Metabolomic analysis for diagnosis of oral squamous cell carcinoma using machine learning

Abstract

Oral squamous cell carcinoma (OSCC) undergoes significant metabolic reprogramming. Adopting metabolomics to identify altered metabolic enzymes and metabolites holds promise for the early and precise diagnosis of OSCC. However, most current workflows rely on single-perspective modeling and lack task-aware prioritization. They underextract metabolomic signals, limit classification robustness, and impede a traceable transition from untargeted discovery to targeted panels. Therefore, we propose Metabolomics-based Integrated Information Learning (MIIL), an interpretable framework for OSCC diagnosis and premalignant screening. MIIL prioritizes task-relevant metabolites to derive a compact biomarker panel for targeted assay development, while strengthening classification via decision-level fusion of heterogeneous learners. Initially, we collect 120 oral mucosa tissues from 40 OSCC patients, including 40 Cancer (CA) samples, 40 Margin-1 specimens (M1), and 40 Margin-2 tissues (M2). Moreover, MIIL conducts untargeted metabolomics analysis to identify the most significant differential metabolites contributing to OSCC diagnosis. Subsequently, targeted metabolomics techniques are exploited for in-depth analysis of amino acid metabolites. Finally, integrated learning models are combined via decision-level fusion of linear, probabilistic, and margin-based signals, supporting accurate OSCC classification. The final results demonstrate that this method excels in the CA.vs.M1, CA.vs.M2, and M1.vs.M2 diagnostic tasks, achieving mean accuracies of 88.33%, 91.67%, and 85.00%, respectively. Trial registration: Chinese Clinical Trial Registry (ChiCTR), ChiCTR2200064861; registered on 2023-04-23.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04988-0.

Introduction

Oral squamous cell carcinoma (OSCC) is a common malignancy among head and neck cancers, with a five-year survival rate of approximately 60%. This relatively low survival rate is primarily due to the lack of efficient diagnostic methods and the current inability to identify reliable tumor markers or biological differential indicators [1–3]. Research has shown that the development of OSCC is accompanied by significant metabolic reprogramming, similar to other types of malignancies. To date, more than 100 metabolites have been identified as dysregulated during the malignant progression of OSCC, including lactate, glutamate, sialic acid, histamine, polyamines, choline, and carnitine, and significant differences in the levels of these metabolites have been confirmed among normal individuals, patients with precancerous lesions, and patients with OSCC [4, 5]. Therefore, the utilization of metabolomics technology to screen and identify differential metabolites in the carcinogenic process of oral mucosa provides important scientific evidence and a new research perspective for the early diagnosis and treatment of OSCC.

Although metabolomics can determine hundreds of metabolites in clinical samples, the complexity of data processing and interpretation remains a significant challenge [6]. The integration of mass spectrometry technology with Artificial Intelligence (AI) in recent years has enabled the identification and validation of metabolite biomarkers, providing a foundation for rapid OSCC screening [6–9]. In particular, machine learning, an AI technology widely applied in numerous fields of biomedical science, has demonstrated unique advantages in the automated analysis of complex data [10–13]. Machine learning is particularly suited for the analysis of omics data, constructing prognostic models, identifying biomarkers, and classifying patients for precision medicine [10]. These advancements not only deepen our understanding of the metabolic reprogramming characteristics of OSCC but also offer potential targets for the development of novel diagnostic methods. However, the application of machine learning in analyzing oral mucosal metabolomics data and developing potential biomarkers remains underexplored, leaving substantial room for further investigation.

Therefore, this study develops a Metabolomics-based Integrated Information Learning (MIIL) for intelligent OSCC diagnosis, which combines stacked statistical algorithms and AI methods to comprehensively analyze metabolomics data. By leveraging advanced statistical techniques, dimensionality reduction, and deploying integrated machine learning models, the MIIL framework can effectively distinguish cancerous tissues from normal tissues. These identified biomarkers not only facilitate the differentiation of malignant from benign tissues but also serve as early diagnostic tools for OSCC.

Motivation

In summary, the motivations of our study are defined as follows:

(1) Current metabolomics learning pipelines often rely on a single modeling perspective, which can be brittle in small, heterogeneous cohorts. Linear models capture effect direction and magnitude but may miss nonlinear separation. Probabilistic models reflect distributional shifts yet can be sensitive to feature dependence assumptions. Margin-based classifiers emphasize separability but may overreact to sampling noise. Consequently, a single model may underexploit the metabolomic signal and yield incomplete mining.

(2) Existing pipelines rarely prioritize metabolites in a task-driven manner across clinically relevant comparisons. Consequently, targeted panels are frequently curated without a traceable, data-driven selection rationale grounded in untargeted findings. This gap weakens interpretability and constrains the feasibility of scalable clinical screening.

Contribution

Therefore, the contributions of our research are summarized as follows:

1. We present an end-to-end and interpretable framework for OSCC diagnosis and premalignant screening. The framework translates untargeted findings into a compact amino-acid panel that informs targeted assay development. It also strengthens classification through decision-level fusion of heterogeneous learners. This design preserves biomarker-level traceability and supports clinically scalable translation.

2. We propose a metabolomics-based integrated learning framework that consolidates complementary metabolite evidence from heterogeneous learners. It aggregates linear effect signals, class-conditional probability estimates, and margin-based separability scores at the decision level, yielding a consensus prediction. This integration improves feature prioritization and classification robustness, supporting more reliable metabolic signature discovery and prediction.

3. Our study links untargeted discovery to targeted quantification. It robustly prioritizes amino acids across clinically relevant comparisons. We then derive a compact amino-acid panel for OSCC diagnosis and premalignant screening, facilitating future development of a targeted and high-throughput diagnostic/screening assay.

Related work

Traditional metabolomics analysis for OSCC diagnosis

Metabolomics explores the metabolic pathways within organisms by monitoring the alterations in their metabolic products following exposure to stimuli or disturbances or as a function of time. This approach has been leveraged for the early diagnosis of various cancers, including breast cancer, gastric cancer, and pancreatic cancer [14–16]. Research findings have demonstrated that, even in the early stages of tumor development, there are significant differences in the metabolic profiles when compared to those of healthy individuals [4]. This suggests that cancer induces a systemic metabolic response, highlighting the potential of metabolomics in cancer diagnosis and the understanding of its underlying mechanisms [17].

