Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2026 Jul 19;91(7):e71308. doi: 10.1111/1750-3841.71308

Vibrational Spectroscopy Predicts Antimicrobial Activity of Orange Peels: A Case Study on Batch‐to‐Batch Variation in Food By‐Product Valorization

Xinyuan Zhang 1, Chi Shu 3, Zhiwei Huang 2,3,✉, Dan Li 1,2,✉
PMCID: PMC13381815  PMID: 42473120

ABSTRACT

Citrus peel is a major agro‐industrial by‐product rich in bioactive metabolites, but batch‐to‐batch variation in antimicrobial activity limits its consistent valorization. This study developed a rapid machine learning‐assisted spectroscopic approach to predict the antimicrobial activity of orange peel by‐products. Fifteen citrus cultivars, including 10 sweet oranges and 5 mandarins, were analyzed using attenuated total reflectance Fourier‐transform infrared spectroscopy (ATR‐FTIR) and Raman spectroscopy. Antimicrobial activity was classified into high‐ and low‐activity groups, and five classification models were developed, including support vector machine, k‐nearest neighbor, decision tree, naïve Bayes, and bagged tree algorithms. ATR‐FTIR spectroscopy showed stronger predictive performance than Raman spectroscopy. The best ATR‐FTIR model was SVM, achieving an accuracy of 0.91, sensitivity of 0.87, specificity of 0.95, precision of 0.93, and F1‐score of 0.90. Its high specificity indicates a low risk of falsely selecting weak antimicrobial batches, which is critical for practical screening. In contrast, Raman‐based models performed less effectively, with the highest accuracy of 0.59 from BT and the highest F1‐score of 0.58 from SVM; the latter showed a sensitivity of 0.83 and a specificity of 0.23. Variable importance analysis identified spectral regions associated with functional groups related to phenolics, flavonoids, and carbohydrates as important contributors to antimicrobial prediction. Ultra‐performance liquid chromatography coupled with quadrupole time‐of‐flight mass spectrometry (UPLC‐Q‐TOF‐MS) further supported the spectral interpretation by showing that flavonoids, phenolic acids, and related metabolites were enriched in high‐activity samples. These findings demonstrate that vibrational spectroscopy combined with machine learning can provide a rapid and scalable screening strategy for evaluating antimicrobial potential in orange peel by‐products.

Practical Applications

This study provides a rapid screening approach to evaluate the antimicrobial potential of orange peel by‐products using vibrational spectroscopy and machine learning. The method could help citrus‐processing and food industries identify promising batches of citrus peel for value‐added applications, such as natural antimicrobial ingredients or food safety‐related product development, while reducing reliance on time‐consuming bioassays.

1. Introduction

The agri‐food industry generates substantial quantities of organic waste. Among these, citrus processing waste accounts for a substantial fraction, generating millions of tons of peel waste worldwide each year (Sharma et al. 2022). Oranges, one of the most widely cultivated citrus fruits worldwide, accounted for approximately 48.82 million metric tons during the 2023/2024 market year, representing half of citrus production (Shahbandeh 2024). Many studies have demonstrated that orange peels possess strong antimicrobial potential, primarily due to their high content of flavonoids, phenolic acids, and essential oils. For instance, Shehata et al. (2021) characterized ethanolic extracts of sweet orange peel, rich in hesperidin and narirutin, and demonstrated their in vitro inhibitory effects against foodborne pathogens. Sabry et al. (2024) reported the antibacterial activity of orange peel extracts against Listeria monocytogenes and Salmonella spp., reinforcing their relevance as natural bio‐preservatives. These findings highlight the potential of citrus waste as a natural antimicrobial source for sustainable food preservation.

Batch‐to‐batch variation is a major challenge in valorizing fruit by‐products because their chemical composition is strongly influenced by cultivar, growing conditions, harvest maturity, and post‐harvest practices (Liu et al. 2012). These factors lead to substantial fluctuations in key antimicrobial constituents, including flavanone glycosides, polymethoxylated flavones, and volatile terpenes, which in turn cause inconsistent antimicrobial efficacy across peel batches (Valente et al. 2010). Such variability complicates quality control and limits reproducibility in downstream applications. Recent studies highlight both the extent of chemical variation and its impact on antimicrobial performance, underscoring the need for robust compositional profiling and predictive tools to ensure consistent bioactivity (Ben Hsouna et al. 2023; Šafranko et al. 2023).

Conventional characterization and antibacterial evaluation of bioactive compounds rely on solvent‐intensive extraction and time‐consuming microbiological assays, which not only hinder the rapid utilization of fresh peels but also exacerbate the impact of batch‐to‐batch variation by delaying timely assessment of their fluctuating chemical and antimicrobial properties (Patra et al. 2022; Zhang et al. 2018). In contrast, vibrational spectroscopy offers a cleaner and non‐destructive alternative (Cebi et al. 2023). Techniques such as attenuated total reflectance Fourier‐transform infrared (ATR‐FTIR) and Raman spectroscopy can rapidly capture molecular fingerprints reflecting the chemical composition of samples without extensive sample preparation procedures (Petersen, Yu, and Lu 2021; Valand et al. 2020).

Therefore, this study hypothesized that ATR‐FTIR and Raman spectral fingerprints can capture chemical differences among citrus peel batches and classify them according to antibacterial activity. Based on this hypothesis, this study aimed to establish a vibrational spectroscopy–based machine learning framework for rapid prediction of the antibacterial potential of citrus peels while accounting for batch‐to‐batch variation. Orange peels (Citrus sinensis) were used as the primary model system, and mandarin peels (Citrus reticulata) were included to extend the scope of the analysis. Supervised machine learning algorithms were used to correlate the spectral fingerprints generated by ATR‐FTIR and Raman spectroscopy, respectively, with the antibacterial data against Salmonella, which is a major foodborne pathogen of global concern, causing widespread gastrointestinal infections and significant public health concerns across the food supply chain (Mkangara 2023). ATR‐FTIR and Raman spectroscopy were compared in terms of model accuracy and chemical interpretability (Hadjiivanov et al. 2021). Variable importance in projection (VIP) analysis was conducted to highlight the key wavenumber regions contributing most to antibacterial prediction, and non‐targeted UPLC‐Q‐TOF‐MS metabolite profiling was employed to support the chemical interpretation of the spectroscopy‐based predictions.

