Enhancing the Understanding of Aroma Formation during Coffee Roasting Using DHS-GC×GC-TOFMS and Chemometrics

Abstract

This study integrates advanced chromatographic techniques and chemometric analysis to deepen the understanding of volatile and semivolatile compound formation during coffee roasting. Dynamic headspace extraction (DHS) was used to capture these compounds at three distinct roasting stages. The volatiles were analyzed and identified using comprehensive two-dimensional gas chromatography coupled with time-of-flight mass spectrometry (GC×GC-TOFMS), a high-resolving power technique to reliably annotate the numerous features present in the aroma-profile. Partial least-squares discriminant analysis (PLS-DA) was used to handle the large and complex data generated by GC×GC-TOFMS. This method successfully classified samples based on the roasting stages. Evaluation of the loadings using a variable importance in projection (VIP) approach revealed that 99 aroma-related volatiles varied significantly during coffee roasting, including carboxylic acids (e.g., acetic acid) and pyrazines such as 2-ethenyl-5-methyl pyrazine and ethyl pyrazine. These compounds exhibited distinct kinetic profiles, providing a detailed understanding of the chemical transformations occurring during roasting. This research describes an important methodology for monitoring aroma-related volatiles that may be useful for personalized coffee roasting, delivering key insights into the complex development of coffee aroma.

1. Introduction

Coffee is celebrated not only for its stimulating properties but also for its captivating aroma, which develops during the roasting process. , Research into coffee’s aroma dates to the early 20th century, when researchers began identifying over 29 volatile compounds in roasted coffee. , Since then, advancements in gas chromatography (GC) and mass spectrometry (MS) have revealed that both green and roasted coffee contain over 1,000 volatile and semivolatile organic compounds (VOCs and SVOCs), each contributing to its complex aroma profile. As the catalog of these compounds expands, researchers have shifted their focus to determining which of these are sensorially active and how they influence the aroma of roasted coffee. −

Home coffee roasters provide enthusiasts with the unique opportunity to personalize the aroma profile of their coffee by experimenting with different roasting techniques and parameters. Unlike prepackaged commercial coffee, which often offers a fixed flavor profile, roasting at home allows for precise control over key variables such as temperature, roasting time, and air flow, all of which significantly impact the development of aromatic compounds. By adjusting these factors, users can influence the formation of specific volatile and semivolatile organic compounds. For instance, a lighter roast might highlight floral and fruity notes due to the preservation of delicate compounds like linalool and aldehydes, while a darker roast can enhance the presence of chocolatey or nutty aromas through the increased formation of Maillard reaction products such as pyrazines and furans. Additionally, home roasters can experiment with different green coffee varieties, blending techniques, and roast profiles, enabling them to tailor the aroma and flavor to their exact preferences.

While much of the research has focused on roasted coffee, the composition of VOCs in raw coffee beans has also gained attention. Despite being present in lower concentrations, green coffee VOCs provide valuable insights for identifying potential off-flavors before roasting and predicting the final aroma quality. Roasting induces the formation of key odoriferous compounds, including aldehydes, pyrazines, phenols, and thiols, which are responsible for the distinctive aromas associated with coffee. The roasting process itself involves exposing raw coffee beans to high temperatures, triggering intricate chemical and physical changes such as moisture loss and density reduction.

The volatile profile of roasted coffee is intricate, with numerous VOCs and SVOCs that interact to create its characteristic aroma. Although extensive studies have been conducted on the chemical transformations during roasting and the design of roasting equipment, ,,− the field remains far from fully understood. Several variables influence the formation of these aroma compounds, including the time and temperature profiles used, the unique properties of the beans, and cultivation and processing techniques. Thus, the roasting process is often considered a craft or an art.

To develop a scientific foundation for optimizing coffee aroma, researchers have explored various analytical techniques to monitor the evolution of these compounds during roasting. Some studies have correlated specific VOCs concentrations with sensory analyses, while others have used model systems to simulate the chemical transformations occurring during roasting. , However, findings from these simplified models must be interpreted with caution, as they may not capture the complexity of reactions in real coffee beans. A more elucidative approach to study the transformations occurring in raw coffee during roasting can be achieved by analyzing the volatile plume released throughout the actual roasting process.

Traditional extraction techniques like solid-phase microextraction (SPME) are commonly used to profile VOCs in food matrices. However, dynamic headspace (DHS) extraction offers a significant advantage for monitoring the evolution of coffee aroma during roasting. In DHS, the volatile and semivolatile compounds are continuously captured on a sorbent tube, providing higher extraction capacity and sensitivity compared to SPME. These compounds are then thermally desorbed and analyzed using gas chromatography, making DHS ideal for real-time profiling.

Despite the use of conventional one-dimensional gas chromatography (1D-GC) in analyzing coffee, ,,,, it often struggles with peak overlap due to the complex nature of the sample. Comprehensive two-dimensional gas chromatography coupled with mass spectrometry (GC×GC-MS) addresses this limitation by providing enhanced peak separation and a more detailed characterization of the coffee matrix. While some approaches have tried to monitor roasting without GC, ,− these methods face challenges in compound identification, as condensed spectral data cannot be easily deconvoluted. In contrast, using GC×GC with electron ionization and commercial MS libraries such as NIST improves the reliability of analyte identification. Given the vast amount of data generated by DHS-GC×GC-MS, manual analysis of each chromatographic peak would be inefficient and could lead to the loss of critical information. Therefore, chemometric tools are necessary for effective data mining.

In this study, the main objective is to develop an analytical setup for monitoring the volatile profile during the coffee bean roasting process. In this scenario, a DHS-GC×GC-TOFMS method was evaluated for monitoring aroma-related volatiles produced during coffee roasting, which may provide important insights into the complex development of coffee aroma. Sampling of aroma-related volatiles was performed during three distinct roasting stages. Using a custom-built lab roaster, five subgroups of Coffea arabica were analyzed. VOCs and SVOCs were captured using DHS and analyzed by GC×GC-TOFMS. Partial least-squares discriminant analysis (PLS-DA) successfully differentiated the samples across the roasting stages, while the Variable Importance in Projection (VIP) approach identified key compounds with significant concentration variations. This integrated methodology provided a powerful tool for real-time profiling of aroma compounds, contributing with valuable insights into coffee roasting chemistry and supporting the development of optimized roasting strategies.

