HPLC-Based Chemometric Analysis for Coffee Adulteration

Article HPLC-Based Chemometric Analysis for Coﬀee Adulteration

Wai Lok Cheah and Mingchih Fang *

Department of Food Science, College of Life Science, National Taiwan Ocean University, No 2, Beining Rd., Keelung City 20224, Taiwan; [email protected] * Correspondence: [email protected]; Tel.: +88-62-2462-2192 (ext. 5123)

Received: 2 June 2020; Accepted: 30 June 2020; Published: 4 July 2020

Abstract: Coffee is one of the top ten most adulterated foods. Coffee adulterations are mainly performed by mixing other low-value materials into coffee beans after roasting and grinding, such as spent coffee grounds, maize, soybeans and other grain products. The detection of adulterated coffee by high performance liquid chromatography (HPLC) is recognized as a targeted analytical method, which carbohydrates and other phenolic compounds are usually used as markers. However, the accurate qualitation and quantitation of HPLC analyses are time consuming. This study developed a chemometric analysis or called non-targeted analysis for coffee adulteration. The HPLC chromatograms were obtained by direct injection of liquid coffee into HPLC without sample preparation and the identification of target analytes. The distinction between coffee and adulterated coffee was achieved by statistical method. The HPLC-based chemometric provided more characteristic information (separated compounds) compared to photospectroscopy chemometric which only provide information of functional groups. In this study, green Arabica coffee beans, soybeans and green mung beans were roasted in industrial coffee bean roaster and then ground. Spent coffee ground was dried. Coffee and adulterants were mixed at different ratio before conducting HPLC analysis. Principal component analysis (PCA) toward HPLC data (retention time and peak intensity) was able to separate coffee from adulterated coffee. The detection limit of this method was 5%. Two models were built based on PCA data as well. The first model was used to differentiate coffee sample from adulterated coffee. The second model was designed to identify the specific adulterants mixed in the adulterated coffee. Various parameters such as sensitivity (SE), specificity (SP), reliability rate (RLR), positive likelihood (+LR) and negative likelihood ( LR) were applied to evaluate the performances − of the designed models. The results showed that PCA-based models were able to discriminate pure coffee from adulterated sample (coffee beans adulterated with 5%–60% of soybeans, green mung beans or spent coffee grounds). The SE, SP, RLR, +LR and LR for the first model were 0.875, 0.938, − 0.813, 14.1 and 0.133, respectively. In the second model, it can correctly distinguish the adulterated coffee from the pure coffee. However, it had only about a 30% chance to correctly determine the specific adulterant out of three designed adulterants mixed into coffee. The SE, RLR and LR were − 0.333, 0.333 and 0.667, respectively, for the second model. Therefore, HPLC-based chemometric analysis was able to detect coffee adulteration. It was very reliable on the discrimination of coffee from adulterated coffee. However, it may need more work to tell discern which kind adulterant in the adulterated coffee.

Keywords: coﬀee; adulteration; chemometrics; PCA; food fraud

Foods 2020, 9, 880; doi:10.3390/foods9070880 www.mdpi.com/journal/foods Foods 2020, 9, 880 2 of 11

