Plasma acylcarnitine dysregulation associated with CPT1-mediated metabolism contributes to oral carcinogenesis

Abstract

Growing evidence suggests that lipid metabolic reprogramming occurs in oral cancer (OC). We characterized circulating metabolic alterations associated with lipid metabolic reprogramming in OC. Semi-targeted and targeted plasma metabolomic profiling was performed in a discovery cohort (182 OC, 364 healthy controls [HC]) followed by independent external validation (52 OC, 52 HC). Machine learning approaches were used to identify plasma metabolites with strong discriminatory performance between patients with OC and healthy controls. Targeted validation confirmed consistent metabolic changes. Namely, three medium-chain acylcarnitines—decanoyl-, octanoyl-, and hexanoylcarnitine—were markedly downregulated in OC plasma, demonstrating consistent and reproducible discrimination during external validation (AUC = 0.941, 95% CI: 0.877–0.988). Pathway enrichment analysis further suggested altered β-oxidation and glycerophospholipid metabolism in OC. We also observed elevated carnitine palmitoyltransferase 1 (CPT1) expression in OC tissues and cells. Moreover, pharmacological inhibition of CPT1 suppressed OC cell growth and altered acylcarnitine profiles. These findings support an association between circulating acylcarnitine alterations and CPT1-related lipid metabolic reprogramming in OC. Furthermore, they provide biological context supporting the association between circulating metabolic alterations and CPT1-related lipid metabolic reprogramming in OC.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-52309-w.

Introduction

Oral cancer (OC) is one of the most common malignancies, with 377,713 new cases of lip and oral cavity cancer and 177,757 deaths being reported in 2020 by the GLOBOCAN project¹. The 5-year survival rate for OC ranges from 15% to 50% for patients with stages III and IV disease, but exceeds 80% if early diagnosed (stages I or II)^2,3. Indeed, early detection of OC markedly improves survival and quality of life, as is associated with a reduced need of invasive/aggressive treatments and better response to therapy. Currently, diagnosis relies on multiple biopsies and histopathological assessments, which can be burdensome for patients⁴. Therefore, identifying non-invasive biomarkers of OC is crucial for improving diagnosis and prognosis.

Progression of oral squamous cell carcinoma, which accounts for over 90% of all OC cases⁵, reportedly occurs under conditions of limited oxygen and nutrient supply⁶. Additionally, lipid reprogramming was reported to support tumor growth by increasing fatty acid uptake and synthesis, while also altering lipid storage⁷. During carcinogenesis, creatine phosphokinase levels decrease and fatty acids are diverted towards the rapid proliferation of cancer cells via fatty acid oxidation (FAO)^6,8. Carnitine palmitoyltransferase (CPT), a key enzyme regulating FAO, was also reported to be upregulated in cancer cells exposed to adipocytes or fatty acids⁹. Of note, dysregulation of acylcarnitine metabolism, which is mediated by CPT1 at the outer mitochondrial membrane¹⁰, was associated with the progression of various cancers, including gastric, colorectal, and lung adenocarcinoma¹¹. Despite its importance, only three studies have investigated lipid-related carcinogenesis in OC using non-invasive methods^12–14. Levels of glycerophospholipids, particularly phosphatidylcholine (PC) and phosphatidylethanolamine (PE), are reportedly reduced, whereas sphingolipid levels are elevated in the sera and plasma of patients with OC compared with those in the controls^12–14. However, these studies lack validation sets and focus solely on OC marker detection instead of metabolism or function, thereby limiting their clinical relevance.

Metabolites reflect physiological and pathological states, offering valuable insights into tumorigenesis¹⁵. In the early stages of OC, systemic metabolic reprogramming and dysregulation of lipid metabolism in cancer cells lead to the release of specific metabolites into the bloodstream^16,17, suggesting that plasma metabolites can be used as proxies for assessing local tumor activity. Machine learning can leverage clinical omics data to build predictive models, facilitating the identification of discriminative metabolic features and potentially relevant disease mechanisms¹⁸. This study aimed to identify, quantify, and validate lipid and non-lipid metabolites in plasma associated with OC using multiple complementary approaches, including machine learning-based analyzes and targeted metabolomics. Understanding the metabolic context of these circulating metabolites may inform future translational and clinical applications.

Methods

Study participants and sample collection

This study included a discovery cohort and an independent external validation cohort. The discovery cohort comprised 182 patients with OC and 364 healthy controls (HCs) selected from the National Cancer Center (NCC) Bio Bank in a 1:2 case-to-control ratio matched for age and sex. OC cases were newly diagnosed and involved diverse anatomical subsites within the oral cavity, including the tongue, floor of the mouth, buccal mucosa, gingiva, palate, retromolar region, alveolar ridge, and lips.

