Abstract
Food is a more complex system than commonly perceived, comprising tens of thousands of molecules whose compositions and interactions ultimately shape human perception. To conceptualize this multifaceted nature, we frame food complexity across three interconnected layers: the molecular composition that defines its chemical foundation, the component interactions that shape food properties, and the perceptual responses that arise from human sensory systems. This review discusses how machine learning is advancing our ability to decode each of these layers, together with multimodal and data-fusion frameworks. Understanding these three layers may enable more accurate prediction of food properties, guide food product innovation, and deepen our scientific understanding of food.
Similar content being viewed by others
Introduction
Food is inherently complex, consisting of thousands of chemical components that interact and ultimately shape its flavor, texture, nutrition, and consumer acceptance1,2. Despite advances in analytical food chemistry and sensory science, we still lack a clear understanding of how molecular composition, component interactions, and human perception are interconnected across scales3,4.
At the chemical level, food contains thousands of compounds with diverse structures and physicochemical properties, including volatility, solubility, bioactivity, and bioavailability. Even small structural differences, such as isomerism, can lead to pronounced changes in functional or sensory behavior5. Although more than 77,000 food-related chemicals have been reported, fewer than one-tenth have annotated flavor or bioactivity information6,7. This sparse knowledge base makes manual interpretation difficult and highlights the value of machine learning (ML), which can predict molecular properties directly from structure and reveal how specific features shape a chemical’s sensory or functional role in food.
Beyond individual chemicals, food properties also arise from interactions among components. These interactions can create emergent behaviors that are not predictable by examining single compounds in isolation. Well-known examples include the balance created jointly by sugars, acids, and volatiles in fruits8, or the enhancement of flavor richness through lipid-aroma interactions in dairy matrices9. Such interaction patterns are further modified by ingredient composition, processing conditions, and storage environments, contributing to the variability observed across foods10. Analytical tools such as gas chromatography-mass spectrometry (GC-MS) and liquid chromatography-mass spectrometry (LC-MS) help identify and quantify complex molecular mixtures11, whereas sensor systems like e-noses and e-tongues mimic human chemosensation to rapidly profile collective molecular signals12,13. When combined with such approaches and ML, these multimodal data support rapid assessment of quality, authenticity, and origin and provide scalable tools for interpreting interaction-driven complexity14.
Human perception adds yet another dimension. The same chemical stimulus can elicit distinct sensory responses across individuals, influenced by genetic variation, cultural background, and personal experience15. Advances in neuroinformatics technologies, including electroencephalography (EEG), functional near-infrared spectroscopy (fNIRS), and functional magnetic resonance imaging (fMRI), provide richer psychological and physiological markers for studying these perceptual differences16. For example, fMRI and fNIRS can be used to identify specific brain regions activated by different flavor stimuli, and EEG can capture, in real time, the rapid changes in brain electrical activity associated with perception17. Integrating ML with these neurophysiological datasets offers a powerful means to model emotions and preferences elicited by food18,19, and to identify the neural features most relevant to perception20. Such approaches deepen our understanding of how humans experience food beyond its chemical and physical properties.
Although previous reviews have examined ML applications in food science, most have focused on a single data type or addressed only one layer of food complexity11,21,22. As a result, the integration of multimodal data within unified ML frameworks remains limited. In this review, we conceptualize food complexity across three interconnected dimensions: (1) the chemical diversity that defines the building blocks of food; (2) the interactions among components that give rise to the emergent properties of foods; and (3) the perceptual processes through which humans experience food (Fig. 1). Unveiling these dimensions demands large, heterogeneous datasets, motivating growing interest in data fusion strategies that integrate information across sources to tackle high-dimensional or sparse data and enhance model accuracy23,24,25. Building on this motivation, the review aims to summarize how ML has been applied to unveil each dimension of food complexity; examine data fusion approaches that unify disparate datasets; and highlight the opportunity of incorporating neuroinformatics data to deepen our understanding of food perception and improve predictive performance.
Level 1 captures the chemical complexity of food, where diverse chemical structures give rise to distinct bioactivities and flavor characteristics. Level 2 represents the interaction level, encompassing the dynamic physical, chemical, and matrix-dependent interactions among food components that shape quality, safety, and stability. Level 3 reflects perceptual complexity, arising from the multisensory integration of taste, smell, trigeminal sensations, and higher cognitive processes that together determine the eating experience.
Representative machine learning algorithms for unveiling food complexity
ML provides a diverse set of algorithms that can extract patterns, uncover relationships, and make predictions from the heterogeneous datasets that characterize food complexity. Depending on the data structure and objective, different algorithmic families, such as supervised and unsupervised learning, offer complementary strengths for classification, regression, and clustering.
As illustrated in Fig. 2, ML workflows in food research typically involve data collection, preprocessing, feature engineering, and model training. Within this pipeline, supervised algorithms like support vector machines (SVM), decision trees (DT), random forests (RF), and k-nearest neighbors (KNN) have been widely adopted for tasks such as quality prediction, authenticity verification, and sensory property modeling26. Unsupervised techniques, including clustering and dimensionality reduction, are often used to explore data structure, extract latent features, and guide early-stage data processing27. This section introduces the core ML algorithms most relevant to decoding food complexity and their representative applications.
The process begins with data collection, integrating information from databases, sensory evaluations, and instrumental analyses. Next, data cleaning and preprocessing standardize formats, resolve inconsistencies, and harmonize multi-source inputs. Feature engineering follows, involving either manual descriptor construction or automated representation learning through deep models. In model development, algorithms are trained, optimized, and validated to capture structure-function relationships. Finally, the resulting models are applied to predict properties across the molecular, interaction, and perception levels, enabling a multiscale understanding of food complexity.
SVM is a margin-based classifier that identifies the decision boundary maximizing the separation between classes in a transformed feature space28. It addresses nonlinear classification issues using kernel functions, including linear, radial basis, and sigmoid kernels. For example, Gerhardt et al. (2019) developed an SVM model to effectively differentiate among various types of virgin olive oil, including extra-virgin, virgin, and lampante olive oil29. Similarly, Sun et al. (2022) developed an SVM model that integrates fruit metabolomic data to predict consumer preferences, thereby facilitating consumer fruit-flavor selection and aiding breeding programs in developing more popular varieties30.
DTs use hierarchical, rule-based splits to partition data into homogeneous groups for classification or regression31. RFs and XGBoost extend DTs by constructing large ensembles of trees and aggregating their predictions to improve robustness and reduce overfitting32. Wang et al. (2021) developed an RF model to predict the retention index and flavor attributes (aromatic, bitter, sulfury, and other) of molecules in beer, achieving satisfactory performance with an R2 of 0.9633. Hu et al. (2023) developed an RF model to determine the relationship between key aroma components and the sensory properties of fragrant peanut oils34.
KNN classifies or predicts samples based on the labels or values of the most similar instances in the feature space35. Wu et al. (2019) developed a KNN model using electronic nose (e-nose) signals preprocessed with fuzzy discriminant principal component analysis (PCA) as input to differentiate between various Chinese liquor types and identify inferior or counterfeit liquids based on flavor, achieving a classification accuracy of 98.3%36. In addition to the previously mentioned ML algorithms, partial least-squares (PLS) regression, K-means clustering, and naïve Bayes have also been used to investigate food properties7.