2. Materials and Methods

2.1. Sample Collection and Preparation of Aqueous Peel Extracts

15 citrus cultivars representing diverse geographic origins were included in this study, comprising 10 sweet oranges (Citrus sinensis) and 5 mandarins (Citrus reticulata and Citrus clementina). Cultivar details are listed in Figure 1a and Table S1, and their geographic distribution is shown in Figure 1b. Fruit peels were manually separated from the fruit and cut into approximately 1 cm × 1 cm pieces. The samples were freeze‐dried at −40°C for 48 h using a freeze dryer (Buchi, Switzerland), then pulverized into powder using a high‐speed grinder (Cornell, USA). The resulting powder was sieved through a stainless‐steel mesh (1.70 mm) to ensure a consistent particle size for subsequent analysis. Barnfield and Mandarin Mini citrus peels used for external validation were obtained from independent batches different from those included in the original 15‐cultivar dataset. In addition, two additional independent cultivars, Hongmeiren (HMR) and China Navel (CN), were collected to further evaluate the model performance on unseen samples. The overall analytical framework is systematically detailed in Figure 2, illustrating the comprehensive methodology that links antibacterial activity indicators derived from broad samples, spectral features obtained through Raman and ATR‐FTIR spectroscopy, predictive outcomes generated by machine learning models, and final metabolite profiling validated using UPLC‐Q‐TOF‐MS.

FIGURE 1.

FIGURE 1

Open in a new tab

Geographic origins of orange samples. (a) Geographic distribution of the 10 orange types and 5 mandarin samples collected for antibacterial and spectroscopic analysis. (b) Samples were sourced from major citrus‐producing regions across North America (California, USA), South America (Brazil), Africa (Egypt, South Africa), Europe (Spain), and Asia (China, Israel).

FIGURE 2.

FIGURE 2

Open in a new tab

Data processing and statistical analysis workflow for antibacterial activity prediction and metabolite interpretation.

To prepare the aqueous peel extraction, the freeze‐dried powder was suspended in sterile deionized (DI) water at a ratio of 1:20 (w/v) and incubated in a shaking water bath (Julabo, Germany) at 80°C for 30 min. The mixtures were centrifuged at 12,000 ×g for 10 min, and the resulting supernatants were collected and sterilized through 0.22 µm membrane filters (Sartorius, Germany).

2.2. Antibacterial Assay

S. enterica serovar Montevideo BAA710 purchased from the American Type Culture Collection (ATCC) was revived by streaking onto tryptic soy agar (TSA; Oxoid, USA) plates, followed by two consecutive subcultures. To prepare the bacterial suspension for the experiment, bacterial colonies were scraped off from the TSA plates and suspended in phosphate‐buffered saline (PBS; Thermo Fisher, USA) and adjusted to 0.5 McFarland standard (∼1.5 × 108 CFU/mL). A working inoculum of 1.0 × 106 CFU/mL was obtained by diluting 100 µL of the bacterial suspension in 15 mL of 2× tryptic soy broth (TSB; Oxoid). In 96‐well plates, 100 µL of bacterial suspension was mixed with 100 µL of peel extract or sterile DI water (negative control) and incubated at 37°C for 24 h. Bacteria growth was assessed by measuring the optical density at 600 nm (OD600). To further assess external validity, the BAA710‐trained model was challenged using independent bacterial strains beyond the training target. These included Gram‐negative non‐Salmonella strains, namely Escherichia coli O157:H7 ATCC35150, generic E. coli ATCC15097, and Pseudomonas fluorescens T11, as well as the Gram‐positive strain Staphylococcus aureus ATCC6538, together with three additional Salmonella strains, namely S. Typhimurium ATCC14028, S. Enteritidis ATCC13076, and S. Heidelberg ATCC8326. With the exception of P. fluorescens T11, which was a self‐isolate from our laboratory, all other strains were purchased from ATCC. This design enabled evaluation of model transferability both across species and across Gram type, as well as within the Salmonella genus relative to the trained strain, S. Montevideo BAA710.

The plates were shaken linearly for 1 s at 567 cycles/min with a 3 mm amplitude before measurement to ensure uniform mixing of the well contents. OD600 was recorded in endpoint mode (8 reads per well). Growth inhibition rates were calculated as:

InhibitionRate=1−OD600sampleOD600control×100%

All measurements were performed in triplicate, and the mean values were used for the following analysis.

2.3. Molecular Fingerprinting by Vibrational Spectroscopy

Raman spectra were acquired using a Renishaw InVia micro‐Raman system (Renishaw, UK) with a 785 nm diode laser (10 mW). A 50× objective (NA = 0.75, Leica NPLAN EPI) was used to focus the laser onto the sample. Approximately 10 ± 5 mg of peel powder was recorded over the range of 100–3600 cm− 1 with a step size of 0.6 cm− 1 and 10 s acquisition time. For each sample, spectra were acquired at 5 different points, with 9 replicate measurements collected in total to account for local heterogeneity and instrumental variation, and the spectra were averaged to obtain a representative spectrum for subsequent analysis.

ATR‐FTIR spectra were collected using a spectrophotometer equipped with a diamond ATR crystal (PerkinElmer, USA). Approximately 10 ± 5 mg of peel powder was analyzed in the range of 400–4000 cm− 1. For each sample, spectra were acquired at 5 different points (16 scans were averaged per sample), and each point was measured in triplicate to account for local heterogeneity and instrumental variation. Before each sample, a background spectrum was recorded and automatically subtracted during acquisition. To ensure real‐time accuracy and compensate for instrumental drift, the background was measured every 30 min throughout the analysis.

2.4. UPLC‐Q‐TOF‐MS Analysis of Peel Extracts

To compare compositional differences, three samples with high antibacterial activity (Barnfield, Morocco, and Egypt) and three with low antibacterial activity (Gold Mandarin, Mandarin Mini, and South Africa) were selected for UPLC‐Q‐TOF‐MS metabolomic analysis. Before analysis, the freeze‐dried peel powders were reconstituted in DI water, kept at −20°C overnight to ensure complete rehydration of the freeze‐dried matrix and to promote precipitation of particulates and other poorly soluble, centrifuged at 16,000 ×g for 5 min at 4°C, and filtered through 0.22 µm polytetrafluoroethylene (PTFE) membranes (Sartorius, Germany). Quality control (QC) samples were prepared by pooling equal aliquots from each extract (Chan et al. 2021).

The prepared samples underwent random analysis using a Xevo G2‐XS QTOF MS system (Waters, Milford, MA, USA) coupled with a CORTECS T3 C18 column (2.1 × 100 mm, 2.7 µm; Waters Pacific, Singapore). The column temperature was set at 30°C, with mobile phases consisting of 0.1% formic acid in water and acetonitrile (ACN), at a flow rate of 0.4 mL/min. A 2 µL injection volume was used, with the autosampler maintained at 10°C. The electrospray ionization (ESI) source was operated in both positive (ESI+) and negative (ESI−) modes. MS data were acquired in auto MS/MS mode over an m/z range of 100–1500, at a scan rate of five spectra per second under extended dynamic range conditions. The solution profile was set as follows: 0–2 min, 5% acetonitrile; 2–10 min 5% acetonitrile to 30% acetonitrile; 10–15 min, 30% acetonitrile to 80% acetonitrile; 15–15.5 min, 80% acetonitrile to 95% acetonitrile; 15.5–16.5 min, 95% acetonitrile; 16.5–16.6 min, 95% acetonitrile to 5% acetonitrile; 16.6–20.1 min, 5% acetonitrile.

