Glycolysis (2 mol APT/mol glucose) is a less efficient pathway for ATP production than mitochondrial OXPHOS (36 mol ATP/mol glucose). However, the rate of glycolysis is 10~100 times faster than the rate of mitochondrial TCA cycle and OXPHOS. Therefore, the amount of ATP produced by the two metabolic pathways of glucose is similar at the same time. In tumor cells, glucose is metabolized by glycolysis, producing lactate and nicotinamide adenine dinucleotides under aerobic conditions. The low-yielding but high-rate ATP production mode is more conducive to the competition of tumor cells for nutrients and meets their energy demand for rapid growth. Tumor cells take up a large amount of glucose from the environment with a strong competitive advantage, ensuring their energy supply and self-growth, depriving immune cells of glucose utilization, and inhibiting their tumor cell-killing effect. In addition, glucose metabolism including glycolysis, pentose phosphate pathway, hexosamine pathway, and glycogen synthesis are reprogrammed in tumor cells (62). As shown in Figure 2 , pyruvate kinase M2 (PKM2), which is highly expressed in tumors and promotes the Warburg effect, is upstream of the decision point between glycolytic and oxidative metabolism. PKM2 converts phosphoenolpyruvate (PEP) into pyruvate, which can be metabolized either to lactate or acetyl-CoA (63). Moreover, PKM2 also functions as a transcriptional coactivator, PKM2 interacts with HIF-1α in the nucleus and binds to the PD-L1 promoter region, enhancing the expression of PD-L1 in tumor cells (64). The interaction between the immune checkpoint PD-L1 and B7-H3 activates the classic aerobic glycolytic pathway PI3K-AKT mTOR in tumor cells (65). Moreover, due to glucose deprivation and downregulation of mTOR activity, the glycolytic signaling pathway PI3K-AKT mTOR is inhibited in T cells (66). In addition, the CTLA-4 pathway competitively inhibits CD28-mediated co-stimulation and reduces Akt phosphorylation and activation, thereby impairing T-cell glucose metabolism and mitochondrial remodeling (67). Moreover, the interaction between immune checkpoints and their ligands (such as PD-1/PD-L1 and CTLA-4/CD86) further participates in the metabolic reprogramming of tumor cells and immune cells (68). In summary, the synergistic effect of tumor metabolism regulation and inhibition of immune checkpoints in the TME can activate the host anti-tumor immune response.

Compared with normal cells, tumor cells utilize a large amount of glucose to produce lactate under aerobic conditions and provide the amino acids and intermediate metabolites of the pentose phosphate required for tumor cell proliferation. The accumulation of lactate in the TME has profound effects on immune cells. In Figure 3A , tumor cells consume large amounts of glucose, which is associated with poor CD4⁺ and CD8⁺ T cell infiltration (69). High glucose depletion leads to the production and secretion of lactate into the TME, where lactate acts in an immunosuppressive manner, reduces the cytolytic activity of natural killer (NK) cells, and enhances PD-1 expression and the immunosuppressive capacity of Treg. In addition, lactate increases the frequency of MDSC in tumors and spleen and induces “M2-like” polarization in tumor-associated macrophages. Additionally, the accumulation of lactate in tumors leads to a pH decrease and hinders the ability of dendritic cells (DC) to recognize and present antigens, as well as the stability of antigen MHC-I complexes ( Figure 4 ) (20). Acidification also reduces the ability of mannan receptors (MR) to bind to antigens, inhibits glycolysis, and promotes monocyte differentiation into monocyte-derived dendritic cells. Moreover, lactate in tumors inhibits toll-like receptor 3 (TLR3) and stimulator of interferon genes (STING) leading to interferon-γ reduction, accelerates antigen degradation, and impairs cross-presentation. Pilon-Thomas et al. (70) reported that neutralizing the acidity of TME with bicarbonate can increase T cell infiltration and improve response to immunotherapy in immune checkpoint inhibition and adoptive cell transfer therapy.

