Volatile Constituents of Cymbopogon citratus (DC.) Stapf Grown in Greenhouse in Serbia: Chemical Analysis and Chemometrics

Abstract

The present study investigated the volatile constituents of Cymbopogon citratus (lemongrass) grown in a greenhouse environment in Serbia, marking the first commercial cultivation of the plant for essential oil production in the region. The essential oils and hydrolates obtained through steam distillation were analyzed via gas chromatography–mass spectrometry (GC-MS), and the resulting chemical data were further processed using chemometric methods. This study applied quantitative structure retention relationship (QSRR) analysis, employing molecular descriptors (MDs) and artificial neural networks (ANNs) to predict the retention indices (RIs) of the compounds. A genetic algorithm (GA) was used to select the most relevant MDs for this predictive modeling. A total of 29 compounds were annotated in the essential oils, with geranial and neral being the dominant components, while 37 compounds were detected in the hydrolates. The ANN models effectively predicted the RIs of both essential oils and hydrolates, demonstrating high statistical accuracy and low prediction errors. This research offers valuable insights into the chemical profile of lemongrass cultivated in temperate conditions and advances QSRR modeling for essential oil analysis.

1. Introduction

Cymbopogon citratus (DC.) Stapf., commonly known as lemongrass, is a perennial plant belonging to the Poaceae family. It is rhizomatous, with a large number of glabrous, bright bluish-green, and strap-like leaves (up to 2.5 cm wide, and 90 cm long) [1]. The multiple resources of lemongrass include the leaves as flavoring agents (for teas, soups, and curries, meat dishes, and seafood), as well as a source of fiber (for textile products such as clothes, bags, and hats, as well as for paper production, which is made after the extraction process of the essential oil) [2,3]. However, the most significant economic value is its leaves’ essential oil, which is widely applied in the food industry as a preservative and flavor agent; in aromatherapy, cosmetics, and pharmaceuticals; as well as a natural insecticide and repellent [4,5]. In addition to its essential oil, during C. citratus distillation process, hydrolate is produced as a byproduct that has high additional value and good antioxidant and antimicrobial activities [6,7].

C. citratus is native to South East Asia (India, Thailand, Bangladesh, and Malaysia), and has been well adopted around the globe in areas with tropical and subtropical climates in Africa, Australia, as well as South and North America [3]. In temperate areas, it is often grown as an ornamental plant, but, since it is sensitive to frost, it is necessary to store the bulbous shoot base with roots for wintering [8]. So far, there are no available data on the commercial cultivation of this plant for the production of essential oil in Europe. Bearing in mind the global climate changes, increasing temperatures during the summer, and milder winters, we tried to grow C. citratus in greenhouses in the climatic region of Serbia for three years. The obtained plant material was processed by semi-industrial steam distillation, analyzed by gas chromatography–mass spectrometry (GC-MS), and further processed by chemometric tools to fully evaluate the quality of the obtained essential oil.

The quantitative structure retention relationship (QSRR) concept elucidates the connection between a chemical compound’s observed structure and its predicted physicochemical or biological properties [9,10,11,12,13,14,15,16,17]. Molecular descriptors (MDs) encode the structure of chemical compounds into numerical values, facilitating their explanation [18]. Recent publications have explored the application of QSRR in GC-MS analysis [14,15,16,17,18,19,20,21,22,23,24,25]. Mathematical models representing the relationship between MDs and retention time can be established using various machine learning algorithms [26,27]. In this study, artificial neural networks (ANNs) were employed due to their proven efficacy in previous research [9,28].

This paper aimed to develop a novel QSRR model for predicting the retention indices (RIs) of the chemical compounds in C. citratus essential oil and hydrolates obtained via steam distillation and analyzed by GC-MS. The genetic algorithm (GA) variable selection method and ANN models were utilized for this purpose.

2. Materials and Methods

2.1. Plant Material and Growing Conditions

Cymbopogon citratus (DC.) Stapf. was grown in a greenhouse in the village of Banatska Topola, Vojvodina province, Serbia (45°40′18.25” N, 20°27′53.91” E). According to our knowledge and available data, this is the first time that this plant has been grown commercially in our country.

The planting material was obtained by producing seedlings from seeds purchased via online shop https://planthouse.hr/ (accessed on 3 March 2022) in March 2018. The seeds were sown in March, and the experiment was set up in April of the same year. The experiment was arranged in three repetitions—three greenhouses with an area of 240 m² (8 × 30 m) and approximately 700 plants/greenhouse (plant density 0.8 × 0.6 m) (Figure 1).