In recent years, advances in computational capacity and the availability of large-scale datasets have enabled the application of deep learning7. Deep learning refers to a family of neural-network-based models capable of learning hierarchical representations from large, high-dimensional datasets26,37. In contrast to traditional ML techniques that require manual feature selection, deep learning models can autonomously learn hierarchical representations, thereby facilitating their deployment in complex data analyses.
Artificial neural networks (ANNs) consist of interconnected computational units that transform input features through weighted connections and nonlinear activation functions38. ANNs have been used to predict the taste attributes of chemicals based on molecular descriptors, thereby facilitating the rapid screening of novel flavorings39. In recent years, significant advances have been made in the development of ANN architectures, including convolutional neural networks (CNNs), graph neural networks (GNNs), and recurrent neural networks (RNNs), which have been proposed to address more complex tasks in food analysis.
CNNs use convolutional filters to extract spatially localized features and are widely applied to grid-like data such as images40. In food flavor research, CNNs are frequently used in the analysis of e-nose and electronic tongue (e-tongue) related tasks41,42. Wu et al. (2019) developed a CNN model to predict the perceptual pleasantness of odors based on e-nose data, achieving an accuracy of > 90% on the test dataset43. Zhang et al. (2019) integrated CNNs with neuroinformatics data to enhance flavor recognition by analysing brain electrical signals in response to various aroma stimuli, including coffee, lemon, and vanilla odors44. Xia et al. (2024) developed a CNN model that effectively decodes EEG signals corresponding to the taste of sour, sweet, bitter, and salty foods, thereby facilitating the prediction of food taste perception45.
GNNs operate on graph-structured data by iteratively aggregating information from neighboring nodes and edges46. In contrast to CNNs, which operate on a fixed set of relationships between pixels in image data, GNNs are designed to handle dynamic relationships between elements, aggregating neighboring information to update node representations. When predicting the taste of molecules, Song et al. (2023) observed that whereas most models perform reasonably well in predicting the umami taste, GNNs stand out for their superior accuracy in predicting bitter and sweet tastes47.
RNNs model sequential data by maintaining hidden states that capture temporal dependencies across time steps48. Qi et al. (2023) used an RNN combined with a multilayer perceptron to enable rapid identification of umami peptides49. However, RNNs face limitations like vanishing and exploding gradients, which hinder performance on long sequences50. To address this, advanced variants like long short-term memory (LSTM) networks were developed. LSTMs enhance RNNs by introducing gating mechanisms that regulate information flow and preserve long-range dependencies51. For instance, Jiang et al. (2023) applied an LSTM encoder to extract features from peptide sequences for umami recognition52. LSTMs have also been used in natural language processing tasks in food research, including flavor term extraction from whisky reviews and personalized recipe generation based on user preferences53,54.
Transformer architecture leverages self-attention mechanisms to model global dependencies in sequences while enabling highly parallelizable training55. This design markedly enhances model parallelism, thereby improving the training efficiency on large-scale datasets. Transformer-based models typically generate hidden state representations for each input token, thereby capturing the range of features learned from the input text. For example, Chew et al. (2022) developed a variant of the transformer model to analyse Instagram posts, and this approach exhibited superior precision and recall relative to baseline models in identifying brands and flavors of electronic nicotine delivery systems56. Additionally, large language models (LLMs) built on Transformer architectures excel at processing multimodal inputs (from text to images) to identify dish types and flavor profiles, while generating customized recipes that meet specific requirements57.
Generative models learn the underlying probability distribution of data to create new samples that resemble those in the training set. By training on extensive datasets containing both flavor chemometric data and human sensory panel evaluations, these models learn to construct latent mathematical representations that encode fundamental patterns2. The trained model can generate synthetic but chemically plausible profiles to augment limited experimental datasets or predict emergent characteristics from novel ingredient combinations by interpolating in the learned space. For example, Queiroz et al. (2023) proposed deep generative models to design new flavor molecules based on their molecular structures58. Aleixandre et al. (2025) presented a generative diffusion network to create new aromas with specified characteristics, based on mass spectrometry data of essential oils59. Zhang et al. (2026) developed a conditional variational autoencoder based on KFO-Atlas to produce category-targeted aroma formulations and validate its outputs by blinded human sensory evaluation60. These aspirational directions highlight the capacity of generative modeling to accelerate discovery in food, although most of them remain contingent on further validation with experimental and sensory data.
Overall, the aforementioned algorithms differ in complexity, data requirements, and representational power, each offering distinct advantages for food research (Table 1). Traditional ML methods like SVM, RF, and KNN are well-suited for structured, small-to-medium datasets and are often used for classification, regression, and feature selection tasks. In contrast, deep learning models excel at capturing complex, nonlinear relationships in large, high-dimensional, and multimodal datasets. They can autonomously learn hierarchical representations, making them ideal for analyzing unstructured data such as sensor signals, molecular graphs, peptide sequences, and textual descriptions. Generative models go a step further by enabling the design of novel molecules and the generation of synthetic data.
Machine learning unveils food chemical complexity
Mapping the uncharted chemical space of foods with machine learning
While databases such as KFO-Atlas, FooDB, FlavorDB, and AdditiveChem collectively catalog tens of thousands of food-related compounds as summarized in Table 2, much of the chemical diversity of food remains uncharted, particularly secondary metabolites, processing-induced compounds, and minor constituents with sensory or health relevance. Early efforts such as FoodMine demonstrated the value of systematic literature mining by extracting over 7000 quantified measurements for garlic and cocoa; however, its reliance on manual curation highlighted the limitations of scalability61.
Recent advances in natural language processing, especially large language models (LLMs), now offer automated routes to expanding the molecular inventory. Domain-adapted LLMs can extract chemical entities, normalize molecular identifiers, and associate compounds with functional annotations from large-scale scientific corpora. Models such as GPT-4 have been shown to identify food-related molecules, metabolic intermediates, and bioactive compounds from a large set of publications at speeds impossible for manual curation62.
Beyond literature, ML is also broadening the search for food-relevant metabolites through genome mining. ML-driven tools such as antiSMASH63 and PRISM64 can analyze biosynthetic gene clusters to infer the structures of secondary metabolites that may have nutritional, flavor, or preservative functions. This is particularly relevant for identifying previously uncharacterized molecules from fermented foods, edible fungi, and underutilized plant species.
Predicting multidimensional properties of food-related chemicals
Among the tens of thousands of known food-related molecules, fewer than 10% have annotated flavor or bioactivity information. This sparse coverage has motivated growing interest in ML as a means to infer molecular properties from structural information.
Traditional approaches, such as quantitative structure-activity relationship (QSAR) models, originally established the link between molecular structure and functional behavior using handcrafted physicochemical descriptors65,66. Fingerprints and descriptor-based models remain effective for small to medium datasets and have been widely applied to tasks such as odor classification, as demonstrated by Shang et al.’s (2017) work using over 1000 descriptors from Dragon software to predict odor characteristics67.
However, descriptor-based strategies inherently depend on manually engineered features, limiting their ability to capture higher-order structural patterns. This has driven a shift toward deep learning approaches that learn representations directly from raw molecular structures (Table 3). GNNs, for example, treat atoms and bonds as nodes and edges, allowing the model to capture the topological and geometric properties essential for molecular function. Pred-O3, developed by Ollitrault et al. (2024), predicted 23 odor notes and 109 human olfactory receptor interactions from 5802 food-derived odorants, revealing structure-odor relationships missing from traditional fingerprints68. Similarly, Lee et al. (2023) created a principal odor map that positioned over 500,000 potential odorants, of which only ~5000 have been characterised, highlighting a vast unexplored chemical-sensory landscape69. Emerging work has also explored the use of large language models for molecular property prediction. Song et al. (2024) fine-tuned GPT-3.5 and GPT-4 on SMILES strings for taste classification, with GPT-4 achieving 86% accuracy, illustrating LLMs’ potential to complement graph-based modeling70.
