Talanta 187 (2018) 348–356

Contents lists available at ScienceDirect

Talanta
journal homepage: www.elsevier.com/locate/talanta

Rum classification using fingerprinting analysis of volatile fraction by T

headspace solid phase microextraction coupled to gas chromatography-mass
spectrometry
José Raúl Belmonte-Sáncheza, Simona Gherghela,b,c, Javier Arrebola-Liébanasa,
⁎
Roberto Romero Gonzáleza, José Luis Martínez Vidala, Ivan Parkinc, Antonia Garrido Frenicha,
a
Research Group “Analytical Chemistry of Contaminants”, Department of Chemistry and Physics, Research Centre for Agricultural and Food Biotechnology (BITAL),
University of Almería, Agrifood Campus of International Excellence, ceiA3, E-04120 Almería, Spain
b
UCL Department of Security and Crime Science, 35 Tavistock Square, London WC1H 9EZ, United Kingdom
c
UCL Department of Chemistry, 20 Gordon Street, London WC1H 0AJ, United Kingdom

A R T I C LE I N FO A B S T R A C T

Keywords: In this study, targeted and untargeted analyses based on headspace solid phase microextraction coupled to gas
Rum chromatography-mass spectrometry (HS-SPME-GC-MS) method were developed for classifying 33 different
Classification commercial rums. Targeted analysis showed correlation of ethyl acetate and ethyl esters of carboxylic acids with
Volatile organic compounds aging when rums of the same brand were studied, but presented certain limitations when the comparison was
Multivariate analysis
carried out between different brands. To overcome these limitations, untargeted strategies based on un-
SPME-GC-MS
supervised treatments, such as hierarchical cluster analysis (HCA) and principal component analysis (PCA), as
well as supervised methods, such as linear discriminant analysis (LDA) were applied. HCA allowed distin-
guishing main groups (with and without additives), while the PCA method indicated 40 ions corresponding to 13
discriminant compounds as relevant chemical descriptors for the correct rum classification (PCA variance of
88%). The compounds were confirmed based on the combination of retention indexes and low and high-re-
solution mass spectrometry (HRMS). Using the obtained results, LDA was carried out for the analytical dis-
crimination of the remaining rums based on manufacturing country, raw material type, distillation method,
wood barrel type and aging period and 94%, 91%, 92%, 95% and 94% of rums, respectively, were correctly
classified. The proposed methodology has led to a robust analytical strategy for the classification of rums as a
function of different parameters depending on the rum production process.

1. Introduction mash is distilled using heat in copper pot stills or in stainless steel
columns to obtain a high content of ethanol, which inevitably leads to
Rum is a fairly aromatic spirit, obtained exclusively from sugar cane the loss of some aroma compounds [6,7]. Additionally, different dis-
juice or molasses, and then subjected to the processes of alcoholic fer- tillation methods can be applied, such as continuous and batch dis-
mentation, distillation and aging. This spirit represents a widely pop- tillation (e.g. Jamaican “heavy rums” typically made by batch distilla-
ular alcoholic beverage with a high world consumption rate (more than tion) [8]. The resulting distillate is diluted with pure demineralized
1 billion of litres per year) and an expected increase of 1.9% in volume water to obtain an alcohol percentage of around 35–40%, which is then
terms over 2016–2021. [1,2]. aged in oak barrels previously used for whiskey or brandy production
The complex elaboration of this type of alcoholic beverage makes it [9,10]. The aging step gives rum its characteristic flavor as a large
an attractive object of study. Differences in the production process are number of new compounds emerge. Ethyl esters are generated as a
known to lead to wide variability in its composition, although this result of the high percentage of ethanol, while a number of different
variation has not been fully understood yet [3,4]. The production compounds such as whiskey lactone, vanillin and 2-methoxyphenol can
process begins with the fermentation of the chosen raw material, which form because of the interaction with the wood barrels [3]. Additionally,
leads to the formation of a number of volatile compounds, such as al- as rum matures, it generally gains golden hues as a result of the tannins
cohols, ethyl esters and aldehydes, among others [5]. The resulting from the barrel staves [11]. After an aging period, typically of at least 1

⁎
Corresponding author.
E-mail address: agarrido@ual.es (A. Garrido Frenich).

https://doi.org/10.1016/j.talanta.2018.05.025
Received 22 March 2018; Received in revised form 4 May 2018; Accepted 7 May 2018
Available online 08 May 2018
0039-9140/ © 2018 Elsevier B.V. All rights reserved.
J.R. Belmonte-Sánchez et al. Talanta 187 (2018) 348–356