Molecular biology research has demonstrated that the onset and progression of OSCC is a multi-stage process that encompasses genetic mutations, epithelial dysplasia, carcinoma in situ, and invasive carcinoma stages. Currently, the diagnosis of OSCC predominantly relies on histopathological examination, lacking non-invasive, convenient, and highly sensitive diagnostic methods. Metabolomics, as a quantitative analytical approach, has shown a closer correlation with tumor phenotypic characteristics compared to genomics and proteomics. It plays a crucial role in the field of malignant tumor research, aiming to discover and identify key differential metabolites during the development of cancer. This research direction seeks to achieve early detection, diagnosis, and treatment of cancer through the early identification of differential metabolites, thereby significantly reducing the incidence and mortality rates of cancer. The application of metabolomics in this context holds immense significance in alleviating the societal and medical burden caused by OSCC. Kitabatake K et al. distinguished patients with oral leukoplakia and a control group of healthy individuals, as well as patients with leukoplakia with epithelial dysplasia, through saliva metabolite analysis [18]. Their discovery of metabolomic biomarkers assists in assessing the risk of cancerous transformation in oral leukoplakia, providing new insights into non-invasive screening methods. Hu Z et al. employed a two-sample Mendelian randomization approach to explore the impact of blood metabolites on OSCC, identifying 14 metabolites positively associated with the risk of OSCC and 15 protective metabolites [19]. This discovery sheds light on the potential role of blood metabolites in the development of OSCC and offers new biomarkers for its diagnosis and treatment.

Metabolomic research using human tissue specimens not only provides direct and accurate metabolic information but also reveals spatial metabolic heterogeneity and disease-specific alterations, which are of significant importance for early diagnosis, pathological analysis, and personalized treatment of diseases. In a study involving 28 paired OSCC and mucosal margin tissue samples at varying distances from the tumor, Yang et al. employed untargeted metabolomics and identified nine amino acids (e.g., alanine, proline, serine) as candidate metabolic biomarkers for abnormal proliferation of the oral mucosal epithelium. Further targeted metabolomic analysis revealed that six amino acids (e.g., aspartic acid, asparagine, proline) could serve as diagnostic markers for the abnormal proliferation of the oral mucosal epithelium, with aspartic acid and asparagine showing particularly significant clinical diagnostic value [20].

Amino acids, as the fundamental building blocks of macromolecular proteins, are crucial for the synthesis of glutathione and nucleotides, nitrogen metabolism, and cellular proliferation. Within the complex matrices of tissues, serum, and saliva, amino acids exhibit a wide dynamic range of concentrations. The variations in their endogenous levels are intimately linked to the onset and progression of cancer [21]. The molecular composition of samples can delineate various pathological and physiological alterations. Despite the complexity of amino acid metabolism within these samples, studying these amino acid changes can substantially facilitate the evaluation of molecular mechanisms in OSCC [22]. Therefore, the quantitative and qualitative analysis of amino acids in OSCC tissue samples holds significant clinical importance for identifying potential diagnostic biomarkers and exploring disrupted metabolic pathways associated with oral mucosal carcinogenesis.

Recent metabolomics research based on machine learning

Recent advancements in artificial intelligence have shown significant promise in the diagnosis and management of various cancers with metabolomics analysis [14, 23–27]. Sun et al. provided a comprehensive evaluation of preprocessing techniques for metabolomics data, including MS-based and NMR-based methods, and their applications in microbiome and metabolome research, highlighting their respective advantages and limitations [23]. Ning et al. conducted a cross-cohort integrative analysis of 9 metagenomic and 4 metabolomics cohorts, identifying consistent gut microbiota characteristics and significant metabolite alterations [24]. Chen et al. developed a 10-metabolite diagnostic model for gastric cancer through targeted metabolomics analysis of 702 plasma samples [14]. Zhao et al. identified aberrations in ’alanine, aspartate, and glutamate metabolism’ using NMR and MS-based metabolomics on matched ESCC tissues and biofluids from 560 participants, with NMR-based panels of five metabolites outperforming clinical serological markers for early-stage ESCC detection [25]. Buergel et al. utilized NMR spectroscopy-derived metabolomic profiles from 117,981 participants to train a neural network predicting the risk of 24 common conditions, indicating significant associations and predictive value over clinical variables for eight common diseases [26]. Lewis et al. employed genome-scale metabolic flux balance analysis models derived from transcriptomic and genomic datasets of 915 TCGA patient tumors to identify and experimentally validate metabolic biomarkers that differentiate radiation-sensitive from radiation-resistant tumors [27].

Despite the growing interest in metabolomics and machine learning for cancer diagnostics, studies specifically focusing on OSCC remain limited. This gap underscores the necessity for further research to establish robust, automated diagnostic models tailored for OSCC, which can leverage the strengths of both metabolomics and artificial intelligence to improve early detection and treatment outcomes.

Method

Overall

To improve OSCC diagnosis and premalignant screening from metabolomics data, we propose Metabolomics-based Integrated Information Learning (MIIL), which integrates statistical screening with machine-learning models. As shown in Fig. 1, MIIL screens metabolomic features to identify candidate biomarkers associated with each diagnostic task (e.g., amino acids). We then quantify the selected metabolite class using targeted assays and train classifiers for diagnosis. This integration improves biomarker validation and diagnostic performance for OSCC, making it a robust tool for early detection and precise clinical diagnosis.

Fig. 1.

Framework of the MIIL pipeline for OSCC diagnosis and premalignant screening. a First, paired tissues were collected as CA, M1, and M2, with matched pathology. b Then untargeted LC-MS/MS profiled metabolites and shortlisted differential features based on VIP, FC, and statistical significance. c The resulting signatures were summarized by KEGG pathway annotation, motivating amino-acid prioritization. d Targeted MRM quantified amino acids, followed by feature filtering using FC and P-value criteria. e Subsequently, logistic regression, Naive Bayes, and SVM were integrated via ensemble learning to construct MIIL classifier. f Performance was assessed for CA.vs.M1 and CA.vs.M2 (OSCC diagnosis) and for M1.vs.M2 (early-lesion screening)

Data collection

This study was approved by the Ethics Committee of the Beijing Stomatological Hospital, Capital Medical University, with informed consent secured from all subjects involved. We collected data from patients with OSCC who were admitted for surgical treatment at the department of oral and maxillofacial surgery from January 2023 to February 2024. Inclusion criteria were as follows: (1) patients aged 18 years and older, regardless of gender; (2) patients diagnosed with squamous cell carcinoma via pathological biopsy; (3) patients who underwent surgical treatment during hospitalization; (4) patients who had not received chemotherapy, radiotherapy, or other tumor treatments before surgery. Exclusion criteria included: (1) patients with severe underlying diseases unable to tolerate surgery; (2) patients with mental disorders; (3) recurrent tumors; (4) female patients who were pregnant or breastfeeding.

