Combined epigenomic landscapes of 5mC, 5hmC, and 6mA modifications in papillary thyroid carcinogenesis

Abstract

Papillary thyroid carcinoma (PTC), though genetically characterized, lacks comprehensive epigenomic profiling. Here we present the first combined analysis of 5-methylcytosine (5mC), 5-hydroxymethylcytosine (5hmC), and 6-methyladenine (6mA) in PTC pathogenesis. Using mDIP-seq on matched tumor and adjacent normal tissues from six treatment-naïve patients, we mapped genome-wide distributions of these modifications, revealing distinct spatial patterns and tumor-specific differentially methylated regions (DMRs) with modification-specific functional partitioning. Silhouette-guided feature selection refined 7,117 DMRs to a 187-locus diagnostic panel achieving robust tumor-normal discrimination, with chromosomes 16 and 19 as recurrent epigenetic hotspots (2.1–7.9-fold enrichment). Integration of TCGA data revealed mutation-dependent epigenetic-transcriptional associations: BRAF-mutant tumors showed selective upregulation of epigenetically dysregulated genes, whereas RAS-mutant tumors displayed distinct transcriptional patterns. Co-expression network analysis highlighted disrupted transcriptional coordination in tumors. This work establishes a multidimensional epigenetic framework for PTC with combinatorial diagnostic signatures and condition-specific associations between DNA modifications and driver mutations (BRAF vs. RAS).

Combined profiling of 5mC, 5hmC, and 6mA in papillary thyroid carcinoma reveals modification specific epigenomic landscapes, a 187-locus diagnostic panel, and BRAF/RAS mutation dependent transcriptional reprogramming.

Introduction

Thyroid cancer, originating in thyroid follicular cells, is the most common primary endocrine malignancy. Papillary thyroid carcinoma (PTC), characterized by its papillary histology, accounts for more than 80% of cases^1,2. Over recent decades, the incidence of thyroid cancer has steadily increased, largely attributable to advancements in diagnostic technologies³. Although PTC is generally associated with a favorable prognosis, challenges remain in achieving early detection and accurately predicting recurrence, underscoring the need for more comprehensive molecular characterization. Genetic studies have revealed a high prevalence of activating somatic alterations in effectors of the mitogen-activated protein kinase (MAPK) signaling pathway, including BRAF and the RAS point mutations, which promote constitutive pathway activation and tumor initiation^2,4. However, while the genetic landscape of PTC has been extensively characterized, elucidation of its epigenetic landscape remains relatively scarce.

Emerging evidence highlights the pivotal role of epigenetic dysregulation, particularly DNA modifications, in tumorigenesis⁵. The canonical modification 5-methylcytosine (5mC), maintained by DNA methyltransferases (DNMTs), is enriched at silenced gene promoters and repetitive elements, where it mediates transcriptional repression⁶. Although aberrant 5mC has often been closely linked to the silencing of tumor suppressor genes in cancer, recent research demonstrates that DNA methylation also plays complex roles outside promoters, such as enhancers and gene bodies⁷. In contrast, 5-hydroxymethylcytosine (5hmC), generated via ten-eleven translocation (TET)-mediated oxidation of 5mC, serves dual roles as both a demethylation intermediate and as an independent epigenetic mark that regulates enhancer activity^8,9. Both 5mC and its oxidized derivative 5hmC contribute to tumorigenesis through mechanisms including tumor suppressor silencing, oncogene activation, genomic instability, and the acquisition of aggressive disease phenotypes^10,11. More recently, 6-methyladenine (6mA)—traditionally considered a prokaryotic modification—has been identified as a dynamic epigenetic mark in eukaryotes. It is implicated in transposon suppression, chromatin accessibility, and other regulatory processes¹². In human cancers, 6mA has been associated with altered gene expression, stemness, and metastasis^13,14. Notably, growing evidence points to functional crosstalk between 6mA and cytosine modifications (5mC/5hmC), involving coordinated chromatin regulation at distal elements or mutual exclusivity at specific loci^15,16. The discovery of a bifunctional dioxygenase capable of demethylating both 5mC and 6mA further underscores the potential for integrated regulation of these modifications¹⁷. Recent studies have begun to map such modification dynamics in cancer. For example, Zhou et al. observed markedly reduced 6mA compared with 5mC in pancreatic cancer, suggesting discordant regulation of these marks during tumorigenesis¹⁸. Although individual alterations in DNA methylation have been preliminarily examined in thyroid cancer^19,20, the integrated landscape of 5mC, 5hmC, and 6mA and their interplay during PTC tumorigenesis remains largely unexplored.

