Artificial Intelligence in Metabolomics: A Current Review

Abstract

Metabolomics and artificial intelligence (AI) form a synergistic partnership. Metabolomics generates large datasets comprising hundreds to thousands of metabolites with complex relationships. AI, aiming to mimic human intelligence through computational modeling, possesses extraordinary capabilities for big data analysis. In this review, we provide a recent overview of the methodologies and applications of AI in metabolomics studies in the context of systems biology and human health. We first introduce the AI concept, history, and key algorithms for machine learning and deep learning, summarizing their strengths and weaknesses. We then discuss studies that have successfully used AI across different aspects of metabolomic analysis, including analytical detection, data preprocessing, biomarker discovery, predictive modeling, and multi-omics data integration. Lastly, we discuss the existing challenges and future perspectives in this rapidly evolving field. Despite limitations and challenges, the combination of metabolomics and AI holds great promises for revolutionary advancements in enhancing human health.

1. Introduction

Metabolomics, broadly including lipidomics, is a powerful approach to study the metabolome, compositions and interactions of small molecule metabolites within a biological system [1, 2]. Metabolome impacts or is impacted by virtually every cellular process, and altered metabolism is one of the hallmarks for many diseases, nutrition intervention, environmental exposure, etc. [3] The metabolites of interest in metabolomics include sugars, amino acids, lipids, and other organic compounds essential for cellular functions [4]. In terms of metabolic reactions, they can be broadly categorized into two major types: catabolism, breaking down molecules to yield energy; and anabolism, synthesis of compounds for cellular functions [5]. In addition, metabolism participates in deactivation, detoxification, and elimination of exogenous substances [6]. To study the whole metabolome, metabolomics measures thousands of metabolite compositions and abundances in parallel, which has emerged as a highly useful approach to examine metabolic pathways, biochemical reactions, and physiological transformations happening within biological systems [7–9].

Metabolomics often results in large datasets. The two most common analytical technologies in metabolomics are nuclear magnetic resonance (NMR) spectroscopy and mass spectrometry (MS), and MS is often coupled to liquid or gas chromatography (LC-MS, GC-MS) [10]. These technologies can detect hundreds to thousands of metabolites and/or metabolic features from a typical biological sample, paving the way for large-scale exploration of a wide range of metabolome [11]. Based on the large datasets, metabolomic studies can comprehensively analyze the metabolic profiles of biological systems, providing insights into the biochemical processes underpinning internal and/or external stimuli [11–14].

Metabolomics has shown great potential in a wide variety of applications. For example, metabolomics was used to identify metabolic signatures associated with specific diseases, and the potential metabolite biomarkers were used for disease diagnosis, prognosis, and monitoring [8, 15–17]. Roche group and Scalbert group reviewed the applications of metabolomics to improve human health, covering various diseases and human nutrition sciences [14, 18]. Armitage and Ciborowski, as well as Xie et al., discussed how metabolomics was used in cancer research, including the identification of cancer biomarkers and metabolic pathways associated with cancer mechanisms [19, 20]. Botas et al. reviewed metabolomics studies in neurodegenerative diseases, such as Alzheimer’s and Parkinson’s diseases [21]. In addition, metabolomics can discover new drug candidates and study drug metabolism, toxicology, efficacy, and potential side effects [22–25]. Kell et al. reviewed the applications of metabolomics in targeted drug discovery [22]. Robertson et al. discussed the challenges and opportunities of metabolomics in drug metabolism and pharmaceutical development [24]. Furthermore, Chen et al. reviewed metabolomics in studying the metabolic products of human gut microbiome [25]. We also published a comprehensive review of recent studies investigating xenobiotic exposures of wide interest and their effects on the gut microbiome and metabolome [26]. Notably, metabolomics offers a functional assessment of cellular activity, and from a systems biology point of view, it can be integrated with various omics domains (such as genes and proteins) to provide a “big picture” of a biological system [27–29]. Both metabolomics itself and this complex interplay require sophisticated data analytical approaches.

Artificial intelligence (AI) makes uses of machines and/or software to perform tasks that typically require human intelligence, such as understanding natural languages, recognizing patterns, making decisions, and solving problems [30–32]. When using complex datasets, AI has demonstrated remarkable proficiency in recognizing hidden patterns, discerning informative features amidst noisy backgrounds, and delivering accurate predictions [33, 34]. These capabilities make it exceptionally well suited for metabolomic data analysis. Indeed, innovative AI algorithms have been developed to process raw mass-spectrometry data, detect chromatogram peaks, identify metabolites, discover biomarkers, build predictive models, construct pathways, and other tasks that challenge traditional statistical approaches [35–37]. In addition, AI shows significant promises in integrative multi-omics analysis, thereby advancing the field of systems biology.

Current literatures have comprehensive reviews of metabolomics techniques and applications [3, 24, 38–42], as well as development and applications of AI in fundamental and applied sciences [43–47]. In this article, we present a systematic review of AI techniques and applications in metabolomic studies in the framework of systems biology. We first introduce the AI concept, history, and landscape of the AI algorithms that are integral to big data analysis. We then present a comprehensive review of AI applications across different aspects of metabolomics studies, highlighting the unique characteristics of metabolomics data and important considerations for effective analysis and appropriate interpretation. We further expand the discussion to underscore the roles of AI in integrative multi-omics analysis. Finally, we discuss the existing challenges and future perspectives in this rapidly evolving field, outlining potential directions for AI-driven metabolomics.

2. AI History, Concepts, and Techniques in Big Data Analysis

Since November 2022, many of us have been greatly impressed by Chat Generative Pre-trained Transformer (ChatGPT), which uses natural language processing (NLP) to create humanlike conversational dialogues [48–50]. ChatGPT is a typical example of AI to simulate and replicate human intelligence, and in fact, AI can do much more.

As shown in Figure 1, the origin of AI can be tracked back to 1950s when the field of computer science just emerged. In 1956, the term “artificial intelligence” was formally introduced at the historic Dartmouth Conference [51]. Early AI systems were used for symbolic reasoning, problem-solving, mathematics, and logical thinking [52]. Interestingly, one of the first AI applications, the Dendral project (1965), was developed to analyze MS data to identify organic compounds [53]. Between 1970s-1980s, the expert systems, as one of the first practical applications of AI, were created by Edward Feigenbaum to simulate decisions of human experts [54, 55]. For example, MYCIN (1976) and Prospector (1983) were well-known expert systems used for medical diagnosis and analyzing geological data in mineral exploration, respectively [35, 56]. Since 1990s, AI techniques, mainly including machine learning (ML) and neural networks (NNs), have gained wide applications in data analysis and pattern recognition. For instance, ML was used for sequence analysis and identifying genetic markers in genomics, and clustering algorithms were employed in chemistry for compound classification [57–59]. Cartwright discussed the impact of ML in chemistry, covering applications in molecular modeling, property prediction, and drug discovery [60]. From 2010s to now, the explosion of big data and advances in computing power facilitated the use of deep learning (DL) techniques. Deep NNs were employed in image analysis for tasks like medical imaging and astronomy [61, 62]. In climate science, AI is able to analyze large datasets to model and predict complex climate patterns [63, 64]. In addition, AI has been used to develop and control robots for data collection in extreme environments, such as deep-sea exploration and planetary exploration [65]. AI-assisted robotic systems have automated high-throughput experimentation in biology and chemistry [66, 67]. Further, AI was used to optimize quantum algorithms and extract valuable information from scientific literature using NLP, which is highly useful in precise and high-throughput drug discovery [68].