The external validation dataset comprised 52 patients with OC and 52 HCs. Patients with OC were recruited from the Department of Oral and Maxillofacial Surgery at Seoul National University Dental Hospital and had pathologically confirmed squamous cell carcinoma. HCs were recruited in a 1:1 ratio matched for sex from cancer-free individuals visiting the health screening center or outpatient clinic at the National Cancer Center.

In both cohorts, only adults aged < 19 years of age were eligible for study inclusion. Additionally, all HCs were cancer-free at enrolment. In the validation cohort, patients with prior or synchronous cancers were excluded. To reduce potential selection bias, controls were selected using predefined matching criteria and eligibility requirements and were recruited during the same enrolment period as the corresponding case groups.

Demographic data, including sex, were self-reported by participants via questionnaires. All participants underwent overnight fasting before blood sampling. Venous blood was collected in K2 EDTA tubes (BD Vacutainer; BD Biosciences) and plasma was separated by centrifugation at 845 ×g for 20 min at 4 °C and preserved at − 80 °C for metabolomic analysis. The study followed the Declaration of Helsinki and was approved by the Institutional Review Board of the Seoul National University Hospital (No. NCC2016-0147) and of the National Cancer Center (No. NCC2018-0217, NCC2019-0116); written informed consent was obtained from all participants prior to their inclusion in the study.

Semi-targeted polar metabolomic profiling

This procedure was used to profile a broad panel of polar metabolites in plasma, including amino acids, acylcarnitines, and other small polar compounds. Semi-targeted metabolite analysis was performed using two ultra-performance liquid chromatography-triple quadrupole-mass spectrometry systems equipped with an electrospray ionization source (Agilent 1290 Infinity LC and 6490 Triple Quadrupole MS; Agilent Technologies). Data were acquired and analyzed using the MassHunter Workstation software (Ver B. 06.00; Agilent Technologies). Plasma samples (50 µL) were extracted with 500 µL chloroform: methanol (2:1 [v/v]) and 100 µL water, followed by centrifugation to precipitate proteins and separate the phases. Polar metabolites were vacuum-dried at room temperature and subsequently reconstituted in 200 µL of 20% methanol. A 1-µL aliquot was injected into the system and chromatographic separation was achieved using a Scherzo SM-C18 column (2 × 100 mm, 3 μm; Imtakt), which is suitable for the analysis of polar metabolites with diverse ionic properties. The protocol included a binary gradient at a flow rate of 0.2 mL/min over 20 min. The gradient comprised 0.1% formic acid in water (solvent A) and 0.1% formic acid in methanol (solvent B). Profiling was performed in dynamic multiple reaction monitoring mode. Quality controls, created by pooling equal amounts of all samples, were analyzed before and after every five samples to ensure stability and reproducibility.

Targeted polar metabolomic profiling

This targeted procedure was applied to quantitatively validate selected polar metabolite candidates identified from the semi-targeted discovery analysis. Plasma samples (20 µL) were extracted using 30 µL water and 150 µL acetonitrile, incubated at 4 °C for 1 h to facilitate protein precipitation, and centrifuged at 4,500 ×g and 4 °C for 10 min. The aqueous supernatant was then diluted at a 1:5 ratio (sample: solvent) using a solution containing 75% acetonitrile and 25% water before targeted profiling. Targeted profiling was performed using an ACQUITY UPLC coupled with an Xevo TQ-XS system equipped with an electrospray ionization source (Waters). The MassLynx software (V 4.2, Waters) was used for data acquisition and analysis. Liquid chromatography separations were performed for 10 min using the same analytical column and flow rates as during semi-targeted profiling. The mobile phase comprised 0.1% formic acid in water (solvent A) and 0.1% formic acid in acetonitrile (solvent B). Gradient elution began with 5% B for 2 min, which was increased to 100% B over 5 min, and was maintained at 100% B for 3 min. The chromatography column was equilibrated with 5% B for 1.5 min. Quantification was performed in multiple reaction monitoring mode.

Lipidomic profiling of non-polar metabolites

This extraction procedure was used to analyze non-polar lipid species in plasma, including triglycerides, phosphatidylcholines, lysophosphatidylcholines, diglycerides, ceramides, and free fatty acids, in positive and negative ionisation modes. Plasma samples (20 µL) were extracted using 30 µL water and 250 µL Splash Lipidomix (Avanti Research) in isopropyl alcohol (1:49 [v/v]). The samples were mixed for 10 min and stored at 4 °C for 2 h to facilitate protein precipitation. After centrifugation at 4,500 ×g and 4 °C for 10 min, the lipid supernatant was transferred to an analytical plate. Lipid profiling was performed using an ACQUITY UPLC coupled with an Xevo G2-XS Q-TOF MS system equipped with an electrospray ionization source (Waters). Data acquisition and analysis were performed using the MassLynx software (V 4.2). Chromatographic separation was performed using an ACQUITY UPLC CSH C18 column (2.1 × 100 mm, 1.7 μm, Waters) at 55 °C and 0.4 mL/min flow rate. The mobile phase in positive mode was 10 mM ammonium formate and 0.1% formic acid in water: acetonitrile (40:60 [v/v], solvent A) and 0.1% formic acid in isopropyl alcohol: acetonitrile (90:10 v/v, solvent B). In the negative mode, the mobile phase was 10 mM ammonium formate in water: acetonitrile (60:40 [v/v], solvent A) and isopropyl alcohol: acetonitrile (90:10 [v/v], solvent B). Samples were separated at a flow rate of 0.4 mL/min for 20 min. Leucine enkephalin was used as reference in the lock-spray technique for accurate mass measurement.

