power - FRAP). Similarly, PC2 explained be stressed that as data were successfully analyzed by PCA, a linear
another 30% of variability in the original responses and separates the algorithm, LDA or PLS-DA was the logical way to try.
juices based on FRAP, gallic acid, and 5-O-caffeoylquinic acid. PC3 and However, in some cases, the differentiation between classes is not so
PC4 explain only 11% of data variance and barely does not differentiate clear (Fig. 3A) and outliers (one or more observation point(s) that is/are
the juice samples. The factor loadings from PC3 and PC4 were very low unusually distant from the other observations) can be detected in the
(except for quercetrin/luteolin and ellagic acid, respectively). Factor dataset. In this case, the researcher cannot expect a straightforward
loadings lower than 0.60 indicate that those variables that do not fit separation between classes. Almost perfect segregation was obtained
well with the factor solution should possibly be dropped from the when all samples are analyzed after outliers removal (in synthetic data)
analysis, especially if the projection of samples on a factor-plane is using only two principal components (PCs), as shown in Fig. 3B.
based on a 2-dimensional graph. As a final comment, the first two PCs Fidelis et al. (2017) evaluated multiple juices from different bota-
explain about 81% of data variance but there remains room for about nical origins (fruits and other vegetables) in relation to some classes of
19% unexplained variation. phenolics/bioactive compounds (tannins, total phenols, flavonoids,
Once the representative PCs were found, on the basis of samples ortho-diphenols, flavonols, total anthocyanins, and betalains), physi-
differentiation/grouping and variance explained, loading analysis is cochemical properties (pH, soluble solids, and acidity), and antioxidant
started in order to find the underlying relationships in the original data effects (Fe2+ chelating properties, antiradical effect (DPPH, ABTS, and
structure. In this step loading could be visualized as a regression vector FRAP)), Folin-Ciocalteu's reducing capacity, and total reducing capa-
(a vector of correlation coefficients between the original variables with city. A total of 570 data points (38 juices and 15 responses) were
each PC-score). The positive factor loadings indicate that the factor will analyzed for patterns using PCA, which explained 72% of data varia-
be higher in the positive axis of that PC. For example, for DPPH, a factor bility with 2 PC and it was possible to pinpoint the juices with higher
loading of 0.69 was obtained with PC1, which means that the samples bioactive compounds and antioxidant activity. PLS-DA was used to
located in the right-hand side (i.e., violet stars) of the graph have higher discriminate juice groups and authors were able to separate Citrus juices
mean DPPH values than the samples located in the left-hand side (i.e., from Super juices (made with berries) with correct classification rates
red stars). Similarly, the negative factor loadings indicate that the factor above 73%, while data-driven SIMCA, which is a one-class classification
will be higher in the positive axis of that PC. For example, for method, was able to discriminate the juices samples with accuracy
(−)-epicatechin a factor loading of −0.75 was obtained for PC1, higher than 86%. In this research, authors concluded that the use of DD-
meaning that the samples located in the right-hand side (i.e., violet SIMCA may be of interest when the authentication of juices based on
stars) of the graph have lower mean concentrations than the samples phenolic compounds and antioxidant activity need to be performed,
located in the left-hand side (i.e., red stars). especially in quality control programs in the juice industry.
As a complementary analysis, as an illustrative example, PCA data Kalaycıoğlu, Kaygusuz, Döker, Kolaylı, and Erim (2017) used PCA to
may be compared to correlation coefficients (Table 2). As shown, the explore only n = 10 Turkish honeybee pollens from distinct origins
antioxidant activity measured by three different assays (i.e., ABTS,
FRAP, and DPPH) is mainly correlated (p < .05) to caffeic acid, Table 2
(−)-epicatechin, (+)-catechin, quercetin, and luteolin. FRAP also Illustrative correlation coefficients to help in the interpretation of the example shown in
correlated significantly with gallic acid and 5-O-caffeoylquinic acid. In Fig. 2.

