Hepatocellular carcinoma (HCC) ranks among the malignancies with the highest mortality rates globally. Current clinical surveillance strategies primarily rely on abdominal ultrasound combined with alpha-fetoprotein (AFP) testing; however, the insufficient sensitivity of AFP often leads to the missed diagnosis of a significant number of patients in the early stages of the disease. In recent years, liquid biopsy technologies have flourished, with serum N-glycomics garnering particular attention due to its capacity to systematically reflect hepatic synthetic function.

In January 2026, a team led by Haojie Lu from Fudan University, in collaboration with other researchers, published a study in Nature Communications titled "Large-scale serum N-glycomics tracks N-glycosylation dynamics in hepatocellular carcinoma progression and enables early diagnosis." By analyzing serum N-glycomic data from 1,074 subjects—integrating glycoproteomics, transcriptomics, and machine learning—this study offers a novel perspective on the precise diagnosis and mechanistic investigation of HCC.

Fig. 1 Overview of study design and workflow. (Fu, et al. 2026)

Study Design and Cohort Size

The study enrolled a total of 1,074 subjects, spanning four distinct disease stages: healthy controls, chronic hepatitis B, liver cirrhosis, and hepatocellular carcinoma. The primary discovery cohort (GZ-I) comprised 744 samples, while two independent external validation cohorts—GZ-II (186 samples) and XZ (144 samples)—were established for verification purposes. The research team employed a high-throughput workflow involving N-glycan release, derivatization, HILIC enrichment, and MALDI-TOF mass spectrometry to systematically profile the serum N-glycome across all samples. Quality control data demonstrated a median coefficient of variation as low as 0.08, with median Pearson correlation coefficients reaching 0.98 both between batches and across cohorts, thereby ensuring the reliability of the data.

Within the GZ-I cohort, the study identified 201 distinct N-glycan compositions; following rigorous screening, 64 structurally defined glycans were selected for downstream analysis. Furthermore, 120 samples underwent additional LC-MS/MS-based glycoproteomic and proteomic profiling, while 186 samples were subjected to derivatization analysis specifically targeting sialic acid linkages.

The Serum N-Glycome: A Sensitive Window into Hepatic Functional Status

As the liver serves as the primary site for the synthesis of serum glycoproteins, alterations in the serum N-glycome can sensitively capture even subtle changes in hepatic function. This study systematically evaluated the associations between N-glycomic profiles and conventional liver function markers, as well as the ALBI score and Child-Pugh classification. The findings revealed:

• N-glycomic features exhibited significant correlations with the ALBI score, with Pearson correlation coefficients for certain glycans approaching 0.5.
• The deterioration of liver function was accompanied by a distinct shift in glycosylation patterns: increased branching, elevated bisecting GlcNAc levels, and enhanced fucosylation, alongside a concomitant decrease in galactosylation and sialylation.
• A random forest model utilizing N-glycomic data to predict ALBI stages achieved an average ROC AUC of 0.892, demonstrating a high degree of concordance between predicted probabilities and actual scores.

Subgroup analyses confirmed that these associations were particularly robust in patients with cirrhosis and HCC, suggesting that serum N-glycomics serves not merely as a disease biomarker, but also as an effective tool for assessing hepatic synthetic and metabolic functions.

Dynamic Remodeling of Glycosylation During HCC Progression

Differential analysis revealed that 48 out of 64 identified glycans were significantly dysregulated during disease progression, exhibiting distinct stage-specific patterns:

• Chronic Hepatitis Stage: Changes in the serum N-glycome were relatively mild.
• Cirrhosis Stage: This stage was characterized primarily by the downregulation of non-fucosylated tri-/tetra-antennary multisialylated glycans.
• Hepatocellular Carcinoma (HCC) Stage: This stage exhibited significant upregulation of fucosylated tri-/tetra-antennary multisialylated glycans; notably, this feature was specific to HCC and was not observed in cirrhosis.

The study clustered the 48 differentially expressed glycans into five Glycan Co-expression Modules (GCMs), each exhibiting a distinct expression trajectory:

• GCM1: Fucosylated, fully sialylated tri-/tetra-antennary glycans, specifically upregulated only in HCC.
• GCM2: The non-fucosylated counterparts of the glycans in GCM1, downregulated in cirrhosis.
• GCM3: Smaller glycans containing bisecting GlcNAc, elevated in both cirrhosis and HCC.
• GCM4: Two hybrid-type monosialylated glycans, generally upregulated across all liver disease stages.
• GCM5: High-mannose, hybrid, or immature complex glycans, downregulated only in HCC.

At the structural level, HCC patients presented a composite phenotype characterized by a decrease in bi-antennary glycans, an increase in tri-/tetra-antennary glycans, reduced galactosylation, and elevated levels of bisecting GlcNAc and fucosylation. Analysis of sialic acid linkages further revealed a specific reduction in α2,6-linked sialylation in patients with cirrhosis.

