A Novel Technique Using Confocal Raman Spectroscopy Coupled with PLS-DA to Identify the Types of Sugar in Three Tropical Fruits

Abstract

Tropical fruits such as cherimoya, soursop, and pineapple share sugars (glucose, fructose, and sucrose) in common but may differ in the content of other phytochemicals. In the present work, confocal Raman spectroscopy and partial least squares discriminant analysis (PLS-DA) were used to establish a classification model among the three fruits and to evaluate the effect of pre-processing methods on the model’s performance. The Raman spectra showed that glucose was present in the fruits in the 800–900 cm⁻¹ band and the 1100–1200 cm⁻¹ band. While sucrose was present in the bands of 1131.22 cm⁻¹, 1134.44 cm⁻¹, and 1133.37 cm⁻¹ in the three fruits, fructose was present in the bands of 1464.22 cm⁻¹, 1467.44 cm⁻¹, and 1464.22 cm⁻¹ in cherimoya, soursop, and pineapple. The accuracy of the PLS-DA model varied according to the pre-processing methods used. The Savitzky–Golay first derivative method produced a model with 98.69–100% and 100% precision on the training and prediction data, respectively.

1. Introduction

Tropical fruits contain many biologically active compounds such as polysaccharides, polyphenols, vitamins, dietary fiber, and carotenoids [1]. Cherimoya (Annona cherimola) fruit is rich in fermentable carbohydrates, especially glucose and fructose [2]. This species is probably native to the area between Ecuador and Peru and has been cultivated since 1200 BC during the Inca Empire period [3]. Similarly, soursop (Anonna muricata L.) is native to tropical America and Africa [4]; its pulp constitutes more than 80% of the fruit and is composed mainly of water, non-reducing sugars, and carbohydrates [5]. Pineapple (Ananas comosus L. Merr) is one of the most popular fruits in the world, widely distributed in tropical regions, and highly dependent on its attractive aroma and sweet taste [6]. Pineapple is an economically important crop in tropical and subtropical regions, and its sweetness determines its quality [7]. Moreover, its soluble sugars such as sucrose, glucose, and fructose are responsible for the sweet taste of fresh-cut pineapple [6]. However, they also have other phytochemical components that can cause differences among these three fruits.

Traditionally, the most commonly used techniques for quantifying sugars in food are liquid and gas chromatography, nuclear magnetic resonance, infrared spectroscopy, spectrophotometry, polarization measurement, and hyperspectral imaging [8,9], which are limited by their invasiveness, intensive use of resources, and negative environmental impact [10,11]. Therefore, in recent years, there has been an increasing need to explore non-destructive and in situ methods that have been reported to work as well as high-performance liquid chromatography [12]. One such non-destructive method is Raman spectroscopy (RS), that enables to detect the internal quality of fruits and vegetables without destroying them [11,12,13]. In addition, the use of RS is increasing, especially in food safety. This technology makes it possible to detect peak signals by exposing the sample to a laser light [14]. This new technology is based on the detection of frequency changes and variations in the intensity of scattered light when a sample interacts with a laser light source [13], obtaining more selective spectra because it provides narrower spectral bands with abundant and well-resolved chemical information that is easy to interpret [12,13].

Andersen et al. [15] mention that RS has become a robust method to assess the content of individual sugars; for example, previous studies have been reported on the application of RS in the quantification of sugars from aqueous solutions, allowing the discrimination of fructose, glucose, maltose, and sucrose with laser excitation at 785 nm [9,12]. This technique is promising because it does not require pretreatment of fruit samples such as berries [9,15]. In addition, RS provides spectroscopic fingerprints that provide information related to specific chemical bonds or functional groups that are useful for analyzing samples qualitatively and quantitatively [11], which is used with chemometric methods to discriminate foods according to their composition [16,17]. In addition, the combination of RS and chemometric methods, was used for the identification of agricultural product varieties in grains, edible oils, and honey [18]. Also, they were used for the discrimination of foreign fats and oils in milk cream, and yogurt [19]; for the rapid quantification of starch in agricultural products [20]; and for the differentiation of conventional and transgenic cotton seed genotypes [21].