The raw data were processed using Progenesis QI (Waters, UK). Peak detection, alignment, and deconvolution were performed with default settings, and features with relative standard deviation (RSD) > 30% in QC samples were excluded. For positive ionization analysis, adduct ions of [M+H]+, [M+NH4]+, [M+Na]+, and [M+K]+ were chosen, and [M–H]– and [M–H2O–H]– were selected for negative ionization analysis. Compounds were tentatively identified by comparing accurate mass (±5 ppm) and MS/MS fragmentation patterns with online databases. Differential metabolite analysis was performed between the high and low antibacterial extracts, with significance thresholds set at |fold change| ≥ 1.5 and adjusted p < 0.05 (false discovery rate, FDR‐corrected).

Partial least‐squares discriminant analysis (PLS‐DA) was then employed to minimize the influence of metabolite magnitudes and distinguish between classes and identify variables contributing to class differentiation. ANOVA calculations (p < 0.05) were performed using Progenesis QI, and an S‐Plot was generated in the EZinfo platform (Umeå, Sweden). Key features of the identified components were cross‐referenced against several databases, including the Kyoto Encyclopedia of Genes and Genomes (KEGG), the Human Metabolome Database (HMDB), Phenol‐Explorer, and the METLIN MS/MS libraries. Prior to multivariate modeling, the UPLC‐Q‐TOF peak intensities were normalized and Pareto‐scaled (mean‐centered and divided by the square root of the standard deviation) to reduce the influence of large‐abundance ions while preserving relevant biological variability. Based on the bacteria inhibition rates obtained earlier, the three extracts with high activity and the three with low activity were designated as two representative comparison groups. These six samples were subjected to PLS‐DA to examine whether their metabolic profiles formed distinct clusters and to identify discriminatory metabolites contributing to group separation.

2.5. Data Processing and Statistical Analysis

The OD600 values from triplicate measurements were averaged, and bacteria inhibition rates (%) were calculated relative to untreated controls. All data were processed and statistically analyzed using Prism 9 (GraphPad Software, USA).

Raman and ATR‐FTIR spectra were preprocessed before chemometric analysis. Spectral smoothing was performed using a Savitzky–Golay filter with a third‐order polynomial. A third‐order polynomial baseline correction was then applied to minimize background distortions. Vector normalization was subsequently conducted to standardize intensity across samples (Beier and Berger 2009). Finally, the processed spectra were mean‐centered before multivariate statistical analysis, including principal component analysis. All preprocessing was carried out in MATLAB R2023b (MathWorks, USA).

Principal component analysis (PCA) was first applied to the standardized ATR‐FTIR and Raman spectra to visualize clustering patterns, assess sample variability, and detect potential outliers; for machine learning model development, the number of retained PCs was optimized during inner cross‐validation, with candidate values generated from the training‐set PCA and capped at 20 PCs. To establish relationships between vibrational spectra and antibacterial activity, five supervised classification algorithms, including support vector machine (SVM), k‐nearest neighbors (KNN), decision tree (DT), naive bayes (Bayes), and bagged trees (BT), were compared for their ability to predict antibacterial activity from Raman and ATR‐FTIR spectra. Samples were classified using an inhibition‐rate threshold of 0.53, which was selected based on the central distribution of antibacterial activity values and used to maintain class balance. The final dataset contained 7 high‐activity samples and 8 low‐activity samples. Model performance was evaluated using 30 repeated grouped hold‐out tests with a 70:30 training‐to‐test split at the biological‐sample level. In each repetition, all replicate spectra derived from the same biological sample were assigned to the same split to reduce data leakage. Hyperparameter optimization (Table S3) was performed using grid search, in which predefined candidate hyperparameter values were systematically evaluated for each model. The parameter combination with cross‐validation performance within the training set was selected for final model construction. External validation samples were withheld from all stages of model development, including threshold determination, model training, hyperparameter optimization, and internal performance evaluation, thereby serving as an independent dataset for final model validation.

Model outputs from Raman and ATR‐FTIR datasets were compared based on accuracy, sensitivity, and specificity. Key discriminatory wavenumbers were matched with known functional groups to interpret biochemical relevance. To further interpret spectral–bioactivity relationships, VIP scores were calculated to identify the most influential spectral regions (Kawamura et al. 2017). The dominant vibrational features identified by VIP analysis were subsequently validated through UPLC‐Q‐TOF‐MS metabolite profiling, confirming the presence of flavonoids, phenolic acids, and organic acids associated with antibacterial activity.

To explore the association between metabolites and bacteria inhibition performance, the Spearman rank correlation coefficient (r) was calculated between the average relative intensity of each metabolite and the corresponding inhibition rates across samples (Worley and Powers 2013). The contribution scores were calculated by weighing the Spearman correlation against normalized abundance variation. The compounds were selected based on the combined coefficient value and distribution score.

3. Results

3.1. Variability in Antibacterial Activity Among Citrus Peel Extracts

To assess the antibacterial effect of citrus peel extracts, we quantified the Salmonella growth reduction rates across 10 orange cultivars and 5 mandarin cultivars collected from major citrus‐producing regions worldwide, including North America, South America, Europe, Africa, and the Asia–Pacific region (Liu et al. 2012). This broad geographical coverage captures natural variability in cultivar genetics, climate, and cultivation environments that may influence bioactive composition. For each cultivar, three independent batches were obtained, and each batch was analyzed in triplicate. As shown in Figure 3a, the bacteria growth reduction rates of citrus peel extracts exhibited a unimodal distribution centered around 53%, with a mean of 53.14% and a median of 53.34%, with a 2.96% standard deviation. Based on this central tendency, a threshold of 53% was selected to classify samples into two activity levels while maintaining statistical balance. Accordingly, as illustrated in Figure 3b, samples were ranked by inhibition rate (%) and grouped into high and low antibacterial activity categories. Extracts exceeding 53% were classified as high‐activity, whereas those below this value were designated as low‐activity.

FIGURE 3.