Amino acids are essential for maintaining the rapid proliferation of tumor cells. Besides as substrates for protein synthesis, amino acids play an important role in energy production, lipid and nucleic acid synthesis, and cellular redox homeostasis maintenance. The increased demand for amino acids results in tumor cells becoming strongly dependent on exogenous amino acids or reprogramming of amino acid metabolism. Alterations in amino acid metabolism can be used not only as clinical biomarkers of cancer progression but also as therapeutic targets.

Glutamine is an important metabolic fuel that helps meet the high demand for energy, biosynthetic precursors, and reducing agents in rapidly proliferating cancer cells (71). In Figure 2 , glutamine is transported to the cytoplasm through the solute carrier family 1, member 5 (SLC1A5) (72), and is converted to glutamate in the mitochondria through a deamination reaction catalyzed by glutaminase. Glutamate is converted to the TCA cycle intermediate α-ketoglutarate (α-KG) by glutamate dehydrogenase (73). α-KG is a critical metabolite that serves in both ATP production and in replenishing TCA cycle intermediates (74). Glutamine metabolism is essential for developing effector T cells (75). The overexpression of inhibitory receptors (such as CTLA-4 or PD-1) can inhibit the upregulation of glucose and glutamine metabolism after TCR involvement and co-stimulation (76, 77). The interaction between PD-1 and PD-L1 or PD-L2 can inhibit the PI3K-Akt mTOR pathway (78) and disrupt T cell metabolic reprogramming, including glutaminolysis (79). Huang et al. (80) reported that SLC1A5 overexpression can stimulate the growth and survival of colon cancer cells. In Figure 3B , glutamine depletion in cancer cells decreases the activation and infiltration of CD8⁺ T cells and enhances the recruitment of MDSC by increasing the secretion of granulocyte-colony stimulating factor (69).

Arginine is a versatile amino acid. In addition to being a building block in protein synthesis, it is a precursor of nitric oxide (NO), creatine, and polyamines. The arginine metabolic pathway is shown in Figure 2 , and there are two main metabolic pathways: Arginine is metabolized into citrulline and NO by nitric oxide synthase, and into ornithine and urea by arginase. Ornithine is an important resource for putrescine, which is a key precursor for polyamines. NO plays a diverse role in tumorigenesis and metastasis, and the promotion or inhibition of tumors depends on sensitivity to NO, exposure time, and NO concentration (81). Low concentrations of NO (<100 nmol/L) may promote tumor development by inhibiting apoptosis and stimulating endothelial cell proliferation. In contrast, high concentrations of NO (400-1000 nmol/L) would disrupt the cell cycle and accelerate cellular apoptosis (82). Because tumor cells rely on arginine in the TME, depriving tumors of arginine has emerged as a novel anti-tumor therapy and has shown encouraging efficacy in clinical trials against certain types of cancer. However, simply depriving arginine not only starves tumor cells but also impairs the anti-tumor immune response. Geiger et al. (83) reported that the reduction of intracellular arginine negatively affected the survival of T cells. When arginine is completely depleted in the medium, the cell cycle of T cells is arrested in the G₀-G₁ phase. Arginine supplementation, while enhancing the anti-tumor immune response, may support tumor growth. Both deprivation and supplementation alter arginine levels in the TME, which influences all cells. To achieve tumor-specific arginine restriction without influencing immune cells, an in-depth study of differences in arginine metabolism between tumor cells and immune cells is needed.

Tryptophan, an essential amino acid, plays a vital role in protein biosynthesis and serves as a precursor for the synthesis of a variety of important bioactive compounds. As shown in Figure 2 , tryptophan is first converted to N-formyl kynurenine, which is deformylated by arylformamidase to kynurenine (Kyn) (84). Kyn then activates the transcription factor aryl hydrocarbon receptor (AhR). Activated AhR could upregulate levels of PD-1 in T cells, inhibiting the anti-tumor immune response (85). Tryptophan metabolism and its key enzymes affect a variety of cell biology functions, including immune response, cell proliferation, and migration, through interactions with downstream molecules or pathways. Le Naour et al. (86) found that tryptophan is associated with restricting immune response in the TME and is often upregulated in human tumors. Targeted therapy based on tryptophan metabolism provides a new and potentially advantageous therapeutic strategy for cancer.