year, the containers are opened for an optional blending step, where 2.2. Samples
rums of different ages are mixed to obtain specific organoleptic char-
acteristics. Lastly, the colour and flavor of rums can be further modified For this study, a total of 33 commercial rums were purchased from
by adding colorants and flavorings. Therefore, rums can be classified different local liquor stores (Almería, Spain). The rums were manu-
depending on the raw material, fermentation process, distillation pro- factured in 10 different countries: Cuba (5 samples), Dominican
cess, aging period, type of barrel used, blending technique, alcohol Republic (8 samples), Grenade (1 sample), Guatemala (3 samples),
strength and possible addition of additives. Jamaica (2 samples), Nicaragua (3 samples), Republic of Mauritius (2
Because of a lack of clear legislation around labelling, terms loosely samples), Spain (6 samples), Trinidad & Tobago (1 sample) and
related to aging periods, such as “Añejo”, “Dorado”, “Premium”, “Super Venezuela (2 samples). All samples were stored in a refrigerator (4 °C)
Premium” or “Reserve” are often used by rum manufacturers without prior to analysis, in their original glass bottles. Information about the
an actual quantitative/qualitative justification. Moreover, the age rum production from the official website of rum manufacturers as well
statement on labels is often not representative of the actual age, as as from the label, and assigned codes for each rum are summarized in
blending of rums of different ages is carried out. According to legisla- Table 1. It should be pointed out that information about aging, raw
tion from both the European Union and the United States, the age material and distillation process was not provided by all manufacturers.
statement on the label needs to refer to the youngest rum in the bottle When the information was not available, this was recorded as NA.
[12,13]. However, in other countries, such as Canada, it can refer to the
oldest rum [14]. Therefore, the development of methods that allow the 2.3. Sample preparation and HS-SPME procedure
reliable characterization of rums and an increased confidence of the
consumers in this type of products in terms of authenticity is needed. Prior to HS-SPME-GC-MS analysis, the rum bottles were left to reach
Nowadays, numerous methods have been described for the classi- room temperature for 1 h. After that, they were opened for the first
fication of alcoholic beverages based on the analysis of the volatile time. Rums with common origin were analysed equally across the
composition [9]. For that purpose, gas chromatography (GC) coupled to sampling sequence according to a block design in order to guarantee
mass spectrometry (MS) has been one of the most frequently used their comparability and lack of potential analytical bias. Three re-
technique [9]. In recent years, headspace solid phase microextraction plicates of each bottle were analysed.
(HS-SPME) has become the extraction method of choice. The combi- Blanks which consisted of a mixture of Milli-Q water (J.T. Baker)
nation of HS-SPME and GC-MS has been applied to different matrices and ethanol (Sigma-Aldrich; San Louis, MO, USA) at a ratio of 63:37 v/v
such as wine [15–19], beer [20–24], tea beers [25] and other popular were prepared to simulate the alcohol content in a typical commercial
spirit beverages, such as whiskey [26,27], gin [28], or cocktail bitters rum. Blanks were analysed between each brand for various specific
[29]. purposes: (i) to check the potential contamination generated by the
However, to our knowledge, rum studies are less frequent and they septum (blank correction during the statistical treatment of the data),
have been generally limited to the comparison of this type of spirit with (ii) to evaluate potential carry over effect in the fiber and (iii) for ad-
their South American analogue (cachaça) [30], to ascertain a specific ditional cleaning up of the fiber.
geographic origin (Cuban rums from non-Cuban rums) [31] or to the For SPME extraction, different combinations of the selected para-
identification of some aroma indicators [6,32–35]. Due to the com- meters that are known to affect the fiber performance (sample volume,
plexity and variability of rum preparation, their classification re- incubation time, extraction temperature, extraction time, and stirring
presents an analytical challenge. speed) were applied in order to maximize the number and the intensity
To overcome this, multivariate analysis has been commonly em- of volatile compounds extracted. Finally, ten mL of each rum sample
ployed for other such complex matrices in order to take advantage of were placed into a 20-mL glass vial fitted with a magnetic cap and a
the huge amount of data obtained from the GC-MS analysis. PTFE/silicone septum of 1.5 mm thickness. After 5 min of preheating
Unsupervised chemometric techniques as principal component analysis the sample at 65 °C (continuous stirring, 250 rpm), the SPME fiber was
(PCA) [36,37] as well as hierarchical cluster analysis (HCA) [38,39] exposed to the sample headspace for an adsorption time of 30 min with
have been commonly used for a preliminary inspection of the data. constant stirring (250 rpm).
Further supervised classification methods, such as linear discriminant After extraction, the fiber was inserted into the GC injector using a
analysis (LDA) [40,41] have been successfully applied for chemometric 0.8 mm dedicated SPME liner to allow thermal desorption of the ana-
analysis, as well as for the classification of different types of beverages lytes at a temperature of 250 °C for 2 min. The compounds were des-
or foods [42]. orbed into the injector in splitless mode for 2 min, prior to the GC-MS
The aim of this study has been the classification of various types of analysis. After desorption, a fiber cleaning step was carried out for 6
rums by developing a comprehensive and robust analytical strategy for additional min with an increased split rate of 100:1.
the analysis of the volatile/semi-volatile compounds. After simple and
completely automated HS−SPME−GC−MS analyses, the raw data 2.4. GC-QqQ-MS analysis
were processed applying available statistical tools for targeted and
untargeted analysis. For exploratory data analysis, unsupervised che- A Scion GC system equipped with an autosampler (Bruker
mometric techniques using unlabelled data were applied. Afterwards, Corporation, Freemont, CA, USA) was used for chromatographic ana-
supervised techniques were applied to achieve rums classification based lyses. Polydimethylsiloxane (100 µm film thickness) SPME fibers were
on the chemical correlations between samples. obtained from Supelco (Bellefonte, Pennsylvania, USA). After their
conditioning following manufacturer's recommendations, the fibers
were used without any further modification. A VF-5 ms capillary
2. Materials and methods column (30 m x 0.25 mm i.d. x 0.25 µm film thickness) from Varian
(Palo Alto, California, USA) was utilized for GC separation. Helium was
2.1. Reagents used as carrier gas at a constant flow rate of 1 mL/min (36.7 cm/s linear
velocity). An untreated fused silica capillary column (2 m x 0.25 mm)
Ethanol HPLC grade was obtained from J.T. Baker (Deventer, from Supelco was used as pre-column.
Holland). C7-C40 saturated alkanes standard mix (1000 µg/mL in n- Mass spectrometric detection was performed by a triple quadrupole
hexane) were supplied by Supelco (Bellefonte, PA, USA). Scion QqQ-MS/MS (Bruker) operating in electron ionization mode (EI,
70 eV). Mass spectral data of the total ion chromatograms (TICs) and
Kovats retention index (KI) of rum samples were compared to the NIST

349
J.R. Belmonte-Sánchez et al. Talanta 187 (2018) 348–356

Table 1
Code, bottle labelling and manufacturer information for the rum samples analysed.
Brand code Bottle code Origin Aging Raw material Distillation method Barrel

A A_5Y Cuba 5 years Molasses Stainless steel columns Oak barrel

A_7Y 7 years Molasses Stainless steel columns Oak barrel
B B_1 Anejo Molasses N/A American oak barrel
B_2 Young Molasses N/A American oak barrel
B_SYRUP1 N/A Molasses N/A American oak barrel
C C_7Y Dominican Republic 7 years Molasses Stainless steel columns American oak barrel
C_10Y 10 years Molasses Stainless steel columns American oak barrel
C_15Y 15 years Molasses Stainless steel columns American oak barrel
D D Anejo N/A N/A N/A
D_R2 Anejo N/A N/A N/A
D_HONEY1 N/A N/A N/A N/A
E E Old reserve N/A N/A N/A
F F Anejo N/A N/A N/A
G G_1 Grenada Double aged. Old Molasses N/A Mixed (American oak barrel+ French oak
reserve barrel)
G_2 Trinidad & Tobago Double aged. Molasses Stainless steel columns Mixed (American oak barrel+ French oak
Overproof barrel)
H H_8Y Guatemala 8 years Sugarcane juice Stainless steel columns American oak barrel
H_12Y 12 years Sugarcane juice Stainless steel columns American oak barrel
H_18Y 18 years Sugarcane juice Stainless steel columns American oak barrel
I I_5Y Jamaica 5–10 years Molasses Copper pot stills American oak barrel
I_12Y 12 years Molasses Copper pot stills American oak barrel
J J_7Y Nicaragua 7 years Molasses Stainless steel columns American oak barrel
J_12Y 12 years Molasses Stainless steel columns American oak barrel
J_18Y 18 years Molasses Stainless steel columns American oak barrel
K K_MAND1 Republic of Mauritius Double aged Sugarcane juice Stainless steel columns N/A
K_VAN1 Double aged Sugarcane juice Stainless steel columns N/A
L L_HONEY1 Spain N/A N/A N/A N/A
M M Anejo Molasses N/A N/A
M_R2 Molasses N/A N/A
N N Dorado N/A N/A N/A
P O Anejo N/A N/A N/A
O P N/A N/A N/A N/A
Q Q_3Y Venezuela 3 years Molasses Mixed (Steel column+a bit of American oak barrel
copper)
Q_10Y 10 years Molasses Mixed (Steel column+a bit of French oak barrel
copper)

1: additives added.
2: same rum type purchased at a different liquor store.
N/A; information not available.

