Food Chemistry 331 (2020) 127329

Contents lists available at ScienceDirect

Food Chemistry
journal homepage: www.elsevier.com/locate/foodchem

Effect of roasting degree of coffee beans on sensory evaluation: Research T

from the perspective of major chemical ingredients
Guilin Hua,b, Xingrong Penga, Ya Gaoa,b, Yanjie Huanga,b, Xian Lia,b, Haiguo Sua,b,
⁎
Minghua Qiua,b,
a
State Key Laboratory of Phytochemistry and Plant Resources in West China, Kunming Institute of Botany, Chinese Academy of Sciences, Kunming 650201, Yunnan, China
b
University of Chinese Academy of Sciences, Beijing 100049, China

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

Keywords: As the most consumed beverage in the world, the material basis of the sensory quality for roasted coffee beans
Roasted coffee beans has always received much attention. The objective of the present study was to clarify the physical morphology
Chemical ingredients changes, main chemical ingredients and cupping scores of arabica coffee beans of different roasting degrees, by
Sensory indicator scanning electron microscopy (SEM), nuclear magnetic resonance (NMR) and sensory analysis, respectively.
Multivariate analysis
Statistical analysis of the data by multivariate analysis demonstrated that trigonelline, sugars, malate, quinic
Cluster heatmap
acids, γ-butyro-lactone and acetate have the potential to be new roasting markers. Additionally, in all the sensory
Sensory molecular network
indicators, body and acidity were found to be susceptible to roasting degree. Basing on cluster heatmap and
sensory molecular network, the complex relationships between sensory indicators and ingredients were dis-
cussed. The results of partial least squares regression (PLSR) showed that the content of the main coffee in-
gredients can be used to predict the body score.

1. Introduction substances, which can directly decide the quality of the beverage
(Baggenstoss, Poisson, Kaegi, Perren, & Escher, 2008; Steen, Waehrens,
As the second most frequently consumed drink after water, coffee, Petersen, Munchow, & Bredie, 2017, Liu et al., 2019). Previous studies
which has been grown in more than 70 countries, is closely related to have shown that some of the main components, especially sugars and
the lives of billions of people around the world and has become the chlorogenic acids (CGAs) (Farah, De Paulis, Moreira, Trugo, & Martin,
second largest traded commodity worldwide after petroleum (Butt & 2006; Sittipod, Schwartz, & Paravisini, 2019), will be reduced by par-
Sultan, 2011). The quality of coffee beverages is influenced by multiple ticipating in the reactions, however, the main chemical composition
factors such as altitude, soil, climate, processing procedures, roasting changes at different roasting stages have not been well elucidated.
degree, and brewing methods; among all the factors, roasting plays a Beverage quality can be usually determined by sensory analysis in
key role in coffee beverage quality (Dutra, Oliveira, Franca, Ferraz, & which a panel of trained, specialized “cuppers” evaluates coffee quality
Afonso, 2001). using either a table with scoring values (scoring method) or a sensory
Once green beans are roasted, intricate physical and chemical lexicon (descriptive method) (Worku, Duchateau, & Boeckx, 2016). The
changes occur. Physical changes are mainly reflected in the dramatic most widely adopted evaluation standard is the “Coffee Cupping Protocol
changes in the shape, water content, density, color, and internal of Specialty Coffee Association of America”, which includes ten sensory
structure of beans (Schenker, Handschin, Frey, Perren, & Escher, 2000). indicators: aroma, flavor, aftertaste, acidity, body, overall, clean up,
For the observation of the microscopic appearance of coffee beans, SEM uniformity, sweetness and balance.
is the most effective method. Although Schenker et al. (2000) have tried Given that sensory analysis is affected by a variety of subjective
to observe the structure of roasted coffee beans using SEM, the mor- factors which may lead to injustice (Romano et al., 2014; Worku et al.,
phology of coffee beans at different roasting levels has not been clearly 2016), a theoretical alternative method is to find the material basis of
explained. The chemical changes are marked with Maillard reaction the abovementioned sensory indicators, and based on these substances,
and caramelization reaction to produce pleasant or unpleasant to objectively evaluate the quality of coffee. In the past 30 years,

⁎
Corresponding author at: State Key Laboratory of Phytochemistry and Plant Resources in West China, Kunming Institute of Botany, Chinese Academy of Sciences,
Kunming 650201, Yunnan, China.
E-mail address: [email protected] (M. Qiu).

https://doi.org/10.1016/j.foodchem.2020.127329
Received 11 January 2020; Received in revised form 24 May 2020; Accepted 10 June 2020
Available online 13 June 2020
0308-8146/ © 2020 Elsevier Ltd. All rights reserved.
G. Hu, et al. Food Chemistry 331 (2020) 127329