Regarding chemometric tools for food discrimination based on spectral data [16,22], principal component analysis (PCA) and partial least squares discriminant analysis (PLS-DA) are considered robust linear classification methods [23]. PCA is an unsupervised method that reduces the data to find similarities between samples [24]. On the other hand, PLS-DA is a supervised method that classifies samples into predefined categories by establishing an internal relationship between spectra (X-block) and a classifier (y-variable) [24,25]. These two methods, known for their speed and effectiveness in providing clear sample separation, can predict a sample class by a model-based regression [22,26]. Therefore, in this research, the effect of data pre-processing methods on model performance was evaluated, and the applicability of RS combined with PCA and PLS-DA to identify the types of sugar from three tropical fruits was studied.

2. Materials and Methods

2.1. Reagents

Glucose, fructose, and sucrose standards (≥99.5% purity) were bought from Sigma-Aldrich (St. Louis, MO, USA).

2.2. Fruit Sampling

Four kilograms each of soursop (Anonna muricata L.), cherimoya (Annona cherimola), and pineapple (Ananas comosus L. Merr) were obtained from a local market in the province of Bagua, in the Amazonas region of Peru (5° 38′ 21″ S, 78° 31′ 54″ W, 408 m.a.s.l.).

The samples were taken to the Laboratorio de Investigación en Ingeniería de los Alimentos y Poscosecha of the UNTRM, where five fruits (without bruises or marks) of each variety were selected. Then, the fruits were manually separated from the peels, and the pulp alone was placed in 50 mL Falcon tubes and stored in deep freezing conditions (Eppendorf, Premium 0410, Hamburg, Germany). The samples were then lyophilized (Labconco, 710402010 model, Kansas City, MO, USA) at 0.008 bar and −84°C for 18 h [27].

2.3. Raman Spectroscopy (RS)

A Horiba Raman spectrometer system (Horiba Scientific, XploRA plus, Montpellier, France) equipped with a 532 nm and 785 nm laser and an Olympus BX 41 microscope was used in this study [28]. The spectrometer was equipped with a 1200 line/mm grating during the spectral recording, and a 532 nm laser was focused on the sample on the microscope stage through a 50 LWD (long working distance) objective (Olympus, Tokyo, Japan). The Raman scattering signals were detected by a charge-coupled device (CCD) detector, with a detection range from 200 to 1700 cm⁻¹ in extended mode. The measurement was performed with an integration time of 5 s, with 10 spectral accumulations and 50% filter laser power. The spectral data were collected using the LabSpec 6.7.1 (Horiba, France) software (Supplementary Material S1). According to Huang et al. [29], the spectral resolution was 3 cm⁻¹; and the calibration was checked by 520.47 cm⁻¹ line of silicon [10]. Following Kolašinac et al. [10], five replicates of each fruit variety were used in this study to overcome sample inhomogeneity. Ten spectra were recorded for each replicate (50 spectra for each fruit variety), resulting in 150 spectra. The Raman peaks were assigned according to the published literature.

2.4. Chemometric Analysis

Raman spectral data were processed according to Ditta et al. and Gersony et al. [30,31]. The data were then smoothed, and a baseline correction, vector normalization, and substrate removal were performed using the Solo+MIA software 8.1 (Eigenvector, Research, Inc. Wenatchee, WA, USA). Dimensionality reduction of spectral intensities was performed using principal component analysis (PCA) and partial least squares discriminant analysis (PLS-DA).

PCA was performed to form clusters per type of fruit. Hotelling’s T² and Q residual statistics were used to identify and remove the anomalous spectra that were considered outliers [25]. Hotelling’s T² and Q residual statistics were calculated at 95% confidence. To perform the partial least squares discriminant analysis (PLS-DA), all data were divided into the calibration and test sets (75% and 25% of the samples, respectively) using the Kennard–Stone approach [10]. The highest values of sensitivity (SE), specificity (SP), and precision (P) were used to evaluate the goodness of the models based on the training and prediction data sets. Precision (P) indicates what fraction of predictions are positive as a positive class; sensitivity (SE), or true positive rate, refers to the percentage of samples of a given class that the model correctly recognizes as belonging to that class; specificity (SP), or true negative rate, refers to samples not belonging to a given class that the model correctly rejects [32].

