Toward smart and in-situ mycotoxin detection in food via vibrational spectroscopy and machine learning

Abstract

Recent advances in vibrational spectroscopy combined with machine learning are enabling smart and in-situ detection of mycotoxins in complex food matrices. Infrared and spontaneous Raman spectroscopy detect molecular vibrations or compositional changes in host matrices, capturing direct or indirect mycotoxin fingerprints, while surface-enhanced Raman spectroscopy (SERs) amplifies characteristic mycotoxins molecular vibrations via plasmonic nanostructures, enabling ultra-sensitive detection. Machine learning further enhances analysis by extracting subtle and unique mycotoxin spectral features from information-rich spectra, suppressing noise, and enabling robust predictions across heterogeneous samples. This review critically examines recent sensing strategies, model development, application performance, non-destructive screening, and potential application challenges, highlighting strengths and limitations relative to conventional methods. Innovations in portable, miniaturized spectrometers integrated with cloud computation are also discussed, supporting scalable, rapid, and on-site mycotoxin monitoring. By integrating state-of-art vibrational fingerprints with computational analysis, these approaches provide a pathway toward sensitive, smart, and field-deployable mycotoxin detection in food.

Highlights

• •Breakthroughs in smart mycotoxin detection using vibrational spectroscopy are discussed.
• •Portable spectrometers, integrated with could computing and machine learning techniques are explored.
• •Studies on direct and indirect mycotoxins detections are discussed.
• •The non-destructive screening capabilities in mycotoxin detections are highlighted.

1. Introduction

“Food contamination” typically refers to the presence of hazardous substances, such as pathogenic microorganisms or toxic compounds, that pose a risk to food safety and make it unsafe for consumption. These contaminants can be biological, chemical, or physical in nature, posing significant health risks ranging from illness to even death (WHO, 2022; Yamato Scale, 2023). Common biological contaminants in food include pathogens such as Salmonella enterica, Staphylococcus aureus, Escherichia coli O157:H7, and Clostridium perfringens. Concurrently, significant chemical contaminants such as mycotoxins, pesticides, and other extraneous chemicals are also prevalent (Eyvazi et al., 2021) These contaminants can enter food supply chains through various routes, thereby compromising the safety of food products. Moreover, food contamination can have global repercussions, affecting the health of consumers across far-reaching locations. Therefore, to mitigate these risks, regulatory bodies such as the Food and Drug Administration of the USA, the Chinese Pharmacopoeia Commission, and the European Commission have established maximum residue levels (MRLs) and their regulatory guidelines for most food contaminants.

Mycotoxins—naturally occurring secondary metabolites produced by fungi such as Aspergillus, Fusarium, and Penicillium—are particularly concerning among the myriad of contaminants due to their widespread occurrence and potential for severe health impacts, posing significant risks to both human and animal health (Z. Wu et al., 2021). These toxic compounds can contaminate animal feed and human food supplies during postharvest processing, storage, and transportation, primarily due to fungal infections (Kuyu & Tola, 2018). The production of mycotoxins is closely linked to the availability of nutrients, particularly carbon and nitrogen sources, within the host environment. For instance, Aspergillus species thrive on simple sugars, leading to elevated levels of aflatoxins (AFs) production. Furthermore, mycotoxins are resistant to complete elimination during conventional food processing (Caceres et al., 2020; Fountain et al., 2016; Navale et al., 2021). Among the various mycotoxins, AFs, ochratoxin A (OTA), zearalenone (ZEA), fumonisins (FBs), and deoxynivalenol (DON) are particularly concerning due to their frequent occurrence and detrimental health effects.

Aflatoxins (AFs) are a group of mycotoxins primarily biosynthesized by fungi within the Aspergillus genus. Predominant variants include AFB₁, AFB₂, AFG₁, and AFG₂, which commonly contaminate agricultural commodities such as peanuts, tree nuts, figs, maize, and rice (Ajmal et al., 2022)_.. Hydroxylated metabolites of AFB₁ and AFB₂, known as AFM₁ and AFM₂ respectively, are frequently detected in milk and other dairy products (Schrenk et al., 2020). These toxins are highly toxic, with a well-established link to increased risk of hepatocarcinogenesis, particularly when co-exposure occurs with hepatitis B virus infection (Moloi et al., 2024). Global regulatory limits for AFs in food vary: the U.S. Food and Drug Administration (FDA) sets an action level of 20 parts per billion (ppb) for total AFs, while the European Union (EU) has a stricter limit of 5 ppb for AFB₁ specifically (Yao et al., 2024). Common analytical methods for detecting and quantifying AFs include enzyme-linked immunosorbent assay (ELISA) (Yoshinari et al., 2024), chromatographic techniques (Liu et al., 2024), ultraviolet (UV) spectroscopy (Kılıç et al., 2024), fluorescence spectroscopy (Bartolić et al., 2022), and immunochemical assays (Raysyan et al., 2020).

Deoxynivalenol (DON), a mycotoxin belonging to the trichothecene group, is the most widely distributed Fusarium toxin globally. It is commonly found in corn, wheat, oats, and barley, where it causes toxic effects such as feed refusal, gastroenteritis, hemorrhage, immunosuppression, and bone marrow damage (Bianchini & Bullerman, 2014). Regulatory limits for DON, as set by the U.S. FDA, include 10 ppm in grains and grain by-products (88 % dry matter basis), 1 ppm in finished wheat products, 5 ppm in swine feed, and 30 ppm in distiller grains (U.S. Food & Drug Administration, 2010). Similar to the detection of aflatoxins (AFs), the predominant analytical techniques for quantifying deoxynivalenol (DON) in food matrices are ELISA and liquid chromatography equipped with diode array or mass spectrometry detectors (Pedroso Pereira et al., 2019; Tahoun et al., 2021).

Ochratoxins are secondary metabolites produced by certain Penicillium and Aspergillus species (Bui-Klimke & Wu, 2015). Among them, Ochratoxin A (OTA) is the most toxic and predominant, frequently contaminating cereals such as oats, rye, wheat, and barley (J. Wei et al., 2019). Classified as a possible human carcinogen by the International Agency for Research on Cancer (IARC), OTA exhibits immunotoxic, carcinogenic, teratogenic, and nephrotoxic effects. Regulatory limits for OTA vary across food categories within the European Union (EU), with particularly stringent limits for infant foods (maximum 0.5 ppb). Conversely, the U.S. Food and Drug Administration (FDA) has not yet established specific regulatory thresholds for OTA in food products (Gil-Serna et al., 2019). Common analytical methods for detecting OTA include ELISA, lateral flow immunoassays, and fluorescence-detection chromatography (Fadlalla et al., 2020).

In summary, action levels of mycotoxins are typically set in parts per billion (ppb) or parts per million (ppm) to ensure food safety and regulatory compliance. This highlights the need for reliable analytical methods to detect and quantify mycotoxins in food commodities. While conventional techniques, such as ELISA, PCR-based methods, and chromatography-MS, offer high sensitivity and accuracy, their application is often limited by the need for costly, advanced instrumentation, complex sample preparation, and susceptibility to matrix interference. Consequently, there is a growing demand for alternative methods that can overcome these limitations while meeting regulatory requirements.

To overcome these challenges, vibrational spectroscopic techniques, such as mid-infrared (MIR), near-infrared (NIR), and Raman spectroscopy, coupled with machine learning techniques, have emerged as promising alternatives for rapid, specific, sensitive and affordable detection of mycotoxins. Once a predictive algorithm is established, users can obtain automated predictions, either quantitative or qualitative, simply by collecting a spectrum, making these methods particularly user-friendly for field mycotoxin screening applications. Moreover, portable spectrometers and advanced cloud computing techniques have attracted significant attention for future research and practical implementations due to their ease of use, autonomous functionality and field-deployable capabilities (Fig. 1). Although machine learning-powered vibrational spectroscopic techniques show strong potential for mycotoxin detection, most existing reviews remain limited to specific food matrices (e.g., grains) or general algorithm development, with almost none providing systematic comparisons to conventional reference methods. Comprehensive reviews focusing on smart and in-situ applications for mycotoxin detection are still lacking, particularly those addressing chemically spiked (mimicking cross-contaminated) and naturally fungi-infected samples, non-destructive detection strategies, and advanced sensing and sampling approaches for rapid acquisition of food-specific fingerprint signals. Furthermore, the critical role of machine learning in enabling robust feature extraction, noise reduction, and automated and real-time prediction in these contexts has yet to be fully evaluated.

Fig. 1.

Characteristics and future prospects in smart and in-situ mycotoxin detection in food by vibrational spectroscopy.

This review addresses these gaps by focusing recent advancements in vibrational spectroscopic techniques for detecting mycotoxins by MIR, NIR, Raman, and surface-enhanced Raman spectroscopy (SERs) combined with machine learning techniques in different food matrices, showing their potential in detecting mycotoxins at action levels set by regulatory agencies. It critically examines sensing strategies, model development, application performance, non-destructive screening, and key challenges, while highlighting strengths and limitations relative to conventional methods. Moreover, the review explores the feasibility of translating these techniques from laboratory environments to practical field use, emphasizing developments in miniaturized spectrometer technology and the integration of cloud-based data processing. Such rapid and reliable spectroscopic approaches offer significant promise in protecting consumers from the harmful effects of mycotoxins and advancing global food security.