As these methods advance, ML is increasingly positioned to accelerate the early-stage prediction of molecular sensory and functional properties, reducing reliance on lengthy experimental screening. Equally important, interpretable AI frameworks such as SHAP47 offer new opportunities to illuminate the structural features most responsible for molecular behavior, bridging predictive modeling with mechanistic understanding.
Integrating machine learning with instrumental analysis to decode food component interactions
Examining food properties, including quality grading, geographical origin, and freshness, facilitates food product development and quality control (Fig. 3a). Traditionally, these tasks rely on labor-intensive, time-consuming experiments, such as sensory evaluation panels and metabonomics analyses, which are often limited by human subjectivity, high costs, and low throughput. Moreover, the complexity of food matrices and the interactions among various components pose challenges for manual interpretation. In this context, ML offers a powerful tool by learning complex, nonlinear relationships from instrumental data.
a Application of machine learning using instrumental analysis data and the representative artificial neural network (ANN) algorithm for predicting food characteristics. b Machine learning modeling with EEG data. EEG data can be manually preprocessed to extract temporal, spatial, or spectral features, which can then be used in conventional machine learning frameworks, such as support vector machines (SVMs) and random forests (RFs). Alternatively, deep learning algorithms, such as convolutional neural networks, enable automatic feature extraction and prediction, offering an advanced approach to decoding sensory responses to food.
Chromatography-based modeling
Chromatographic techniques such as GC-MS, LC-MS, and GC-IMS spectrometry provide detailed molecular fingerprints of foods and are widely used to profile volatiles, semi-volatiles, and other key constituents. When coupled with ML, these datasets enable rapid and accurate analysis of product origin, quality, and sensory attributes.
Recent studies demonstrate that statistical classifiers and modern ML models can decode subtle variations in chromatographic profiles for tasks such as geographical origin discrimination in honey71, regional classification of Atractylodes lancea72, and prediction of α-acid content in hops73. Similar approaches have been used to assess rancidity in walnuts74 and sensory grade in Sauvignon Blanc wine75. Together, these examples illustrate how ML enhances the interpretive power of chromatographic datasets by uncovering latent relationships between volatile patterns and food properties.
Chromatography-based ML models are also increasingly used to complement sensory evaluation. While sensory panels remain essential, they suffer from fatigue, cultural variability, and limited throughput15. ML can serve as an initial screening tool to map chemical profiles to sensory descriptors, as in studies linking GC-MS data to aroma attributes in peanut oils76 or predicting beer flavor preferences by integrating chemical, sensory, and consumer data77. These approaches help form a hybrid workflow in which ML provides rapid, objective screening, while human assessors calibrate and validate final outputs.
Sensor-based modeling
E-noses and e-tongues extend the capability of instrumental analysis by capturing holistic responses to odorants and tastants through sensor arrays. When integrated with ML, these systems offer high-throughput, reproducible alternatives to traditional sensory evaluation (Fig. 3a).
E-nose signals together with ML have been used to classify products such as Chinese baijiu36, discriminate brewing stages78, recognize beer odors79, and detect spoilage volatiles80. Advanced deep learning methods further improve e-nose performance. Shi et al. (2019) proposed a CNN-SVM hybrid for precise beer odor recognition, replacing the fully connected layer of a CNN with an SVM to enhance predictive capability and better capture complex sensory patterns79. Xiong et al. (2021) developed a convolutional spiking neural network to identify spoilage odors, integrating residual networks with spiking neurons to convert continuous sensor signals into discrete pulses. This design reduces redundant computations, requires fewer parameters than a one-dimensional-CNN, occupies less memory, and achieves >84% accuracy, making it well-suited for e-nose devices with limited computational resources80.
Although these approaches effectively model overall odor profiles, ML models that can predict odor activity values (OAVs) or quantify multi-compound synergistic, masking, or enhancing effects remain scarce. Traditional methods for studying aroma interactions, such as σ-τ diagrams, OAV calculations, S-curves, or distribution-based metrics81,82, require intensive mathematical formulation and often fail to capture nonlinear behaviors in real food matrices. ML offers new opportunities to learn these complex relationships directly from large datasets of known interactions, yet such datasets remain limited and models for multi-component interaction prediction are still in their infancy.
E-tongue systems show similar potential. ML-enabled classifiers have been developed for tea authentication83 and for predicting Pu’er tea storage time using transfer-learning approaches that leverage pre-trained models on large temporal signal datasets84. Data augmentation strategies, such as introducing controlled noise, also help mitigate small sample sizes frequently encountered in sensor-based studies84.
Despite many applications, it is worth noting that taste stimuli dissolve in saliva before reaching their respective receptors. Once dissolved, the components of saliva can interact with these stimuli and their receptors, thereby influencing taste perception85. However, solvents commonly used in food research with e-tongues often exhibit notable differences in their physicochemical properties compared to those of human saliva. Such information has been less considered in ML modeling. It is recommended that future research using the e-tongue account for the effects of saliva and consider using artificial saliva as a solvent to more accurately replicate oral environmental conditions.
Data fusion modeling
Insights from chromatography- and sensor-based modeling show that individual analytical platforms capture only fragments of the complex interactions within food matrices. Chromatography provides detailed molecular fingerprints, particularly for volatiles, whereas e-noses, e-tongues, and other sensor systems capture holistic olfactory and gustatory responses. Integrating these complementary data streams can substantially improve food property prediction by representing a broader spectrum of chemical and sensory information. Yet this richness comes with challenges, including high dimensionality, redundancy, and noise.
Data fusion addresses these limitations by combining multimodal information to enhance model robustness and predictive performance86. Fusion strategies typically operate at three levels: low, mid, and high (Fig. 4a)87,88. Low-level fusion merges raw or preprocessed signals into a unified input matrix. Mid-level fusion extracts and concatenates informative features across modalities to preserve complementary structure while reducing dimensionality. High-level fusion integrates predictions from separate models trained on different data types.
a Combining multi-source heterogeneous data at low, mid, and high levels. b Data fusion pipeline involving electroencephalogram (EEG) and functional near-infrared spectroscopy (fNIRS) data, utilizing both early fusion and late fusion strategies to improve predictive accuracy in understanding food perception. Note: E-nose electronic nose, E-tongue electronic tongue, GC-MS gas chromatography-mobility spectrometry, LC-MS liquid chromatography-mass spectrometry, VIS visible spectroscopy, NIR near-infrared spectroscopy, HIS hyperspectral imaging spectroscopy.
Applications across food quality and flavor research illustrate the advantages of this approach. For example, combining nuclear magnetic resonance, LC-MS, and GC-MS data has revealed synergistic insights into compositional changes in heat-treated apple juice89. Studies fusing MS and NIR signals demonstrate that mid-level fusion often yields higher accuracy than low-level approaches when predicting sensory attributes such as bitterness and grassiness in olive oil90. Multisensor fusion of e-nose and e-tongue signals similarly enhances flavor modeling by jointly capturing olfactory and taste information, enabling accurate classification, quality monitoring, and regression tasks91,92. Incorporating visual data from electronic eye systems further broadens sensory representation, as shown in Longjing tea evaluation, integrating aroma, taste, and color93.