FIGURE 3

Open in a new tab

Distribution and variability of antibacterial activity among citrus peel extracts. (a) Frequency distribution of inhibition rates showing a unimodal pattern. (b) Bacteria growth inhibition rates of orange and mandarin peel extracts from 15 different origins, illustrating both within‐origin variability and clear separation between high‐ and low‐activity cultivars.

3.2. Vibrational Spectral Fingerprints and Machine Learning Prediction

For each of the 15 orange and mandarin cultivars, the peel samples were freeze‐dried at −40°C for 48 h and ground into fine powders. Approximately 0.01 g of peel powder was used for both ATR‐FTIR and Raman measurements. Each cultivar was analyzed in triplicate, and for each replicate, three different surface points were selected, with five repeated scans per point to minimize instrumental variation and improve spectral reliability.

The Raman and ATR‐FTIR spectra of orange peel extracts from 15 orange related cultivars exhibited distinct but complementary vibrational fingerprints (Figure 4a,b). Characteristic Raman bands at 1600–1650 cm− 1 and 1000–1200 cm− 1 indicated aromatic C = C and C–O–C stretching of flavonoids and phenolic glycosides (Edwards 2006), while ATR spectra showed strong O–H (∼3300 cm− 1) and C = O (∼1700 cm− 1) absorptions typical of polyphenols and carbohydrates (Valand et al. 2020).

FIGURE 4.

FIGURE 4

Open in a new tab

Vibrational spectral profiles and machine learning classification performance of citrus peel extracts. (a, b) Average Raman/ATR‐FTIR spectra of 15 orange peel extracts from different origins. (c, d) Classification performance of five machine learning algorithms (SVM, KNN, DT, Bayes, BT) using ATR‐FTIR/Raman spectral features.

Although the major spectral features were broadly shared among cultivars, visual inspection alone was insufficient to determine whether subtle variations in peak intensity, shape, or overall spectral pattern were associated with antibacterial activity. Therefore, supervised machine learning models were further applied to evaluate whether Raman and ATR‐FTIR fingerprints could be used to classify extracts with different antibacterial potentials.

Five binary classification models, including SVM, KNN, DT, Bayes, and BT, were evaluated to predict antibacterial activity from Raman and ATR‐FTIR spectra (Figure 4c,d). For Raman‐based models (Figure 4d), classification accuracies ranged from 0.49 to 0.59, with the BT (0.59) performing the best, showing balanced precision and recall. However, the overall F1‐scores (∼0.46–0.58) suggest moderate discrimination ability, reflecting higher signal variability and noise in Raman spectra compared to ATR‐FTIR models. In contrast, ATR‐FTIR models achieved consistently better performance (Figure 4c), with accuracy between 0.68 and 0.91. The SVM model yielded the highest accuracy (0.91) and strong precision (0.93), reducing the risk of falsely selecting weak antibacterial candidates. Meanwhile, the DT and BT also performed robustly (accuracy 0.81–0.87). These results indicate that ATR spectral data provided more stable features for antibacterial prediction.

3.3. External Validation

The validation dataset was used to assess the internal model. In contrast, an independent dataset was prepared for external validation to evaluate further the robustness and generalizability of the developed machine learning models (Steyerberg and Harrell 2016). External validation was performed using independent batches of citrus peel waste, including Barnfield, Mandarin Mini, HMR, and CN, collected separately from those used in the original dataset, even where some samples represented the same cultivar (Barnfield and Mandarin Mini). The developed prediction model correctly classified the majority of these samples, achieving an overall prediction accuracy of 92% (Figure S2).

Using the same inhibition‐rate threshold of 0.53 established for the BAA710‐based model, samples were classified as high antibacterial when the inhibition rate was above 0.53 and low antibacterial when it was below 0.53. However, strains other than S. Montevideo BAA710, including both other Salmonella serovars and non‐Salmonella bacteria (Gram‐positive and Gram‐negative), were not considered applicable to the present model.

3.4. Feature Contribution and PLS‐DA Clustering

Figure 5a,b display the PLS‐DA score plots derived from UPLC‐Q‐TOF‐MS metabolite data in positive and negative ionization modes. Each point denotes one citrus‐peel sample, colored by geographic origin and grouped by antibacterial activity intensity; ellipses denote 95% confidence regions. PLS‐DA was used primarily as a dimension‐reduction and exploratory visualization approach. The PLS‐DA‐derived three dimension scores were then used as the three‐dimensional score space for PERMANOVA to evaluate group‐level differences between the low‐ and high‐antibacterial activity groups. PERMANOVA indicated significant differences in both positive ionization mode (F = 6.80, P = 0.002) and negative ionization mode (F = 4.15, P = 0.014), suggesting that the two activity groups differed in their overall metabolomic profiles. Given the limited number of samples used for UPLC‐Q‐TOF‐MS analysis, these score plots were interpreted as exploratory visualizations rather than independently validated predictive models. The concordant trends observed across ionization modes provide supportive chemical evidence for the spectroscopy‐based classification results. To enable further feature interpretation across models, VIP scores were calculated on the same dataset. It was to quantify the contribution of each wavenumber to antibacterial prediction across different machine learning frameworks. In Figure 5c, the VIP tables show the wavenumbers most influential in the Raman and ATR‐FTIR models. For Raman spectroscopy, the top VIP values are concentrated around 3200–3550 cm− 1, corresponding to O–H stretching vibrations from hydroxyl groups engaged in hydrogen bonding. These regions reflect the abundance of hydrophilic compounds such as flavonoids, phenolic acids, and sugars (Edwards 2006). In the ATR‐FTIR model, the highest VIP scores appear at 1650–1750 cm− 1 (C = O stretching of carbonyl and ester groups) and 1000–1300 cm− 1 (C–O–C and C–O stretching), which are characteristic of glycosidic and carboxylic structures (Stuart 2004). The strong contribution of these polar functional regions indicates that hydroxyl‐rich and carbonyl‐rich compounds are key spectral markers of antibacterial potency.

FIGURE 5.

FIGURE 5

Open in a new tab

Spectral feature importance and multivariate discrimination of orange and mandarin peel samples. (a, b) UPLC‐Q‐TOF‐based PLS‐DA score (positive, negative) plot of clustering pattern consistent with two different antibacterial effect levels. (c) VIP tables highlighting the wavenumbers most contributing to discrimination in Raman and ATR models.

Machine learning classification based on vibrational spectra successfully distinguished citrus peel samples with high and low antibacterial activities. To further interpret these spectral discriminations and verify their chemical basis, feature importance and metabolite validation analyses were conducted through UPLC‐Q‐TOF‐MS.