2. Vibrational spectroscopy and machine learning techniques

2.1. MIR, NIR and Raman spectroscopic techniques

Vibrational spectroscopy, comprising IR and Raman techniques, is based on transitions between quantized vibrational energy states of molecules resulting from the interaction between radiation emitted by a light source and the sample material (Fig. 2a-c, example spectra of soybean meal that is susceptible to mycotoxin contamination). Molecular-level interactions are especially critical for identifying mycotoxins, which generate diagnostic spectral signatures due to their unique chemical structures. Vibrational spectroscopic sensors combined with multivariate data analysis, have shown potential to deliver the fingerprinting capabilities, reliability and accuracy of expensive, lab-based instruments in portable, rugged, easy-to-use systems designed for field analysis of food contaminants, including mycotoxins. MIR (400-4000 cm^-1) and NIR (12,500–4000 cm⁻¹) are the most popular IR spectroscopy in studies focusing on mycotoxins (Hackshaw et al., 2020). MIR spectroscopy yields fairly narrow bands corresponding to specific fundamental vibrations (Fig. 2a), describing as infrared “fingerprint” for (bio)chemical substances with high sensitivity and selectivity. For a molecule to be IR active, its dipole moment has to be changed, which requires a lack of symmetry in its structure; therefore, polar groups, such as C O (ketones, aldehydes, carboxylic acids, esters, and amides related to mycotoxins), and O—H (alcohols, phenols, or carboxylic acids associated with mycotoxins) exhibit strong MIR absorption (Yao et al., 2020). These IR-active groups can serve as direct molecular targets for mycotoxin detection, with their vibrational features correlating to both the identity and concentration of the toxin. However, a significant challenge in FT-IR spectroscopy is the strong interferences from water when analyzing mycotoxins in liquid food samples, which can obscure important biochemical information in lipids (3000–3500 cm⁻¹) and Amide I (∼1650 cm⁻¹) regions (Hackshaw et al., 2020). Techniques such as lowering the pathlength by using the attenuated total reflectance (ATR) sampling technique, dehydrating the sample, subtracting a water spectrum for the acquired sample spectrum, and using a D2O solution can address this issue.

Fig. 2.

Spectra of soybean meal collected using different spectroscopic techniques: (a) MIR, (b) NIR, and (c) Raman Spectrometers (Rodriguez-Saona et al., 2020); Typical settings for mycotoxins screening of (d) attenuated total reflection (ATR) in MIR spectroscopy, (e) diffuse reflection in NIR spectroscopy, and (f) transmission in NIR spectroscopy; Representative spectra of direct aflatoxins screening by (g) MIR and (h) Raman spectroscopy.

In contrast to the sharper “fingerprinting” bands produced by MIR spectroscopy, NIR spectroscopy is based on overtone and harmonic molecular vibrations, producing broader bands (Fig. 2b) that typically exhibit lower sensitivity for the direct detection of contaminants like mycotoxins (Hackshaw et al., 2020). However, these broader features confer unique advantages for quantitative analysis, particularly in high-moisture and liquid food samples, due to NIR's lower water absorptivity (Beć et al., 2020). This is especially important for mycotoxin monitoring in commodities like grains and dairy, which are prone to contamination. The low absorptivity in the NIR region enables the use of higher-penetration probe radiation beams, facilitating direct (in-line) and rapid transmission through intact food materials (Inagaki et al., 2017). This enhanced penetration depth is vital for detecting internally localized mycotoxins in bulk commodities. Additionally, NIR spectroscopy can sample larger volumes and heterogeneous areas, providing more representative spectra by averaging over variable distributions, in contrast to MIR and Raman techniques (Pudełko et al., 2020). This spatial coverage effectively addresses the heterogeneous distribution of mycotoxins in agricultural matrices. Consequently, these advantages have led to the development and implementation of NIR spectroscopy for indirectly monitoring mycotoxins by assessing compositional changes in contaminated agri-food products.

Raman spectroscopy is based on inelastic scattering, where an energy exchange results in a shift in the wavelength of scattered radiation (Yao et al., 2020). However, spontaneous Raman signals are inherently weak, with only 1 in 10⁸ photons undergoing inelastic scattering, as most light undergoes elastic scattering at the incident light frequency. This inherent weakness requires sensitive detection strategies for identifying trace levels of mycotoxins (Hackshaw et al., 2020). Despite this, Raman spectroscopy is advantageous for screening mycotoxins due to its reliance on fundamental vibrations associated with specific chemical functional groups, which enables precise fingerprinting and qualitative analysis of mycotoxins. For a molecule to be Raman active, it must undergo a change in polarizability during symmetric vibrations, a property that can be directly exploited for toxin detection. Raman spectroscopy thus complements MIR spectroscopy: whereas MIR (Fig. 2a) produces intense signals from polar functional groups due to dipole moment changes, Raman (Fig. 2c) is more sensitive to backbone structures and symmetric bonds governed by polarizability changes (Hashimoto et al., 2019). Furthermore, Raman offers superior spatial resolution (∼1 μm²) compared to infrared methods (∼10 μm²), enabling more detailed mapping of mycotoxin distribution within heterogeneous food matrices. Its low water absorptivity also minimizes interference in aqueous samples, giving it an edge over MIR and NIR for analyzing liquid foods such as dairy products with little or no sample preparation. An additional advantage is its ability to perform non-invasive measurements through transparent packaging, such as glass (Li-Chan, 2010).

To overcome the weak spontaneous Raman signals associated with mycotoxin detection, longer acquisition times can be used for spectral collections; however, this approach may risk sample damage due to prolonged exposure to the high-powered laser (X. Yuan et al., 2018). An alternative method to amplify the Raman signal is surface-enhanced Raman spectroscopy (SERs), which utilizes nanoscale corrugated metal surfaces (typically gold or silver) to support plasmon resonances. This technique can greatly enhance the Raman signal by up to ∼10⁸ or higher, facilitating the detection of trace mycotoxin contaminants and enabling the identification of otherwise undetectable toxins (Langer et al., 2020). Another challenge in employing Raman spectroscopy for screening mycotoxins is fluorescence interference from the food matrix, which causes baseline drift and can obscure the mycotoxin fingerprinting information. Overcoming this interference is crucial for accurate mycotoxin analysis. Solutions include performing mathematical subtraction, switching to a longer-wavelength laser (e.g., 1064 nm), applying SERS technology, or pre-treating the sample with the laser beam (referred to as “photobleaching” or “bleaching”) (D'Acunto, 2019; Hackshaw et al., 2020). Recent studies have shown significant progress in using spontaneous Raman and SERS for fingerprinting mycotoxins, underscoring their potential for on-site and real-time trace detection in food safety assurance. Overall, MIR, NIR, and Raman/SERs each offer distinct advantages for mycotoxin detection. MIR provides sharp, high-sensitivity fingerprinting of polar functional groups from mycotoxin structures; NIR allows deeper penetration of high-moisture and heterogeneous samples for quantifying mycotoxins; and Raman and SERs technique offer detailed mapping, high spatial resolution, low water interference and ultra-sensitive detection through SERs, making it especially suitable for liquid matrices and non-invasive measurements of mycotoxins. Together, these complementary features highlight the versatility of vibrational spectroscopy for diverse food safety applications. Compared to conventional methods for detecting mycotoxins, vibrational spectroscopic techniques offer several additional advantages (Fig. 1). They are highly portable, enabling field-based or on-site analysis compared to bulky MS instrumentation; non-destructive, allowing direct measurement of samples without extensive preparation; capable of high-throughput screening, particularly for bulk or heterogeneous samples using NIR; and provide rapid data acquisition (around 10 s for NIR and SERs), supporting near real-time monitoring. Furthermore, they are relatively environmentally friendly, avoiding the use of solvents or reagents commonly required in chromatographic or immunoassay methods. When combining with machine learning, these techniques can further enhance data analysis and predictive capabilities.

2.2. Machine learning techniques

The spectrum generated from spectroscopic techniques requires the development of robust algorithms for identifying biomarkers related to mycotoxins within the complex food matrices (Abraham & Kellogg, 2021). Manual calibration is particularly challenging due to overlapped noises, peaks and intrinsic chemical variations within the spectrum. To address these challenges, multivariate data analysis has been incorporated into spectra analysis, enhancing the identification process and future smart implementation (Goyal et al., 2024). Pre-processing and transformation techniques are essential to minimize undesirable effects on spectral data and increase the reliability of contamination detection models (Mousa et al., 2022). Common pre-processing and transformation techniques that used in mycotoxins detection are illustrated in Fig. 3, encompassing centering and scaling, normalization, detrending, baseline correction, smoothing, derivatives, MSC (Multiplicative Scatter Correction) and SNV (Standard Normal Variate). However, there are no criteria for selecting the optimal pre-processing and transformation techniques in mycotoxins detection yet, likely due to the varying interactions between specific mycotoxins and different food matrices, thus a comprehensive understanding of the data is recommended to optimize and prevent overprocessing (Pan et al., 2022).

Fig. 3.