Lipids are essential nutrients for cells, the main structural components of cell membranes, signal molecules, and energy providers. The most common lipids include fatty acids, triglycerides, sphingolipids, phospholipids, and cholesterol (87). In addition, abnormal fatty acid metabolism is related to the occurrence and development of tumors, such as liver cancer, gastric cancer, ovarian cancer, and cervical cancer (88–90).

The sterol regulatory element binding protein pathway in lipid metabolism controls lipid synthesis in cells (91). In Figure 2 , De novo synthesis of lipids can promote high expression of PD-L1 and tumor immune suppression through palmitoylation of the immune checkpoint PD-L1 (92). In addition, PD-1 stimulates the AMP-activated protein kinase activity, inhibits glycolysis, and promotes fatty acid oxidation by upregulating activated CD4⁺ T cells, thereby inhibiting the development of effector T cells (93). The lipid transporter protein CD36 mediates intracellular fatty acid uptake and lipid droplet growth (94). CD36-mediated fatty acid uptake in tumor-infiltrating CD8⁺ T cells activates lipid peroxidation and iron apoptosis, while CD36 deficiency inhibits tumor growth (95). In addition, CD36 on the surface of CD8⁺ TILs takes up oxidized low-density lipoprotein and induces lipid peroxidation, thereby promoting CD8⁺ T cell dysfunction in tumors (96). Cholesterol metabolism regulates the antitumor activity of CD8⁺ T cells. Inhibition of acetyl-CoA acetyltransferase-1 increases plasma membrane cholesterol levels and enhances the effector and proliferation function of CD8⁺ cells (97).

After acquiring mass spectrometry data, several processing steps are typically performed before the data can be annotated. These steps include format conversion, peak detection, peak alignment, calculating peak areas, and filtering to eliminate redundancies (111, 146).

Several deep learning algorithms are particularly suitable for preprocessing mass spectrometry data. For example, convolutional neural networks (CNNs) excel at capturing local features by detecting patterns such as peaks and characteristic signals through convolution operations (43, 112, 147). Furthermore, CNNs achieve scale invariance through pooling operations, enabling the model to recognize the same features at different scales, thus enhancing robustness and generalization (148). Melnikov et al. (111) developed the “peakonly” algorithm, utilizing CNNs to enhance peak detection and integration in LC-MS data for metabolomics, achieving high precision and flexibility in managing noisy peaks. Similarly, research by Risum A.B (117)., Kantz (119), and Lim (120) demonstrates the application of CNNs for tasks such as peak identification and alignment. Improved RNN structures, like long short-term memory (LSTM) networks and gated recurrent units (GRUs), can capture long-range dependencies, aiding in analyzing temporal correlations in mass spectrometry data (149). For instance, Li and colleagues developed ChromAlignNet based on LSTM, a deep learning model that significantly improves peak alignment in GC-MS data, outperforming existing methods and requiring no user input for reference chromatograms and parameters (118). Metabolomics spectral and mass spectrometry data, essentially continuous two-dimensional data, can be encoded for processing by large language models. However, there currently lacks a dedicated natural language processing model for this preprocessing task, which warrants attention.

Metabolite annotation involves translating spectral patterns into matching chemical molecules. The most common approach involves matching mass spectra (GC-MS’s electron ionization (EI) and LC-MS’s tandem MS) with reference library spectra of known molecules (150). This step is currently the most widely applied machine learning application in metabolomics, as AI can enhance both matching and reference library generation.

The main steps in spectral matching involve comparing the query spectrum with the library spectrum and calculating similarity scores. Machine learning models, such as the DNN-based SIRIUS 4 (132), the MLP-based MSNovelist (142), and the CNN-based DeepCCs (124), are widely used for spectral prediction, fragmentation prediction, and structure identification. Notably, large language models (LLMs), a type of deep learning AI model, excel in various tasks, particularly natural language processing (NLP) (151). LLMs, such as OpenAI’s GPT-X and Google’s BERT, are composed of neural networks with numerous parameters and trained on vast amounts of unlabeled data via self-supervised or semi-supervised learning (151). These models are also used in protein structure prediction and drug screening (151).