1. Introduction The coffee bean is one of the most widely traded agricultural commodities and its consumption has increased rapidly [1]. Many gourmets say what you choose to drink in the morning tells what kind of person you are. The methods of coffee roasting, brewing and drinking are currently hyped up to an art in certain specialty coffees. Animal-passed coffee beans may be one of the most poetic examples. Based on the Database of United States Pharmacopeia Convention (USP) [2], coffee is the top ten adulterated food products due to its high commercial value and the shortage of coffee beans worldwide. Coffee adulteration may be performed by changing the quality of beans or adding other low-cost coffee and non-coffee materials, such as spent coffee grounds, corn, barley, maize, soybeans and other grains which bear a resemblance to coffee beans in term of color, particle size and texture [3,4]. Roasted soybeans have been found to be a good adulterant of coffee, because their color, flavor and aroma are similar to coffee. The making process has even be patented [5]. The contamination of low-priced robusta species in claimed 100% Arabica coffees is another means of coffee fraud. There are also cases of coffee products being recalled due to presence of undeclared drug contents, namely, sildenafil and tadalafil [6]. In order to protect consumers and ensure food safety, the authenticity assessment of coffee products is a positive way of approaching the problem of coffee fraud. Several analytical methods have been proposed to identify the adulteration of ground roasted coffee, including gas chromatography-mass spectrometry (GC-MS) [3], high-performance liquid chromatography (HPLC) [7–9], ultraviolet-visible spectrophotometry (UV-Vis) [10], ultra-performance liquid chromatography (UPLC) [11], Fourier-transform infrared spectroscopy (FTIR) [12] and voltammetric electronic tongue [13]. The above methods apply either targeted- or non-targeted analysis methods. Targeted analysis methods are broadly used to detect food contaminations—as well adulterations—which focus on detecting one or more compounds in a specimen. Specific targeted methods are generally complicated but can detect analytes in part per trillion (ppt) levels in complex matrices [14]. However, most of the adulterations are unknown additive compounds, so the targeted analysis is not always effective. Non-targeted methods such as chemometric analyses are the combination of emerging analytical methods and statistic software to detect food adulterations [15]. Non-targeted analysis treats food data as fingerprints. These indicate authenticity and provide early warning information for adulterated food. Therefore, the development of an HPLC-based non-targeted analysis to detect coffee adulteration is paramount. In a review of the literature, detecting coffee adulteration by using liquid chromatography belongs to a target analysis method that analyzes compounds (carbohydrates) in coffee and adulterated coffee to accomplish the distinction [7–9,11]. In contrast, spectroscopy methods have evolved as non-target analytical methods to detect coffee adulteration. For example, one study used ultraviolet-visible (UV-Vis) spectroscopy and successive projection algorithms (SPA) to construct a linear discriminant analysis (LDA) model for the identification of adulterants in ground roasted coffee [10]. In many coffee researches, coffee beans were roasted in convection ovens in laboratories to the desired color and intensity. The heating/roasting effect of a conventional oven was slow and inhomogeneous. Hence, the roasting time was long—40 min on average for a batch [12,16], while a commercial coffee bean roaster only requires 12–15 min. In order to simulate the commercial roasting method and enhance the representativeness of the sampling, industrial roaster was used in the study. In addition, many chemometric spectroscopy methods for coffee adulteration have been conducted on the powdered form of coffee [12], or extracted by solvents such as CDCl3 for nuclear magnetic resonance (NMR) analysis [17]. However, coffee is normally consumed in liquid form, and hot water is the only solvent. In this study, commercial roasting and brewing methods were applied for coffee sample preparation. Non-targeted analysis, coupled with HPLC was developed for the discrimination among roasted coffee, common adulterants and mixtures of coffee and adulterants. The statistical method used in this research was principal component analysis (PCA). Two models were developed in order to evaluate the ability of the method to detect coffee adulteration, as well as to identify the adulterants in the adulterated sample. Foods 2020, 9, x FOR PEER REVIEW 3 of 11 Foods 2020, 9, 880 3 of 11 2. Materials and Methods 2. Materials and Methods 2.1. Coffee and Adulterants 2.1. Co ee and Adulterants Greenff Arabica coffee beans (Chanchamayo, Peru) were purchased from Ho Hsin Beans Trading Co., GreenLtd (Taichung, Arabica co ffTaiwan).ee beans (Chanchamayo,Soybeans and Peru)green weremung purchased beans were from obtained Ho Hsin Beansfrom Tradinga local Co.,supermarket. Ltd (Taichung, Spent Taiwan). coffee grounds Soybeans were and supplied green mung by beanschain 85 were °C obtainedBakery Cafe from shop a local in supermarket.Taiwan and Spentkept frozen coffee (− grounds12 °C) until were use. supplied by chain 85 ◦C Bakery Cafe shop in Taiwan and kept frozen ( 12 ◦C) until use. 2.2.− Roasting 2.2. Roasting Spent coffee grounds were defrosted at room temperature for 18 h, and then dried in a convectionSpent co ovenffee grounds(Model wereDK-500DT, defrosted Bioman, at room New temperature Taipei, Taiwan) for 18 h, at and 100 then °C driedfor 5 inh. a In convection order to ovensimulate (Model the roasting DK-500DT, methods Bioman, of commercial New Taipei, coffee Taiwan) in this at 100 study,◦C foran 5industrial-grade h. In order to simulate800-N coffee the roastingbean roaster methods (Yang of Chia commercial Machine co Works,ffee in this Taichung, study, an Taiwan) industrial-grade was used. 800-N Coffee co beansffee bean (500 roaster g), soybeans (Yang Chia(500 g) Machine and green Works, mung Taichung, beans (500 Taiwan) g) were was each used. subjected Coffee beans to the (500 coffee g), soybeansbean roaster (500 at g) 180 and °C. green The mungtemperature beans (500went g) down were eachto the subjected turning topoint the coandffee went bean up roaster to 160 at 180°C ◦inC. 8 The min, temperature which was went the downdehydration to the turning process point of coffee and wentbeans up and to adulterants. 160 ◦C in 8 min, The whichfirst cracking was the sound dehydration happened process in the of cocoffeeffee beansbeans andat 11.5 adulterants. min and 195 The °C. first The cracking other materials sound happened did not crackle. in the co Coffeeffee beans beans, at 11.5soybeans min and and 195 green◦C. Themung other beans materials were removed did not crackle. to an air-cooling Coffee beans, chambe soybeansr when and the green temperature mung beans rose were to 210 removed °C, 230 to an°C air-coolingand 240 °C, chamber respectively. when theThe temperature roasting times rose towere 210 ◦alsoC, 230 different◦C and 240for ◦eachC, respectively. sample: 11–14 The roastingmin for timesroasting were coffee also beans, different 16–18 for min each for sample: roasting 11–14 the minsoybeans for roasting and 16–17 coff minee beans, for roasting 16–18 min the green for roasting mung thebeans. soybeans The roasting and 16–17 end-point min for was roasting determined the green by mung the appearance beans. The roastingin color end-pointto the desired was determinedbrown. The byroasting the appearance curves of coffee in color beans, to the soybeans desired and brown. green The mung roasting beans curves are shown of co inffee Figure beans, 1. soybeansSamples were and greengrinded mung by electric beans aregrinder shown (Model in Figure 600,1 .Yang Samples Chia were Machine grinded Works, by electric Taichung, grinder Taiwan) (Model after 600, roasting Yang Chiaand submitted Machine Works, to color Taichung, evaluation. Taiwan) after roasting and submitted to color evaluation.

Figure 1. Roasting curves of green Arabica cocoffeeﬀee beans,beans, soybeanssoybeans andand greengreen mungmung beans.beans.

2.3.2.3. Color Evaluation AA tristimulustristimulus colorimetercolorimeter (TC-1800 MK-II, Tokyo DenshokuDenshoku Co., Tokyo, Japan)Japan) waswas usedused toto measuremeasure thethe luminosityluminosity (CIELAB(CIELAB L*)L*) ofof thethe samples,samples, whichwhich waswas thethe mostmost relevantrelevant parameterparameter aboveabove a*a* andand b*b* forfor roastedroasted cocoffeeffee beansbeans [[18].18]. ItIt waswas alsoalso successfullysuccessfully employedemployed asas aa referencereference forfor roastingroasting degreedegree [[16].16]. Comparing the value of luminosity, thethe degreedegree ofof cocoffee-roastingffee-roasting cancan bebe classifiedclassified intointo threethree groups:groups: light light (23.5 (23.5 << L*L* < <25.0),25.0), medium medium (21.0 (21.0 < L*< L*< 23.5)< 23.5) and and dark dark (19.0 (19.0 < L*<

Foods 2020, 9, 880 4 of 11

2.4. Experimental Design Coffee beans, soybeans and green mung beans were each from three roasting batches (for three replicates). Spent coffee grounds were obtained from three different dates. Ten pure materials of each batch including coffee beans, soybeans, green mung beans and spent coffee grounds were prepared as pure samples—in total 4 10 3 = 120 samples (4 materials, 10 replicates, 3 batches). The adulterated × × samples were prepared by intentionally mixing coffee beans and adulterants at different ratios (coffee: adulterants, 95:5, 90:10, 80:20, 60:40, 40:60% w/w) in every batch (3 3 5 = 45, 3 batches, 3 adulterants, × × 5 concentrations). Cross-batch mixtures were also prepared (3 5 = 15, 3 adulterant cross-batches, × 5 concentrations). In total, 60 adulterated samples were prepared. All 180 samples were submitted to HPLC analysis.