considerable works have been devoted to the discovery of aromatic produced in Pu'er City, Yunnan Province, China and all belong to
volatiles produced during roasting, and have led to the identification of Arabica (Catimor) species. The roasting degree was mirrored by the
over 1000 volatile organic compounds (VOC) (Colzi et al., 2017; Barie, color value of the beans: LRB, 80–95; MRB, 60–75; DRB, 40–55.
Bucking, Stahl, & Rapp, 2015; Lindinger et al., 2008). In addition to the
research on aroma, there are also reports on the material basis of 2.3. Morphological characterization of coffee beans with different roasting
acidity, body and bitterness (Santos, Lopo, Rangel, & Lopes, 2016; degrees
Blumberg, Frank, & Hofmann, 2010; Frank, Zehentbauer, & Hofmann,
2006; Rizzi, Boekley, & Ekanayake, 2004). Due to the intricacies of the For SEM characterization, three green beans, MRBs, and DRBs with
roasting products, there is still a long way to match the indicators of uniform shape were selected, respectively. The beans were cut in the
sensory evaluation with the trace ingredients produced in roasting; direction of the axis to expose the inner section, and the inner section
actually, even the relationship between the main ingredients and sen- was cut into 1–2 mm thick slices. For each bean, 2–3 sections were
sory indicators has not been fully understood, which makes it hard to taken for SEM observation. The surface was uniformly plated with
realize the coffee quality evaluation model based on chemical compo- silver and then placed under a SEM (CARL ZEISS) to collect images.
sitions. To best of our knowledge, there is no precedent for directly SEM magnification was ranged from 40 to 600 times.
using the content of chemical components to predict the sensory score.
In the past few decades, in addition to direct use of instrumentation
2.4. Detection of major chemical constituents by 1H NMR
combined with expert assessment to find flavor components, re-
presented by gas chromatography–olfactometry-mass spectrometry
2.4.1. Extraction solvent screening
(GC-O-MS) (Zou, Liu, Song, & Liu, 2018), instrumental detection com-
Before formal 1H NMR analysis, the extraction solvent was first
bined with multivariate analysis has also been tried for the excavation
screened. The solvents of four different polarities, n-hexane (b. p
of flavor substance and quality control of roasted coffee beans (Sittipod
68.7 °C), ethyl acetate (b. p 77.2 °C), ethanol (b. p 78.4 °C), and water
et al., 2019). For example, proton transfer reaction-mass spectrometry
(b. p 100.0 °C), were used to extract coffee of three roasting degrees
was employed in the discrimination of coffee beans of different roasting
(from the same green bean sample), respectively. In total, 500 g beans
degree (Romano et al., 2014), and near infrared spectroscopy was
of three roasting degrees were crushed and passed through an 80 mesh
exploited as an analytical tool for on-line monitoring of acidity during
sieve. For each individual experiment, 50 g (M0) of coffee powder and
coffee roasting (Worku et al., 2016).
200 mL (V0) of extraction solvent were added to a 500 mL Erlenmeyer
As complicated the aromatic contents of coffee are, nearly all of the
flask, and extracted in a 60 °C water bath for 1 h with the aid of ul-
volatiles are derived from nonvolatile ingredients of green beans, which
trasound. After that, 10 mL (V1) of the extraction solution was removed
breakdown and react during roasting, forming a complex mixture
and concentrated under vacuum. The weight of the concentrate (M1)
(Hashim & Chaveron, 1995). Therefore, in the current study, the main
was accurately weighed and the extraction rate (Y) was calculated by
chemical ingredients were chosen as research subjects to clarify their
the following formula:
production and changes during roasting, as well as assessing the link
between these changes and the quality of coffee beverages. Specifically, V0 M1
Y= × 100%
the microscopic appearance and main chemical ingredients of coffee V1 M0
beans of different roasting degree were clarified with SEM and 1H NMR,
respectively. On the other hand, influences of roasting on sensory in-
dicators were studied. Further, an in-depth study of the relationship 2.4.2. 1H NMR signal acquisition
between major chemical ingredients and sensory evaluation was carried Six samples were randomly selected from three roasting degree bean
out and a sensory molecular network based on major chemical in- samples, respectively. A total of 18 samples were obtained for 1H NMR
gredients was constructed. Finally, an exploratory study was conducted analysis. For each sample, the beans were pulverized and sieved
to build a sensory scores prediction model based on the characteristic through an 80 mesh sieve to obtain uniform powder. Accurately
signal of the main coffee ingredients. weighed samples of 40.0 mg were placed in a 100 mL centrifuge tube,
and 800 μL D2O was added. The samples were treated ultrasonically in
2. 2.Material and methods a water bath at 80 °C for 1 h and centrifuged at 4000 rpm for 15 min.
Then, 450 μL of the supernatant solution was stored in a nuclear
2.1. Materials and instruments magnetic tube at a constant temperature of 4 °C. The collection of 1H
NMR data was performed on the same day as sample extraction on a
D2O (99.9%) for NMR detection was purchased from Saen Chemical Bruker DRX-600 MHz NMR instrument. 1H NMR sampling parameters
Technology (Shanghai) Co., Ltd. Samples were ground by a Jiuyang were as follows: sampling data point 655536; peak width: 12,019 HZ;
JYL-B060 grinder. Centrifugation was performed on an 80–2 benchtop acquisition time: 2.7 s; relaxation time 10 s; number of scans 4.
centrifuge (Shanghai Medical Devices Co., Ltd.). The Bruker DRX- Presaturation method was used for water peak suppression.
1
600 MHz NMR instrument (Bruker, Zurich, Switzerland) was used to H NMR data was displayed in MestReNova and the main chemical
detect 1H NMR spectra. SEM images were acquired with a Sigma300 ingredients were identified against literature and standards (Wei,
(CARL ZEISS) field emission scanning electron microscope. The malic Furihata, Hu, Miyakawa, & Tanokura, 2011). The characteristic che-
acid (99.0%), citric acid (99.0%), and quinic acid (98.0%) for 1H NMR mical shifts of the main compounds were integrated to obtain their
spectra confirmation were purchased from J&K Chemical Co., Ltd. absolute peak areas, and the result are shown in Table S2.
Trigonelline (98.0%) was purchased from Acmec Biochemical Co. Ltd,
Shanghai. Sucrose (98.0%), glucose (98.0%) and xylose (98.0%) were 2.5. Sensory evaluation
purchased from Aladdin Reagent (Shanghai) Co., Ltd. Caffeine (99.0%)
was prepared and tested for purity by HPLC. The sensory analysis was conducted by four Q-Grader panelists
trained on the SCAA cupping protocol (Association, 2015). A total of 10
2.2. Coffee bean samples indicators including aroma, flavor, aftertaste, acidity, body, balance,
overall, clean cup, uniformity and sweetness were scored during the
All sample coffee beans for cupping and 1H NMR detection, in- sensory evaluation. The sensory indicators were evaluated on a scale
cluding 24 light roasted bean (LRB) samples, 24 moderate roasted bean from 6 to 10 with 0.25 increments and the sum of all ten sensory in-
(MRB) samples and 24 dark roasted bean (DRB) samples, were dicators’ scores was the final score of an individual sample.