LLMs’ capabilities in handling complex data, pattern recognition, and reading scientific literature make them suitable for spectral matching and metabolite annotation. For example, Fakouri Baygi et al. (141) developed IDSL_MINT, a cheminformatics deep learning framework based on transformer models, which predicts molecular fingerprint descriptors from MS/MS fragmentation spectra and facilitates the training of predictive models from various mass spectral libraries. RT-Transformer is another deep neural network model that uses graph attention networks and 1D-Transformer to predict retention times via various chromatographic methods, showing competitive performance and excellent scalability, thus increasing the accuracy of metabolite identification in liquid chromatography (138).

Although still in the early exploration stages and limited by the number of tokens in semantic libraries, LLMs present a promising solution for handling larger databases and various forms of metabolomics annotation sources.

We are in an era of data explosion, where the advancement of sequencing technologies has accumulated an unimaginable volume of data over the past few decades. This includes genomics, transcriptomics, epigenomics, proteomics, and microbiomics. Several successful examples have demonstrated the powerful effect of integrating multi-omics in biological and medical research. For instance, Chakraborty et al. (152) integrated ChIP-Seq and RNA-Seq data to study head and neck squamous cell carcinoma (HNSCC), finding that tumor-specific histones H3K4me3 and H3K27ac are associated with transcriptional changes in HNSCC driver genes, such as EGFR, FGFR1, and FOXA1. Additionally, Vaske et al. (153) used the path recognition algorithm (PARADIGM) integrated with genomic model data to infer the activity of patient-specific biological pathways from multi-omics data, better identifying changes in tumor-related metabolic pathways in glioblastoma multiforme (GBM) and breast cancer datasets. Mo et al. (154) hypothesized that different molecular phenotypes can be predicted through a set of orthogonal latent variables, which represent different molecular driving factors. They proposed a new framework that uses generalized linear regression to construct a joint model of categorical and numerical variables (both continuous and discrete) from integrated genomics, epigenomics, and transcriptomics data (154). In another study, integrating metabolomics and transcriptomics revealed molecular perturbations underlying prostate cancer. The metabolite sphingosine demonstrated high specificity and sensitivity for distinguishing prostate cancer from benign prostatic hyperplasia. Downstream of sphingosine, impaired sphingosine-1-phosphate receptor 2 signaling represents a loss of tumor suppressor function and a potential key oncogenic pathway for therapeutic targeting (155).

The combined effect of multi-omics analysis exceeds the sum of its parts, thus necessitating the integration and joint analysis of various omics data. However, different omics data have distinct analysis workflows and information densities. For example, transcriptomics data provide gene expression levels characterized by their quantitative nature, while proteomics and metabolomics data are obtained through mass spectrometry, which is more complex and diverse (156).

This integration presents several challenges, requiring robust computational methods to handle these challenges and effectively extract inter-omics relationships from heterogeneous, large-scale, and noisy biological data generated across different platforms, technologies, tissues, and species (149, 156). AI is a promising approach to address these challenges.

In metabolomics, AI technologies can be applied in multiple areas, including data preprocessing, feature extraction, pattern recognition, and data integration. For instance, deep learning algorithms can identify feature peaks in mass spectrometry data, enhancing the accuracy of metabolite identification (111, 117, 149). Additionally, machine learning methods can be used for the integrative analysis of multi-omics data, helping to uncover complex biological networks and pathways (156).

The analysis of metabolomic data and the integration of multi-omics typically involve steps such as correcting batch effects, inferring networks, reducing data dimensionality, and recognizing patterns. Batch effects, which result from differences in experimental conditions and personnel, can mask true biological variations, thereby impacting the accuracy and reliability of analyses (157).