2. Material and Methods

2.1. Samples

Five batches containing 1 kg of unroasted coffee beans were provided by a Regional Coffee Growers Cooperative, all belonging to the Coffea arabica category but from different subgroups. Upon receipt, the coffee samples were stored in a light-protected environment under controlled temperature conditions. Table S1 details the characteristics provided by the cooperative for each type of coffee analyzed. The percentage of impurities and moisture complies with the regulations of the Ministry of Agriculture, Livestock, and Supply (MAPA) for unroasted coffee beans.

2.2. Coffee Roasting

In each experiment, approximately 8.2 ± 0.1 g of raw coffee beans were roasted in the laboratory. The heating program selected for roasting was based on a previous study by our group. A 1040 W Mini Retro electric popcorn popper, model RHP 310 (Nostalgia Products Group, Green Bay, WI, USA), was modified for this purpose. The main modifications included adding a potentiometer to the power cord, allowing control of the energy supplied to the popcorn popper’s electric heating system, and separating the power supply from the ventilation system, which is responsible for rotating the beans.

Based on previously obtained results, wherein it was investigated the impact of roasting on VOC associated with changes in the coffee aroma and the potential impact of analyzing the samples several days after roasting, this study adopted an intermediate roasting program with a total heating time of 8 min, where the power increased from 70% to 100% after the first 2 min.

2.3. Extraction

The extraction of volatile and semivolatile organic compounds was performed using dynamic headspace (DHS) sampling with 5 mm-OD glass thermal desorption (TD) tubes (Part No. 2414-1025) packed with a sorbent phase of Tenax TA and Carboxen (GL Sciences B.V. – Eindhoven, The Netherlands). The TD tubes were preconditioned according to the manufacturer’s recommendations. Additional information regarding the extraction performance of commonly used sorbent materials in TD tubes is available elsewhere for interested readers. −

A sampling port was added to the popcorn popper by replacing the original lid with a custom 3D-printed cover, as shown in Figure . This modification ensured an insulated environment so that the VOCs produced during roasting could be carried through an opening in the lid and be sorbed by the liner used. Additionally, the temperature was controlled using a TC4S module with a Pt100 probe (Autonics – Sao Paulo, SP, Brazil).

1.

Experimental setup used for roasting raw coffee and capturing volatile (VOC) and semivolatile organic compounds (SVOC) produced during the process. Analyte sampling and preconcentration was attained using dynamic headspace (DHS) extraction.

Three distinct roasting stages were monitored during the 8 min roasting of raw coffee beans. The first monitored stage occurred from 0 to 2:40 min (time 1), the second from 2:40 to 5:20 min (time 2), and the third from 5:20 to 8:00 min (time 3). At the end of each stage, the liner was replaced with another preconditioned liner, totaling three liners per roasting. All liners were sealed immediately after removal from the system and were only opened when introduced into the OPTIC-4 multimode injector (GL Sciences B.V., Eindhoven, The Netherlands) for the thermal desorption step in the gas chromatograph.

As shown in Figure , a setup consisting of a vacuum pump, a needle valve, and inert Tygon S3 hoses (Saint Gobain) was used to generate a constant flow of 100 mL/min, exiting the roaster and passing through the liner with the sorbent phase. Each of the five available raw coffee samples was roasted in duplicate, resulting in a total of 30 measurements.

2.4. TD-GC×GC-TOFMS

Analyte introduction was performed using thermal desorption. After sampling, the TD tube was placed in the OPTIC-4 module. Analyte desorption was achieved using a fast-heating rate from 40 to 260 °C at 20 °C/s. The injection modes were evaluated in split and splitless conditions. The optimized method was performed using a 2 min splitless program. Helium was used as the carrier gas at a constant flow rate of 1.0 mL/min.

GC×GC-TOFMS analyses were conducted on an Agilent 8890 gas chromatograph coupled with a Pegasus BT 4D time-of-flight mass spectrometer (LECO Corporation – St. Joseph, MI, USA). Modulation was performed using a consumable-free QUADJET thermal modulator. For further information on modulation techniques, please read , . The chromatographic columns consisted of a first-dimension (¹D) column of 30 m × 0.25 mm × 0.25 μm Rxi-5MS and a second-dimension (²D) column of 2.2 m × 0.15 mm × 0.15 μm Rxi-17Sil MS (Restek Corporation – Bellefonte, PA, USA). The ²D column was housed in a secondary GC oven.

The 1D oven was heated from 50 to 260 °C at 4 °C/min, with a final isotherm of 5 min. The 2D oven and the modulator were programmed with offsets of +5 °C and +20 °C above the primary oven, respectively. The modulation period was 5 s.

The transfer line to the TOFMS and the ion source were maintained at 275 and 250 °C, respectively. Electron ionization at 70 eV was used in all experiments. The acquisition rate was 100 spectra/s with a mass range of 50–600 m/z.

2.5. Identification

GC Image software (Zoex – Houston, TX, USA) was used for the tentative identification of analytes through mass spectrum similarity searches using the NIST20 MS library (National Institute of Standards – Gaithersburg, MD, USA). LTPRI (linear temperature-programmed retention index) filtering was also utilized in the analyte identification process. Identification was performed using a minimum similarity of 75% between the experimental mass spectra and those in the NIST library, with a deviation of ±50 LTPRI units from the NIST values. Odor descriptors were obtained from the Good Scents Company database, which offers olfactometry data from pure compounds. ,,

Peak detection was performed using GC Image. Feature assignment was conducted using template matching based on marker peaks and chemical logic for chromatographic alignment. ,

2.6. Data Processing

The MATLAB R2021b environment (MathWorks – Natick, MA, USA) was used for chemometric analysis. The processed peak tables were exported from GC Image in “.csv” format and concatenated to form a matrix X, containing all independent variables for modeling. The data set comprised 169 features from 30 roasted coffee measurements. The matrix X (30 × 169) was preprocessed using autoscaling in PLS Toolbox 9.0 (eigenvector Research Inc. – Wenatchee, WA, USA).