3. Results and Discussion

3.1. Characterization of the Raman Spectra of Fruits

Diverse peaks were identified in these fruits with different intensities (Figure 1). The core of the pineapple contains significantly higher amounts of fructose and glucose [33]. Meanwhile, the soursop contains carbohydrates, proteins, folate, calcium, phosphorus, iron, and vitamin C, and [34,35,36] reported that the spectrum of sucrose presents the characteristic bands of α−glucose (847 cm⁻¹) and β−fructose (868 cm⁻¹). In this work, glucose was present with the highest intensity in cherimoya and pineapple, at 855.15 cm⁻¹ and 851.93 cm⁻¹, respectively, and with the lowest intensity in soursop (854.08 cm⁻¹). Schulz and Baranska [36] identified other peaks related to sugars in carrots (1462, 1126, and 840 cm⁻¹).

The assignments of molecule vibration associated with the sugars in tropical fruits were made according to the published literature (

Table 1. Assignments of Raman peaks for sugar standards and cherimoya, soursop, and pineapple.

Assignment 1	Literature	Fructose	Glucose	Sucrose	Cherimoya	Soursop	Pineapple
δ (CH2)	1460 [37]	1473.93	1471.74	1472.81	1464.22	1467.44	1464.22
ρ (CH2)	1346 [37]	1356.80	1359.73	1355.73	1355.73	1352.50	1353.58
ν (CC), ν (CO), β (COH)	1142 [37]	1133.37	1133.37	1133.37	1131.22	1134.44	1133.37
ν (CC), ν (CO)	1074 [37]	1080.73	1081.81	1080.73	1078.58	1085.03	1085.03
CH, COH bending	924 [38]	932.49	929.27	933.57	928.20	928.20	918.53
ν (CH)	867 [38]	859.45	853.00	860.52	855.15	854.08	851.93
β (CCO)	596 [39]	594.12		595.20	601.64		595.12
β (CCO), β (CCC)	527 [39]	536.12		537.19	536.12	535.04	535.04
β (CCC)	465 [39]	456.63	455.55	457.70	453.40	455.55	455.55
β (CCC),β (CCO),β (OCO)	424 [38]	416.88	415.81	416.88	417.95

¹ ν–stretching, β–in-plane bending.

). Glucose is one of the most common hexoses with aldehydes in its structure [37], containing a carbonyl group (C-O stretch) [38]. This sugar was found at 1131.22 cm⁻¹, 1134.44 cm⁻¹, and 1133.37 cm⁻¹ in cherimoya, soursop, and pineapple, respectively. Fructose, also known as ‘fruit sugar’, occurs in fruits and plants in a free state and as a building block of sucrose [37]. The Raman bands at 1472–1454 cm⁻¹, corresponding to CH2 scissor vibrations [39], contain peaks at 1464.22 cm⁻¹, 1467.44 cm⁻¹, and 1464.22 cm⁻¹ in cherimoya, soursop, and pineapple, respectively; nevertheless, the peak corresponding to fructose’s β(CCO) vibration was absent in soursop. Although RS allowed the identification of peaks associated with sugars in the three tropical fruits, it was not possible to quantify them. Therefore, it is also suggested that other techniques be used to complement the study. One technique that could be useful to validate the results is gas chromatography–mass spectrometry (GC-MS) analysis and profiling with the HS-SPME technique, since it is an effective tool due to its ability to recover and characterize compounds as mentioned by Mihaylova et al. [40].

