Metabolomics: Serum Panel Spots Early Head and Neck Cancer
Most head and neck squamous cell carcinomas (HNSCC) are diagnosed after local spread, and circulating tumor DNA picks up only 20 to 50% of early-stage cases. A blood test that catches Stage I disease would change that math. Writing in eBioMedicine, a Shanghai Jiao Tong University team ran serum metabolomics on 938 people at three hospitals, trained an eight-metabolite classifier on one cohort, and carried it unchanged to two others, one of them measured with a separate isotope-labeled assay.
Key Takeaways
Serum metabolomics separated head and neck cancer from healthy controls at three Chinese hospitals, including patients with Stage I tumors.
A head and neck cancer metabolite panel stayed accurate after switching from relative LC-MS/MS peak areas to absolute isotope-labeled measurements, a step many panels skip.
Oral premalignant lesions scored close to cancer, so a serum metabolite test for head and neck cancer fits pre-biopsy triage, not definitive diagnosis.
Three Hospitals, Two Metabolomics Assays
Discovery used 248 treatment-naive patients and 250 controls from Shanghai, with fasting serum drawn before surgery. Zhengzhou supplied 90 patients and 81 controls for external validation. Beijing added 119 patients, 120 controls, and 30 people with oral potentially malignant disorders (OPMDs), lesions such as leukoplakia that can turn malignant. Scheduled multiple reaction monitoring (MRM) returned a polar metabolite panel of 183 compounds, which LASSO and recursive feature elimination cut to eight for a random forest. Beijing samples went instead through a targeted metabolomics assay with stable isotope-labeled internal standards, and the model was applied without retraining.
Key Findings
35 metabolites shifted in discovery serum: Cystine and ornithine rose, choline and trimethylamine N-oxide fell, and arginine biosynthesis topped the pathway list.
The model traveled well: Choline, arginine, cystine, glutamic acid, hypoxanthine, glycerophosphocholine, acetylcysteine, and p-aminobenzoic acid gave an area under the curve (AUC) of 0.990 internally and 0.901 in Zhengzhou (sensitivity 91.1%, specificity 80.2%).
Assay transfer held up: On isotope-labeled Beijing data the AUC was 0.980, and Stage I tumors reached 0.927 in Zhengzhou and 0.988 in Beijing.
Precancer is the weak point: HNSCC versus OPMDs gave an AUC of only 0.710, and at 5% prevalence the positive predictive value was 49.1%.
Figure 1. Study design. Panel A lists the Shanghai discovery cohort (248 HNSCC, 250 controls), the Zhengzhou external cohort (90 HNSCC, 81 controls), and the Beijing disease-control cohort (119 HNSCC, 120 controls, 30 OPMDs). Profiling by LC-MS/MS, differential metabolite analysis, and clustering across Stages I to IV appear in panel B. Panel C traces the machine learning pipeline behind the eight-metabolite model, while panel D intersects HMDB annotations with TCGA transcriptomes to build a 16-gene panel defining C1 and C2 subtypes. Adapted from Ma et al. (2026), eBioMedicine.
Arginine Leaves the Blood and Stays in the Tumor
Serum arginine fell to 0.65-fold of control while ornithine climbed, the pattern expected when arginase converts one into the other, and CPTAC tumor proteomics showed arginase-1 up and the synthesis enzymes ASS1 and ASL down. Then the direction flipped. In 61 paired tissue samples, arginine, choline, and glycerophosphocholine were higher in tumor than in adjacent tissue, which the authors read as local retention draining the circulation. Geiger and colleagues (2016) showed that arginine availability governs T cell survival, so the idea carries immunological weight. With tissue from a single center, it remains a hypothesis.
A Transcriptomic Bridge Built From Databases
The group intersected 383 metabolite-associated genes from the Human Metabolome Database with 1,532 genes upregulated in TCGA head and neck RNA-seq data, ending at a 16-gene panel. Clustering split 504 tumors into C1 and C2. C2 carried more TP53 mutations, fewer CD8+ T cells, and worse survival (P = 0.042), and the C2-like cell line SCC-25 was roughly 59-fold more sensitive to gemcitabine than the C1-like CAL-27. None of these transcripts came from the patients who gave serum. Two cell lines cannot carry a treatment claim.
From Cohort Z-Scores to a Single-Patient Result
The quiet hinge here is normalization: each cohort was Z-scored against its own mean, which let a model trained on relative MRM peak areas run unchanged on isotope-dilution concentrations. When we move a discovery panel into targeted serum metabolomics at Dalton, we anchor scores to a fixed pooled reference serum instead, because a score that shifts with a batch's case mix can't be issued for one patient. That can cost some apparent accuracy, but it's the version of multi-omics biomarker discovery that survives real screening.
Frequently Asked Questions
What is serum metabolomics used for in cancer diagnostics?
Serum metabolomics measures hundreds of small molecules, such as amino acids and nucleosides, in a blood sample. In oncology it is used to find metabolite panels that flag cancer and to show how tumors rewire nutrient use.
Can a blood test detect early-stage head and neck cancer?
Not yet in routine care. An eight-metabolite serum panel reached an AUC of 0.927 for Stage I disease at an independent hospital, but the design was retrospective and case-control, which tends to overstate accuracy.
How much serum and instrument time does a targeted metabolomics panel need?
Very little sample. The validated assay here used 30 microliters of serum and about 20 minutes of LC-MS/MS time per sample. Most of a program's cost sits in recruiting independent validation cohorts, not in instrument time.
Conclusion
A serum signature centered on arginine turnover, holding across three hospitals and two quantification methods, is now credible for HNSCC. What nobody has shown is that it separates cancer from the inflammatory and premalignant oral disease that crowds a real head and neck clinic, or that it works outside Chinese cohorts. If you're building something similar, freeze the normalization reference before the first external sample arrives; retrofitting it later means revalidating the model.
Related Reading
See how we run these analyses in one lab: Dalton's multi-omics CRO services.
Citation
Ma, H., Zhao, X., Li, Z., Zhang, Z., Pan, X., Du, M., Wan, Q., Zhang, C., Wang, X., Zhao, H., Feng, Z., Zhang, X., & Chen, W.-T. (2026). Serum metabolomics profiling in head and neck squamous cell carcinoma: A multicentre diagnostic study. eBioMedicine, 130, 106379. https://doi.org/10.1016/j.ebiom.2026.106379
Note
This blog post summarizes findings from the above-cited research. Figures are adapted from the original publication. For full details, please refer to the source article.
By Seungjun Yeo, CEO at Dalton Bioanalytics. Specializing in multi-omics mass spectrometry for drug discovery and biomarker research.