Figure 1.

History of artificial intelligence and its connections to metabolomics.

Currently, AI continues to rapidly grow and diversify, with applications in various disciplines, from biology and chemistry to astronomy, climate science, and beyond [69]. Petersen et al. reviewed the use of AI in drug discovery, biomarker identification, and personalized medicine [70]. Lai et al. reviewed how AI and ML are used for analyzing genomics data, including gene expression, DNA sequencing, and variant calling [71]. Impressively, AI has outperformed human intelligence in tasks that demand processing vast volumes of data, discerning intricate relationships, and making rapid decisions, where human cognitive processes may be comparatively slower or more prone to errors [72]. As the technology advances, it is expected that in the near future AI will bring revolutionary improvements in many sectors including banking, transportation, healthcare, etc.

2.1. Machine Learning (ML)

As shown in Figure 1, ML is a subfield of AI that focuses on developing algorithms and models that enable computers to learn from the data and improve their performance without manual intervention [73]. There are 3 major types of ML algorithms, supervised learning, unsupervised learning, and reinforcement learning (Table 1).

Table 1.

Major types of ML algorithms

Supervised Learning

Linear classifiers: linear regression, logistic regression (LogReg), mixed-effect regression, and other generalized linear models, partial least squares (PLS) regression, partial least squares discriminant analysis (PLS-DA), regularized regression

Non-linear classifiers: K-nearest neighbors (KNN), decision trees, random forest (RF), gradient boosting, support vector machine (SVM), naive Bayes, artificial neural network (ANN)

Unsupervised Learning

Clustering: K-means clustering, hierarchical clustering, self-organizing map (SOM)

Dimension reduction: principal component analysis (PCA), Independent Component Analysis (ICA), t-Distributed Stochastic Neighbor Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP)

Reinforcement Learning

Q-learning, Deep Q-Networks (DQN), Policy Gradients, Actor-Critic, Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradients (DDPG), Trust Region Policy Optimization (TRpO)

In supervised learning, the objective is to learn the mapping from input features to output labels [74, 75]. Once a model is trained, it can make predictions on new, unseen data. Supervised learning has been widely used in metabolomics studies. For example, Catharino’s group combined MS with supervised ML analyses to discover biomarkers predictive of body weight [76]. Tiedt et al. used supervised ML and found a set of four metabolites that accurately distinguished between patients with ischemic stroke and those with stroke mimics [77]. Notably, ensemble learning can be employed to improve predictions from multiple models, further enhancing the overall performance and robustness [78].

Unsupervised learning does not use output labels to guide the learning process. The objective is to unveil inherent structures in the data, solely based on the input features [79]. Clustering is a common unsupervised learning task, which aims to partition data into groups based on their similarities. The discovered clustering patterns can aid in identifying novel sample subtypes, exploring correlated features, detecting anomalies, and revealing other distribution characteristics worthy of further examination [80]. Another unsupervised learning application is dimensionality reduction, which transforms data from a high dimensional space onto a low dimensional space while preserving the most meaningful information of the original data [81]. Data in the reduced dimensions can be easily visualized to inspect structures, and the transformed features have lower redundancy and higher information content compared to the untransformed ones [82]. Dimension reduction algorithms and clustering algorithms are often utilized together to explore and visualize structures in a metabolomic dataset, such as grouping metabolites with similar concentration profiles [83, 84], discovering sample subsets with similar metabolic profiles [85], identifying novel compounds [86], and detecting unexpected patterns [87].

In reinforcement learning, the task is to make sequential decisions that are robust to stochasticity, uncertainty, partial observability, and other challenges in a complex environment [88]. The algorithm optimizes a model via interactions with the environment in trials and errors, during which desired behaviors are rewarded and undesired ones are penalized [89]. While reinforcement learning has achieved great successes in fields like robotics, game playing, and autonomous systems, its application in systems biology is still in its early stages [90]. The demand for substantial amounts of data and extensive training makes it computationally intensive and impractical in biomedical scenarios with limited data and high sample complexity [91]. Therefore, simplification of complex real-world biological mechanisms into more manageable systems is frequently necessary. For instance, Hu et al. selectively incorporated the influxes/outfluxes of six bile acids across six organs as parameters in a differential equation and employed reinforcement learning to approximate the body’s adjustment, treating homeostasis as a reward signal [92]. This model successfully predicted the disease progression of primary sclerosing cholangitis.

2.2. Deep Learning (DL)

As shown in Figure 1, DL is a subset of ML algorithms that use multiple layers of interconnected artificial neurons to simulate the way how the human brain works [93]. Compared to traditional ANNs where neurons are organized in a few fully-connected feedforward layers, DL models feature diverse types of neurons arranged in deep hierarchical structures [94]. This design allows DL models to automatically capture complex information within the data, and it has been used for supervised, unsupervised, or reinforcement learning tasks that are challenging for traditional ML approaches [95]. Leveraging these powerful and versatile DL technologies, myriads of models have been built to analyze and integrate genomic, transcriptomic, and epigenetic data [96, 97], as well as to develop novel drugs [98].

Recently, these DL models have also begun to find applications in metabolomics. For example, a convolutional neural network (CNN) can autonomously construct informative features from spatiotemporally distributed signals, eliminating the need for time-consuming manual feature engineering [99]. Kim et al. designed a CNN to process ¹H-¹³C NMR spectra, achieving precise detection of peaks in complex mixtures [100]. ChromAlignNet uses the recurrent neural network (RNN) technology, which allows information to be memorized and passed over multiple layers to form feedback loops [101], and it demonstrated superior accuracy over traditional methods to align GC-MS chromatograms [102]. Choudhury et al. built kinetic models for dynamical metabolism analysis [103] using a generative adversarial network (GAN) that enables a competition between two NNs, one trained on real samples and the other trained on synthesized artificial samples, to enhance model performance [104].

Attention mechanisms can significantly improve the performance of a DL model by selectively focusing on important elements in the input [105]. A groundbreaking development in DL architecture was the introduction of a model solely based on attention, known as Transformer [106]. The Transformer model utilizes a self-attention mechanism, allowing it to weigh different parts of the input sequence, which enables it to effectively capture long-range dependencies and surpass traditional RNNs in various tasks [107]. In fact, ChatGPT, as mentioned above, harnesses the Transformer architecture to comprehend and generate human-like texts [108]. Using the Transformer framework, Shrivastava et al. developed a MassGenie tool to explore millions of fragmentation patterns in silico, which showed an unprecedented ability to identify de novo small molecules from MS data [109].