After controlling for liver function parameters through covariate analysis, 30 glycans remained significantly differentially expressed, demonstrating that these glycosylation alterations are not merely passive concomitants of liver dysfunction but are, rather, intimately linked to specific pathological processes.

N-Glycome Molecular Subtypes: Capturing Tumor Heterogeneity

Based on unsupervised consensus clustering, patients with hepatocellular carcinoma (HCC) were classified into three distinct N-glycome molecular subtypes, each characterized by unique clinical features:

• Subtype 1: This subtype was enriched in small, immature glycans within GCM3, showed generally lower levels of large, multisialylated, afucosylated glycans in GCM2, and was associated with relatively poor liver function.
• Subtype 2: This subtype exhibited reduced levels of large, multisialylated, fucosylated glycans in GCM1. Its glycomic profile most closely resembled that of non-HCC controls, suggesting it may represent an early-stage carcinoma phenotype with relatively preserved liver function.
• Subtype 3: This subtype was characterized by elevated levels of GCM1 glycans and reduced levels of GCM5 glycans, presenting a glycomic profile complementary to that of Subtype 1.

While Subtypes 1 and 3 exhibit no significant differences in terms of liver function or tumor stage, their diametrically opposed glycosylation patterns suggest the potential involvement of distinct pathogenic mechanisms. Patients in Subtype 2 are more frequently found in earlier pathological stages, thereby providing a potential basis for clinical risk stratification.

Site-Specific Glycoproteomics: Deciphering the Molecular Origins of the Serum Glycome

Serum glycomics reflects the collective signal of circulating glycoproteins; however, precisely which proteins—and which specific sites—contribute to disease-associated glycans? Through a data-dependent acquisition (DDA)-based glycoproteomic analysis of 120 samples, this study identified 3,057 glycopeptides, 2,824 site-specific glycans, 276 glycosylation sites, and 168 glycoproteins.

By constructing pseudo-glycome profiles (generated by aggregating the signals of identical glycan compositions across all identified sites and proteins), the study revealed that the average Pearson correlation coefficient between the pseudo-glycome and the actual serum glycome reached an impressive 0.95. This finding demonstrates that data at the site-specific level can effectively account for and explain observed serum glycome phenotypes. More importantly, different glycans exhibit highly heterogeneous site origins:

• H5N4S2 (one of the most abundant glycans in serum): This glycan showed uniform contributions from multiple sites.
• H5N5F1S1: This glycan was primarily derived from IGHM N46 and IGHA2 N92/N207.
• Tri-/tetra-antennary fucosylated multisialylated glycans: These glycans were primarily carried by the A1AG1 N93 site.
• Immature hybrid-type glycans: These glycans were primarily associated with IGHG1 N180/N299.

Differential analysis distinguished two driving mechanisms:

• Glycosylation-dominant mechanism: For instance, H5N3S1 is significantly elevated in chronic hepatitis B, primarily originating from the N869 site of the A2MG protein. Although total A2MG protein levels showed only a mild increase, the H5N3S1 glycoform at this specific site increased substantially, demonstrating that alterations in the regulation of glycosyltransferases themselves drove the observed changes in the glycome.
• Protein-dominant mechanism: For instance, the elevation of tetra-antennary fucosylated glycans—such as H7N6F1S2—in HCC is primarily driven by an increased abundance of the acute-phase response protein A1AG1.

This finding suggests that serum glycomics data must be interpreted in conjunction with changes in protein abundance to accurately decipher their underlying biological origins.

At the Transcriptomic Level: Global Activation of the N-Glycosylation Pathway in Tumor Tissues

Utilizing the TCGA-LIHC dataset and seven GEO datasets, this study systematically characterized the expression profiles of 307 glycogenes involved in the N-glycosylation pathway within HCC tumor tissues. Differential analysis revealed that 118 glycosylation-related genes were significantly upregulated, while only two were downregulated, indicating a state of global activation:

• Initiation Step: DPAGT1, members of the ALG family, and components of the oligosaccharyltransferase complex (e.g., STT3A, DDOST) were significantly upregulated, suggesting that the tumor enhances the efficiency of N-glycan precursor synthesis and transfer.
• ER Quality Control: The expression of chaperones such as CANX and CALR was elevated.
• Branching and Extension: MGAT1 was upregulated approximately 1.4-fold; MGAT4A/B were upregulated approximately 2-fold; MGAT5 (a key enzyme catalyzing the formation of β1,6 branches) was upregulated more than 3-fold; and MGAT3 (involved in bisecting GlcNAc formation) was upregulated approximately 5-fold.
• Terminal Modification: The core fucosyltransferase FUT8 was upregulated approximately 1.7-fold; the Lewis antigen fucosyltransferase FUT3 was upregulated approximately 4-fold; the α2,3-sialyltransferases ST3GAL2/3/4 were upregulated 1.7- to 3.3-fold; and ST6GAL1—which predominates in α2,6-sialylation—maintained a high level of expression.