While multimodal fusion outperforms unimodal methods, practical implementation remains challenging. Differences in data formats, resolutions, scales, and naming conventions can hinder integration and reduce model generalizability. Food matrices also exhibit heterogeneity and temporal variability, complicating alignment across platforms. Preprocessing strategies, including wavelet filtering, Savitzky-Golay smoothing, independent component analysis, dynamic time warping, correlation-optimized warping, and retention-time correction, help mitigate noise and reconcile instrumental shifts94,95. Batch effects arising from platform discrepancies can be reduced using standard normal variate scaling, multiplicative scatter correction, or empirical-Bayes adjustments. At a broader level, adherence to the Findable, Accessible, Interoperable, Reusable (FAIR) data principles enhances metadata harmonization and reproducibility96. Incorporating these procedures is critical for enabling accurate, scalable, and comprehensive food property profiling.
Integrating machine learning with neuroinformatics to decode perception complexity
Human flavor perception emerges from the integration of taste, smell, texture, visual cues, and trigeminal sensations, processed across distributed neural circuits. Recent advances in neuroinformatics, particularly EEG, fNIRS, and fMRI, provide non-invasive ways to probe these processes and link subjective experiences to objective neural responses16. Each modality offers complementary strengths: EEG captures rapid neural dynamics, fNIRS enables naturalistic monitoring of cortical haemodynamics, and fMRI provides fine-grained spatial localization. Together, these tools create an opportunity to build ML frameworks that decode perceptual complexity directly from brain signals and move toward individualized food prediction.
Machine learning for decoding EEG-based flavor responses
EEG is a neurophysiological technique that records brain electrical activity by measuring the electrical potential difference between the scalp and skull using electrodes placed on the scalp. These signals represent temporal variations in electrical activity across distinct brain regions. Owing to its high temporal resolution and relatively low experimental cost, EEG has been extensively used in food research to enhance understanding of the relationship between food flavor perception and individuals’ cortical processing97.
Using EEG, Yang et al. (2023) identified neurophysiological indicators associated with four tastes (sour, sweet, bitter, and salty) and their respective intensities98. They discovered that different taste qualities could be distinguished within 250-1,500 ms of stimulation, and the alpha and theta frequency bands show greater sensitivity to different tastes than do the delta, beta, and gamma bands98. These responses reflect not only sensory processing but also reward valuation and cognitive control mechanisms.
Nevertheless, analyzing extensive EEG data to discern distinctive patterns and correlations with specific tastes presents considerable challenges. The application of ML techniques could prove instrumental in addressing this challenge. Figure 3b shows the ML model for food-flavor-induced EEG signals, with an emphasis on feature extraction, model construction, and prediction. Classical ML models, including decision trees, SVMs, and k-nearest neighbors, have been successfully applied to tasks such as classifying sweet vs. non-sweet stimuli99 and distinguishing fresh vs. non-fresh foods based on spectral features20.
Deep learning further advances EEG analysis by automatically extracting multiscale temporal-spatial representations. Multi-scale CNNs enable robust taste-category recognition across sour, sweet, bitter, salty, and umami stimuli19, and data-augmentation techniques such as spatiotemporal reconstruction mitigate issues of limited sample size45. ML models have also been used to predict olfactory pleasantness from EEG signatures, with graph-based features showing particularly strong performance18,100. Collectively, these studies highlight the emerging potential of EEG-based ML systems to decode affective and perceptual dimensions of food experience.
Machine learning for decoding fNIRS and fMRI signals
Besides EEG, fNIRS and fMRI are widely used non-invasive techniques that help to evaluate food perception by providing physiological indicators of brain activity. These technologies offer higher spatial resolution than EEG, enabling the observation of brain activity through hemodynamic changes. In particular, fNIRS uses near-infrared light to penetrate the skull and quantify the alterations in cerebral cortical blood oxygenation. Although its spatial resolution is inferior to that of fMRI, fNIRS is relatively portable and places fewer constraints on participants during use, making it well-suited for practical applications101. In contrast, fMRI measures brain activity by detecting blood-oxygen-level-dependent signals, offering exceptional spatial resolution that allows precise localisation of cognitive functions within the brain cortex101. Both fNIRS and fMRI facilitate the examination of brain responses to flavor stimuli during sensory evaluation, thereby providing valuable insights into the neural processing of sensory information102,103,104,105.
Integrating ML with these imaging modalities improves classification performance and reduces reliance on manual feature engineering. CNNs trained on resting-state fMRI data can discriminate between participant groups106, and SVM models have identified brain networks distinguishing responses to high-calorie foods (potato chips) and low-calorie foods (zucchini)107. These approaches demonstrate how data-driven methods can extract perceptual information that is not readily apparent in univariate analyses.
Multimodal neuroinformatics and data fusion for food perception prediction
While various neuroinformatics technologies can extract numerous features, including time, frequency, time-frequency, and spatial features, the characteristics derived from different neuroinformatics modalities inherently possess unique strengths and limitations. For instance, EEG provides rapid responsiveness but lacks spatial localisation. In contrast, fMRI provides high spatial resolution but is highly susceptible to head and body movements. Meanwhile, fNIRS enables monitoring of the cerebral cortex surface with greater tolerance to body movements, though it has slower responsiveness and lower spatial resolution108,109. A meta-LDA classifier that combines EEG band power changes with fNIRS hemoglobin concentration data can markedly enhance the accuracy of motor imagery classification relative to the use of EEG data alone110. This suggests that combining neuroinformatics technologies through multimodal data integration, supported by advanced ML techniques, could leverage their complementary strengths.
Multimodal fusion methods in neuroinformatics can be categorized as symmetric or asymmetric. Symmetric fusion treats all modalities equally, while asymmetric fusion prioritizes one modality as a reference or constraint for another111. Feature fusion can also be classified as early or late, depending on the timing of integration. Fig. 4b illustrates the data fusion strategies for EEG and fNIRS data. Early fusion combines data from multiple sources into a unified format before processing, suitable for strongly correlated data, whereas late fusion integrates results after independent analyses, ideal for independent data sources112. For example, Sun et al. (2020) extracted features from fNIRS and EEG signals, concatenated them to create enhanced feature vectors, and classified participants’ emotional states while watching videos using an SVM algorithm113. Furthermore, an increasing number of studies have examined the influence of auditory and visual cues on food perception114,115.
Challenges and opportunities
ML at the molecular layer benefits from growing databases that support the prediction of sensory attributes, receptor activities, and toxicological profiles. However, progress is constrained by three major gaps. First, molecular databases contain substantial redundancy and inconsistent annotations due to overlapping curation sources116, complicating data integration and reducing modeling reliability. Second, sensory annotations remain narrowly focused on olfaction and taste, while trigeminal and chemesthetic dimensions are largely undocumented, with PungentDB being a rare resource linking molecules to TRP channels117 (Table 2). Third, generative models produce promising molecular suggestions but lack interpretability and are rarely validated experimentally, limiting real-world applicability58,59.