In this study, we employed liquid chromatography-tandem mass spectrometry (LC-MS/MS) for untargeted metabolomic analysis to screen the metabolic profile characteristics of 60 OSCC tissue samples, encompassing a total of 20 cases, each comprising three samples. The collection criterion for each case was defined as Cancer (tumor tissue, CA), Margin-1 (0.5 cm from the tumor boundary, representing the dysplastic epithelial zone, M1), and Margin-2 (2.0 cm from the tumor boundary, representing the normal tissue, M2). In our institution, Margin-1 (M1, 0.5 cm from the tumor boundary) and Margin-2 (M2, 2.0 cm from the tumor boundary) were defined as fixed sampling anchors to capture a reproducible spatial metabolic gradient from tumor-adjacent dysplastic tissue to distal tissue approximating normal mucosa. This fixed-distance scheme was implemented to standardize sampling across patients and should not be interpreted as a universal ’safe margin’ threshold, which may vary by surgical practice and anatomical context. Biologically, oral field cancerization is characterized by a continuum of molecular and histopathological changes extending beyond the tumor edge, supporting the rationale for sampling a near margin and a distal reference tissue. Distance-related metabolic alterations in OSCC surgical margins have been reported previously [20, 28–30], consistent with our design. Notably, histopathology in our cohort confirmed that CA corresponded to squamous cell carcinoma, M1 to epithelial dysplasia, and M2 to histologically normal mucosa (Fig. 1a).

For each patient, two adjacent tissue samples were collected: one for histopathological diagnosis and the other for metabolomic analysis. Additionally, this research involved the recruitment of 20 more cases (each including three samples: Cancer (CA), Margin-1 (M1), and Margin-2 (M2), collected in the same manner) for multiple reaction monitoring (MRM)-based targeted quantitative analysis of amino acids. This sample size is consistent with the scale commonly used in OSCC tissue metabolomics studies, and the within-subject paired design (CA/M1/M2 per patient) increases statistical efficiency by controlling inter-individual variability [31, 32]. All tissue samples designated for metabolomic analysis were rapidly frozen in liquid nitrogen within 30 min post-surgery and stored at C until further processing. Additional patient-level clinical variables (e.g., BMI, tumor site, smoking/alcohol history, and major comorbidities including hypertension and diabetes) are summarized for each cohort in Additional File 6: Table S3 (untargeted LC-MS/MS cohort) and Additional File 7: Table S4 (targeted MRM cohort).

Specimens intended for pathological examination were fixed in formalin solution for 24–48 h, subsequently undergoing routine paraffin embedding and hematoxylin and eosin (H&E) staining. The pathological diagnosis was conducted by the pathologist. The diagnostic outcome revealed that all samples within the Cancer (CA) group were identified as OSCC. The Margin-1 (M1) group was diagnosed with epithelial dysplasia, whereas the Margin-2 (M2) group was histologically recognized as normal tissue (Fig. 1a).

Metabolomics analysis

Untargeted metabolomics process

Untargeted metabolomics analysis was performed using a Vanquish UHPLC system (ThermoFisher, Germany) coupled with an Orbitrap Q ExactiveTM HF mass spectrometer or an Orbitrap Q ExactiveTM HF-X mass spectrometer (Thermo Fisher, Germany). Before LC-MS/MS analysis, samples were prepared as follows: tissues (100 mg) were individually ground using liquid nitrogen, and the homogenate was resuspended in prechilled 80% methanol with thorough vortexing. The samples were then incubated on ice for 5 min, followed by centrifugation at 15,000 g at C for 20 min. A portion of the supernatant was diluted to a final concentration containing 53% methanol using LC-MS grade water. Subsequently, the samples were transferred to fresh Eppendorf tubes and centrifuged again at 15,000 g at C for 20 min. The resulting supernatant was injected into the LC-MS/MS system for analysis. For the positive ion mode, mobile phase A consisted of 0.1% formic acid, while mobile phase B was methanol. For the negative ion mode, mobile phase A was 5 mM ammonium acetate, and mobile phase B was methanol. Molecular formulas were predicted based on molecular ion peaks and fragments, and the data were compared against the mzCloud (https://www.mzcloud.org/), mzVault, and Masslist databases to remove background ions. The raw quantification results were normalized using the following formula: sample raw quantification value / (sum of sample metabolite quantification values/sum of QC1 sample metabolite quantification values), yielding relative peak areas. Compounds with a coefficient of variation greater than 30% in relative peak areas within QC samples were excluded, leading to the final identification and relative quantification of metabolites.

A pooled QC sample was generated by mixing equal aliquots of all study samples and was injected before, during, and after the analytical run to monitor system stability (30 QC injections in total; Additional File 1: Table S1). Three blank injections were included, where the first two served as system washes and the third blank was used for background subtraction. Features with QC CV > 30% were removed during preprocessing. As all samples were acquired in a single batch and QC clustering indicated stable performance (Additional File 2: S1 and Fig. S1), no additional drift correction (e.g., QC-RLSC/LOESS) or batch correction (e.g., ComBat) was applied.

After the untargeted metabolomics process, this study collects a metabolomics-based OSCC dataset containing samples, denoted as , where each sample comprises features. In addition, the corresponding one-hot label matrix is , indicating the class membership of each sample. To better illustrate, the metabolomics dataset in this research is partitioned into groups of positive samples and negatives according to their categories (CA.vs.M1, CA.vs.M2, M1.vs.M2), where and represent the sample numbers of each group, respectively.

Differential metabolite analysis

The LC-MS/MS workflow yielded approximately 3000 metabolic features. The high dimensionality of this metabolomics dataset necessitated extensive computational effort for its analysis. To address this challenge, we implemented a combination of differential metabolite analysis algorithms to filter and identify the metabolites that contribute most significantly to OSCC diagnosis, providing a foundation for subsequent targeted metabolomics analyses. To screen all extracted metabolites, we retained features that simultaneously satisfied the following decision rule:

1

where denotes the variable-importance score derived from a PLS-DA model, quantifies group-wise abundance shifts, and returns the two-sided significance level from a two-sample t-test. The full definitions and computational procedures of these metrics are provided in Additional File 3: S2.

Prior to PLS-DA, metabolite intensities were log2-transformed, followed by mean-centering and unit-variance scaling (autoscaling/UV scaling). PLS-DA was used as an exploratory multivariate visualization, and model validity was assessed by reporting R2/Q2 together with 200-permutation tests (see Additional File 4: S4 and Fig. S2-4). In addition, p-values were adjusted for multiple testing using the Benjamini-Hochberg false discovery rate (BH-FDR) procedure, and the corresponding q-values are reported in Additional File 5: Table S2A-S2F. For exploratory feature screening, we applied nominal p-values to preserve sensitivity, while BH-FDR adjusted q-values were also computed and reported to enable robustness assessment under multiple-testing control.