Metabolomics data processing and analysis

Metabolomic data were processed and analyzed using MetaboAnalyst 5.0 (https://www.metaboanalyst.ca/). Missing values were imputed using a limit-of-detection-based approach, defined as one-fifth of the minimum peak intensity, which was applied to account for low-abundance features below the detection threshold. Differential metabolites were selected based on a false discovery rate (FDR)-adjusted P-value of < 0.05 and a variable importance in projection score of > 1.0. Heatmaps and volcano plots were created for differentially expressed metabolites. receiver operating characteristics (ROC) curve analysis was used to identify potential OC-related markers and assess their performance. The area under the ROC curve (AUC) was used to determine marker discriminatory effectiveness, with an AUC > 0.80 indicating high sensitivity and specificity. Pathway analysis was performed using an FDR cut-off of P < 0.05 in MetaboAnalyst to visualize major metabolic networks.

Gene expression analysis in OC

Expression of CPT-family genes (CPT1A, CPT1B, CPT1C, and CPT2) and carnitine-acylcarnitine translocase (CACT) in human tissues was analyzed using Gene Expression Omnibus (GEO) microarray datasets (oral squamous cell carcinoma, SLC25A20 [GDS1584], and squamous cell carcinoma of the tongue and matched normal margins [GDS4562]) publicly available from the GEO database (https://www.ncbi.nlm.nih.gov/geo/). Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) was used as internal control to calculate relative expression patterns.

Measurement of CPT1 levels

CPT1 content was assessed using a commercially available enzyme-linked immunosorbent assay kit (mbs724213; MybioSource). Briefly, diluted (1:5) plasma samples were added to 96-well plates along with the standards. After adding the conjugate, the plates were incubated at 37 °C for 1 h, then washed 5 times. Next, 50 µL of substrates A and B were added to each well, followed by incubation in the dark for 10 min. The reaction was terminated by adding 50 µL of stop solution and absorbance at 450 nm was measured using a UV/vis spectrometer (SPECTROstar Nano; BMG Labtech). CPT1 content was determined by comparing sample absorbance values to the standard curve at 450 nm.

Cell lines

Human normal gingival fibroblast (HGF-1) and oral squamous cell carcinoma (YD-10B, CAL27, and SCC1) cell lines were provided by Dr. Yun-Hee Kim and Yuh Soeg Jung (National Cancer Center, Gyeonggi-do, Republic of Korea). HGF-1, CAL27, and SCC1 cells were maintained at 5% CO₂ and 37 °C in Dulbecco’s modified Eagle’s medium supplemented with 10% fetal bovine serum (Hyclone), 100 U/mL penicillin, and 100 µg/mL streptomycin (Gibco). YD-10B cells were cultured at 5% CO₂ and 37 °C in Roswell Park Memorial Institute-1640 medium supplemented with 10% fetal bovine serum, 100 U/mL penicillin, and 100 µg/mL streptomycin.

Immunohistochemistry

The Institutional Review Board (No. NCC2018-0217) approved the use of OC tissues prepared from paraffin-embedded blocks of six tumor samples. Core biopsies (2 mm in diameter) were obtained from the tumor area of these blocks. Tissue section (3-µm thick) were dried at 56 °C for 1 h. Staining was performed using an automated Discovery XT instrument (Ventana Medical Systems) with the Chromomap DAB detection kit (Roche). The tissue sections were deparaffinized, rehydrated, and washed with reaction buffer. Antigen retrieval was performed by heating in citrate buffer (pH 6.0) at 95 °C for 20 min. The primary antibody against CPT1 (dilution 1:200; 15184-1-AP, Proteintech) was added and incubated at room temperature for 32 min, followed by incubation with the UltraMap anti-rabbit horse radish peroxidase-conjugated secondary antibody (Roche Diagnostics) for 16 min. Images were obtained using Vectra Polaris (PerkinElmer) and analyzed using the inForm software (PerkinElmer). CPT1 expression was quantified using H-score [1 × (% cells 1+) + 2 × (% cells 2+) + 3 × (% cells 3+)], yielding a score range of 0–300¹⁹.