The details of the soil conditions are given in

Table 1. Soil chemical properties of top layer (0–30 cm).

pH		CaCO3 ***	Humus ****	Total Nitrogen *****	P2O5 ******	K2O ******
in KCl *	in H2O **
7.46	8.39	0.84%	4.18%	0.268%	30.03 mg/100 g soil	44.15 mg/100 g soil

* Potential acidity was determined in suspension soil with potassium chloride (10 g:25 cm³), potentiometrically with a pH meter; ** active acidity was determined in suspension soil with water (10 g:25 cm³), potentiometrically with a pH meter; *** free calcium carbonate was determined volumetrically with a Scheibler calcimeter; **** Humus content was determined by the oxidation of organic matter using the Tyurin method; ***** total nitrogen content was determined using CHNS elemental analysis after dry combustion using a VarioEL III analyzer (Elementar Analysensysteme GmbH, Hanau-Germany); ****** readily available phosphorus and potassium in the soil was determined by the extraction of ammonium lactate (AL) spectrophotometrically at a wavelength of 830 nm.

, while the weather conditions (average minimal and maximal monthly temperature and insolation) are given in Figure 2. The temperature during summer was harmonized with the outside temperature, but the sides of the greenhouse were opened as needed for ventilation. During the winter, the temperature was kept above 0 °C, using agricultural polyethylene foil to cover the plants. However, during the winter, a significant number of C. citratus leaves withered, but, after removing the dried parts, the plants successfully regenerated for the new growing season.

Fertilization was coordinated with the chemical analysis of the soil; watering was carried out by drip irrigation during the growing season; during the winter, the plants were not watered. Diseases and pests were not observed during cultivation, so there was no need to apply plant protection preparations.

Voucher specimens were determined by Dr. Milica Rat and deposited in the BUNS Herbarium collection (Department of Biology and Ecology, Faculty of Science, University of Novi Sad) under number 2-1406.

2.2. Harvest and Postharvest Processing

The harvest of C. citratus was performed by hand by cutting leaves at 10 cm from the soil. The harvested plant material was dried in a solar dryer until a constant weight. The distillation of the essential oil was performed with a steam distillation system (Institute of Field and Vegetable Crops Novi Sad, Serbia).

Steam distillation (4 h), as a separation process, uses steam from the boiling water generated by a high-pressure boiler (Ventilator Ltd., Zagreb, Croatia), which passes through the plant material in a plumbing stainless-steel distillation vessel (Inox Ltd., Bački Petrovac, Serbia). The vapor of the volatiles was carried via pipelines to a condenser and was further collected in a Florentine flask (Iskra Ltd., Poreč, Croatia). The volatiles were not miscible with water and therefore floated on the surface of the water phase, and could be easily separated by decantation. The collected essential oil was placed into a separatory funnel overnight with the addition of sodium sulphate and, finally, was filtered using filter paper to remove impurities and particulate matter, ensuring a clear sample for further analysis.

However, some volatile compounds were soluble in water and gave it specific organoleptic properties. Additional methods, such as simultaneous steam distillation and extraction by a Likens–Nickerson-type glassware apparatus, make their isolation possible. This system used two round-bottom flasks, one with 400 mL of hydrolate and another with 5 mL of dichloromethane. After 2 h of heating (both flasks had their own heating mantle), volatile compounds (recovered essential oil) were isolated from the hydrolate, and further used for analysis.

2.3. Analysis of Volatile Compounds

The following instruments were used in this study (all Agilent, Santa Clara, CA, USA): HP 5890 series II gas chromatograph combined with flame ionization detector (GC-FID), HP 5973 mass selective detector (GC-MS), and 19091S-433 HP-5MS capillary column fused silica (length, 30 m; film thickness, 0.25 μm; inner diameter, 0.25 mm). The following conditions were used: linearly programmed temperature range, 60–300 °C; a rate of 3 °C/min; carrier gas, helium; inlet pressure, 19.6 psi; constant pressure mode; 1 mL/min flow rate at 210 °C; splitless injections with a volume of 1 μL. The contents of the chemicals of the Cymbopogon citratus essential oils and hydrolates from the three growing seasons are given in percent (%). The linear retention index (RI_exp) was used to identify volatile compounds in relation to C8-C32 n-alkanes and compared with the literature data (RI_lit) according to Adams4 and NIST17 databases, as well as the RI library of the Faculty of Chemistry, University of Belgrade.

2.4. Quantitative Structure Retention Relationship (QSRR) Analysis

The molecular structure datasets representing C. citratus essential oil and hydrolates, obtained through hydrodistillation and analyzed via GC-MS, were provided in the format of concise ASCII files (with the .SMI extension, indicating Simplified Molecular Input Line). These molecular data were sourced from the PubChem database (https://pubchem.ncbi.nlm.nih.gov (accessed on 3 March 2022)). To compute the specific molecular descriptors for each chemical compound [29], we utilized the freely available PaDel-descriptor software, ver. 2.21 [30]. Due to the extensive volume of data for each compound, a genetic algorithm (GA) [31,32] was employed. The calculations were conducted using Heuristic Lab, ver. 3.3.16, 2018 (https://dev.heuristiclab.com/trac.fcgi/ (accessed on 3 March 2022)) to identify the most pertinent molecular descriptors for predicting retention indices (RIs).