(2014) mass spectra database. positive electron ionization (EI) source was operated at 70 eV at a
At the beginning of the analysis, the column temperature was set to temperature of 200 °C with a transfer line temperature of 250 °C. High-
35 °C, and the temperature was increased to 100 °C at a 4 °C/min rate, resolution mode was operated at a 60000 full width at half maximum
and then to 250 °C (hold 20 min) at a rate of 20 °C/min. The total run (FWHM) resolving power (m/z 207).
time was 43.75 min. SPME extraction was operated using an identical fiber selected for
The QqQ mass spectrometer was operated in full scan mode. The GC-QqQ-MS analysis (polydimethylsiloxane, 100 µm film thickness).
temperatures of the transfer line, manifold, and ionization source were Quantitation was performed using Xcalibur 4.1 and TraceFinder 4.1
set to 280, 40, and 280 °C, respectively. The electron multiplier voltage software (Thermo Fisher Scientific, Les Ulis, France).
was set to 1600 V (+200 V offset above the value obtained in the auto-
tuning process). Mass peak widths set in the first and third quadrupole
were of m/z 1.5 and 2.0, respectively. The analysis was carried out in 2.6. Pre-processing and data treatment
the range of mass/charge ratios of m/z 50–400.
The volatile compounds of interest were identified against a com- The GC−MS raw data (XMS file format) were converted to CDF file
mercial library (NIST14) and by the use of GC retention indices. format using the Openchrom software (Armonk, NY, USA) [44]. Then,
Retention time of each volatile was converted to the Kovats retention the dataset was pre-processed with MZmine 2.23 software (Norwood,
index using C7-C40 n-alkanes as references and verified with those MA, USA) [45]. The MZmine's“3D Viewer” tool offered a three-dimen-
reported in the literature. A retention index window of ± 20 was ap- sional representation of the total ion chromatogram (TIC) in order to
plied to MS peak identification assignment [43]. work with useful ranges of intensity, retention time and m/z ranges,
(Fig. S-1, Supporting Information). The pre-processement treatment
2.5. HRMS-Q-Exactive analysis consisted of the optimisation of the following steps: (i) centroid mass
detection, (ii) chromatogram builder, (iii) chromatographic deconvo-
As an additional confirmation method, a Q-Exactive-GC hybrid lution, (iv) alignment, (v) peak list, (vi) duplicate peak filter and (vii)
quadrupole Orbitrap mass spectrometer (Q-Exactive™, Thermo Fisher gap filling. The results were stored as a CSV format file.
Scientific, Bremen, Germany) was used with the same chromatographic The data set was imported as a XML file using Excel, version 2013
conditions as the GC-QqQ-MS analysis reported above. A VF-5 ms ca- (Redmond, WA, USA) and then, it was processed by normalising the ion
pillary column (30 m x 0.25 mm i.d. x 0.25 µm film thickness) from intensities between 0 and 1: [value-minimum value]/[maximum value-
Varian (Palo Alto, California, USA) was used for GC separation. The minimum value].

350
J.R. Belmonte-Sánchez et al. Talanta 187 (2018) 348–356

2.7. Processing and multivariate data analysis rum samples (i.e. C_15Y) contained lower amounts of ethyl acetate or
ethyl esters than younger rum samples of different brands (i.e., A_7Y,
All statistical analyses in this study were conducted using the SPSS I_5Y, H_8Y and H_12Y).
23.0 software package (SPSS Inc, Chicago, IL, USA). For sample clas- An evaluation of the presence of the target compounds in the dif-
sification and discrimination, HCA, PCA and LDA were applied. ferent rums did not provide a clear correlation between their presence
in the studied samples and other factors such as origin, raw material,
3. Results and discussion distillation type, barrel type, etc. Therefore, to overcome the limitations
observed in the targeted analysis, further chemometrical tools were
3.1. HS-SPME-GC-MS method evaluated for untargeted analysis.