2
G. Hu, et al. Food Chemistry 331 (2020) 127329

2.6. Data processing cells, the green coffee beans are very hard (Fig. 1A, 1D), so they can be
regarded as aggregates of microreactor units, which provide consider-
Multivariate analysis was performed using SIMCA 14.1 (Umetrics, able support during roasting pressure buildup (Baggenstoss et al.,
Sweden). For the main chemical ingredients data set (n = 18), cen- 2008). The continued increase in temperature is accompanied by de-
tralized was performed for the 1H NMR absolute peak area, then, gradation of flavor precursor compounds such as sugars, amines, and
principal component analysis (PCA) was performed. The chemical in- CGAs to produce a large amount of flavor compounds. It is generally
gredients data set was centered and normalized for partial least squares believed that when the temperature rises to approximately 154 °C, the
discriminant analysis ((PLS-DA)). For the cupping score data set Maillard reaction between sugars and amines begins, and the Maillard
(n = 72), the data were centered and normalized, after which, PLS-DA reaction products, coffee melanoidins, make the color of the beans
was performed to explore the effect of roasting degree on sensory darker.
evaluation. The data set for the body score prediction was also cen- The resulting volatile flavor compounds and CO2 accumulate in the
tralized before PLS modeling. Analysis of variance (ANOVA) was con- cavity inside the coffee bean, causing pressure inside the beans to rise.
ducted for the significance examination of all the models. Roasted coffee beans can withstand theoretical pressures up to 16 bar
The Pearson correlation matrix between 7 cupping indicators and (Baggenstoss et al., 2008). When the vapor pressure and the CO2
compounds was calculated by SPSS 16.0. The cluster heatmap was pressure inside the beans build up, the “1st crack” occurs. The “blow
drawn based on the numpy, matplotlib and seaborn packages in python holes” produced at this time become the outlets of the flavor com-
3.5, and the sensory molecular network was drawn in Cytoscape 3.6.1. pounds inside the beans. As shown in Fig. 1B, 1E, after the “1st crack”
(MRB), a loose pore-like structure was observed under SEM, the mi-
3. Results and discussion cropores gradually became dense from the periphery to the middle, and
an obvious crack appeared in the middle. Oil droplets can be observed
3.1. The production and change of coffee flavor substances around and at the bottom of the pore structure.
The combustion causes the CO2 to continue to accumulate. After
3.1.1. Changes in microstructure and main chemical ingredients during reaching a certain pressure, the “2nd crack” occurs and the interior
roasting begins to coke or carbonize. Previous study found that although there
The change in microstructure caused by roasting has major influ- were high pressure conditions during the roasting process, no signs of
ence on the final quality of coffee beverage. During roasting, the green cell wall rupture were observed in the SEM image (Baggenstoss et al.,
beans are heated at 200–240 °C for 10–15 min. External temperature, 2008; Schenker et al., 2000), which may be due to the change of the cell
roasting time, coffee bean size, shape, water content and other factors wall from the glass state to a more elastic state at high temperatures.
will affect the change of coffee microstructure and the occurrence of From our observation, after the “2nd crack”, the pore size did not
chemical reactions, thus affecting the generation and release of flavor change significantly, but the crack in the middle became obviously
substances. longer, the oily substance on the surface increased, and the cell struc-
Due to unusually thick cell walls and the tight alignment between ture in the center of the bean began to be destroyed (Fig. 1C, F).

Fig. 1. SEM image of coffee beans. Green bean (A, D). Medium roasted beans (between 1st crack and 2nd crack, B, E). Dark roasted beans (after 2nd crack, C, F). (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