2.5. Chromatograms Acquisition of Brewed Coffee and Adulterated Coffee A Shimadzu LC-2040C Plus (Shimadzu Co., Kyoto, Japan) equipped with an Agilent Zorbax SB-Phenyl C18 column (150 mm 4.6 mm, 5 µm) and a PDA (photodiode-array) detector was used × in the measurement. The mobile phases were pure water (solvent A) and methanol (solvent B). The gradient elution was operated as flowing: 0–4 min: 2% B, 4–8 min: 4%–10% B, 8–13 min: 10%–20% B, 13–20 min: 20%–35% B, 20–27 min: 35%–90% B, 27–27.5 min: 90%–2% B, 27.5–30 min: 2% B. 1 The injection volume and flow rate were 20 µL and 1 mL min− , respectively. The PDA was set at scan range between 200 and 800 nm; the chromatogram was monitored at 254 nm. Coffee and adulterated coffee samples were well-brewed in order to simulate commercially brewed coffee. The water temperature was controlled in the range of 91–94 ◦C. Two grams each of ground coffee beans and adulterated coffee samples were brewed with 25 mL water for 10 min. Each sample was centrifuged at 1000 g for 5 min and then filtered with No. 1 filter paper. The filtrate was passed × through a 0.2-µm PTFE filter and then transferred into a 1.5 mL vial for analysis.

2.6. Statistical Analysis SPSS version 22 (Statistical Product and Service Solutions, IBM Co., New York, NY, USA) was used to evaluate the color measurements between the samples of coffee and adulterated coffee in order to ensure uniformities among batches, adulterants and mixtures. An independent t-test was used in this analysis. Principle component analysis (PCA) with normalized HPLC chromatogram (retention time alignment of peaks) data was applied to distinguish adulterated coffee from roasted coffee samples. For PCA analysis, data metrices were constructed (each row corresponded to a sample; each column represented peak intensity at a given retention time) and handled by MATLAB software version 2018a (MathWorks, Natick, MA, USA) with the Classification Learner software application. Two models were built to us the PCA data. The first model was built with 180 samples (30 for pure coffee, 30 for pure soybeans, 30 for pure green mung beans, 30 for pure spent coffee grounds, and 60 for adulterated coffee) aimed on the authentication of coffee samples. The second model was built with 150 samples (180 samples from the first model, excluding 30 samples of pure coffee) to test the ability to identify the adulterants. In the model experiments, 60% of the samples were randomly selected as training set; the others were treated as the test set. The performances of the studied models were characterized by evaluating the quality of figure of merit (FOM). Figure of merit correspond to numeric parameters such as specificity (SP), sensitivity (SE) and reliability rate (RLR) [19]. The definition of the specificity (SP) was the ratio of true-negative samples (TN) to the sum of false-positive samples (FP) and the total number of known-negative samples (equation 1). It provided a measure of how well the model can predict samples of the class of controls. The sensitivity (SE) defined as the ratio of true positive samples (TP) to the sum of false-negative samples (FN) and the total number of known positive samples (equation 2). It provided a measure of Foods 2020, 9, 880 5 of 11 how well the model correctly identify samples of a given class [20]. The reliability rate (RLR) provided an overview of the trueness of the model and its definition was shown in equation 3 [12].

TN Speciﬁcity (SP) = (1) (TN + FP)

TP Sensitivity (SE) = (2) (TP + FN) FN FP RLR = 1 + = SP + SE 1 (3) − TP + FN TN + FP − In the first model, the TP was defined as the adulterated sample and was successfully detected by the system. The TN was the pure coffee sample or other pure adulterants and was successfully detected. The FP was the pure coffee sample but was misclassified as adulterated sample. The FN was the adulterated sample but was misclassified as pure coffee or other adulterants. In the second model, the TN was the pure adulterant and successfully detected. The FP was the pure adulterant and misclassified as adulterated coffee. TP was defined as the adulterated sample and the specific adulterant was successfully identified. The FN was the specific adulterant in the adulterated coffee which was classified as the other adulterant in the adulterated coffee or the pure adulterant. The positive likelihood ratio (+LR) and the negative likelihood ratio ( LR) were also applied to − evaluate the performance of the studied models. The definitions of +LR and LR are listed in equation − 4 and equation 5 [21]. A bigger value of +LR gave the stronger confidence of the positive result, while a smaller value of LR presented the higher value of the testing results [22]. − SE + LR = (4) 1 SP − 1 SE LR = − (5) − SP

3. Results and Discussion

3.1. Coffee Roasting and Color Measurement The color measurement was performed to evaluate whether the appearance of the roasted adulterants and the coffee were similar after roasting and grinding. Table1 shows the L* value of roasted coffee (ground) was significantly different from roasted soybeans, green mung beans and spent coffee grounds (p < 0.05) and spent coffee grounds was the darkest one (L* was the lowest 18.0). Similar studies of roasted coffees and adulterants in convection ovens with precisely controlled temperatures, such as 240 ◦C 11 min for coffee, 240 ◦C 30 min for corn and 250 ◦C 28 min for barley resulted in color similarity [16]. In this study, soybeans and green mung beans required higher roasting temperatures and time to reach the desired color. The industrial-grade coffee bean roaster continually increased the temperature over time during the roasting and it was not easy to control, compared to convection oven-roasting in which there is a constant temperature during the whole roasting procedure. We tried to roast the adulterants at higher temperatures and for longer times. In these overheated roasting conditions, a similar color was obtained, but the flavor was burnt. Therefore, we decided to roast the adulterants in coffee-bean way. The final temperature was slightly increased to 230 ◦C and 240 ◦C for soybeans and green mung beans, respectively. The maximum temperature of the commercial roaster for coffee was 240 ◦C. The roasting time for adulterants was extended to around 18 min as shown in Figure1. The condition resulted in good-smelling soybeans and green mung beans; the soybean especially smelled like coffee. Though the color of adulterants and coffee was different, there was no significant difference among all adulterated coffees in which coffee was mixed with 20% adulterants (p > 0.05). Foods 2020, 9, x FOR PEER REVIEW 6 of 11 Foods 2020, 9, 880 6 of 11 Table 1. L* values of coffee beans, adulterants and adulterated coffee beans.