3. AI in Metabolomics

3.1. Overview

Applying AI models on metabolomics data suits well because of large datasets and complex patterns among them. The integration of AI in metabolomics has greatly improved data analysis, biological interpretation, and predictive statistical modeling [110, 111]. As a result, AI-assisted metabolomics has made significant discoveries in diverse applications including disease diagnosis [112], precision medicine [113], drug discovery [114], etc.

Figure 2 depicts the general metabolomics workflow. Metabolomics is a comparative science, and it needs samples from two or more groups under different endogenous and/or exogenous conditions [10]. Various kinds of biological samples have been used, such as blood, urine, tissue, and cells. These samples are subject to untargeted (broad screening) and/or targeted (only metabolites of interest) metabolomics analysis. NMR and LC/GC-MS are common analytical technologies to measure a wide range of metabolites, generating large datasets. These analytical platforms require different sample preparation protocols to extract metabolites and improve detection [115–117]; for example, chemical derivatization is often performed to volatilize the compounds prior to GC-MS detection [118]. After collection of the raw data, data will be extracted, which includes steps of background correction, peak picking, peak deconvolution, peak integration, etc. In addition, metabolites are identified by comparing their experimental data to chemical standards, reference libraries or databases, and their concentrations can be measured in targeted measurement [119]. Prior to statistical analysis, data pretreatment involves baseline correction, binning, scaling, and/or normalization. In statistical modeling, both univariate and multivariate statistics can be used to discover significant metabolites. These metabolites are then mapped to metabolic pathways and integrated with other omics data to reconstruct the molecular networks. An optimal set of metabolites and/or genes/transcripts/proteins can be obtained to accurately distinguish samples under different conditions [120]. Further, internal and external cross validations are required to ensure study rigor in metabolomics studies [121]. In the following sections, we will discuss in details how AI can be applied to different steps in metabolomics.

Figure 2.

A general workflow in metabolomics.

3.2. AI in Analytical Detection

LC-MS and GC-MS data contain abundances of not only identified metabolites, but also those without annotations [122]. To identify these unknows, AI is highly efficient to compare the mass spectra and retention times (RTs) from the experiments with authentic chemical standards as well as those present in existing databases [123]. For example, Jang et al. constructed a standalone software named AI-SIDA (AI-screener of illicit drugs and analogues), which provided a very powerful platform for screening unknown erectile dysfunction (ED) drugs and analogues [124]. The OngLai algorithm was developed to detect homologous series using the RDKit [125]. Entropy of AI was used for the classification of stilbenoid compounds [126]. In addition, AI algorithms have demonstrated exceptional capabilities to predict compound structures, even in cases where no prior reference exists [127]. AlphaFold2 (AF2) was developed by DeepMind to predict 3D structures of chemical compounds with an atomic-level accuracy. These multifaceted AI approaches for metabolite identification not only accelerate the annotation process but also are promising to discover novel compounds that have never been studied before [128].

AI algorithms were also utilized to analyze complex spectral data. For example, Zhang et al. developed the DeepSpectra model to learn patterns from raw spectroscopic data and improve the quantitative performance [129]. The results showed that DeepSpectra generated improved quantitation results than conventional linear and nonlinear calibration approaches, through minimizing artifacts, retaining useful patterns, and providing better repeatability and accuracy.

3.3. AI in Data Extraction and Preprocessing

The primary goal of data extraction in metabolomics is to convert complex, high-dimensional raw data into a format that can be used for downstream analysis and interpretation [111, 130]. CNNs have been proven particularly effective at these tasks, such as processing, selecting, and integrating chromatographic peaks with strong local correlations, outperforming other methods by an average increase of 6% to 24% in classification accuracy [131, 132]. In addition, CNN exhibits less dependency on the data preprocessing methods compared to other chemometric techniques [133], as well as the ability to deconvolute overlapping peaks in complex chromatograms [134]. Serrano group introduced a novel ANN model derived from PUNNs [135]. This model significantly improved peak selection, which makes it effective in quantifying analytes with overlapped chromatographic peaks from a single detector instrument [135].

The use of RT, along with the m/z ratio, can enhance metabolite identification in untargeted metabolic profiling [136, 137]. The accuracy of peak picking is high when the RT of a metabolite is known; however, RT prediction is challenging when the metabolite is unknown. To address this challenge, many ML methods have been developed. Fiehn’s group integrated five different ML algorithms into the Retip R package, and they observed a significant decrease (68%) in the number of candidate structures when conducting a search for all isomers in mouse plasma samples [138]. In the initial phase of RT prediction, structural information is encoded into a vector format. This can be achieved using techniques such as quantitative structure–retention relationship (QSRR) models, or molecular fingerprints [139–141]. Liapikos et al. developed a model using four ML regression algorithms (Bayesian Ridge Regression, Extreme Gradient Boosting Regression, and Support Vector Regression (SVR) using both linear and non-linear kernels), in which the molecular descriptors were employed to represent the physical, chemical, and structural properties of the molecules in suitable mathematical structures [142]. In 2019, Bouwmeester et al. carried out a comprehensive comparison of various ML methods for predicting RTs in LC [143]. 151 features were extracted from the simplified molecular input line entry system (SMILES) notation to train seven models, incorporating both linear and nonlinear approaches. Stand-alone models, such as ANN and SVM, demonstrated high accuracy in the RT prediction task. In addition, the results highlighted the superior performance of ensemble methods that combined multiple ML tools.

Metabolomics data may have systematic errors, including batch differences, longitudinal drifts, instrument-to-instrument variation, etc. To minimize these technical variances, Kehoe et al. developed a sparse SVM model to address batch effects that are a common challenge in metabolomics studies with many samples [144]. Shen et al. developed the SVR method to reduce unwanted intra- and inter-batch variations and integrate multiple data batches in large-scale metabolomics studies, which significantly improved the performance of downstream multivariate statistical analyses in terms of classification and prediction accuracy [145]. The ability of ML methods to handle non-linear data makes them suitable for capturing complex patterns and removing systematic errors in the data. Fan et al. performed systematic error removal using the random forest (SERRF) approach. They showed that SERRF outperformed 15 other commonly used normalization methods, significantly reducing the unwanted systematic variation and revealing the biological variance of interest [146].

Class imbalance, in which each group has a very different sample size, may introduce severe biases into model fitting and performance assessment. To address this issue, oversampling techniques such as Synthetic Minority Over-sampling Technique (SMOTE) [147], as well as undersampling techniques such as random subsampling [148] and bootstrapping [149] are commonly used. These techniques are employed to balance classes in the training data but not in the validation and testing data, such that the resulted performance reflects the original sample size ratios. Gui compared over-sampling (OS), under-sampling (US), and SMOTE, and as a result, SMOTE technique was shown to have the minimum value of Cost [150].