3.2. Multivariate Analysis

Before performing any chemometric analysis, the raw spectra needed to be corrected because they contained fluorescence, as in Figure 2a. For this purpose, the spectra underwent a baseline correction with a seventh-order polynomial fitting [41,42], resulting in the spectra of Figure 2b, where the characteristic peaks of the, soursop, and pineapple chemical composition are distinguished. Raw spectra are available in the Supplementary Data file.

3.2.1. Principal Component Analysis

PCA is a chemometric method that allows us to explore patterns [43] and is almost always the first analysis performed on a multivariate data set [44]. The acquired spectral data set may be disturbed by random noise during acquisition; therefore, pre-processing methods are necessary to improve the PCA results. Pre-processing methods can reduce the random noise and improve the spectral characteristics of interest [45]. Figure 3 shows the results of the four pre-processing methods applied to the cherimoya, soursop, and pineapple spectral data sets. The mean centering method (Figure 3a) produced the least visible changes in the corrected spectra compared to the Savitzky–Golay first (Figure 3b), second (Figure 3c) and third (Figure 3d) derivative method. These effects can be observed in the intensities and positions of the peaks.

The changes in the model performance indicators are presented to highlight the effect of the pre-processing methods on the model [45].

Table 2. Effect of four data pre-processing methods on the PCA model for cherimoya, soursop, and pineapple.

Pre-Processing Method	PC	RMSECV	% Variance Captured Total	Q Residual	Hotelling’s T2
Mean centering	3	254.18	76.05	23.95	76.05
First derivative of Savitzky–Golay	4	44.06	92.59	7.41	92.59
Second derivative of Savitzky–Golay	5	25.86	91.10	8.90	91.10
Third derivative of Savitzky–Golay	5	39.32	81.06	18.94	81.06

reports the root mean square error for cross-validation (RMSECV), which is an internal validation error estimated according to De Almeida et al. [46]. Table 2 shows that when the SG first derivative pre-processing method was used, the PCA model was composed of four principal components that captured 92.59% of the total data variance and 7.41% of the Q residuals (

Table 3. Results of the PCA model with different pre-processing methods.

Pre-Processing Methods	RMSEC	RMSECV	Principal Component Number	Eigenvalue of Cov (X)	% Variance Captured This PC	% Variance Captured Total
Mean centering	180.47	254.18	1	1.29 × 107	39.29	39.29
			2	6.36× 106	19.34	58.64
			3	5.72 × 106	17.42	76.05
First derivative of Savitzky–Golay	31.30	44.06	1	2.17 × 106	68.52	68.52
			2	3.81 × 105	12.02	80.54
			3	2.75 × 105	8.66	89.20
			4	1.08 × 105	3.39	92.59
Secondderivative ofSavitzky–Golay	19.20	25.86	1	5.47 × 105	54.96	54.96
			2	1.90 × 105	19.06	74.01
			3	8.61 × 104	8.65	82.67
			4	4.62 × 104	4.65	87.31
			5	3.77 × 104	3.79	91.10
Thirdderivative ofSavitzky–Golay	29.15	39.31	1	5.12 × 105	47.57	47.57
			2	1.72 × 105	15.96	63.53
			3	8.88 × 104	8.24	71.77
			4	6.25 × 104	5.80	77.57
			5	3.75 × 104	3.48	81.06

). Mata et al. [21] used a plot of Hotelling’s T² vs. Q residuals to eliminate four spectra identified as outliers for the discrimination between transgenic and conventional cotton seeds. Figure 4 was obtained after removing two outliers and shows that cherimoya was separated from the other fruits using the four pre-processing methods. However, none became an outlier, since it was at most 95%.

The mean centering pre-processing method produced a three-component PCA model with a captured variance of 76.05%, which is the lowest captured variance among the models (Table 2). In the score plot, a better sample separation can be observed due to the formation of three well-defined clusters (Figure 5). Cherimoya and soursop, which belong to the Annonaceae family [47], form clusters closer to each other because they share many Raman peaks (including sugars), such that their chemical compositions are similar, unlike pineapple, which belongs to another family (Bromeliaceae) and has a different chemical composition. The separation of the clusters may be caused by the differences in the fruits’ phytochemical compound content, such as polyphenols, vitamins, fiber, and carotenoids. The loading plot of the first principal component produced by the four pre-processing methods (Figure 5) shows more pronounced peaks in the 1000–1200 cm⁻¹ band, where the characteristic glucose peak (common to the three tropical fruits) is found.