Table 1. L* values of coffee beans, adulterantsLuminosity and Measurement adulterated coResultffee beans. Analyte L* (Mean ± SD, Luminosityp-Value (t-Test Measurement of Coffee and Result Other Analytes) Analyte n = 3) Coffee beans L* (Mean 20.8 ± 0.8SD, n = 3) p-Value (t-Test of Coffee and Other Analytes) ± CoffSoybeansee beans 27.5 20.8 ± 0.40.8 0.00 ± GreenSoybeans mung beans 30.5 27.5 ± 1.20.4 0.00 0.00 ± GreenSpent mung coffee beansgrounds 18.0 30.5 ± 0.21.2 0.00 0.00 ± SpentCoffee co beansffee grounds+ soybeans * 21.4 18.0 ± 0.70.2 0.10 0.00 ± CoffeeCoffee beans beans + green+ soybeans mung *beans * 21.5 21.4 ± 0.60.7 0.07 0.10 ± CoCoffeeffee beans beans+ +green spent mungcoffee beansgrounds * * 20.5 21.5 ± 0.40.6 0.28 0.07 ± Coffee beans + spent coffee grounds * 20.5 0.4 0.28 * 20% roasted adulterant± in coffee beans. * 20% roasted adulterant in coffee beans. 3.2. HPLC Chromatograms 3.2. HPLC Chromatograms Applications of HPLC for the detection of adulterated coffee were mainly based on targeted methods.Applications Analytes such of HPLC as oligosaccharides, for the detection monosaccharides, of adulterated co fftrigonelline,ee were mainly and/or based nicotinic on targeted acid weremethods. applied Analytes as markers such asto oligosaccharides,identify adulterants monosaccharides, in coffee [7–9,11]. trigonelline, In this study, and/or HPLC nicotinic with acid non- were targetedapplied analysis as markers was to developed identify adulterants to differentiate in coff coffeeee [7– 9adulteration.,11]. In this study, HPLC HPLC analysis with was non-targeted directly performedanalysis was after developed sample extraction to differentiate and the co generatedffee adulteration. chromatogram HPLC was analysis utilized was as directly a “fingerprint”. performed Figureafter sample2 shows extraction the HPLC and chromatograms the generated of chromatogram pure coffee, pure was adulterants utilized as aand “fingerprint”. adulterated Figurecoffee2 containingshows the 20% HPLC soybeans. chromatograms Coffee beans of pure (Figure coffee, 2A), pure soybeans adulterants (Figure and adulterated2B) and green co ffmungee containing beans (Figure20% soybeans. 2C) showed Co ffveryee beans distinct (Figure chromatograms2A), soybeans because (Figure each2B) material and green presented mung beanstheir own (Figure special2C) compounds,showed very such distinct as chromatogramscaffeine, tannic becauseacid, lino eachleic material acid, nicotinic presented acid, their chlorogenic own special compounds,acids and trigonellinesuch as caff eine,in coffee tannic [23]. acid, Soybeans linoleic acid, contained nicotinic isoflavones acid, chlorogenic [24]. acidsGreen and mung trigonelline beans incontained coffee [23 ]. polyphenols,Soybeans contained polysaccharides isoflavones and [ 24peptides]. Green [25]. mung However, beans contained the chromatograms polyphenols, of coffee, polysaccharides spent coffee and groundspeptides (Figure [25]. However, 2D) and adulterated the chromatograms coffee beans of co ff(Figureee, spent 2E) co showedffee grounds high similarities. (Figure2D) and Chemometric adulterated approachescoffee beans were (Figure employed2E) showed to discriminate high similarities. chromatograms Chemometric generated approachesfrom different were adulterants employed and to adulterateddiscriminate coffee chromatograms beans. generated from different adulterants and adulterated coffee beans.

FigureFigure 2. 2.HPLCHPLC chromatograms chromatograms (overlap (overlap of ofthree) three) of of(A ()A coffee) coffee beans, beans, (B ()B soybeans,) soybeans, (C ()C green) green mung mung beans,beans, (D (D) spent) spent coffee coffee grounds grounds and and (E ()E adulterated) adulterated coffee coffee beans beans containing containing 20% 20% soybeans. soybeans.