AI is especially suited for exploring complex patterns and interpreting non-linear effects within such data. AI techniques, by learning the hidden structures and relationships in large, high-dimensional datasets, help researchers uncover complex biological signals that might be overlooked by traditional statistical methods (151). This capability is crucial for effectively integrating diverse omics data, providing a comprehensive view of biological systems. ComBat, an algorithm operating within an empirical Bayesian framework, is commonly used for batch correction (127). It adjusts batch effects by estimating and modifying batch-specific parameters to align them closer to pooled estimates, harmonizing data across different batches while preserving biological signals. Another sophisticated method for batch correction is the Mutual Nearest Neighbors (MNN) technique (128), which uses shared nearest neighbor information to correct batch discrepancies. By aligning nearest neighbors across batches, MNN effectively harmonizes datasets.

In terms of data correlation network, MMvec, inspired by word2vec, uses matrix factorization to analyze the co-occurrence patterns of microbes and metabolites, offering insights such as the metabolites produced by microbes in specific diseases (135). Similarly, Similarity Network Fusion (SNF) integrates multiple omics datasets by creating separate networks for each data type and merging them using non-linear fusion techniques based on message-passing theory (158). This integration deepens as the networks converge through successive iterations. Network-Based Multi-group Data Integration (NetICS) provides a strategy for integrating diverse group data in tumor gene sequencing (159). It utilizes network diffusion models on directed functional interaction networks to predict the impacts of genetic, epigenetic, and miRNA variations on downstream genes and proteins.

Multi-omics datasets encompass a wide range of data types and sizes, from gene or metabolite abundances in hundreds of samples to additional dimensions like sample origin and clinical data in single-cell studies, posing the ‘curse of dimensionality’ (160). Moreover, the high correlation among variables can introduce multicollinearity, complicating the identification of significant biological markers. Specific machine learning methods for metabolomics, such as PCA (112), t-SNE (161), and NMF (162), facilitate data reduction and visualization. However, deep learning and large language models surpass these methods by offering superior capabilities for feature learning and contextual understanding, thus expanding potential applications in complex biological data analysis.

Metabolomics not only provides crucial insights into metabolic pathways and physiological states but also plays a significant role in disease diagnosis, drug development, and nutritional research. However, the development of biomarkers faces considerable challenges, as traditional methods are time-consuming, costly, and often lack accuracy (163, 164). The rapid advancement of AI technologies, particularly machine learning and deep learning, has significantly increased the efficiency and accuracy of biomarker development. Machine learning and deep learning algorithms, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs) (112), and ANNs (112), excel in handling complex and high-dimensional data, capturing patterns and relationships that traditional methods might overlook. Compared with conventional models, the incorporation of these technologies has improved the predictive capabilities of biomarkers.

For instance, in a study on gastric cancer diagnosis and prognosis prediction, Chen et al. (165) utilized a machine learning random forest model to analyze plasma samples from 702 participants across multiple centers. They developed a diagnostic model comprising ten metabolites, achieving a sensitivity of 0.905, significantly surpassing the sensitivity of traditional cancer protein marker methods (less than 0.40) (165). Similarly, in a study by Kuwabara et al. (166) liquid chromatography-mass spectrometry (LC-MS) and the alternative decision tree (ADTree) algorithm were used to analyze 2602 saliva samples for colorectal cancer detection. Their model achieved an AUC value of 0.870 in distinguishing colorectal cancer from healthy controls, whereas traditional detection methods typically have AUC values around 0.70 (166).

The methods for establishing predictive and diagnostic models in metabolomics primarily include traditional machine learning algorithms and deep learning algorithms (167, 168). Among traditional machine learning algorithms, logistic regression (LR) is widely used for binary classification problems due to its simplicity, ease of interpretation, and fast computation (169). It has shown good performance in disease prediction based on metabolites. For example, in a study on endometrial cancer (EC) by Bahado-Singh et al. (170) a logistic regression model was developed using multiple metabolites and demographic characteristics. This model combined C14:2, phosphatidylcholine with acyl-alkyl residue sum C38:1 (PCae C38:1), and 3-hydroxybutyric acid, achieving an AUC (95% CI) of 0.826 (0.706-0.946), with a sensitivity of 82.6% and specificity of 70.8% (170). However, logistic regression is less capable of modeling complex non-linear relationships and may underperform when handling high-dimensional data with complex interactions.

