Genome-wide 5-hydroxymethylcytosine in circulating cell-free DNA reflects DNA methylation dynamics for non-invasive detection of early hepatocellular carcinoma

Introduction

Hepatocellular carcinoma (HCC) is one of the most prevalent malignancies and a leading cause of cancer-related mortality worldwide. Despite recent advances in diagnosis and treatment, the early detection of HCC remains a major clinical challenge (1). Owing to its insidious onset, approximately 80% of HCC patients are diagnosed at advanced stages, thereby missing the optimal therapeutic window (2). Liquid biopsy has emerged as a promising non-invasive strategy for tumor surveillance by analyzing circulating tumor components such as cell-free DNA (cfDNA), circulating tumor DNA (ctDNA), exosomes, and circulating tumor cells (CTCs). For instance, surveillance of ctDNA mutations and aberrant methylation patterns has shown good accuracy in early-stage HCC detection (3,4).

5-Hydroxymethylcytosine (5-hmC), generated through the TET enzyme-mediated oxidation of 5-methylcytosine (5-mC), serves as a stable and dynamic epigenetic mark that reflects active DNA demethylation processes (5). Accumulating evidence has indicated that 5-hmC alterations in cfDNA are highly informative for cancer diagnosis, including in HCC (6-8). However, most existing studies have focused on either cfDNA profiling alone or tissue-level methylation analysis in isolation. A comprehensive integrative approach that simultaneously interrogates 5-hmC changes in cfDNA and their corresponding 5-mC alterations in tissues, and further links these epigenetic features to tumor initiation and prognosis, remains insufficiently explored. This gap limits our ability to fully understand the role of 5-hmC in early hepatocarcinogenesis and to translate its potential into clinically applicable biomarkers. Therefore, this study aims to develop and validate a cfDNA 5-hmC-based signature for early HCC detection, and to systematically evaluate its diagnostic performance and its potential biological and prognostic relevance through integrative analysis with tissue-level 5-mC profiles. We present this article in accordance with the TRIPOD reporting checklist (available at https://cco.amegroups.com/article/view/10.21037/cco-2026-1-0022/rc).

Methods

Study subjects

A cfDNA 5-hmC cohort of 2,554 subjects was employed in our analysis [GSE112679 (6)]. The cohort enrolled 388 patients with benign liver lesions (BLL), 286 patients with a history of chronic hepatitis B (CHB), 106 patients with liver cirrhosis (LC), 1,204 patients with HCC, and 570 healthy controls (6). An independent validation cohort of cfDNA 5-hmC was obtained from the Cell-Free Epigenome Atlas (CFEA) (9). Alternatively, the cfDNA 5-mC data was obtained from GSE129374 (10). The tissue 5-mC data (HM450k) was available at The Cancer Genome Atlas (TCGA) and previous studies [GSE113017 (11), GSE89852 (12), and GSE136319 (13)]. Corresponding clinical, molecular, biochemical indicators data of TCGA samples were downloaded from the cBioPortal (https://www.cbioportal.org/). Complete dataset information used in this study was shown in table available at https://cdn.amegroups.cn/static/public/cco-2026-1-0022-1.xlsx. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Data preprocessing

For the 5-hmC data, raw count matrices were processed using the DESeq2 package to construct DESeqDataSet objects. Gene-level expression values in fragments per kilobase per million mapped reads (FPKM) were subsequently calculated using the fpkm function in DESeq2 (11). For the CFEA cohort, gene-level 5-hmC counts were first quantified using deepTools (14), followed by the same normalization procedure described above. For standard β value 5-mC profiles, we removed the 5-mC sites with NA values in over 70% of samples and the missing values were supplemented using the knnInputation function in the R package “DMwR” (15).

Differential DNA methylation analysis

We first mapped the 5-mC sites in the Illumina Infinium HumanMethylation450 BeadChip (450K) to the human genome assembly GRCh37 (hg19) (16). The 5-mC sites located in the 10 kb upstream of genes’ transcription start site (TSS) and 10kb downstream of the transcription end sites (TESs) were used to characterize the DNA methylation at gene levels. We identified the differentially methylated 5-mC sites between stage I/II tumors and adjacent tissues in the TCGA HCC cohort using the Mann-Whitney U test. P values were corrected using the Benjamini-Hochberg (BH) procedure (q-value). The mean methylation levels of 5-mC sites with q-value <0.01 were employed to assess the corresponding gene methylation levels.