Foods 2020, 9, x FOR PEER REVIEW 7 of 11

3.3. Principal Component Analysis (PCA) The PCA-scores scatter plot derived from normalized HPLC chromatograms is shown in Figure 3. Pure coffee, soybeans, green mung beans and spent coffee grounds are well-separated. The coffee sample was more to the positive of PC1 and the others were more to negative of PC1. The successful separation among coffee, soybean and green mung bean was as expected, because each material contained different compounds and presented a specific chromatogram. The separation between coffee and spent coffee grounds was very good as well. Detailed elements from PCA analysis indicated that some small peaks appeared between 20–25 min in HPLC chromatograms of spent coffee grounds (Figure S1) contributed most to the discrimination between spent coffee grounds and coffee. The PCA values of adulterated coffees were disorderly around pure coffee. The 60% adulterated coffee obtained values close to zero in PC1 and close to adulterants as well. The various mixing ratios (5%, 10%, 20%, 40% and 60%) are displayed in different colors in Figure 3. The results indicated that it was possible to detect coffee adulteration with various adulterants in admixture as low as 5% (w/w). ThisFoods study2020, 9 ,was 880 able to use HPLC chromatograms for chemometric analysis. However, in Figure7 2, of it 11 is clear that overlapped HPLC chromatograms did not perfectly match each other. In many chemometric applications, the spectra—especially the most adopt FTIR spectra—are almost identical 3.3. Principal Component Analysis (PCA) to each other in the same test group. The uses of HPLC chromatograms as fingerprints for chemometricThe PCA-scores analysis scatteris powerful plot derived because from HPLC normalized chromatograms HPLC chromatogramsprovide more distinct is shown and in detailed Figure3 . information,Pure coffee, due soybeans, to the peaks green represented mung beans to anddifferent spent compounds. coffee grounds While are us well-separated.ing FTIR for chemometric The coffee analysis,sample wasthe morepeaks to only the positiverepresent of different PC1 and functi the othersonal weregroups. more Figure to negative 4 shows of the PC1. FTIR The spectra successful of roastedseparation coffee among beans coandffee, soybeans. soybean They and were green simila mungr in bean certain was degree; as expected, compared because to Figure each 2A,B material the chromatogramscontained different of compoundscoffee and andsoybean presented were avery specific different. chromatogram. However, The the separation system betweenstability coandffee chromatogramand spent coffee reproducibility grounds was veryof HPLC good was as well. not as Detailed good as elements spectrophotometric from PCA analysis method indicated (e.g., FTIR). that Insome this smallstudy, peaks the appeareddetection betweenlimit of 20–25adulterants min in in HPLC coffee chromatograms was set at 5%. of Compared spent coffee to grounds many chemometric(Figure S1) contributed analyses with most IR to spectra, the discrimination the detection between limit was spent below coff 1%ee grounds[26]. and coffee.

FigureFigure 3. 3. PCA-scoresPCA-scores scatter scatter plot plot of of coffee coffee beans beans in in compar comparisonison to to soybeans, soybeans, green green mung mung beans, beans, spent spent coffeecoffee grounds grounds and and adulterated adulterated coffee coffee containing containing differe differentnt adulterants adulterants in in differe differentnt mixing mixing ratios ratios (5%, (5%, 10%,10%, 20%, 20%, 40% 40% and and 60%). 60%).

The PCA values of adulterated coffees were disorderly around pure coffee. The 60% adulterated coffee obtained values close to zero in PC1 and close to adulterants as well. The various mixing ratios (5%, 10%, 20%, 40% and 60%) are displayed in different colors in Figure3. The results indicated that it was possible to detect coffee adulteration with various adulterants in admixture as low as 5% (w/w). This study was able to use HPLC chromatograms for chemometric analysis. However, in Figure2, it is clear that overlapped HPLC chromatograms did not perfectly match each other. In many chemometric applications, the spectra—especially the most adopt FTIR spectra—are almost identical to each other in the same test group. The uses of HPLC chromatograms as fingerprints for chemometric analysis is powerful because HPLC chromatograms provide more distinct and detailed information, due to the peaks represented to different compounds. While using FTIR for chemometric analysis, the peaks only represent different functional groups. Figure4 shows the FTIR spectra of roasted co ffee beans and soybeans. They were similar in certain degree; compared to Figure2A,B the chromatograms of co ffee and soybean were very different. However, the system stability and chromatogram reproducibility of HPLC was not as good as spectrophotometric method (e.g., FTIR). In this study, the detection limit of adulterants in coffee was set at 5%. Compared to many chemometric analyses with IR spectra, the detection limit was below 1% [26]. Foods 2020, 9, x FOR PEER REVIEW 8 of 11 Foods 2020, 9, 880 8 of 11 Foods 2020, 9, x FOR PEER REVIEW 8 of 11

Figure 4. FTIR (ATR, Ge) spectra of roasted coffee beans and soybeans. Figure 4. FTIR (ATR, Ge) spectra of roasted cocoffeeffee beansbeans andand soybeans.soybeans. 3.4. Discriminatory Power of Models Built 3.4. Discriminatory Power of Models Built 3.4. DiscriminatoryTwo models Powerwere built of Models in this Built study. The first model was designed to distinguish pure coffee formTwo adulterated models werecoffee. built The in second this study. model The was first design modeled to was identify designed the adulterants to distinguish that pure exist co inff theee Two models were built in this study. The first model was designed to distinguish pure coffee formadulterated adulterated coffee. co ffAee. total The of second 180 sample model sets was were designed divided to identifyinto two thegroups adulterants including that a existtraining in the set form adulterated coffee. The second model was designed to identify the adulterants that exist in the adulterated(60% data) and coff ee.a test A totalset (40% of 180 of data). sample We sets used were the divided training into data two set groupsto calculate including the parameters a training and set adulterated coffee. A total of 180 sample sets were divided into two groups including a training set (60%to generate data) and the a models test set for (40% the of first data). and We second used themodels. training The data test setsetto was calculate used to the test parameters the classification and to (60% data) and a test set (40% of data). We used the training data set to calculate the parameters and generateresults of the the models models. for A theconfusion first and matrix second was models. used to Theillustrate test set the was test used results. to test the classification to generate the models for the first and second models. The test set was used to test the classification resultsIn of the the first models. model, A the confusion classesmatrix 1–5 were was coffee, used tosoybean, illustrate green the testmung results. bean, spent coffee ground results of the models. A confusion matrix was used to illustrate the test results. and Inadulterated the first model, coffee. the Of classes 18 pure 1–5 were coffee coff samplesee, soybean, (true green class mung 1), 2 bean, samples spent were coffee classified ground and as In the first model, the classes 1–5 were coffee, soybean, green mung bean, spent coffee ground adulteratedadulterated cocoffee.ffee. OfOf 1836 pureadulterated coffee samples coffees (true class 1),5), 22 samplessamples werewere classifiedclassified asas adulteratedpure coffee. and adulterated coffee. Of 18 pure coffee samples (true class 1), 2 samples were classified as coTheffee. training Of 36 adulteratedset contained co 2ff false-positiveees (true class (FP) 5), 2samples samples (classifying were classified pure ascoffee pure as co adulteratedffee. The training coffee), adulterated coffee. Of 36 adulterated coffees (true class 5), 2 samples were classified as pure coffee. set70 containedtrue negative 2 false-positive (TN) samples (FP) (classifying samples (classifyingunadulterated pure sample coffee correctly), as adulterated 2 false-negative coffee), 70 true(FN) The training set contained 2 false-positive (FP) samples (classifying pure coffee as adulterated coffee), negativesamples (TN)(classifying samples adulterated (classifying coffee unadulterated as pure coffee) sample andcorrectly), 34 true positive 2 false-negative (TP) samples (FN) (classifying samples 70 true negative (TN) samples (classifying unadulterated sample correctly), 2 false-negative (FN) (classifyingadulterated adulterated coffee correctly) coffee (Figure as pure 5A). coffee) These and para 34 truemeters positive were (TP) used samples to calculate (classifying the discriminatory adulterated samples (classifying adulterated coffee as pure coffee) and 34 true positive (TP) samples (classifying copowerffee correctly) and are summarized (Figure5A). These in Table parameters 2. Then, weredata from used tothe calculate test set were the discriminatory applied to the power established and adulterated coffee correctly) (Figure 5A). These parameters were used to calculate the discriminatory arefirst summarized model. The results in Table are2. shown Then, data in Figure from 5B. the Three test set FP were and three applied FN to were the obtained. established The first first model. model power and are summarized in Table 2. Then, data from the test set were applied to the established Thewas results able to are correctly shown indistinguish Figure5B. Three21 adulterated FP and three samples FN were from obtained. 24 adulterated The first coffees. model wasHowever, able to 3 first model. The results are shown in Figure 5B. Three FP and three FN were obtained. The first model correctlypure coffees distinguish were recognized 21 adulterated as adulterated samples from coffe 24es, adulterated and 3 adulterated coffees. However, coffees were 3 pure recognized coffees were as was able to correctly distinguish 21 adulterated samples from 24 adulterated coffees. However, 3 recognizedpure coffees. as adulterated coffees, and 3 adulterated coffees were recognized as pure coffees. pure coffees were recognized as adulterated coffees, and 3 adulterated coffees were recognized as pure coffees.