this sense, if the main objective is to check for association between Responses DPPH ABTS FRAP
bioactive compounds and functional properties, correlation analysis
should be carried out. DPPH 1
For instance, Pearson's correlation coefficients or Spearman's rank ABTS 0.899 1
FRAP 0.946 0.947 1
correlation coefficients are the choices for normally distributed data Gallic acid 0.564* 0.529* 0.608
and for data do not conform to the normal distribution, respectively (de Caffeic acid 0.895 0.911 0.935
Oliveira et al., 2015). 5-O-caffeoylquinic acid 0.523* 0.518* 0.622
As a final comment on this topic, there is no scientific need to (+)-Epicatechin 0.875 0.812 0.804
(+)-Catechin 0.926 0.874 0.935
perform PCA or HCA for data sets that have a similar conclusion as the
Quercetin 0.873 0.924 0.901
one shown in the above-mentioned example. However, if the number of Quercetrin 0.425* 0.378* 0.333*
responses and samples is quite large and data are quite complex (i.e., Luteolin 0.788 0.829 0.845
NMR spectra), PCA is highly indicated. Ellagic acid 0.238* 0.356* 0.458*
dos Santos et al. (2017) quantified 13 phenolic compounds in 96
Note: * denotes p > .05 while the other correlation coefficients present p < .05.

86
D. Granato et al. Trends in Food Science & Technology 72 (2018) 83–90

composition and functional properties of aqueous extracts of fermented

rooibos (Aspalathus linearis). For this purpose, a 22 factorial design with
three central points was used to manufacture beverages in which some
phenolic acids and flavonoids were quantified using LC-MS/MS, the
antioxidant activity (ABTS, FRAP, and total reducing capacity), and the
inhibition of α-amylase and α-glucosidase were determined. As a large
amount of data were generated (210 data points), authors performed a
PCA to reduce dimensionality of the data. Authors verified that rooibos
extracted at 85 °C, regardless of the extraction time, presented the
highest levels of phenolic compounds, in vitro antioxidant activity, and
highest inhibition of the digestive enzymes. Although correlation
coefficients were calculated to know which compounds exerted the in
vitro antioxidant effect, PCA was effective in showing the best techno-
logical conditions to produce the infusions with higher bioactive com-
pounds.
Farag, Ezzat, Salama, and Tadros (2016) studied the anti-acet-
ylcholinesterase activity and bioactive compounds of four sweet basil
species (Ocimum basilicum, Ocimum africanum, Ocimum americanum and
Ocimum minimum) by ultra-performance liquid chromatography quad-
rupole time of flight mass spectrometry (UPLC/qTOF/MS), PCA was
used as exploratory tool and OPLS-DA was used for its further analysis.
Twenty one hydroxycinnamic acids, 4 benzoic acid conjugates, 14 fla-
vonoid conjugates, 2 alcohols, 5 acyl sugars, 4 triterpenes and 12 fatty
acids were identified in the extracts. Using these responses, authors
applied PCA and HCA to pinpoint the sweet basis species with higher
anti-acetylcholinesterase activity: O. americanum, O. africanum, and O.
basilicum. Additionally, OPLS-DA was used to distinguish between O.
basilicum (official drug) from O. americanum, with more than 96% of
data variability explained by the classification model.

2.2. Hierarchical cluster analysis

HCA is a clustering method which explore the organization of

samples in groups and among groups depicting a hierarchy (Lee & Yang,
2009). The result of HCA is usually presented in a dendrogram, a plot
which shows the organization of samples and its relationships in tree
Fig. 3. Principal components analysis, PCA, to project different samples (i.e., fruits from
different varieties) based on some selected responses: outlier detection with no separation
form. There are two main approaches to resolve the grouping problem
between varieties (A), no outliers with a clear separation between fruit varieties (B). in HCA, agglomerative or divisive (Fig. 4).
In the first one, each sample is initially considered a cluster, and
subsequently pairs of clusters are merged. In divisive approach algo-
based on organic acids, carbohydrates, 14 minerals, total phenolic
rithm start with one cluster including al samples, recursive splits are
content, and antioxidant activity measured by the DPPH assay. Not
performed. Clustering is achieved by use of an appropriate metric of
surprisingly, the first three principal components explained 71% of data
samples distance (usually, Euclidean, Mahalanobis or Manhattan dis-
variability and authors claim they “classified” the pollen samples ac-
tance) and linkage criterion among groups. Complete, single and
cording to the geographical origin of the samples (less than five samples
average and Ward's linkage are the more common variants of linkage
per class, in which, n = 2 chestnuts, n = 1 oak, n = 1 Abana, n = 1
criterions. Ward's method, based in optimal value of a target function, is
Bayburt, n = 1 Balikesir, n = 1 buckwheat, and n = 3 Anzer). How-
a possible choice (Granato, Karnopp, & van Ruth, 2015).
ever, results are untrustworthy when such low number of samples are
HCA has also been extensively used to evaluate the multivariate
available, so conclusions based on the PC plots should be pondered.
association between bioactive compounds and bioactivity of foods,
According to de Oliveira et al. (2015), for PCA, at least five responses
beverages and their extracts. For instance, Viapiana, Struck-Lewicka,
and five objects (samples) need to be part of the dataset.
Konieczynski, Wesolowski, and Kaliszan (2016) used HCA aiming to
Santos et al. (2016) used PCA to reveal the effects of time
associate the relationship between phenolic composition measured by
(5–10 min) and extraction temperature (65–85 °C) on phenolic
HPLC with the in vitro antioxidant activity (FRAP and DPPH assays) of