Modeling interactions among food ingredients is far more complex than predicting single-molecule properties. First, most ML studies still focus on small volatile compounds because they have well-defined structures, richer databases, and established analytical workflows. In contrast, interactions involving macromolecules, such as lipids, proteins, and polysaccharides, remain largely unexplored, even though they play critical roles in flavor release, stability, and texture. This gap arises from the structural heterogeneity of macromolecules, the strong dependence of interactions on matrix conditions (pH, temperature, ionic strength), the scarcity of high-resolution in situ datasets, and the difficulty of representing molecular and supramolecular features within unified ML frameworks. Second, although databases such as KFO-Atlas, Open Food Facts, Recipe1M, FoodRepo, and FoodData Central map food-ingredient relationships, current computational approaches offer limited chemical interpretability and struggle to incorporate reaction dynamics and spatial constraints60 (Table 2). Third, empirical data generation remains a major bottleneck: matrix effects, chromatographic variability, and isomer ambiguity compromise measurement stability and hinder reproducible quantification5. Inter-laboratory differences and inconsistent labeling further reduce dataset comparability118. Moreover, food systems evolve continuously through oxidation, enzymatic transformations, and Maillard reactions29,89, yet available datasets are predominantly static snapshots that fail to capture these temporal dynamics.
Despite growing interest in neuroinformatics, its application to food perception remains limited due to multiple structural and methodological barriers. First, large-scale EEG, fNIRS, and fMRI datasets specifically focused on flavor are scarce, and each modality has inherent limitations: EEG lacks spatial precision; fNIRS and fMRI are expensive, sensitive to motion artifacts, and difficult to deploy in naturalistic settings110. Small sample sizes, inconsistent acquisition protocols, heterogeneous hardware configurations, and culturally narrow participant pools further undermine reproducibility and model generalizability. Neuroinformatics signals are inherently noisy and highly susceptible to environmental and physiological interference, making the collection of high-quality data outside controlled laboratories challenging. Second, the field lacks standardized preprocessing pipelines and evaluation criteria. These inconsistencies reduce cross-study comparability and limit the transferability of models across populations and contexts. At the modeling level, explainability remains insufficient: existing frameworks mainly serve methodological inspection rather than providing actionable insights for product development, sensory evaluation, or consumer applications. Third, most current studies examine olfactory or gustatory pathways in isolation. However, real-world flavor perception is inherently multisensory, shaped by trigeminal stimulation as well as visual, auditory, and oral tactile cues.
Looking ahead, progress in understanding food complexity requires coordinated advances across the molecular, interaction, and perceptual levels. At the chemical level, databases should expand beyond olfactory and gustatory data to include trigeminal responses and receptor-level information. High-throughput screening, combined with carefully curated, standardized datasets, will be essential for mapping the vast space of uncharacterized food molecules and improving structure-function understanding. At the interaction level, standardization across labs, open data sharing, and unified ontologies are crucial for making data comparable and enabling reproducible models. These efforts should go hand in hand with multi-scale analytical approaches capable of capturing dynamic changes during processing and storage. At the perceptual level, there is a pressing need for large, culturally diverse neuroimaging and sensory datasets. Federated learning offers a promising solution by enabling multi-center model training without exchanging raw neurophysiological data, thereby safeguarding participant privacy while improving model generalizability across populations and experimental settings; in parallel, synthetic data generation, using generative models such as variational autoencoders, GANs, or diffusion-based approaches, can help alleviate small-sample limitations by augmenting training data with statistically realistic neural signals, reducing overfitting to site-specific noise119,120. There also remains considerable scope to develop end-user-centered, explainable AI systems that tailor model outputs to the needs of different stakeholders. For example, developing interfaces that map neural features onto sensory attributes familiar to product developers, or consumer-oriented visual summaries that communicate how flavor cues influence predicted affective responses121. Furthermore, the development of industry-wide standards, rigorous validation frameworks, and systematic cost-benefit and return-on-investment assessments will be essential for narrowing the gap between laboratory findings and real-world applications. Finally, incorporating modalities including olfactory, gustatory, trigeminal, visual, and auditory sensations through multimodal data fusion and ML frameworks will enable more comprehensive decoding of flavor responses and support future developments in personalized nutrition, virtual tasting, and immersive multisensory applications.
Advances in ML will support progress across all three layers. Ongoing development of flexible deep learning models is essential to combine diverse data types, including chemical, sensory, and neural signals, while automated feature engineering and domain-specific explainable AI can lessen dependence on manual processing and enhance interpretability. Creating multisensory models that identify nonlinear interactions among sensory pathways will be vital. Ultimately, close collaboration among food scientists, chemists, data scientists, neuroscientists, and sensory experts will accelerate the development of shared resources, standardized workflows, and integrated modeling tools, thereby advancing the field’s understanding of food complexity.
Conclusions
This review outlines the transformative potential of ML in decoding the three interrelated layers of food complexity: (1) the structural and physicochemical properties of individual food molecules; (2) the food properties arising from interactions among food components; and (3) the neurophysiological processes underlying human food perception. By harnessing a growing ecosystem of molecular, instrumental, and neuroinformatics data, ML provides a powerful means to bridge micro-level chemical data with macro-level sensory experiences.
We summarised the application of advanced ML and deep learning architectures, including CNNs, GNNs, and transformers, which are particularly well-suited to modeling the complex, nonlinear, and multimodal nature of food-related data. We also examined the growing role of neuroinformatics technologies such as EEG, fNIRS, and fMRI in decoding individual perceptual responses, and how their integration with ML can advance personalised food experience modeling. Moreover, we emphasised that integrating diverse data modalities through data fusion across chromatographic, sensor-based, imaging, and neurophysiological platforms offers a critical solution to challenges such as data sparsity, heterogeneity, and noise. These strategies enable the development of more accurate and generalisable predictive models for food quality assessment, flavor profiling, consumer preference prediction, and product authentication.
Looking forward, progress will depend on three priorities: (1) building comprehensive and standardized multimodal databases, including underrepresented dimensions such as trigeminal perception; (2) developing interpretable and hybrid ML models that balance predictive accuracy with mechanistic insight; and (3) fostering interdisciplinary collaboration across food science, chemistry, neuroscience, and data science to establish shared platforms and validation pipelines. By aligning technical development with rigorous standards and collaborative practices, we expect ML to not only advance the scientific understanding of food complexity but also accelerate its translation into and consumer-relevant applications.
Data availability
No datasets were generated or analysed during the current study.
References
Menichetti, G., Barabasi, A. L. & Loscalzo, J. Chemical complexity of food and implications for therapeutics. N. Engl. J. Med. 392, 1836–1845 (2025).
Cui, Z. et al. Artificial intelligence and food flavor: How AI models are shaping the future and revolutionary technologies for flavor food development. Compr. Rev. Food Sci. Food Saf. 24, e70068 (2025).
Zhang, D. et al. Domain knowledge, just evaluation, and robust data standards are required to advance AI in food science. Trends Food Sci. Technol. 164, 105272 (2025).
Zhang, D. Practical guide for food scientists to build AI: Data, algorithms, and applications. Food Chem. 499, 147281 (2026).
Qian, J. et al. ChemSweet: An AI-driven computational platform for next-gen sweetener discovery. Food Chem. 463, 141362 (2025).
Avellaneda-Tamayo, J. F., Chavez-Hernandez, A. L., Prado-Romero, D. L. & Medina-Franco, J. L. Chemical multiverse and diversity of food chemicals. J. Chem. Inf. Model 64, 1229–1244 (2024).
Kou, X. et al. Data-driven elucidation of flavor chemistry. J. Agric. Food Chem. 71, 6789–6802 (2023).
Liu, X., Gmitter, F. G., Grosser, J. W. & Wang, Y. Effects of rootstocks on the flavor quality of huanglongbing-affected sweet orange juices using targeted flavoromics strategy. RSC Adv. 13, 5590–5599 (2023).