FigureFigure 5. 5.Confusion Confusion matrixmatrix ofof thethe firstfirst modelmodel thatthat usedused toto discriminatediscriminate cocoffeeffeefrom fromadulterated adulterated co coffeeffee forfor ( A)) training setset andand (BB)) test test set set (class (class 1—coffee; 1—coffee; class 2—soybeans; classclass 3—green3—green mungmung beans;beans; class class Figure 5. Confusion matrix of the first model that used to discriminate coffee from adulterated coffee 4—spent4—spent co coffeeffee grounds; grounds; class class 5—adulterated 5—adulterated co coffeeffee beans). beans); for A) training set and B) test set (class 1—coffee; class 2—soybeans; class 3—green mung beans; class 4—spent coffee grounds; class 5—adulterated coffee beans);

Foods 2020, 9, 880 9 of 11

Table 2. Estimated ﬁgure of merit (FOM) of the ﬁrst and second model.

SP SE RLR +LR LR − First model Training set 0.972 0.944 0.916 33.7 0.0576 Test set 0.938 0.875 0.813 14.1 0.133 Second model Training set 1.00 0.583 0.583 – 0.417 Test set 1.00 0.333 0.333 – 0.667 SP—speciﬁcity; SE—sensitivity; RLR—reliability rate; +LR—positive ratio likelihood; LR—negative − Foodsratio 2020, likelihood. 9, x FOR PEER REVIEW 9 of 11

TheThe secondsecond modelmodel waswas designeddesigned forfor distinguishingdistinguishing whichwhich kindkind ofof adulterantsadulterants werewere inin thethe adulteratedadulterated cocoffees.ffees. TheThe resultsresults areare shownshown inin FigureFigure6 .6. In In the the training training set, set, the the pure pure soybeans, soybeans, green green mungmung beansbeans andand spentspent cocoffeeffee groundsgrounds werewere correctlycorrectly classified,classified, whilewhile thethe adulteratedadulterated cocoffeesffees werewere partiallypartially classified.classified. For For example, example, in in12 12soybean-adulterated soybean-adulterated coffees coff ees(class (class 4), 2 4),were 2 wereclassified classified as green- as green-mung-bean-adulteratedmung-bean-adulterated and one and was one cl wasassified classified as spent-coff as spent-coee-grounds-adulteratedffee-grounds-adulterated (Figure (Figure 6A). In6 A).the Intest the set test of set the of second the second model, model, all all36 36pure pure adulterants adulterants (12 (12 for for each) each) were were distinguished thethe fromfrom adulteratedadulterated cocoffee.ffee. However, However, the the second second model model was was no nott able able to to identify identify the the adulterants adulterants correctly correctly in inthe the adulterated adulterated coffees. coffees. For example, For example, in 8 green-m in 8 green-mung-bean-adulteratedung-bean-adulterated coffees co(Figureffees (Figure6B, class6B, 5), classonly 5),2 were only correctly 2 were correctly identified identified as green-mung-be as green-mung-bean-adulterated,an-adulterated, one was onemisclassified was misclassified as soybean- as soybean-adulteratedadulterated and 5 were and classified 5 were classified as spent-coffee-gr as spent-coounds-adulterated.ffee-grounds-adulterated. In the second In the secondmodel, model,the FN thesamples FN samples were many were and many had anda high had chance a high of chancemisjudging of misjudging the adulterants the adulterants existed in the existed adulterated in the adulteratedcoffee. coffee.