EWAS analysis

We obtained trait-associated 5-mC site sets from EWAS Atlas (17). Next, we enriched the differentially methylated 5-mC sites into each trait using Fisher’s exact test. 5-mC sites with q-value <0.01 and were involved in the analysis. Traits with P value <0.01 were considered as the significantly enriched trait.

Differential hydroxymethylation analysis

To identify highly variable genes at the 5-hmC level, we calculated the coefficient of variation (CV) for each gene. The top 10% genes (n=1,910) with the highest CV across all the samples were employed to recognize the differentially hydroxymethylated genes (DHGs) between early HCC and non-HCC states (BL, CHB, LC, and healthy), respectively, using DESeq2 (18). Genes with adjusted P value <0.01 were identified as the DHGs.

Identification of early HCC 5-hmC signature (eHMS)

Firstly, the GSE112679 cohort was divided into two datasets: 2/3 of cases in the training set and 1/3 of cases in the validation set (table available at https://cdn.amegroups.cn/static/public/cco-2026-1-0022-1.xlsx). Then, we integrated an ensemble learning framework to identify the eHMS in the context of different sample states (BL, CHB, LC, and healthy). In the framework, we first constructed a univariate logistic regression model to test the relation between sample states and 5-hmC levels. Genes with P value <0.01 were identified as the candidate predictive markers. Next, we trained an XGBoost classifier using the candidate markers with default parameters implemented in the xgb.train function of the xgboost package (max_depth =6, learning_rate =0.3, min_split_loss =0, reg_lambda =1, reg_alpha =0, tree_method = “hist”, nrounds =100). The area under the curve (AUC) was employed to evaluate the signature accuracy. To obtain more robust markers, 10-fold cross-validation was applied 1,000 times. This process resulted in 10×1,000 output signatures. We ranked the signature importance by calculating the “overrepresented” score (OVS), i.e.,

OVSi=t10000∗∑j=1tAUCi[1]

where t denotes the number of times the signature has been represented, AUC reflects the model prediction accuracy. A larger OVS indicates a higher prediction accuracy of the overrepresented signature. We considered the signature with the highest OVS as the optimal eHMS for different disease states (eHMS-H, -BLL, -CHB, and -LC). We selected all the signature genes across disease states and derived meta-signatures for distinguishing non-HCC and early HCC cases (n=19). The multivariate logistic regression model was constructed in the training set and calculated the eHMS score as ∑n=1n=19αn∗h, where α and h represent the logistic regression coefficient and 5-hmC levels of the n^th gene of eHMS, respectively.

Construction of MeHMS score

We applied the eHMS for distinguishing HCC and non-HCC cases at the cfDNA and tissue levels in 5-mC datasets. For the cfDNA 5-mC cohort [GSE129374 (10)], we divided the samples (n=44, 22 LCs and 22 tumors) equally into a training set (n=22) and a test set (n=22). We then retrained the logistic regression model and calculated the 5-mC-related eHMS (MeHMS) score in the 5-mC cfDNA training set and TCGA-HCC cohort, respectively. And the accuracy of the MeHMS score was tested in the cfDNA testing set and other tissue validation cohorts.

Pathway correlation analysis

We integrated the 2,790 biological pathways from Kyoto Encyclopedia of Genes and Genomes (KEGG) (19), REACTOME (20), WikiPathways (21), and MetaCyc (22). We calculated the enrichment score (ES) of each pathway based on the single-sample gene set enrichment analysis (ssGSEA) in the TCGA HCC cohort (23). Subsequently, Spearman’s correlation test was employed to assess the relationship between ES and MeHMS scores in adjacent tissues and different stage tumors, respectively. P values were adjusted by the BH procedure (q-value). The strength of correlation (CS) was calculated as:

CS=−log10(q)*|Rho|[2]

Hypoxic tumor microenvironment analysis

The tumor hypoxia score (H) was calculated based on the transcription levels of 51-gene signature from the previous study of Buffa et al. (24). The tumor microenvironment-enriched cell types were assessed by xCell (25). We applied the linear regression model to test for the relationship between tumor microenvironment and hypoxia score, i.e.,

Cj∼α∗H+β+ε[3]

where C represents the ES of microenvironment cell j, β denotes the co-variables, including the patient’s race, age, and gender. The resulting P values were adjusted for multiple testing using the BH procedure (q-value). Cells with a q-value <0.05 were considered hypoxia-related microenvironment cells. These cells were subsequently involved in Spearman’s correlation analysis between microenvironmental abundance and MeHMS score.