3.4. AI in Statistical Modeling

Given preprocessed data, statistical modeling aims to uncover complicated metabolic interactions, identify novel biomarkers, and improve our understanding of various biological processes [151–153]. To select suitable AI algorithms, adjust the configuration of a model effectively, and design new algorithms, a profound understanding of the characteristics inherent in metabolomics data is imperative. Key aspects include:

• Curse of dimensionality [154]: While metabolomic experiments can profile a vast array of metabolites often numbering in the hundreds to thousands, the high costs often limit the number of samples examined. This results in datasets with the number of variables/metabolites much larger than the number of samples, which is prone to cause data overfitting.

• Complex relationship between metabolites [155]: As different metabolites are jointly involved in a multitude of biochemical reactions and pathways, their concentrations may show various levels of correlation.

• Large variations [156]: Metabolite levels can change rapidly in response to intrinsic and extrinsic stimuli. The biological variation among individuals or experimental replicates may overshadow subtle metabolic changes of interest. In addition, covariates, such as age, gender, and demographic features, also affect metabolite levels. These covariates shall also be considered together with the metabolites in feature selection and predictive modeling.

• Ambiguity in metabolic causality: Metabolic changes could either cause or result from phenotypic changes. Given the diversity of metabolites involved, it is challenging to ascertain which ones should be modeled as dependent variables and which as independent variables.

A variety of strategies have been developed to address these challenges at each stage of the analysis process, mainly including feature extraction and selection, as well as modeling and evaluation.

3.4.1. Feature Extraction and Selection

Feature extraction and feature selection both aim to identify the most relevant and insightful features while reducing the data dimensionality [157]. Feature extraction accomplishes this goal by transforming the original features into a smaller set of artificial features, typically using the unsupervised learning algorithms described in Section 2.1. However, the interpretation of the transformed artificial features can be challenging. In contrast, feature selection employs supervised ML algorithms to pinpoint metabolites associated with the phenotype/condition of interest [158].

Figure 3 and Table 2 summarize four major categories of feature selection strategies. Filter methods, encompassing various univariate and multivariate statistical tests, are the most commonly used strategies due to their simplicity and efficiency [20, 159, 160]. However, the selected features may not be the best predictors to distinguish samples of different phenotypes. Wrapper methods address this issue by adding an evaluation step that uses the selected metabolites to build a predictive model and choosing the best-performing set. For example, Gaul et al. combined a recursive feature elimination (RFE) step with an SVM evaluator, which generated a 16-metabolite predictive model to diagnose early-stage serous epithelial ovarian cancer [161]. Embedded methods, such as decision tree, RF, and least absolute shrinkage and selection operator (LASSO) [162], simultaneously perform feature selection and model training. As an example, LASSO was applied to construct significant lipid profiles for different molecular subtypes of breast cancer [163]. When these methods individually cannot produce satisfactory results, hybrid techniques to combine them could be used. To predict renal cell carcinoma status, Bifarin et al. first used filter techniques to narrow down the feature set from 7,147 NMR and MS features to 128 based on t-test p-values, fold changes, and correlation tests [164]. They then utilized wrapper approaches, specifically RF-RFE and PLS, to further refine this selection to the top 20 features. Ultimately, a vote-based approach was used to identify a highly effective panel of seven metabolites [164].

Figure 3.

Illustration of the four categories of feature selection strategies.

Table 2.

Comparison of the four categories of feature selection strategies

	Filter	Wrapper	Embedded	Hybrid
Advantages	Computationally efficient; independent of model.	Tailored to model performance; can capture feature interactions.	Efficient as part of model training; directly linked to model’s predictive power; reduces overfitting risk.	Balances efficiency and accuracy; flexible.
Limitations	May miss important interactions; not tailored to specific model performance.	Computationally intensive; risk of overfitting; dependent on specific model.	Tied to specific model; can be complex; interpretability may vary.	computationally intensive; requires careful balance and tuning.
Best Use in Metabolomics	Feature reduction before modeling in very large data sets with high dimensionality.	When computational resources allow and model performance is critical, it may produce smaller, well-defined feature sets.	Suited for datasets where model integration is beneficial and /or when specific models are preferred.	When a single method is insufficient and/or when a balanced approach is needed.
Examples	t-test, correlation test, linear regression, ANOVA, etc.	Forward selection, backward elimination, RFE, etc.	LASSO, decision trees, etc.	Combines filter, wrapper, and embedded methods.

3.4.2. Predictive Modeling

Table 3 compares commonly used ML algorithms with specific examples of their applications in metabolomic modeling. AI offers significant advantages over traditional statistical methods for modeling complex, non-linear relationships between features and responses. For targeted metabolomics studies where the datasets are relatively small, or when interpretability is crucial, methods that provide clear insights into data relationships, such as decision trees or traditional statistical regression, are preferable. However, for untargeted metabolomics studies where the datasets are large, or when predictive performance is paramount and interpretability is less critical, complex “black box” models like DL may be more suitable. Iterative testing is crucial for selecting the optimal model. Initially, a variety of models are tested, from simple ones like linear regression to more complex types such as RF, SVM, and DL, to establish baseline performance. Based on these initial results, model selection will be refined by tuning hyperparameters and re-evaluating performance. This structured approach ensures that the chosen method optimally addresses the study objectives, such as classifying samples based on phenotypes or conditions, or predicting quantitative outcomes.

Table 3.

Comparison of ML algorithms for predictive modeling

Algorithm	Advantages	Limitations	Examples
PLS-DA	Supporting modeling multiple metabolites as features or responses.	Assumptions about data distribution.	Fabritiis et al. (2023) [166]
RF	Providing feature importance scores, robust to overfitting.	Can be biased with categorical variables, large trees can be complex.	Chen et al. (2021) [167]
SVM	Flexible with kernel choice.	Choice of kernel critical, computationally intensive for large datasets.	Kang et al. (2022) [168]
KNN	Intuitive, simple, effective in data with clear local structures.	Sensitive to k and distance metric, struggles with high-dimensional data.	Wang et al. (2022) [169]
CNN	Excellent for capturing spatial hierarchies, robust to translation and rotation.	Requiring large training datasets, computationally intensive.	Kim et al. (2022) [100]
RNN	Effective for modeling time-dependent data, captures temporal dependencies.	Requiring large training data, prone to vanishing gradient problem, complex to train.	Li et al. (2019) [102]
GAN	Capable of generating realistic new samples, enhancing data diversity.	May produce unrealistic samples, challenging to train.	Choudhury et al. (2022) [103]

The need for large training datasets has historically limited the applications of DL in metabolomics [165]. Recently, however, Choudhury et al. employed GAN to overcome this limitation by generating synthetic data that closely mimic real metabolite profiles [103]. By augmenting a small sample set with a large synthetic dataset, the Reconstruction of Kinetic Models using Deep Learning (REKINDLE) model was trained. This approach allows REKINDLE to explore a vast parameter space, ensuring that sufficient diversity is incorporated in the learned model.