In this research, the GA discerned the most suitable molecular descriptors for constructing a robust model to predict the RIs of the compounds present in C. citratus essential oils and hydrolates. We scrutinized the correlation between the descriptors and identified collinear ones through factor analysis.

2.5. Artificial Neural Network (ANN)

This paper utilized a multilayer perceptron (MLP) model, comprising input, hidden, and output layers, owing to its established capability to approximate nonlinear functions [33]. For the ANN modeling, the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm was employed [34]. To enhance the performance of the ANN, both input and output data were normalized. The experimental database was partitioned randomly into three subsets: training (60%), testing (20%), and validation (20%), facilitating ANN modeling [35].

The outcomes of the ANN, including weight values, are contingent upon the initial assumptions of the parameters necessary for constructing and fitting the ANN [36,37]. Various network topologies were explored, encompassing hidden neuron counts ranging from 5 to 20 [38]. The training process involved running the network 100,000 times with random initial values for weights and biases [39]. Optimization relied on minimizing validation errors.

Statistical analysis of the data was primarily conducted using Statistica 10 software [40].

2.6. Statistical Tests of the Model Fit

Statistical tests are instrumental in evaluating a model’s accuracy by assessing the correspondence between real data and the model’s predictions. Utilizing fit metrics enables a comprehensive comparison of all models, facilitating the identification of the most suitable one, as exemplified in studies by Bakshaev and Rudzkis [41] and Maydeu-Olivares [42].

In assessing the efficacy of the higher heating value modeling, several crucial statistical parameters were computed, including the chi-square test (Χ²; Equation (1)), mean bias error (MBE; Equation (2)), mean percentage error (MPE; Equation (3)), root mean square error (RMSE; Equation (4)), and absolute average relative deviation (AARD; Equation (5)).

$$x^2={\displaystyle \frac{{\displaystyle \sum _{i=1}^N(x_{pre,i}}-x_{\exp ,i}{)}^2}{N-n}}$$

(1)

$$MBE={\displaystyle \frac{1}{N}}\cdot {\displaystyle \sum _{i=1}^N(x_{pre,i}-x_{\exp ,i})}$$

(2)

$$MPE={\displaystyle \frac{100}{N}}\cdot {\displaystyle \sum _{i=1}^N\left(\frac{\left|x_{pre,i}-x_{\exp ,i}\right|}{x_{\exp ,i}}\right)}$$

(3)

$$RMSE={\left[{\displaystyle \frac{1}{N}}\cdot {\displaystyle \sum _{i=1}^N{\left(x_{pre,i}-x_{\exp ,i}\right)}^2}\right]}^{1/2}$$

(4)

$$AARD={\displaystyle \frac{1}{N}}\cdot {\displaystyle \sum _{i=1}^N\left|\frac{x_{\exp ,i}-x_{pre,i}}{x_{\exp ,i}}\right|}$$

(5)

where x_exp,i is the experimental value; x_pre,i is the predicted values calculated by the model; N and n are the numbers of observations and constants, respectively.

2.7. Global Sensitivity Analysis

Yoon’s interpretation method was used to determine the relative influence of the MDs on the RI [43]. This method was applied on the basis of the weight coefficients of the developed ANN.

3. Results

The yield of dry plant mass per greenhouse varied, and there was a visible decline with the aging of the crop (

Table 2. Main yield parameters of C. citratus grown in the greenhouse conditions from three seasons.

	1st Season	2nd Season	3rd Season
Dry leaf yield (kg)/greenhouse	102.00 ± 7.90	83.00 ± 1.88	58.00 ± 3.83
Essential oil yield (L)/greenhouse	0.93 ± 0.06	0.77 ± 0.05	0.50 ± 0.05
Wight of 100 mL essential oil (g)	86.63 ± 7.63	86.90 ± 5.00	87.20 ± 8.43

Results are expressed as average value of three measures ± standard deviation (SD).

). The same trend was noted in the case of essential oil yield per greenhouse, while the weight of 100 mL of C. citratus essential oil was between 86.63 and 87.20 g.

A total of 29 compounds were annotated in the essential oil of C. citratus, which accounted for 98.5–99.8% of the essential oil (

Table 3. Chemical composition of Cymbopogon citratus essential oils from three growing seasons.