This study provides the first combined profiling of 6mA, 5mC, and 5hmC dynamics across matched tumor and adjacent normal tissues from six treatment-naïve patients with PTC to the best of our knowledge. By applying multi-omics analysis, we delineate methylation-specific patterns associated with PTC progression, identify pathway-level convergence through functional genomics, and characterize relationships between methylation signatures and transcriptional reprogramming, supported by validation in The Cancer Genome Atlas (TCGA) cohort. Our findings reveal previously unrecognized epigenetic architectures that define malignant transformation, establishing a novel mechanistic framework for understanding PTC pathogenesis within a multidimensional epi-transcriptomic context.

Methods

Patients and samples

Fresh tumor tissues and matched adjacent normal tissues were collected from six treatment-naïve patients with PTC undergoing curative surgery at the First Affiliated Hospital of Wannan Medical College. Immediately after resection, all specimens were snap-frozen in liquid nitrogen and stored at –80 °C until further processing. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Ethics Committee of the First Affiliated Hospital of Wannan Medical University (approval No. 2022–34). Written informed consent was obtained from all participants prior to sample collection. All ethical regulations relevant to human research participants were followed. Table 1 summarizes the clinicopathological characteristics of the cohort, classified according to AJCC 8th edition staging criteria. All patients were female (age range: 35–57 years) with histologically confirmed PTC, comprising three bilateral and three unilateral cases. None had received radioiodine therapy or systemic treatments before sample collection. For multidimensional epigenomic profiling, we performed duplicate sequencing runs for each of the 12 matched tumor-normal tissue pairs (6 patients × 2 tissue types) across three DNA modifications. This experimental design generated 72 raw sequencing datasets (12 pairs × 3 modifications × 2 technical replicates), which were subsequently merged during bioinformatic processing.

Table 1.

Clinical and pathological characteristics of the patients

PatientID	Gender	Age	Pathological diagnosis	Laterality	AJCC 8th edition pTNM stage
S3	Female	45	PTC	Right	pT1aN0Mx
S4	Female	57	PTC	Right	pT1N0Mx
S5	Female	35	PTC	Bilateral	pT1N0Mx
S12	Female	57	PTC	Bilateral	pT1aN0Mx
S10	Female	48	PTC	Left	pT1N1Mx
S11	Female	44	PTC	Bilateral	pT1N0Mx

mDIP-seq library construction and sequencing

mDIP-seq (multiplex DNA immunoprecipitation followed by sequencing) is a high-throughput and cost-efficient method for profiling multiple DNA modifications across large-scale samples by using the single-stranded DNA (ssDNA) barcoding technique we developed previously^21,22. Briefly, genomic DNA (1–5 μg) was processed for mDIP-seq following an established protocol with critical modifications, 10% of purified DNA was retained as an input control for each sample, and IgG controls were included in all immunoprecipitation experiments to assess antibody specificity. RNase-treated DNA was fragmented to 200–400 bp using focused ultrasonication (Covaris ME220; 70 W, 20% duty factor, 1000 cycles/burst, 130 s at 4 °C). After denaturation and adapter ligation, multiplexed immunoprecipitation was performed by incubating DNA overnight at 4°C with species-specific antibodies: anti-5mC (5 μg; Thermo Fisher MA538432), anti-5hmC (3.5 μg; Abcam ab106918), and anti-6mA (8 μg; Synaptic Systems 202003) (Supplementary Table 1). Antigen-antibody complexes were captured using 30 μL Dynabeads™ Protein A/G (Thermo Fisher 26162), rigorously washed, and heat-eluted. Immunoprecipitated DNA was converted to double-stranded libraries using KAPA HiFi HotStart ReadyMix, ligated to Illumina-compatible adapters, and amplified for 10–20 cycles (qPCR-optimized) with indexed primers (Supplementary Table 2). Final libraries were quantified using an Agilent Bioanalyzer and sequenced on an Illumina NovaSeq 6000 platform (150 bp paired-end; Jiangsu Biotech Public Service Platform, Nanjing). Biological replicates were processed independently at all stages. Comprehensive quality control metrics for all samples, including alignment rates, duplication levels, GC content, percent pass filter, and adapter content, are provided in Supplementary Data 1.

Bioinformatics analysis of mDIP-seq data

The mDIP-seq data were processed through a standardized bioinformatics pipeline. Raw paired-end reads were subjected to quality control and adapter trimming using Trim Galore (v0.6.10) with default parameters, retaining sequences with Phred quality scores ≥30. Processed reads were aligned to the reference genome GRCh38/hg38 (UCSC Genome Browser) using Bowtie2 (v2.5.1), and resulting alignments were converted into sorted binary files using SAMtools (v1.6). Duplicate reads were removed with Picard tools (v2.27.5), retaining only unique alignments. Genome-wide methylation enrichment peaks were identified using MACS2 (v2.2.7.1) with an effective genome size of 2.91 × 10⁹bp (GENCODE GRCh38). Normalized methylation profiles were generated with deepTools (v3.5.1) at 10-bp resolution using reads per genomic content (RPGC) normalization.