Tumor stemness evaluation

Stemness scores of TCGA-HCC samples were calculated by the previous study of Malta et al. (26). The stemness scores were measured by epigenetic (methylome and transcription-factor binding sites) and gene expression features, respectively. Based on the two independent stemness indices, we performed Spearman’s correlation test between stemness score and MeHMS score.

Survival analysis

To investigate the relationship between eHMS and prognostic risk, we performed the multivariate Cox regression analysis based on patients’ disease-specific survival (DSS) in the TCGA cohort. The risk score of eHMS was calculated as:

Riskscore=∑i=1n=19βi*Mi[4]

where β_i and M_i represents the regression coefficient and methylation levels of the i^th gene in eHMS. We subgrouped the tumors into the high- and low-risk groups based on the median risk score. The statistical significance between the two groups was calculated using the log-rank test. Additionally, the risk score was employed for overall, relapse-free, and progression-free survival analysis.

Statistics analysis

The statistical analysis was performed using R, version 4.1.2. P values less than 0.05 (two-sided) were considered statistically significant. Detailed statistical methods are described in the corresponding sections where they are applied.

Results

Overview of cfDNA hydroxymethylation during HCC initiation

DNA methylation at cytosine residues (5-mC) is a well-recognized epigenetic modification, and its oxidation by TET enzymes generates 5-hmC, a more dynamic and stable marker that reflects active epigenomic remodeling (Figure 1A). Emerging evidence suggests that aberrant 5-hmC patterns in cfDNA can serve as sensitive indicators of early tumorigenesis. To systematically characterize the dynamics of cfDNA hydroxymethylation during hepatocarcinogenesis, we analyzed a large cohort of 2,554 subjects from Cai et al. (6), which encompassed five representative hepatic states: healthy controls, BLL, chronic hepatitis B (CHB), liver cirrhosis (LC), and HCC. From this dataset, we identified 1,910 highly variable genes (top 10% by CV) at the 5-hmC level (‘Methods’). Notably, the distribution of cfDNA 5-hmC signals in these genes was markedly distinct between HCC and non-HCC cases (Figure 1B), with overall 5-hmC levels consistently elevated in HCC compared with other states. Similar increases were also observed in tumor tissues relative to adjacent non-tumor tissues (Figure 1C), suggesting globally active hydroxymethylation in the cfDNA and tissue of HCC patients.

Figure 1 The cfDNA hydroxymethylation landscape in HCC. (A) Schematic of conversion of 5-mC to 5-hmC. 5-mC is converted to 5-hmC through the TET enzymes. (B) Density plot showing the average 5-hmC levels of cfDNA among healthy, BLL, CHB, LC, early HCC, and late HCC samples. (C) Box plot showing the average 5-hmC levels between HCC and adjacent tissues. (D) Box plot showing the average 5-mC levels between different tumor stages and adjacent tissues in the TCGA-HCC cohort. (E) Box plot showing the average 5-mC levels between LC and LC with HCC tumors in GSE129374. (F) Bar plot showing the enriched EWAS traits. (G) Venn plot showing the intersections of healthy-, LC-, CHB-, and BLL-DHGs. Bar plot indicating the number of DHG shown on the right. (H) Sankey plot showing the enriched KEGG pathways of each type DHG. *, P<0.05; **, P<0.01; ***, P<0.001, as calculated by the Mann-Whitney U test (C-E). 5-hmC, 5-hydroxymethylcytosine; 5-mC, 5-methylcytosine; BLL, benign liver lesions; cfDNA, cell-free DNA; CHB, chronic hepatitis B; DHG, differentially hydroxymethylated gene; EWAS, epigenome-wide association study; HCC, hepatocellular carcinoma; KEGG, Kyoto Encyclopedia of Genes and Genomes; LC, liver cirrhosis; TCGA, The Cancer Genome Atlas.