Common data processing and transformation techniques, unsupervised and supervised machine learning algorithms for screening mycotoxins using vibrational spectroscopy.

Chemometric algorithms comprise unsupervised and supervised learning (Fig. 3), that provide maximum relevant chemical/biochemical information by analyzing multivariate data, with the goal of identifying patterns within the data to predict the properties or classify unknown samples based on their chemical profile, allowing for regression, categorization and classification of samples. In mycotoxin analysis, these algorithms specifically identify subtle patterns that indicate contamination levels or the presence of specific toxin classes within complex spectral data. Unsupervised learning explores the data to find intrinsic structures without the consideration of class labels, which is useful for identifying intrinsic groups in the data, suspicion or general exploration (Yao et al., 2024). Most commonly unsupervised learning techniques that include principal component analysis (PCA) and hierarchical clustering analysis (HCA) (Peris-Díaz & Krężel, 2021). On the other hand, supervised learning discovers patterns in the data that relate data attributes to a target (class) attribute, utilizing to predict target attribute values in future data instances. Within mycotoxin detection, the target attribute is typically the presence/absence, concentration, or type of mycotoxin. Supervised learning techniques, including classification and regression, involve several key steps when integrated with spectrometers for detecting mycotoxins. This process begins with the selection of a training set and an external validation set. The training set consists of samples with known class labels, such as mycotoxin positive vs mycotoxin negative. The training set is also used to optimize parameters characteristic. Following this, variable selection is performed to retain informative variables related to mycotoxins while eliminating those that introduce noise or lack discriminative power. Common techniques for variable selection include forward stepwise, backward stepwise, and genetic algorithms. A predictive algorithm is then constructed using the training set, establishing relationships between spectroscopic variables and the class or levels of the mycotoxin contamination in the samples, with internal validation. Finally, the algorithm is validated using an independent set of data, which helps assess its performance in detecting mycotoxins in new, unseen samples.

Different supervised classification methods vary in their approach and their performances are influenced by inherent data patterns. For example, Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA) are classical Bayes classification methods based on Bayes' theorem, which assume feature independence within each class. LDA assumes linear separability between classes, such as distinguishing aflatoxin contamination levels where class 1 represents zero or below the cut-off and class 2 represents levels above the cut-off) (Kim et al., 2023a, Kim et al., 2023b). The efficacy of LDA in mycotoxin detection is often challenged by matrix complexity in food samples, where co-occurring metabolites can obscure spectral signatures of low-concentration toxins (Shi et al., 2024). In contrast, QDA allows for quadratic boundaries, accommodating differing class distributions, such as classifying almonds based on five distinct aflatoxin B1 concentrations (Kim et al., 2023a; Mishra et al., 2024). To address collinearity issues in the data, methods such as Soft Independent Modeling of Class Analogy (SIMCA) (Yao et al., 2024), Artificial Neural Networks (ANNs) (Camardo Leggieri et al., 2021), Support Vector Machines (SVM) (H. Zhu et al., 2024), and Random Forest (RF) (Aref et al., 2024) are frequently employed. These techniques are well-suited for capturing complex, non-linear relationships between spectroscopic data and mycotoxins levels, delivering robust performance. Additionally, regression-based classification techniques, such as Partial Least Squares Discriminant Analysis (PLS-DA) (Lotfy et al., 2021) and Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) (Yao et al., 2025), are commonly used to model the relationship between spectroscopic data and mycotoxin-related class labels. These methods leverage regression to enhance classification accuracy in complex multivariate datasets. Deep learning architectures now enable real-time identification of masked mycotoxins by extracting latent patterns from Raman spectra, significantly improving early warning capabilities in grain supply chains(Y. Ran et al., 2025). Additionally, integrating multi-model approaches with single-cell spectroscopy could revolutionize on-site monitoring of toxigenic fungal phenotypes, addressing the urgent need for rapid diagnostics in climate-stressed agricultural systems(L. Liu et al., 2025).

In practical applications, the primary question often concerns obtaining a quantitative answer, how much mycotoxin is present in food. Regression techniques are widely utilized for this purpose by modeling the relationship between input variables, spectroscopic data and mycotoxin concentrations, enabling accurate quantification. Critically, the co-occurrence of multiple mycotoxins in agricultural products (e.g., aflatoxins, fumonisins, and alternariol) necessitates methods capable of resolving complex spectral overlaps and synergistic matrix effects, which directly impact quantification accuracy (C. Wei et al., 2024). Multiple Linear Regression (MLR) is commonly employed for datasets with linear relationships, using the Ordinary Least Squares method to minimize the sum of squared residuals and estimate the relationship between spectroscopic data and mycotoxin concentrations (Peromingo et al., 2020). However, MLR cannot handle missing data well, which may yield unstable estimates with highly correlated predictors. MLR also requires at least K + 1 rows, which is challenging for aflatoxin sensor data, where K often exceeds 1000. As alternatives, linear regression methods such as Partial Least Squares Regression (PLSR) and its variant, Orthogonal Partial Least Squares Regression (OPLSR), are highly effective for handling high-dimensional sensor data. These methods reduce dimensionality and focus on predictive variation, which have been frequently reported in recent studies for mycotoxin detections (Kim et al., 2023a; Ong et al., 2022). However, due to the complexity of the biochemical information obtained by spectrometers, spectroscopic data can exhibit non-linear relationships with mycotoxins. To address this, non-linear regression methods, such as Support Vector Regression (SVR) that leverages kernel-based techniques, Random Forest Regression (RFR) using decision tree ensembles, and Artificial Neural Network Regression (ANNR) utilizing neutral networks, are employed to deal with non-linearity allowing for a better fit of the data. Though many recent publications (Camardo Leggieri et al., 2021; Inglis et al., 2024) have employed non-linear techniques for detecting mycotoxins concentrations, these methods are often criticized for being “black-box” algorithms, which can limit their interpretability. Moreover, the selection of classification and regression methods is influenced by factors such as the characteristics of the data, the complexity of mycotoxin interactions, and the desired level of prediction accuracy. Consequently, recent research has employed various techniques to compare their performance, aiming to identify the most appropriate approach for specific applications. Ultimately, method choice must balance regulatory compliance (e.g., detection limits meeting EU thresholds for infant foods) with technical feasibility, especially for rapid on-site screening versus laboratory confirmation (Liu et al., 2024).

3. Mycotoxin analysis by MIR spectroscopy

The first study to propose the potential of FT-MIR for detecting fungal contamination was published in 1992 (Greene et al., 1992). Since then, researchers have explored FT-IR for analyzing mycotoxin contamination, with recent studies incorporating advanced machine learning techniques, as summarized in Table 1. The production of mycotoxins involves the metabolization of nutrient sources such as carbohydrates, proteins, and fats, leading to compositional changes in host crops. Consequently, previous studies have examined the sensitivity and specificity of detecting mycotoxins both directly through chemical spiking and indirectly through these compositional changes, with respect to their maximum regulatory limits. The ATR sampling technique has been widely used in mycotoxin screening (Fig. 2d), where ground (destructive) or liquid samples are simply placed on ATR elements (approx. 2-3 mm in diameter), allowing analysis of 2-3 μm of the sample for single Reflection ATR (Freitag et al., 2022a). The ease of sample handling by ATR has simplified MIR spectra acquisition and benefited the applications of MIR in detecting mycotoxins. To improve the spectral reproducibility of the ground samples, sieving fractions were often included in the study to standardize the particle size for screening mycotoxins (Kos et al., 2007). In addition, a subsampling step is often introduced in the FT-MIR mycotoxin analysis (De Girolamo, von Holst, et al., 2019a).

Table 1.

Summary of detecting mycotoxins by using MIR spectroscopy.

Mycotoxins	Instrument	Food/Matrix	Chemometrics	Results	Reference
DON AFB1	Benchtop FT-MIR	Maize Peanut	Bagged decision tree	Accuracy: 79 % (DON, Cut-off value: 1.75 ppm) and 85 % (DON, 0.5 ppm) for maize Accuracy: 77 % (AFB1, 8 ppb) for peanuts	(Kos et al., 2016)
AFs	Portable FT-MIR	Peanut	SIMCA OPLS-DA PLSR	Accuracy: 89.6 % (SIMCA), 86.2 %(OPLS-DA) with a cut-off value of 3 ppb Rv = 0.85 RMSEP = 96.02 ppb	(Yao et al., 2025)
AFs	Portable FT-MIR	Peanut	SIMCA	Accuracy:100 % with a cut-off value of 30 ppb	(Yao et al., 2024)
AFs	Benchtop FT-MIR	Extract of Chicken feed and food grains	PLSR	ACN extract: R2cv = 0.99 75 % MeOH extract: R2cv = 0.99	(Salisu et al., 2022b)
AFs	Benchtop FT-MIR	Brown rice	LDA PLSR	Accuracy = 90.6 % (AFs: ≤5 ppb, 5-300 ppb, >300 ppb) AFs (0-2363.6 ppb): Rv = 0.95, RMSEP = 232 ppb	(F. Shen et al., 2018)
AFs	Benchtop FT-MIR	Peanut	A discriminant analysis approach	Accuracy = 98.6 % (AFs: ≤20 ppb, 20-1200 ppb, >1200 ppb)	(Kaya-Celiker et al., 2014)
AFB1	Benchtop FT-MIR	Peanut oil	PLS-DA SVM	Accuracy: 94.6 % (Presence or absence of aflatoxin, PLS-DA) Accuracy: 98.2 % (Presence or absence of aflatoxin, SVM)	(Song et al., 2021)
AFs	Benchtop FT-MIR	Maize	KNN, LDA, PLS-DA	Accuracy: 100 %, 92 % and 92 % (KNN, LDA and PLS-DA) (AFs: <20 ppb, 20-200 ppb, 300-450 ppb, 550-700 ppb, >850 ppb)	(Lee et al., 2015)
DON	Benchtop FT-MIR	Wheat	MLR	Rv = 0.87 RMSEP = 1.90 ppm SEP = 1.65 ppm	(Abramović et al., 2007)
OTA	Benchtop FT-MIR	Dried vine fruit	PLS	R2 = 0.95 RMSEP = 43.4 %	(Galvis-Sánchez et al., 2007b)
OTA	Benchtop FT-MIR	Durum wheat	PLS-DA, PC-LDA	Accuracy: 96 % (FT-MIR) Cut-off value: 2 μg/kg	(De Girolamo, von Holst, et al., 2019b)
DON	Benchtop FT-MIR	Wheat bran	PLS-DA, PC-LDA	Accuracy: 86 % (PLS-DA, Cut-off value: 400 ppb) and 87 % (PC-LDA)	(De Girolamo, Cervellieri, et al., 2019b)