4. AI-assisted Metabolomics in Systems Biology

Metabolomics is one of the omics fields in systems biology. Systems biology is an interdisciplinary area to understand the big picture of a biological system, often at the levels of organism, tissue, and cell [170–174]. As shown in Figure 4, a fundamental process in biology involves the expression of genetic information (genotype, encoded in DNAs) as observable characteristics (phenotype) [175, 176]. This transition is governed by various molecular and cellular processes. To initiate it, genes are expressed in a multi-step process that includes transcription and translation, synthesis of RNAs from DNAs and proteins from RNA sequences, respectively [177]. Although not all genes are always expressed, the resulting proteome is crucial in determining phenotypes. For example, the proteasome is a highly sophisticated protease complex that is responsible for the degradation of proteins, and it plays an important role in regulating various cellular metabolism [178, 179]. The proteasome is extensively involved in protein turnover, regulation of specific metabolic pathways, and control of key metabolic processes [180]. At the phenotype level, metabolome provides significant insights into how genetic information and environmental factors (co-)influence the composition and abundance of metabolites (small molecules) in cells, tissues, and organisms [181–185]. With combined input from genome, transcriptome, proteome, and metabolome, systems biology studies the interactions among these components and aims to gain a comprehensive understanding of a biological system.

Figure 4.

Systems biology: from genotype to phenotype in a biological system.

4.1. Metabolic Network and Pathway Analysis

AI can assist in the construction and analysis of metabolic networks and pathways [186–188]. AI algorithms were used to provide a comprehensive understanding of how metabolic pathways were regulated and perturbed in different conditions [189]; therefore, it gained a deeper understanding of the complex interactions and networks of metabolites in biological systems [151, 152]. ML algorithms, especially DL techniques, are applied to metabolic pathway and network analysis to address the complexity and high-dimensional nature of the data [111, 190]. Jin et al. showed that graph DL was a novel approach to process network structures and analyze large, multidimensional, and complex biological data [188].

Elovici group used a ML model to map metabolic pathways onto the metabolite correlation networks and compute network properties for each pathway [191]. Using this RF model, they predicted previously unidentified pathways and successfully validated the predictions in vivo [191]. Their approach involved associating metabolites with pathways and using measured metabolite correlations to calculate feature vectors for each pathway. These feature vectors were generated using a combination of statistical, graph-related, and correlation network metrics. The RF model was then trained to classify the activity of 169 organism-related active pathways from the MetaCyc database [192], along with 85 non-active pathways and 85 randomly selected metabolite combinations. Notably, this approach is limited to identified metabolites and predefined pathways during the training. In another study by Hosseini and colleagues, pathway activity was weighed based on the likelihood of metabolites being connected to the pathway [193]. A generative model was developed to establish links between pathway activity probabilities, metabolites, and measured spectral masses. This approach emphasized metabolites that were unique to a specific pathway, resulting in novel predictions compared to standard enrichment analysis.

Metabolic flux analysis (MFA) is also broadly included in the big field of metabolomics, and MFA provides a dynamic description of metabolic networks in a biological system [194, 195]. MFA tracks the footprint of isotopically labeled substrates, such as ¹³C₆-glucose, in metabolic pathways to examine the impacts of internal or external stimuli [196]. Using the ANN approaches, Antoniewicz et al. predicted the fluxes of ¹³C-labeled metabolites in mammalian gluconeogenesis [197]. Compared to traditional methods that require many samples (>200), ANN algorithms, including multiple linear regression (MLR), PCR, and PLS, performed much better for flux prediction using the mass isotopomer data.

4.2. Integrative Analysis of Multi-omics Data

Integrating metabolomics data with other omics data presents unique challenges due to the inherent differences in metabolites compared to genes, proteins, and other molecular components. Table 4 presents the data characteristics in different omics that need to be accounted for in integrative analysis. First, harmonizing data scales is challenging because the data units vary by omics types and instruments [198]. For example, metabolites are measured in relative abundance (e.g., counts per second, cps, in MS) or absolute quantitation (mM, μM, or ng/mL), while genomic sequences are reported as A/T/C/G characters, and gene expression abundance is reported as continuous values or discrete counts. In addition, the diversity of chemical structures and the presence of isomers and modifications introduce uncertainties in integrated analysis [199], which are higher in metabolomics than other omics as most genetic variants and transcripts can be accurately mapped to unique genomic regions [200]. The sparsity of metabolomics data, where many metabolites have dramatically different concentrations, complicates the process of finding meaningful associations with other omics data [201]. Furthermore, metabolites often participate in multiple pathways, making it difficult to link changes in metabolite levels to specific genetic or protein-level changes due to the pathway crosstalk. High biological variation of metabolites, influenced by genetics, diet, and environmental factors, adds another layer of complexity [202].

Table 4.

Characteristics of various omics data in systems biology

Feature	Metabolomic Data	Genomic Data	Transcriptomic Data	Proteomic Data	Microbiome Data
Nature	Full set of small molecule metabolites in a biological sample	Entire DNA sequence of an organism, including all genes	Full set of RNA transcripts from the genome at a specific time	Complete range of proteins produced in a system	Genetic material of all microbes in a specific environment
Stability	Highly dynamic; changes rapidly in response to internal and external factors	Relatively stable; changes only due to mutations	Dynamic; varies with time, conditions, and tissue type	Dynamic; influenced by post-translational modifications and conditions	Dynamic; varies with environmental factors and host interactions
Dynamics	Shows real-time metabolic activity and state	Static representation of potential cellular capabilities	Indicates active gene expression at a given moment	Reflects the proteins currently present and functional	Reflects the composition and functions of microbial communities
Information	Biochemical activity and metabolic state	Genetic potential and hereditary information	Gene expression levels and activity	Functional state at the protein level	Microbial diversity, community structure, and potential functions
Applications	Disease diagnosis, drug metabolism, nutrition	Hereditary diseases, genetic predispositions	Disease mechanisms, biomarker discovery	Disease mechanisms, drug targets, biomarkers	Human health, disease states, environmental studies
Challenges	Diversity of metabolites, sensitivity of detection	Data size, interpretation of variants	Temporal and spatial complexity, data volume	Protein modifications, complexity of interactions	Diversity of species, functional interpretation

Therefore, it is critically important to choose appropriate data integration methods that accommodate the unique characteristics of metabolomics data. Currently, new AI methods and tools are facilitating the integration of metabolomics with other omics to extract meaningful insights [203]. For example, ML and network-based approaches can help identify connections between genes, enzymes, and metabolites, with the potential of discovering previously unknown metabolic pathways [152, 186]. In addition, ML methods have been used to analyze multi-omics data, resulting in groundbreaking discoveries in cancer molecular biology [204]. These AI approaches can lead to a deeper understanding of biological mechanisms, and in the case of diseases, the discovery of novel therapeutic targets and more effective treatments.