Cell viability

HGF-1 (4 × 10⁴ cells/mL), YD-10B, CAL27, and SCC1 (3 × 10⁴ cells/mL) were cultured in 96-well plates. After 24 h, the cells were incubated with 200 µM etomoxir (236020, Sigma-Aldrich) or dimethyl sulfoxide for 48 h. 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (Duchefa Biochemie) solution was added to each well to a final concentration of 5 mg/mL and the mixture was incubated at 37 °C for 6 h. Formazan pellets were dissolved in 2-propanol (67-63-0, Merck) and absorbance was measured using a VERSA Max microplate reader (Molecular Devices) at 540 and 650 nm (reference wavelength).

Metabolites analysis by machine learning

A case-control design was used to analyze blood metabolites in patients with OC via machine learning. Machine learning analyses were implemented in Python (version 3.9.12). In the large discovery set, eXtreme Gradient Boosting, Gradient Boosting Machine (GBM), LightGBM, and Random Forest were applied for data mining and to identify key markers from among the 48 differently expressed metabolites identified in the semi-targeted discovery analysis. Multiple machine learning models were compared during the discovery stage to prioritize robust candidate metabolites. Model performance was evaluated using five-fold cross-validation within the discovery set. This cross-validation was selected to balance model stability and sample size considerations. The best-performing model for OC diagnosis was developed using LightGBM combined with a gradient-based one-sided sampling algorithm. LightGBM was chosen due to its superior ability to efficiently handle large datasets and perform feature selection²⁰. The top 20 significant metabolites were selected based on feature importance. A Random Forest model was then used to evaluate the selected markers in the discovery and validation sets. Random Forest outperformed other models, such as support vector machine and logistic regression, in terms of accuracy and overall performance metrics²¹, making it the optimal choice for evaluating the selected markers. Model performance was assessed using five key metrics: accuracy, AUC, F1 score, sensitivity, and specificity.

Statistical analysis

Statistical analyzes, including Chi-square, Student’s t, and Fisher’s exact tests, and multivariate logistic regression, were performed in Python (version 3.9.12). Results are presented as the mean ± standard deviation or n (%), with significance set at P < 0.05. Age, sex, smoking, and alcohol consumption were included as covariates a priori to reduce residual confounding, given their established relevance to OC risk. False discovery rate (FDR) adjustment was applied during differential metabolite selection and pathway analysis in MetaboAnalyst, as described above.

Results

Clinical and diagnostic characteristics of the study population

The characteristics of the discovery and external validation datasets, comprising a total of 650 participants, are shown in Fig. 1A. The average age range of both groups was 60–69 years, and approximately 65% of the participants were male (Table S1). The proportion of current smokers and drinkers did not differ significantly between the two datasets. The discovery dataset had a higher frequency of T2 stage and tongue-site occurrences. In contrast, the validation dataset was characterized by a high prevalence of T1 and T2 tumors, with the alveolar ridge being the most common site.

Fig. 1.

Data processing and identification of differential plasma metabolites. (A) Overall workflow of the metabolomics study. (B) Heatmap showing relative abundances of 48 differential metabolites in healthy controls (HC) and patients with oral cancer (OC). (C) Volcano plot comparing HC and OC. Each point represents an individual metabolite. Significant metabolites were defined using a fold change (FC) threshold of 1.0 on the x-axis and a false discovery rate (FDR)-adjusted P value < 0.05 on the y-axis. Negative log₂ FC values (blue) indicate lower metabolite levels in OC, whereas positive values (red) indicate higher levels in OC. (D) Multivariable logistic regression analysis showing odds ratios (ORs) and 95% confidence intervals for associations between metabolites and OC in the discovery semi-targeted dataset, adjusted for age, sex, smoking, and alcohol consumption. (E) Corresponding analysis in the discovery targeted dataset. See also Tables S2 and S3.

Semi-targeted profiling of polar candidate metabolites

Metabolomic analysis of 182 OC and 364 HC samples from the discovery dataset revealed 96 annotated metabolites, of which 48 were differentially abundant (Fig. 1B; Table S2). Table S3 summarizes the metabolites and lipid species analyzed by each workflow. The volcano plot in Fig. 1C reveals significantly lower levels of octanoylcarnitine, hexanoylcarnitine, and decanoylcarnitine, but higher levels of hypoxanthine in OC than in HC samples. After adjusting for sex, age, smoking, and drinking, odds ratios (ORs) were calculated (Fig. 1D; Table S4). Eighteen metabolites showed significant differences in logistic regression analysis. Higher tertiles of hypoxanthine and pyridoxamine were linked to increased OC risk (ORs: 33.5 and 8.91, respectively), whereas higher octanoylcarnitine and decanoylcarnitine levels were associated with decreased OC risk (OR: 0.01).