Fig. 4. HCA dendrogram for agglomerative algorithm (A) and di-

visive algorithm grouping flow (B).

87
D. Granato et al. Trends in Food Science & Technology 72 (2018) 83–90

concentrations of bioactive compounds and antioxidant activity indices.

Sánchez-Salcedo et al. (2016) used HCA as a tool to propose a
polyphenolic fingerprint of white (Morus alba L., n = 4) and black
(Morus nigra L., n = 4) mulberry leaves clones. UHPLC-MS identified 31
phenolic compounds, mostly important 20 flavonoids, in more than 120
spectrums analyzed, a very high number of variables for such low
number of samples. Ward's method based on Euclidean distance gen-
erated three major groups, first characterized by 4 clones of both spe-
cies presenting high amount of caffeic acid-hexoside, caffeoylquinic
acid and kaempferol-malonyl-rutinoside and low content of O-hexoside
flavonols. Only one clone of Morus nigra formed the second group, re-
presenting caffeic acid and cryptochlorogenic acid as characteristics.
The last group was formed of 3 clones of two mulberry species, pre-
senting high flavonols containing O-hexoside and a low content of
caffeic acid.
To study the geographical influence on phenolic content and anti-
oxidant activity in Napirira bean (Phaseolus vulgaris L.), Fan and Beta
(2017) applied an unsupervised pattern recognition method (HCA)
based on the Euclidean distance and Ward's method. Total phenolic
content, antioxidant activity, and phenolic compounds (protocatechuic
acid, p-hydroxybenzoic acid, catechin, p-coumaric acid, ferulic acid and
sinapic acid) were analyzed in eighteen bean samples from four loca-
tions in Malawi. HCA was able to differentiate 3 major groups: group 1
clustered samples with high contents of phenolic compounds and an-
tioxidant activity which were from the high-altitude region; group 2
clustered samples that presented low contents of phenolic compounds
and antioxidant activity which were from a lower-altitude region; and
group 3 contained samples with intermediate values of phenolic com-
pounds and antioxidant activity and included samples from both in-
termediate regions of Malawi. As a conclusion, HCA was a useful tool to
associate the phenolic compounds/antioxidant activity with the culti-
vation region.
A good example where algorithm configuration could be decisive to
obtain a valid conclusion is illustrated by Kaškonienė, Ruočkuvienė,
Kaškonas, Akuneca, and Maruška (2015). Authors analyzed the total
phenolic and flavonoids contents, antioxidant activity and individual
phenolic compounds (gallic acid, caffeic acid, ferulic acid, 2-hydro-
xycinnamic acid, rutin naringenin and quercetin) in 14 pollen samples
collected in the Baltic region (Latvia and Lithuania) and two others
from Spain and China. Data were treated by HCA using both Spearman's
distance and Euclidean distance. Samples were clustered in two groups
according to the antioxidant activity. Similarly, Euclidean distance
clustered the samples into three major groups according to the geo-
graphical regions with clear differences in the phenolic composition. As
a conclusion the choice of distance function is not a trivial matter and
should be tested when HCA is applied. The use of the only one clus-
tering technique (i.e., k-means or tree-clustering), amalgamation rule
(i.e., single linkage, complete linkage, or Ward's method), and distance
measure (i.e, Euclidean, Manhattan, 1- Pearson r) is not recommended.
Nayik and Nanda (2016) analyzed the minerals, phenolic compo-
sition and antioxidant activity of n = 37 unifloral honeys from Kashmir,
India. PCA and HCA were used to assess the effects of the botanical
origins of those samples based on the quality parameters and verified
Fig. 5. Example of PCA (A) and HCA (B, C) applied to a data set composed of n = 20 fruit
samples (A and B) according to the concentrations of total phenolics, carotenoids, anti- that PCA was able to group the samples according to the origin (apple,
oxidant activity measured by the ORAC assay, and inhibition of lipase and α-amylase. cherry, saffron and wild bush). The authors claimed that “minerals
presented the highest discriminating power” in PCA while samples were
“classified” using HCA. The terms “discriminating power” and “classi-
19 chamomile commercial samples (Matricaria chamomilla L.). Overall,
fication” are related to supervised chemometric tools, such as LDA/
caffeic, ferulic, and syringic acids were the most effective phenolics in
QDA or SIMCA, among other techniques (Kasprzyk, Depciuch, Grabek-
exerting antioxidant activity in the herbal extracts (MeOH:H2O, 80:20
Lejko, & Parlinska-Wojtan, 2018; Mapelli-Brahm, Hernanz-Vila, Stinco,
v/v). Linear correlation coefficients were also calculated to display a
Heredia, & Meléndez-Martínez, 2018; Popek, Halagarda, & Kursa,
mathematical proof of such findings (r > 0.70, p < .05). HCA and
2017). Therefore, such terms should be avoided when PCA or HCA are
PCA were used with the aim to tentatively “classify” the commercial
employed.
samples based on the HPLC fingerprint but no differentiation between
samples was achieved. This study shows that PCA/HCA methods not
always provide sufficient means to group samples according to the