3
G. Hu, et al. Food Chemistry 331 (2020) 127329

3.1.2. Exploring changes in the main chemical ingredients based on 1H order of CGA isomers was CQAs > diCQAs > FQAs (Moon, Yoo, &
NMR Shibamoto, 2009). As shown in Fig. 2, although they degraded more
To further clarify the changes in the main chemical ingredients in slowly than sugars, CGAs were significantly reduced as the roasting
beans during roasting, 1H NMR was adopted in the analysis of roasted degree increased, and some CGAs remained in DRB. Contrary to CGAs,
beans. First, the extraction solvent screening experiment was carried the quinic acids (quinic acid and syllo-quinic acid), as the main de-
out. n-hexane, ethyl acetate, ethanol and water were used to extract gradation products of CGAs, increased with roasting degree. According
coffee bean of three roasting degrees, respectively, and their extraction to Moon and Shibamoto (2010) quinic acids will degrade with dark
rate and 1H NMR profiles are shown in Figs. S4 and S5, respectively. roasting to form catechol, phenol, benzonic acid and 2-furyl-methanol;
Interestingly, the extraction rate of n-hexane, ethyl acetate and ethanol however, in this study, the degradation was not detected.
increased with the roasting degree, which may be due to the reaction of There were different findings in the change in caffeine during
carbohydrates during roasting to form small or moderately polar mel- roasting, Rodarte, Abrahao, Pereira, and Malta (2009) found that caf-
anoidins. Although these macromolecular compounds do not show feine was not degraded at any roasting degree. However, according to
obvious 1H NMR signals, their content can increase with roasting de- the research result of Hecimovic, Belscak-Cvitanovic, Horzic, and
gree, resulting in an increase in the organic solvent extraction rate. The Komes (2011) based on Arabica and Robusta coffee beans, LRB con-
extraction rate of water was much higher than that of the other three tained the highest overall content of caffeine, which exhibited a de-
solvents, and no significant difference was detected in the water ex- crease with intensified roasting. Our results showed that during dark
traction rates of the three roasting degrees. 1H NMR spectra of the roasting, although not particularly obvious, caffeine did have a de-
extracts of the four solvents showed that the water extract contained creasing trend (Fig. 2).
the most abundant types of compounds. Sugar, caffeine, CGAs, and Previous studies have reported that trigonelline degraded during
trigonelline are all very water soluble (Upadhyay, Ramalakshmi, & Rao, roasting (Rodarte 2007; Baggenstoss et al., 2008), and was largely de-
2012; Jeszka-Skowron, Zgola-Grzeskowiak, & Grzeskowiak, 2015), but composed upon thermal treatment into N-methylpyridinium, nicotinic
are very low soluble in n-hexane, ethyl acetate. Although caffeine and acid and methylnicotinate (Lang, Yagar, Eggers, & Hofmann, 2008;
CGAs have good solubility in hot ethanol, sugars, as the most important Stadler et al., 2002). The 1H NMR spectra showed that trigonelline
flavor precursor compounds in coffee, are almost insoluble in ethanol. experienced no significant change in LRB or MRB samples, and the
Among the four solvents, water has the highest boiling point, which abovementioned degradation mainly occurred in DRB (Fig. 2). Similar
means that when water is used as extraction solvent, not only can the to trigonelline, choline was experienced significantly reduce after dark
main chemical components in the coffee be extracted to the utmost roasting.
extent, but also the error caused by the evaporation of the solvent can The N-methylpyridine, including 2-methyl-; 2,3-dimethyl-; 2,5-di-
be reduced. In addition, hot water is also the extraction solvent in coffee methyl-; 2,6-dimethyl-; trimethyl-and tetramethyl pyradine were de-
brewing. Therefore, although a small portion of the fat-soluble com- tected in coffee beans, among which, 2-methylpyridine had the largest
ponent may be neglected, in consideration of various advantageous, ratio (Hashim & Chaveron, 1995). They were reported as a kind of
water was finally determined as the extraction solvent for the next Maillard reaction products of sugars and amines or the degradation
experiment. products of trigonelline (Hashim & Chaveron, 1995). N-methylpyridine
Wei et al. (2011) proposed the rapid identification of the main began to form and accumulate during the medium roasting phase of the
chemical components in roasted coffee beans using NMR technology, sample, and achieved the highest content after dark roasting (Fig. 2).
which was successfully applied by Ciaramelli, Palmioli, and Airoldi Additionally, acetate and γ-butyro-lactone significantly increased in
(2019) to identify the variety and origin of coffee. Considering that the DRB samples.
compounds in roasting are substantially dynamic, and the change speed
of different compounds vary from their chemical or physical properties, 3.1.3. Roasting degree marker mining basing on multivariate analysis
we intend to further reveal the changes profile in the main chemical To accurately describe the influence of roasting on the content of
ingredients of roasted coffee beans using 1H NMR technology. the main coffee ingredients, the characteristic chemical shift intervals
Six samples were randomly selected from each roasting degree, and of the compounds were integrated to obtain the absolute peak area
a total of 18 samples were obtained for 1H NMR detection. Based on the information of the compounds (Table S2), and further PCA and PLS-DA
detection method of Wei et al. (Fig. S6), and further confirmed with the analysis was conducted.
1
H NMR data of the standards (Fig. S7), the characteristic signals of PC1 and PC2 of the PCA analysis model explained 80.0% and 16.4%
sugars, CGAs, caffeine, trigonelline, quinic acids, acetate, formate, ci- of variance, respectively. In the score plot, three groups of samples were
trate, malate, γ-butyro-lactone, choline, 2-furyl-methanol and N-me- divided into three clusters from bottom left to top right in increasing
thyl-pyridinium were identified from the extracts (Fig. 2, Table S1). order of roasting (Fig. S8), and no outlier existed in all samples. The
The pyrolysis of sugars and polysaccharides gave rise to a wide PLS-DA model fit four principal components (p < 0.01). The goodness
range of volatiles including aliphatic carbonyls, alcohols, acids, furans of fit of the model was R2X = 0.938, R2Y = 0.944, and the predictive
and cyclic diketones (Steen et al., 2017). Simultaneous interaction with power was Q2 = 0.866. The score plot of the PLS-DA model showed
a nitrogen-containing fragment could give rise to pyrazines, pyridines, that the samples of three groups were well distinguished from left to
pyrroles and, in some cases, imidazoles (Liu et al., 2019). Although right in descending order of roasting (Fig. 3A). The corresponding
these reactions can violently occur in lightly roasting degree, obviously loading plot showed that compounds marked with blue, including
they were not going thoroughly, for there were plenty of sucrose and CGAs, trigonelline, sucrose, other sugars, choline, and caffeine, had the
other sugars remaining in LRB samples (Fig. 2). The sugars were sub- highest content in LRB, while N-methyl-pyridinium, quinic acids, γ-
stantially degraded after moderate roasting, and completely dis- butyro-lactone, malate, 2-furyl methanol and acetate, which are
appeared in DRB samples. marked with brown, had the highest content in DRB.
CGAs can be mainly divided into caffeoylquinic acids (CQAs) with 3 The clarification of the changes in the main chemical ingredients of
isomers (3-, 4-, and 5-CQA), feruloylquinic acids (FQA) with 3 isomers roasted beans is also meaningful to the selection of roasting degree
(3-, 4-, and 5-FQA), and dicaffeoylquinic acids (diCQAs) with 3 isomers markers. Although detecting the external color of the beans is still the
(3,4-diCQA; 3,5-diCQA; 4,5-diCQA) (Moeenfard, Rocha, & Alves, main means of determining roasting degree (Ruosi et al., 2012), con-
2014). They are lost as a consequence of the thermal breakage of sidering the influence of geography, climate, variety and other factors
carbon-carbon covalent bonds, resulting inisomerization in the initial on the flavor precursor compounds, it is hard to become an unified
roasting stages and epimerization, lactonization, and degradation in the standard; thus, more factors must be concluded to conduct more ac-
later stages (Farah et al., 2006; Sittipod et al., 2019). The decreasing curate and unified assessment. Due to their characteristic changes