To further explore the epigenetic context, we examined corresponding 5-mC profiles in the TCGA-HCC cohort. Tumors generally exhibited reduced 5-mC levels compared to adjacent non-tumor tissues (Figure 1D), consistent with widespread demethylation accompanying hydroxymethylation gain. Independent validation in another cfDNA methylation dataset [Hlady et al. (10)] confirmed that HCC patients showed significantly lower 5-mC than LC patients (Figure 1E). Differential 5-mC analysis in TCGA tumors revealed sites strongly enriched in epigenome-wide association study (EWAS) recognized HCC risk traits, including diabetes, alcohol consumption, environmental exposure, and high saturated-fat diet (Figure 1F). These findings imply that epigenomic alterations bridge external risk factors and early tumor initiation. We next identified DHGs in cfDNA by comparing early HCC with non-HCC states (Figure S1). The largest number of DHGs was observed between healthy and early HCC subjects, with the majority being upregulated (Figure 1G). Functional annotation showed that dysregulated DHGs were significantly enriched in pathways related to cancer, immune regulation, cell signaling, and metabolism (Figure 1H). Particularly, 5-hmC-enriched metabolic pathways such as alcoholism (hsa05034), butanoate metabolism (hsa00650), and diabetes (hsa04950) were highlighted, all of which have established links to HCC risk. These results delineate a distinct cfDNA hydroxymethylation landscape during HCC initiation, demonstrate its interplay with corresponding 5-mC alterations, and underscore the promise of cfDNA 5-hmC as a non-invasive biomarker for early HCC surveillance.

Circulating cfDNA 5-hmC signature in early HCC diagnosis

Because early-stage HCC often presents with an insidious onset and CHB and LC represent high-risk conditions, we employed an ensemble learning framework to identify a robust cfDNA-based 5-hmC signature for early HCC detection (‘Methods’). Four condition-specific signatures were first derived (eHMS-H, eHMS-BLL, eHMS-CHB, and eHMS-LC), which were subsequently integrated into a 19-gene meta-signature, termed eHMS (Figure 2A). Based on this signature, an eHMS score was calculated, which was significantly elevated in both early- and late-stage HCC compared with healthy, BLL, CHB, and LC groups (Figure 2B), highlighting its potential utility for surveillance. We next evaluated the diagnostic performance of eHMS in comparison with a previously reported 32-gene 5-hmC signature by Cai et al. (6). Across multiple disease-state comparisons, eHMS consistently achieved superior accuracy. Specifically, the AUCs for distinguishing early HCC from healthy, BLL, CHB, and LC were 92.3%, 82.5%, 78.8%, and 81.3%, respectively (Figure S2). When pooling all non-HCC groups versus early HCC, the eHMS achieved an AUC of 89.4% (Figure 2C). Importantly, in an external validation cohort from the CFEA (10), eHMS maintained strong diagnostic performance with an AUC of 93.5% (table available at https://cdn.amegroups.cn/static/public/cco-2026-1-0022-1.xlsx). These results establish eHMS as a reliable and high-performing cfDNA 5-hmC signature, supporting its application as a non-invasive biomarker for early HCC surveillance.

Figure 2 eHMS contributes to the early HCC detection at both 5-hmC and 5-mC levels. (A) The flowchart of identifying eHMS is based on ensemble learning. (B) Density plot showing the eHMS score among healthy, benign BL, CHB, LC, early HCC, and late HCC samples. P values were calculated by the Mann-Whitney U test. (C) The performance of the eHMS and Cai et al. signature (32-gene 5-hmc) in distinguishing non-HCC and HCC samples in training and validation sets. 5-hmC, 5-hydroxymethylcytosine; 5-mC, 5-methylcytosine; AUC, area under the curve; BLL, benign liver lesions; CHB, chronic hepatitis B; CI, confidence interval; DHG, differentially hydroxymethylated gene; eHMS, early HCC 5-hmC signature; HCC, hepatocellular carcinoma; LC, liver cirrhosis.

eHMS is capable of interrogating tumor tissues at the 5-mC level

cfDNA is DNA fragments in the blood released from dying cells in tissues (27). Hence, we further traced eHMS to the DNA methylation profiles of HCC tissues using the TCGA methylome data (Figure 3A). Among the 19 genes of eHMS, 13 showed significant hyper-hydroxymethylation and 6 displayed hypo-hydroxymethylation in cfDNA from early HCC cases (Figure S3A). We then evaluated the 5-mC levels of these genes based on the mean β-values of their corresponding differentially methylated 5-mC sites (Figure S3B). In total, 16 genes were differentially methylated between tumors and adjacent tissues, and these were used to derive a methylation-based eHMS score (MeHMS) (table available at https://cdn.amegroups.cn/static/public/cco-2026-1-0022-1.xlsx). Consistent with the cfDNA-derived eHMS, the MeHMS score was significantly higher in tumors across all stages compared with adjacent tissues (Figure 3B; Figure S3C). When benchmarked against classical HCC marker alpha-fetoprotein (AFP), the MeHMS score achieved superior diagnostic performance (AUC =99.9%) in distinguishing tumors from adjacent tissues (Figure 3C). Robust predictive power of MeHMS was further validated across three independent cohorts (Figure 3D). Additionally, recalculation of the MeHMS score using cfDNA 5-mC data from an independent dataset confirmed its utility, showing superior performance in distinguishing LC from HCC (AUC =89.5%) (Figure 3E). These findings indicate that the eHMS signature is reliable at both the 5-hmC and 5-mC levels, underscoring its robustness as a biomarker for HCC detection.