88
D. Granato et al. Trends in Food Science & Technology 72 (2018) 83–90

2.3. Overall comments on PCA and HCA especially of a large data set is analyzed. However, the indiscriminate
use of multivariate exploratory statistical techniques (PCA and HCA) to
Both PCA and HCA are usually used concomitantly in studies cov- assess the association between bioactive compounds and in vitro func-
ering bioactive compounds and functional properties. To illustrate what tional properties is criticized as the results will be, in most cases, a sine
is widely seen in published articles, consider the following: n = 20 qua non observation. When appropriate, the researcher should bear in
samples coming from two fruits (A and B) are analyzed for the con- mind that the correlation between the content of chemical compounds
centrations of total phenolics, carotenoids, antioxidant activity mea- and bioactivity could be duly discussed using simple correlation coef-
sured by the oxygen radical absorbance capacity (ORAC) assay, and ficients.
inhibition of amylase and lipase. Results were analyzed using PCA and
the 2D projection is given in Fig. 5A: it is possible to see a defined Acknowledgements
cluster containing fruit “B” and another group containing most “A”
fruits. However, there are n = 3 “A” samples that are far from the main Daniel Granato acknowledges CNPq for a productivity grant (pro-
“A” group. One could say they are outliers simply by looking at the cess 303188/2016-2). J. S. Santos and G. B. Escher thank CAPES/
projection, but this cannot be done as PCA does not “classify” objects. In Fundação Araucária for their Ph.D scholarships.
Fig. 5B, HCA was applied using the Ward's method as the amalgamation
rule and Euclidean distances were calculated between fruits. Using a References
linkage distance of 15, only two groups are formed, one containing the
“A” fruits and the other containing all “B” fruits. Similarly, if a distance Acierno, V., Alewijn, M., Zomer, P., & van Ruth, S. M. (2018). Making cocoa origin
of 5 is considered, there are 1 group containing the “B” fruits and two traceable: Fingerprints of chocolates using flow infusion - electro spray ionization -
mass spectrometry. Food Control, 85, 245–252.
other groups containing the “A” fruits, which is similar to the PCA re- Aloglu, A. K., Harrington, P. B., Sahin, S., Demir, C., & Gunes, M. E. (2017). Chemical
sults. However, if a linkage distance of 1.5 is considered, a total of 6 profiling of floral and chestnut honey using high-performance liquid chromatography-
small groups can be visualized. Using this simple example, it is possible ultraviolet detection. Journal of Food Composition and Analysis, 62, 205–210.
Andrić, F., Bajusz, D., Rácz, A., Šegan, S., & Héberger, K. (2016). Multivariate assessment
to conclude that HCA is an arbitrary method and should be used for of lipophilicity scales—computational and reversed phase thin-layer chromato-
exploratory purposes only. Additionally, neither PCA nor HCA creates a graphic indices. Journal of Pharmaceutical and Biomedical Analysis, 127, 81–93.
“mathematical model” for classification and authentication purposes. Aquino, L. F. M. C., Silva, A. C. O., Freitas, M. Q., Felicio, T. L., Cruz, A. G., & Conte-
Junior, C. A. (2014). Identifying cheese whey an adulterant in milk: Limited con-
Rather, they only project or display the objects under investigation
tribution of a sensometric approach. Food Research International, 62, 233–237.
based on selected responses and grouping of samples may be identified Beebe, K. R., Pell, R. J., & Seasholtz, M. B. (1998). Chemometrics: A practical guide (1st ed.).
by the user. Moreover, neither PCA nor HCA provides a statistical sig- New York: Wiley & Sons.
Brereton, R. G. (2014). A short history of chemometrics: A personal view. Journal of
nificance of such similarities (Andrić, Bajusz, Rácz, Šegan, & Héberger,
Chemometrics, 28(10), 749–760.
2016). Brereton, R. G. (2015). Pattern recognition in chemometrics. Chemometrics and Intelligent
If the aim is to find an association between bioactive compounds Laboratory Systems, 149, 90–96.
and functional properties using HCA, the method may be applied Brown, S. D. (2017). The chemometrics revolution re-examined. Journal of Chemometrics,
31(1), e2856.
(Fig. 5C). It is an easy and straightforward result: total phenolics and Chiesa, L., Panseri, S., Bonacci, S., Procopio, A., Zecconi, A., Arioli, F., et al. (2016).
carotenoids are associated with ORAC and inhibition of α-amylase. Authentication of Italian PDO lard using NIR spectroscopy, volatile profile and fatty
Conversely, the inhibition of lipase does not seem to be associated with acid composition combined with chemometrics. Food Chemistry, 212, 296–304.