4
G. Hu, et al. Food Chemistry 331 (2020) 127329

Fig. 2. 1H NMR spectrum identification of main chemical ingredients in roasted bean extracts. 1H NMR spectrum is arranged according to the roasting degree, with
LRB in front, MRB in the middle, and DRB in the end. The characteristic signals of sugars, CGAs, caffeine, trigonelline, quinic acids, acetate, formate, citrate, malate,
γ-butyro-lactone, choline, 2-furyl-methanol, lipid, and N-methyl-pyridinium were identified from the extracts.

during roasting, CGAs, free amino acids, alkylpyrazines, 2-furfuryl al- three roasting degree were included. To intuitively see overall change,
cohol and 5-methylfurfural/2-acetylfuran ratio have been proposed as the aroma, flavor, aftertaste, acidity, body, balance, and overall were
roasting markers (Dutra et al., 2001; Hashim & Chaveron, 1995; plotted with heatmap (Fig. S10). On the whole, in addition to body
Dorfner, Ferge, Yeretzian, Kettrup, & Zimmermann, 2004; Ruosi et al., score, which increases as the degree of roasting, the MRB samples had
2012). The variable importance in projection (VIP) value of the PLS-DA higher cupping scores in terms of aroma, flavor, aftertaste, acidity,
model, first published by Wold and Johansson (1993), is a parameter of balance, and overall than LRB and DRB.
screening biomarkers in metabolomics, and the variable with VIP value To quantify the extent to which the roasting degree affects the score
greater than 1 can be considered as a differential variable between of the evaluation, multivariate data analysis was further conducted.
groups (Farres, Platikanov, Tsakovski, & Tauler, 2015). Among the VIP Based on aroma, flavor, aftertaste, acidity, body, balance, and overall
values corresponding to the first component of the current PLS-DA score of 72 samples, a 72*7 data matrix was constructed, and the matrix
discriminant model, nine (class) compounds, including N-methyl-pyr- was further analyzed by PLS-DA analysis. The goodness of fit of the
idinium, quinic acids, CGAs, γ-butyro-lactone, other sugars, trigonel- model was R2X = 0.9, R2Y = 0.825 the predictive power was
line, 2-furyl-methanol, sucrose, and acetate had VIP values greater than Q2 = 0.785, and the p value for ANOVA was lower than 0.01. As seen
1 (Fig. S9), indicating that these compounds play important roles in the from the score plot, three different categories of coffee can be well
differentiation of roasting degree. classified (Fig. 4A). In this experiment, VIP value was used to determine
In addition to the roasting markers already reported in previous the contribution of the input sensory indicators to the discriminant
articles, from the results of PLS-DA analysis, other ingredients, such as model. It can also directly reflect the influence of the roasting degree on
trigonelline, sugars, malate, quinic acids, γ-butyro-lactone and acetate sensory indicators. The model fitted a total of four principal compo-
also have the potential to be markers of roasting degree. As discussed in nents, and VIP values of variables in each component are shown in
section 3.1.2, the contents of sugars (sucrose and other sugars) and Fig. 4B. In the first component, VIP values of body, acidity, balance, and
quinic acids in beans of different roasting degree were varied sig- flavor were greater than 1. Among other three main components, only
nificantly, making them ideal markers for distinguishing roasting de- VIP values of body and acidity were greater than 1, and total VIP value
gree. Although obvious changes occurred mainly in DRB, the content of indicated that the influence of roasting degree on sensory indicators can
trigonelline, γ-butyro-lactone and acetate have the potential to become be ranked as: body > acidity > balance > flavor > balance >
auxiliary indicators for evaluating the degree of roasting. It is worth overall > aroma.
noting that although the content of γ-butyro-lactone in roast coffee is
relatively low, the significant changes make it a great contribution to
the discriminant model. 3.3. Relationship between main chemical ingredients and sensory indicators

In this section, the data set including the main chemical ingredients,
3.2. Sensory evaluation cupping score and color value for the 18 samples was used to calculate
the Pearson correlation coefficient of the variables (Tables S4 and S5).
To minimize human error in sensory evaluation, all 72 samples of The cluster heatmap shown in Fig. 5A was drawn based on the

5
G. Hu, et al. Food Chemistry 331 (2020) 127329

Fig. 3. The score plot (A) and corresponding loading plot (B) of the first two dimensions of PLS-DA analysis based on the 1H NMR data of samples. (n = 18).

calculation result. Usually, the correlation strength of the variables is leads to a lower cupping score. Indeed, a large body of literature re-
judged by the following range of values: correlation coefficient 0.8–1.0, ported that sugars can produce various flavor compounds that are fa-
extremely strong correlation; 0.6–0.8, strong correlation; 0.4–0.6, vorable for cupping by the Maillard reaction during roasting (Liu et al.,
moderate correlation; 0.2–0.4, weak correlation; 0.0–0.2, weakly cor- 2019; Velasquez, Pena, Bohorquez, Gutierrez, & Sacks, 2019). In ad-
related or uncorrelated. To intuitively clarify the complex relationship dition, from the results of cluster analysis, it can be seen that the effect
between compounds, sensory indicators, total cupping score and color of choline on sensory quality was very similar to that of sugars.
value, and variables with moderate to extremely strong correlation Trigonelline, CGAs and caffeine were under the adjacent branches
were screened to construct a sensory molecular network, in which of the cluster analysis, indicating they have similar correlation between
correlation was represented by lines of different colors and thicknesses content and sensory evaluation. In previous reports, the degradation
(Fig. 5B). products of trigonelline and CGAs along with caffeine were considered
As the maximum amount of ingredients present in coffee water to be related to the bitterness of coffee drinks (Locas & Yaylayan, 2004).
extract, the content of sugars (sucrose and other sugars) was extremely CGAs can be thermally transformed into the bitter-tasting CGA lactones
strongly negatively correlated with body and was strongly negatively 5-O-caffeoyl-muco-γ-quinide, 3-O-caffeoyl-γ-quinide, 4-O-caffeoyl-
correlated with balance, aftertaste, flavor, and overall. A reasonable muco-γ-quinide, 5-O-caffeoyl-epi-δ-quinide, 4-O-caffeoyl-γ-quinide,
explanation is that the insufficient Maillard reaction of sucrose resulting 3,4-O-dicaffeoyl-γ-quinide, 4,5-O-dicaffeoyl-muco-γ-quinide, and 3,5-
in the inadequate formation of favorable flavor compounds, indirectly O-dicaffeoyl-epi-δ-quinide (Frank et al., 2006; Rizzi et al., 2004).