Note: DON: deoxynivalenol; AF: aflatoxin; OTA: ochratoxin A; SIMCA: Soft independent modeling of class analogy; OPLS-DA: Orthogonal partial least squares discriminant analysis; PLSR: partial least squares regression; ACN: acetonitrile; MeOH: methanol; LDA: linear discriminant analysis; Rv: correlation coefficient of validation; RMSEP: root mean-square error of prediction; PLS-DA: partial least squares discriminant analysis; SVM: support vector machine; KNN: k-nearest neighbors; SEP: standard error of prediction; PC-LDA: principal component-linear discriminant analysis;

Chemical spiked samples have been evaluated for the potential of using MIR spectroscopy for detecting mycotoxins directly. Song et al. highlighted the capability of FT-MIR-ATR in identifying the peanut oil that spiked with AFB₁ (0.02 μg/mL) by using PLS-DA and SVM algorithm (Song et al., 2021). The trace amount of AFB₁ was barely discernable in the spectrum due to the predominant absorption peaks of peanut oil. The performance of the nonlinear SVM model was superior to that of the PLS-DA model, demonstrating AFB₁ detection accuracies of 98.2 % and 94.6 % respectively. Additionally, FT-MIR-ATR has been combined with PLSR for rapid quantification of total aflatoxins that spiked in commercial feeds and food grains (limit of the method <5.0 ng/g), reporting with prediction accuracy of R²cv (coefficient of determination of cross-validation) =0.99 from acetonitrile extracts and R²cv = 0.99 from 75 % methanol extracts (Salisu et al., 2022a). Recently, our group conducted a spiking study by contaminating individual peanut kernels with an AFs mixture (Fig. 2g). Supervised classification successfully extracted the AFs chemical fingerprint (30–400 ppb) from the complex food matrix, achieving 100 % accuracy in external validation (Yao et al., 2024). This discrimination was strongly associated with the C C stretching vibrations of the ring structures of aflatoxins. So far, only these three studies have demonstrated that FT-MIR has the potential to detect mycotoxins around regulatory levels. However, these studies may face significant challenges due to chemical composition changes in fungi-contaminated samples, which can alter the food matrix background in-real application scenarios. Additionally, some studies had limited sample sizes in their validated models and measured spiked samples in different batches due to the characteristics of FTIR-ATR measurements. These constraints underscore the need for further research using fungal-infected samples to enable more reliable field measurements.

Fungal-infected kernels have been studied to assess the potential of MIR spectroscopy for detecting mycotoxins, by generating classification algorithms (De Girolamo, Cervellieri, et al., 2019a; Kos et al., 2016; F. Shen et al., 2018). Kos et al. reported the detection of AFB₁ and DON in peanuts and maize respectively and generated bagged decision tree algorithms that gave 77 % prediction accuracy for AFB1 (8 ppb) and 85 % prediction accuracy for DON (0.5 ppm). Similarly, FT-MIR was combined with PLS-DA (Partial least squares-discriminant analysis) and PC-LDA (principal components-linear discriminant analysis) to detect OTA in durum wheat (De Girolamo, von Holst, et al., 2019b) and DON in wheat bran (De Girolamo, Cervellieri, et al., 2019b). Moreover, FT-MIR coupled with PLSR or MLR (multiple linear regression) was used to determine AFs in brown rice (F. Shen et al., 2018) and peanuts (Kaya-Celiker et al., 2014), DON in wheat (Abramović et al., 2007), OTA in dried vine fruit (Galvis-Sánchez et al., 2007a). Furthermore, fungal contamination and mycotoxin production was reported to associate with the spectral changes in the protein, fat, and carbohydrate absorption regions (Kaya-Celiker et al., 2014; Kos et al., 2016). Applications of MIR spectroscopy have illustrated sufficient sensitivity and specificity for detecting mycotoxins in fungal-infected kernels, allowing for having high sampling throughput with rapid sampling handling time (1-2 min per sample) and providing reliable and accurate results. However, to date, no studies have reported using FT-MIR for mycotoxin determination focusing on intact single kernels. The limited sampling area of ATR elements (approx. 2-3 mm in diameter) may not be able to capture enough representative chemical information from the surface (couple microns) of food crops, challenging its applications for non-destructive analysis.

The integration of machine learning with MIR spectroscopy has represented a transformative advancement in mycotoxin detection, offering rapid, non-destructive screening of agricultural commodities. This combination leverages MIR's ability to capture molecular fingerprints specific to mycotoxins and detect compositional changes in fungal-infected crops caused by mycotoxin metabolism, while machine learning algorithms enhance analytical precision and scalability (Kabir et al., 2025). As discussed, machine learning algorithms such as non-linear SVM and PLS-DA have demonstrated high detection accuracy, even in complex matrices where subtle spectral variations are often obscured by dominant components. However, several challenges remain in fully realizing the potential of machine learning-MIR systems. Variability in sample composition and preparation can compromise spectral reproducibility, and preprocessing steps like grinding and sieving are often necessary to standardize samples. Additionally, data limitations, particularly the need for large, diverse mycotoxin datasets, hinder model robustness and generalization. For example, geospatial ML models for mycotoxin prediction require multi-year, multi-location datasets to account for environmental factors, while small sample sizes risk underfitting (Castano-Duque et al., 2023). Furthermore, recruiting naturally infected samples will be essential to improve model accuracy and reliability.

4. Non-destructive and destructive evaluations of mycotoxins by NIR spectroscopy

NIR spectroscopy has been used to predict mycotoxin contaminations in different agricultural commodities (Table 2), which is used more frequently compared to MIR spectroscopy. Diffuse reflection setting (Fig. 2e) is the predominant NIR spectral acquisition mode in mycotoxin studies. Basically, in this setting, sample materials are placed in a NIR transparent sampling cup, and the cup is placed on the spectrometer to collect a spectrum, only requiring a few seconds. The detector and the NIR light source are typically underneath the same side of the sample. Advanced diffuse reflection NIR spectrometers allow subsampling by rotating the sample cup, enabling spatial averaging over a variety of angles for heterogeneous crop samples while the spectral acquisition (Freitag et al., 2022b). Diffuse reflection NIR spectra are usually plotted as log (1/Reflectance (%)) vs wavelength, to transform the reflectance into absorbance for quantitative analysis (Workman Jr. & Weyer, 2012). However, the direct detection of low concentrations of mycotoxin signals in NIR spectra is hard, due to the overlapped and combinations of spectral information from major constituents of the samples (Freitag et al., 2022b). As mycotoxigenic fungi grow on the crop, the fungi metabolism results in changes in the major constituents. Thus, mycotoxin can be indirectly detected by using NIR spectroscopic methods (Levasseur-Garcia, 2012).

Table 2.

Summary of detecting mycotoxins by using NIR spectroscopy.