A HS-SPME procedure was developed to carry out the analysis of the 3.3. Pre-processing and dataset treatment
volatile/semivolatile compounds in rums. For that, a typical commer-
cial 7-years old rum was used. In the first stage, the raw data export to CDF format was visualized
The 100 µm polydimethylsiloxane (PDMS) SPME fiber was selected using the 3D Viewer tool of MZmine 2.23. Details of the MZmine steps
as it has been extensively used for extracting volatile/semi-volatile and parameters are summarized in Table S-1 (Supporting Information).
compounds with a wide polarity range [46], some applications in al- Once the alignment was completed and before the “Gap Filling”
coholic drinks such as wine [47] and rum [6,35]. process, the extraneous ions detected in all the samples (corresponding
For the GC–QqQ-MS analysis, the starting oven temperature was set to SPME fiber, rubber septum, column bleeding, etc.) were substrated.
to 35 °C to allow the elution of the most volatile compounds. In order to For that purpose, a blank sample was analysed between each analysis of
avoid the chromatographic co-elution of high concentrations of low every type of commercial rum. This allowed the possibility of sub-
molecular weight alcohols, such as ethanol and minor volatile com- tracting the ions detected by the software in these samples (known
pounds, each sample was monitored in full scan mode in the m/z range interfering masses), that could lead to errors in classification. These
50–400. ions were mostly represented by cyclosiloxane signals from the SPME
fiber (m/z 222, 296, 370), siloxanes from the sample vial septa (m/z 73,
3.2. Analysis of targeted compounds 207, 281), and phthalates from plasticizers contamination (m/z 149).
The resulting data from MZmine were exported as a CSV file con-
Targeted analysis was focused on the determination of some se- taining a matrix giving the ion intensity values for each rum sample. In
lected compounds traditionally present in rum samples and already the time window from 1.20 to 30.00 min for a mass range m/z from 50
reported as relevant compounds for the organoleptic properties of such to 400, when 96 analyses (33 different rum samples with three re-
spirit drinks [3,4,6]. For this purpose, ethyl acetate, furfural, 5-hydro- plicates each, except 3 rum samples that were studied in duplicate)
xymethylfurfural, 2,5-furandicarboxaldehyde, vanillin, 3-methyl-1-bu- were performed, 231 variables (number of ions detected as the relevant
tanol, whiskey lactone and the ethyl esters of the majoritarian car- ones by MZmine from the GC-MS data) were obtained. Thus, a total of
boxylic acids between C8 and C16 (individual and as sum of all them) 22176 data points were processed and used for multivariate analysis.
were monitored and found in most of the studied rum samples. Their
identification was carried out by comparison of their mass spectra with 3.4. Unsupervised methods: Exploratory data analysis
NIST mass spectrum library and considering the retention index tabu-
lated in bibliography [43]. An exploratory data analysis by HCA was carried out. The HCA was
Six of the studied rum brands were available at different ages. For applied in order to reduce the dimensions of the data set by grouping
them, the peak areas of various discriminant compounds were plotted relatively similar samples in one cluster and relatively dis-similar ob-
to check the content variation among the samples. The data for ethyl jects in another. Besides, the HCA enabled to check the repeatability of
acetate and the sum of the even ethyl esters between C8 and C16 are the analyses, detect anomalous values and discard any outlier re-
shown in Fig. 1. The data obtained for the individual ethyl esters are plicates. At this step, four outlier replicates out of a total of 96 rum
provided in the Supporting Information Fig. S2. It can be observed that analyses were discarded. However, a minimum of two replicates per
with a few exceptions, there is a trend for an increased peak area with rum sample were always considered.
rum aging within the same brand. For example, ethyl acetate and the In this first step, the HCA study classified satisfactorily five samples
sum of the ethyl esters were found at higher amounts in older rums (I, J, containing additives such as honey, syrup or flavoring (labelled as va-
C and Q samples). However, these compounds content was not sa- nilla and tangerine flavored rums). The dendrogram obtained is shown
tisfactorily correlated with aging between samples of different brands in Fig. 2. Their chromatographic profiles showed relevant differences
and therefore, they cannot be used as proper chemical indicators of regards to the rest of rum samples. Therefore, and after concluding
aging in rums. For example, in both plots it can be observed that some about the easy discrimination of rum samples containing additives,

Fig. 1. Bar chart representation of ethyl acetate (left) and the sum ethyl esters between C8 and C16 (right) among different rum samples. Arrows were added to
highlight aging trends within the same brands.

351
J.R. Belmonte-Sánchez et al. Talanta 187 (2018) 348–356

Fig. 2. Dendogram obtained by HCA using the total rum batch. See Table 1 for sample code.

these kind of samples were disregarded to facilitate further chemo- Therefore, a new m/z range was investigated (m/z 35–400) for this
metric classification of the rest of the rums. particular case obtaining an adequate library spectrum comparison
Applying the factor analysis method, where the data was submitted (Fig. S3 in Supporting Information) and a 100% HRF score. However, it
to an orthogonal rotation (varimax rotation), a total of 40 variables (m/ should be noted that the use of a lower mass range has caused sig-
z) were selected as the most discriminant ions (score values higher than nificant interferences for high volatility compounds (retention time
0.80) for further PCA. The factor analysis scores obtained for all ions lower than 5 min) due to the presence of high concentrations of ethanol
are shown in Table S-2 (Supporting Information). (Fig. S4 in Supporting Information).
PCA using the selected variables was applied to transform the high- Two of the studied chromatographic peaks (retention time 18.4 and
dimensional variables into a small number of orthogonal factors, 19.5 min) were initially identified by NIST as ionene (1,2,3,4-tetra-
principal components (PCs), whilst accounting for the largest variance. hydro-1,1,6-trimethyl- naphthalene). However, HRF study of the HRMS
PCA three-dimensional representation (Fig. 3) provided 88% of the data provided more reliable confirmation for the peak at 18.4 min (HRF
total variance explained on the three first PCs. In this plot, PC1, PC2 score of 100.00%). The chromatographic peak at retention 19.5 min
and PC3 account for 36%, 36% and 15% of the individual variance, also presented a high HRF score (99.53%) and probably it is a ionene-
respectively. derivative, but not included in the NIST library. Ionene has been re-
ported as a pyrolysis degradation of carotenoid products [49] that have
been reported as present in sugarcane [50]. Therefore, ionene and its
3.5. Identification of the most discriminant compounds related compound can be produced during molasses production.
For the identification of the compound at RT 11.8 min, further
These most discriminant ions were initially assigned to 13 volatile discussion was required. After GC-QqQ-MS analysis, NIST library search
organic compounds based on their retention indexes and mass spec- showed trans-2-tetrahydro-5-methyl-furanmethanol as the main tenta-
trometric data using the NIST database (match factor higher than 750). tive identification option. However, further HRMS Q-Exactive Orbitrap
Further identification studies by HRMS with Q-Exactive Orbitrap were analysis identified tetrahydro-2H-pyran-2-methanol as the first choice.
carried out and their mass spectra evaluated by a high-resolution fil- Both compounds present the same formula (C6H12O2), exact mass (m/z
tering (HRF) tool from the Tracefinder software. Initially, the HRF as- 116.08318) and very similar mass spectrometric profile. The main
signed tentative identification was based on the use of high-resolution difference between both spectra is the distribution of ions in the cluster
mass spectra but using the traditional spectral matching at unit re- (m/z 55–60) shown in Fig. 4. The cluster profile obtained in the ex-
solution. Later, all unique combinations of atoms from these candidate perimental spectrum fits with the one shown for tetrahydro-2H-pyran-
precursors are generated and matched to m/z peaks using narrow mass 2-methanol. This compound also presented a high HRF score of 100%.
tolerances [48]. The HRF scores obtained for the candidates ranged Table 2 shows the 13 selected compounds with their corresponding
from 99.50% to 100.00%, confirming 11 of the compounds previously identified compounds, empirical formula, selected ions, retention time,
assigned by GC-QqQ-MS analysis using retention index and NIST da- KI value, HRF score and identification methods used. It can be readily
tabase search (Table 2). Although different identification methods were observed that all ions from PC1 and PC2 belong to six ethyl ester
applied for the same compound, further discussion was necessary for compounds (hexadecanoic acid, ethyl ester; (E)-9-octadecenoic acid,
final confirmation of the three hesitant results. ethyl ester; tetradecanoic acid, ethyl ester; octanoic acid, 3-methylbutyl
The initial confirmation of ethyl acetate presented certain doubts ester; decanoic acid, ethyl ester; octanoic acid, ethyl ester). Ethyl ester
because of the monitored mass range (m/z 50–400). Its predominant m/ compounds, formed by the reaction of ethanol with acyl-CoA, are well-
z 43 (CH3CO+) was not observed in those experimental conditions.