3.2.2. Partial Least Square Discriminant Analysis

Since the discriminative power of a classification model is unknown, it is necessary to analyze it using different pre-processing methods [48].

Table 4. Classification results of the PLS-DA model.

Pre-Processing Method	Sweet Fruit	Training Data Set			Prediction Data Set
		P 1 (%)	SE 2 (%)	SP 3 (%)	P (%)	SE (%)	SP (%)
Mean centering	Cherimoya	93.06	93.87	97.98	100.00	100.00	100.00
	Soursop	97.02	97.95	96.97	100.00	100.00	100.00
	Pineapple	100.00	98.00	100.00	100.00	100.00	100.00
First derivative of Savitzky–Golay	Cherimoya	100.00	97.43	100.00	100.00	100.00	100.00
	Soursop	98.68	100.00	98.66	100.00	100.00	100.00
	Pineapple	100.00	100.00	100.00	100.00	100.00	100.00
Second derivative of Savitzky–Golay	Cherimoya	96.68	54.83	98.11	100.00	12.50	100.00
	Soursop	84.44	92.00	83.05	80.00	100.00	75.00
	Pineapple	100.00	89.28	100.00	100.00	100.00	100.00
Third derivative of Savitzky–Golay	Cherimoya	98.13	71.79	98.61	100.00	20.00	100.00
	Soursop	87.94	97.22	86.67	80.59	69.23	83.33
	Pineapple	98.68	100.00	98.67	100.00	100.00	100.00

¹ Precision (P), ² Sensitivity (SE) or true positive rate, ³ Specificity (SP) or true negative rate.

shows the results of the PLS-DA model applied to tropical fruits under the four pre-processing methods studied (all results are in

Table 5. Result of the PLS-DA model with different pre-processing methods.

Pre-Processing Method	Modeled Class	Cherimoya	Soursop	Pineapple
Mean centering	Sensitivity (Cal)	0.959	1.000	1.000
	Specificity (Cal)	0.929	0.949	1.000
	Sensitivity (CV)	0.939	1.000	1.000
	Specificity (CV)	0.909	0.949	1.000
	Class. Err (Cal)	0.0557617	0.0252525	0
	Class. Err (CV)	0.0760668	0.0252525	0
	RMSEC	0.31426	0.273395	0.217022
	RMSECV	0.333692	0.289513	0.229203
	Bias	1.66533× 10−16	1.66533× 10−16	1.11022 × 10−16
	CV Bias	0.00127205	−0.00184987	0.000577824
	R2 Cal	0.554066	0.662499	0.78946
	R2 CV	0.501205	0.622938	0.766066
First derivative of Savitzky–Golay	Sensitivity (Cal)	0.974	1.000	1.000
	Specificity (Cal)	0.972	0.987	1.000
	Sensitivity (CV)	0.897	0.972	1.000
	Specificity (CV)	0.958	0.987	1.000
	Class. Err (Cal)	0.0267094	0.00666667	0
	Class. Err (CV)	0.0721154	0.0205556	0
	RMSEC	0.26207	0.197014	0.198749
	RMSECV	0.311281	0.252941	0.209272
	Bias	−0.00915475	0.00408662	0.00506812
	CV Bias	0.00775709	−0.0145767	0.00681962
	R2 Cal	0.699008	0.822952	0.819861
	R2 CV	0.595265	0.717618	0.800737
Secondderivative ofSavitzky–Golay	Sensitivity (Cal)	0.548	1.000	0.964
	Specificity (Cal)	0.906	0.831	1.000
	Sensitivity (CV)	0.548	0.960	0.929
	Specificity (CV)	0.887	0.831	1.000
	Class. Err (Cal)	0.272976	0.0847458	0.0178571
	Class. Err (CV)	0.28241	0.104746	0.0357143
	RMSEC	0.397685	0.343363	0.242586
	RMSECV	0.400828	0.351006	0.244279
	Bias	0.0964188	−0.131557	0.0351384
	CV Bias	0.0958097	−0.13477	0.0389599
	R2 Cal	0.366397	0.538207	0.740784
	R2 CV	0.356427	0.518102	0.740114
Thirdderivative ofSavitzky–Golay	Sensitivity (Cal)	0.821	0.944	1.000
	Specificity (Cal)	0.889	0.813	0.987
	Sensitivity (CV)	0.821	0.917	1.000
	Specificity (CV)	0.875	0.813	0.973
	Class. Err (Cal)	0.145299	0.121111	0.00666667
	Class. Err (CV)	0.152244	0.135	0.0133333
	RMSEC	0.348922	0.32159	0.208805
	RMSECV	0.361874	0.336387	0.234444
	Bias	−0.00584702	0.000421702	0.00542532
	CV Bias	−0.00528552	−0.00200115	0.00728667
	R2 Cal	0.465947	0.528059	0.801175
	R2 CV	0.425925	0.486149	0.749947