3.5. UPLC‐Q‐TOF‐MS‐Based Metabolite Annotation and Activity‐Associated Compounds

Figure S1 and Figure 6b summarize the compound‐class distribution of the 36 tentatively identified compounds in citrus peel extracts. Flavonoids represented the dominant class, accounting for 44.4% of the identified metabolites, followed by terpenoids at 19.4%, coumarins at 11.1%, and smaller proportions of lignans, organic acids, and phenylpropanoids, each contributing approximately 8.3%. This distribution indicates that the metabolite profile of citrus peel extracts was enriched in oxygenated and structurally diverse secondary metabolites, particularly flavonoids and related phenolic compounds. The predominance of flavonoids is chemically relevant because these compounds commonly contain hydroxyl, carbonyl, aromatic, and glycosidic structures, which may contribute to the antibacterial potential of citrus peel extracts. These functional groups are also strongly represented in ATR‐FTIR spectral regions associated with O–H, C = O/C = C, and C–O vibrations.

FIGURE 6.

FIGURE 6

Open in a new tab

Correlation‐based ranking and distribution of compounds associated with antibacterial activity. (a) Top ranked compounds showing the strongest positive associations with antibacterial activity, based on Spearman correlation coefficients and contribution values. (b) Compounds identification, abundance patterns, and chemical classification of citrus peel extracts. (c) Heatmap showing the relative abundance patterns of 36 identified compounds across six citrus peel samples with different antibacterial activities.

In Figure 6a, correlation analysis further identified several compounds strongly associated with bacteria inhibition rates. Citric acid, naringin, and two identified coumarin‐type compounds (Compound 22 and Compound 24) exhibited the highest Spearman correlation coefficients (r = 0.94), followed by corchoionoside C and additional flavonoid‐related metabolites (r = 0.83, r = 0.89). Notably, both organic acids and flavonoid glycosides emerged as major contributors, suggesting that antibacterial performance is driven not by a single dominant compound but by coordinated quantitative variation among functionally complementary compounds. The heatmap further showed that the relative abundance of these compounds differed across citrus peel samples, with highly antibacterial samples generally enriched in several positively correlated compounds compared with weakly antibacterial samples (Figure 6c). These findings support a synergistic model in which acidic stress and phenolic‐mediated membrane or oxidative disruption collectively contribute to antibacterial efficacy (Rami et al. 2021; Coban 2020; Rempe et al. 2017).

4. Discussion

This study established a vibrational spectroscopy–machine learning framework for the rapid evaluation of antibacterial activity in aqueous extracts of citrus peel. Although previous studies have shown that vibrational spectroscopy can effectively reflect the chemical composition of plant‐ and food‐derived matrices, enabling successful applications in compositional profiling and prediction (Bureau 2009; Cozzolino 2022), as illustrated by the use of FTIR‐ATR for rapid prediction of phytochemical composition in fresh fig (Hssaini et al. 2022); and while the antimicrobial activities of citrus‐derived metabolites have been extensively investigated (Coban 2020; Hernández Figueroa et al. 2024); relatively few studies have integrated vibrational spectroscopy with functional antibacterial assessments, particularly in the context of food waste valorization and the development of sustainable screening strategies. By comparing ATR‐FTIR and Raman spectroscopy, we demonstrated that vibrational fingerprints can be directly correlated with biological functionality. The classification analyses showed that ATR‐FTIR achieved higher predictive accuracy and stronger correlations with bacteria inhibition rates than Raman spectroscopy in the rapid screening of antibacterial potential in citrus peel by‐product extracts. The integration of VIP analysis and UPLC‐Q‐TOF‐MS metabolite validation further confirmed that polar bioactive compounds, particularly flavonoid glycosides and organic acids, were the major contributors to antibacterial potency.

The PLS‐DA score plots (Figure 5a,b) revealed a clear separation between high and low antibacterial samples. Despite their largely comparable metabolite compositions, the high and low antibacterial samples could be distinctly separated. This indicates that small but functionally critical differences in the concentrations of key metabolites, rather than broad compositional changes, are responsible for the divergent antibacterial performances. Importantly, this finding highlights a fundamental challenge in valorizing citrus by‐products and other food waste streams: visual inspection or basic compositional measurements cannot reliably identify materials with bioactive potential, as most samples appear macroscopically similar despite underlying chemical variability. Therefore, advanced analytical approaches such as vibrational spectroscopy combined with chemometrics are beneficial to use for discriminating decisive chemical variations, enabling more informed, data‐driven decisions in selecting food waste streams for high‐value applications (Petersen et al. 2021).

The present findings align with prior research demonstrating that flavonoids and phenolic acids from citrus peels possess antibacterial activity (Sinurat et al. 2020; Saleem et al. 2023). Unlike many earlier studies that relied on solvent‐intensive extraction and conventional microbiological assays, our work adopts a solvent‐free, non‐destructive spectroscopic and chemometric approach. Eliminating organic solvents reduces chemical waste, energy consumption, and safety risks, while enabling rapid, scalable, and environmentally benign screening strategies. This solvent‐free framework aligns with cleaner production principles by minimizing resource input and facilitating data‐driven valorization of residues. Organic solvent extraction often enriches fewer polar compounds, including polymethoxylated flavones and lipophilic terpenes, thereby altering the natural compositional balance of citrus peel matrices. In contrast, water extraction preferentially retains polar flavonoid glycosides and organic acids, which are more representative of real food‐processing conditions and are closely associated with antibacterial functionality in aqueous environments (Munir et al. 2024; Usman et al. 2022). Furthermore, our observation that ATR‐FTIR outperforms Raman spectroscopy is consistent with literature indicating that FTIR is more responsive to polar functional groups and functional group accumulation, while Raman is more sensitive to non‐polar bonds (Hadjiivanov et al. 2021; Anshari et al. 2025). Collectively, these results support ATR‐FTIR as an efficient, sustainable tool for high‐throughput screening of bioactive compounds in complex plant matrices, bridging conventional compositional analysis and functional bioactivity prediction. Several physicochemical factors can further explain the superior predictive performance of ATR‐FTIR. First, the antibacterial metabolites identified in this study (flavonoid glycosides, organic acids, and oxygenated coumarins) contain abundant polar functional groups (O–H, C = O, and C–O–C). These vibrations generate stronger IR absorption bands than Raman scattering, making ATR‐FTIR more sensitive to subtle compositional variation (Schulz and Baranska 2007; Lu et al. 2011; Coates 2000). Second, ATR‐FTIR captures bulk chemical vibrations across the entire extract layer in contact with the ATR crystal, which reduces issues such as fluorescence and pigment interference that often degrade Raman spectra of plant matrices (Wei et al. 2015; Petersen et al. 2021). Besides, IR absorption intensities typically show more linear concentration dependence for polar phytochemicals, enabling stronger quantitative relationships between spectral features and bacteria inhibition rates (Valand et al. 2020). In contrast, Raman signals from these molecules are weaker and more susceptible to local heterogeneity and matrix effects (Krysa et al. 2022; Schulz and Baranska 2007). But the lower predictive performance of Raman spectroscopy in the present study does not necessarily indicate a lack of analytical value. Rather, it suggests that Raman spectroscopy was less aligned with the dominant discriminatory signals in this system. Given its different sensitivity to molecular bonding environments, Raman spectroscopy may still provide useful structural information that is not prominently reflected by ATR‐FTIR (Farber et al. 2019; Hanzelik et al. 2025). Therefore, although ATR‐FTIR was more effective as a functional screening tool in this study, Raman spectroscopy may still retain value as a complementary analytical technique for characterizing citrus peel chemistry from a different vibrational perspective.