Data processing using PLS-DA allowed correlation of the area of volatile organic compounds in the measurements with each of the three studied roasting stages (property of interest). In this study, values of 1 were assigned to measurements from the first roasting stage (time 1), values of 2 to measurements from the middle stage (time 2), and values of 3 to measurements from the final stage (time 3). Thus, a vector representing the property of interest y (30 × 1) was formed.

The model for correlating experimental data (X) with the property of interest (y) was constructed, and the number of latent variables was selected based on the analysis of mean calibration and cross-validation errors (Venetian blind method). The model’s classification performance was evaluated using a confusion matrix, considering the values of true positive (TP), true negative (TN), false positive (FP), and false negative (FN).

Permutation tests were used to help identify an overfit model as well as to determine whether the given model was significantly different from one built under the same conditions but with random data. If the modeling conditions were overfitting, they would often produce a fit to random data that was better than expected. Permutation tests were performed using PLS Toolbox and confirmed the absence of overfitting (Supporting Information). In general, values less than 0.05 indicate model significance at the 95% confidence level.

3. Results and Discussion

3.1. TD-GC×GC-TOFMS

The change in coffee bean color, transitioning from a greenish hue to a brownish one throughout the roasting process, is the most distinct change that can be observed during this stage. However, it is not only the color of the coffee beans that undergoes alterations during roasting. This change is accompanied by evolutions in the volatile profile of the beans due to a range of chemical reactions promoted by heating, including Maillard reactions, Strecker degradation, caramelization, degradation of chlorogenic acids, and lipid oxidation.

The chromatograms obtained for the coffee samples at roasting times 1 and 3 (Figure A,B) reveal an evolution in the volatile profile of the coffee beans as the roasting time progresses. Additionally, the chromatograms of all 30 measurements indicate the presence of approximately 169 volatile compounds, which were trapped by the liners during roasting and tentatively identified. The analytes, as presented in Table , include aromatic hydrocarbons, alcohols, ketones, linear and branched hydrocarbons, nitrogenous compounds, esters, carboxylic acids, aldehydes, terpenes, and terpenoids, among others, consistent with previous studies. ,, Odor descriptors for each compound were obtained from refs ,, . A blank run of the setup used in the study was conducted to eliminate any potential memory effects.

2.

Examples of GC×GC-TOFMS chromatograms for different roasting time intervals of a raw coffee sample using DHS with a mixed sorbent phase (Tenax TA and Carboxen). (A) Time 1; (B) Time 3. The chromatogram at Time 1 corresponds to volatiles produced between 0 and 2:40 min, while at Time 3, it represents volatiles produced between 5:20 and 8:00 min. (C) Chromatogram of a coffee sample, highlighting (white dots) the 99 features identified using the VIP values approach in the PLS-DA model. Additionally, the chemical structures of some selected VIP compounds were also depicted in the figure.