Figure 3 Validation of the MeHMS score in distinguishing HCC from non-HCC. (A) Circosplot showing the eHMS signature’s mean 5-hmC levels and their corresponding 5-mC levels (TCGA cohort) between non-HCC and HCC samples. (B) Violin plot showing the MeHMS score between adjacent tissues and tumors with different stages. (C-E) The performance of MeHMS in distinguishing adjacent tissues and HCC samples in validation-TCGA (C), validation-tissue (D), and validation-cfDNA (E) cohorts. *, P<0.05; **, P<0.01; ***, P<0.001, as calculated by the Mann-Whitney U test (A,B). 5-hmC, 5-hydroxymethylcytosine; 5-mC, 5-methylcytosine; AFP, alpha-fetoprotein; ALB, albumin; AUC, area under the curve; CI, confidence interval; eHMS, early HCC 5-hmC signature; HCC, hepatocellular carcinoma; MeHMS, 5-mC-related eHMS; TCGA, The Cancer Genome Atlas.

eHMS prompts metabolic plasticity in the tumor microenvironment

To further explore the biological significance of the MeHMS score, we performed pathway enrichment analyses across tumor progression stages in the TCGA cohort (‘Methods’). The MeHMS score was strongly associated with multiple metabolic processes, particularly glycometabolic pathways (Figure 4A). In stage I tumors, genes implicated in glycogen metabolism and regulation, including PPP2R family members and GYG1, were significantly upregulated (Figure 4B), suggesting that the MeHMS score reflects enhanced glycogen turnover during early tumor development. Consistent with recent reports that glycogen accumulation can suppress Hippo signaling and promote HCC initiation (28), we found that tumors with high MeHMS scores exhibited alterations in pathways regulating Hippo signaling (Figure 4B), and this association was most evident in stage I disease (Figure S4). We further observed that tumors with higher MeHMS scores were enriched for hypoxia signatures (P=3.5×10⁻⁴; Figure 4C), supporting the role of hypoxia in metabolic reprogramming during early hepatocarcinogenesis. MeHMS was also negatively correlated with the global microenvironment score derived from xCell (P=2.2×10⁻⁹; Figure 4D), reflecting that the MeHMS score captures the complexity of the tumor microenvironment, with stromal cell signatures showing the strongest associations (Figure S5). Alternatively, the hypoxic microenvironment is also enriched in some stem cell characteristics (hematopoietic stem cells and mesenchymal stem cells), which are obviously linked to MeHMS scores (Figure 4E). Further parsing of the relations between tumor stemness and MeHMS score showed that the eHMS was a promising indicator (Figure 4F) in both stemness epigenetic features (Rho =0.69, P value <2.2e−16) and gene expression features (Rho =0.45, P value <2.2e−16) (‘Methods’) (26). The crosstalk among MeHMS score, metabolism, hypoxic microenvironment, and tumor stemness collectively suggested that the MeHMS score is capable of indicating the metabolic plasticity in the tumor microenvironment.

Figure 4 eHMS is involved in metabolic plasticity in the tumor microenvironment. (A) Biological pathways are correlated with MeHMS score in different tumor stages of TCGA cohort. Glycometabolic metabolism-related pathways were labeled on the right. (B) Dysregulated genes in glycogen synthesis and degradation are associated with tumor hypoxia scores. In the panel, the bar plot shows the fold change of significantly differentially expressed genes at the transcription levels on the left. The bubble plot shows Spearman’s correlation coefficient between gene transcription and tumor hypoxia score on the right. (C,D) Scatter plot showing the association between MeHMS score and (C) hypoxia score/(D) microenvironment score, respectively. (E) Heatmap showing the crosstalk among microenvironment cells, MeHMS score, and Hypoxia score. (F) MeHMS score shows a significant correlation with the tumor stemness score at both epigenetic and gene expression levels. ****, P<0.0001; NS, not significant. In the top panel, P values were calculated using the Mann–Whitney U test to compare tumor samples at each stage with normal samples, while in the bottom panel, P values were derived from Spearman correlation analysis. eHMS, early HCC 5-hmC signature; HCC, hepatocellular carcinoma; KEGG, Kyoto Encyclopedia of Genes and Genomes; MeHMS, 5-mC-related eHMS; TCGA, The Cancer Genome Atlas.