Identification of potential OC diagnostic markers

To prioritize key metabolites for OC detection, four machine learning models were compared using the 48 differentially expressed metabolites identified in the semi-targeted discovery analysis. Among these models, LightGBM showed the best performance in the discovery set (Table S5), with an AUC of 0.9952 (Fig. 2A, B), sensitivity of 0.9782 ± 0.0081, and specificity of 0.9853 ± 0.0104. Using the LightGBM model, the top 20 significant metabolites, ranked in decreasing order of importance, were glycerophosphocholine, propionylcarnitine, taurine, and glutamate (Fig. 2E). Predictive odds and SHapley Additive exPlanations (SHAP) values were analyzed across 5-fold cross-validation (Fig. 2C, D), showing stable classification performance and consistent contributions of key metabolites. ROC analysis was also conducted to further evaluate whether the 48 metabolites could distinguish between OC and HC (Fig. S1, Table S2). Eleven metabolites with AUC > 0.75 were identified (Fig. S1A–K) and combinations of these metabolites were also assessed (Fig. S1L–N). The model distinguished HC from OC with 94.4% accuracy (Fig. S1L), which increased to 97.0% upon multivariate ROC analysis of the fourth model, which combined 10 metabolites (Fig. S1M). The top 10 metabolites matched individual biomarkers (Fig. S1N). Based on previous analyzes (variable importance in projection, FDR P-value, logistic regression, machine learning, and AUC), 7 candidate markers—octanoylcarnitine, decanoylcarnitine, hexanoylcarnitine, glycerophosphocholine, acetylcarnitine, taurine, and hypoxanthine— were selected for further analysis.

Fig. 2.

Machine learning-based prioritization of candidate metabolites in the discovery set. (A–E) Semi-targeted metabolite analysis and (F–J) targeted metabolite analysis. (A, F) Partial dependence (POD) plots. (B, G) Receiver operating characteristic (ROC) curves generated using Light Gradient Boosting Machine (LightGBM). (C, H) SHapley Additive exPlanations (SHAP) plots from a representative fold. (D, I) Performance metrics across five-fold cross-validation, including AUC, accuracy, sensitivity, specificity, F1 score, recall, and precision. (E, J) Feature importance plots showing the top-ranked metabolites according to model contribution. LightGBM, light gradient-boosted machine.

Targeted validation of selected polar metabolites

Consistent with the results of semi-targeted analysis, octanoylcarnitine, decanoylcarnitine, hexanoylcarnitine, glycerophosphocholine, and acetylcarnitine levels were lower in OC than in HC samples, whereas taurine and hypoxanthine levels were higher (Fig. S2A–G, Table S2). Multivariate logistic regression confirmed these findings, aligning with the ORs obtained from the semi-targeted analysis (Fig. 1E; Table S4). The LightGBM model showed strong OC/HC discrimination for these seven metabolites, with AUC = 0.97 (sensitivity: 0.9273, specificity: 0.9651) (Fig. 2F–I). Glycerophosphocholine, hypoxanthine, taurine, and acylcarnitines were identified as the most important metabolites in order of significance (Fig. 2J). Nevertheless, individual AUCs for octanoylcarnitine, hexanoylcarnitine, and decanoylcarnitine were > 0.8, supporting their selection as candidates for subsequent validation (Fig. 3A–C; Table S2).

Fig. 3.

Validation of three candidate acylcarnitine biomarkers across discovery and external validation cohorts. Individual metabolite performance for (A) decanoylcarnitine, (B) octanoylcarnitine, and (C) hexanoylcarnitine. Multivariable ROC curves based on the combined three-metabolite model in the (D) discovery cohort and (E) external validation cohort. See also Table S2.

To validate the predictive performance of octanoylcarnitine, hexanoylcarnitine, and decanoylcarnitine, an external validation dataset of 52 OC and 52 HC samples was analyzed. As in the discovery set, the levels of these metabolites were lower in OC than in HC (Fig. S2H–J), with individual AUCs > 0.8 (Fig. 3A–C). Multivariate ROC analysis using a Random Forest model showed high predictive accuracies of 0.946 in the discovery set and 0.941 in the independent external validation set, supporting their potential as diagnostic markers for OC (Fig. 3D, E). This external validation cohort was used for targeted validation of predefined candidate metabolites, not for repeated broad metabolomic or lipidomic profiling.

Global lipidomic profiling and pathway analysis

As three acylcarnitines were identified as potential biomarkers of OC, a comprehensive lipid profiling was performed. Increase levels of 11 triglycerides and decreased levels of 4 triglycerides, 3 diglycerides, 12 free fatty acids, 5 PCs, and 5 lysophosphatidylcholines (LysoPCs) were detected in the OC samples as compared with HCs (Fig. 4A, B). Notably, 20:4 and 20:5 free fatty acids were significantly downregulated. Enrichment analysis revealed 39 pathways involved primarily in amino acid and lipid metabolism, including glycerophospholipid metabolism and mitochondrial fatty acid β-oxidation (Fig. 4C, D; Table S6). Glycerophospholipid metabolism, which connects choline, acetylcholine, and glycerophosphocholine to acylcarnitine (FDR P = 1.43 × 10^− 32, pathway impact = 0.1858), was of particular interest (Fig. 4E).

Fig. 4.