No.	Compound	RIlit	RIexp	RIpred.	RIerror	R.T. (min)	1st Season	2nd Season	3rd Season
1	6-Methyl-5-hepten-2-one	986	983	1036.071	−53.071	7.270	4.50 ± 0.09	2.50 ± 0.00	2.00 ± 0.08
2	Myrcene	988	988	1050.141	−62.141	7.424	22.50 ± 2.13	15.60 ± 0.91	25.50 ± 0.98
3	cis-β-Ocimene	1032	1034	1041.030	−7.030	8.994	0.40 ± 0.00	0.20 ± 0.02	0.20 ± 0.01
4	trans-β-Ocimene	1044	1044	1041.030	2.970	9.379	0.30 ± 0.02	0.20 ± 0.01	0.10 ± 0.00
5	6,7-Epoxymyrcene	1088	1091	1050.141	40.859	11.115	0.10 ± 0.01	0.10 ± 0.00	nd
6	Rosefuran	1106	1097	1136.960	−39.960	11.342	nd	0.10 ± 0.01	nd
7	Linalool	1095	1097	1137.469	−40.469	11.366	1.10 ± 0.04	0.70 ± 0.07	0.80 ± 0.08
8	Perillene	1002	1099	1164.838	−65.838	11.435	0.40 ± 0.03	0.30 ± 0.00	0.20 ± 0.00
9	allo-Ocimene	1128	1126	1049.858	76.142	12.597	0.10 ± 0.01	nd	nd
10	exo-Isocitral	1140	1143	1125.798	17.202	13.270	0.10 ± 0.00	tr	nd
11	trans-Chrysanthemal	1153	1147	1203.735	−56.735	13.472	0.10 ± 0.01	0.10 ± 0.01	nd
12	Citronellal	1148	1150	1176.408	−26.408	13.597	0.40 ± 0.04	0.30 ± 0.00	0.30 ± 0.02
13	trans-Pinocamphone	1158	1155	1149.154	5.846	13.872	nd	nd	0.10 ± 0.00
14	cis-Isocitral	1160	1161	1149.154	11.846	14.098	0.60 ± 0.04	0.80 ± 0.01	0.90 ± 0.06
15	cis-Pinocamphone	1172	1168	1149.154	18.846	14.448	nd	nd	0.50 ± 0.00
16	Rosefuran epoxide	1173	1173	1229.515	−56.515	14.587	0.10 ± 0.01	0.10 ± 0.01	0.20 ± 0.01
17	trans-Isocitral	1177	1180	1201.643	−21.643	14.884	1.00 ± 0.08	1.30 ± 0.02	1.40 ± 0.13
18	Citronellol	1223	1226	1187.474	38.526	16.913	0.50 ± 0.04	0.40 ± 0.03	nd
19	Neral	1227	1241	1180.087	60.913	17.562	27.10 ± 0.75	32.30 ± 2.28	28.70 ± 0.29
20	Geraniol	1249	1252	1197.480	54.520	18.085	3.10 ± 0.12	2.80 ± 0.05	1.80 ± 0.15
21	Geranial	1264	1271	1180.087	90.913	18.916	34.70 ± 1.40	40.70 ± 3.56	35.60 ± 0.82
22	2-Undecanone	1294	1292	1233.893	58.107	19.820	0.40 ± 0.04	0.10 ± 0.00	tr
23	Thymol	1289	1297	1336.532	−39.532	19.978	nd	0.10 ± 0.00	nd
24	Carvacrol	1298	1301	1212.075	88.925	20.221	nd	0.10 ± 0.01	nd
25	Geranyl acetate	1379	1382	1383.747	−1.747	23.801	0.30 ± 0.00	0.40 ± 0.01	0.40 ± 0.03
26	trans-Caryophyllene	1408	1417	1451.016	−34.016	25.326	0.20 ± 0.02	0.20 ± 0.01	0.30 ± 0.00
27	trans-α-Bergamotene	1432	1433	1403.273	29.727	26.012	0.20 ± 0.00	0.20 ± 0.01	0.20 ± 0.01
28	2-Tridecanone	1497	1494	1371.961	122.039	28.536	0.20 ± 0.01	0.10 ± 0.01	nd
29	Caryophyllene oxide	1582	1581	1536.723	44.277	32.118	0.10 ± 0.00	0.10 ± 0.00	0.10 ± 0.01
	Sum						98.5	99.8	99.3

RI_lit—retention index in the literature; RI_exp—retention index experimentally obtained on HP-5MS capillary column; RI_pred.—predicted retention index. RI_error = RI_exp − RI_pred., nd/not detected, tr—trace.

). The dominant compounds during all three years were geranial (34.7–40.7%) and neral (27.1–32.3%), followed by myrcene (15.6–25.5%).

A total of 37 components were annotated in the C. citratus hydrolates, 2 of which were unidentified (10 largest peaks). The dominant compounds were geranial (26.5–32.6%) and neral (18.4–26.1%), followed by 6-methyl-5-hepten-2-one (17.6–23.7%). A complete analysis of the C. citratus hydrolate is given in

Table 4. Chemical composition of Cymbopogon citratus hydrolates from three growing seasons.