eHMS contributes to patients’ prognosis and disease progression

We next investigated the prognostic relevance of eHMS in HCC. Kaplan-Meier survival analysis demonstrated that patients classified into the eHMS high-risk group had significantly worse outcomes compared with those in the low-risk group, including shorter DSS (P<0.001), relapse-free survival (RFS, P=0.041), and progression-free survival (PFS, P=0.02), with a similar trend observed for overall survival (OS, P=0.18) (Figure 5A). To further validate these findings, we performed multivariate Cox regression analysis adjusting for age, sex, and stage. The eHMS risk score emerged as an independent prognostic factor across multiple survival endpoints, consistently associated with higher hazard ratios for DSS, OS, RFS, and PFS (Figure 5B). Together, these results indicate that eHMS not only stratifies patients by recurrence and progression risk but also provides an independent predictor of long-term survival outcomes in HCC.

Figure 5 Prognostic value of eHMS in HCC patients. (A) Kaplan-Meier curves of tumor samples stratified by the median risk score of eHMS in disease-specific, overall, relapse-free, and progression-free survival, respectively. (B) Forest plot showing the result of multivariate Cox-regression analysis for correlation between the eHMS-risk groups and the disease-specific, overall, relapse, and progression-free survivals after adjusted diagnosis age, gender, and race. *, P<0.05; **, P<0.001; ***, P<0.0001. AIC, Akaike information criterion; DSS, disease-specific survival; eHMS, early HCC 5-hmC signature; HCC, hepatocellular carcinoma; HR, hazard ratio; OS, overall survival; PFS, progression-free survival; RFS, relapse-free survival.

Discussion

Recent evidence has highlighted that aberrant 5-hmC patterns in cfDNA are not only markers of tumor presence but also reflective of underlying risk exposures that predispose to hepatocarcinogenesis. In our analysis, DHGs were enriched in pathways linked to established HCC risk factors, including alcohol consumption, diabetes, environmental exposure, and high-fat dietary intake. This suggests that circulating 5-hmC alterations may capture the epigenetic footprints of these carcinogenic stimuli before overt tumor development. Consistent with prior reports that metabolic and lifestyle factors contribute to early epigenome remodeling in liver (29,30), our findings support the potential of cfDNA 5-hmC profiling as an integrated readout of both genetic and environmental risk in individuals at high risk for HCC. We hence established a 19-gene cfDNA 5-hmC signature (eHMS) for early HCC detection and further derived a 5-mC-related counterpart (MeHMS) for tissue-based validation. Together, these dual-layer epigenetic biomarkers demonstrated robust and consistent performance across cfDNA and tissue datasets, with a trend toward improved sensitivity and specificity compared with AFP-based assessments (30).

By linking cfDNA hydroxymethylation with tissue methylation states, this study provides integrative evidence that non-invasive epigenetic signatures can faithfully represent tumor biology. Importantly, the ability of eHMS to discriminate early HCC from high-risk backgrounds such as chronic hepatitis B and cirrhosis underscores its potential clinical utility in surveillance programs, where current imaging and serological tests often fail (31). Beyond diagnostic value, the association of MeHMS with hypoxia- and stemness-related features suggests that early HCC epigenomes may capture microenvironmental reprogramming (32,33), although further functional validation is needed. The observed prognostic significance further supports eHMS as tools for both early detection and patient stratification.