The AI integration of multi-omics mainly uses approaches of sequential integration and simultaneous integration. Sequential integration adopts a step-wise approach, where omic datasets are initially analyzed individually or in specific combinations, followed by the integration of their findings in subsequent stages. The sequential integration method allows for many types of data analyses, even in cases where omic measurements for the same samples are unavailable. It is particularly suitable for bulk datasets, as demonstrated by a previous study that successfully integrated ATAC-seq and RNA-seq data to identify critical genes and regulatory pathways associated with the neuroprotection of S-adenosylmethionine (SAM) against perioperative neurocognitive disorder (PND) [205].

Simultaneous integration involves the parallel analysis of multiple omic layers, allowing the analysis of all available omic data at once within a single modeling step. While this approach requires that data originate from the same biological samples or individuals, it is a powerful method for deriving valuable insights into cellular functions by capturing the interplay between different omic datasets. These integration methods can be broadly categorized into concatenation, model-based, and transformation-based approaches [206].

• Concatenation-based integration methods involve the direct merging of data matrices from individual omics domains, creating a comprehensive multi-omic dataset. This integrated data is then used as the input for various AI techniques. Concatenation-based integration can also incorporate unsupervised methods, as demonstrated in a study on muscle-invasive bladder cancer (MIBC) [207].

• Model-based integration methods involve a multi-stage process where individual models are first developed for different omics data types, and then these models are combined into a joint model, allowing for the integration of heterogeneous data sources. Model-based integration methods can be further categorized as either supervised or unsupervised. As an illustrative example, Multi-omics Supervised Autoencoder (MOSAE) was applied to the TCGA Pan-Cancer dataset and to effectively predict four distinct clinical outcomes [208].

• Finally, transformation-based integration methods begin by converting individual omics datasets into graphs or kernel matrices, which are then combined to create a unified representation and capture the relationships between diverse omics data. The versatility of these methods is a significant advantage, allowing them to integrate various omic types effectively. For instance, the fast multiple kernel learning for dimensionality reduction (fMKL-DR) method employs kernel matrix transformation and an SVM classifier to effectively stratify samples [209]. Weighted-nearest neighbor (WNN) method is an example of transformation-based unsupervised algorithm that effectively integrates single-cell multimodal datasets [210].

4.3. Biological and Functional Integration

AI algorithms were used to incorporate biological and functional information. Because of the intercorrelation between metabolites and biological variations, it is often challenging to identify reproducible and generalizable biomarkers. A RF model was developed to derive a feature importance score based on the existing biological knowledge [211]. This updated RF model was used in recognizing features, and it provided more discriminating power. In addition, metabolites were ranked using prior knowledge, and this rank was successfully used to guide feature selection and predictive modeling [212, 213].

AI methods were also used to improve biological interpretation. For example, local interpretable model-agnostic explanations (LIME) [214] and SHapley Additive exPlanations (SHAP) analysis [215] were used to interpret significant variables within the disease context. In addition, strategies, such as simplifying complex models and prioritizing functionally important features, were employed to enhance interpretability. Further, unsupervised learning methods combined with visualization techniques, such as PCA followed by t-SNE plot, as well as clustering results displayed in dendrograms are useful to uncover hidden structures in the data.

5. AI-assisted Metabolomics in Health

5.1. Disease Biomarker Discovery

AI has found many applications in disease diagnosis and classification in the field of metabolomics. Many diseases induce altered metabolism, resulting in measurable changes in the levels of metabolites. As an example, it has been well known that cancer cells preferentially use glycolysis as their source of energy, a phenomenon known as the Warburg effect [216]. In subjects with congestive heart failure, cardiomyocytes resume the fetal metabolic program that favors glucose versus fatty acids [217]. AI algorithms have the capacity to identify distinct metabolite patterns from different groups, and these patterns can subsequently serve as valuable tools for disease diagnosis and progression monitoring [218, 219]. Table 5 summarizes the recent applications of AI-assisted metabolomics in human health. Turi et al. and Jendoubi et al. reviewed ML approaches for analyzing metabolic profiles, focusing on biomarker discovery and disease diagnosis [126, 220]. In particular, AI techniques were used for identifying metabolite biomarkers for various diseases, such as cancer, cardiovascular diseases (CVD), and Alzheimer’s disease (AD) [221–223].

Table 5.