Cal: calibration; CV: cross-validation; RMSEC: root mean square error calibration, RMSECV: root mean square error calibration of cross-validation.

). The mean centering pre-processing method and the first derivative of the Savitzky–Golay (SG) pre-processing method show the best prediction results, with a precision of 100% in all tropical fruits. On the training data set, the precision for mean centering and the first derivative of the SG pre-processing method was 93.06–100% and 98.68–100%, respectively. With the second and third derivatives of the SG pre-processing method, the precision was 80–100% and 84.44–100% on the prediction and training data sets, respectively. These values are higher than those reported by Kolašinac et al. [10], whose precision was 19.35–100% on the training data for classifying five varieties of Balkan paprika. These results varied because of the pre-processing method applied. Moreover, in the study of meat carcass classification using Raman spectroscopy, Logan et al. [32] found that the best PLS-DA model correctly classified 94% of grass-fed versus grain-fed carcasses. According to Table 4, the first derivative of the SG pre-processing method produced the best results in the PLS-DA classification model for the training data set, as it correctly classifies the samples in 98.68–100% of cases (P), the samples belonging to its class in 97.43–100% of cases (SE), and the samples not belonging to its class in 98.66–100% of cases (SP). All indicators for the prediction data set were 100%.

We also analyzed the classification power of the PLS-DA model through the receiver operating characteristic (ROC) curves (Figure 6). The ROC curves of cherimoya, soursop, and pineapple were located closer to the upper left corner of the ROC space, which indicates that the first derivative of the SG pre-processing method produces a PLS-DA model with optimal performance, since it can correctly classify the tropical fruits in the classes to which they belong. In all three fruits, the value of the area under the curve (AUC) was one. These results confirm the ROC curves.

4. Conclusions

Sucrose was detected in the three fruits in a spectral region between 1131 and 1134 cm⁻¹. Glucose was detected with 855.15 cm⁻¹ (cherimoya) and 851.93 cm⁻¹ (pineapple) peaks. Finally, fructose appeared with peaks between 1464 and 1467 cm⁻¹ in the three fruits. These results show the high sensitivity of RS for detecting sugars. Therefore, RS is an emerging technology that can provide robustness to data analysis obtained from conventional technologies such as chromatography or spectrophotometry. However, access to chromatography equipment is expensive and not sustainable.

Likewise, RS complemented by chemometric tools such as PCA and PLS-DA, as used in this study, made it possible to establish a high-precision classification model to discriminate among three fruits, thus being a method that can be used to discriminate fruits based on spectral fingerprints using non-destructive technology. Finally, since the fruits have sugar types in common, it is assumed that the discriminatory efficiency of PLS-DA is based not only on the sugar type but also on the content of polyphenols, vitamins, dietary fiber, carotenoids, and other chemical compounds. However, we recommend further studies to corroborate these findings.