Faccia, M. The flavor of dairy products from grass-fed cows. Foods 9, 1188 (2020).
Zhang, D. et al. Data-driven prediction of molecular biotransformations in food fermentation. J. Agric. Food Chem. 71, 8488–8496 (2023).
Zhang, D. et al. Unveiling the chemical complexity of food-risk components: A comprehensive data resource guide in 2024. Trends Food Sci. Technol. 148, 104513 (2024).
Jiang, S., Ni, C., Chen, G. & Liu, Y. A novel data fusion strategy based on multiple intelligent sensory technologies and its application in the quality evaluation of Jinhua dry-cured hams. Sens. Actuators B: Chem. 344, 130324 (2021).
Ross, C. F. Considerations of the use of the electronic tongue in sensory science. Curr. Opin. Food Sci. 40, 87–93 (2021).
Feng, Y. et al. A mechanistic review on machine learning-supported detection and analysis of volatile organic compounds for food quality and safety. Trends Food Sci. Technol. 143, 104297 (2024).
Sipos, L., Nyitrai, Á, Hitka, G., Friedrich, L. F. & Kókai, Z. Sensory panel performance evaluation—Comprehensive review of practical approaches. Appl. Sci. 11, 11977 (2021).
Han, P. Advances in research on brain processing of food odors using different neuroimaging techniques. Curr. Opin. Food Sci. 42, 134–139 (2021).
Zhao, Q. et al. An advance in novel intelligent sensory technologies: From an implicit-tracking perspective of food perception. Compr. Rev. Food Sci. Food Saf. 23, e13327 (2024).
Becerra, M. A. et al. Odor Pleasantness Classification from Electroencephalographic Signals and Emotional States. In: Advances in Computing (eds Jairo E. Serrano C. & J. C. Martínez-Santos). Springer International Publishing (2018).
Gao, H. et al. Basic taste sensation recognition from EEG based on multiscale convolutional neural network with residual learning. IEEE Trans. Instrum. Meas. 72, 1–10 (2023).
Cui, Z. et al. TastePeptides-EEG: An ensemble model for umami taste evaluation based on electroencephalogram and machine learning. J. Agric. Food Chem. 71, 13430–13439 (2023).
Qiao, G. et al. Food recommendation towards personalized wellbeing. Trends Food Sci. Technol. 156, 104877 (2025).
Jiao, X. et al. Artificial intelligence in smart seafood safety across the supply chains: Recent advances and future prospects. Trends Food Sci. Technol. 163, 105161 (2025).
Dähne, S. et al. Multivariate machine learning methods for fusing multimodal functional neuroimaging data. Proc. IEEE 103, 1507–1530 (2015).
Shi, P. et al. AI-driven exploration of microbial resources in fermented foods. Trends Food. Sci. Technol. 167, 105450 (2026).
Yu, S. et al. Characterization of selected Chinese soybean paste based on flavor profiles using HS-SPME-GC/MS, E-nose and E-tongue combined with chemometrics. Food Chem. 375, 131840 (2022).
Tseng, Y. J., Chuang, P.-J. & Appell, M. When machine learning and deep learning come to the big data in food chemistry. ACS omega 8, 15854–15864 (2023).
Sarker, I. H. Machine learning: Algorithms, real-world applications and research directions. SN Comput. Sci. 2, 160 (2021).
Cortes, C. & Vapnik, V. Support-vector networks. Mach. Learn. 20, 273–297 (1995).
Gerhardt, N. et al. Quality assessment of olive oils based on temperature-ramped HS-GC-IMS and sensory evaluation: Comparison of different processing approaches by LDA, kNN, and SVM. Food Chem. 278, 720–728 (2019).
Sun, Z. et al. An exploration of pepino (Solanum muricatum) flavor compounds using machine learning combined with metabolomics and sensory evaluation. Foods 11, 3248 (2022).
Breiman, L., Friedman, J., Olshen, R. A. & Stone, C. J. Classification and regression trees. (CRC Press, 1984).
Zhang, S. et al. A transfer learning approach to predict the combined toxicity of mycotoxins with limited data. Food Biosci. 74, 108060 (2025).
Wang, Y. et al. Prediction of flavor and retention index for compounds in beer depending on molecular structure using a machine learning method. RSC Adv. 11, 36942–36950 (2021).
Hu, B. et al. Unraveling the relationship between key aroma components and sensory properties of fragrant peanut oils based on flavoromics and machine learning. Food Chem.: X 20, 100880 (2023).
Cover, T. & Hart, P. Nearest neighbor pattern classification. IEEE Trans. Inf. Theory 13, 21–27 (1967).
Wu, X. et al. Discrimination of Chinese liquors based on electronic nose and fuzzy discriminant principal component analysis. Foods 8, 38 (2019).
Zhang, D. et al. Discovery of toxin-degrading enzymes with positive unlabeled deep learning. Acs Catal. 14, 3336–3348 (2024).
Zornetzer, S. F., Davis, J. & Lau, C. An introduction to neural and electronic networks. (Academic Press, San Diego, 1990).
Lee, J., Song, S. B., Chung, Y. K., Jang, J. H. & Huh, J. BoostSweet: Learning molecular perceptual representations of sweeteners. Food Chem. 383, 132435 (2022).
LeCun, Y. Gradient-based learning applied to document recognition. Proc. IEEE 86, 2278–2324 (1998).
Potărniche, I.-A., Saroși, C., Terebeș, R. M., Szolga, L. & Gălătuș, R. Classification of food additives using UV spectroscopy and one-dimensional convolutional neural network. Sensors 23, 7517 (2023).
Wang, S. et al. Synergetic application of an E-tongue, E-nose and E-eye combined with CNN models and an attention mechanism to detect the origin of black pepper. Sens. Actuators A: Phys. 357, 114417 (2023).
Wu, D., Luo, D., Wong, K.-Y. & Hung, K. P. O. P.-C. N. N. Predicting odor pleasantness with convolutional neural network. IEEE Sens. J. 19, 11337–11345 (2019).
Zhang, X., Hou, H. & Meng, Q. EEG-based odor recognition using channel-frequency convolutional neural network. In: 2019 Chinese Control Conference (CCC)). IEEE (2019).
Xia, X., Yang, Y., Shi, Y., Zheng, W. & Men, H. Decoding human taste perception by reconstructing and mining temporal-spatial features of taste-related EEGs. Appl. Intell. 54, 3902–3917 (2024).
Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M. & Monfardini, G. The Graph Neural Network Model. IEEE Trans. Neural Netw. 20, 61–80 (2009).
Song, Y. et al. A Comprehensive Comparative Analysis of Deep Learning Based Feature Representations for Molecular Taste Prediction. Foods 12, 3386 (2023).
Rumelhart, D. E., Hinton, G. E. & Williams, R. J. Learning representations by back-propagating errors. Nature 323, 533–536 (1986).
Qi, L. et al. Umami-MRNN: Deep learning-based prediction of umami peptide using RNN and MLP. Food Chem. 405, 134935 (2023).
Das, S., Tariq, A., Santos, T., Kantareddy, S. S. & Banerjee, I. Recurrent neural networks (RNNs): Architectures, training tricks, and introduction to influential research. Mach. Learn. Brain Disord. 197, 117–138 (2023).
Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural Comput. 9, 1735–1780 (1997).