352
J.R. Belmonte-Sánchez et al. Talanta 187 (2018) 348–356

Fig. 3. PCA representation using the most discriminant ions.

known for playing an important role in the organoleptic properties of [53]. Additionally, ionene and tetrahydro- 2H-Pyran-2-methanol were
fermented beverages, due to their fruity aroma [51]. For PC3, high also identified as highly discriminant in PC3.
discrimination power was observed for low molecular weight com-
pounds, such as ethyl acetate, diethoxymethane and 1,1 diethox- 3.6. Supervised analyses
ybutane. Ethyl acetate provides fruity and brandy notes to beverages
[52]. Franitza et al. has already reported the presence of 1,1-diethoxy- After variables selection, LDA was applied in order to classify the
3-methyl- butane in two different rums, providing intense fruity aroma rums according to different groups by maximizing the ratio of between-

Table 2
Most discriminant ions for PC1, PC2, and PC3 with their score. retention time and corresponding tentatively identified compounds.
PC Compound Empirical m/z (score) RT (min) KIa HRF (%)b Identification method
formula

1 Hexadecanoic acid, ethyl C18H36O2 185.2 (0.94); 241.2 (0.94); 284.3 (0.93); 213.2 (0.93); 157.2 24.10 1993 100.00 KI. Libc. HRF
ester (0.89); 115.1 (0.88); 88.1 (0.86); 55.1 (0.85);
1 (E)-9-Octadecenoic acid ethyl C20H38O2 180.2 (0.91); 112.2 (0.90); 222.3 (0.87); 264.2 (0.86); 137.1 25.17 2170 100.00 KI. Lib. HRF
ester (0.85); 109.1 (0.85); 169.1 (0.84)
1 Tetradecanoic acid, ethyl C16H32O2 88.1 (0.86); 55.1 (0.82) 23.05 1789 100.00 KI. Lib. HRF
ester
2 Octanoic acid, 3-methylbutyl C13H26O2 70.2 (0.95); 127.2 (0.92) 20.89 1441 99.83 KI. Lib. HRF
ester
2 Decanoic acid, ethyl ester C12H24O2 157.1 (0.92); 115.1 (0.92); 88.1 (0.81); 55.1 (0.91) 20.44 1389 99.50 KI. Lib. HRF
2 Octanoic acid, ethyl ester C10H20O2 88.1 (0.85); 127.2 (0.85); 57.2 (0.85) 18.17 1194 99.86 KI. Lib. HRF
3 Ionene C13H18 131.1 (0.94); 159.1 (0.91); 144.1 (0.90); 116.1 (0.85); 91.1 18.4 1211 100.00 KI. Lib. HRF
(0.85); 113.1 (0.83)
3 Butane, 1,1-diethoxy-3- C9H20O2 103.1 (0.94); 73.1 (0.94); 75.1 (0.94) 9.2 918 99.82 KI. HRF
methyl-
3 Ionene-derivative C13H18 159.1 (0.90) 19.5 1295 99.53 HRF
3 Unknown – 130.2 (0.89) 18.2 1197 – –
3 Ethyl Acetate C4H8O2 54.7 (0.87) 2.3 611 100.00 KI. Lib. HRF
3 2H-Pyran-2-methanol, C6H12O2 85.1 (0.86) 11.8 988 100.00 KI. Lib. HRF
tetrahydro-
3 Methane, 1.1-diethoxy- C5H12O2 103.1 (0.84) 3.3 668 – KI. Lib

a
KI; Kovats retention indices.
b
HRF (High-Resolution Filtering score); percentage of the spectrum obtained by HRMS Orbitrap that can be explained by combination of accurate mass, library
matching and percentage of explained ions observed.
c
Lib; Identification based on mass spectrometric data using the NIST database (match factor higher than 750).

353
J.R. Belmonte-Sánchez et al. Talanta 187 (2018) 348–356

Fig. 4. Comparison between theoretical mass spetrum of tetrahydro-2H-pyran-2-methanol (a), theoretical mass spetrum of trans-tetrahydro-5-methyl-2-fur-
anmethanol (b) and the obtained experimental spectrum by HRMS Q-Exactive Orbitrap (c) with zoom-in on the m/z 55–60 cluster (d).

class variance and minimizing the ratio of within-class variance. the manufacturing country.
Classification groups were selected based on the main information When raw material was used as classificatory criterion (Fig. 5.a),
about the rum elaboration, typically provided by rum manufacturers. 97% of rums were correctly classified. The molasses group was clearly
The leave-one-out method [54] was used as cross-validation pro- separated from the sugar cane group, while the group containing the
cedure to evaluate the prediction ability (Q2) for each LDA model using rums where no information was available were placed close to the
the previously selected variables. For a good predictability, the differ- molasses group. The LDA data distribution reveals that the NA rums
ence between R2 (coefficient of determination) and Q2 value should not with no information about their production process could be elaborated
exceed 0.3 and poor robustness of the model is usually suspected when from molasses. The cross-validation process, which used only the sugar
that happens [55,56]. The rums that had no production information cane group and the molasses group, provided a prediction ability Q2
available for the respective classifying group were not used to ensure an = 0.91 indicating that clear differences between these two groups
accurate prediction. provide a good prediction ability of this model.
For the country group, the 94% of the rums with 10 different origins When the classification was performed according to distillation
were correctly classified (R2 = 0.94) (Fig. S-5 in Supporting method, where four categories were used (copper pot stills, stainless
Information). When cross-validation was carried out, the LDA model steel columns, a mixture, and NA) (Fig. S-5 in Supporting Information),
had a prediction ability of Q2 = 0.73. This cross-validation value can be 92% of the rums were correctly classified. The cross-validation showed
explained because the composition of rum is mostly dependent on the a prediction ability of Q2 = 0.87 of the rums. Rums made in copper pot
manufacturing process due to its complexity and variability, rather than stills were clearly differentiated from the rest. The other three

354
J.R. Belmonte-Sánchez et al. Talanta 187 (2018) 348–356

Fig. 5. LDA representation showing the classification of 28 rums based on (a) raw material and (b) aging period.