Genome-wide profiling and functional annotation of differential 5mC/5hmC/6mA modifications

ChIP-seq data processing and differential analysis were conducted using DiffBind (v3.10.1) for peak read quantification, normalization, and identification of 400-bp tumor-normal differentially methylated regions (DMRs) for 5mC, 5hmC, and 6mA (FDR < 0.05, |log₂ fold change | >1). Functional annotation of differential regions was performed using ChIPseeker (v1.36.0) with the hg38 genome annotation database. Differential regions landscapes were visualized through multiple approaches. Volcano plots were generated to highlight significantly enriched or depleted sites. UMAP dimensionality reduction was applied for sample clustering, and silhouette analysis was used to evaluate cluster separation. For genome-wide unsupervised embedding, the DiffBind-derived consensus 400-bp region matrix (sample × region, normalized read counts; hg38) was used as input: mark-specific UMAPs were computed independently for 5mC, 5hmC, and 6mA, and an integrated embedding was generated by concatenating the three mark-specific 2D coordinates into a single matrix followed by a second UMAP. DMR-based embeddings used the differential subset of the same consensus matrix. These analyses were implemented using umap (v0.2.10.0), ggplot2 (v3.5.1), and factoextra (v1.0.7). Cross-modification integration was performed with GenomicRanges (v1.52.1), enabling joint analyses of genomic coordinates across the three methylation marks. Gene ontology (GO) enrichment of differential regions was performed using rGREAT package (v2.2.0) with the “GO: Biological Process” ontology. Significantly enriched terms were filtered, and the top six terms were visualized as horizontal bar plots. The x-axis corresponded to -log₁₀(p-value) (except for 5mC regions, where fold enrichment was used due to near-zero p-values). All statistical analyses were conducted in R (v4.3.3). Non-redundant de novo motifs were identified from DMRs using HOMER (v5.1) with default parameters (findMotifsGenome.pl). Motifs were ranked by ascending pvalue, and the top 3 motifs were selected for visualization.

Feature selection and dimensionality reduction based on silhouette analysis

To optimize feature selection and enhance clustering performance, we implemented an iterative feature elimination strategy guided by silhouette coefficient analysis. Initially, the dataset was projected into two-dimensional space using UMAP for dimensionality reduction. The contribution of each feature to UMAP axes was quantified using Pearson correlation, and feature importance was ranked by summing absolute correlation values across both dimensions (“threshold_sum”). Features were then iteratively removed in ascending order of threshold_sum, starting from the least contributive. At each step, clustering quality was reassessed using the silhouette coefficient, which compared UMAP-derived clusters with biologically defined tissue labels. The silhouette coefficient ranges from −1 to +1. Following established guidelines, values above 0.70 are generally considered to indicate strong and well-separated cluster structure, values between 0.50 and 0.70 reflect reasonable but moderate separation, and values below 0.25 suggest poorly defined or overlapping clusters²³. Iterative removal was terminated once the feature set dropped below 30 variables or when further elimination reduced clustering performance. The optimal feature subset was defined as the configuration yielding the highest average silhouette score, corresponding to maximal separation between tissue-specific clusters in the reduced-dimensionality space. Cluster separation was further validated by two complementary resampling procedures applied to the UMAP embeddings of each feature panel (all regions, DMRs, and silhouette-selected panels). For the non-parametric bootstrap, samples were resampled with replacement to generate B = 1000 pseudo-datasets matched in size to the original cohort, UMAP projection and silhouette computation were repeated for each replicate, and the mean silhouette together with the 95% bootstrap confidence interval were reported. For the label-permutation test, tumor/normal labels were randomly shuffled without replacement while keeping the UMAP embedding fixed, and the silhouette coefficient was recomputed for n = 1000 permutations to construct the null distribution.

Correlation heatmap and co-expression network analysis

For both tumor and matched normal thyroid tissue samples, we calculated all pairwise Pearson correlation coefficients among 37 selected epigenetic regulators, comprising 11 thyroid cancer hub genes, 7 genes associated with 5mC, 9 with 5hmC, and 10 with 6mA. Correlation patterns were visualized as hierarchical clustered heatmaps generated with complete linkage clustering. Triangular correlation matrices were rendered on a blue–white–red gradient, with annotation tracks added to distinguish functional gene groups. To further explore gene–gene relationships, we constructed co-expression networks by connecting gene pairs with absolute correlation coefficients above a defined threshold. Networks were visualized using Kamada–Kawai force-directed layouts, where nodes were colored by functional category and edges were weighted proportionally to |Pearson’s r | ², thereby highlighting both functional grouping and correlation strength.