Chung, I. M., Kim, J. K., Yang, J. H., Lee, J. H., Park, S. K., Son, N. Y., et al. (2017). Effects
any of the responses. Although HCA shows the existence of association of soil type and organic fertilizers on fatty acids and vitamin E in Korean ginseng
between responses, however it does not provide a measure of the as- (Panax ginseng Meyer). Food Research International, 102, 265–273.
sociation (qualitative approach). One alternative to overcome this Erasmus, S. W., Muller, M., Butler, M., & Hoffman, L. C. (2018). The truth is in the iso-
topes: Authenticating regionally unique South African lamb. Food Chemistry, 239,
limitation is to calculate the correlation coefficient and provide a 926–934.
quantitative measure of the correlation between responses. As a matter Fan, G., & Beta, T. (2017). Discrimination of geographical origin of Napirira bean
of fact, the inhibition of lipase is not correlated to the concentrations of (Phaseolus vulgaris L.) based on phenolic profiles and antioxidant activity. Journal of
Food Composition and Analysis, 62, 217–222.
total phenolics (r = −0.022, p = .927), carotenoids (r = 0.213, Farag, M. A., Ezzat, S. M., Salama, M. M., & Tadros, M. G. (2016). Anti-acet-
p = .367), and ORAC (r = 0.304, p = .193). In this case, the use of HCA ylcholinesterase potential and metabolome classification of 4 Ocimum species as
is near meaningless as correlation coefficients are robust enough to determined via UPLC/qTOF/MS and chemometric tools. Journal of Pharmaceutical
and Biomedical Analysis, 125, 292–302.
draw the association between the chemical composition and the func-
Fidelis, M., Santos, J. S., Coelho, A. L. K., Rodionova, O. Y., Pomerantsev, A., & Granato,
tional properties of the fruits. D. (2017). Authentication of juices from antioxidant and chemical perspectives: A
Although PCA and HCA are very useful to study the data structure feasibility quality control study using chemometrics. Food Control, 73, 796–805.
and find similarities among samples, in most cases, linear correlation Garrido-Delgado, R., Muñoz-Pérez, M. A., & Arce, L. (2018). Detection of adulteration in
extra virgin olive oils by using UV-IMS and chemometric analysis. Food Control, 85,
coefficients would render very similar interpretations of the results. 292–299.
Indeed, it is widely known and recognized that higher levels of phenolic Giannetti, V., Mariani, M. B., Mannino, P., & Marini, F. (2017). Volatile fraction analysis
compounds will render a higher antioxidant activity measured by by HS-SPME/GC-MS and chemometric modeling for traceability of apples cultivated
in the Northeast Italy. Food Control, 78, 215–221.
chemical reactions in vitro (Guo, Sun, Yu, & Qi, 2017; Lv, Zhang, Shi, & Granato, D., de Araújo Calado, V. M., & Jarvis, B. (2014). Observations on the use of
Lin, 2017). Another main disadvantage of using PCA/HCA in those statistical methods in food science and technology. Food Research International, 55,
studies is the real applicability of the observations: it seems that most 137–149.
Granato, D., Karnopp, A. R., & van Ruth, S. M. (2015a). Characterization and comparison
researchers only use PCA and HCA to increment their data analysis of phenolic composition, antioxidant capacity and instrumental taste profile of juices
rather than to explain the mechanisms of action and have a strong and from different botanical origins. Journal of the Science of Food and Agriculture, 95(10),
in-depth discussion based on a solid hypothesis. In fact, in the field of 1997–2006.
Granato, D., Koot, A., Schnitzler, E., & van Ruth, S. M. (2015b). Authentication of geo-
bioactive compounds, when in vitro assays are used, it is somewhat graphical origin and crop system of grape juices by phenolic compounds and anti-
obvious that almost all carotenoids and phenolic compounds will exert oxidant activity using chemometrics. Journal of Food Science, 80(3), C584–C593.
antioxidant activity. In this case, correlation coefficients should be Granato, D., Koot, A., & van Ruth, S. M. (2015c). Geographical provenancing of purple
grape juices from different farming systems by proton transfer reaction mass spec-
calculated and results analyzed.
trometry using supervised statistical techniques. Journal of the Science of Food and
Agriculture, 95(13), 2668–2677.
3. Final comments and recommendations Granato, D., Margraf, T., Brotzakis, I., Capuano, E., & Ruth, S. M. (2015d).
Characterization of conventional, biodynamic, and organic purple grape juices by
chemical markers, antioxidant capacity, and instrumental taste profile. Journal of
The use of PCA and HCA in food chemistry studies has increased in Food Science, 80(1), C55–C65.
the past years because the results are easy to interpret and discuss,