No.	Compound	RIlit	RIexp	RIpred.	RIerror	R.T. (min)	1st Season	2nd Season	3rd Season
1	3-Methyl-2-butenal	784	785	1009.350	−224,350	3.109	0.30 ± 0.00	0.30 ± 0.02	nd
2	Hexanal	801	796	848.331	−52,331	3.288	0.10 ± 0.00	0.10 ± 0.00	nd
3	2,2-Dimethyl-3(2H)-furanone	834	831	898.210	−67,210	3.875	0.40 ± 0.02	0.40 ± 0.04	0.40 ± 0.04
4	3Z-Hexenol	850	847	832.563	14,437	4.126	0.40 ± 0.04	0.70 ± 0.00	tr
5	5,5-Dimethyl-2(5H)-furanone	952	949	909.646	39,354	6.364	0.10 ± 0.00	nd	nd
6	Benzaldehyde	952	959	968.564	−9564	6.614	0.10 ± 0.01	nd	nd
7	6-Methyl-5-hepten-2-one	981	985	1039.808	−54,808	7.368	23.70 ± 2.17	17.60 ± 1.33	20.90 ± 1.19
8	Dehydro-1,8-cineole	995	990	1026.268	−36,268	7.493	3.50 ± 0.21	4.20 ± 0.39	4.00 ± 0.34
9	p-Cymene	1020	1023	1085.259	−62,259	8.605	0.10 ± 0.01	nd	0.10 ± 0.01
10	1,8-Cineole	1026	1030	1006.697	23,303	8.852	0.10 ± 0.01	0.10 ± 0.01	tr
11	cis-β-Ocimene	1032	1033	1011.734	21,266	9.035	nd	nd	0.20 ± 0.01
12	Benzene acetaldehyde	1036	1042	980.087	61,913	9.300	0.20 ± 0.00	0.20 ± 0.02	0.10 ± 0.01
13	2,5,5-Trimethyl-3-hexyn-2-ol #	/	1056	965.121	90,879	9.889	nd	nd	0.20 ± 0.01
14	cis-Linalool oxide (furanoid)	1067	1071	1073.100	−2100	10.358	0.60 ± 0.03	0.60 ± 0.02	0.90 ± 0.07
15	trans-Linalool oxide (furanoid)	1084	1088	1073.100	14,900	10.973	0.60 ± 0.02	0.60 ± 0.04	0.70 ± 0.06
16	Linalool	1095	1100	1152.415	−52,415	11.424	2.80 ± 0.27	2.00 ± 0.03	2.80 ± 0.07
17	trans-2,8-p-Mentha-dien-1-ol	1118	1118	1105.124	12,876	12.221	nd	0.40 ± 0.03	0.30 ± 0.00
18	cis-p-Mentha-2,8-dien-1-ol	1133	1133	1134.732	−1732	12.880	0.30 ± 0.03	0.30 ± 0.03	0.30 ± 0.01
19	Ipsdienol	1140	1145	1217.318	−72,318	13.377	nd	0.10 ± 0.00	nd
20	trans-Chrysanthemal	1153	1149	1143.801	5199	13.519	nd	0.10 ± 0.01	0.10 ± 0.01
21	trans-Pinocamphone	1158	1156	1095.148	60,852	13.876	nd	nd	0.30 ± 0.03
22	p-Mentha-1,5-dien-8-ol	1166	1166	1116.561	49,439	14.259	nd	1.10 ± 0.03	4.60 ± 0.08
23	cis-Pinocamphone	1172	1168	1095.148	72,852	14.471	nd	nd	0.80 ± 0.00
24	Terpinen-4-ol	1174	1173	1114.630	58,370	14.634	nd	nd	0.10 ± 0.00
25	Menthol	1167	1172	1180.735	−8735	14.552	nd	0.10 ± 0.01	nd
26	trans-Isocitral	1177	1183	1174.781	8219	14.978	nd	0.10 ± 0.00	0.30 ± 0.03
27	NI-1 *	/	1183	1207.915	−24,915	15.068	1.60 ± 0.11	nd	nd
28	p-Cymen-8-ol	1179	1184	1053.132	130,868	15.093	nd	nd	0.90 ± 0.01
29	α-Terpineol	1186	1186	1201.723	−15,723	15.121	nd	0.10 ± 0.00	nd
30	NI-2 **	/	1191	1210.783	−19,783	15.377	4.60 ± 0.27	nd	nd
31	Citronellol	1223	1227	1350.247	−123,247	16.972	0.60 ± 0.02	nd	nd
32	Nerol	1227	1229	1228.622	0378	17.023	nd	0.40 ± 0.03	nd
33	Neral	1235	1241	1242.116	−1116	17.609	18.40 ± 1.56	26.10 ± 0.73	20.00 ± 1.31
34	Geraniol	1249	1256	1228.622	27,378	18.202	nd	5.20 ± 0.24	0.10 ± 0.01
35	Piperitone	1249	1253	1222.984	30,016	18.164	5.00 ± 0.20	nd	tr
36	Geranial	1270	1270	1242.116	27,884	18.968	26.50 ± 2.19	32.60 ± 0.02	31.20 ± 2.59
37	Geranyl formate	1300	1302	1278.547	23,453	20.278	0.10 ± 0.01	nd	nd
	Sum						90.1	93.4	89.3