Nevertheless, several limitations should be noted. The sample diversity included in this study was limited by the citrus varieties and origins available in Singapore retail markets during the collection period. Consequently, the dataset may not fully represent the wider variability present across cultivars, growing regions, and seasonal batches; future studies with larger and more diverse datasets are needed to enhance the robustness (Xu et al. 2023; Zhang et al. 2022; Li 2025). However, the selected two citrus categories were orange and mandarin, as these were the dominant and most consistently accessible commercial types, making them suitable for a controlled comparison of peel antibacterial activity and spectral characteristics. Second, more rigorous metabolomics, isotopic labeling, or bioassay‐guided fractionation would further clarify causal contributions. Future research should move beyond identification toward establishing causal relationships between specific metabolites (or metabolite subclasses) and antibacterial potency, rather than attributing functional effects to broad chemical categories (Cushnie and Lamb 2005; Adetuyi and Farombi 2021; Wang et al. 2019). Third, while the antibacterial effects of the peel extracts were evaluated against multiple Gram‐positive and Gram‐negative bacteria, including several additional Salmonella strains, the present model was trained only on the activity measured against S. enterica serovar Montevideo BAA710. The current framework, therefore, remains target‐specific, and its predictive capacity should not be generalized beyond the trained strain at this stage. In this respect, the study has shown the feasibility of integrating vibrational spectroscopy with machine learning to predict antibacterial activity within a defined biological system. This provides a solid foundation for future model expansion toward multi‐strain or multi‐species prediction of antibacterial effects, thereby advancing more efficient and scalable evaluation strategies for complex food matrix systems.

Taken together, this study demonstrates that integrating ATR‐FTIR and Raman spectroscopy with machine learning offers a rapid, solvent‐free, and data‐driven approach for predicting antibacterial potential of citrus peel extracts. The method supports cleaner production by reducing chemical use, analytical time, and waste generation, and it provides a scalable platform for sustainable valorization of fruit‐processing residues into natural antimicrobial ingredients.

5. Conclusion

This study systematically evaluated batch‐to‐batch variation in the antimicrobial activity of orange peel by‐products and developed a spectroscopy‐assisted ML strategy for rapid activity prediction. Among 15 orange related cultivars from major citrus‐producing regions, clear differences in antimicrobial activity against S. enterica serovar Montevideo BAA710 were observed, supporting the need for batch‐level screening before value‐added utilization. By integrating ATR‐FTIR, Raman spectroscopy, and grouped ML validation, this study demonstrated that spectral fingerprints could be linked to functional antimicrobial performance. Compared with Raman spectroscopy, ATR‐FTIR showed greater practical potential for antimicrobial screening because it provided more reliable classification of high‐ and low‐activity batches, reducing the false selection of weak antimicrobial batches. These results suggest that infrared‐sensitive functional groups may better capture chemical variation associated with antimicrobial activity in the tested orange peel samples. VIP analysis identified spectral regions associated with phenolics, flavonoids, carbohydrates, and related functional groups as important contributors to classification. UPLC‐Q‐TOF‐MS further supported this interpretation, indicating that high‐activity samples were characterized by higher contributions from flavonoids, phenolic acids, and other related compounds. Overall, this study provides a practical framework for rapid, solvent‐free screening of antimicrobial potential in citrus peel by‐products.

From a practical perspective, the proposed spectroscopy‐assisted approach can serve as a rapid preliminary screening tool for citrus peel batches before conventional antibacterial assays are conducted. Antibacterial assays remain necessary for final activity confirmation, but they are relatively time‐consuming and labor‐intensive when large numbers of commercial batches need to be evaluated. Vibrational spectroscopy requires limited sample preparation, uses little or no solvent, and rapidly generates chemical fingerprints that can be used for machine‐learning‐based batch prioritization. Therefore, this method may help identify batches with higher antibacterial potential at an early screening stage and reduce the workload associated with repeated bioassays.

Author Contributions

Xinyuan Zhang: conceptualization, methodology, investigation, writing – original draft. Chi Shu: validation, resources. Zhiwei Huang: supervision, writing – review and editing, resources. Dan Li: conceptualization, writing – review and editing, supervision, project administration, validation.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supplementary Material: jfds71308‐sup‐0001‐SuppMat.docx

JFDS-91-0-s001.docx (789.2KB, docx)

Acknowledgments

This research was financed by a department fund at the Department of Food Science and Technology, National University of Singapore (E‐160‐00‐0015‐01, Food Microbial Safety Research) and National University of Singapore (Suzhou) Research Institute (Key Technology Research and Industrialization of Food for Special Medical Use and Future Foods).

Contributor Information

Zhiwei Huang, Email: biehzw@nus.edu.sg.

Dan Li, Email: dan.li@nus.edu.sg.

Data Availability Statement

Data will be made available on request.