Summary of AI-assisted Metabolomics in Health

AI Algorithm	Strength	Biomarker	Analytical Platform	Sample	Application	Data Source	Reference
SVM-RFE	Effective in high-dimensional spaces	16 diagnostic metabolites	UPLC-MS	Serum	Epithelial ovarian cancer	N/A	Gaul et al. (2015) [161]
LogReg	N/A	63 metabolites	RP-HPLC-QQQ/MS	N/A	Bladder cancer	N/A	Kouznetsova et al. (2019) [253]
FCBF	N/A	5 biomarkers	Targeted LC/MS	Lung tumor tissue	Lung cancer	N/A	Xie et al. (2021) [20]
ML-based LDA	N/A	10 ratios of metabolites, 2 ratios of metabolites	LC-MS/MS	Tumor tissue	Cancer	N/A	Wallace et al. (2020) [254]
DL, RF, SVM, RPART, LDA, PAM, GBM	N/A	20 metabolites, 8 pathways	GC-TOF/MS	Tumor tissue	Breast cancer	http://www.ncbi.nlm.nih.gov/geo.	Alakwaa et al. (2018) [255]
OPTICS	Hierarchical, easily interpretable data structure	Imaging	N/A	Tumor tissue	Cancer	http://www.ebi.ac.uk/metabolights/MTBLS415	Inglese et al. (2017) [224]
Multiple LogReg, ADTree	Higher accuracy	31 metabolites	CE-TOF-MS, LC-QQQ-MS	Saliva	Breast cancer	N/A	Murata et al. (2019) [256]
RF	N/A	5 biomarkers	LTQ-XL Orbitrap Discovery	Plasma	Weight gain	http://dx.doi.org/10.21227/k446-fp12	Dias-Audibert et al. (2020) [76]
RF	N/A	17 discriminatory metabolites	LC-MS/MS	Stool	Cirrhosis	metabolomic data were deposited at GNPS	Oh et al. (2020) [257]
LogReg	N/A	68 metabolites	RPLC-MS	Plasma	Human pregnancy metabolome	https://doi.org/10.21228/M81H58	Liang et al. (2020) [258]
RF, LDA, LogReg, KNN, naive Bayes, SVM	N/A	4 metabolites	RP/UPLC-MS/MS, HILIC/HPLC-MS/MS, LC-Q-Exactive high resolution MS	Serum	Ischemic stroke	N/A	Tiedt et al. (2020) [77]
RFE, PLS, RF, KNN	Optimization of hyperparameters	10 metabolites	LC-MS	Urine	Renal cell carcinoma	http://dx.doi.org/10.21228/M8P97V	Bifarin et al. (2021) [164]
MUVR	Simultaneously identify minimal-optimal and all-relevant variables	13 metabolites	LC-MS	Serum	Gout and asymptomatic hyperuricemia	N/A	Shen etal. (2021) [259]
LogReg, RF, SVM	Effective in handling high-dimensional data	26 metabolites and lipids	HILIC-LC-MS, RP-LC-MS	Serum	Rheumatoid arthritis	N/A	Luan etal. (2021) [260]
RF	Highly conserved, predictive chemical feature subsets for taxonomic identity	34 metabolites	LC-MS qTOF MS	Gut bacteria	Human gut microbiota	N/A	Han etal. (2021) [261]
LogReg	N/A	5 purine metabolites	UPLC-Q/TOF MS, UHPLC-TQ MS/MS	Serum	Coronary artery disease	N/A	Jung etal. (2021) [262]
LDA, SVM, RF, LogReg	N/A	58 metabolites	GC-MS, UHPLC-TQ MS/MS	Serum	Diabetic kidney disease	N/A	Liu etal. (2021) [263]
SVM, RF	Easily build non-linear classifiers; effectively reduce the high-variance	N/A	GC-MS, LC-MS	Plasma	Major depressive disorder	N/A	Liu etal. (2016) [264]
ANN, O2PLS, PCA, PLS	Nonlinear MSA	11 compounds	HPLC-HRMS	Plant extracts	Anti-inflammation	N/A	Chagas-Paula et al. (2015) [265]
RF	Built-in feature selection, Variable importance scores are useful for biomarker prioritization	22 serum proteins, 7 metabolites	UPLC-MS/MS	Serum	COVID-19	https://www.iprox.org/(IPX0002106000andIPX0002171000)	Shen et al. (2020) [266]
Gradient Boosting, RF	N/A	Top 20 ion features	LC/Q-TOF, LC-MS/MS	Nasopharyngeal	Influenza	N/A	Hogan et al. (2021) [267]
AdaBoost, GDB, RF, XRF, PLS, SVM	N/A	19 discriminant biomarkers	HESI-Q Exactive Orbitrat-MS	Plasma	COVID-19	https://10.5281/zenodo.4329381	Delafiori et al. (2021) [268]
RF	N/A	N/A	GC-MS	Pericarp tissue	Tomato	N/A	Toubiana et al. [191]
MCR-ALS, PCA, k-Means	N/A	N/A	GC-MS	Saffron	Saffron extract	N/A	Aliakbarzadeh et al. (2016) [269]
EDNN	superior to the conventional methods in regression, based on RMSEs	N/A	NMR	Fish muscle	Fish	N/A	Asakura et al. (2018) [270]

AI was used to improve cancer diagnosis and treatment. Inglese et al. developed a novel 3D MS imaging technology, which is a powerful tool to investigate the chemical and biological interactions occurring in tumor tissues [224]. Inspection of the 3D data derived from DL and cluster analysis is a straightforward and robust approach to identify the presence of tumor cells, which is extremely difficult by visual inspection [224]. In addition, ML algorithms have been applied to analyze images to detect various types of cancer at an early stage [225]. Bebas et al. used classification methods, including SVM, KNN, RF, and DL, and the best results were achieved using the SVM classifier with the highest specificity and sensitivity [226]. Takats, Nicholson, and Glen groups made use of a deep unsupervised NN, specifically parametric t-SNE, and they projected a 3D-desorption electrospray ionization (DESI)-MS dataset from a human colorectal adenocarcinoma biopsy onto a 2-dimensional manifold, which enabled the identification of clusters that may not be discernible with linear methods [224]. Further, AI not only improves the accuracy of diagnosis but also allows for quicker and more effective treatments. Elemento group reviewed the recent progress in the applications of AI in oncology, which highlighted the potential of AI to be used in clinical settings [227].

AI is also very useful in clinical decision-making and management for other diseases, such as CVD and strokes [228, 229]. Diverse AI methodologies, including robotics, ML, and NLP, have been employed for the examination of CVD [228]. Chang group compared the strengths and limitations of both traditional and AI approaches in CVD biomarker discovery, and AI algorithms were found to have high potential in risk stratification and early disease detection [230]. In addition, a number of advanced ML models have been developed to detect biomarkers of coronary artery disease, coronary artery stenosis, myocardial ischemia, and stroke lesion outcomes [231–235]. Juarez-Orozco et al. used ML to identify predictors for recognizing patients who were likely to have myocardial ischemia and an elevated risk of Major Adverse Cardiovascular Events (MACE) [236].

AI, particularly ML, has been used for extracting dependable predictors and automatically classifying diverse dementia subtypes, such as mild cognitive impairment (MCI) that is likely to progress to AD [223]. Li group reviewed AI methods that have been successfully used for the diagnosis and prognosis of AD at the MCI or preclinical stages [237]. Battista et al. summarized ML methods that can automatically classify patients with AD, even in the early stages of the condition, and these AI approaches can identify an optimal combination of neuropsychological predictors [238].

5.2. Personalized Medicine & Drug Discovery

AI can assist in the development of personalized medicine by analyzing an individual’s metabolome [239]. Personalized medicine aims to optimize medical treatments for individual patients based on their unique characteristics, including their age, sex, genetics, lifestyle, and metabolism. AI-driven robotic metabolomics is able to analyze large and complex datasets, offering insights into complex gene-protein-metabolite interactions for targeted drug discovery and disease treatments [240]. Particularly, AI-driven robotic metabolomics automates labor-intensive tasks such as sample preparation and data acquisition, which increases data robustness and accuracy [240–242]. This approach is not only cost-effective, reducing the likelihood of ineffective treatments, but also empowers patients by involving them more directly in their healthcare decisions, providing specific health status information and potential risks [243]. Azad et al. suggested that ML/AI could be integrated with metabolomics for automated clinical diagnosis, treatment, and prognosis [244]. As a result, AI-assisted metabolomics can help generate personalized recommendations for medication, diet, and lifestyle modifications, which has the potential to significantly improve patient outcomes and reduce healthcare costs [245].

In addition, metabolomics can examine metabolic effects of potential drugs, based on which AI can predict how drugs interact with metabolic pathways and prioritize drug candidates for further investigation [246]. David et al. provided a guideline for the practice of AI in drug discovery [247]. AI-driven analysis has been used to determine the most effective drug dosages, safety, and efficacy. Martinelli reviewed the ML models, such as elastic net regression and RF, which have been used to study metabolism and provide invaluable insights to determine attrition rates and optimize drug efficacy [248]. Similarly, ML algorithms have been applied to expedite the process of identifying potential drug candidates and predicting their effectiveness. Tanoli et al. described the use of supervised ML models for three levels of predictions related to drug repurposing [249]. ML has also been used to predict the specific effects of drugs on cancer [250]. In a study conducted by McDermott and Garnett groups, the responses of 1,001 human cancer cell lines were tested on 265 anticancer drugs, establishing a comprehensive pharmaceutical landscape in cancer [251]. Varshney group showed that AI can address many disadvantages during the classical drug development, such as too many candidates and being slow in determining drug side effects, through reducing human intervention and bridging the translational gap between drugs and diseases [252].