Global lipidomic profiling and pathway enrichment analysis in oral cancer. (A) Workflow of global lipidomic profiling. (B) Volcano plot comparing lipid metabolites between HC and OC. Each point represents one lipid species. Significant lipids were defined using an FC threshold of 1.0 and an FDR-adjusted P value < 0.05. Blue points indicate lower lipid levels in OC, whereas red points indicate higher levels. (C) Overview of pathway analysis performed using MetaboAnalyst. Node color represents the P value, and node size reflects pathway impact. (D) Pathway enrichment results. (E) Key pathways related to glycerophospholipid metabolism and mitochondrial fatty acid β-oxidation. See also Table S6.

Expression of the CPT family of genes in OC

The levels of acylcarnitine derivatives, which are crucial for fatty acid β-oxidation in cancer¹¹, were significantly altered in plasma samples from patients with OC. To explore whether these circulating changes were associated with CPT-related lipid metabolic features in tumors, we examined the expression profiles of CPT1A, CPT1B, CPT1C, CPT2, and CACT in oral squamous cell carcinoma (Fig. S3A–E). CPT1A was upregulated in OC, whereas CACT was downregulated. Enzyme-linked immunosorbent assay of 102 HC and 58 OC samples showed higher CPT1 levels in OC, although the difference was not statistically significant (P = 0.157) (Fig. S3F).

CPT1 expression in OC tissues and cells

Immunohistochemical analysis of samples from six patients showed higher CPT1 staining in OC tissues than in adjacent normal tissues, with an average H-score fold change of 1.72 (P < 0.001) (Fig. 5A–C). To investigate whether elevated CPT1 in OC was associated with increased acylcarnitine levels, the concentrations of short-, medium-, and long-chain acylcarnitines were compared between normal and cancer oral cells. Significantly higher acylcarnitine levels were detected in oral cancer cells (Fig. 5D–F). Moreover, pharmacological inhibition of CPT1 using etomoxir reduced OC cell viability by approximately 0.39-fold (P < 0.001) without affecting normal cells (Fig. 5G). Etomoxir treatment also increased short-chain acylcarnitine levels (C0, C2, and C6), whereas it decreased medium-chain (C8 and C10) and long-chain (> C12) acylcarnitine levels in OC cells (Fig. 5H). These results suggest that CPT1 is associated with OC cell growth and altered cellular metabolism of medium- and long-chain acylcarnitines in vitro.

Fig. 5.

CPT1 expression in oral cancer tissues and cells and effects of CPT1 inhibition on acylcarnitine profiles. (A) Representative immunohistochemical images of CPT1 expression in oral tumor tissues (100× magnification; scale bar, 100 μm). (B) Mean staining intensity and (C) quantified H-scores of CPT1 expression in tumor tissues. Error bars indicate mean ± SEM from three independent regions per sample. Comparison of normal and oral cancer cells showing normalized concentrations of (D) short-chain, (E) medium-chain, and (F) long-chain acylcarnitines. (G) Cell viability of HGF-1, YD-10B, CAL27, and SCC1 cells following treatment with vehicle (dimethyl sulfoxide) or 200 µM etomoxir for 48 h, assessed using the MTT assay. (H) Fold changes (etomoxir-treated vs. untreated) in cellular acylcarnitine levels. P < 0.05; **P < 0.001.

Discussion

Timely and accurate diagnosis of OC is critical, as delayed detection remains a major contributor to OC-related mortality. However, imaging and biopsies often fail to detect early-stage OC. Metabolomic profiling offers a non-invasive strategy to capture tumor-related metabolic alterations that may complement existing diagnostic approaches, particularly in individuals at increased risk of OC. Consistent with previous metabolomics studies^12–14, the present findings confirm systemic metabolic alterations associated with OC development. We identified a panel of plasma acylcarnitines (decanoyl-, octanoyl-, and hexanoylcarnitine) with strong OC discriminatory performance (AUC = 0.941). In addition, our lipidomic and functional data support an association between these circulating alterations and CPT1-related lipid metabolic reprogramming in OC.

Acylcarnitines transport fatty acids into the mitochondria, initiating their oxidation and producing acetyl-CoA^22,23. Hypoxia and acidosis are common in solid tumors, with acidosis-induced FAO being crucial for cancer progression¹¹. FAO also activates fatty acid synthase and adversely affects mitochondrial redox states, promoting cancer cell proliferation²³. Herein, the medium-chain acylcarnitines, hexanoylcarnitine (C6:0), octanoylcarnitine (C8:0), and decanoylcarnitine (C10:0), were found to be downregulated in patients with OC. Previous studies have reported low levels of short-chain fatty acids in OC plasma¹⁴ and highlighted the role of acylcarnitines in mitochondrial dysfunction and hepatocellular carcinoma²⁴. Moreover, downregulation of octanoylcarnitine has been identified as a highly predictive marker for breast cancer in a Korean study²⁵, whereas downregulation of decanoylcarnitine has been recognized as a promising diagnostic marker for renal cell carcinoma²⁶. Collectively, these results support the potential relevance of the abovementioned metabolites in OC development. They also suggest that perturbation of medium-chain acylcarnitine metabolism may represent a recurring feature across multiple cancer types, although the magnitude and direction of these alterations may vary depending on tumor type and biospecimen. However, the observed reduction in plasma medium-chain acylcarnitines should be interpreted with caution, as circulating acylcarnitine levels may also reflect systemic metabolic processes, including hepatic metabolism, inflammation, dietary influences, gut microbiome composition, altered metabolite transport, and other extra-tumoral factors.