89
D. Granato et al. Trends in Food Science & Technology 72 (2018) 83–90

Granato, D., Nunes, D. S., & Barba, F. J. (2017). An integrated strategy between food of statistical software in food science and technology: Advantages, limitations and
chemistry, biology, nutrition, pharmacology, and statistics in the development of misuses. Food Research International, 75, 270–280.
functional foods: A proposal. Trends in Food Science & Technology, 62, 13–22. de Oliveira, C. C., de Araújo Calado, V. M., Ares, G., & Granato, D. (2015). Statistical
Guo, Y., Sun, L., Yu, B., & Qi, J. (2017). An integrated antioxidant activity fingerprint for approaches to assess the association between phenolic compounds and the in vitro
commercial teas based on their capacities to scavenge reactive oxygen species. Food antioxidant activity of Camellia sinensis and Ilex paraguariensis teas. Critical Reviews in
Chemistry, 237, 645–653. Food Science and Nutrition, 55, 1456–1473.
Jandrić, Z., & Cannavan, A. (2017). An investigative study on differentiation of citrus Oliveri, P., & Downey, G. (2012). Multivariate class modeling for the verification of food
fruit/fruit juices by UPLC-QToF MS and chemometrics. Food Control, 72, 173–180. authenticity claims. Trends in Analytical Chemistry, 35, 74–86.
Kaiser, H. F. (1960). The application of electronic computers to factor analysis. Oliveri, P., & Simonetti, R. (2016). Chemometrics for food authenticity applications. In G.
Educational and Psychological Measurement, 20, 141–151. Downey (Ed.). Advances in food authenticity testing (pp. 701–728). (1st ed.).
Kalaycıoğlu, Z., Kaygusuz, H., Döker, S., Kolaylı, S., & Erim, F. B. (2017). Characterization Amsterdam: Elsevier.
of Turkish honeybee pollens by principal component analysis based on their in- Opatić, A. M., Nečemer, M., Lojen, S., Masten, J., Zlatić, E., Šircelj, H., et al. (2018).
dividual organic acids, sugars, minerals, and antioxidant activities. LWT – Food Determination of geographical origin of commercial tomato through analysis of
Science and Technology, 84, 402–408. stable isotopes, elemental composition and chemical markers. Food Control. http://
Karabagias, I. K., Louppis, A. P., Karabournioti, S., Kontakos, S., Papastephanou, C., & dx.doi.org/10.1016/j.foodcont.2017.11.013.
Kontominas, M. G. (2017). Characterization and geographical discrimination of Paneque, P., Morales, M. L., Burgos, P., Ponce, L., & Callejón, R. M. (2017). Elemental
commercial Citrus spp. honeys produced in different Mediterranean countries based characterisation of Andalusian wine vinegars with protected designation of origin by
on minerals, volatile compounds and physicochemical parameters, using chemo- ICP-OES and chemometric approach. Food Control, 75, 203–210.
metrics. Food Chemistry, 217, 445–455. Popek, S., Halagarda, M., & Kursa, K. (2017). A new model to identify botanical origin of
Kaškonienė, V., Ruočkuvienė, G., Kaškonas, P., Akuneca, I., & Maruška, A. (2015). Polish honeys based on the physicochemical parameters and chemometric analysis.