6. Challenges and Future Perspectives

Although it is of increasing interest to use AI in metabolomics, there are challenges to analyze metabolomics data using AI approaches. As mentioned above, one of the main challenges is curse of dimensionality. Training a reliable AI model demands a substantial number of samples. In statistics, a common guideline suggests that around 20 samples are needed for each parameter to be estimated. AI models often comprise a considerable number of parameters. For instance, in a RF model with 100 decision trees each containing an average of 10 nodes, there are 1,000 parameters to be estimated (as the splitting cutoff at each node is a parameter). Training such a model would necessitate 20,000 samples. However, in many applications, the number of samples falls short of this criterion, or may even be smaller than the number of parameters, leading to a situation known as the curse of dimensionality. Under this curse, the model is inevitably prone to overfitting the data, capturing noise and intricacies specific to the training data but failing to generalize well to new, unseen data. To mitigate the curse of dimensionality, it is essential that the number of model parameters is kept lower than the number of samples. In cases where increasing the sample size is not feasible, another approach is to reduce the complexity of the model. This can be achieved by, for example, decreasing the number of trees in a RF model or reducing the number of neurons in an ANN model. In addition, it is also recommended to reduce the number of input features. To further combat overfitting, robust validation methods and diverse, well-annotated datasets are essential, ensuring that AI models are trained to generalize effectively. Stringent internal cross validation should be embedded in AI approaches, and when possible, external multi-center cross validation using a large sample set should be performed to validate the significant metabolites established by AI.

The complexity of AI models can lead to interpretation challenges, making it difficult to understand the rationale behind specific predictions or recommendations. Developing more transparent AI models can help understand how decisions are made, fostering trust and practical utility in clinical settings [271]. Another significant challenge is the requirement for extensive computational resources and expertise in AI, which may not be readily available in all healthcare settings. Improving access to computational resources and fostering interdisciplinary collaborations among AI experts, bioinformaticians, and clinicians can facilitate the adoption of AI in diverse healthcare environments. In addition, the dynamic nature of AI models, while beneficial in rapidly adapting to new data, can also lead to model drift over time, necessitating continuous monitoring and updating. This weakness highlights the need for a balanced approach in integrating AI into metabolomics to ensure that the benefits of advanced data analysis are harmonized with practical applications. Regularly updating AI models and employing adaptive learning algorithms can help prevent model drift, maintaining the relevance and accuracy of the analysis over time.

Further, the transition of AI-assisted metabolomics to routine clinical practices faces hurdles, including technological and scientific challenges, practicality, regulatory compliance, and integration into existing healthcare frameworks. The resulted AI models need to have friendly interfaces for end-users such as clinicians, ideally as plug-ins in electronic health record software that can be directly used in clinical decision-making. Data security and patient privacy should also be protected, similar to other omics data. By tackling these challenges, AI-assisted metabolomics can be more effectively used to enhance patient care and treatment outcomes.

7. Conclusions

In this review, we reviewed recent applications of AI in the field of omics, with a special focus on metabolomics. Metabolomics studies the metabolome in systems biology, and it produces large datasets examining hundreds to thousands of metabolites. AI was designed to work with large datasets, and it has the potential to bring “human intelligence” to metabolomics, but with computers and software packages. Recently, various AI algorithms have been developed and used in metabolomics to improve data collection, data extraction, and statistical modeling. AI algorithms can efficiently and accurately identify metabolites, detect changes in their concentrations, and determine their biological significance. Based on previous successful applications, it is expected that AI-assisted metabolomics has the potential to revolutionize disease prevention, diagnosis, and treatment.

Highlights.

• We provide a recent overview of the methodologies and applications of artificial intelligence (AI) in metabolomics.

• We introduce metabolomics and AI algorithms that are commonly used in big data analyses.

• Recent applications of AI-assisted metabolomics in the context of systems biology and human health are summarized.

• Challenges and future perspectives for the development and applications of AI approaches in metabolomics are discussed.

Abbreviations:

AI

artificial intelligence

ML

machine learning

DL

deep learning

MS

mass spectrometry

NMR

nuclear magnetic resonance

LC

liquid chromatography

GC

gas chromatography

CNNs

convolutional neural networks

RNNs

recurrent neural networks

GANs

generative adversarial networks

NLP

natural language processing

HRMS

high-resolution mass spectrometry

QSRR

quantitative structure–retention relationship

NN

neural networks

ANN

artificial neural networks

PCA

principal component analysis

ICA

independent component analysis

UMAP

uniform manifold approximation and projection

t-SNE

t-distributed stochastic neighbor embedding

SVM

support vector machines

SVR

support vector regression

RF

random forest

SOM

self-organizing maps

KNN

k-nearest neighbors

DQN

deep q-networks

PPO

proximal policy optimization

DDPG

deep deterministic policy gradients

TRPO

trust region policy optimization

SMILES

simplified molecular input line entry system

MFA

metabolic flux analysis

MLR

multiple linear regression

PLS

partial least square

AD

Alzheimer’s Disease

MCI

mild cognitive impairment

PLS-DA

partial least squares regression–discriminant analysis

LogReg

logistic regression

ADTree

alternative decision tree

MSA

multivariate statistical analysis

XRF

extreme random forest

GDB

gradient tree boosting

ADA

ADA tree boosting

EDNN

ensemble deep neural network

VAEs

variational autoencoders

ANOVA

analysis of variance

REKINDLE

reconstruction of kinetic models using deep learning

ChatGPT

chat generative pre-trained transformer

SAM

S-adenosylmethionine

PND

perioperative neurocognitive disorder

MIBC

muscle-invasive bladder cancer

MOSAE

multi-omics supervised autoencoder

WNN

weighted-nearest neighbor

fMKL-DR

fast multiple kernel learning for dimensionality reduction

DESI

desorption electrospray ionization

CVD

cardiovascular disease

MACE

major adverse cardiovascular events

RT

retention time

LIME

local interpretable model-agnostic explanations

SHAP

SHapley Additive exPlanations

AF2

AlphaFold2

ED

erectile dysfunction

SMOTE

synthetic minority over-sampling technique

OS

over-sampling

US

under-sampling

SERRF

systematic error removal using random forest

LASSO

least absolute shrinkage and selection operator

MUVR

multivariate methods with unbiased variable selection in R

FCBF

fast correlation-based filter

AdaBoost

adaptive boosting

RFE

recursive feature elimination

MCR-ALS

multivariate curve resolution-alternating least squares