1. Analytes Identified in the Volatile and Semivolatile Fraction of Coffee Samples throughout the Roasting Process .

	Analyte identity	M.F.	ID Level	CAS	Odor	VIP value
1	(1-butylheptyl) benzene	C17H28	3	4537–15–9	-	1.60
2	(1-butyloctyl)-benzene	C18H30	2	2719–63–3	-
3	(1-methylpropyl)-benzene	C10H14	3	135–98–8	-
4	(1-methylpropyl)-cyclohexane	C10H20	2	758–01–7	-	1.31
5	(1-propylnonyl)-benzene	C18H30	2	2719–64–4	-	1.93
6	(2-methylbutyl) cyclohexane	C11H22	3	54105–77–0	-
7	(Z)-14-methyl-8-hexadecen-1-ol	C17H34O2	3	-	-	1.47
8	1-(2-ethylphenyl) ethanol	C10H14O	3	-	-	1.50
9	1-(2-furanyl)-ethanone	C6H6O2	3	1192–62–7	powerful balsamic-sweet odor with a tobacco-like, almost narcotic pungency and floral undertones of balsamic-cinnamic character
10	1-(2-methyl-1H-pyrrol-3-yl) ethanone	C7H9NO	3	-	-	1.77
11	1-(2-methylphenyl)-ethanone	C9H10O	2	577–16–2	sweet, hawthorn powdery, anisic, coumarinic, phenolic, burnt	1.01
12	1-(4-tert-butylphenyl) propan-2-one	C13H18O	3	81561–77–5	-	1.13
13	1-(acetyloxy)-2-propanone	C5H8O3	2	592–20–1	fruity-buttery, somewhat sour	1.89
14	1-(hydroxyphenyl)-2-methylpropane	C10H14O	3	-	-
15	1-(methoxymethyl)-4-methylnaphthalene	C13H14O	3	71235–76–2	-
16	1,1-dimethyl-2-propyl-cyclohexane	C11H22	3	81983–71–3	-
17	1,2,3,4-tetrahydro-5-methyl-naphthalene	C11H14	2	2809–64–5	-
18	1,2,3,4-tetrahydro-naphthalene	C10H12	2	119–64–2	-	1.76
19	1,2,3,4-tetramethyl-benzene	C10H14	3	488–23–3	-	1.76
20	1,2,3,5,6,7-hexahydro-1,1,2,3,3-pentamethyl-4H-inden-4-one	C14H22O	2	33704–61–9	diffusive, spicy, musk-like with strong floral note
21	1,2,3-trimethyl-(1α,2α,3β)-cyclohexane	C9H18	2	7667–55–2	-
22	1,2,3-trimethyl-benzene	C9H12	2	526–73–8	-
23	1,2,4,5-tetramethyl-benzene	C10H14	2	95–93–2	weak, hazelnut	1.36
24	1,2,4-trimethyl-3-nitrobicyclo[3.3.1]nonan-9-one	C12H19NO3	3	129967–65–3	-	1.50
25	1,2,4-trimethylbenzene	C9H12	2	95–63–6	-
26	1,2-dimethyl-benzene	C8H10	3	95–47–6	geranium	1.45
27	1,3,5-trimethylbenzene	C9H12	3	108–67–8	-	1.28
28	1,3-bis(1,1-dimethylethyl)-benzene	C14H22	2	1014–60–4	-	2.30
29	1,3-diethyl-5-methyl-benzene	C11H16	2	2050–24–0	-	1.14
30	1,7-dimethyl-naphthalene	C12H12	2	575–37–1	-	1.08
31	16-oxokahweol	C22H28O4	2	81760–47–6	-	2.03
32	1-butyl-cyclohexene	C10H18	2	3282–53–9	-
33	1-ethyl-2,3-dimethylcyclohexane	C10H20	3	7058–05–1	-	1.18
34	1-ethyl-2-methyl-benzene	C9H12	2	611–14–3	-
35	1-ethyl-2-methyl-cyclohexane	C9H18	2	3728–54–9	-	1.58
36	1-ethyl-3-methyl-benzene	C9H12	2	620–14–4	-
37	1-ethylidene-1H-indene	C11H10	2	2471–83–2	-	1.02
38	1-isopropyl-2,3-dimethylcyclopentane	C10H20	3	489–20–3	-
39	1-methyl-2-propyl-cyclohexane	C10H20	2	4291–79–6	-
40	1-methyl-3-propyl-benzene	C10H14	2	1074–43–7	-
41	1-methyl-3-propyl-cyclohexane	C10H20	3	4291–80–9	-	2.16
42	1-methyl-4-(1-methylbutyl)-cyclohexane	C12H24	2	54411–00–6	-
43	1-methyl-4-(2-methyloxiranyl)-7-oxabicyclo [4.1.0] heptane	C10H16O2	2	96–08–2	mentholic
44	1-methyl-4-propyl-benzene	C10H14	2	1074–55–1	-
45	1-methyl-naphthalene	C11H10	2	90–12–0	naphthyl, chemical, medicinal, camphoreous	1.26
46	1-tetradecene	C14H28	3	1120–36–1	-	2.00
47	2,2-bis[5-(1,4-benzodithiafulven-6-yl)-2-thienyl]propane	C27H20S6	3	-	-	1.38
48	2,3-dihydro-1,3-dimethyl-1H-indene	C11H14	3	4175–53–5	-	1.38
49	2,3-dihydro-1,6-dimethyl-1H-indene	C11H14	3	17059–48–2	-	1.06
50	2,3-dihydro-4,7-dimethyl-1H-indene	C11H14	3	6682–71–9	-	1.32
51	2,3-dihydro-5-methyl-1H-indene	C10H12	2	874–35–1	-	1.04
52	2,4-decadienal	C10H16O	2	25152–84–5	orange-like, sweet and fresh-citrusy, fried, oily
53	2,4-dimethyl-1-heptene	C9H18	2	19549–87–2	-	1.56
54	2,5-dimethyl-pyrazine	C6H8N2	2	123–32–0	roasted, nutty, grassy, “cornnuts”
55	2,6-dibutyl-1,4-benzoquinone	C14H20O2	3	-	-	1.13
56	2,9-dimethyl-decane	C12H26	3	1002–17–1	-
57	2-ethenyl-5-methyl-pyrazine	C7H8N2	2	13925–08–1	coffee	1.32
58	2-ethyl-1,3-dimethylcyclohexane	C10H20	3	7045–67–2	-
59	2-ethyl-1,4-dimethyl-benzene	C10H14	2	1758–88–9	-	1.42
60	2-ethyl-1H-isoindole-1,3(2H)-dithione	C10H9NS2	3	35373–06–9	-	2.09
61	2-ethyl-3-methyl-pyrazine	C7H10N2	3	15707–23–0	nutty, earthy, must, bread, peanut	1.67
62	2-ethyl-6-methyl-pyrazine	C7H10N2	2	13925–03–6	roasted potato
63	2-ethylhexyl salicylate	C15H22O3	2	118–60–5	orchid, sweet, balsam
64	2-furancarboxyaldehyde	C5H4O2	3	98–01–1	sweet, woody, baked bread, caramel-like	1.14
65	2-furanmethanol	C5H6O2	2	98–00–0	slightly caramellic, warm-oily (well correlated with the underirable burnt and bitter note of dark-roasted coffees)	1.04
66	2-hexyldec-1-en-3-yne	C16H28	3	69520–29–2	-	1.17
67	2-methyl-tetrahydro-furan-3-one	C5H8O2	2	3188–00–9	sweet, brown, bready, nutty	1.89
68	2-methyl-3-buten-2-ol acetate	C7H12O2	2	24509–88–4	-	1.39
69	2-methyl-decane	C11H24	2	6975–98–0	-	1.00
70	2-methyl-propanoic acid 1-(1,1-dimethylethyl)-2-methyl-1,3-propanediyl ester	C16H30O4	3	74381–40–1	-	1.60
71	2-methyl-propanoic acid 3-hydroxy-2,4,4-trimethylpentyl ester	C12H24O3	3	74367–34–3	-
72	2-nonadecanone	C19H38O	2	629–66–3	-	1.27
73	2-nonenal	C9H16O	2	18829–56–6	green, cucumber, aldehydic, fatty, citrus
74	2-propenylidene-cyclobutene	C7H8	3	52097–85–5	-	1.44
75	2-syn-methyl-decalin	C11H20	2	2958–76–1	-	1.31
76	2-tetradecyloxirane	C16H32O	3	7320–37–8	-
77	2-undecenoic acid	C11H20O2	3	15790–94–0	-
78	3,4,5-trichloro-6-methylcyclohepta-2,4,6-trien-1-one	C8H5Cl3O2	3	-	-	1.54
79	3,5-dimethyl-octane	C10H22	2	15869–93–9	-	1.70
89	3,6-dimethyl octane	C10H22	2	15869–94–0	-
81	3,7-dihydro-1,3,7-trimethyl-1H-purine-2,6-dione	C8H10N4O2	3	58–08–2	odorless	1.03
82	3,7-dimethyl-nonane	C11H24	2	17302–32–8	-
83	3-ethylbicyclo [4.4.0] decane	C12H22	3	-	-
84	3-methyl-decane	C11H24	2	13151–34–3	-