Chemometric analysis of bee pollen based on volatile and phenolic compound com- LWT - Food Science and Technology, 77, 482–487.
positions and antioxidant properties. Food Analytical Methods, 8(5), 1150–1163. Qannari, E. M. (2017). Sensometrics approaches in sensory and consumer research.
Kasprzyk, I., Depciuch, J., Grabek-Lejko, D., & Parlinska-Wojtan, M. (2018). FTIR-ATR Current Opinion in Food Science, 15, 8–13.
spectroscopy of pollen and honey as a tool for unifloral honey authentication. The Ropodi, A. I., Panagou, E. Z., & Nychas, G. J. E. (2016). Data mining derived from food
case study of rape honey. Food Control, 84, 33–40. analyses using non-invasive/nondestructive analytical techniques; determination of
Lee, I., & Yang, J. (2009). Common clustering algorithms. In S. D. Brown, R. Tauler, & B. food authenticity, quality & safety in tandem with computer science disciplines.
Walczak (Eds.). Comprehensive chemometrics (pp. 577–618). Oxford, England: Trends in Food Science & Technology, 50, 11–25.
Elsevier. Sánchez-Salcedo, E. M., Tassotti, M., Del Rio, D., Hernández, F., Martínez, J. J., & Mena,
Liu, N., Koot, A., Hettinga, K., de Jong, J., & van Ruth, S. M. (2018). Portraying and P. (2016). (Poly)phenolic fingerprint and chemometric analysis of white (Morus alba
tracing the impact of different production systems on the volatile organic compound L.) and black (Morus nigra L.) mulberry leaves by using a non-targeted UHPLC–MS
composition of milk by PTR-(Quad)MS and PTR-(ToF)MS. Food Chemistry, 239, approach. Food Chemistry, 212, 250–255.
201–207. Santos, J. S., Deolindo, C. T. P., Fujita, A., Genovese, M. I., Daguer, H., Valese, A., et al.
Lund, J. A., Brown, P. N., & Shipley, P. R. (2017). Differentiation of Crataegus spp. guided (2016). Effects of time and extraction temperature on phenolic composition and
by nuclear magnetic resonance spectrometry with chemometric analyses. functional properties of red rooibos (Aspalathus linearis). Food Research International,
Phytochemistry, 141, 11–19. 89, 476–487.
Luo, K., Shi, Q., & Feng, F. (2017). Characterization of global metabolic profile of Zhi-Zi- dos Santos, W. N. L., Sauthier, M. C. S., dos Santos, A. M. P., Santana, D. A., Azevedo, R. S.
Hou-Po decoction in rat bile, urine and feces after oral administration based on a A., & Caldas, J. C. (2017). Simultaneous determination of 13 phenolic bioactive
strategy combining LC–MS and chemometrics. Journal of Chromatography B, 1040, compounds in guava (Psidium guajava L.) by HPLC-PAD with evaluation using PCA
260–272. and Neural Network Analysis (NNA). Microchemical Journal, 133, 583–592.
Lv, H., Zhang, Y., Shi, J., & Lin, Z. (2017). Phytochemical profiles and antioxidant ac- Szymanska, E., Gerretzen, J., Engel, J., Geurts, B., Blanchet, L., & Buydens, L. M. C.
tivities of Chinese dark teas obtained by different processing technologies. Food (2015). Chemometrics and qualitative analysis have a vibrant relationship. Trends in
Research International, 100, 486–493. Analytical Chemistry, 69, 34–51.