Notably, although the β1,4-galactosyltransferases B4GALT2/3/4/6 were all upregulated, B4GALT1—the predominant β1,4-galactosyltransferase in human liver—was not significant in the TCGA-LIHC dataset but has been reported elsewhere to be downregulated in HCC tissues. This may explain why an overall decrease in galactosylation was still observed at the glycomic level. Survival analysis further revealed that specific glycosylation-related genes, such as B4GALT2 and B4GALT3, were significantly associated with patient prognosis.

Machine Learning Diagnostic Models

This study selected 26 high-abundance glycans with a missing rate below 1% to construct four diagnostic models using an AutoML framework:

• Model 1 (HCC vs. Healthy Controls): Internal validation AUC of 0.93; performance remained robust during external validation.
• Model 2 (HCC vs. Chronic Hepatitis B): Internal validation AUC of 0.90.
• Model 3 (HCC vs. Cirrhosis): Internal validation AUC of 0.84.
• Model 4 (HCC vs. All Non-HCC Cases): Internal validation AUC of 0.87.

Key performance highlights include:

• All three pairwise models demonstrated highly consistent performance across two independent external cohorts, exhibiting only a marginal decline, thereby demonstrating the models' strong generalization capabilities.
• The models were capable of detecting up to 80% of AFP-negative HCC cases, significantly outperforming AFP alone (with DeLong test P-values consistently < 0.05).
• Combining the glycan models with AFP did not result in improved performance, suggesting that the N-glycome alone captures the diagnostic information provided by AFP.
• The models maintained acceptable diagnostic efficacy for Stage I HCC (TNM classification), with Model 1 achieving an AUC of 0.859, thereby demonstrating potential for early-stage screening.

Furthermore, the Model 4-derived probability score (PP4) showed significant correlations with ALBI staging, Child-Pugh classification, and TNM staging, suggesting that it serves not only to distinguish between benign and malignant conditions but also to reflect hepatic functional reserve and tumor burden.

SHAP interpretability analysis revealed that H6N5F1S3, H5N4F1, and H3N4F1 consistently ranked as the top three most important features across all models, closely followed by H5N2, H6N5S3, and H5N4S1. Training a model using these six glycans alone resulted in virtually no significant decline in performance, demonstrating that—through synergistic, non-linear interactions—they constitute the diagnostic core of the system.

Clinical Significance and Future Outlook

Through the utilization of large-scale cohorts and multi-omics integration, this study systematically elucidated the dynamic patterns of the serum N-glycome during the pathogenesis and progression of HCC. Its core value is manifested across three levels:

First, the study confirms that the serum N-glycome serves as a sensitive indicator for assessing liver function; the changes in galactosylation and sialylation reflected therein are closely correlated with hepatocyte synthetic capacity and Golgi functional integrity, holding promise as a complement to existing liver function scoring systems.

Second, the study elucidates the dual origins of alterations in the serum glycome associated with HCC: these encompass both protein-driven effects—arising from changes in the abundance of acute-phase proteins such as A1AG1—and glycosylation-driven effects, stemming from changes in specific glycoforms at sites such as A2MG N869. This deconstruction of multi-level regulation provides a more precise molecular framework for understanding tumor-associated glycosylation.

Finally, a machine learning model based on 26 specific glycans was validated across multiple external cohorts, demonstrating diagnostic performance superior to that of AFP—particularly among AFP-negative patients—thereby offering a clinically viable screening tool for non-invasive surveillance.

Admittedly, the current study is primarily based on a Chinese cohort of patients with hepatitis B-related HCC; future research will require further validation in Western populations with HCC associated with non-alcoholic fatty liver disease or alcoholic liver disease. Furthermore, the lack of longitudinal follow-up data currently precludes a direct assessment of the model's capacity to predict the risk of progression from cirrhosis to HCC. As the throughput of glycomics assays continues to rise and AI algorithms undergo iterative optimization, serum N-glycome analysis is poised to become an integral component of a precision monitoring system for HCC, driving the evolution of liver disease management from reliance on traditional protein biomarkers toward a multi-dimensional molecular diagnostic paradigm.

Related Services & Products

• Glycan Release Service
• Glycan Separation and Purification Service
• Glycan Derivation Service
• Glycomic Characterization Service
• Glycomic Quantitation
• Flow Cytometry Profiling
• Glycoprotein Enrichment
• Glycoprotein Quantification
• Glycoprotein Structure Analysis
• Cancer Cell Glycolytic Analysis
• Cancer Cell TCA Cycle Analysis
• Cancer Cell PPP Analysis
• Gene-level Regulation Analysis of Cancer Glucose Metabolism
• Glucose Metabolism-Microenvironment Crosstalk Analysis in Cancer
• Cancer Glucose Metabolism In Vivo Analysis
• N-Glycan
• O-Glycan

Reference

1. Fu, B., et al. (2026). Large-scale serum N-glycomics tracks N-glycosylation dynamics in hepatocellular carcinoma progression and enables early diagnosis. Nature Communications. DOI: 10.1038/s41467-026-68579-x.