Statistics and reproducibility

All statistical analyses were performed using R (v4.3.3). Technical replicates were defined as independent library preparations and sequencing runs from the same DNA sample, which were subsequently merged during bioinformatic processing to increase sequencing depth. No custom code was used; all analyses relied on publicly available software packages as described in Methods.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Combined epigenomic profiling reveals distinct methylation landscapes in PTC

Using our recently developed multiplex and multifactorial DNA modification sequencing method, mDIP-seq, we first established a multidimensional epigenomic landscape to profile genome-wide distributions of 5mC, 5hmC, and 6mA across 12 matched tumor-normal tissue pairs from six treatment-naïve patients with PTC. This generated 36 datasets (12 pairs × 3 modifications) (Fig. 1A). Hierarchical clustering revealed that samples segregated primarily by DNA modification type rather than tissue origin (tumors vs. normal tissue). Correlation analysis supported this observation, showing strong inter-sample concordance for 5hmC and 6mA, whereas 5mC exhibited minimal correlation with either mark (Fig. 1B). UMAP analysis confirmed this modification-specific clustering pattern transcending the tumor-normal dichotomy (Fig. 1C). Despite clustering being dictated mainly by modification type, tumor tissues consistently exhibited systematic shifts from their matched normal counterparts within each modification group, indicating malignancy-associated remodeling across all three epigenetic landscapes. Silhouette coefficient analysis quantified moderate clustering robustness across modifications (mean coefficient = 0.56), with 5mC profiles exhibiting the highest distinctiveness (Supplementary Fig. 1A). Genome-wide density mapping revealed divergent genomic distributions: 5mC was predominantly enriched in promoter-proximal regions (±1 kb from transcription start sites, TSSs), consistent with its role in direct transcriptional regulation²⁶. Conversely, 5hmC and 6mA were preferentially enriched in introns (Supplementary Fig. 1B). Analysis of transcription factor (TF)-binding sites reinforced this divergence, showing strong 5mC enrichment near TSSs (Supplementary Fig. 1C). These findings suggest that while 5mC contributes directly to promoter activity and transcription initiation, 5hmC and 6mA may participate in modulating transcriptional elongation.

Fig. 1. Integrated epigenomic profiling reveals distinct methylation landscapes in PTC.

A Schematic workflow for epigenomic profiling of thyroid cancer, including experimental design and analysis pipelines. B Correlation heatmap with hierarchical clustering analysis of DNA modifications across all samples. C UMAP plot illustrating the spatial patterns of methylation modifications in tumor (filled circles) and normal (hollow circles) tissues, colored by methylation type (5mC: green; 5hmC: purple; 6mA: yellow). D UMAP projection of tumor-specific 5mC modifications for 5mC-specific clustering (tumor: red; normal: blue). E UMAP projection of tumor-specific 5hmC modifications for 5hmC-specific clustering. F UMAP projection of tumor-specific 6mA modifications for 6mA-specific clustering. G Combined UMAP analysis integrating all three modifications (5mC, 5hmC, 6mA). H IGV visualization of 5mC signals of PDLIM7, 5hmC signals of SPI1, and 6mA signals of CRABP1 in the tumor and normal thyroid tissue. UMAP embeddings in (C–G) were computed using the DiffBind-derived consensus 400-bp region matrix (sample × region, normalized read counts; hg38) as input.

Modification-specific tumor–normal differences varied in discriminative capacity. Tumor-specific 5mC modifications formed spatially discrete clusters (Fig. 1D), supported by a high silhouette coefficient (0.64), indicating robust tumor-normal discrimination (Supplementary Fig. 1D). This stark divergence aligns with 5mC’s established role as a key regulator of CpG island hypermethylation in oncogenic pathways²⁷. In contrast, 5hmC displayed diffuse spatial patterns (Fig. 1E) and poor clustering efficacy (silhouette coefficient=0.17, Supplementary Fig. 1E), reflecting its limited discriminative power in PTC. Meanwhile, 6mA exhibited intermediate discriminative ability, with partial spatial gradients in tumor tissues (Fig. 1F) and moderate clustering validity (silhouette coefficient = 0.42, Supplementary Fig. 1F). Combined analysis of all three marks (Fig. 1G) produced a silhouette coefficient of 0.61 (Supplementary Fig. 1G), marginally lower than that of 5mC alone (0.64). This slight reduction reflects feature redundancy between 5hmC and 6mA, consistent with their high intercorrelation. Nonetheless, their integration revealed complementary contributions, suggesting context-dependent contributions in PTC epigenome remodeling. Additionally, genome-wide peak distributions for each modification were visualized across replicates, allowing comparative assessment of methylation profiles (Supplementary Fig. 2). IGV snapshots confirmed the tumor-specific alterations of modifications on several typical cancer-related genes: PDLIM7 (increased 5mC), SPI1 (reduced 5hmC), and CRABP1 (increased 6mA) (Fig. 1H). These locus-specific signatures support the utility of multidimensional epigenomic profiling for identifying candidate biomarkers in thyroid cancer.