References

  1. Adetuyi, B. O. , and Farombi E. O.. 2021. “6‐Gingerol, an Active Constituent of Ginger, Attenuates Lipopolysaccharide‐Induced Oxidation, Inflammation, Cognitive Deficits, Neuroplasticity, and Amyloidogenesis in Rat.” Journal of Food Biochemistry 45, no. 4: 1–14. 10.1111/jfbc.13660. [DOI] [PubMed] [Google Scholar]
  2. Anshari, R. , Tsuboi M., Sato H., Tashiro K., and Ozaki Y.. 2025. “Raman and ATR‐FTIR Unmask Crystallinity Changes and Carboxylate Group and Vinyl Group Accumulation in Natural Weathering Polypropylene Microplastics.” Scientific Reports 15, no. 1: 2518. 10.1038/s41598-025-85837-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Beier, B. D. , and Berger A. J.. 2009. “Method for Automated Background Subtraction From Raman Spectra Containing Known Contaminants.” The Analyst 134, no. 6: 1198. 10.1039/b821856k. [DOI] [PubMed] [Google Scholar]
  4. Ben Hsouna, A. , Sadaka C., Generalic Mekinic I., et al. 2023. “The Chemical Variability, Nutraceutical Value, and Food‐Industry and Cosmetic Applications of Citrus Plants: A Critical Review.” Antioxidants 12, no. 2: 481. 10.3390/antiox12020481. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Bureau, S. , Ruiz D., Reich M., et al. 2009. “Application of ATR‐FTIR for a Rapid and Simultaneous Determination of Sugars and Organic Acids in Apricot Fruit.” Food Chemistry 115, no. 3: 1133–1140. 10.1016/j.foodchem.2008.12.100. [DOI] [Google Scholar]
  6. Cebi, N. , Bekiroglu H., and Erarslan A.. 2023. “Nondestructive Metabolomic Fingerprinting: FTIR, NIR and Raman Spectroscopy in Food Screening.” Molecules 28, no. 23: 7933. 10.3390/molecules28237933. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Chan, M. Z. A. , Lau H., Lim S. Y., Li S. F. Y., and Liu S. Q.. 2021. “Untargeted LC‐QTOF‐MS/MS Based Metabolomics Approach for Revealing Bioactive Components in Probiotic Fermented Coffee Brews.” Food Research International 149: 110656. 10.1016/j.foodres.2021.110656. [DOI] [PubMed] [Google Scholar]
  8. Coates, J. 2000. “Interpretation of Infrared Spectra, A Practical Approach.” In Encyclopedia of Analytical Chemistry, 5606. John Wiley & Sons Ltd. [Google Scholar]
  9. Coban, H. B. 2020. “Organic Acids as Antimicrobial Food Agents: Applications and Microbial Productions.” Bioprocess and Biosystems Engineering 43, no. 4: 569–591. 10.1007/s00449-019-02256-w. [DOI] [PubMed] [Google Scholar]
  10. Cozzolino, D. 2022. “An Overview of the Successful Application of Vibrational Spectroscopy Techniques to Quantify Nutraceuticals in Fruits and Plants.” Foods 11, no. 3: 315. 10.3390/foods11030315. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Cushnie, T. P. , and Lamb A. J.. 2005. “Antimicrobial Activity of Flavonoids.” International Journal of Antimicrobial Agents 26, no. 5: 343–356. 10.1016/j.ijantimicag.2005.09.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Edwards, H. G. M. 2006. “Spectra–Structure Correlations in Raman Spectroscopy.” In Handbook of Vibrational Spectroscopy 3:1839–1870. 10.1002/9780470027325.s4103. [DOI] [Google Scholar]
  13. Farber, C. , Li J., Hager E., et al. 2019. “Complementarity of Raman and Infrared Spectroscopy for Structural Characterization of Plant Epicuticular Waxes.” ACS Omega 4, no. 2: 3700–3707. 10.1021/acsomega.8b03675. [DOI] [Google Scholar]
  14. Hadjiivanov, K. I. , Panayotov D. A., Mihaylov M. Y., et al. 2021. “Power of Infrared and Raman Spectroscopies to Characterize Metal‐Organic Frameworks and Investigate Their Interaction With Guest Molecules.” Chemical Reviews 121, no. 3: 1286–1424. 10.1021/acs.chemrev.0c00487. [DOI] [PubMed] [Google Scholar]
  15. Hanzelik, P. P. , Gergely S., Abonyi J., and Kummer A.. 2025. “Data Fusion of Spectroscopic Data for Enhancing Machine Learning Model Performance.” Digital Chemical Engineering 17: 100271. 10.1016/j.dche.2025.100271. [DOI] [Google Scholar]
  16. Hernández Figueroa, R. H. , López‐Malo A., and Mani‐López E.. 2024. “Antimicrobial Activity and Applications of Fermentates From Lactic Acid Bacteria—A Review.” Sustainable Food Technology 2, no. 2: 292–306. 10.1039/d3fb00241a. [DOI] [Google Scholar]
  17. Hssaini, L. , Razouk R., and Bouslihim Y.. 2022. “Rapid Prediction of Fig Phenolic Acids and Flavonoids Using Mid‐Infrared Spectroscopy Combined With Partial Least Square Regression.” Frontiers in Plant Science 13: 782159. 10.3389/fpls.2022.782159. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Kawamura, K. , Tsujimoto Y., Rabenarivo M., Asai H., Andriamananjara A., and Rakotoson T.. 2017. “Vis‐NIR Spectroscopy and PLS Regression With Waveband Selection for Estimating the Total C and N of Paddy Soils in Madagascar.” Remote Sensing 9, no. 10:1081. 10.3390/rs9101081. [DOI] [Google Scholar]
  19. Krysa, M. , Szymańska‐Chargot M., and Zdunek A.. 2022. “FT‐IR and FT‐Raman Fingerprints of Flavonoids—A Review.” Food Chemistry 393: 133430. 10.1016/j.foodchem.2022.133430. [DOI] [PubMed] [Google Scholar]
  20. Li, J. , Brenner M., Pierides I., et al. 2025. “Machine Learning and Data‐Driven Inverse Modeling of Metabolomics Unveil Key Processes of Active Aging.” npj Systems Biology and Applications 11, no. 1: 103. 10.1038/s41540-025-00580-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Liu, Y. Q. , Heying E., and Tanumihardjo S. A.. 2012. “History, Global Distribution, and Nutritional Importance of Citrus Fruits.” Comprehensive Reviews in Food Science and Food Safety 11, no. 6: 530–545. 10.1111/j.1541-4337.2012.00201.x. [DOI] [Google Scholar]