Upon the above differential metabolite analysis methods, we identified 145, 349, and 277 differential metabolites in the CA.vs.M1, CA.vs.M2, and M1.vs.M2 comparisons, respectively. Subsequent statistical analysis revealed that the majority of these metabolites were amino acids, peptides, and their analogs. Given the prominence of amino acids among the identified metabolites, we conducted further targeted metabolomic analyses specifically focusing on this class of metabolites. This targeted approach allowed for more detailed characterization and quantification, enhancing our understanding of their roles in the metabolic reprogramming associated with OSCC.

Targeted metabolomics process

During the targeted metabolomics process, chromatographic analysis was conducted using an ACQUITY UPLC BEH Amide column (2.1 × 100 mm, 1.7 m) to achieve efficient separation of amino acid metabolites. The mobile phase consisted of phase A (5 mM ammonium acetate aqueous solution containing 0.1% formic acid) and phase B (acetonitrile containing 0.1% formic acid). Mass spectrometric analysis was performed using an electrospray ionization source (ESI) in positive ionization mode. MRM was employed for mass spectrometric scanning, ensuring specific detection of target amino acids and internal standards. Targeted amino-acid quantification was performed in MRM mode and reported as absolute concentrations (g/g tissue). Each analyte was quantified against its matched internal standard. The analyte-internal standard mapping, calibration equations, and quantification limits are provided in Additional File 8. Concentration results for all study samples are provided in Additional File 9: Table S5, and QC injections with their quantitative stability monitoring are provided in Additional File 1: Table S1. Assay performance including accuracy/precision, stability, and recovery is summarized in Additional File 10: Table S6. Blank/wash injections were included in the analytical sequence to minimize and monitor potential carryover. A standard curve was established via LC-MS to evaluate linearity, with the limit of quantification (LOQ) determined by a signal-to-noise ratio (S/N 10). Precision was indicated by intraday and interday RSDs 15%. Sample stability within 24 h exhibited an RSD 15%. The stability of QC sample detection, with an RSD 15%, confirmed that the method is stable and reliable.

Finally, we conducted differential metabolite screening on the quantified outputs from targeted metabolomics. Metabolites were retained if they satisfied the following rule:

2

Given the selected index set and the quantified feature matrix , we formed the classifier input by subsetting the corresponding columns:

3

Machine learning-based classification for OSCC diagnosis and premalignant screening

Machine learning (ML) is well suited for classifying OSCC-related metabolic profiles and identifying discriminative biomarkers. In this study, ML models are used to distinguish CA.vs.M1, CA.vs.M2, and M1.vs.M2 based on metabolomic features.

Therefore, our MIIL model employs an ensemble learning strategy, integrating a diverse set of advanced machine-learning techniques to improve OSCC diagnosis. The ensemble combines sub-classifiers, including logistic regression, Naive Bayes, and Support Vector Machine (SVM), leveraging their complementary strengths to enhance diagnostic accuracy and robustness. By synergistically addressing the complexities of OSCC classification, the MIIL model demonstrates a significant capability to provide reliable and efficient diagnostic support in clinical applications. Following the differential metabolite analysis process, the data and corresponding labels used as inputs for the machine learning classifiers are defined as and , respectively.

Given the inherent limitations of individual classifiers, we incorporate ensemble learning to amalgamate the outputs of multiple classification algorithms, thereby mitigating their respective shortcomings while amplifying their strengths. Let denote the base learners, instantiated as logistic regression, Naive Bayes, and SVM in this study. For each specimen , each base learner outputs a class-probability vector. We aggregate these probabilities via soft voting to obtain the final probability vector and infer the final diagnosis accordingly:

4

5

where denotes the aggregated probability of assigning specimen to class c, and is the final predicted label obtained by selecting the class with the maximum aggregated probability. The detailed formulations of each base learner are reported in Additional File 3: S3.

Experiment

Dataset and implementation

This study incorporated a total of 120 tissue samples. Of these, 60 were subjected to LC-MS/MS for untargeted metabolomic analysis, while the remaining 60 underwent MRM for targeted amino acid analysis. The clinical characteristics of the patients in this study are detailed in Table 1. The experimental analysis was divided into three tasks: CA.vs.M1, CA.vs.M2, and M1.vs.M2, with 40 samples per task. Notably, in the untargeted metabolomics task, we analyzed the positive- and negative-ion modes separately. For each classification task, train-test partitioning was performed at the patient level to prevent information leakage. In each run, 30% of patients were randomly assigned to the test set. We conducted five independent runs with different random seeds. Within each train/test split, feature screening was performed using the training set only.

Table 1.

Clinical characteristics of patients with OSCC in this study

Group	Gender		Age	Clinical stage
	Male	Female		I-II	III-IV
LC-MS/MS	11	9	61.95 ± 12.14	5	15
MRM	10	10	64.70 ± 8.44	9	11

OSCC oral squamous cell carcinoma; LC-MS/MS liquid chromatography-tandem mass spectrometry; MRM multiple reaction monitoring. Values are counts unless otherwise indicated. Age is presented as mean ± standard deviation (years). Clinical stage was categorized as I-II versus III-IV according to routine clinical staging practice

The experiments in this study were conducted using Python 3.9.0. Key algorithmic libraries applied in our workflow included NumPy for numerical computations, pandas for data manipulation and analysis, sklearn (scikit-learn) for implementing machine learning algorithms, and matplotlib for data visualization. These libraries collectively provided a comprehensive and efficient framework for processing metabolomics data, training and validating machine learning models, and visualizing the results.

Performance parameters

To comprehensively evaluate the diagnostic performance of the Metabolomics-based Integrated Information Learning (MIIL) model for OSCC, we selected several critical performance metrics, including Accuracy, Recall, Precision, and F1-score. These metrics are defined as follows:

6

7

8

9

where TP denotes true positives, FP false positives, TN true negatives, and FN false negatives, respectively.

Results

Results of untargeted metabolomics analysis

Analysis of differential metabolites. As shown in Fig. 2, Venn diagrams were employed to illustrate the unique and shared differential metabolites across different groups. In the positive and negative ion modes, 21 and 4 differential metabolites, respectively, were common among the CA, M1, and M2 groups. In addition, permutation testing (200 permutations) indicated no obvious overfitting across comparisons, with negative Q2-intercepts in the validation plots (Additional File 4: S4 and Fig. S2-4).

Fig. 2.