Differential DNA modification signatures and functional implications in PTC

To resolve cancer-specific epigenomic changes, we identified tumor-specific differentially methylated regions (DMRs) for 5mC, 5hmC, and 6mA (FDR < 0.05, |log₂ fold change | > 1). Figure 2A–C presents the volcano plots for 5mC, 5hmC, and 6mA DMRs, respectively, with red indicating upregulation and blue indicating downregulation. Notably, 5mC exhibited a higher number of upregulated regions (Fig. 2A), while 5hmC showed the opposite trend with more downregulated regions (Fig. 2B). 6mA demonstrated both upregulation and downregulation, but with a predominance of upregulated regions (Fig. 2C). These modification-specific trends suggest divergent roles in PTC progression. Heatmap visualization of DMRs confirmed robust separation of tumor and normal tissues across all marks (Fig. 2D–F). Dimensionality reduction highlighted the enhanced discriminative power of DMRs: silhouette coefficients increased from 0.64 (global 5mC), 0.42 (6mA), and 0.17 (5hmC) to 0.71, 0.67, and 0.92, respectively (Fig. 2G–I; Table 2). Strikingly, integration of all three marks yielded the highest classification accuracy (silhouette = 0.95; Fig. 2J), underscoring their combined potential as diagnostic biomarkers. These findings collectively demonstrate that context-specific DMRs, rather than global methylation patterns, drive the epigenomic divergence in PTC, with combinatorial analysis offering enhanced resolution for tumor stratification. Analysis of differential modification distributions revealed significantly stronger enrichment patterns of DMRs compared to the baseline genomic profiles (Supplementary Fig. 1B, C). To characterize the dominant genome-wide distribution patterns of these differential modifications, we compared the genomic locations of 5mC, 5hmC, and 6mA DMRs against genome-wide background distributions generated using randomly selected genomic regions (Fig. 2K, L). This analysis demonstrated that differential 5mC alterations exhibit the most prominent enrichment relative to background. Specifically, 5mC DMRs showed a statistically significant increase within the 0–1 kb window flanking the TSS (Transcription Start Site), consistent with a strong association with promoter-proximal regions. In contrast, the distributions of both differential 5hmC and differential 6mA regions near the TSS were substantially similar to the background distribution, showing no significant enrichment relative to background in this promoter-proximal window. Similarly, transcription factor binding site (TFBS) density analysis confirmed the strongest colocalization near the TSS occurs with differential 5mC, while both differential 5hmC and 6mA show minimal TFBS association in this region relative to background.

Fig. 2. Differential DNA modifications and integrated epigenomic signatures in PTC.

Volcano plots of DMRs for 5mC (A), 5hmC (B), and 6mA (C) (FDR < 0.05, |log₂FC | > 1). DMRs are plotted as log₂ fold change (x-axis) versus −log₁₀(FDR) (y-axis). Hypermethylated regions are highlighted in red and hypomethylated regions in blue. Heatmaps of z-score-normalized methylation levels at DMRs for 5mC (D), 5hmC (E), and 6mA (F) across tumor and normal thyroid samples. Red indicates high methylation; blue indicates low methylation. UMAP projections showing clustering of tumor (red) and normal (blue) samples based on DMRs for 5mC (G), 5hmC (H), and 6mA (I). J Integrated UMAP projection combining DMR features from 5mC, 5hmC, and 6mA for tumor (red) and normal (blue) samples. K Stacked bar plot showing the genomic distribution of regions harboring differential 5mC, 5hmC, or 6mA modifications in PTC samples and the distribution of size-matched randomly sampled genomic regions (hg38) used as background. L Genomic distribution of differentially methylated regions (DMRs) for 5mC/5hmC/6mA relative to TSS and the distribution of size-matched randomly sampled genomic regions (hg38) used as background.

Table 2.