Jiang, J. et al. A machine learning method to identify umami peptide sequences by using multiplicative LSTM embedded features. Foods 12, 1498 (2023).
Miller, C., Hamilton, L. & Lahne, J. Sensory descriptor analysis of whisky lexicons through the use of deep learning. Foods 10, 1633 (2021).
Gona, S. N. R. & Marellapudi, H. Suggestion and invention of recipes using bi-directional LSTMs-based frameworks. SN Appl. Sci. 3, 1–17 (2021).
Vaswani, A. et al. Attention is all you need. Adv. neural Inf. Process. Syst. 30, 6000–6010 (2017).
Chew, R., Wenger, M., Guillory, J., Nonnemaker, J. & Kim, A. Identifying electronic nicotine delivery system brands and flavors on Instagram: natural language processing analysis. J. Med. Internet Res. 24, e30257 (2022).
Ma, P. et al. Large language models in food science: Innovations, applications, and future. Trends Food Sci. Technol. 148, 104488 (2024).
Queiroz, L. P. et al. Generating flavor molecules using scientific machine learning. ACS omega 8, 10875–10887 (2023).
Aleixandre, M., Prasetyawan, D. & Nakamoto, T. Generative diffusion network for creating scents. IEEE Access 13, 57311–57321 (2025).
Zhang, J. et al. Molecular atlas of key food odorants reveals structured aroma organization and enables generative aroma design. bioRxiv (2026).
Hooton, F., Menichetti, G. & Barabási, A.-L. Exploring food contents in scientific literature with FoodMine. Sci. Rep. 10, 16191 (2020).
Schilling-Wilhelmi, M. et al. From text to insight: large language models for chemical data extraction. Chem. Soc. Rev. 54, 1125–1150 (2025).
Blin, K. et al. antiSMASH 8.0: extended gene cluster detection capabilities and analyses of chemistry, enzymology, and regulation. Nucleic Acids Res. 53, W32–W38 (2025).
Skinnider, M. A. et al. Comprehensive prediction of secondary metabolite structure and biological activity from microbial genome sequences. Nat. Commun. 11, 6058 (2020).
Zhao, Y., Xia, Y., Yu, Y. & Liang, G. QSAR in natural non-peptidic food-related compounds: Current status and future perspective. Trends Food Sci. Technol. 140, 104165 (2023).
Kar, S., Roy, K. & Leszczynski, J. in Advances in QSAR Modeling: Applications in Pharmaceutical, Chemical, Food, Agricultural and Environmental Sciences (ed K. Roy) 203-302 (Springer International Publishing, 2017).
Shang, L., Liu, C., Tomiura, Y. & Hayashi, K. Machine-learning-based olfactometer: Prediction of odor perception from physicochemical features of odorant molecules. Anal. Chem. 89, 11999–12005 (2017).
Ollitrault, G. et al. Pred-O3, a web server to predict molecules, olfactory receptors and odor relationships. Nucleic Acids Res. 52, W507–W512 (2024).
Lee, B. K. et al. A principal odor map unifies diverse tasks in olfactory perception. Science 381, 999–1006 (2023).
Song, R., Liu, K., He, Q., He, F. & Han, W. Exploring Bitter and Sweet: The Application of Large Language Models in Molecular Taste Prediction. J. Chem. Inf. Model. 64, 4102–4111 (2024).
Karabagias, I. K. & Nayik, G. A. Machine learning algorithms applied to semi-quantitative data of the volatilome of citrus and other nectar honeys with the use of HS-SPME/GC–MS analysis, lead to a new index of geographical origin authentication. Foods 12, 509 (2023).
Gan, Y. et al. Using HS-GC-MS and flash GC e-nose in combination with chemometric analysis and machine learning algorithms to identify the varieties, geographical origins and production modes of Atractylodes lancea. Ind. Crops Prod. 209, 117955 (2024).
Brendel, R., Schwolow, S., Rohn, S. & Weller, P. Gas-phase volatilomic approaches for quality control of brewing hops based on simultaneous GC-MS-IMS and machine learning. Anal. Bioanal. Chem. 412, 7085–7097 (2020).
Chakraborty, P. et al. Non-destructive method to classify walnut kernel freshness from volatile organic compound (VOC) emissions using gas chromatography-differential mobility spectrometry (GC-DMS) and machine learning analysis. Appl. Food Res. 3, 100308 (2023).
Zhu, W., Benkwitz, F. & Kilmartin, P. A. Volatile-based prediction of sauvignon blanc quality gradings with static headspace–gas chromatography–ion mobility spectrometry (SHS–GC–IMS) and interpretable machine learning techniques. J. Agric. Food Chem. 69, 3255–3265 (2021).
Bi, K., Zhang, D., Qiu, T. & Huang, Y. GC-MS fingerprints profiling using machine learning models for food flavor prediction. Processes 8, 23 (2020).
Schreurs, M. et al. Predicting and improving complex beer flavor through machine learning. Nat. Commun. 15, 2368 (2024).
Li, Y. et al. Physicochemical parameters combined flash GC e-nose and artificial neural network for quality and volatile characterization of vinegar with different brewing techniques. Food Chem. 374, 131658 (2022).
Shi, Y. et al. A deep feature mining method of electronic nose sensor data for identifying beer olfactory information. J. food Eng. 263, 437–445 (2019).
Xiong, Y. et al. An odor recognition algorithm of electronic noses based on convolutional spiking neural network for spoiled food identification. J. Electrochem. Soc. 168, 077519 (2021).
Niu, Y. et al. Characterization of odor-active volatiles and odor contribution based on binary interaction effects in mango and vodka cocktail. Molecules 25, 1083 (2020).
Niu, Y., Zhang, J., Xiao, Z. & Zhu, J. Evaluation of the perceptual interactions between higher alcohols and off-odor acids in Laimao Baijiu by σ–τ plot and partition coefficient. J. Agric. food Chem. 68, 14938–14949 (2020).
Ren, G., Li, T., Wei, Y., Ning, J. & Zhang, Z. Estimation of Congou black tea quality by an electronic tongue technology combined with multivariate analysis. Microchem. J. 163, 105899 (2021).
Yang, Z. et al. Employment of an electronic tongue combined with deep learning and transfer learning for discriminating the storage time of Pu-erh tea. Food Control 121, 107608 (2021).
Martin, L. E., Gutierrez, V. A. & Torregrossa, A.-M. The role of saliva in taste and food intake. Physiol. Behav. 262, 114109 (2023).
Li, Q. et al. Machine learning technique combined with data fusion strategies: A tea grade discrimination platform. Ind. Crops Prod. 203, 117127 (2023).
Azcarate, S. M., Ríos-Reina, R., Amigo, J. M. & Goicoechea, H. C. Data handling in data fusion: Methodologies and applications. TrAC Trends Anal. Chem. 143, 116355 (2021).
Jin, G. et al. Tracing the origin of Taiping Houkui green tea using 1H NMR and HS-SPME-GC–MS chemical fingerprints, data fusion and chemometrics. Food Chem. 425, 136538 (2023).
Alves Filho, E. G. et al. An integrated analytical approach based on NMR, LC–MS and GC–MS to evaluate thermal and non-thermal processing of cashew apple juice. Food Chem. 309, 125761 (2020).
Borràs, E. et al. Prediction of olive oil sensory descriptors using instrumental data fusion and partial least squares (PLS) regression. Talanta 155, 116–123 (2016).