FigureFigure 6.6. ConfusionConfusion matrix ofof secondsecond modelmodel thatthat usedused toto identifyidentify thethe adulterantsadulterants existedexisted inin thethe adulteratedadulterated cocoffeeffee forfor ((AA)) training training setset andand ((BB)) testtest setset (class(class 1—soybeans;1—soybeans; classclass 2—green2—green mungmung beans;beans; classclass 3—spent3—spent cocoffeeffee grounds;grounds;class class4—adulterated 4—adulterated cocoffeeffee containing containing soybeans; soybeans; classclass 5—adulterated 5—adulterated cocoffeeffee containing containing green green mung mung beans; beans; class class 6—adulterated 6—adulterated co coffeeffee containing containing spent spent co coffeeffee grounds). grounds);

FigureFigure of merit merit (FOM) (FOM) was was used used to characterize to characterize the performances the performances of the ofstudied the studied models. models.Statistic Statistic equations were applied (Equation 1–5) for the calculation of SP, SE, RLR, +LR and LR. equations were applied (Equation 1–5) for the calculation of SP, SE, RLR, +LR and −LR. In first model,− InSP first (0.938) model, and SPSE (0.938)(0.875) andof the SE test (0.875) set were of the sati testsfied. set The were reliability satisfied. rate The (RLR) reliability was 0.813 rate (RLR)proved was the 0.813trueness proved of first the model trueness [12]. of Positive first model likelihood [12]. Positive ratio (+LR) likelihood was used ratio to eval (+LR)uate was whether used tothe evaluate positive whetherresult was the correctly positive tested. result The was high correctly +LR value tested. (>10 The) strongly high + LRindicated value (the>10) positive strongly result indicated tested was the positivecorrect [22]. result The tested obtained was correct +LR of [ 22the]. first The obtainedmodel were+LR 33.7 of theand first 14.1 model for training were 33.7 set andand 14.1test forset, training set and test set, respectively (Table2). On the other hand, the LR values were low (0.0576 for respectively (Table 2). On the other hand, the −LR values were low (0.0576− for training set and 0.133 trainingfor test set) set andindicated 0.133 there for test was set) strongly indicated evidence there was to prove strongly the TN evidence result towa proves detected the TN correctly result was[22]. detectedThus, the correctly performance [22]. Thus, of first the performancemodel was ofgood. first modelIn the was second good. model, In the secondlower model,SE values lower were SE observed—0.583 and 0.333 for the training and test sets, respectively. The lower SE values indicated that the ability of the model to detect the positive sample was low. This model performed 100% in prediction of TN samples and 0 case of FP sample. The obtained SPs were 1.00 for either training or test set. Therefore, the RLRs of the second model were same as the SEs. However, the second model’s SP value was satisfied (means the second model detected the pure adulterants correctly), but the aim of this model was to identify the adulterant mixed in the adulterated coffee (positive sample). Therefore, the SE value was more important and relevant to the ability of the model. On the other hand, the −LR values were 0.417 and 0.667 for training set and test set, respectively, which represented no evidence to prove that the true negative result was correct [22]—even if the SP value was 1.0. This indicated that the −LR was better than the SP for explaining the performance of the model. Thus, the second model was only suitable for the discrimination of pure coffee and

Foods 2020, 9, 880 10 of 11 values were observed—0.583 and 0.333 for the training and test sets, respectively. The lower SE values indicated that the ability of the model to detect the positive sample was low. This model performed 100% in prediction of TN samples and 0 case of FP sample. The obtained SPs were 1.00 for either training or test set. Therefore, the RLRs of the second model were same as the SEs. However, the second model’s SP value was satisfied (means the second model detected the pure adulterants correctly), but the aim of this model was to identify the adulterant mixed in the adulterated coffee (positive sample). Therefore, the SE value was more important and relevant to the ability of the model. On the other hand, the LR values were 0.417 and 0.667 for training set and test set, respectively, which − represented no evidence to prove that the true negative result was correct [22]—even if the SP value was 1.0. This indicated that the LR was better than the SP for explaining the performance of the − model. Thus, the second model was only suitable for the discrimination of pure coffee and adulterated coffee, while the ability of identifying adulterants existed in the adulterated coffee was low.

4. Conclusions HPLC-based chemometric analysis was performed and able to distinguish coffee from adulterated coffee with detection limit around 5%. This study simulated commercial roasting conditions in sample preparation and liquid sample was obtained by coffee brewing. The chemometric analysis applied chromatograms (information of compounds) instead of photospectra (information of functional groups) studied in most chemometric analysis. The experimental model was able to separate pure coffee and adulterated coffees. However, the performance of telling what kind of adulterants existed in the adulterated coffee still needs to be improved.

Supplementary Materials: The following are available online at http://www.mdpi.com/2304-8158/9/7/880/s1, Figure S1: HPLC chromatogram of coffee and spent coffee grounds sample for 20–25 min of retention time. Author Contributions: Investigation, W.L.C.; methodology, W.L.C.; project administration, M.F.; writing—original draft, W.L.C.; writing—review & editing, M.F. All authors have read and agreed to the published version of the manuscript. Funding: This research was funded by MOST, Grant Number 107–2320-B-019–001. Conflicts of Interest: The authors declare no conflicts of interest.

References

1. International Coffee Organization. Coffee Market Report: November 2019. Available online: http: //www.ico.org/documents/cy2019-20/cmr-1119-e.pdf (accessed on 9 February 2020). 2. Song, H.Y.; Jang, H.W.; Debnath, T.; Lee, K.G. Food Chemicals Codex, 8th ed.; United States Pharmacopeia: Rockville, MD, USA, 2013. 3. Oliveira, R.C.; Oliveira, L.S.; Franca, A.S.; Augusti, R. Evaluation of the potential of SPME-GC-MS and chemometrics to detect adulteration of ground roasted coffee with roasted barley. J. Food Compos. Anal. 2009, 22, 257–261. [CrossRef] 4. Toci, A.T.; Farah, A.; Pezza, H.R.; Pezza, L. Coffee Adulteration: More than Two Decades of Research. Crit. Rev. Anal. Chem. 2015, 46, 83–92. [CrossRef][PubMed] 5. Hullah, W.; Trail, A.W.; Mills, D.; Cringle, J.; Drive, P.; Albrecht, S.; Ave, M. Instant Coffee Substitute from Soybeans and Method of Making. U.S. Patent H673, 1989. 6. FDA Recalls. Market Withdrawals, & Safety Alerts. 2019. Available online: https: //www.fda.gov/safety/recalls-market-withdrawals-safety-alerts/brian-richardson-dba-tha-pink-issues- voluntary-nationwide-recall-kopi-jantan-tradisional-natural (accessed on 19 March 2020). 7. Domingues, D.S.; Pauli, E.D.; de Abreu, J.E.; Massura, F.W.; Cristiano, V.; Santos, M.J.; Nixdorf, S.L. Detection of roasted and ground coffee adulteration by HPLC by amperometric and by post-column derivatization UV–Vis detection. Food Chem. 2014, 146, 353–362. [CrossRef] 8. Pauli, E.D.; Barbieri, F.; Garcia, P.S.; Madeira, T.B.; Acquaro, V.R.; Scarminio, I.S.; Da Camara, C.A.P.; Nixdorf, S.L. Detection of ground roasted coffee adulteration with roasted soybean and wheat. Food Res. Int. 2014, 61, 112–119. [CrossRef] Foods 2020, 9, 880 11 of 11