85	3-methyl-nonane	C10H22	3	5911–04–6	-
86	3-tetradecyne	C14H26	2	60212–32–0	-
87	4,5-dimethyl-nonane	C11H24	3	17302–23–7	-
88	4-methyl-nonane	C10H22	2	17301–94–9	-	1.85
89	4-tert-butyl-cyclohexanol acetate	C12H22O2	2	10411–92–4	fresh, fruity, floral, pine
90	4-tert-butylcyclohexyl acetate	C12H22O2	2	32210–23–4	woody, floral, oily, fruity, herbal	1.09
91	5-(4-methylphenoxy) methyl-2-amino-1,3,4-thiadiazoles	C10H11N3OS	3	-	-	1.59
92	5-ethyl-2-methyl-octane	C11H24	3	62016–18–6	-	1.10
93	5-methyl-2(3H)-furanone	C5H6O2	2	591–12–8	sweet, coconut, tobacco and coumarin-like nuances	1.78
94	5-methyl-2-furancarboxaldehyde -	C6H6O2	2	620–02–0	sweet-spicy, warm, slightly caramellic
95	6-ethyl-2-methyl-octane	C11H24	3	62016–19–7	-
96	7,9-ditert-butyl-1-oxaspiro(4,5)deca-6,9-diene-2,8-dione	C17H24O3	2	82304–66–3	-	1.02
97	7-tetradecyne	C14H26	2	35216–11–6	-
98	8-methyl-1-decene	C11H22	2	61142–79–8	-	2.34
99	9,12-octadecadienoic acid	C18H32O2	3	60–33–3	faint fatty
100	9,12-octadecadienoic acid, methyl ester	C19H34O2	2	2462–85–3	-	1.08
101	9-octadecenoic acid	C18H34O2	3	112–80–1	fatty, waxy, lard, fried, tallow	1.53
102	acetate 2-furanmethanol	C7H8O3	2	623–17–6	sweet, green banana peel, estery
103	acetic acid	C2H4O2	3	64–19–7	pungent, stinging sour	1.41
104	acetophenone	C8H8O	2	98–86–2	sweet, mimosa, acacia, floral, almond
105	amberonne (isomer)	C16H26O	2	-	-	1.00
106	amberonne (isomer)	C16H26O	2	-	-	1.38
107	benzenemethanol	C7H8O	2	100–51–6	light floral, rose, phenolic, balsamic	1.12
108	bicyclo [4.1.0] heptan-2-one	C7H10O	3	5771–58–4	-
109	bicyclo [5.3.0] decane	C10H18	2	16189–46–1	-
110	biphenyl	C12H10	2	92–52–4	metallic, cinnamon-like, bergamot- and neroli-like	1.23
111	camphor	C10H16O	2	464–48–2	camphoreous
112	cyclohexanone	C6H10O	3	108–94–1	minty, acetone
113	decahydro-1,6-dimethyl-naphthalene	C12H22	3	1750–51–2	-	1.33
114	decahydro-trans-naphthalene	C10H18	2	493–02–7	-	1.04
115	decanal	C10H20O	3	112–31–2	sweet, waxy, orange peel, citrus, floral	1.02
116	diacetate 1,2-ethanediol	C6H10O4	3	111–55–7	green floral estery alcoholic
117	diethyl phthalate	C12H14O4	2	84–66–2	odorless
118	dihydromyrcenol	C10H20O	2	18479–58–8	fresh, lime cologne, herbal, air marine	1.46
119	dodecanal	C12H24O3	2	112–54–9	soap, waxy, aldehydic, citrus, green, floral	1.31
120	dodecanamide	C12H25NO	3	1120–16–7	-	2.40
121	eicosanoic acid methyl ester	C21H42O2	2	1120–28–1	-	1.24
122	ethylbenzene	C8H10	2	100–41–4	-	1.54
123	ethyl-pyrazine	C6H8N2	2	13925–00–3	peanut, butter, musty, woody, roasted, cocoa	1.11
124	heptane, 4-methyl-	C8H18	3	589–53–7	-
125	hexadecanoic acid, methyl ester	C17H34O2	3	112–39–0	oily, waxy, fatty, orris	1.30
126	hexanal	C6H12O	2	66–25–1	very powerful, penetrating, fatty-green, grassy. Cut grass and unripe fruit	2.80
127	hexane, 3-ethyl-2,5-dimethyl-	C10H22	2	52897–04–8	-
128	homosalate	C16H22O3	2	118–56–9	mild menthol
129	indane	C9H10	2	496–11–7	-	1.04
130	isobutyl ether	C8H18O	3	628–55–7	-
131	isopropyl myristate	C17H34O2	3	110–27–0	faint, oily, fatty	1.53
132	levomenthol	C10H20O	2	2216–51–5	peppermint, spicy, cooling, camphoreous	1.47
133	lilial	C14H20O	2	80–54–6	fresh, floral, lilac, powdery	1.21
134	limonene	C10H16	2	138–86–3	citrusy, lemon-like, fresh and sweet odor
135	linalool oxide	C10H18O2	2	60047–17–8	powerful sweet-woody, penetrating odor with floral-woody-earthy undetones	1.31
136	longifolene	C15H24	2	475–20–7	sweet, woody, rose, medicinal	1.15
137	methyl (2-pyrimidyl) amine	C5H7N3	3	931–61–3	-	1.98
138	methyldecahydronaphthalene	C11H20	3	2958–75–0	-	1.04
139	methyl-pyrazine	C5H6N2	2	109–08–0	nuttty, cocoa, roasted, coffee, chocolate
140	naphthalene	C10H8	2	91–20–3	pungent, choking dry-tarry odor of moderate to poor tenacity	1.06
141	nonanal	C9H18O	3	124–19–6	strong, soap-like, metallic	1.92
142	octadecanoic acid	C18H36O2	2	57–11–4	odorless, mild, fatty, waxy	1.56
143	octahydro-1H-indene	C9H16	2	496–10–6	-	1.15
144	octahydro-1H-indene	C9H16	2	3296–50–2	-
145	octahydro-2,2,4,4,7,7-hexamethyl-1H-indene	C15H28	3	54832–83–6	-	1.49
146	octanoic acid	C8H16O2	2	124–07–2	oily rancid, sweat-like, repulsive even in dilution	1.35
147	o-tert-butyl cyclohexyl acetate	C12H22O2	2	88–41–5	fruit, woody, green, apple, herbal	1.55
148	p-cymene	C10H14	2	99–87–6	typical kerosene-like odor in high cocnentration
149	pentadecanal	C15H30O	2	2765–11–9	fresh, waxy
150	pentanoic acid, 2,2,4-trimethyl-3-hydroxy-, isobutyl ester	C12H24O3	3	244074–78–0	-	1.06
151	pentanoic acid, 3-mehtyl	C6H12O2	3	105–43–1	animalic, sharp acidic, cheesy, green with a fruit swweaty nuance
152	pentyl-cyclohexane	C11H22	3	4292–92–6	-
153	phenylacetaldehyde	C8H8O	2	122–78–1	green, floral, and sweet
154	phenylmethane	C7H8	2	108–88–3	sweet-gassy odor, milder than that of benzene	1.81
155	phthalic acid, bis(2-ethylhexyl) ester	C24H38O4	2	117–81–7	-
156	phthalic acid, isobutyl 3-methylbutyl ester	C17H24O4	2	-	-
157	phytol	C20H40O	2	253686–88–3	-	1.14
158	phytone	C18H36O	2	502–69–2	oily, herbal, jasmin, celery, woody	1.29
159	p-mentha-1,5,8-triene	C10H14	2	21195–59–5	roasted	1.49
160	propyl- benzene	C9H12	2	103–65–1	-
161	propyl-cyclohexane	C9H18	2	1678–92–8	-
162	protoanemonine	C5H4O2	3	108–28–1	-	1.20
163	pseudocumene	C9H12	3	95–63–6	-
164	tetrachloroethylene	C2Cl4	2	127–18–4	-	1.28
165	tetradecanoic acid	C14H28O2	2	544–63–8	waxy, fatty, soapy, coconut	1.38
166	thiofanox sulfone	C9H18N2O4S	3	39184–59–3	-
167	trans-13-octadecenoic acid	C18H34O2	2	693–71–0	-	1.60
168	undecanal	C11H22O	2	112–44–7	waxy, aldehydic, citrus, watermelon	1.93
169	α-hexyl-cinnamaldehyde	C15H20O	2	101–86–0	-