In the present study, CPT1-related evidence in OC was supported by multiple complementary observations, including CPT1A upregulation in publicly available transcriptomic datasets, higher CPT1 staining in OC tissues, and altered acylcarnitine profiles in OC cells. These findings are consistent with previous reports linking CPT1 upregulation to altered cancer metabolism^9,27,28. These findings also align with those of studies showing CPT1 overexpression in lung²⁹, gastric³⁰, prostate³¹, ovarian³², and breast³³ cancers. Notably, previous studies have also implicated fatty acid oxidation and CPT1-related metabolic reprogramming in head and neck malignancies, including CPT1A-dependent fatty acid oxidation in nasopharyngeal carcinoma and broader lipid metabolism reprogramming in head and neck cancer, further supporting the relevance of our findings in the head and neck cancer context^27,28.

Increased CPT1 expression in OC tissues and altered acylcarnitine profiles in plasma and cells collectively support an association between CPT1-related lipid metabolic reprogramming and OC. Fatty acids entering cells are converted to acyl-CoA, which is essential for transport and β-oxidation³⁴. Acyl-CoA is subsequently converted to acylcarnitines via CPT1, which also facilitates the transport of long-chain fatty acids³⁵. In our previous study, CPT1 knockdown³⁶, as well as treatment of OC cells with the CPT1 inhibitor, etomoxir, significantly suppressed OC cell growth (YD-10B, CAL27, and SCC1) without affecting non-cancerous cells (HGF-1). Additionally, our findings revealed increased levels of short-, medium-, and long-chain acylcarnitines in OC cells compared with healthy cells. Etomoxir treatment reduced medium- and long-chain acylcarnitine levels while increasing short-chain acylcarnitine levels, consistent with inhibition of β-oxidation³⁷. However, as etomoxir has potential off-target effects on mitochondrial metabolism, these findings should be interpreted with caution. Thus, elevated CPT1 in OC may be associated with intracellular acylcarnitine accumulation and altered β-oxidation, suggesting a potential role in OC metabolic reprogramming.

Acylcarnitines transported by CPT1 are shuttled into the mitochondria via CACT and reconverted into acyl-CoA and carnitine by CPT2 in the mitochondrial inner membrane^34,38,39. The resulting acyl-CoA undergoes β-oxidation to generate acetyl-CoA, which enters the tricarboxylic acid cycle, acting as an immediate energy source^34,35. In mitochondria, dysregulated β-oxidation leads to excessive accumulation of acetyl-CoA, thereby contributing to tumor progression²³. These findings are consistent with an association between elevated CPT1 in OC and intracellular accumulation of medium- and long-chain acylcarnitines, which may reflect altered β-oxidation. This intracellular shift may be associated with the reduced levels of circulating medium-chain acylcarnitines observed in this study, consistent with enhanced intracellular fatty acid oxidation; however, the direct contribution of tumor CPT1 activity to circulating metabolite levels warrants further investigation. Although CPT2 and CACT levels were not assessed in OC tissues in this study, previous research has demonstrated the crucial role of CPT2 in cancer cell proliferation in epithelial ovarian, gastrointestinal, and triple-negative breast cancers, suggesting that similar mechanisms may underlie OC progression⁴⁰.

We further investigated lipid and non-lipid metabolic pathways related to OC development, focusing on acylcarnitine, glycerophospholipids, and FAO using global lipid profiling. Diacylglycerol, free fatty acids, PC, and LysoPC levels were lower in OC samples. The levels of 11 triglycerides were elevated in OC, whereas four triglycerides showed decreased levels. This finding aligns with that of Wang et al.., who reported reduced levels of PC, LysoPC, PE, and phosphatidylglycerol in OC¹², and of Yang et al.., who observed downregulation of LysoPC (18:3), LysoPC (20:4), LysoPE (20:3/0:0), and lysosphingomyelin (d18:1) levels¹³. LysoPC acyltransferase 1 and phospholipase A2 convert PC⁴¹, a major plasma membrane lipid, to LysoPC. This signaling molecule regulates cell proliferation, inflammation, and cancer invasion^42,43. Overexpression of LysoPC acyltransferase 1 promotes cancer growth by altering membrane PC levels⁴⁴. Dysregulation of glycerophosphodiester phosphodiesterase 5, which converts LysoPC to glycerophosphocholine, is linked to high PC and low glycerophosphocholine levels in aggressive tumors⁴⁵. Decreased glycerophosphocholine levels are also observed in breast and ovarian cancers⁴⁶. Inhibition of phospholipase domain-containing protein 8 disrupts phospholipid reprogramming in triple-negative breast cancer, affecting PC and glycerophosphocholine levels⁴⁷. LysoPC, reconverted from PC via choline, is converted into fatty acids that enter the cells⁴⁸, where acyl-CoA is synthesized by acyl-CoA synthase. It then combines with carnitine to form acylcarnitines via CPT1⁴⁹. Lower levels of plasma acylcarnitine and fatty acids, together with increased CPT1 levels, may reflect altered mitochondrial transport and β-oxidation associated with OC. A schematic overview integrating the main findings of this study with literature-based metabolic pathways is illustrated in Fig. 6.