Mycotoxins	Instrument	Food/Matrix	Chemometrics	Results	Reference
AFs AFB1	Benchtop FT-NIR	Maize	PCA-DA	Accuracy: 97.4 % for aflatoxin B1 and 100 % for total AFs with a cut-off value of 2 ppb	(Bailly et al., 2024)
AFs	Benchtop FT-NIR	Brown rice	LDA PLSR	Accuracy = 96.9 % (≤5 ppb, 5-300 ppb, >300 ppb) AFs (0-2363.6 ppb): Rv = 0.95, RMSEP = 231 ppb	(F. Shen et al., 2018)
AFs	Benchtop FT-NIR	Maize	KNN, LDA, PLS-DA	Accuracy: 100 %, 96 % and 96 % (KNN, LDA and PLS-DA) (AFs: <20 ppb, 20-200 ppb, 300-450 ppb, 550-700 ppb, >850 ppb)	(Lee et al., 2015)
AFs	Benchtop FT-NIR	Corn	PLS-DA, CARS-PLSDA, PLSR	Accuracy: 91.11 % (PLS-DA) and 97.78 % (CARS-PLSDA) (Detect aflatoxigenic fungus-infected kernels) Rpre = 0.91 RMSEP = 284.27 ppb	(Tao, Yao, Zhu, et al., 2019)
AFB1	Benchtop FT-NIR	Peanut	PLS-DA, RF-PLS-DA,	Accuracy: 88.57 % (PLS-DA, Cut-off value: 20 ppb), 92.86 % (PLS-DA, 100 ppb), 90.00 % (RF-PLS-DA, 20 ppb) and 94.29 % (RF-PLS-DA, 100 ppb)	(Tao, Yao, Hruska, et al., 2019)
FBs	Benchtop FT-NIR	Maize	Decision Tree	Accuracy: 82 % (Cut-off value: 4 ppm)	(Levasseur-Garcia et al., 2015b)
FB1, FB2, ZEN	Benchtop FT-NIR	Maize	PLSR	FBs: R = 0.81 RMSEP = 659 ppb ZEN: R = 0.99 RMSEP = 69.4 ppb	(Tyska et al., 2021b)
DON	Benchtop FT-NIR	Wheat bran	PLS-DA, PC-LDA	Accuracy: 86 % (PLS-DA, Cut-off value: 400 ppb) and 87 % (PC-LDA)	(De Girolamo, Cervellieri, et al., 2019b)
DON	Benchtop FT-NIR	Wheat grain	PLSR	R2 = 0.83, RMSECV = 6.18 ppm	(Peiris et al., 2017)
DON	Benchtop FT-NIR	Whole wheat meals	LDA, PLS-DA	Accuracy: 87.10 % (LDA) and 93.55 % (PLS-DA) (Cut-off value: 1 ppm)	(Zhang et al., 2021)
DON	Benchtop FT-NIR	Barley	PLS-DA, PLSR	PLS-DA (cut-off value: 1.25 ppm): 90.9 % (specificity) and 81.9 % (sensitivity) PLSR: RMSEP = 1.61 ppm	(dos Caramês et al., 2020)

Note: FB: fumonisin B; ZEN: zearalenone; PCA-DA: Principal Component Analysis—Discriminant Analysis; CARS-PLSDA: competitive adaptive reweighted sampling-PLS-DA; Rpre: correlation coefficient of prediction; RF-PLS-DA: random frog-PLS-DA; RMSECV: root mean-square error of cross-validation.

Mycotoxins in fungi-infected intact kernels have been studied by NIR spectroscopy, which requires a large sample volume due to the low absorbance coefficient of the overtones (Dvořáček et al., 2012a; Levasseur-Garcia et al., 2015a; Levasseur-Garcia & Kleiber, 2015). Cecile Levasseur-Garcia et al. applied NIR spectroscopy (400-2498 nm) to detect FBs contaminations in fungi-infected intact maize kernels, ensuring the safety compliance to the EU regulation limit (FBs, 4 ppm) (Levasseur-Garcia et al., 2015b). Dvořáček et al. used an FT-NIR (1000-2500 nm) spectrometer to non-invasively determine the DON amount in infected intact wheat grains. Performances of R_cv (correlation coefficient of cross-validation) = 0.88 and SEP (standard error of prediction) of 6.23 mg/kg for predicting DON content of 0-90 mg/kg were achieved (Dvořáček et al., 2012b). These studies achieved the goal of detecting mycotoxin contamination using a small pile of randomly selected intact kernels (which filled the sampling cup for measurements) nondestructively by NIR spectroscopy. However, recent studies have reported the highly diverse nature of fungal contamination in crops, indicating that even one kernel with extremely high mycotoxin levels could significantly impact the analysis results for an entire batch, potentially exceeding regulatory limits (C. Zhang et al., 2017). Therefore, a well-designed sampling technique should be applied to the analysis based on the sample types and application environments. Additionally, variability in sample size, composition, and structure within intact kernels may lead to inconsistencies in spectral data. To address this challenge, a comprehensive and well-trained algorithm would be particularly required for its real-world applications.

Milled kernel samples or even sieved samples are often investigated to increase the sample homogeneity for building NIR spectroscopic models (De Girolamo, Cervellieri, et al., 2019b; Giacomo & Stefania, 2013; Tyska et al., 2021a; Zhang et al., 2021). Tyska et al. predicted the FBs and ZEN in ground maize samples with standardized particle size, by using a benchtop NIR system (400-2500 nm). Likewise, De Girolamo et al. explored FT-NIR to predict DON content in sieved ground common and drum wheat. Although there are spectral differences between different wheat species, the developed PLSR model gave Rv (correlation coefficient of validation) = 0.76 and RMSEP (root mean-squeare error of prediction) = 379 ppb, allowing to predict the DON content below the maximum limit set by the European Commission (De Girolamo et al., 2009). Overall, the homogeneity helps in reducing the variability in sample size, composition and structure, and obtaining more representative results, minimizing the risk of missing localized contamination that could occur with intact kernels. Using milled samples can enhance the detection sensitivity and ensure compliance with regulatory limits.

Comparative studies were conducted by using both MIR and NIR spectroscopy for detecting mycotoxins in crops. It might be expected that MIR would have greater performance than NIR on detecting mycotoxins due to their superior fingerprinting capability of MIR. In this regard, previous studies provided inconsistent comparison results. Lee et al. demonstrated that fundamental vibrations detected by MIR outperformed the analysis of combinations and overtones of vibrations from NIR. They observed enhanced performance in detecting aflatoxins in maize using both MIR and Raman spectroscopy compared to NIR (Lee et al., 2015). However, Shen et al. evaluated MIR and NIR spectroscopy in detecting aflatoxins in brown rice and reported that NIR spectroscopy outperformed MIR spectroscopy based on correlation coefficients and root mean-square errors (F. Shen et al., 2018). De Girolamo et al. found that NIR spectroscopy performed better on detecting DON, while performed slightly worse than MIR spectroscopy on detecting OTA (De Girolamo, Cervellieri, et al., 2019b; De Girolamo, von Holst, et al., 2019b). These findings underscore the varying detectability of different mycotoxins depending on the spectroscopic technique used, suggesting the importance of selecting the appropriate method based on the specific mycotoxin of interest. Futuremore, the type of crop and its physical characteristics can influence the effectiveness of MIR and NIR, impacting their relative performances. Notably, diffuse reflection NIR spectrometers often exhibit superior performance compared to ATR-MIR spectrometers, possibly due to the small ATR sampling windows (DI: ∼2 mm) and short penetration depth (DI: ∼2 μm) in the latter, which can impact the representativeness of the collected spectrum. Researchers have thus focused on optimizing working flows including grounding and sieving samples to enhance spectral reproducibility in ATR-MIR spectrometers.

As discussed, the integration of NIR spectroscopy and machine learning has significantly advanced mycotoxin analysis in agricultural products, offering rapid and accurate alternatives to traditional chromatographic methods. This synergy addresses critical food safety challenges while introducing new computational considerations. ML algorithms optimize feature selection from NIR spectral data, improving detection sensitivity. For example, combining NIR with a PLS-DA model for evaluating enniatin in barley grain achieved excellent sensitivity and specificity (100 % and 94.2 %, respectively), demonstrating NIR's potential as a promising tool for monitoring emerging mycotoxins (dos Caramês et al., 2020). Similarly, an SVM model significantly enhanced detection performance, with the optimized NIR-based model achieving a correlation coefficient of 0.9761 for the quantitative analysis of peanut aflatoxin B₁ (Li et al., 2023). Despite these advancements, similar to MIR-based machine learning models, NIR models require large, annotated datasets for effective training, which are often scarce for emerging mycotoxins. Limited datasets hinder model validation robustness and increase the risk of overfitting. Additionally, overlapping NIR spectra can reduce classification accuracy, as observed in prior studies (Williams et al., 2012). Multi-mycotoxin detection remains particularly challenging due to the divergent chemical properties of different mycotoxins, necessitating customized extraction protocols.

5. Spontaneous Raman scattering and SERs evaluations on screening mycotoxins

5.1. Recent advances of Raman and SERs in mycotoxins screening

Spontaneous Raman scattering has demonstrated its capability for rapid detection of mycotoxins in crops (Table 3) (Lee et al., 2014, Lee et al., 2015; Y. Liu et al., 2009). Lee et al. used a benchtop Raman system with a 785 nm laser to detect AFs-contaminated maize samples (Lee et al., 2015). The spectral data were analyzed by using supervised classification approaches (KNN (K-nearest neighbors), LDA (linear discriminant analysis), and PLS-DA) and regression approach (MLR (multiple linear regression)). It was found that the classification accuracy was 70.7-96.6 %, and bands for β(C—H) ring mode (1400-1500 cm⁻¹), ν(C—O) mode (1080 cm⁻¹), and ν(C—C) mode (956 cm⁻¹) were significantly influenced from the mycotoxin contamination. The MLR regression models explained most of the variance (83-92 %) and exhibited low errors. Similarly, our group recently conducted a spiking study using an AFs mixture and peanut kernels. Utilizing a handheld Raman system with a 1064 nm laser, we achieved 80 % accuracy in discriminating peanut kernels with spontaneous AFs Raman fingerprints (Fig. 2h) at a concentration of 100 ppb during external validation (Yao et al., 2024). This discrimination was closely associated with the C O stretching vibrations of the aflatoxins' structure. Liu et al. revealed that ground wheat and barley samples with DON could be classified based on contamination amounts (low from high) (Y. Liu et al., 2009). A Raman system equipped with a 1064 nm laser was used in this study. A simple intensity-intensity plot as the discriminant analysis approach was able to identify the samples with DON (> 2 mg/kg (wheat samples), and > 6 mg/kg (barley samples)). However, the sensitivity of spontaneous Raman scattering in detecting mycotoxins is criticized, for instance, it is not sensitive enough to detect DON in ∼μg/kg levels. Thus, SERs has been investigated as an ultrasensitive approach to detect trace levels of mycotoxins.