^aLinear temperature programming retention indices (LTPRI) were obtained from NIST. All analytes were identified considering a minimum mass spectrum similarity of 75% and deviations less than ± 50 LTPRI units from the NIST Web Book. Analytes identified by the VIP values approach (Section ) in the volatile and semivolatile fraction of coffee samples during the roasting procedure are also highlighted. Note: VOCs are listed in alphabetical order. ID Level: Description available at https://doi.org/10.1007/s11306-007-0082-2.

To ensure that all chemical information obtained by DHS-GC×GC-TOFMS is adequately interpreted, it is essential to employ appropriate data processing approaches, such as chemometrics. This allows patterns among the samples to be recognized, preventing the loss of any relevant information in the study.

3.2. Data Processing

A peak table-based approach was adopted, enabling robust statistical and multivariate analysis with reduced computational demands. Feature assignment utilized smart templates based on marker peaks and chemical logic for chromatographic alignment. , More specifically, this workflow defined peak regions to establish chromatographic windows where analytes were consistently detected and matched using spectral information across chromatograms, generating comprehensive chemical profiles. In the current study, data processing was performed using PLS-DA, a widely employed approach in supervised analyses. If the generated model is successful, it will have calculated a threshold value to differentiate and classify the samples among the three present classes (class 1 – time 1; class 2 – time 2; and class 3 – time 3).

One of the initial steps in constructing the model was determining the number of latent variables (LVs) to be used. The chosen number of LVs corresponds to the dimensions of the subspace in which the original data will be projected. The selection of the number of LVs was based on the mean classification errors for cross-validation (Venetian blind). In this scenario, a PLS-DA model with 3 latent variables was built.

The confusion matrix of the PLS-DA model (Table S2) assists in identifying that the VIP (Variable Importance in Projection) values of the model contain information related to each of the three modeled roasting stages (i.e., three classes) and, consequently, to compounds that underwent concentration variations throughout roasting. Since the primary objective was to identify VIP scores through the PLS-DA model, permutation tests were used to evaluate overfitting. As shown in Table S3, such results indicated the absence of an overfitted PLS-DA model. In this context, the VIP values for each class were exported from the model, and those above the value of 1 were selected. As a result, 99 VIPs were identified, referring to variables (compounds) from the data matrix that are statistically significant for classifying samples into each class.

Therefore, 99 organic compounds among all the analytes trapped in the liners and analyzed by DHS-GC×GC-TOFMS exhibited statistically significant changes in their profiles during the roasting of raw coffee beans. All 99 compounds identified by this approach are highlighted in the last column of Table , and their respective chromatographic peaks can be visualized in the chromatogram presented in Figure C.

After listing all compounds using the VIP values approach, they were grouped into chemical classes, including aromatic hydrocarbons, alcohols, ketones, terpenoids, terpenes, hydrocarbons, carboxylic acids, nitrogen-containing compounds, esters, aldehydes, and purine. Figure illustrates the changes in the areas of these chemical classes over the three monitored roasting stages (time 1, time 2, and time 3).

3.

Roasting profiles of selected VOCs and SVOCs, grouped by chemical classes. Feature selection was based on the VIP variable selection approach. Different profiles can be identified: (A, B) analytes that show an increasing peak area during roasting; (C) intermediary analytes; and (D) carboxylic acids and purines, which exhibit a continuous decrease in peak area.

The development of coffee aroma and taste by roasting remains a highly empirical and handcrafted activity. In this scenario, the development of online approaches exploring mass spectrometric techniques to follow and obtain real-time information along the roasting process is highly desirable. Some of these researchers used online chemical ionization , and photoionization, ,, mass spectrometry, while others developed interesting setups to study the complex network of chemical changes that occur in a single bean. ,, Moreover, an intriguing approach dynamically profiled coffee volatiles in the roasting plume exiting the roaster hood using a styrene-divinylbenzene resin (XAD-2). The interested reader is directed elsewhere for other studies in changes in volatilomics caused by roasting of food matrices. −

During the roasting step, the coffee bean can be seen as a pressurized reactor in which some of the reactions first observed by the chemist Louis-Camille Maillard in 1912 can occur. The reactions bearing his name (Maillard reactions) involve the reaction between reducing sugars and amino acids or low-molecular-weight peptides to form several compounds. As proposed by Nursten et al., Maillard reaction products may include (a) “simple” sugar dehydration/fragmentation compounds (furans, pyrones, cyclopentenes, carbonyl compounds, and acids); (b) “simple” amino acid degradation products (aldehydes and sulfur-containing molecules); and (c) volatiles produced by a cascade of further interactions (pyrroles, pyridines, imidazoles, pyrazines, oxazoles, thiazoles, and compounds from aldol condensations).