RI_lit—retention index in the literature; RI_exp—retention index experimentally obtained on HP-5MS capillary column; RIpred— predicted retention index. RI_pred.—predicted retention index. RI_error = RI_exp − RI_pred., nd/not detected, tr—trace. ^# The identification was conducted according to the RI library of the Faculty of Chemistry, University of Belgrade: m/z 43 (100), 82 (65), 125 (45), 110 (32), 54 (17), 41 (15), 107 (12), 67 (8), 53 (8), 83 (8); * NI-1 m/z 135 (100), 43 (55), 132 (43), 117 (39), 91 (33), 115 (28), 65 (12), 150 (12), 136 (10), 92 (10); ** NI-2 m/z 59 (100), 94 (96), 79 (89), 91 (46), 77 (29), 93 (26), 43 (21), 119 (15), 41 (11), 134 (11).

.

Figure 3 presents a graphical comparison between the experimentally obtained retention indices (RIs) of the C. citratus essential oil and hydrolate and the retention time indices predicted by the ANN model. The red dots represent the geranial retention indices (RI_exp. and RI_pred. coordinates).

The most significant discrepancies between the experimentally measured retention indices (RIs) and those predicted by the ANN model for C. citratus essential oil during the training cycle were observed with geranial, the major compound. Additionally, discrepancies were found for minor compounds such as allo-ocimene, which was present only in the first season at 0.1% but not detected in the second or third seasons, similar to carvacrol, which was detected only in the second season (0.1%). In contrast, during the validation cycle, another minor compound, 2-tridecanone, also showed notable deviations.

For the C. citratus hydrolates, notable discrepancies between the experimental and ANN-predicted RIs were observed for ipsdienol, cis-pinocamphone, p-cymen-8-ol, geraniol, piperitone, and geranyl formate during the training cycle; 3-methyl-2-butenal and citronellol during the testing cycle; and 2,5,5-trimethyl-3-hexyn-2-ol and geranial during the validation cycle. Except geranial, all other compounds were minor and not crucial for the quality of C. citratus hydrolate.

Consequently, for the RI-prediction ANN model of C. citratus essential oil, GA identified six significant MDs: AATS4v, GATS5m, SPC-5, Sv, BIC4, and MIC0 (

Table 5. MDs for RI predictive model for C. citratus essential oil.

MD	MD Group	MD Meaning
AATS4v	Autocorrelation	Average Broto–Moreau autocorrelation—lag 4/weighted by van der Waals volumes
GATS5m	Autocorrelation	Geary autocorrelation—lag 5/weighted by mass
SPC-5	Chi path cluster	Simple path cluster, order 5
Sv	Constitutional	Sum of atomic van der Waals volumes (scaled on carbon atom)
BIC4	Information content	Bond information content index (neighborhood symmetry of 4th order)
MIC0	Information content	Modified information content index (neighborhood symmetry of 0th order)

). The correlation matrix among these descriptors is depicted in

Table 6. The correlation coefficient matrix for the selected descriptors by GA for C. citratus essential oil.

	GATS5m	SPC-5	Sv	BIC4	MIC0
AATS4v	0.200 p = 0.339	0.094 p = 0.655	0.172 p = 0.411	0.258 p = 0.213	0.038 p = 0.856
GATS5m		−0.076 p = 0.720	0.167 p = 0.425	−0.038 p = 0.857	0.022 p = 0.918
SPC-5			0.095 p = 0.651	0.372 p = 0.067	0.189 p = 0.366
Sv				−0.066 p = 0.754	−0.050 p = 0.811
BIC4					0.176 p = 0.399

Molecular descriptors are given in Table 5; p-value.

.

Similarly, for the RI-prediction ANN model of C. citratus hydrolates, the following MDs were deemed most crucial: AATS0p, MATS2c, MATS5s, GATS3v, SpAbs_Dze, SM1_Dzi, and CIC2 (

Table 7. MDs for RI-prediction model for C. citratus hydrolate.

MD	MD Group	MD Meaning
AATS0p	Autocorrelation	Average Broto–Moreau autocorrelation—lag 0/weighted by polarizabilities
MATS2c	Autocorrelation	Moran autocorrelation—lag 2/weighted by charges
MATS5s	Autocorrelation	Moran autocorrelation—lag 5/weighted by I state
GATS3v	Autocorrelation	Geary autocorrelation—lag 3/weighted by van der Waals volumes
SpAbs_Dze	Barysz matrix	Graph energy from Barysz matrix/weighted by Sanderson electronegativities
SM1_Dzi	Barysz matrix	Spectral moment of order 1 from Barysz matrix/weighted by first ionization potential
CIC2	Information content	Complementary information content index (neighborhood symmetry of 2nd order)

), and their correlation matrix is depicted in

Table 8. The correlation coefficient matrix for the selected descriptors by GA for C. citratus hydrolate.