The overlap of differential metabolites across pairwise comparisons in untargeted LC-MS/MS. Numbers indicate metabolite counts within each region. a Positive-ion mode. b Negative-ion mode

Hierarchical clustering analysis (HCA) is conducted on all differential metabolites obtained from the pairwise comparisons, normalizing and clustering the relative quantification values of the differential metabolites (Fig. 3). The clustering analysis revealed that certain metabolites were abundantly present in specific samples. Additionally, a significant positive correlation was observed between Adenylosuccinic acid and 2-Methylpentanedioic acid.

Fig. 3.

Hierarchical clustering of differential metabolites in untargeted metabolomics. Heatmaps show normalized relative peak areas for metabolites identified as differential. Rows correspond to metabolites and columns to samples; the color scale encodes scaled abundance. Dendrograms summarize hierarchical clustering based on the normalized profiles. a Positive-ion mode. b Negative-ion mode

KEGG pathway annotation. The analysis in Fig. 4 revealed that a significant number of metabolites are involved in amino acid metabolism, carbohydrate metabolism, lipid metabolism, and the metabolism of cofactors and vitamins, indicating the prominence of these metabolic pathways in the sample. Additionally, a high number of metabolites were associated with pathways related to infectious diseases, particularly bacterial infections, and endocrine and metabolic diseases, which may be linked to the malignant transformation of oral mucosa. The KEGG pathway annotation results reflect the distribution of metabolites across various categories, with a particularly large proportion involved in metabolic pathways, suggesting that metabolic processes were predominantly detected in this study.

Fig. 4.

KEGG pathway annotation of differential metabolites identified from untargeted LC-MS/MS. Differential metabolites were mapped to KEGG to summarize their functional distribution across major biological categories. Bar lengths indicate the number of annotated metabolites assigned to each pathway term. a Positive-ion mode. b Negative-ion mode

KEGG enrichment analysis. The results of the KEGG pathway enrichment analysis for the differential metabolites are shown in Table 2. Comparing the CA group with M1, the Cysteine and methionine metabolism pathway is the most enriched, followed by the Arginine and proline metabolism pathway. The Aminoacyl-tRNA biosynthesis pathway is more enriched in the comparison between CA and M2, while the ABC transporters pathway is more enriched between M1 and M2.

Table 2.

The KEGG enrichment analysis

MapID	MapTitle	P-value
map00260	Glycine, serine and threonine metabolism	0.43
map00270	Cysteine and methionine metabolism	0.43
map00520	Amino sugar and nucleotide sugar metabolism	0.43

MapID denotes the KEGG pathway identifier; MapTitle is the pathway name. P-value is based on the set of differential metabolites identified in this study

Results of targeted metabolomics analysis

PCA analysis. Principal component analysis (PCA) was utilized to analyze the amino acid metabolomic data derived from oral mucosal tissue samples [33]. The results of the PCA analysis are illustrated in Fig. 5, where PC1 and PC2 together explain 81.58% of the total variance. Additionally, the samples exhibit a distinct separation trend in the PCA plot, with the CA group samples primarily distributed on the left side of the plot, while the M1 and M2 group samples are located on the right side. This distribution pattern may reflect differences in the amino acid metabolic characteristics among the various groups. Specifically, the CA group samples show lower scores on PC1, whereas the M1 and M2 group samples demonstrate higher scores on PC1. This observation suggests that significant alterations in amino acid metabolism may occur during the progression of oral mucosal carcinogenesis.

Fig. 5.

The PCA score plot of amino acid metabolomic data. Each point represents one sample and is colored by group (CA, M1, and M2). Ellipses denote the 95% confidence regions for each group

Clustering analysis. Hierarchical clustering analysis (HCA) was performed on all identified amino acids, with their concentration values normalized prior to clustering. The results indicate distinct differences in amino acid metabolic patterns across the different groups (CA, M1, M2). Furthermore, within-group trends show similar variations in amino acid concentrations among the samples. Amino acids exhibiting similar metabolic patterns may be associated with the progression of oral mucosal carcinogenesis. Certain amino acids in the red high-expression and blue low-expression regions in Fig. 6 may collectively participate in the metabolic processes of precancerous lesions.

Fig. 6.

Hierarchical clustering heatmap of amino acid metabolomic data quantified by targeted MRM. Rows correspond to amino acids and columns to individual samples. Values are scaled abundances. Dendrograms summarize similarity-driven clustering of amino-acid profiles

Metabolite characterization and screening. The identification of differential metabolites was based on fold change (FC) and P-value thresholds. Comparative analyses identified 17 differential amino acids for CA.vs.M1 and 21 for CA.vs.M2 (Additional File 11: Table S7). In addition, 8 differential amino acids were detected for M1.vs.M2. These results indicate distinct amino-acid metabolic profiles across the three groups. The observed between-group differences may reflect metabolic alterations associated with lesion progression and the severity of epithelial dysplasia.

MIIL classification performance and ablation analysis. We adopt ensemble learning to classify samples based on differential metabolites. Multiple metrics are reported to quantify diagnostic performance. As shown in Tables 3, 4, and 5, MIIL achieves accuracies of 88.33%, 91.67%, and 85.00% for the CA.vs.M1, CA.vs.M2, and M1.vs.M2 tasks, respectively. These results underscore the potential of MIIL for automated OSCC diagnosis and premalignant screening.

Table 3.

Summary of diagnostic results in CA.vs.M1 task

Model	Accuracy	Recall	Precision	F1-score
LR	81.67±3.73	83.33±16.67	82.67±10.31	81.50±5.43
NB	85.00±3.73	86.67±13.94	85.48±9.08	84.94±4.59
SVM	81.67±10.87	76.67±25.28	85.95±9.00	78.92±14.88
MIIL w/o feature selection	80.00±4.56	80.00±21.73	82.67±10.31	78.83±8.66
MIIL with vote stable features	86.67±4.56	90.00±9.13	85.48±9.08	87.12±4.24
MIIL(ours)	88.33±4.56	93.33±9.13	85.95±9.00	88.91±4.14

LR logistic regression, NB, Naive Bayes, SVM support vector machine, MIIL (ours) our proposed model, F1-score harmonic mean of precision and recall. Accuracy, recall, precision and F1-score are reported as mean ± SD (%)

Table 4.

Summary of diagnostic results in CA.vs.M2 task

Model	Accuracy	Recall	Precision	F1-score
LR	88.33±9.50	93.33±9.13	85.24±10.16	88.97±8.96
NB	90.00±6.97	96.67±7.45	87.14±12.53	90.93±5.89
SVM	90.00±3.73	96.67±7.45	86.43±8.89	90.71±2.86
MIIL w/o feature selection	90.00±3.73	96.67±7.45	86.43±8.89	90.71±2.86
MIIL with vote stable features	91.67±5.89	96.67±7.45	89.29±10.71	92.25±5.11
MIIL(ours)	91.67±5.89	96.67±7.45	89.29±10.71	92.25±5.11

LR logistic regression, NB Naive Bayes, SVM support vector machine, MIIL (ours) our proposed model, F1-score harmonic mean of precision and recall. Accuracy, recall, precision and F1-score are reported as mean ± SD (%)

Table 5.