Among the four principal aliphatic acids (formic, acetic, glycolic, and lactic) that may be generated from glucose and fructose fragmentation, , acetic acid (#103, Figure C) was successfully trapped and detected in this study. As with the other carboxylic acids (Figure D), this pungent compound exhibited a decreasing evolutionary profile along the roasting time. The decrease in acetic acid content, due to degradation or evaporation, during further roasting has been previously observed. Additionally, dark-roasted arabicas are prone to exhibit less acidity, with bitterness becoming the major taste. , Tetradecanoic acid (#165) and octanoic acid (#146), both depicted in Figure C, were also found in other studies, the latter being found in Colombian, Santos, and green Mexican arabica coffees and in a commercial coffee powder, exhibiting an oily rancid and sweat-like odor. #165 peak profile evolution over the course of the roasting procedure is illustrated in Figure A. Its decreasing trend over time follows the general pattern observed for the carboxylic acid group. Additionally, the oleic acid (#101) and stearic acid (#142) ratio was already mentioned as a possible indicator of robusta in coffee blends. However, the present study identified a different distribution profile between these two free fatty acids, with stearic acid exhibiting a greater proportion than oleic acid instead of an equivalent (arabica) or smaller (robusta), as previously reported for the coffee samples they analyzed.

4.

Distinct peak profiles of relevant compounds identified at each roasting stage. (A) Tetradecanoic acid, which exhibits a continuous decrease in peak intensity over the course of the roasting process; (B) hexanal, which increases initially and then slightly decreases; and (C) dihydromyrcenol, showing a continuous increase throughout the roasting experiment.

As part of the advanced stage of the Maillard reaction network, the Strecker degradation has preeminent importance in the formation of the coffee aroma spectrum. It involves the reaction between dicarbonyl compounds and amino acids to form a variety of products, including aldehydes, alkylpyrazines, aminoketones, carbon dioxide, and ketones. ,,, Ketones represent an important class of VOCs for coffee aroma. In general, they may represent 21.5% of the whole fraction of compounds that constitute the “coffee aroma”. Among them, many cyclic ketones, such as the VIP-selected (#11) 1-(2-methylphenyl)-ethanone, Figure C, may have their origin traced from carbohydrate dehydration/fragmentation. , Imparting a coumarinic, phenolic, and burnt aroma in coffee samples, #11 and the other VIP-selected ketones with their interesting aroma descriptors, such as 1-(2-furanyl)-ethanone, (#9, tobacco-like odor, almost narcotic pungency, and floral undertones) and acetophenone (#104, mimosa, acacia, and floral odors), more than double their total amount throughout the roasting stages (Figure B).

Maillard reaction also provides important sulfur-containing chemicals for coffee aroma, such as thiazoles and thiols. Their key role in the flavors of roasted and cooked foods has been recognized for years. The present experimental setup could successfully trap and detect other sulfur-containing volatile compounds, such as dithione (#60) and the thiodiazole derivative (#91). Unfortunately, none of their odor descriptors could be found; however, it is well-known that sulfur volatiles are essential for the characteristic impact of roasted coffee, with at least 100 compounds being identified in coffee. , Apart from Maillard reactions, their origin may also be associated with natural metabolic pathways of the coffee plant and storage of the roasted bean.

Pyrazines constitute another important chemical family derived from the Maillard reaction network. As mentioned for the sulfur-containing compounds, pyrazines are important contributors to the flavor of raw, roasted, and cooked foodstuffs, and there are a few approaches developed to obtain them not only from natural sources (e.g., extraction and concentration) but also by developing and exploring modern biotechnology processes. The initial model experiments tried to elucidate pyrazine synthesis mechanisms through reactions between α-amino acids, carbohydrates, and small oxygen-containing molecules. All these movements aim to cope with the increasing demand for adding these nitrogen-containing compounds to foods. We have observed that the onset of roasting is accompanied by an increase in pyrazine concentration. This is directly related to the coffee’s aroma profile, as several pyrazines contribute to roasted and coffee-like notes, among others. Among the six pyrazines trapped and identified in this study, 2,5-dimethylpyrazine (#54) (roasted and nutty odors) is used in breakfast cereals, and 2-ethyl-3-methylpyrazine (#61, Figure C) (musty, earthy, bread, and peanut odors) is used in peanut products, popcorn, and bread. Considering the three VIP-selected pyrazines (#57, #61, and #123), all of them have already been previously reported in coffee samples. , Apart from different roasting conditions and being a robusta coffee sample, 2-ethenyl-5-methylpyrazine (#57, Figure C) (coffee odor) exhibited a similar evolutionary profile as observed by Silwar and Lüllmann. Its peak area increased along the roasting process (from stage 1 to 2) and then remained almost the same until the end of the roasting period. Grouping the pyrazines with other VIP-selected nitrogen-containing compounds, they exhibited the evolutionary profile of initial growth and then their total area value remained almost constant until the end of the experiment (Figure A).

Aldehydes were one of the chemical groups that presented an evolutionary behavior of an initial rise followed by a gradual decline, Figure C. Considering the roasting scenario, one of their origins may be traced to the oxidative degradation of amino acids along their interaction with sugars at elevated temperatures. Another established route to aldehyde formation is through the autoxidation of unsaturated fatty acids. Considering the recognized presence of these lipids in coffee, it was interesting to not only visually observe a few very small oily droplets on the roasted bean surface at the end of the roasting period but also detect some unsaturated fatty acids in the current study (such as linoleic acid (#99), oleic acid (#101), and trans-13-octadecenoic acid (#167)). Some aldehydes mentioned in the literature that have their origin traced to linoleic acid and methyl linoleate autoxidation include hexanal (#126, Figure C) and 2,4-decadienal (#52). Besides being identified in the roasted product, the VIP-selected hexanal can be found in raw coffee beans and be involved in the staling process of roasted coffee beans when stored in an oxygen-containing atmosphere. Like the other aldehydes, hexanal also exhibited a general peak profile evolution of initial increase followed by a slightly declineas illustrated in Figure B. With its characteristic green odor notes, 2-nonenal (#73) formation is also linoleic acid-related, being mentioned as a compound that helps balance some flavor notes, such as acidic, sour, and caramel.