Comparative analysis of feature selection performance across genomic regions

	All regions		Differential regions		Selected regions
	Silhouette score	No. of features	Silhouette score	No. of features	Silhouette score	No. of features
5mC	0.64	200,084	0.71	5857	0.91	80
5hmC	0.17	169,998	0.92	879	0.96	64
6mA	0.42	671,630	0.67	381	0.90	43
All	0.61	1,041,712	0.95	7117	0.90	187

Gene Ontology (GO) enrichment analysis revealed distinct biological processes associated with differentially methylated regions. In 5mC-associated regions, genes were enriched in immune regulatory pathways, including negative regulation of interferon-alpha/beta production and regulation of the pentose-phosphate shunt (Fig. 3A). These findings suggest that 5mC dysregulation may facilitate immune evasion by silencing interferon-driven antitumor responses, while simultaneously enhancing metabolic reprogramming to sustain redox balance and proliferation. De novo motif analysis further identified enrichment of IRF5, a key interferon signaling regulator²⁸, and ZFP57, a genomic imprinting regulator²⁹ reinforcing the roles of 5mC in immune suppression and metabolic control (Fig. 3B). For 5hmC, GO terms highlighted developmental processes and negative regulation of wound healing (Fig. 3C), consistent with aberrant activation of embryonic programs or dysregulated tissue repair mechanisms during dedifferentiation and stromal remodeling. Corresponding motif enrichment for HOXA1 and RNA Binding Motif Protein 5 (RBM5) (Fig. 3D) implicates aberrant developmental signaling and RNA-processing defects as drivers of invasive phenotypes^30,31. In contrast, 6mA-modified regions were enriched for genes regulating cell differentiation and biosynthetic processes (Fig. 3E), implicating this mark in the loss of thyroid-specific differentiation and hyper-activation of anabolic pathways to fuel proliferation. Motif analysis revealed enrichment of FEZF2 and ZBTB6 (Fig. 3F), transcriptional regulators associated with stemness maintenance and cell cycle dysregulation^32,33, along with pervasive poly-T tracts that may contribute to genomic instability and telomere dysfunction. Collectively, these distinct methylation-specific signatures align with hallmarks of thyroid oncogenesis, including immune suppression, metabolic reprogramming, dedifferentiation, and proliferative dysregulation.

Fig. 3. Integrative analysis of Gene Ontology enrichment and de novo motifs associated with differential DNA modification regions in PTC.

A GO Biological Process enrichment for genes linked to 5mC differential regions, visualized as horizontal bar plots (top 6 terms) using rGREAT (fold enrichment mapped to x-axis). B De novo motif enrichment analysis of 5mC regions, displaying top-ranked transcription factor binding motifs. C GO Biological Process enrichment for 5hmC-associated genes, showing horizontal bar plots of the six most enriched terms. D Top-ranked de novo motifs identified in 5hmC differential regions, ranked by statistical significance. E GO Biological Process enrichment for 6mA-modified regions, illustrated as horizontal bar plots of the six highest-ranked terms. F De novo motif analysis of 6mA regions, highlighting the top three enriched motifs.

Silhouette-guided feature selection optimizes tumor-normal discrimination

To address the challenge of high-dimensional epigenomic datasets, we applied an iterative silhouette-guided feature selection strategy (Fig. 4A). Initial DMR-based signatures achieved strong but feature-heavy discriminative power, with silhouette scores of 0.71 for 5mC (5857 loci), 0.92 for 5hmC (879 loci), and 0.67 for 6mA (381 loci), while the combined 7117-locus panel reached 0.95 (Table 2). Although effective, such high feature counts limit both interpretability and clinical translation. Feature reduction substantially improved clustering fidelity while retaining biological resolution. Optimized panels retained only 80 (5mC), 64 (5hmC), and 43 (6mA) loci. This refinement significantly improved silhouette scores across individual modifications, 5mC increased by 28% (from 0.71 to 0.91), 5hmC by 4.3% (0.92–0.96), and 6mA by 34% (0.67–0.90). UMAP visualizations confirmed the preserved or enhanced separation using selected features for individual modifications (Fig. 4B–D) and their combination (Fig. 4E). Quantitative silhouette analysis further demonstrated superior separation efficiency compared to both whole-genome and DMR-based approaches (Fig. 4F). Notably, the combined 187-feature panel preserved nearly the same discriminative capacity as the full 7,117-DMR set (silhouette = 0.90 vs. 0.95), while reducing feature complexity by 38-fold, the full details of the 187 loci used in the diagnostic panel have now been included as Supplementary Data 2. To assess the robustness of these results given the limited sample size, we performed bootstrap resampling (B = 1000) and permutation testing (n = 1000) on UMAP embeddings for each panel. All four panels achieved permutation pvalues ≤ 0.004 with narrow bootstrap 95% confidence intervals, confirming that the observed cluster separation significantly exceeds chance expectation (Supplementary Table 3). Furthermore, individual-level log₂(Tumor/Normal) enrichment ratios computed for each of the six patients across the selected regions revealed highly concordant directional changes, indicating that the identified loci reflect reproducible tumor-associated epigenetic alterations rather than patient-specific outliers (Supplementary Fig. 3). This striking balance between parsimony and accuracy underscores the power of silhouette-guided optimization, positioning it as a practical framework for both mechanistic interrogation and clinical translation of epigenomic biomarkers.