	MATS2c	MATS5s	GATS3v	SpAbs_Dze	SM1_Dzi	CIC2
AATS0p	−0.153 p = 0.404	−0.230 p = 0.206	−0.224 p = 0.218	−0.005 p = 0.978	0.309 p = 0.085	0.073 p = 0.690
MATS2c		0.085 p = 0.646	0.239 p = 0.188	0.056 p = 0.759	0.087 p = 0.636	0.287 p = 0.111
MATS5s			0.019 p = 0.917	0.223 p = 0.220	−0.050 p = 0.786	0.187 p = 0.306
GATS3v				0.239 p = 0.187	−0.098 p = 0.594	0.292 p = 0.105
SpAbs_Dze					0.259 p = 0.152	0.183 p = 0.315
SM1_Dzi						−0.004 p = 0.985

Molecular descriptors are given in Table 7; p-value.

. As previously noted, the correlation among MDs was scrutinized, and collinear descriptors were identified through factor analysis. In the GA calculation, only one of the correlated descriptors was retained.

The calibration and predictive capability of a QSRR model should be tested through model validation. The most widely used coefficient of determination (r²) can provide a reliable indication of the fitness of a model. Thus, it was employed to validate the calibration capability of the QSRR model.

The ANN technique was used to build RI-prediction models to explore the nonlinear relationship between the RIs and the selected descriptors.

Table 9. ANN model summary (performance and errors), for training, testing, and validation cycles.

Network	Performance			Error			Training Algorithm	Error Function	Activation
	Training	Testing	Validation	Training	Testing	Validation			Hidden	Output
MLP 6-10-1	0.819	0.241	1.000	1274.580	1427.840	2957.379	BFGS 5	SOS	Exp.	Exp.
MLP 7-11-1	0.872	0.746	1.000	1060.173	9060.502	2018.322	BFGS 12	SOS	Log.	Exp.

Performance term represents the coefficients of determination, while error terms indicate a lack of data for the ANN model.

shows the statistical results of the MLP 6-10-1 (for predicting the RIs for C. citratus essential oil) and MLP 7-11-1 (for C. citratus hydrolates) networks.

The Table 9 presents the performance results for two developed Multi-Layer Perceptron (MLP) neural network architectures: MLP 6-10-1 (essential oils) and MLP 7-11-1 (hydrolate). For MLP 6-10-1, the training accuracy was 0.819, while the testing accuracy was markedly lower at 0.241, with the validation accuracy reaching 1.000, indicative of potential overfitting. The model exhibited a training error of 1274.580, a testing error of 1427.840, and a validation error of 2957.379. Both MLPs utilized the BFGS optimization algorithm, with the sum of squares (SOS) as the error function and exponential functions for both hidden and output layers. In comparison, MLP 7-11-1 demonstrated an improved training accuracy of 0.872, a testing accuracy of 0.746, and a validation accuracy of 1.000, also suggesting overfitting. Its training error was 1060.173, the testing error was substantially higher at 9060.502, and the validation error was 2018.322. This model employed the BFGS algorithm, sum of squares (SOS) as the error function, and exponential activation functions in the hidden layers, with logarithmic activation functions in the output layer. Both models exhibited discrepancies between training and testing performance, underscoring potential issues with generalization.

The additional statistical tests to test model accuracy indicated good fitting to the experimentally obtained RI values. For the essential oil, the chi-squared statistic was 3114.39, reflecting the goodness of fit of the model. The root mean square error (RMSE) was 54.68, indicating the average deviation between the predicted and observed values. The mean bias error (MBE) was 7.48, suggesting systematic overestimation, while the mean percentage error (MPE) was 3.85%, denoting a tendency toward overestimation. The sum of squared error (SSE) was 74,745.34, and the average absolute residual deviation (AARD) was 3.85, both highlighting the extent of prediction errors. The coefficient of determination (r²) was 0.82, demonstrating that the model explained 82% of the variability in the data. Skewness was 0.40, indicating a mild rightward asymmetry, and kurtosis was −0.96, suggesting a platykurtic distribution with fewer extreme outliers. The mean value was 7.48, the standard deviation (StDev) was 55.28, and the variance (Var) was 3056.11.

For the hydrolate, the chi-squared statistic was 4500.61, suggesting a less optimal model fit compared to that of the essential oil. The RMSE was 66.03, reflecting a more significant average deviation from the observed values. The MBE was −1.48, indicating a slight underestimation overall, and the MPE was 4.68%, demonstrating a tendency toward overestimation. The SSE was 139,518.88, and the AARD was 4.56, both revealing more significant prediction errors. Skewness was −1.10, reflecting a leftward asymmetry, and kurtosis was 2.95, indicating a leptokurtic distribution with more extreme outliers. The mean was −1.48, StDev was 67.07, and Var was 4498.36.

The predicted RIs, presented in Figure 3, confirmed the adequate prediction capabilities of the constructed ANN by showing the relationship between the predicted and experimental RI values.