Mapelli-Brahm, P., Hernanz-Vila, D., Stinco, C. M., Heredia, F. J., & Meléndez-Martínez, Tavares, K. M., Lima, A. R., Nunes, C. A., Silva, V. A., Mendes, E., Casal, S., et al. (2016).
A. J. (2018). Isoprenoids composition and colour to differentiate virgin olive oils Free tocopherols as chemical markers for Arabica coffee adulteration with maize and
from a specific mill. LWT - Food Science and Technology, 89, 18–23. coffee by-products. Food Control, 70, 318–324.
Martínez, E. B., Ramos, E. F., Hernández, N. P., Vallejo, L. G. Z., Ruano, N. V., Ponce, M. Tian, Y., Yan, C., Zhang, T., Tang, H., Li, H., Yu, J., et al. (2017). Classification of wines
V., et al. (2017). 1H NMR-based metabolomic fingerprinting to determine metabolite according to their production regions with the contained trace elements using laser-
levels in serrano peppers (Capsicum annum L.) grown in two different regions. Food induced breakdown spectroscopy. Spectrochimica Acta Part B: Atomic Spectroscopy,
Research International, 102, 163–170. 135, 91–101.
Mehretie, S., Al Riza, D. F., Yoshito, S., & Kondo, N. (2018). Classification of raw Torkashvand, A. M., Ahmadi, A., & Nikravesh, N. L. (2017). Prediction of kiwifruit
Ethiopian honeys using front face fluorescence spectra with multivariate analysis. firmness using fruit mineral nutrient concentration by artificial neural network
Food Control, 84, 83–88. (ANN) and multiple linear regressions (MLR). Journal of Integrative Agriculture, 16(7),
Müller-Maatsch, J., Schweiggert, R. M., & Carle, R. (2016). Adulteration of anthocyanin- 1634–1644.
and betalain-based coloring foodstuffs with the textile dye ‘Reactive Red 195’ and its Viapiana, A., Struck-Lewicka, W., Konieczynski, P., Wesolowski, M., & Kaliszan, R.
detection by spectrophotometric, chromatic and HPLC-PDA-MS/MS analyses. Food (2016). An approach based on HPLC-fingerprint and chemometrics to quality con-
Control, 70, 333–338. sistency evaluation of Matricaria chamomilla L. commercial samples. Frontiers in Plant
Munck, L., Nørgaard, L., Engelsen, S. B., Bro, R., & Andersson, C. A. (1998). Science, 7, 1–11.
Chemometrics in food science – a demonstration of the feasibility of a highly ex- Wang, X., Zeng, Q., Contreras, M. M., & Wang, L. (2017). Profiling and quantification of
ploratory, inductive evaluation strategy of fundamental scientific significance. phenolic compounds in Camellia seed oils: Natural tea polyphenols in vegetable oil.
Chemometrics and Intelligent Laboratory Systems, 44, 31–60. Food Research International, 102, 184–194.
Nayik, G. A., & Nanda, V. (2016). A chemometric approach to evaluate the phenolic Zhu, W., Wang, X., & Chen, L. (2017). Rapid detection of peanut oil adulteration using
compounds, antioxidant activity and mineral content of different unifloral honey low-field nuclear magnetic resonance and chemometrics. Food Chemistry, 216,
types from Kashmir, India. LWT - Food Science and Technology, 74, 504–513. 268–274.
Nunes, C. A., Alvarenga, V. O., Sant'Ana, A. S., Santos, J. S., & Granato, D. (2015). The use

90