Table 3.

Summary of detecting mycotoxins by using spontaneous Raman spectroscopy and SERs.

Mycotoxins	Spontaneous/SERS	Food/Matrix	Chemometrics	LODs	Reference
AFs	Spontaneous (1064 nm laser)	Peanut	SIMCA	<100 ppb	(Yao et al., 2024)
AFs	Spontaneous (785 nm laser)	Maize	LDA, LSVM, QDA, PLSR	<10 ppb	(Kim et al., 2023b)
AFB1	Spontaneous (785 nm laser)	Maize	SVM	2.62 ppb	(Deng et al., 2022)
AFs	Spontaneous (785 nm laser)	Maize	KNN, LDA, PLSDA, MLR	<20 ppb	(Lee et al., 2015)
AFB1	SERs (Au NPs, 785 nm laser)	Wheat and corn flour	N/A	0.85 ppb	(Liu, Vanmol, et al., 2020)
DON	SERs (AgNPs, 785 nm laser)	Corns, kidney beans, oats	N/A	Corns and kidney beans: 10−6 M Oats: 10−4 M	(J. Yuan et al., 2017a)
FB1, FB2, FB3	SERs (Ag dendrites, 785 nm laser)	Maize	KNN, MLR	N/A	(Lee and Herrman, 2016)
ZEN, OTA, AFB1	SERs (Polystyrene microsphere-encapsulated SERs, 785 nm laser)	Mixed solution	N/A	0.159 fg/L, 2.015 fg/L, and 1.561 fg/L for ZEN, OTA and AFB1	(Yang et al., 2022)
AFB1	SERs (Au@AgNSs-Fe3O4@AuNFs on an Al-plate, 633 nm laser)	Peanut oil	N/A	0.40 pg/mL	(He et al., 2020b)
OTA	SERs (Au@Au-AgNNSs in solution, Au@Au-AgNNSs-4-MBA-aptamer, 633 nm laser)	Red wine	N/A	0.004 ng/mL	(Shao et al., 2018)
AFB1	SERs (Au NBPs-AAO, handheld 785 nm laser)	Peanuts	N/A	0.5 μg/L	(Lin et al., 2020)
Patulin	SERs (MIP-ir-AuNPs in solution, 633 nm laser)	Blueberry sauce, grape sauce, orange juice	N/A	5.37 × 10−12 M	(L. Wu et al., 2019b)
Patulin	SERs (MIP-ir-Au/PDMS/AAO template, 785 nm)	Orange juice, grapefruit jam, blueberry jam	N/A	8.5 × 10-11 M	(Y. Zhu et al., 2020b)
OTA	SERs (Ag-capped nanopillars substrates, 780 nm laser)	Wine	N/A	White wine: 155 ppb Red wine: 306 ppb	(Rostami et al., 2020)
OTA	SERs (Gold on a glass slide, 660 nm laser)	Standard solution	PCA, OPLS, PLSR	N/A	(Gillibert et al., 2018)
OTA, DON	SERs (Film over nanosphere, 785 nm laser)	Standard solution	N/A	N/A	(Rodriguez et al., 2020)
DON, FB1	SERs (Au nanopillar arrays, 785 nm laser)	Standard solution	PCA	DON: <1 ppm FB1 < 1.225 ppm	(Liu, Wen, et al., 2020)
AFB1, OTA	SERs (AgNPs on a silicon wafer, 785 nm laser)	Cocoa beans	CAR-PLS, GA-PLS	AFB1: 14.15 pg/mL OTA: 2.63 pg/mL	(Kutsanedzie et al., 2020)

Notes: AF: Aflatoxin; DON: deoxynivalenol; FB: fumonisin B; OTA: ochratoxin A; ZE: zearalenone; KNN: k-nearest neighbor; LDA: linear discriminant analysis; LSVM: linear support vector machines; SVM: support vector machine; QDA: quadratic discriminant analysis; PLSDA: partial least squares discriminant analysis; MLR: multiple linear regression; PCDA: principal component discriminant analysis; PCR: principal components regression; PLSR: partial least squares regression; LODs: limits of detections; PCA: principal component analysis; NPs: nanoparticles; CAR-PLS: competitive adaptive reweighted sampling-partial least squares; OPLS: orthogonal projection on latent structure; GA-PLS: genetic algorithm-PLS; DSNB: 5,5-dithiobis(succinimidyl-2-nitrobenzoate); NS: nanosphere; NF:nanoflower; NNS: nanogapped nanostructure; MBA: mercaptobutyramidine; NBPs: nanobipyramids; AAO: anodic aluminum oxide; MIP: molecular imprinted polymer; PDMS: polydimethylsiloxane.

SERs generates significant enhancements in Raman scattering, originating from surface plasm particles (Martinez & He, 2021). Generally, SERs is sensitive to enhance the signal from symmetric vibrations of covalent bonds in non-polar groups, such as S—S, C—C, and C C, having advantages of high sensitivity, non-destructiveness, and rapidity. The possible obstacles in SERs-based techniques have been investigated a lot by researchers to improve the selectivity, accuracy, and precision in detecting mycotoxins (Table 3) (Li et al., 2019a). SERs substrates are directly associated with signal enhancement performance, thus great efforts have been taken to fabricate substrates, utilizing bottom-up and top-down approaches (Fig. 4a) (Logan et al., 2024). Generally, gold and silver nanoparticles are the most popular and extensively used SERs substrates. Liu et al. prepared gold nanoparticles and obtained SERs spectrum of grain extracts, showing that AFB₁ had unique fingerprinting peaks at 1300 cm^-1 with a LOD at 0.85 μg/kg (Fig. 4b) (Liu et al., 2020). Yuan et al. investigated a rapid DON detection approach for corn, kidney beans, and oats samples by using silver nanoparticles. In their study, the LODs were achieved at 10^-6-10^-4 M with the characteristic bands at 855, 1007, 1202, 1430, 1448, and 1680 cm^-1 (J. Yuan et al., 2017a). In addition, anisotropic nanoparticles (such as Ag dendrites, Au nanotriangles, Au nanobipyramids, and Ag nanocubes) and engineered 3D structure substrate (Fig. 4c) were used to detect mycotoxins, with the edges in the particles to create “hot spots” for great enhancements in Raman signal (Z. Wu et al., 2021). For example, Lee & Herrman prepared a silver dendrites SERs substrate for detecting FBs contamination in maize samples and established KNN models with correct classification rates of 70.6-79.4 %. Chemometric quantification models were developed by multiple linear regression, demonstrating strong correlation (K.-M. Lee & Herrman, 2016a). Additionally, SERs tagging detection has been becoming popular with employing parts of metal nanoparticles, protection layers, Raman reporters, and recognition elements (i.e., antibodies and aptamers, Fig. 4d), which allows recognition of specific target molecules and improves the reproducibility and stability of the SERs signals (Cao et al., 2024; He et al., 2020a; Lin et al., 2020). However, interferences from the complex food matrix such as lipids, organic acid, carbohydrates, and protein often make SERs applications of detecting mycotoxins in food products challenging. Thus, molecular imprinted polymer (MIP), microfluidic, supported liquid membrane (SLM), and thin-layer chromatography (TLC) are commonly integrated into the SERs detection to provide accurate and sensitive mycotoxin screening platforms these days (L. Wu et al., 2019a; Y. Zhu et al., 2020a).

Fig. 4.

Schematic of the proposed SERS mechanisms, utilizing (a) bottom-up and top-down fabrication approaches (Logan et al., 2024), with (b) Ag nanoparticles (J. Yuan et al., 2017b), (c) cauliflower-inspired 3D substrates (Li et al., 2019b), (d) aptamers engineered via SERS nanotags and MRS nanoprobe (Cao et al., 2024), (e) integrated into an end-edge-cloud computing framework (Y. Wang et al., 2024) and (f) addressing challenges in IoT-based smart sensors for automated farming (Rajak et al., 2023) and (g) portable spectroscopy sensors (Rodriguez-Saona et al., 2020).