9. Song, H.Y.; Jang, H.W.; Debnath, T.; Lee, K.-G. Analytical method to detect adulteration of ground roasted coffee. Int. J. Food Sci. Technol. 2018, 54, 256–262. [CrossRef] 10. Barbosa, M.F.; Souto, U.T.D.C.P.; Dantas, H.V.; de Pontes, A.S.; Lyra, W.D.S.; Diniz, P.H.G.D.; Araujo, M.C.U.; Da Silva, E.C. Identification of adulteration in ground roasted coffees using UV–Vis spectroscopy and SPA-LDA. LWT 2015, 63, 1037–1041. [CrossRef] 11. Cai, T.; Ting, H.; Jin-Lan, Z. Novel identification strategy for ground coffee adulteration based on UPLC–HRMS oligosaccharide profiling. Food Chem. 2016, 190, 1046–1049. [CrossRef] 12. Reis, N.; Botelho, B.; Franca, A.S.; Oliveira, L.S. Simultaneous Detection of Multiple Adulterants in Ground Roasted Coffee by ATR-FTIR Spectroscopy and Data Fusion. Food Anal. Methods 2017, 10, 2700–2709. [CrossRef] 13. De Morais, T.C.B.; Rodrigues, D.R.; Souto, U.T.D.C.P.; Lemos, S. A simple voltammetric electronic tongue for the analysis of coffee adulterations. Food Chem. 2019, 273, 31–38. [CrossRef] 14. Kaufmann, A.; Butcher, P.; Maden, K.; Walker, S.; Widmer, M. Reliability of veterinary drug residue confirmation: High resolution mass spectrometry versus tandem mass spectrometry. Anal. Chim. Acta 2015, 856, 54–67. [CrossRef] 15. Kaufmann, A. Combining UHPLC and high-resolution MS: A viable approach for the analysis of complex samples? TrAC Trends Anal. Chem. 2014, 63, 113–128. [CrossRef] 16. Reis, N.; Franca, A.S.; Oliveira, L.S. Performance of diffuse reflectance infrared Fourier transform spectroscopy and chemometrics for detection of multiple adulterants in roasted and ground coffee. LWT 2013, 53, 395–401. [CrossRef] 17. Okaru, A.O.; Scharinger, A.; de Rezende, T.R.; Teipel, J.C.; Kuballa, T.; Walch, S.G.; Lachenmeier, D.W. Validation of a Quantitative Proton Nuclear Magnetic Resonance Spectroscopic Screening Method for Coffee Quality and Authenticity (NMR Coffee Screener). Foods 2020, 9, 47. [CrossRef][PubMed] 18. Mendonça, J.C.; Franca, A.S.; Oliveira, L.S. Physical characterization of non-defective and defective Arabica and Robusta coffees before and after roasting. J. Food Eng. 2009, 92, 474–479. [CrossRef] 19. Olivieri, A.C. Analytical Figures of Merit: From Univariate to Multiway Calibration. Chem. Rev. 2014, 114, 5358–5378. [CrossRef] 20. Szymanska, E.; Saccenti, E.; Smilde, A.K.; Westerhuis, J.A. Double-check: Validation of diagnostic statistics for PLS-DA models in metabolomics studies. Metabolomics 2011, 8, 3–16. [CrossRef] 21. Baratloo, A.; Safari, S.; Elfil, M.; Negida, A. Evidence Based Emergency Medicine Part 3: Positive and Negative Likelihood Ratios of Diagnostic Tests. Emergency 2015, 3, 170–171. 22. Sloane, P.D.; Slatt, L.M.; Ebell, M.H.; Jacques, L.B.; Smith, M.A. Essentials of Family Medicine, 5th ed.; Lippincott Williams & Wilkins: Baltimore, MD, USA, 2008; p. 44. 23. Patay-Éva, B.; Bencsik, T.; Papp, N. Phytochemical overview and medicinal importance of Coffea species from the past until now. Asian Pac. J. Trop. Med. 2016, 9, 1127–1135. [CrossRef] 24. Ciabotti, S.; Silva, A.C.B.B.; Juhasz, A.C.P.; Mendonça, C.D.; Tavano, O.L.; Mandarino, J.M.G.; Gonçalves, C.A.A. Chemical composition, protein profile, and isoflavones content in soybean genotypes with different seed coat colors. Int. Food Res. J. 2016, 23, 621–629. 25. Hou, D.; Yousaf, L.; Xue, Y.; Hu, J.; Wu, J.; Hu, X.; Feng, N.; Shen, Q. Mung Bean (Vigna radiata L.): Bioactive Polyphenols, Polysaccharides, Peptides, and Health Benefits. Nutrients 2019, 11, 1238. [CrossRef] 26. Zou, H.; Zhang, W.; Feng, Y.; Liang, B. Simultaneous determination of melamine and dicyandiamide in milk by UV spectroscopy coupled with chemometrics. Anal. Methods 2014, 6, 5865–5871. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).