Summary of diagnostic results in M1.vs.M2 task

Model	Accuracy	Recall	Precision	F1-score
LR	70.00±9.50	56.67±25.28	79.29±13.08	62.94±16.80
NB	80.00±9.50	83.33±11.79	79.29±13.08	80.71±9.32
SVM	76.67±6.97	80.00±18.26	77.67±13.72	77.02±7.40
MIIL w/o feature selection	83.33±13.18	86.67±13.94	81.43±13.08	83.85±13.00
MIIL with vote stable features	81.67±6.97	83.33±0.00	81.90±11.74	82.28±5.79
MIIL(ours)	85.00±6.97	90.00±9.13	83.10±11.12	85.84±6.19

LR logistic regression, NB Naive Bayes, SVM support vector machine, MIIL (ours) our proposed model, F1-score harmonic mean of precision and recall. Accuracy, recall, precision and F1-score are reported as mean ± SD (%)

To dissect the contribution of each component, we conducted ablation analyses that compared single classifiers (LR, NB, and SVM) against the ensemble, removed variable screening (MIIL w/o feature selection), and replaced per-split feature screening with a vote-stable feature set shared across repetitions (MIIL with vote stable features). Under five repeated patient-level splits, metabolites selected in at least four runs were defined as vote-stable and were retained for downstream analysis.

Across the OSCC diagnosis tasks, MIIL consistently outperformed the strongest single-model baseline in terms of accuracy. For CA.vs.M1, MIIL improved accuracy by 3.33 pp over NB. It also exceeded the no-screening variant by 8.33 pp. MIIL further surpassed the vote-stable feature variant by 1.66 pp. For CA.vs.M2, MIIL achieved a 1.67 pp gain over the best single-model baselines and the no-screening variant. It matched the vote-stable feature variant (both 91.67%), indicating that stable feature aggregation already captures much of the discriminative signal for this task.

For premalignant screening (M1.vs.M2), MIIL delivered a larger benefit, improving accuracy by 5.00 pp relative to NB. It also improved upon MIIL w/o feature selection by 1.67 pp and exceeded the vote-stable feature variant by 3.33 pp. These results indicate that both task-aware metabolite prioritization and decision-level fusion contribute to stronger generalization, particularly for the more challenging M1.vs.M2 discrimination.

Confusion matrix analysis. Figure 7 summarizes aggregated confusion matrices across five repeated patient-level holdout runs. Counts were obtained by summing the test-set matrices from all runs. For CA.vs.M1 and CA.vs.M2, CA was defined as the positive class. The classifier yielded few false negatives (2 and 1) and limited false positives (5 and 4). These patterns indicate reliable OSCC discrimination. For M1.vs.M2, M1 was assigned as the positive class. Misclassifications increased (FP = 6, FN = 3), consistent with subtler stage-to-stage differences. Even so, correct predictions remained dominant, supporting feasibility for premalignant screening.

Fig. 7.

Aggregated confusion matrices of our MIIL classifier. Counts were aggregated by summing confusion matrices over five independent runs with different random seeds. a CA.vs.M1. b CA.vs.M2. (c) M1.vs.M2

Sensitivity at clinically relevant specificity. Figure 8 reports sensitivity as fixed specificity scores. For OSCC diagnosis (CA.vs.M1 and CA.vs.M2), clinically relevant specificity can be set at a high level. Sensitivity remained close to 90.00% across most of the specificity grid, indicating stable detection under stringent false-positive control. In contrast, M1.vs.M2 screening exhibited a stronger sensitivity-specificity trade-off. Sensitivity declined at higher specificity targets. Screening typically prioritizes sensitivity over specificity. Under a moderate constraint (specificity = 60.00%), sensitivity reached 93.33%, supporting effective case-finding when higher recall is desired.

Fig. 8.

Sensitivity at fixed specificity for the MIIL classifier. For each target specificity on the x-axis, a decision threshold was determined on the training cohort to maximize sensitivity while satisfying the specified minimum specificity. Sensitivity was then computed on the corresponding test cohort. Points denote the mean sensitivity across five independent runs with different random seeds, and step segments connect adjacent specificity targets. a CA.vs.M1. b CA.vs.M2. c M1.vs.M2

ROC analysis. The discriminative capability of candidate diagnostic biomarkers was quantified via receiver operating characteristic (ROC) analysis of differential metabolites [34]. Figure 9 presents the ROC curves, with the shaded bands indicating the corresponding 95% Confidence Intervals (CIs). For OSCC-oriented diagnosis, the metabolites achieved strong separation in both CA.vs.M1 and CA.vs.M2 comparisons (AUC = 0.8978 and 0.9856, respectively). The relatively tight 95% CIs around these curves further indicate stable discrimination across resampling uncertainty.

Fig. 9.

ROC curves of the MIIL classifier across three tasks. Curves report AUC values, and shaded bands indicate 95% CIs. Additionally, the diagonal line denotes chance-level performance. a CA.vs.M1. b CA.vs.M2. c M1.vs.M2

In contrast, differentiating M1 from M2 represents a premalignant screening and is intrinsically more challenging. Consistent with this setting, the M1.vs.M2 comparison showed reduced separation (AUC = 0.8311) with a broader 95% CI. Even so, the model retained moderate-to-strong discriminative ability, supporting its potential for premalignant screening.

Calibration analysis. We further examined probability calibration using reliability diagrams and Brier scores. Figure 10 contrasts mean predicted probabilities with the observed fraction of positives. The diagonal line indicates perfect calibration. For OSCC-oriented diagnosis, calibration was favorable in CA.vs.M1 and CA.vs.M2. The corresponding Brier scores were 0.109 and 0.067, respectively. The CA.vs.M2 curve closely followed the diagonal, indicating well-calibrated risk estimates.

Fig. 10.

Calibration curves of our MIIL model. Each curve compares mean predicted probability with the observed fraction of positives. The diagonal indicates perfect calibration. All results were summarized over five independent patient-level holdout runs with different random seeds a CA.vs.M1. b CA.vs.M2. c M1.vs.M2

In contrast, M1.vs.M2 exhibited larger deviations from the diagonal. The Brier score increased to 0.170, consistent with the more challenging premalignant screening setting. The curve retains a broadly increasing tendency overall, although deviations across probability bins indicate unstable calibration due to limited samples.