Fig. 4. Silhouette-guided feature selection and genomic patterning of selected regions.

A Schematic workflow for iterative feature selection and dimensionality reduction based on silhouette score analysis. UMAP projections showing tumor-normal discrimination using selected 5mC (B), 5hmC (C), and 6mA (D) regions. E UMAP projection using combined selected regions of 5mC, 5hmC, and 6mA. F Average silhouette scores quantifying cluster separation for tumor/normal samples, comparing models using all regions, DMRs, and silhouette-selected regions for individual modifications (5mC, 5hmC, 6mA) and their combination. G Stacked bar plot showing the genomic distribution of regions harboring selected 5mC, 5hmC, or 6mA modifications in PTC samples and the distribution of size-matched randomly sampled genomic regions (hg38) used as background. H Genomic distribution of selected methylation regions for 5mC/5hmC/6mA relative to TSS and the distribution of size-matched randomly sampled genomic regions (hg38) used as background. I Circos plot visualizing the chromosomal distribution (outer ring: chromosome ideograms) and intra-chromosomal localization of silhouette-selected regions for DNA modifications. Progressing inward from the outermost ring: the distribution of selected 5mC regions, 5hmC regions, and 6mA regions on each chromosome is shown. J Heatmap depicting chromosome-specific enrichment scores (fold change versus genome-wide background proportion) for selected regions of 5mC, 5hmC, and 6mA modifications.

We next characterized the genomic distribution of the silhouette-optimized DMRs. Strikingly, selected 5mC regions were predominantly promoter-proximal, with 31.25% localized within ≤1 kb of TSSs and 43.75% within promoter regions (≤3 kb). In sharp contrast, 5hmC regions showed minimal promoter localization (1.56% within ≤1 kb), instead clustering in distal intergenic regions (45.31%) and intronic regions (43.75%). The 6mA regions displayed an intermediate pattern, with 20.93% situated in promoter-proximal regions (≤2 kb) and 39.53% within introns (Fig. 4G). Consistent with these spatial distributions, TFBS analysis revealed marked enrichment of 5mC regions within the 0–1 kb TSS window, minimal enrichment for 5hmC at promoters but notable density at distal regulatory elements, and moderate enrichment for 6mA in promoter-adjacent regions, reinforcing its role in transcriptional modulation at structurally defined regulatory zones (Fig. 4H). Chromosomal mapping revealed non-random patterns of modification across the genome (Fig. 4I). While 5mC, 5hmC, and 6mA exhibited divergent dysregulation patterns, they shared significant enrichment biases toward specific chromosomes (Fig. 4J). Notably, chromosome 16 showed 2.1-fold (5mC), 5.3-fold (5hmC), and 4.0-fold (6mA) enrichment compared to the genomic background. Similarly, chromosome 19 displayed 7.9-fold (5mC), 4.9-fold (5hmC), and 2.5-fold (6mA) enrichment. Both chromosomes are characterized by exceptionally high CpG island and gene density³⁴, suggesting that their chromatin architecture provides a permissive environment for coordinated deposition of multiple methylation marks. This convergence implies that chromosome-specific features—such as nuclear positioning, local chromatin compaction, or clustering of co-regulated genes—may create privileged landscapes for concurrent DNA modification alterations in thyroid tumorigenesis.

Integration of DNA modification profiles with transcriptional and mutational landscapes

Systematic integration of TCGA-THCA RNA-seq data, somatic mutation profiles, and locus-specific DNA modification landscapes (5mC, 5hmC, 6mA) revealed multi-layered interactions between epigenetic reprogramming, transcriptional dysregulation, and genomic instability in PTC. Transcriptome-wide differential expression analysis revealed extensive dysregulation, with 1144 genes significantly upregulated and 552 downregulated (|log₂FC | > 1, FDR < 0.05), indicating pervasive transcriptional remodeling in tumors (Fig. 5A). Intersection analysis of epigenetic modifications with differentially expressed genes (DEGs) revealed modification-specific transcriptional associations: selected 5mC-modified regions were exclusively enriched at promoters of upregulated oncogenic loci (PDLIM7, PXDN, GABRG3, SAMD11; Fig. 5B). In contrast, selected 5hmC-modified regions overlapped both upregulated (CELF4, CEP170B, ACTN1, ADAMTS12) and downregulated tumor suppressors (NFATC1, CASZ1; Fig. 5C), whereas selected 6mA-modified regions correlated with dysregulation of genes involved in signal transduction and metabolism. This included upregulation of GPRC5B, CARNS1, DMD, MAP2, and RENBP, alongside downregulation of POU2F2 and CRABP1 (Fig. 5D). Mutation profiling identified BRAF as the predominant driver mutation (59% prevalence), with RAS family mutations (NRAS: 8%, HRAS: 3%) occurring in a nearly mutually exclusive manner (Fig. 5E).