Limitations include reliance on retrospective public datasets and potential variability in cfDNA processing. Prospective validation in large, multi-center cohorts will be essential to confirm clinical applicability. Additionally, the exceptionally high diagnostic performance observed for the MeHMS score is derived from tissue-based comparisons, where the contrast between tumor and adjacent non-tumor samples is inherently strong. In contrast, cfDNA-based analyses in blood reflect a heterogeneous mixture of tumor- and non-tumor-derived DNA, which may reduce the signal-to-noise ratio and limit direct translation of tissue-level performance to clinical settings. This distinction also highlights the importance of cautious interpretation when extrapolating tissue-based findings to blood-based applications and underscores the need for further validation of cfDNA-based performance in clinically relevant settings. In addition, the current study was based solely on 5-hmC signal intensity. Future studies integrating sequence-derived fragmentomic features with epigenetic modification signals may further improve the sensitivity and robustness of cfDNA-based detection. In summary, this study demonstrates that genome-wide cfDNA 5-hmC signatures, supported by 5-mC-derived validation, provide a powerful and non-invasive approach for early HCC detection and prognostication. Such dual-layer epigenetic markers hold promise for incorporation into precision surveillance strategies to improve outcomes in at-risk populations.

Conclusions

Our study establishes a cfDNA-based 5-hmC signature as a non-invasive biomarker for early HCC detection. This approach shows promising potential to enhance current surveillance strategies and improve patient management.

Acknowledgments

We thank the State Key Laboratory of Complex, Severe, and Rare Diseases, and the Center for Bioinformatics at the National Infrastructures for Translational Medicine, Peking Union Medical College Hospital, for providing high-performance computing services. All the data used in this manuscript can be obtained from TCGA and GEO.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://cco.amegroups.com/article/view/10.21037/cco-2026-1-0022/rc

Peer Review File: Available at https://cco.amegroups.com/article/view/10.21037/cco-2026-1-0022/prf

Funding: This work was supported by the National Key Research and Development Program of China (No. 2020YFA0803702 to J.P.), and the Peking Union Medical College Hospital Outstanding Young Talent Development Program (No. UBJ11753 to J.P.).