Even though they exhibited a constant increase profile along the roasting period monitored (Figure A), apparently, in general, low-molecular-weight alcohols do not tend to be potent odorants in coffee. Some exceptions may include alcohols from the recognized organoleptically active class of the isoprenoids (i.e., terpenoids). Five of these compounds were detected in the roasting process, with all of them (#31, #118, #132, #135, and #157) being statistically significant for sample classification in the PLS-DA model, since they increased steadily across the three observed roasting stages. Being possibly derived from linalool (already cited as an important odorant for roasted powder of *arabica* coffee) oxidation, linalool oxide (#135) odor description is cited as powerful sweet-woody and a penetrating odor with floral-woody-earthy undertones. #135 was also found in strongly roasted coffee samples and in a raw coffee characterized by a “Rio” off-flavor. , Levomenthol (#132, peppermint and spicy odors) and dihydromyrcenol (#118, Figure C, fresh, herbal, and air marine odors) were other interesting alcohol isoprenoids found in the roasted coffee samples. The steadily increasing profile of #118 during the coffee roasting experiment, in line with the trend observed for its chemical class, is shown in Figure C. Apart from not finding any odor description for it, 16-oxokahweol (#31) was another terpene-derived compound found in the coffee samples. This diterpene alcohol ester may be formed by the esterification of the hydroxyl group at the C-16 position of kahweol. And both diterpenes are cited as present in the lipid fraction of robusta and arabica coffees. In fact, kahweol is cited as very sensitive to acids, heat, and light, being reported as an unstable compound when in its purified form, , which could possibly justify its absence in the DHS-trapped fractions.

Similarly to the alcohols, the terpenes’ evolutionary profile had a prominent increase in the first half of the roasting period, and then their total area value remained almost constant until the end of the roasting period (Figure B). Besides not being a VIP, the monoterpenic compound limonene (#134, citrussy, lemon-like, and fresh) was already widely found in coffee analysis. Being studied as a major contributor to the aroma of specialty coffees, , for example. Among the other two VIP-selected terpenic compounds, an interesting one is p-mentha-1,5,8-triene (#159, Figure C), which is characterized by a roasted odor.

Although aliphatic and aromatic hydrocarbons are among the most abundant chemical families in the analyzed coffee samples, they are generally not considered to have strong odoriferous properties compared to other characteristic coffee constituents mentioned above. As shown in Table , little information is available regarding the odor descriptors of these compound classes. Nevertheless, they may play an indirect role in the aroma and flavor of coffee.

Obtaining such an extensive and rich data set on the identities and behaviors of coffee compounds during the roasting process of Coffea arabica represents a significant step toward a comprehensive understanding of the art of coffee roasting. Based solely on the sensory notes of the compounds listed in Table , it may be possible to achieve a coffee with more floral notes by roasting it somewhere between Times 2 and 3, as this time frame exhibits a higher concentration of compounds associated with these sensory descriptors. Similarly, for those seeking a more traditional coffee profile, roasting up to time frame 2 could be a viable option, as this stage presents the highest concentration of pyrazines, which contribute to roasted and coffee-like odor descriptors. It is important to emphasize that these observations are general insights based on chemical and sensory data, without delving into the complex antagonistic and synergistic interactions among compoundskey variables in this field.

Finally, developing new approaches that integrate chemical analysis and chemometric tools can help bridge existing gaps in coffee research. Moreover, combining such methodologies with additional data sources, such as sensory evaluations, has the potential to enhance coffee quality and refine roasting profiles. This, in turn, can be tailored to factors such as the quality of raw coffee, its origin, and brewing methods.

4. Conclusions

This paper introduces a novel setup and workflow by combining DHS-GC×GC-TOFMS and chemometrics to gain deeper insights into the evolutionary profiles of chemical compounds during the roasting stages of raw coffee beans. The study utilized a reliable experimental setup for real-time VOC and SVOC collection, using reusable and immobilized adsorbent phases for analyte sampling and preconcentration.

The enhanced resolving power of GC×GC-TOFMS was key to avoiding the loss of roasting-related chemical information due to peak overlap, as demonstrated by the complex volatile profiles obtained. Subsequently, PLS-DA combined with a VIP approach proved to be an effective method for studying the 99 organic compounds that underwent statistically significant concentration variations across the roasting stages, revealing distinct evolutionary profiles for each chemical family. In this scenario, the method developed proved to be suitable for monitoring the evolution of the coffee’s volatile profile during roasting.

Our findings undoubtedly contribute to a clearer understanding of the diverse and complex chemical transformations that occur during the roasting of raw coffee beans. This knowledge may help broaden the sensory spectrum of coffee by enabling personalized roast profiles, tailoring aroma and flavor to consumers’ preferences.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.5c01783.

• Information of the coffees used in the study and their respective characteristics (Table S1); confusion matrix for the PLS-DA model (Table S2); probability of model insignificance vs permuted samples (Table S3) (PDF)

A.C.P.: conceptualization, formal analysis, Investigation. C.A.T.: conceptualization, formal analysis. L.W.H.: conceptualization, funding acquisition, Supervision.

This study received funding from the Regional Coffee Growers Cooperative from Guaxupé (Cooxupé), the National Council for Scientific and Technological Development (CNPq) (Grants 302748/2018-0 and 316202/2021-5), the São Paulo Research Foundation (FAPESP) (Grants 14/508673, 20/01064-6, and 22/07406-1), and the Coordination for the Improvement of Higher Education Personnel (CAPES)Finance Code 001. Additionally, the instruments used in this study were provided through funding from ANP/PETROBRAS (Grants 2019/00209-3 and 2019/00210-1, Brazil). The Article Processing Charge for the publication of this research was funded by the Coordenacao de Aperfeicoamento de Pessoal de Nivel Superior (CAPES), Brazil (ROR identifier: 00x0ma614).

The authors declare no competing financial interest.

Published as part of ACS Omega special issue “Chemistry in Brazil: Advancing through Open Science”.