Discussion

In this study, we utilized metabolomic analysis to identify diagnostic biomarkers for normal oral mucosa, epithelial dysplasia, and OSCC. The screening results indicate that in the CA.vs.M1 task, a total of 145 differential metabolites are identified, with 102 being positive and 43 negative. Among these, 34 are amino acids, comprising 28 positive and 6 negative. In the CA.vs.M2 task, 349 differential metabolites are identified, with 224 being positive and 125 negative. Within this subset, 76 are amino acids, including 53 positive and 23 negative. In the M1.vs.M2 task, 277 differential metabolites are identified, with 178 being positive and 99 negative. Among these, 66 are amino acids, with 44 positive and 22 negative. Across all tasks, amino acids consistently represent the largest category of differential metabolites, highlighting their significant role in the metabolic reprogramming associated with OSCC.

Precancerous lesions of the oral mucosa, which present morphological changes in tissue, are associated with an increased probability of transformation into squamous cell carcinoma. Among the recognized precancerous lesions, leukoplakia of the oral mucosa is the most common [35]. Based on histopathological diagnosis, leukoplakia can be categorized into epithelial benign hyperkeratosis and epithelial dysplasia; the latter being a precancerous condition with varying degrees of dysplasia and associated risks of malignancy. Differentiating between simple hyperplasia and dysplastic leukoplakia is crucial for the early detection of malignant transformations. However, aside from a few cases where leukoplakia deterioration may manifest as ulceration and pain, it is challenging to distinguish between these two conditions clinically [36–38]. In this study, we utilized both untargeted and targeted metabolomics methods via LC-MS/MS and MRM to develop the MIIL model. This model combines statistical analysis and AI to analyze key information within metabolomic features. The results demonstrated that our integrated model achieved accuracy rates of 93.33%, 88.33%, and 85.00% for the tasks of discriminating CA from M1, CA from M2, and M1 from M2, respectively. This suggests that the MIIL model, grounded in metabolomics, holds significant value in the screening for early cancerous changes in the oral mucosa and surveillance of populations susceptible to oral mucosal diseases.

To explicitly address which metabolites drive the MIIL model’s decisions, we summarized the amino acid features selected for each diagnostic task (Additional File 11: Table S7). The models consistently prioritized a core set of amino acids, including key nodes in glutamine metabolism (glutamate, glutamine), the serine-glycine-one-carbon axis (serine, glycine), and branched-chain amino acids (leucine, isoleucine, valine) [39–41]. Most notably, a compact subset of only eight amino acids-glutamate, glycine, serine, threonine, histidine, aspartate, arginine, and proline-was sufficient for the model to distinguish dysplastic margin (M1) from normal tissue (M2) with 85.00% accuracy (Table 5). This refined feature set maps directly onto the most significantly dysregulated pathways identified in our KEGG analysis (Fig. 4, Table 2), such as ’Alanine, aspartate and glutamate metabolism’ and ’Glycine, serine and threonine metabolism’. Thus, the AI model does not merely perform pattern recognition; it highlights a mechanistically anchored, minimal set of metabolites that capture key metabolic reprogramming along the carcinogenic continuum [42, 43].

The identification of a compact, high-impact metabolite set-particularly the eight-amino-acid signature for M1.vs.M2 discrimination-supports a direct path toward clinical translation as a streamlined, targeted MRM biomarker panel. Such a panel would substantially reduce the complexity and cost compared with untargeted metabolomics, improving feasibility for routine implementation [44]. Potential applications include: (a) intraoperative margin assessment to complement frozen-section histopathology, especially in equivocal cases [45]; (b) risk stratification of oral potentially malignant disorders by quantifying the metabolic gradient associated with field cancerization to guide surveillance intervals [46]; and (c) a foundation for future non-invasive screening after validation in saliva or plasma [47]. We acknowledge that rigorous external validation, standardization of pre-analytical variables (e.g., ischemia time), and clinically meaningful thresholds are necessary next steps.

We observed that glutamate levels in the CA group were significantly higher than those in both the M1 and M2 groups (Fig. 11). Furthermore, the M1 group exhibited higher glutamate levels compared to the M2 group. The metabolic pathway of glutamine plays a crucial role in maintaining intracellular redox balance by promoting the synthesis of glutathione (GSH) and increasing the levels of reduced nicotinamide adenine dinucleotide phosphate (NADPH), thereby inhibiting the activity of reactive oxygen species (ROS) [22, 48]. Glutamine is transported into the cytoplasm via specific solute carrier (SLC) family transport proteins, and subsequently, it is transported into the mitochondria through variants of the alanine-serine-cysteine transporter 2 (ASCT2). Within the mitochondria, glutamine is deaminated to glutamate by the enzyme glutaminase (GLS). Following this, glutamate is converted to -ketoglutarate (-KG) and oxaloacetate (OAA) through the action of glutamate dehydrogenase 1 (GDH1) and several mitochondrial transaminases [49]. This process not only supports the tricarboxylic acid (TCA) cycle but also promotes ATP production through a series of biochemical reactions [50]. As a metabolite of glutamine, glutamate serves not only as a critical energy source and precursor for the biosynthesis of macromolecules in tumors but also can be converted into other amino acids such as proline, alanine, and aspartate. Under glucose-restricted conditions, glutamate can act as an alternative energy source, enhancing the TCA cycle through the production of -KG [51]. In tumors that exhibit glutamate addiction, phosphoenolpyruvate (PEP) generated during glutamate metabolism serves as a supplement to the PEP in the glucose metabolic pathway, thereby maintaining the equilibrium of tumor cell metabolism [52–54].

Fig. 11.

Group-wise comparison of targeted amino-acid features. Bars represent mean values for each group, and error bars indicate variability across samples a CA.vs.M1. b CA.vs.M2. c M1.vs.M2

Through the correlation analysis of all metabolite concentration changes (Fig. 12), we discovered a significant positive correlation between glycine and serine. As one-carbon donors in the folate cycle, glycine, serine, and their associated enzymes play crucial roles in promoting nucleotide synthesis, methylation reactions, and maintaining redox balance, thereby further contributing to the development of oral mucosal carcinogenesis [55–57]. Glycine, a semi-essential amino acid, not only serves as a major carbon source in the one-carbon metabolism pathway of pyrimidine and purine biosynthesis but also participates in the synthesis of glutathione (GSH) to maintain redox balance, significantly impacting cellular physiological states [58]. Serine, a non-essential amino acid, is vital for tumor growth. The de novo serine synthesis pathway (SSP) begins with 3-phosphoglycerate (3PG) produced during glycolysis and is completed through the catalytic actions of phosphoglycerate dehydrogenase (PHGDH), phosphoserine aminotransferase (PSAT), and phosphoserine phosphatase (PSPH) [59, 60]. As a rate-limiting enzyme, PHGDH is highly expressed in tumor cells to counteract the serine-limited environment. Conversely, the E3 ubiquitin ligase RNF5 inhibits tumor progression by mediating the degradation of PHGDH. Under conditions of ample serine supply, serine palmitoyltransferase (SPT) catalyzes the de novo synthesis of sphingolipids. However, in a serine-restricted environment, SPT switches to using alanine as a substrate to synthesize cytotoxic deoxysphingolipids, thereby exerting an inhibitory effect on tumors [61].