Qiu, S., Wang, J., Tang, C. & Du, D. Comparison of ELM, RF, and SVM on E-nose and E-tongue to trace the quality status of mandarin (Citrus unshiu Marc.). J. food Eng. 166, 193–203 (2015).
Haddi, Z. et al. E-Nose and e-Tongue combination for improved recognition of fruit juice samples. Food Chem. 150, 246–253 (2014).
Xu, M., Wang, J. & Zhu, L. The qualitative and quantitative assessment of tea quality based on E-nose, E-tongue and E-eye combined with chemometrics. Food Chem. 289, 482–489 (2019).
Calvini, R. & Pigani, L. Toward the development of combined artificial sensing systems for food quality evaluation: A review on the application of data fusion of electronic noses, electronic tongues and electronic eyes. Sensors 22, 577 (2022).
Adade, S. Y.-S. S. et al. Advanced food contaminant detection through multi-source data fusion: Strategies, applications, and future perspectives. Trends Food Sci. Technol. 156, 104851 (2025).
Brinkley, S. et al. The state of food composition databases: Data attributes and FAIR data harmonization in the era of digital innovation. Front. Nutr. 12, 1552367 (2025).
Songsamoe, S., Saengwong-ngam, R., Koomhin, P. & Matan, N. Understanding consumer physiological and emotional responses to food products using electroencephalography (EEG). Trends Food Sci. Technol. 93, 167–173 (2019).
Yang, T. et al. Insights into brain perceptions of the different taste qualities and hedonic valence of food via scalp electroencephalogram. Food Res. Int. 173, 113311 (2023).
Romeo-Arroyo, E. et al. Exploratory Research on Sweetness Perception: Decision Trees to Study Electroencephalographic Data and Its Relationship with the Explicit Response to Sweet Odor, Taste, and Flavor. Sensors 22, 6787 (2022).
Xia, X. et al. Recognition of odor and pleasantness based on olfactory EEG combined with functional brain network model. Int. J. Mach. Learn. Cybern. 14, 2761–2776 (2023).
Pereira, J., Direito, B., Luhrs, M., Castelo-Branco, M. & Sousa, T. Multimodal assessment of the spatial correspondence between fNIRS and fMRI hemodynamic responses in motor tasks. Sci. Rep. 13, 2244 (2023).
Okamoto, M. et al. Prefrontal activity during flavor difference test: application of functional near-infrared spectroscopy to sensory evaluation studies. Appetite 47, 220–232 (2006).
Minematsu, Y., Ueji, K. & Yamamoto, T. Activity of frontal pole cortex reflecting hedonic tone of food and drink: fNIRS study in humans. Sci. Rep. 8, 16197 (2018).
Pazart, L., Comte, A., Magnin, E., Millot, J.-L. & Moulin, T. An fMRI study on the influence of sommeliers’ expertise on the integration of flavor. Front. Behav. Neurosci. 8, 358 (2014).
Ai, Y. & Han, P. Neurocognitive mechanisms of odor-induced taste enhancement: A systematic review. Int. J. Gastronomy Food Sci. 28, 100535 (2022).
Ho, M.-C., Shen, H.-A., Chang, Y.-P. E. & Weng, J.-C. A CNN-based autoencoder and machine learning model for identifying betel-quid chewers using functional MRI features. Brain Sci. 11, 809 (2021).
Mendez-Torrijos, A. et al. Snack food as a modulator of human resting-state functional connectivity. CNS Spectr. 23, 321–332 (2018).
Dashtestani, H. et al. Structured sparse multiset canonical correlation analysis of simultaneous fNIRS and EEG provides new insights into the human action-observation network. Sci. Rep. 12, 6878 (2022).
Luo, N., Shi, W., Yang, Z., Song, M. & Jiang, T. Multimodal fusion of brain imaging data: Methods and applications. Mach. Intell. Res. 21, 136–152 (2024).
Fazli, S. et al. Enhanced performance by a hybrid NIRS–EEG brain computer interface. Neuroimage 59, 519–529 (2012).
Uludağ, K. & Roebroeck, A. General overview on the merits of multimodal neuroimaging data fusion. Neuroimage 102, 3–10 (2014).
Zhang, Y.-D. et al. Advances in multimodal data fusion in neuroimaging: Overview, challenges, and novel orientation. Inf. Fusion 64, 149–187 (2020).
Sun, Y., Ayaz, H. & Akansu, A. N. Multimodal affective state assessment using fNIRS+ EEG and spontaneous facial expression. Brain Sci. 10, 85 (2020).
Motoki, K., Spence, C. & Velasco, C. When visual cues influence taste/flavour perception: A systematic review. Food Qual. Preference 111, 104996 (2023).
Taylor, A. J. et al. Factors affecting flavor perception in space: Does the spacecraft environment influence food intake by astronauts? Compr. Rev. food Sci. food Saf. 19, 3439–3475 (2020).
Yonchev, D., Dimova, D., Stumpfe, D., Vogt, M. & Bajorath, J. Redundancy in two major compound databases. Drug Discov. Today 23, 1183–1186 (2018).
Chen, Z., Li, J., Hou, N., Zhang, Y. & Qiao, Y. TCM-Blast for traditional Chinese medicine genome alignment with integrated resources. BMC Plant Biol. 21, 339 (2021).
Liu, R. et al. Systematic investigation into matrix effect compensation in the GC-MS analysis of flavor components using analyte protectants. Talanta 291, 127818 (2025).
Fendor, Z. et al. Federated learning in food research. J. Agric. Food Res. 23, 102238 (2025).
Zheng, W., Yuan, Q., Zhang, A., Lei, Y. & Pan, G. Data augmentation of flavor information for electronic nose and electronic tongue: An olfactory-taste synesthesia model combined with multiblock reconstruction method. Expert Syst. Appl. 272, 126810 (2025).
Zhang, J. et al. Mapping sleep-promoting volatiles in aromatic plants with machine learning: A comprehensive survey of 2300 molecules. Digital Discovery, in press (2026).
Dutta, P., Jain, D., Gupta, R. & Rai, B. Classification of tastants: A deep learning based approach. Mol. Inform. 42, e202300146 (2023).
Vega-Márquez, B., Nepomuceno-Chamorro, I., Jurado-Campos, N. & Rubio-Escudero, C. Deep learning techniques to improve the performance of olive oil classification. Front. Chem. 7, 929 (2020).
Guo, Y., Xia, X., Shi, Y., Ying, Y. & Men, H. Olfactory EEG induced by odor: Used for food identification and pleasure analysis. Food Chem. 455, 139816 (2024).
Acknowledgements
This project was supported by the Collaborative Innovation Center of Fragrance Flavor and Cosmetics, the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (A-8003718-00-00), and the Start-Up Grant of the National University of Singapore (A-0010237-00-00). The authors thank the anonymous reviewers for their valuable comments. The authors acknowledged using an AI tool (ChatGPT) for text polishing and grammar check. The author is fully responsible for the content and conclusions of the manuscript.
Author information
Authors and Affiliations
Contributions
X. K. and D.Z. designed the research. Q.K., J.Z., X. H., X.K., and D.Z. wrote the initial manuscript. J.Z., Q.K., X.K., and D.Z. rechecked the manuscript and participated in manuscript revision. All authors approved the final paper.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Ke, Q., Zhang, J., Huang, X. et al. Machine learning unveils three layers of food complexity. npj Sci Food 10, 87 (2026). https://doi.org/10.1038/s41538-026-00730-w
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s41538-026-00730-w