categories were correctly classified in most of the cases, despite their butane). For the confirmation of the compounds, a strategy based on
proximity in the LDA representation. Copper pot stills are the tradi- the combination of retention indexes, NIST database matching using
tional method to distil rums; however, these days, many manufactures low-resolution mass spectrometry and HRF scores using high-resolution
prefer stainless steel columns for their increased efficiency. With steel spectra obtained by HRMS Q-Exactive Orbitrap was employed.
columns, only one fraction is collected, whereas copper pot stills pro- The 28 traditional rums were classified based on manufacturing
vide manufacturers with the possibility to combine the head, heart and country, raw material, distillation method, barrel type, and aging
tail fractions to obtain specific rums. This elaboration difference can period with data classification values from R2 = 0.94 to R2 = 0.97
explain the clear LDA distinction of the copper pot distillation data. (prediction ability from Q2 =0.65 to Q2 =0.91). LDA results showed
For barrel type (unspecified oak barrel, American oak barrel, French conclusive differences between the classification groups and the im-
oak barrel, NA), 95% of rums were correctly classified rums (Fig. S-5 in portance of enough representative sample for each group. The applied
Supporting Information). A cross-validation value of Q2 = 0.86 was classification strategy allowed for a better understanding of rum com-
obtained. The graphical proximity of the oak and American oak barrels position, including for rums with not accurate label information.
could be an indication that American oak is the typical wood used when Moreover, this technique could be of interest to future investigations,
the manufacturers do not specify the type of barrel utilized. for example, to other spirit beverages.
The last LDA representation (Fig. 5.b) was based on rum aging
(young, medium, old)., The rum aging parameter was created by con- Conflict of interest
sidering rums aged less than 5 years as young, those aged between 5
and 10 years as medium, and rums aged for more than years as old. A The authors declare no conflict of interest.
94% of correctly classified rums were obtained for aging. For cross-
validation, Q2 = 0.65 were obtained. This value can be explained by Appendix A. Supporting information
the blending process and the differences in legislation about age la-
belling. Supplementary data associated with this article can be found in the
online version at http://dx.doi.org/10.1016/j.talanta.2018.05.025.
4. Conclusions
References
The present study has developed an analytical strategy for the
classification of rums depending on different steps from the production [1] Authentic Caribbean Rum, Decade of rum, 〈https://www.acr-rum.com/assets/pdf/
decade-of-rum-english-usa.pdf〉, 2013.
process using the most discriminant compounds of the volatile fraction.
[2] Top 25 Countries by Rum Consumption Per Capita, (n.d.). 〈https://blog.
For that, the targeted and untargeted analysis of rums was evaluated euromonitor.com/2017/08/top-25-countries-by-rum-consumption-per-capita.
using HS-SPME-GC-MS and chemometric tools. The target analysis html〉 (Accessed 20 March 2018).
found some chemical indicators (ethyl acetate and ethyl esters of car- [3] L. Franitza, M. Granvogl, P. Schieberle, Influence of the production process on the
key aroma compounds of rum: from molasses to the spirit, J. Agric. Food Chem. 64
boxylic acids) that could be correlated with aging within the same (2016) 9041–9053, http://dx.doi.org/10.1021/acs.jafc.6b04046.
brand, but presented clear limitations when they were used across [4] J.A. Pino, S. Tolle, R. Gök, P. Winterhalter, Characterisation of odour-active com-
different brands. No other correlations with other parameters were pounds in aged rum, Food Chem. 132 (2012) 1436–1441, http://dx.doi.org/10.
1016/j.foodchem.2011.11.133.
found. On the other hand, the untargeted analysis using chemometric [5] D.R. Cardoso, L.G. Andrade-Sobrinho, A.F. Leite-Neto, R.V. Reche, W.D. Isique,
tools led to the classification of the 33 rums of different brands and M.M.C. Ferreira, B.S. Lima-Neto, D.W. Franco, Comparison between cachaça and
ages. For that, unsupervised (HCA, PCA) and supervised techniques rum using pattern recognition methods, J. Agric. Food Chem. 52 (2004)
3429–3433, http://dx.doi.org/10.1021/jf035262+.
(LDA) were employed. The HCA showed considerable differences be- [6] J. Pino, M.P. Martí, M. Mestres, J. Pérez, O. Busto, J. Guasch, Headspace solid-phase
tween the traditional rums and the rums prepared by addition of honey, microextraction of higher fatty acid ethyl esters in white rum aroma, J. Chromatogr.
syrup and flavoring. This chemometric method offers the potential to A. 954 (2002) 51–57, http://dx.doi.org/10.1016/S0021-9673(02)00167-X.
[7] D.A. Nicol, Rum, in: A.G.H. Lea, J. Piggott (Eds.), Fermented Beverage Prod.
clearly distinguish these rums with additives from the rest. For the
Springer, US, 2003, pp. 263–287.
correct classification of traditional rums, PCA provided 40 ions as re- [8] J.E. Murtagh, Feedstocks, fermentation and distillation for production of heavy and
levant chemical descriptors corresponding to 13 discriminant com- light rums, alcohol Textb. A Ref. beverage, Fuel Ind. Alcohol Ind. (2007) 243–255.
[9] P. Wiśniewska, M. Śliwińska, T. Dymerski, W. Wardencki, J. Namieśnik, Application
pounds (e.g. hexadecanoic acid ethyl ester, octanoic acid ethyl ester,
of gas chromatography to analysis of spirit-based alcoholic beverages, Crit. Rev.
decanoic acid ethyl ester, ethyl acetate and 1,1-diethoxy-3-methyl-