Fig. 12.

Correlation structure among amino-acid metabolite features. Circle color encodes the correlation coefficient, and circle size scales with correlation magnitude. The diagonal indicates self-correlation

Furthermore, leucine and isoleucine also exhibit a positive correlation (Fig. 12), which may be attributed to their similar metabolic pathways in the carcinogenesis of oral mucosa. Both leucine and isoleucine are classified as branched-chain amino acids (BCAAs) and essential amino acids (EAAs), and they are closely involved in the metabolic processes of both normal and tumor cells [21]. BCAAs serve as critical nitrogen sources for the synthesis of macromolecules and are essential for the survival of tumor cells [62]. BCAAs influence protein synthesis either by signaling the nutritional status of the cell or as constituent amino acids for protein construction [41]. The accumulation of BCAAs primarily promotes tumor growth by activating mammalian targets of rapamycin complex 1 (mTORC1) [62]. Specifically, leucine plays a significant role in new protein translation, making it crucial for protein synthesis [63]. The catabolism of BCAAs and their associated enzymes are closely linked to tumorigenesis. This catabolic process is mediated by branched-chain amino acid transaminase 2 (BCAT2), resulting in the production of branched-chain -keto acids (BCKAs), including -ketoisocaproate (-KIC), -keto--methylvalerate (-KMV), and -ketoisovalerate (-KIV). Subsequently, these BCKAs, such as -KIC, can be further metabolized into acetyl-CoA, while -KIV can be transformed into succinyl-CoA. As for -KMV, it can be further metabolized into both acetyl-CoA and succinyl-CoA. These metabolic products actively participate in the tricarboxylic acid (TCA) cycle. Therefore, the breakdown of BCAAs is crucial for the development and progression of OSCC [62]. Given the close association between BCAAs and tumors, alterations in BCAA levels can aid in the early diagnosis of OSCC [64].

The quantitative evaluation summarized in Tables 3 – 5 collectively supports the utility of MIIL for metabolomics-based diagnosis. The single-model versus MIIL ablations suggest that the gains arise from complementary decision cues rather than any single learner. These cues capture distinct facets of the metabolomic signal. This complementarity is particularly beneficial for OSCC diagnosis and premalignant screening in small, heterogeneous cohorts. Additionally, the strong performance across tasks is consistent with the clinical relevance of the amino-acid biomarkers prioritized from untargeted discovery. Comparisons against MIIL w/o feature selection and MIIL with vote-stable features further indicate that the vote-stable amino-acid subset contributes substantially to robustness.

Moreover, ROC analysis confirms strong separability for OSCC-oriented diagnosis, while early-lesion screening is intrinsically harder. We therefore report sensitivity at clinically relevant specificity to align evaluation with practice. In diagnostic tasks (CA.vs.M1, CA.vs.M2), specificity can be set high to suppress false positives. Sensitivity remains close to 90.00% across most specificity targets, supporting confident OSCC identification. In premalignant screening task (M1.vs.M2), the priority shifts toward sensitivity to reduce missed premalignant cases. Performance declines under stringent specificity constraints, which is expected. When specificity is relaxed to 60.00%, sensitivity increases markedly (93.33%).

This study has several limitations that should be considered when interpreting the findings. First, the cohort size is moderate and data collection was confined to a single center. This setting may restrict generalizability across institutions with different case mixes and perioperative practices. Second, the paired CA/M1/M2 design increases statistical efficiency and reduces inter-individual variability. However, it does not substitute for independent external validation. Third, the stage composition differed modestly between the two independently collected cohorts. Several clinical covariates were also recorded inconsistently across cohorts. These factors may introduce residual confounding and affect model transportability. Future work should prioritize larger, multi-center cohorts with harmonized protocols and richer clinical annotation. Stratified analyses and covariate-adjusted modeling will further clarify robustness and clinical applicability.

Conclusion and future scope

In summary, this study presents a metabolomics-based integrated information learning model that effectively integrates metabolomics analysis techniques, statistical methods, and artificial intelligence algorithms enabling comprehensive analysis and supporting OSCC diagnosis and premalignant screening. We first performed LC-MS/MS-based untargeted metabolomics to extract metabolomic features. Subsequently, statistical techniques are exploited to identify differential metabolites, with amino acids being identified as the most effective metabolite category for OSCC diagnosis. Moreover, targeted metabolomics using MRM is applied to specifically extract these key metabolites. Finally, an ensemble learning AI diagnostic method is incorporated to achieve accurate OSCC diagnosis. Therefore, our MIIL algorithm effectively addresses limitations of traditional diagnostic methods and improves clinical diagnostic accuracy.

Given the modest sample size and limited external validation, these results should be regarded as preliminary. Larger, multi-center studies with harmonized clinical annotation and stratified evaluation are warranted in future studies to confirm the metabolic signatures and substantiate the clinical performance of MIIL.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

WY, and XH conceived the study. WY, SL, and XM developed the theory and the code, WY, JR, and SH performed the analysis and wrote the manuscript, with support and supervision from LQ and XH. All the authors approved the final manuscript.

Funding

This work was supported by Beijing Nova Program (20240484547), Beijing Natural Science Foundation (L242131), and Beijing Hospitals Authority’s Ascent Plan (DFL20241501).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of Beijing Stomatological Hospital, Capital Medical University (Approval No. CMUSH-IRB-KJ-PJ-2021-18). Trial registration: Chinese Clinical Trial Registry (ChiCTR), ChiCTR2200064861; registered on 2023-04-23. Written informed consent was obtained from all patients prior to participation, in strict accordance with the Declaration of Helsinki to ensure the protection of patient rights, interests, and privacy. When a patient was unable to provide informed consent independently, consent was obtained from their legally authorized representative. In such cases, the study procedures were also explained to the participant to the extent of their comprehension, and written informed consent was signed jointly by the legal guardian and the investigator responsible for the consent discussion, ensuring that the patient’s will was not contravened.

Consent for publication

Written informed consent for publication of anonymized clinical data and images was obtained from all participants or their legal guardians. All identifying information has been removed to ensure confidentiality.

Conflict of interests

The authors declare no conflict of interest.