6
G. Hu, et al. Food Chemistry 331 (2020) 127329

Fig. 4. The score plot (A) and corresponding loading plot (B) of the first two dimensions of PLS-DA analysis based on the sensory evaluation data of samples.
(n = 64).

Additionally, CGAs or its analogues were found to be important flavor compounds except lipids and acetic acid, showing the rationality to
regulating substances (Sittipod et al., 2019). Similar to sugars, the reflect the content of the main components by detecting color value,
content of these three ingredients was extremely strongly negatively while it only extremely strongly correlated with body in all sensory
correlated with body. In addition, CGAs showed moderately negative indicators.
correlation with balance.
The remaining nine (classes) compounds, including formate, citrate,
malate, 2-furyl-methanol, lipids, γ-butyro-lactone, quinic acids, acetate 3.4. Establishment of a sensory prediction model
and N-methyl-pyridinium, are distributed in the same cluster.
Combined with the score plot and loading plot of the PLS-DA analysis in Although it has always been the mainstream evaluation method for
Fig. 3, the content of these compounds was found to be relatively high coffee quality, sensory evaluation is thought to be fairly subjective and
in MRB or DRB. The correlation heatmap showed that they were all generally less replicable and consistent than physically based mea-
positively correlated with the body score. Among them, 2-furyl-me- surements. Genetic factors, health, age, education, past experiences,
thanol, γ-butyro-lactone, quinic acids, acetate and N-methyl-pyridinium food habits, smoking habit, and cultural and religious patterns can all
showed strong correlation. In addition, these nine compounds showed be factors that interfere with the objective score of the cupping panels
overall low positive correlation with balance, flavor, aftertaste, aroma (Romano et al., 2014; Worku et al., 2016). Therefore, researchers have
and overall. long been exploring alternative methods for evaluating coffee quality
Although acetic acid content was significantly increased in DRB, the (Worku et al., 2016; Romano et al., 2014).
acidity of coffee did not increase. In fact, it is prone to produce bitter In this research, 11 compounds with extremely strong or strong
substances and mask the acid taste of coffee in DRB. For instance, correlations with body (apart from lipids, citrate, malate and formate)
several heterocyclics were suggested as potential bitter-tasting agents: were tested to be used as input variables to establish a body score
2-furyl-methanol, which can be formed by the degradation of sugars prediction model. Data showed that there were high correlations be-
and quinic acid at high temperatures produces the bitter taste of roasted tween input variables (Table S5, marked by yellow), suggesting that the
coffee due to the interaction with dihydroxy-benzene or trihydrox- variables had a strong collinearity and it was not appropriate to es-
ybenzene (Kreppenhofer, Frank, & Hofmann, 2011); 5-hydroxymethyl- tablish the prediction model using the least squares regression (LSR)
2-furanaldehyde, pyrazines, diketopiperazines and various trigonelline method. PLSR method combines the advantages of PCA and LSR, which
thermolysis products are also putative candidates for the key bitter can extract the latent variables (integrated variables) ti and ui in
compounds in coffee (Blumberg et al., 2010; Frank et al., 2006). Ad- the X and Y data sets, respectively. ti and ui can maximize the original
ditionally, not all organic acids produce acid taste; apart from the bitter information in X(Y) dataset, and at the same time, satisfy the maximum
taste of CGAs and its lactone derivatives mentioned above, quinic acids, extent linear correlation. This method can not only solve the problem of
a kind of well-known thermal degradation products of CGAs, were re- multicollinearity but also exhibit good modeling effects when the
ported to exhibit an aspirin-like bitter taste at a threshold level of number of observed samples is small, so it was chosen as the modeling
10 ppm (Frank et al., 2006). method.
In addition to sweetness, clean up and uniformity, which were not The model fits a total of three principal components and the fit of
included in the analysis due to their small differences between samples, the model was R2X = 0.993, R2Y = 0.834; the predictive power of the
dense strong correlations between sensory indicators are shown in model was Q2 = 0.755 (p < 0.01). The scatter plot of t1/t2 showed no
Fig. 5B. Among them, balance with flavor, balance with aftertaste, and outlier present in observed samples (Fig. S11). A linear relationship is
flavor with aftertaste were of extremely strong correlation, and it is showed in the scatter plot of t1 and u1 (Fig. 6A). Additionally, the
likely that these indicators are different descriptions belonging to a predictive ability of the model was further checked by comparing be-
same intrinsic characteristic. Body and acidity showed relatively low tween prediction and observed body value, and the result showed that
correlation with other indicators, indicating that they independently the gaps between the predicted and observed score in all samples were
describe an intrinsic feature and play an irreplaceable role in the sen- within 0.2 (Fig. 6B). All the evidence suggests that it is feasible to
sory evaluation. Another important feature of roasted beans, color predict the body score of coffee beverage by testing the content of the
value, was strongly or extremely strongly correlated with all detected main ingredients.

7
G. Hu, et al. Food Chemistry 331 (2020) 127329

Fig. 5. A) The cluster heatmap for the correlation

between coffee sensory indicators and major che-
mical ingredients. B) The flavor molecular network
showing the correlation between total cupping
score, sensory indicators, color value and major
chemical ingredients. The thickness of the line cor-
responds to the strength of the correlation. Weak
correlation and medium correlation (positive and
negative) are indicated by gray lines; strong and
extremely strong positive correlation are indicated
by blue lines; strong and extremely strong negative
correlation are indicated by pink lines. (For inter-
pretation of the references to color in this figure
legend, the reader is referred to the web version of
this article.)