Random forest (RF), on the other hand, improves model robustness and resistance to overfitting by constructing multiple decision trees and averaging their results, making it suitable for high-dimensional data processing. For instance, in a study on colorectal cancer by Telleria et al., RF and logistic regression models were used to develop an accurate predictive model based on several metabolites. This model combined hemoglobin (Hgb), bilirubin E,E, lactosyl-N-palmitoyl-sphingosine, glycocholenate sulfate, and STLVT, achieving an accuracy of 91.67% (95% CI 0.7753-0.9825), with a sensitivity of 0.7 and specificity of 1 (171).

Support vector machine (SVM) constructs hyperplanes for classification, making it suitable for small samples and high-dimensional data with good generalization ability. In a study on breast cancer by An et al., an SVM model was developed based on 47 metabolites. This model achieved high accuracy in breast cancer prediction (AUC = 1), with an AUC of 0.794 for breast cancer vs healthy controls (HC), and 0.879 for benign vs HC in the testing cohort (172).

Bayesian methods, based on Bayes’ theorem, can handle uncertainty in data and are suitable for small sample learning, although they require high computational complexity and rely on prior information. In a study on early lung cancer by Xie et al., a Naive Bayes algorithm was used to develop a predictive model based on six plasma metabolites. This model significantly distinguished early lung cancer patients from healthy individuals, achieving an AUC of 0.989, with a sensitivity of 98.1% and specificity of 100.0%. The study also identified the top five important metabolites as potential biomarkers for early lung cancer screening (173).

CNN uses combinations of convolutional layers, pooling layers, and fully connected layers to extract spatial features, making them suitable for image and time-series data processing. In metabolomics research, CNNs have demonstrated excellent feature extraction and prediction performance, although they require large amounts of data and high computational resources. In a study on cholangiocarcinoma and pancreatic adenocarcinoma by Urman et al., a neural network (NN) algorithm was used to develop predictive models based on multiple lipid and protein biomarkers. This model differentiated between benign strictures and cholangiocarcinoma patients with an AUC of 0.984, a sensitivity of 94.1%, and a specificity of 92.3%. The same method was also used to distinguish pancreatic adenocarcinoma patients from control groups, achieving an AUC of 0.98, a sensitivity of 88%, and a specificity of 100% (174). Furthermore, the study evaluated the performance of other machine-learning algorithms. The authors used a Bayesian variant of the general linear model (BGLM) and the C5.0 decision tree algorithm on the same data. Both C5.0 and BGLM demonstrated good performance in feature selection and prediction, although their predictive power was slightly lower than that of the NN algorithm.

ANN simulates the structure of human brain neurons and learns complex features and patterns through multiple layers of neurons and weight adjustments (116). ANN has strong feature extraction capabilities and is suitable for processing large-scale data, but it requires extensive data for training and high computational resources, and the model interpretability is poor. In a study on oral cancer by Monedeiro et al., an ANN model was developed based on nine relevant volatile organic compounds (VOCs). These compounds included 1-octen-3-ol, hexanoic acid, E-2-octenal, heptanoic acid, octanoic acid, E-2-nonenal, nonanoic acid, 2,4-decadienal, and 9-undecenoic acid. The model’s performance was assessed using 10-fold cross-validation and receiver operating characteristic curves, achieving an overall accuracy of 90%, with 100% sensitivity and specificity for oral cancer cases (175).

In Table 2 , we summarize the recent applications of different AI technologies in establishing predictive or diagnostic models through metabolomics. Various predictive models have achieved very high performance in their respective cohorts. It is important to note that most studies have employed cross-validation with separated cohorts and lack independent external validation cohorts. Most cohorts consist of around a hundred samples, which may lead to overfitting when using AI models to build predictive models, potentially resulting in poor validation performance in real-world external cohorts (167). The main reasons for the lack of external validation cohorts are data acquisition difficulties and high costs. Establishing comprehensive and interconnected channels and large integrated databases through AI technology may be a feasible approach in the future.