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://cco.amegroups.com/article/view/10.21037/cco-2026-1-0022/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Llovet JM, Kelley RK, Villanueva A, et al. Hepatocellular carcinoma. Nat Rev Dis Primers 2021;7:6. [Crossref] [PubMed]
2. Chen X, Liu HP, Li M, et al. Advances in non-surgical management of primary liver cancer. World J Gastroenterol 2014;20:16630-8. [Crossref] [PubMed]
3. Labgaa I, Villacorta-Martin C, D'Avola D, et al. A pilot study of ultra-deep targeted sequencing of plasma DNA identifies driver mutations in hepatocellular carcinoma. Oncogene 2018;37:3740-52. [Crossref] [PubMed]
4. Kisiel JB, Dukek BA. Hepatocellular Carcinoma Detection by Plasma Methylated DNA: Discovery, Phase I Pilot, and Phase II Clinical Validation. Hepatology 2019;69:1180-92. [Crossref] [PubMed]
5. Tahiliani M, Koh KP, Shen Y, et al. Conversion of 5-methylcytosine to 5-hydroxymethylcytosine in mammalian DNA by MLL partner TET1. Science 2009;324:930-5. [Crossref] [PubMed]
6. Cai J, Chen L, Zhang Z, et al. Genome-wide mapping of 5-hydroxymethylcytosines in circulating cell-free DNA as a non-invasive approach for early detection of hepatocellular carcinoma. Gut 2019;68:2195-205. [Crossref] [PubMed]
7. Li W, Zhang X, Lu X, et al. 5-Hydroxymethylcytosine signatures in circulating cell-free DNA as diagnostic biomarkers for human cancers. Cell Res 2017;27:1243-57. [Crossref] [PubMed]
8. Bachman M, Uribe-Lewis S, Yang X, et al. 5-Hydroxymethylcytosine is a predominantly stable DNA modification. Nat Chem 2014;6:1049-55. [Crossref] [PubMed]
9. Yu F, Li K, Li S, et al. CFEA: a cell-free epigenome atlas in human diseases. Nucleic Acids Res 2020;48:D40-4. [Crossref] [PubMed]
10. Hlady RA, Zhao X, Pan X, et al. Genome-wide discovery and validation of diagnostic DNA methylation-based biomarkers for hepatocellular cancer detection in circulating cell free DNA. Theranostics 2019;9:7239-50. [Crossref] [PubMed]
11. Shimada S, Mogushi K, Akiyama Y, et al. Comprehensive molecular and immunological characterization of hepatocellular carcinoma. EBioMedicine 2019;40:457-70. [Crossref] [PubMed]
12. Kuramoto J, Arai E, Tian Y, et al. Genome-wide DNA methylation analysis during non-alcoholic steatohepatitis-related multistage hepatocarcinogenesis: comparison with hepatitis virus-related carcinogenesis. Carcinogenesis 2017;38:261-70. [Crossref] [PubMed]
13. Cerapio JP, Marchio A, Cano L, et al. Global DNA hypermethylation pattern and unique gene expression signature in liver cancer from patients with Indigenous American ancestry. Oncotarget 2021;12:475-92. [Crossref] [PubMed]
14. Ramírez F, Ryan DP, Grüning B, et al. deepTools2: a next generation web server for deep-sequencing data analysis. Nucleic Acids Res 2016;44:W160. [Crossref] [PubMed]
15. Acuna E. Dprep: Data Pre-Processing and Visualization Functions for Classification. 2015.
16. Harrow J, Frankish A, Gonzalez JM, et al. GENCODE: the reference human genome annotation for The ENCODE Project. Genome Res 2012;22:1760-74. [Crossref] [PubMed]
17. Li M, Zou D, Li Z, et al. EWAS Atlas: a curated knowledgebase of epigenome-wide association studies. Nucleic Acids Res 2019;47:D983-8. [Crossref] [PubMed]
18. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014;15:550. [Crossref] [PubMed]
19. Kanehisa M, Furumichi M, Sato Y, et al. KEGG: integrating viruses and cellular organisms. Nucleic Acids Res 2021;49:D545-51. [Crossref] [PubMed]
20. Jassal B, Matthews L, Viteri G, et al. The reactome pathway knowledgebase. Nucleic Acids Res 2020;48:D498-503. [Crossref] [PubMed]
21. Martens M, Ammar A, Riutta A, et al. WikiPathways: connecting communities. Nucleic Acids Res 2021;49:D613-21. [Crossref] [PubMed]
22. Karp PD, Riley M, Paley SM, et al. The MetaCyc Database. Nucleic Acids Res 2002;30:59-61. [Crossref] [PubMed]
23. Barbie DA, Tamayo P, Boehm JS, et al. Systematic RNA interference reveals that oncogenic KRAS-driven cancers require TBK1. Nature 2009;462:108-12. [Crossref] [PubMed]
24. Buffa FM, Harris AL, West CM, et al. Large meta-analysis of multiple cancers reveals a common, compact and highly prognostic hypoxia metagene. Br J Cancer 2010;102:428-35. [Crossref] [PubMed]
25. Aran D, Hu Z, Butte AJ. xCell: digitally portraying the tissue cellular heterogeneity landscape. Genome Biol 2017;18:220. [Crossref] [PubMed]
26. Malta TM, Sokolov A, Gentles AJ, et al. Machine Learning Identifies Stemness Features Associated with Oncogenic Dedifferentiation. Cell 2018;173:338-354.e15. [Crossref] [PubMed]
27. Fan HC, Blumenfeld YJ, Chitkara U, et al. Noninvasive diagnosis of fetal aneuploidy by shotgun sequencing DNA from maternal blood. Proc Natl Acad Sci U S A 2008;105:16266-71. [Crossref] [PubMed]
28. Liu Q, Li J, Zhang W, et al. Glycogen accumulation and phase separation drives liver tumor initiation. Cell 2021;184:5559-5576.e19. [Crossref] [PubMed]
29. McGlynn KA, Petrick JL, El-Serag HB. Epidemiology of Hepatocellular Carcinoma. Hepatology 2021;73:4-13. [Crossref] [PubMed]
30. Plaz Torres MC, Jaffe A, Perry R, et al. Diabetes medications and risk of HCC. Hepatology 2022;76:1880-97. [Crossref] [PubMed]
31. Villanueva A, Minguez B, Forner A, et al. Hepatocellular carcinoma: novel molecular approaches for diagnosis, prognosis, and therapy. Annu Rev Med 2010;61:317-28. [Crossref] [PubMed]
32. Lin HY, Jeon AJ, Chen K, et al. The epigenetic basis of hepatocellular carcinoma - mechanisms and potential directions for biomarkers and therapeutics. Br J Cancer 2025;132:869-87. [Crossref] [PubMed]
33. Fernández-Barrena MG, Arechederra M, Colyn L, et al. Epigenetics in hepatocellular carcinoma development and therapy: The tip of the iceberg. JHEP Rep 2020;2:100167. [Crossref] [PubMed]