355
J.R. Belmonte-Sánchez et al. Talanta 187 (2018) 348–356

Anal. Chem. 45 (2015) 201–225, http://dx.doi.org/10.1080/10408347.2014. solid-phase microextraction method for the determination of some secondary
904732. compounds of Brazilian sugar cane spirits by gas chromatography, J. Agric. Food
[10] D.W. Lachenmeier, R. Attig, W. Frank, C. Athanasakis, The use of ion chromato- Chem. 49 (2001) 3533–3539.
graphy to detect adulteration of vodka and rum, Eur. Food Res. Technol. 218 (2003) [35] J.A. Pino, Characterization of rum using solid-phase microextraction with gas
105–110, http://dx.doi.org/10.1007/s00217-003-0799-8. chromatography-mass spectrometry, Food Chem. 104 (2007) 421–428, http://dx.
[11] P. Hernández-Orte, E. Franco, C.G. Huerta, J.M. García, M. Cabellos, J. Suberviola, doi.org/10.1016/j.foodchem.2006.09.031.
I. Orriols, J. Cacho, Criteria to discriminate between wines aged in oak barrels and [36] R.S. da Costa, S.R.B. Santos, L.F. Almeida, E.C.L. Nascimento, M.J.C. Pontes,
macerated with oak fragments, Food Res. Int. 57 (2014) 234–241, http://dx.doi. R.A.C. Lima, S.S. Simões, M.C.U. Araújo, A novel strategy to verification of adul-
org/10.1016/j.foodres.2014.01.044. teration in alcoholic beverages based on Schlieren effect measurements and che-
[12] CFR, - Code of Federal Regulations of the United States of America, Alcohol, mometric techniques, Microchem. J. 78 (2004) 27–33.
Tobacco Products and Firearms, Washington DC, 1991. [37] J. Bouhlel, D.J. Bouveresse, S. Abouelkaram, E. Baéza, C. Jondreville, A. Travel,
[13] European Parliament Council of the European Union, 110/2008 of the European J. Ratel, E. Engel, D.N. Rutledge, Talanta comparison of common components
parliament and of the council of 15 January 2008 on the definition, description, analysis with principal components analysis and independent components analysis:
presentation, labelling and the protection of geographical indications of spirit application to SPME-GC-MS volatolomic signatures, Talanta 178 (2018) 854–863,
drinks and repealing Council Regulation (EEC) No 1576/89 Off, 2008. http://dx.doi.org/10.1016/j.talanta.2017.10.025.
[14] Canada Department of Justice, Food and Drug Regulations C.R.C, c 870, 2009. [38] U. Römisch, H. Jäger, X. Capron, S. Lanteri, M. Forina, J. Smeyers-Verbeke,
[15] A.E. Springer, J. Riedl, S. Esslinger, T. Roth, M.A. Glomb, C. Fauhl-Hassek, Characterization and determination of the geographical origin of wines. Part III:
Validated modeling for german white wine varietal authentication based on multivariate discrimination and classification methods, Eur. Food Res. Technol. 230
headspace solid-phase microextraction online coupled with gas chromatography (2009) 31–45.
mass spectrometry fingerprinting, J. Agric. Food Chem. 62 (2014) 6844–6851, [39] A. Montaño, Talanta volatile profiles of green Spanish-style table olives: application
http://dx.doi.org/10.1021/jf502042c. of compositional data analysis for the segregation of their cultivars and production
[16] Z. Xiao, S. Liu, Y. Gu, N. Xu, Y. Shang, J. Zhu, Discrimination of cherry wines based areas, Talanta 169 (2017) 77–84, http://dx.doi.org/10.1016/j.talanta.2017.03.
on their sensory properties and aromatic fingerprinting using HS‐SPME‐GC‐MS and 066.
multivariate analysis, J. Food Sci. 79 (2014) 284–294. [40] A. Biancolillo, R. Bucci, A.L. Magrì, A.D. Magrì, F. Marini, Data-fusion for multi-
[17] Y. Niu, X. Zhang, Z. Xiao, S. Song, K. Eric, C. Jia, H. Yu, J. Zhu, Characterization of platform characterization of an Italian craft beer aimed at its authentication, Anal.
odor-active compounds of various cherry wines by gas chromatography–mass Chim. Acta 820 (2014) 23–31.
spectrometry, gas chromatography–olfactometry and their correlation with sensory [41] J. Saurina, Characterization of wines using compositional profiles and chemo-
attributes, J. Chromatogr. B. 879 (2011) 2287–2293. metrics, TrAC Trends Anal. Chem. 29 (2010) 234–245.
[18] S. Grützmann, V. Caliari, M. Sganzerla, H. Teixeira, Talanta Volatile composition of [42] F. Magagna, L. Valverde-Som, C. Ruíz-Sambl As, L. Cuadros-Rodríguez,
Merlot red wine and its contribution to the aroma: optimization and validation of S.E. Reichenbach, C. Bicchi, C. Cordero, Combined untargeted and targeted fin-
analytical method, Talanta 174 (2017) 752–766, http://dx.doi.org/10.1016/j. gerprinting with comprehensive two-dimensional chromatography for volatiles and
talanta.2017.06.074. ripening indicators in olive oil, Anal. Chim. Acta (2016), http://dx.doi.org/10.
[19] A.K. Hjelmeland, P.L. Wylie, S.E. Ebeler, Talanta A comparison of sorptive extrac- 1016/j.aca.2016.07.005.
tion techniques coupled to a new quantitative, sensitive, high throughput GC – MS / [43] V.I. Babushok, Trends in analytical chemistry chromatographic retention indices in
MS method for methoxypyrazine analysis in wine, Talanta 148 (2016) 336–345, identification of chemical compounds, Trends Anal. Chem. 69 (2015) 98–104,
http://dx.doi.org/10.1016/j.talanta.2015.10.086. http://dx.doi.org/10.1016/j.trac.2015.04.001.
[20] J. Jiao, N. Ding, T. Shi, X. Chai, P. Cong, Z. Zhu, Study of Chromatographic [44] P. Wenig, J. Odermatt, OpenChrom: a cross-platform open source software for the
Fingerprint of the Flavor in Beer by HS-SPME-GC, Anal. Lett. 44 (2011) 648–655. mass spectrometric analysis of chromatographic data, BMC Bioinforma. 11 (2010)
[21] O. Pinho, I.M. Ferreira, L.H. Santos, Method optimization by solid-phase micro- 405, http://dx.doi.org/10.1186/1471-2105-11-405.
extraction in combination with gas chromatography with mass spectrometry for [45] T. Pluskal, S. Castillo, A. Villar-Briones, M. Orešič, MZmine 2: modular framework
analysis of beer volatile fraction, J. Chromatogr. A. 1121 (2006) 145–153. for processing, visualizing, and analyzing mass spectrometry-based molecular