Fig. 6.

Schematic overview of lipid metabolic reprogramming during oral cancer progression. Dashed arrows indicate literature-based pathways or inferred metabolic relationships that were not directly investigated in the present study.

This study has several strengths. We analyzed a relatively large, well-characterized cohort, including both discovery and external validation sets, to ensure robustness of our findings. The identified acylcarnitines could be detected in small plasma volumes, supporting a minimally invasive OC screening method. In addition, we integrated metabolite profiling with tissue- and cell-based mechanistic data, providing insights into CPT1-mediated lipid metabolism in OC.

However, some limitations should be noted. First, although we reported the upregulation of CPT1 in OC tissues and its roles in β-oxidation and acylcarnitine metabolism, downstream components, such as CPT2 and CACT, were not investigated. Second, the findings on acylcarnitine accumulation and CPT1 inhibition were based on in vitro OC cell line models. In addition, the functional role of CPT1 was evaluated using pharmacological inhibition with etomoxir, which has known off-target effects on mitochondrial metabolism. Therefore, further studies using genetic approaches, such as CPT1 knockdown or overexpression, are required to more specifically define its role in OC. Third, plasma metabolite alterations may reflect systemic metabolic changes rather than tumor-specific metabolism alone. Factors such as hepatic metabolism, inflammation, dietary influences, gut microbiome composition, altered metabolite transport, and other extra-tumoral processes may also contribute to circulating acylcarnitine levels. Fourth, this study did not include metabolic flux or secretion analyses to directly connect intracellular CPT1-related metabolism with circulating acylcarnitine changes. Fifth, nested cross-validation was not performed during model development, which should be considered when interpreting the performance of the discovery-stage model. Furthermore, although matched control selection and predefined eligibility criteria were applied, residual selection bias inherent to the case-control design cannot be fully excluded. Finally, immunohistochemistry for CPT1 in OC tissues was performed on a limited sample size; larger cohorts are needed to validate these findings. In addition, future studies should investigate whether de novo lipid synthesis, potentially driven by excess acetyl-CoA, contributes alongside fatty acid oxidation to metabolic reprogramming and tumor growth in OC. Future validation studies should also assess the specificity of the identified metabolic signature for OC relative to other tumor types. Addressing these aspects will provide a more comprehensive understanding of lipid metabolism in OC.

In conclusion, our findings highlight the diagnostic potential of plasma decanoylcarnitine, octanoylcarnitine, and hexanoylcarnitine, and support an association between CPT1 and lipid metabolic alterations in OC progression. These results support the use of circulating metabolites for non-invasive detection and suggest that targeting of metabolic pathways may provide avenues for therapeutic intervention in OC. Further studies are needed to investigate the in vivo roles of downstream metabolic components and better understand their contribution to oral carcinogenesis.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Yeon-Hee Kim: Conceptualization, Writing – Original draft, Formal analysis, Validation, Visualization, Data curation, Methodology, Software, Writing – review & editing. Sung Weon Choi: Investigation, Resources. Jong Ho Lee: Investigation. Joo Yong Park: Investigation. Heesun Cheong: Data curation. Mi Kyung Kim: Conceptualization, Supervision, Funding acquisition, Resources, Project administration, Writing – review & editing. All authors have read and approved the final manuscript.

Funding

This research was supported by the National Cancer Center, Republic of Korea (grant number 2510800). The funders had no role in the study design, data collection, analysis, interpretation, manuscript writing, or decision to submit for publication.

Data availability

The raw and processed metabolomics data supporting the findings of this study are publicly available in the MetaboLights repository (https://www.ebi.ac.uk/metabolights/MTBLS12655). All datasets have been fully de-identified to protect participant privacy and are in accordance with applicable ethical guidelines. Data supporting the figures and statistical analyses are provided within the main text.

Competing interests

The authors declare no competing interests.

Patent applications

This research has led to the registration and filing of the following patents: Biomarker composition containing acylcarnitine metabolite for diagnosis of oral cancer (Korean Patent No. 10-2627818-00-00, European Patent Application No. 22816433.1, and U.S. Patent Application No. 18/566,404).