4. Conclusion method of sensory molecular networks was proposed to reveal the in-
tricate relationship between major compounds and sensory indicators.
The changes in microstructure and main chemical ingredients Multivariate analysis and correlation analysis proved that the con-
during the roasting of coffee beans were clarified by SEM and 1H NMR, tent of coffee's main ingredients did have significant effects on acidity
respectively. 1H NMR combined with multivariate analysis was con- and body, especially the effect on body. The PLS model based on the
firmed to be an effective strategy to monitor roasting degree. main chemical components was used in an exploratory manner for
Trigonelline, sugars, malate, quinic acids, γ-butyro-lactone and acetate prediction of body score and demonstrated good prediction perfor-
showed potential to be used to monitor the roasting degree of coffee mance. Other cupping indicators considered in this research, including
beans. balance, flavor, balance, overall and aroma, were less affected by the
The complexity of the chemical composition in roasted beans de- content of the main component. To replace these indicators, undetected
termines that it is impossible for researchers to describe a sensory volatile components or trace non-volatile components must be con-
feature using one or a kind of compounds. To more accurately clarify sidered.
the material basis of sensory indicators, in the current research, the

8
G. Hu, et al. Food Chemistry 331 (2020) 127329

Fig. 6. The scatter plot of t1 and u1 (A) and prediction and observed body value (B) of the PLS model for the prediction of body score.

CRediT authorship contribution statement Special Fund Project of Pu’er municipal government (2017) and Expert
workstation Project of Dr. QIU (2018) as well as Foundation of State
Guilin Hu: Conceptualization, Methodology, Software, Validation, Key Laboratory of Phytochemistry and Plant Resources in West China
Formal analysis, Investigation, Data curation, Writing - original draft, (P2015-ZZ09).
Writing - review & editing, Visualization. Xingrong Peng: Project ad-
ministration. Ya Gao: Methodology. Yanjie Huang: Software. Xian Li: Appendix A. Supplementary data
Conceptualization. Haiguo Su: Formal analysis. Minghua Qiu:
Investigation, Resources, Supervision, Project administration, Funding Supplementary data to this article can be found online at https://
acquisition. doi.org/10.1016/j.foodchem.2020.127329.

Declaration of Competing Interest References

Baggenstoss, J., Poisson, L., Kaegi, R., Perren, R., & Escher, F. (2008). Coffee roasting and
The authors declare that they have no known competing financial aroma formation: Application of different time-temperature conditions. Journal of
interests or personal relationships that could have appeared to influ- Agricultural and Food Chemistry, 56(14), 5836–5846.
ence the work reported in this paper. Barie, N., Bucking, M., Stahl, U., & Rapp, M. (2015). Detection of coffee flavour ageing by
solid-phase microextraction/surface acoustic wave sensor array technique (SPME/
SAW). Food Chemistry, 176, 212–218.
Acknowledgements Blumberg, S., Frank, O., & Hofmann, T. (2010). Quantitative studies on the influence of
the bean roasting parameters and hot water percolation on the concentrations of
bitter compounds in coffee brew. Journal of Agricultural and Food Chemistry, 58(6),
This research work was supported financially by the National 3720–3728.
Natural Science Foundation of China (Nos. 31670364, U1902206), Butt, M. S., & Sultan, M. T. (2011). Coffee and its consumption: Benefits and risks. Critical
Project of Key New Productions of Yunnan Province (No. 2015BB002), Reviews in Food Science and Nutrition, 51(4), 363–373.