  22. Lu, X. , Rasco B. A., Kang D. H., Jabal J. M., Aston D. E., and Konkel M. E.. 2011. “Infrared and Raman Spectroscopic Studies of the Antimicrobial Effects of Garlic Concentrates and Diallyl Constituents on Foodborne Pathogens.” Analytical Chemistry 83, no. 11: 4137–4146. 10.1021/ac2001498. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Mkangara, M. 2023. “Prevention and Control of Human Salmonella Enterica Infections: An Implication in Food Safety.” International Journal of Food Science 2023: 1–26. 10.1155/2023/8899596. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Munir, H. , Yaqoob S., Awan K. A., et al. 2024. “Unveiling the Chemistry of Citrus Peel: Insights Into Nutraceutical Potential and Therapeutic Applications.” Foods 13, no. 11: 1681. 10.3390/foods13111681. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Patra, A. , Abdullah S., and Pradhan R. C.. 2022. “Review on the Extraction of Bioactive Compounds and Characterization of Fruit Industry By‐Products.” Bioresources and Bioprocessing 9, no. 1: 14. 10.1186/s40643-022-00498-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Petersen, M. , Yu Z., and Lu X.. 2021. “Application of Raman Spectroscopic Methods in Food Safety: A Review.” Biosensors 11, no. 6:187. 10.3390/bios11060187. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Rami, E. , Singh A., and Favzulazim S.. 2021. “An Overview of Plant Secondary Metabolites, Their Biochemistry and Generic Applications.” The Journal of Phytopharmacology 10, no. 5: 421–428. 10.31254/phyto.2021.10523. [DOI] [Google Scholar]
  28. Rempe, C. S. , Burris K. P., Lenaghan S. C., and Stewart C. N. Jr.. 2017. “The Potential of Systems Biology to Discover Antibacterial Mechanisms of Plant Phenolics.” Frontiers in Microbiology 8: 422. 10.3389/fmicb.2017.00422. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Sabry, B. A. , Badr A. N., Mohammed D. M., Desoukey M. A., and Farouk A.. 2024. “Validating the Protective Role of Orange and Tangerine Peel Extracts Foramending Food Safety Against Microorganisms' Contamination Using Molecular Docking.” Heliyon 10, no. 6: e27737. 10.1016/j.heliyon.2024.e27737. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Šafranko, S. , Šubarić D., Jerković I., and Jokić S.. 2023. “Citrus By‐Products as a Valuable Source of Biologically Active Compounds With Promising Pharmaceutical, Biological and Biomedical Potential.” Pharmaceuticals 16, no. 8: 1081. 10.3390/ph16081081. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Saleem, M. , Durani A. I., Asari A., et al. 2023. “Investigation of Antioxidant and Antibacterial Effects of Citrus Fruits Peels Extracts Using Different Extracting Agents: Phytochemical Analysis With in Silico Studies.” Heliyon 9, no. 4: e15433. 10.1016/j.heliyon.2023.e15433. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Schulz, H. , and Baranska M.. 2007. “Identification and Quantification of Valuable Plant Substances by IR and Raman Spectroscopy.” Vibrational Spectroscopy 43, no. 1: 13–25. 10.1016/j.vibspec.2006.06.001. [DOI] [Google Scholar]
  33. Shahbandeh, M. 2024. “Leading Orange Juice Producers Worldwide in 2023/2024 by Country (in 1,000 metric tons)*[Graph].” In Statista. [Google Scholar]
  34. Sharma, P. , Vishvakarma R., Gautam K., et al. 2022. “Valorization of Citrus Peel Waste for the Sustainable Production of Value‐Added Products.” Bioresource Technology 351: 127064. 10.1016/j.biortech.2022.127064. [DOI] [PubMed] [Google Scholar]
  35. Shehata, M. G. , Awad T. S., Asker D., El Sohaimy S. A., Abd El‐ Aziz N. M., and Youssef M. M.. 2021. “Antioxidant and Antimicrobial Activities and UPLC‐ESI‐MS/MS Polyphenolic Profile of Sweet Orange Peel Extracts.” Current Research in Food Science 4: 326–335. 10.1016/j.crfs.2021.05.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Sinurat, J. P. , Manalu P., and Br Karo R. M.. 2020. “Antibacterial Activity of Flavonoid‐Rich Fractions of Citrus Maxima Peel Extract.” Stannum: Jurnal Sains dan Terapan Kimia 2, no. 2: 16–21. 10.33019/jstk.v2i2.1977. [DOI] [Google Scholar]
  37. Steyerberg, E. W. , and Harrell F. E. Jr.. 2016. “Prediction Models Need Appropriate Internal, Internal–External, and External Validation.” Journal of Clinical Epidemiology 69: 245–247. 10.1016/j.jclinepi.2015.04.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Stuart, B. 2004. “Organic Molecules.” In Infrared Spectroscopy: Fundamentals and Applications, 71–93. [Google Scholar]
  39. Usman, I. , Hussain M., Imran A., et al. 2022. “Traditional and Innovative Approaches for the Extraction of Bioactive Compounds.” International Journal of Food Properties 25, no. 1: 1215–1233. 10.1080/10942912.2022.2074030. [DOI] [Google Scholar]
  40. Valand, R. , Tanna S., Lawson G., and Bengtstrom L.. 2020. “A Review of Fourier Transform Infrared (FTIR) Spectroscopy Used in Food Adulteration and Authenticity Investigations.” Food Additives & Contaminants Part A, Chemistry, Analysis, Control, Exposure & Risk Assessment 37, no. 1: 19–38. 10.1080/19440049.2019.1675909. [DOI] [PubMed] [Google Scholar]
  41. Valente, L. M. M. , da Paixão D., do Nascimento A. C., et al. 2010. “Antiradical Activity, Nutritional Potential and Flavonoids of the Cladodes of Opuntia Monacantha (Cactaceae).” Food Chemistry 123, no. 4: 1127–1131. 10.1016/j.foodchem.2010.05.074. [DOI] [Google Scholar]
  42. Wang, X. L. , Wu Z. H., Di L., Zhou F. J., Yan Y. M., and Cheng Y. X.. 2019. “Renoprotective Meroterpenoids From the Fungus Ganoderma Cochlear.” Fitoterapia 132: 88–93. 10.1016/j.fitote.2018.12.002. [DOI] [PubMed] [Google Scholar]
  43. Wei, D. , Chen S., and Liu Q.. 2015. “Review of Fluorescence Suppression Techniques in Raman Spectroscopy.” Applied Spectroscopy Reviews 50, no. 5: 387–406. 10.1080/05704928.2014.999936. [DOI] [Google Scholar]
  44. Worley, B. , and Powers R.. 2013. “Multivariate Analysis in Metabolomics.” Current Metabolomics 1: 92–107. 10.2174/2213235X11301010092. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Xu, P. , Ji X., Li M., and Lu W.. 2023. “Small Data Machine Learning in Materials Science.” npj Computational Materials 9, no. 1: 9–42. 10.1038/s41524-023-01000-z. [DOI] [Google Scholar]
  46. Zhang, Q. W. , Lin L. G., and Ye W. C.. 2018. “Techniques for Extraction and Isolation of Natural Products: a Comprehensive Review.” Chinese Medicine 13: 20. 10.1186/s13020-018-0177-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Zhang, W. , Kasun L. C., Wang Q. J., Zheng Y., and Lin Z.. 2022. “A Review of Machine Learning for Near‐Infrared Spectroscopy.” Sensors 22, no. 24: 9764. 10.3390/s22249764. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Material: jfds71308‐sup‐0001‐SuppMat.docx

JFDS-91-0-s001.docx (789.2KB, docx)

Data Availability Statement

Data will be made available on request.

Articles from Journal of Food Science are provided here courtesy of Wiley

RESOURCES