[22] F. Rodrigues, M. Caldeira, J. de S. Câmara, Development of a dynamic headspace profile data, BMC Bioinform. 11 (2010) 395, http://dx.doi.org/10.1186/1471-
solid-phase microextraction procedure coupled to GC–qMSD for evaluation the 2105-11-395.
chemical profile in alcoholic beverages, Anal. Chim. Acta 609 (2008) 82–104. [46] S. Risticevic, D. Vuckovic, J. Pawliszyn, Solid - Phase Microextraction development
[23] D. Saison, D.P. De Schutter, F. Delvaux, F.R. Delvaux, Optimisation of a complete of fiber solid - phase, 2010.
method for the analysis of volatiles involved in the flavour stability of beer by solid- [47] A. Ruiz-Delgado, F.J. Arrebola-Liébanas, R. Romero-González, R. López-Ruiz,
phase microextraction in combination with gas chromatography and mass spec- A. Garrido Frenich, Headspace solid-phase microextraction coupled to gas chro-
trometry, J. Chromatogr. A. 1190 (2008) 342–349. matography-tandem mass spectrometry for the determination of haloanisoles in
[24] N. Moreira, S. Meireles, T. Brandão, P. Guedes, D. Pinho, Talanta Optimization of sparkling (cava and cider) and non-sparkling (wine) alcoholic beverages, Food
the HS-SPME – GC – IT / MS method using a central composite design for volatile Addit. Contam. - Part A Chem. Anal. Control. Expo. Risk Assess. 33 (2016)
carbonyl compounds determination in beers, Talanta 117 (2013) 523–531, http:// 1535–1544, http://dx.doi.org/10.1080/19440049.2016.1229870.
dx.doi.org/10.1016/j.talanta.2013.09.027. [48] N.W. Kwiecien, D.J. Bailey, M.J.P. Rush, J.S. Cole, A. Ulbrich, A.S. Hebert,
[25] L. Rong, L.J. Peng, C.T. Ho, S.H. Yan, M. Meurens, Z.Z. Zhang, D.X. Li, X.C. Wan, M.S. Westphall, J.J. Coon, High-resolution filtering for improved small molecule
G.H. Bao, X.L. Gao, T.J. Ling, Brewing and volatiles analysis of three tea beers identification via GC/MS, Anal. Chem. 87 (2015) 8328–8335, http://dx.doi.org/10.
indicate a potential interaction between tea components and lager yeast, Food 1021/acs.analchem.5b01503.
Chem. 197 (2016) 161–167, http://dx.doi.org/10.1016/j.foodchem.2015.10.088. [49] R.G. Achari, G. Shaw, R. Holleyhead, Identification of ionene and other carotenoid
[26] C.N. Rhodes, K. Heaton, I. Goodall, P.A. Brereton, Gas chromatography carbon degradation products from the pyrolysis of sporopollenins derived from some pollen
isotope ratio mass spectrometry applied to the detection of neutral alcohol in Scotch exines, a spore coal and the Green River shale, Chem. Geol. 12 (1973) 229–234,
whisky: an internal reference approach, Food Chem. 114 (2009) 697–701. http://dx.doi.org/10.1016/0009-2541(73)90091-0.
[27] J.I. Cacho, N. Campillo, P. Viñas, M. Hernández-córdoba, Talanta cloud point ex- [50] J.D. Fontana, B. Czeczuga, T.M.B. Bonfim, M.B. Chociai, B.H. Oliveira,
traction and gas chromatography with direct microvial insert thermal desorption M.F. Guimarães, M. Baron, Bioproduction of carotenoids: the comparative use of
for the determination of haloanisoles in alcoholic beverages, Talanta 160 (2016) raw sugarcane juice and depolymerized bagasse by Phaffia rhodozyma, Bioresour.
282–288, http://dx.doi.org/10.1016/j.talanta.2016.07.029. Technol. 58 (1996) 121–125, http://dx.doi.org/10.1016/S0960-8524(96)00092-2.
[28] S. Vichi, M. Riu-Aumatell, M. Mora-Pons, J.M. Guadayol, S. Buxaderas, E. López- [51] K.J. Verstrepen, G. Derdelinckx, J.P. Dufour, J. Winderickx, J.M. Thevelein,
Tamames, HS-SPME coupled to GC/MS for quality control of Juniperus communis I.S. Pretorius, F.R. Delvaux, Flavor-active esters: adding fruitiness to beer, J. Biosci.
L. berries used for gin aromatization, Food Chem. 105 (2007) 1748–1754. Bioeng. 96 (2003) 110–118, http://dx.doi.org/10.1016/S1389-1723(03)90112-5.
[29] A.J. Johnson, H. Heymann, S.E. Ebeler, Volatile and sensory profiling of cocktail [52] A.P. Dionísio, G. Molina, D. Souza de Carvalho, R. Dos Santos, J.L. Bicas,
bitters, Food Chem. 179 (2015) 343–354, http://dx.doi.org/10.1016/j.foodchem. G.M. Pastore, Natural flavourings from biotechnology for foods and beverages, Nat.
2015.01.114. Food Addit. Ingred. Flavour. (2012) 231–259, http://dx.doi.org/10.1016/B978-1-
[30] E.S.P. Nascimento, D.R. Cardoso, D.W. Franco, Quantitative ester analysis in cacha 84569-811-9.50011-6.
{ç}a and distilled spirits by gas chromatography− mass spectrometry (GC−MS), J. [53] L. Franitza, M. Granvogl, P. Schieberle, Characterization of the key aroma com-
Agric. Food Chem. 56 (2008) 5488–5493. pounds in two commercial rums by means of the sensomics approach, J. Agric. Food
[31] O.M. Sampaio, R.V. Reche, D.W. Franco, Chemical profile of rums as a function of Chem. 64 (2016) 637–645, http://dx.doi.org/10.1021/acs.jafc.5b05426.
their origin. The use of chemometric techniques for their identification, J. Agric. [54] D.W. Osten, Selection of optimal regression models via cross-validation, J.
Food Chem. 56 (2008) 1661–1668, http://dx.doi.org/10.1021/jf0726841. Chemom. 2 (1988) 39–48, http://dx.doi.org/10.1002/cem.1180020106.
[32] F.W.B. Aquino, L.M. Boso, D.R. Cardoso, D.W. Franco, Amino acids profile of sugar [55] R. Veerasamy, H. Rajak, A. Jain, S. Sivadasan, C.P. Varghese, R.K. Agrawal,
cane spirit (cacha{ç}a), rum, and whisky, Food Chem. 108 (2008) 784–793. Ravichandran Veerasamy, et al., Validation of QSAR models-strategies and im-
[33] Z.L. Cardeal, P.J. Marriott, Comprehensive two-dimensional gas chromatography- portance validation of qsar models -strategies and importance, (n.d.). 〈https://pdfs.
mass spectrometry analysis and comparison of volatile organic compounds in semanticscholar.org/4eb2/5ff5a87f2fd6789c5b9954eddddfd1c59dab.pdf〉
Brazilian cacha{ç}a and selected spirits, Food Chem. 112 (2009) 747–755, http:// (Accessed 24 January 2018).
dx.doi.org/10.1016/j.foodchem.2008.06.057. [56] B. Worley, R. Powers, HHS Public Access, 1 (2015) pp. 92–107. doi:10.2174/
[34] E.A. Nonato, F. Carazza, F.C. Silva, C.R. Carvalho, Z. de L. Cardeal, A headspace 2213235X11301010092.Multivariate.

356