9
G. Hu, et al. Food Chemistry 331 (2020) 127329

Ciaramelli, C., Palmioli, A., & Airoldi, C. (2019). Coffee variety, origin and extraction degradation of less volatile coffee components: Quinic acid, caffeic acid, and
procedure: Implications for coffee beneficial effects on human health. Food Chemistry, chlorogenic acid. Journal of Agricultural and Food Chemistry, 58(9), 5465–5470.
278, 47–55. Moon, J. K., Yoo, H. S., & Shibamoto, T. (2009). Role of roasting conditions in the level of
Colzi, I., Taiti, C., Marone, E., Magnelli, S., Gonnelli, C., & Mancuso, S. (2017). Covering chlorogenic acid content in coffee beans: Correlation with coffee acidity. Journal of
the different steps of the coffee processing: Can headspace VOC emissions be Agricultural and Food Chemistry, 57(12), 5365–5369.
exploited to successfully distinguish between Arabica and Robusta? Food Chemistry, Rizzi, G. P., Boekley, L. J., & Ekanayake, A. (2004). The influence of roasting-derived
237, 257–263. polymeric substances on the bitter taste of coffee brew. Nutraceutical Beverages:
Dorfner, R., Ferge, T., Yeretzian, C., Kettrup, A., & Zimmermann, R. (2004). Laser mass Chemistry, Nutrition, and Health Effects, 871, 229–236.
spectrometry as on-line sensor for industrial process analysis: Process control of Rodarte, M. P., Abrahao, S. A., Pereira, R. G. F. A., & Malta, M. R. (2009). Non-volatile
coffee roasting. Analytical Chemistry, 76(5), 1386–1402. compounds in coffee from the south of Minas Gerais state region submitted to dif-
Dutra, E. R., Oliveira, L. S., Franca, A. S., Ferraz, V. P., & Afonso, R. J. C. F. (2001). A ferent roasting degrees. Ciencia E Agrotecnologia, 33(5), 1366–1371.
preliminary study on the feasibility of using the composition of coffee roasting ex- Romano, A., Cappellin, L., Ting, V., Aprea, E., Navarini, L., Gasperi, F., & Biasioli, F.
haust gas for the determination of the degree of roast. Journal of Food Engineering, (2014). Nosespace analysis by PTR-ToF-MS for the characterization of food and ta-
47(3), 241–246. sters: The case study of coffee. International Journal of Mass Spectrometry, 365, 20–27.
Farah, A., De Paulis, T., Moreira, D. P., Trugo, L. C., & Martin, P. R. (2006). Chlorogenic Ruosi, M. R., Cordero, C., Cagliero, C., Rubiolo, P., Bicchi, C., Sgorbini, B., & Liberto, E.
acids and lactones in regular and water-decaffeinated arabica coffees. Journal of (2012). A further tool to monitor the coffee roasting process: Aroma composition and
Agricultural and Food Chemistry, 54(2), 374–381. chemical indices. Journal of Agricultural and Food Chemistry, 60(45), 11283–11291.
Farres, M., Platikanov, S., Tsakovski, S., & Tauler, R. (2015). Comparison of the variable Santos, J. R., Lopo, M., Rangel, A. O. S. S., & Lopes, J. A. (2016). Exploiting near infrared
importance in projection (VIP) and of the selectivity ratio (SR) methods for variable spectroscopy as an analytical tool for on-line monitoring of acidity during coffee
selection and interpretation. Journal of Chemometrics, 29(10), 528–536. roasting. Food Control, 60, 408–415.
Frank, O., Zehentbauer, G., & Hofmann, T. (2006). Bioresponse-guided decomposition of Schenker, S., Handschin, S., Frey, B., Perren, R., & Escher, F. (2000). Pore structure of
roast coffee beverage and identification of key bitter taste compounds. European Food coffee beans affected by roasting conditions. Journal of Food Science, 65(3), 452–457.
Research and Technology, 222(5–6), 492–508. Sittipod, S., Schwartz, E., & Paravisini, L. (2019). Identification of flavor modulating
Hashim, L., & Chaveron, H. (1995). Use of methylpyrazine ratios to monitor the coffee compounds that positively impact coffee quality. Food Chemistry. https://doi.org/10.
roasting. Food Research International, 28(6), 619–623. 1016/j.foodchem.2019.125250.
Hecimovic, I., Belscak-Cvitanovic, A., Horzic, D., & Komes, D. (2011). Comparative study Specialty Coffee Association of America (2015). Cupping standards. Retrieved January 1,
of polyphenols and caffeine in different coffee varieties affected by the degree of 2015, from http//://www.scaa.org/?page=resources&d=cupping-standards.
roasting. Food Chemistry, 129(3), 991–1000. Stadler, R. H., Varga, N., Milo, C., Schilter, B., Vera, F. A., & Welti, D. H. (2002).
Jeszka-Skowron, M., Zgola-Grzeskowiak, A., & Grzeskowiak, T. (2015). Analytical Alkylpyridiniums. 2. Isolation and quantification in roasted and ground coffees.
methods applied for the characterization and the determination of bioactive com- Journal of Agricultural and Food Chemistry, 50(5), 1200–1206.
pounds in coffee. European Food Research and Technology, 240(1), 19–31. Steen, I., Waehrens, S. S., Petersen, M. A., Munchow, M., & Bredie, W. L. P. (2017).
Kreppenhofer, S., Frank, O., & Hofmann, T. (2011). Identification of (furan-2-yl) me- Influence of serving temperature on flavour perception and release of Bourbon
thylated benzene diols and triols as a novel class of bitter compounds in roasted Caturra coffee. Food Chemistry, 219, 61–68.
coffee. Food Chemistry, 126(2), 441–449. Upadhyay, R., Ramalakshmi, K., & Rao, L. J. M. (2012). Microwave-assisted extraction of
Lang, R., Yagar, E. F., Eggers, R., & Hofmann, T. (2008). Quantitative investigation of chlorogenic acids from green coffee beans. Food Chemistry, 130(1), 184–188.
trigonelline, nicotinic acid, and nicotinamide in foods, urine, and plasma by means of Velasquez, S., Pena, N., Bohorquez, J. C., Gutierrez, N., & Sacks, G. L. (2019). Volatile and
LC-MS/MS and stable isotope dilution analysis. Journal of Agricultural and Food sensory characterization of roast coffees – effects of cherry maturity. Food Chemistry,
Chemistry, 56(23), 11114–11121. 274, 137–145.
Lindinger, C., Labbe, D., Pollien, P., Rytz, A., Juillerat, M. A., Yeretzian, C., & Blank, I. Wei, F. F., Furihata, K., Hu, F. Y., Miyakawa, T., & Tanokura, M. (2011). Two-dimensional
(2008). When machine tastes coffee: Instrumental approach to predict the sensory H-1-C-13 nuclear magnetic resonance (NMR)-based comprehensive analysis of
profile of espresso coffee. Analytical Chemistry, 80(5), 1574–1581. roasted coffee bean extract. Journal of Agricultural and Food Chemistry, 59(17),
Liu, C. J., Yang, N., Yang, Q., Ayed, C., Linforth, R., & Fisk, I. D. (2019). Enhancing 9065–9073.
Robusta coffee aroma by modifying flavour precursors in the green coffee bean. Food Wold S, Johansson A, & (eds)., C. M. (1993). PLS-partial least squares projections to latent
Chemistry, 281, 8–17. structures. ESCOM Science Publishers: Leiden, 523–550.
Locas, C. P., & Yaylayan, V. A. (2004). Origin and mechanistic pathways of formation of Worku, M., Duchateau, L., & Boeckx, P. (2016). Reproducibility of coffee quality cupping
the parent furan – A food toxicant. Journal of Agricultural and Food Chemistry, 52(22), scores delivered by cupping centers in Ethiopia. Journal of Sensory Studies, 31(5),
6830–6836. 423–429.
Moeenfard, M., Rocha, L., & Alves, A. (2014). Quantification of caffeoylquinic acids in Zou, T. T., Liu, J. B., Song, H. L., & Liu, Y. (2018). Discovery of amadori-type conjugates
coffee brews by HPLC-DAD. Journal of Analytical Methods in Chemistry, 2014, 965353. in a peptide maillard reaction and their corresponding influence on the formation of
Moon, J. K., & Shibamoto, T. (2010). Formation of volatile chemicals from thermal pyrazines. Journal of Food Science, 83(6), 1588–1595.

10