A better prediction of RIs was obtained in the training cycle rather than during the testing cycle, which was expected because more chemical compounds’ RIs were used in the calculation compared with in the testing cycles. This is also obvious from Table 6, where the r² values during training set performance are higher than the r² values during the testing cycle.

The obtained results revealed the reliability of the ANN models in predicting the RIs of compounds in C. citratus essential oil and hydrolates determined by GC-MS.

The relative influence of the most important input variables was identified using a genetic algorithm on the RIs or global sensitivity analysis: Yoon’s interpretation method. According to Figure 4a, Sv was the most influential parameter in the RI-prediction model for C. citratus essential oils, with an approximately relative importance of 44.17%, while the influences of GATS5m and MIC0 were 22.15% and 14.83%, respectively. SpAbs_Dze was the most influential parameter for the RI-prediction model for C. citratus hydrolates, with an approximate relative importance of 42.70%, while the influence of GATS3v was 20.37%, as shown in Figure 4b.

4. Discussion

The chemical composition, biological activity, and application of C. citratus have been very well studied [44,45,46,47]. Geranial, neral, and myrcene were the main volatile compounds in C. citratus grown outside the native region in Serbia under greenhouse conditions. The same dominant compounds have been found by other scientists for plants grown under field conditions [48,49,50]. Investigations with C. flexuosus showed that the citral content changes with soil type, altitude, temperature, and rainfall [51]. However, apart from growing conditions, the drying method, distillation method, and conditions also influence essential oil quality [50,52].

There are several studies dealing with the chemical composition of C. citratus hydrolates, and all reporting a fresh aroma with lemon odor, with citral isomers as the dominant compounds [53]. Rodrigues et al. [6], among 15 compounds in C. citratus hydrolate, determined geranial and neral contents of approximately 36.16 and 22.83%, respectively, as the major compounds, while Morales-Aranibar et al. [54], among 12 compounds as dominant, found neral (or β-citral) with 36.15% and geranial (or α-citral) with 33.28% to be the main ones. Additionally, studies showed that C. citratus hydrolate has the potential for use in organic agriculture [54], as well as in the food industry [6,55].

The initial step in QSRR analysis involves estimating and identifying fundamental structural descriptors, which are numerical parameters encoding the structure of chemical compounds obtained through GC-MS analysis [14,19,56]. PaDel-descriptor software, ver. 2.21 was employed to compute these molecular descriptors (MDs), which could represent various aspects of the compounds under investigation and have been proven effective in QSRR studies [57]. Before the genetic algorithm (GA) calculation, factor analysis was conducted to eliminate descriptors with identical or nearly identical values for the examined molecular structures, ensuring only one correlated descriptor remained for GA consideration [58].

A GA was utilized to select the most suitable set of MDs, while the evolution simulation technique was employed to determine the most relevant descriptors [59]. The number of genes in the GA calculation equaled the number of MDs obtained from the PaDel-descriptor, with the initial population of the first generation chosen randomly [60]. The number of genes remained low throughout the GA calculation to maintain a small subset of descriptors, with a minimum 60% probability of generating zero for a gene. The simulation employed crossover (90% probability) and mutation (0.5%) operators. A population size of 100 genes was adopted for the GA, evolving over 50 generations. Evolution ceased when 90% of the generations exhibited the same fitness level.

Moreau–Brotoare spatial autocorrelation descriptors [61] were weighted by van der Waals volumes (AATS4v). This descriptor is determined by the molecular structure and physico-chemical features of atoms [62].

Furthermore, 2D autocorrelation Geary autocorrelation (lag 5/weighted by mass) descriptors are defined by the interatomic distances obtained within the geometry matrix, which is, therefore, determined by the set of atomic characteristics [63].

Moran autocorrelation coefficients are general indices of spatial autocorrelation determined by the weighted atomic property, number of atoms and topological distance between them, and Kronecker delta value [64]. The Moran coefficients utilized in the ANN model were weighted by charges (MATS2c) and I state (MATS5s), as shown in Table 5.

The chi path belongs to the group of connectivity indices that are numerical possibilities of two identical molecules encountering each other and is obtained from the bond accessibilities [65,66,67]. The chi path index used for calculation was the average simple path cluster, order 5 (SPC-5).

5. Conclusions

C. citratus grown under greenhouse conditions in Serbia were used for essential oil and hydrolat production by steam distillation. A total of 29 compounds were detected in the essential oil and 37 in the hydrolate, and geranial and neral were dominant in both. The results demonstrated that the selected molecular descriptors (six for essential oil and seven for hydrolate) were not correlated, as suggested by a correlation coefficient matrix; thus, the descriptors were suitable for QSRR analysis. These descriptors were utilized as inputs for an artificial neural network model (ANN). The ANN model adequately predicted the RIs of the compounds in the C. citratus essential oils and hydrolates. Suitable models with high statistical quality and low prediction errors were derived.