The coupling of machine learning with SERs overcomes some of the key challenges in mycotoxin Raman analysis, such as baseline drifts, signal variations and matrix interferences in food systems. By leveraging machine learning's computational power to decode molecular fingerprints from SERs spectra, this synergy enables label-free, non-destructive detection with unprecedented precision. Machine algorithms extract subtle spectral features that are often imperceptible to conventional analysis, achieving at molar-level sensitivity. Research has shown that polystyrene microsphere-encapsulated SERs aptamer sensors can reduce background noise and enable rapid, portable detection of multiple mycotoxins in actual samples. The LODs for ZEN, OTA, and AFB1were as low as 0.159 fg/L, 2.015 fg/L, and 1.561 fg/L, respectively (Yang et al., 2022). Similarly, machine learning is essential in advancing the detection technology of Raman spectroscopy-based coupled nanosensors. For example, SERS-coupled density functional theory (DFT) and multivariate calibration methods have been applied for accurate quantification of AFB1 in peanut oil, with a correlation coefficient of 0.9332, and the method can be extrapolated to other mycotoxins (Chen et al., 2020). Moreover, spectral heterogeneity in food samples is mitigated by machine learning-driven feature extraction, which lowers the LOD to a certain extent. SERS combined with deep learning models, such as one-dimensional convolutional neural networks (1D CNN) and two-dimensional convolutional neural networks (2D CNN), has proven to be an ultrasensitive and effective strategy for detecting ZEN in corn oil, with LODs of 6.81 × 10 − 4 and 7.24 × 10 − 4 μg/mL, respectively (J. Zhu et al., 2023). Additionally, integrating machine learning algorithms (including artificial neural networks (ANN) and KNN) has enabled precise classification and quantification of DON detection under dual-mode colorimetric/SERS, achieving a classification accuracy of 98.8 % and a mean squared error (MSE) of 0.57 (Sun et al., 2025). In conclusion, the integration of machine learning with Raman/SERS technologies significantly enhances mycotoxin detection. This combination achieves superior sensitivity through noise suppression algorithms and matrix-independent specificity via explainable machine learning teachniques. It outperforms antibody-based methods in speed, cost, and accuracy, meeting the requirements for onsite deployable food safety monitoring. Future work should prioritize edge computing integration and the cross-application of deep learning techniques.

5.2. Challenges of implementing SERs in mycotoxins screening

Although SERs has shown great promise for detecting mycotoxins in food samples due to its high sensitivity and molecular specificity, several challenges hinder its effectiveness and widespread application in screening these contaminants. One significant challenge is related to the chemical structure of mycotoxins. Many mycotoxins are aromatic compounds with few or no ionization groups, such as nitrogen or sulfur, which are crucial for strong interactions with Ag or Au surfaces. This lack of functional groups diminishes the adsorption efficiency of mycotoxins on SERs substrates, leading to weak and inconsistent signal enhancement. Despite advancements in SERs tagging detection have been made for mycotoxins, such as the development of recognition elements (e.g., antibodies and aptamers) and analysis strategies for AFB₁, OTA, and DON, the lack of specific binding reagents has hindered progress in detecting other mycotoxins. Moreover, detecting trace levels of mycotoxins, particularly in the parts-per-billion range, requires highly sensitive and optimized protocols that are not yet standardized across different laboratories. Additionally, the complexity of food matrices poses another challenge, as matrix effects can interfere with the isolation and accurate identification of target mycotoxins. This often necessitates extensive sample preparation to extract mycotoxins or combining SERs with other technologies, adding time to the process and potentially diminishing the rapid detection advantage that SERs offers. Furthermore, while untagged SERs substrates remain the most feasible approach for mycotoxins detection, real-world applications are still limited by the absence of machine learning studies to reliably extract contaminant information from the spectra. Furthermore, the high cost and technical expertise required for producing high-quality SERs substrates and interpreting the complex spectra continue to restrict its routine use. However, automated and customized spectral processing in smart phones integrated with cloud computing could be a potential solution. Overall, to overcome these limitations, advancements in substrate development, more efficient sample preparation, standardized detection protocols, and the integration of automated spectral processing through smartphone applications and cloud computing are essential for the broader adoption of SERs in mycotoxin screening.

6. Potential for miniaturized systems and cloud computing for in-situ smart mycotoxins screening

Driven by advances in photonic technologies and semiconductors, vibrational spectroscopy systems have evolved from bulky benchtop equipment to field-deployable devices, revolutionizing the food industry (Rodriguez-Saona et al., 2020). Commercialization of ruggedized and miniature optical devices (Fig. 1, Fig. 4g) are amenable to be implemented in-field or on-site mycotoxin screening (Villanueva et al., 2023). Traditional NIR chemical analysis was first revolutionized by the introduction of Phazir (Thermo Scientific, Inc.) in 2006. This Phazir analyzer operates in a spectral range of 1600-2400 nm (reflectance mode), by using a tungsten halogen light source, a MEMS (Micro-electromechanical system)-based spatial light modulator, and an InGaAs detector. Afterward, the size of NIR spectrometers has been shrunk continuously to miniaturized with pioneering detectors and wavelength selection techniques (Rodriguez-Saona et al., 2020). In contrast to NIR spectrometers, the size of MIR spectrometers (4000 − 600 cm⁻¹) is limited to “lunchbox” size due to the challenges with moving parts and quantum-type pyroelectric detectors that require cooling (Crocombe, 2018). The miniaturized Raman systems are equipped with different excitation sources that lead to different operational features. To reduce the size of Raman systems, most handheld Raman spectrometers are equipped with a 785 nm excitation laser and a CCD (charge-coupled device) detector (Rodriguez-Saona et al., 2020).

The potential and applicability of portable vibrational spectroscopic approaches for the rapid screening of mycotoxins have been evaluated by researchers. Shen et al. reported that a smartphone-controlled NIR spectrometer operating from 900 to 1700 nm was successfully used to detect FB₁ and FB₂ in ground corn samples (G. Shen et al., 2022). Few studies have employed portable MIR spectrometers for detecting mycotoxins in food products. Kos et al. reported the detection of DON and AFB₁ by using a portable MIR spectrometer, showing 73 %-90 % classification accuracy (Kos et al., 2016). SERs using handheld Raman spectrometers have demonstrated remarkable capabilities for the detection of mycotoxins. SERs based on silver nanoparticles demonstrated the rapid and low-cost detection of DON by using a portable Raman spectrometer (785 nm), achieving LODs of 100 nM, 10^-6 M, and 10^-4 M in sweet corn, kidney beans, and oats respectively (J. Yuan et al., 2017a). TLC combined with SERS by using a portable Raman spectrometer enabled the separation and the on-site detection of four aflatoxins (Qu et al., 2018).

Cloud computing plays a pivotal role in driving the progress of modern agriculture, particularly in advancing precision agriculture (Xia et al., 2021a). By integrating cloud computing (Fig. 4e), a vast amount of spectroscopic data can be centrally stored, allowing for easy assessment and efficient management. Moreover, spectroscopic data can undergo preprocessing, transformation, and modeling within the cloud platform, leading to actionable insights that optimize mycotoxins management practices. Additionally, stakeholders in agriculture such as farmers, researchers, policy makers can easily access the data and algorithms (models), thus fostering precision contaminants screening and collective problem-solving efforts (Langmead & Nellore, 2018). Researchers have significantly contributed to the advancement of cloud computing for applications. For instance, the standardization of detection management platforms (Cui, 2022; Parastar & Shaye, 2015), intelligent data imaginer templates, and on-site standardized detection processes (Zang et al., 2021), along with real time screening and uploading systems (Zhang et al., 2021) have been advanced.

In applications related to vibrational spectroscopic detection and cloud computing, cloud technology has been successfully integrated with near infrared spectrum analysis to develop and verify the algorithms for monitoring apple chilling injury (Xia et al., 2018). Similarly, Xia et al. evaluated various cloud computing artificial intelligence models for classifying greenhouse tomato plants under water stress, using near-infrared spectroscopy (Xia et al., 2021b). Moreover, in the realm of miniaturized systems integrated with cloud computing, Mu et al. demonstrated the use of a high-sensitive smart-phone-based Raman system with cloud network architecture for the material identification (Mu et al., 2019). Chandler et al. enabled a handheld Raman spectrometer with cloud and AI deep learning algorithms for mixture analysis (Chandler et al., 2019). While screening for mycotoxins using these technologies is still in its infancy, the potential for novel portable spectrometers combined with cloud computing in screening mycotoxins represents an exciting area of research. However, challenges such as untrained farmers, limited digital literacy, and data security concerns persist (Fig. 4f) (Rajak et al., 2023). Despite these obstacles, continued exploration of these avenues holds promise for field solutions, particularly in real-time analysis and cloud-based data support.

7. Comparisons between reference methods and vibrational spectroscopy technologies for mycotoxin detection

7.1. Reference methods: characteristics, limits of detection (LODs), and quantification (LOQs) in applications

As mentioned before, ELISA and chromatography coupled with mass spectrometry (LC-MS/MS, GC–MS/MS) are widely used reference methods for detecting mycotoxins (Qi et al., 2022). ELISA operates on the principle that a specific antibody exhibits sensitivity to the unique structural features of a mycotoxin. Recent commercial ELISA kits feature shortened incubation times (∼30 min), leveraging the reaction kinetics between an antigen and an antibody before reaching equilibrium; however, this reduction in incubation duration may result in decreased sensitivity (Picó, 2016). There are three main types of ELISA: direct competitive, indirect competitive, and non-competitive. The primary distinctions lie in antibody–mycotoxin interaction conditions and measurement techniques, each optimized for specific applications (Sharma et al., 2014). ELISA is particularly useful for mycotoxin analysis in resource-limited settings due to its affordability and simplicity (Dib et al., 2023a). Despite being portable and requiring minimal sample volume, ELISA still necessitates sample preparation, typically involving liquid–liquid extraction (LLE) followed by simple steps such as filtration and centrifugation, without further cleanup (Narenderan et al., 2020). ELISA can achieve, but in some cases may not meet, regulatory limits for mycotoxins, such as the US FDA thresholds of 10 ppm for DON in grains and 20 ppb for AFs. For instance, DON is detected by ELISA with an LOD of 54 ppb and an LOQ of 167 ppb in wheat, both exceeding the regulatory limit (Dvořáček et al., 2012c), while AFs show an LOD of 1.75 ppb in maize (Dib et al., 2023b). Importantly, when sample amounts are limited, such as a single peanut kernel, ELISA may lack sufficient sensitivity to quantify mycotoxins on a weight basis (ppb or μg/kg). ELISA's susceptibility to cross-reactivity can further compromise sensitivity and specificity, potentially resulting in false positives or negatives (Goyal et al., 2024). These limitations illustrate a trade-off between simplicity, speed, and reliability. Although ELISA is a well-established technique that is convenient for only relatively rapid screening (∼30 min) in resource-limited laboratories, its results often require careful interpretation in regulatory contexts. Consequently, while ELISA is suitable for preliminary screening, chromatographic methods remain the reference standard for mycotoxin detection, as recommended by regulatory agencies such as the USDA (USDA, 2023). Comprehensive studies assessing ELISA accuracy across a wider range of food commodities are still needed to clarify its practical applicability. Overall, these considerations highlight that the performance of ELISA is highly dependent on mycotoxin type, concentration and sample matrix, emphasizing the need for method selection and careful data interpretation in routine analysis.

Chromatographic methods, though more labor-intensive in terms of sample pretreatment, offer excellent sensitivity and selectivity by including extensive cleanup processes. These steps are crucial for addressing matrix effects caused by the complexity of food compositions (Yang et al., 2020). Cleanup is vital to avoid interference, ensuring that co-elution of compounds does not reduce ionization efficiency, thus preserving data accuracy and repeatability (Krska & Molinelli, 2007). Extraction typically uses organic solvents such as acetonitrile or methanol, followed by various cleanup methods, including solid-phase extraction (SPE), dispersive SPE (DSPE), LLE, QuEChERS, and immunoaffinity columns. For instance, detecting AFs, fumonisin B₁, and T-2 toxin in peanuts, maize, and wheat using LC-MS/MS requires ethanol-water extraction, centrifugation, dilution, purification through immunomagnetic SPE (IMSPE), separation, and desorption (W. Wang et al., 2022). LC-MS/MS offers high selectivity and sensitivity, enabling the simultaneous detection of multiple mycotoxins (Boshra et al., 2024). It can quantify DON in barley with an LOD of 15 ppb and an LOQ of 50 ppb (dos Caramês et al., 2022), as well as AFs in maize with LODs ranging from 0.11 to 0.36 ppb and LOQs from 0.36 to 1.19 ppb (Ouakhssase et al., 2019). GC–MS/MS is highly sensitive for analyzing certain mycotoxins, offering reduced interference and excellent sensitivity through fragmentation. However, it requires a derivatization step that lowers the boiling point of non-volatile compounds, making them easier to volatilize in GC and enhancing the stability of thermolabile compounds (R. Ran et al., 2013). Derivatization methods like silanization, acylation, and akylation often involve hazardous chemicals, and may introduce interference issues (Singh & Mehta, 2020a). When evaluating the overall cost of these methods, factors such as instrumentation, personnel training, consumables, and samples shipping should be considered. LC-MS/MS and GC–MS/MS, while highly sensitive and capable of detecting multiple mycotoxins, are expensive and time-consuming, requiring extensive sample preparation and highly trained personnel due to the complexity of food matrices and diverse chemical structures of mycotoxins (Magoke et al., 2025). The high selectivity and sensitivity of these chromatographic methods are advantageous, but sensitivity can also be influenced by the ionization method used (Singh & Mehta, 2020b).

Food contamination remains a persistent risk throughout the supply chain, making early and accurate detection critical for quality assurance. Errors or misinterpretation in data can result in regulatory non-compliance and consumer safety concerns. While ELISA and chromatographic methods provide complementary strengths in accessibility and analytical accuracy, neither fully meets the growing demand for rapid, large-scale, or on-site monitoring. Recently established approaches, such as ultra-fast HPLC and streamlined sample preparation via QuEChERS (Quick, Easy, Cheap, Effective, Rugged, and Safe) method, promise faster analysis and higher throughput while maintaining sensitivity and selectivity, though they remain constrained by instrumentation requirements. Chromatographic methods remain irreplaceable as gold-standard confirmatory tools, but integrating rapid, non-destructive techniques, such as vibrational spectroscopy, with these reference methods provides a practical solution, enabling cost-effective, on-site screening without compromising analytical reliability. Such integration may bridge the gap between convenient screening and confirmatory accuracy, representing enabling both efficient monitoring and regulatory compliance.

7.2. Comparative performance of vibrational spectroscopic-based methods in applications

Based on the existing research, Raman spectroscopy represents a rapid and precise alternative for mycotoxin detection, and its adaptability to portable devices makes in-situ monitoring feasible. However, its broader implementation is limited by the inherently weak signal from spontaneous Raman scattering, which restricts sensitivity, and by the relatively few studies using portable Raman devices for practical food applications (Xiao et al., 2023). SERs, although promising, is currently less viable for industrial applications due to the high cost of SERs substrates and the technological limitations that have discussed in the previous sections. NIR spectroscopy presents a more cost-effective solution than Raman spectroscopy, offering the flexibility of different pathlengths and suitable for detecting a wide range of mycotoxins non-destructively. However, NIR spectroscopy is challenged by strong band overlap that can influence the sensitivity of results, compared with MIR that provides distinctive fingerprint spectra for directly identifying mycotoxins fingerprints. NIR is particularly effective for detecting compositional changes in food matrices, making it suitable for indirect detection of mycotoxins.

Vibrational spectroscopic methods require reference methods to develop multivariate algorithms for mycotoxin detection. Compared with traditional mass spectrometry-based approaches, most vibrational spectroscopy methods (excluding SERs) generally exhibit lower sensitivity than gold-standard chromatographic techniques, yet they still meet regulatory guidelines. For instance, spontaneous Raman spectroscopy had LODs of 10 ppb for AFs in maize (Kim et al., 2023b) and 301 ppb for DON in wheat brans (Mignani et al., 2016). MIR spectroscopy reported a LOD of 20 ppb for AFs in peanuts (Yao et al., 2024) and a LOQ of 0.18 ppm for DON in wheat (Shen et al., 2019). NIR spectroscopy detected DON with an LOQ of approximately 1 ppm in intact wheat (Dvořáček et al., 2012c). Comparisons among these methods would benefit from more standardized studies using the same matrices and sample preparation procedures. SERs demonstrated exceptional sensitivity, such as an LOD of 0.85 ppb for AFB₁ in wheat and corn flour (Liu, et al., 2020). SERs studies often report results in ng/mL or M units, which complicates direct comparisons with techniques using ppb/ppm units. Moreover, vibrational spectroscopy can achieve comparable or greater sensitivity than certain traditional ELISA-based approaches, while offering a significantly shorter analysis time (<1 min vs ∼30 min), making it highly promising for rapid monitoring of mycotoxin contamination in food safety. Despite these technical advantages, translating vibrational spectroscopy into routine, field-deployable workflows requires continued efforts. Standardized and validated SOPs that integrate rapid spectroscopic screening with machine learning across diverse food matrices are still limited. Practical implementation and scalability particularly require improvements in addressing data variability, model deployment, differences among commodities, and critical stages of the supply chain.

8. Conclusion

Mycotoxins are common food contaminants that cause a range of health issues from illness to even death. The detection of mycotoxins requires comprehensive monitoring of food with reliable and accurate analytical approaches. Spectroscopic techniques, NIR, MIR, and Raman (spontaneous and SERs) combined with machine learning have been proven to be promising techniques for screening mycotoxins by recently published studies. Advancements in microchip technology have propelled a variety of portable vibrational spectroscopic devices and opened the door to on-the-go mycotoxins screening in the food industry. This review highlighted the developments in the detection of mycotoxins by using vibrational spectroscopy and provided an overview of the current state of the art for portable vibrational spectroscopic devices as a promising tool for future applications. Future work will require studies on the feasibility of in-field analysis by using portable spectrometers that integrated with cloud computing.

CRediT authorship contribution statement

Siyu Yao: Writing – review & editing, Writing – original draft, Data curation, Conceptualization. Tong Yu: Writing – original draft, Formal analysis. Alessandra Fantina Victorio Ramos: Writing – original draft. Zhongkun Zhang: Formal analysis. Zulipikaer Rouzi: Data curation. Luis Rodriguez-Saona: Writing – review & editing, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.