Fig. 5. Integrative analysis of transcriptional, mutational, and epigenetic landscapes in PTC.

A Volcano plot of differentially expressed genes (DEGs) from the TCGA-THCA cohort, with upregulated genes in red and downregulated genes in blue. Venn diagrams showing the overlap between genes associated with selected 5mC (B), 5hmC (C), and 6mA (D) regions and DEGs from RNA-seq data. (E) Bar plot of the top 10 mutated genes in TCGA-THCA samples. F Hierarchical clustering analysis of the union of 17 genes associated with selected 5mC, 5hmC, and 6mA regions and DEGs from RNA-seq data, demonstrating the ability to broadly distinguish tumor from normal tissues into two major clusters. Box plots of log2 fold changes relative to wild type, grouped by BRAF (G) or RAS (H) mutation status. Gray represents wild-type samples, and colored represents mutant samples. Statistical significance is denoted by asterisks: *** indicates p < 0.001, ** indicates p < 0.01, * indicates p < 0.05, and “NS” indicates not significant. The p values were calculated using a Wilcoxon rank-sum test comparing log₂ fold changes between wild-type and mutant samples.

Hierarchical clustering of the 17 modification-associated DEGs revealed two molecularly distinct subtypes. Cluster 1, composed exclusively of tumor samples, was significantly enriched for BRAF-mutant cases, whereas Cluster 2, a mixed tumor–normal group, was predominantly associated with RAS-altered tumors (NRAS, KRAS, and HRAS alterations), indicative of robust intratumoral heterogeneity (Fig. 5F). Mutation-stratified differential expression analysis revealed distinct transcriptional programs. In BRAF-mutant tumors, 15 of 17 signature genes were significantly dysregulated compared with BRAF-wild-type counterparts (Wilcoxon rank-sum test p < 0.05), with PXDN and GABRG3 being non-significant exceptions (Fig. 5G, Supplementary Fig. 4). Among these, 13 genes showed marked upregulation, while NFATC1 exhibited paradoxical downregulation. Conversely, RAS-mutant tumors displayed significant expression changes in 13 of 17 genes relative to RAS-wild-type samples (Wilcoxon p < 0.05), with PXDN, CARNS1, NFATC1, and RENBP lacking statistical significance (Fig. 5H, Supplementary Fig. 5). The results are consistent with previous studies, which indicate that BRAF and RAS mutations play significant but distinct roles in the MAPK pathway^2,35. Specifically, BRAF mutations are associated with aggressive clinicopathological features through activation of extracellular matrix (ECM)-related genes, whereas RAS mutations are linked to suppression of immunoglobulin-related genes, highlighting divergent oncogenic programs in PTC³⁶.

Transcriptional correlation and co-expression network analysis reveals distinct gene regulatory patterns in PTC and normal thyroid tissue

In our investigation of the transcriptional landscape of PTC and normal thyroid tissue, we constructed correlation heatmaps and co-expression networks for 37 epigenetic regulators, comprising 7 genes linked to 5mC, 9 genes linked to 5hmC, 10 genes linked to 6mA and 11 established thyroid cancer hub genes. The 26 methylation-associated regulators were directly identified through integrative analysis of our epigenomic data (Figs. 3, 5B–D), filtered by: (i) Functional enrichment of de novo motifs in differentially methylated regions (DMRs), and (ii) Significant overlap between selected methylated regions and differentially expressed genes (DEGs). The 11 thyroid cancer hub genes represent consensus PTC drivers identified through TCGA-THCA and established genetics studies^4,37. In tumor tissue, the correlation heatmap (Fig. 6A and Supplementary Fig. 6) revealed a distinct clustering pattern in which thyroid hub genes aggregated closely together, indicative of coordinated transcriptional regulation. Co-expression networks constructed at correlation thresholds of 0.6 and 0.7 (Fig. 6B, C) further highlighted these strong gene–gene interactions, visualized using Kamada–Kawai force-directed layouts with node coloring by functional group and edge weights proportional to |r | ². Conversely, in normal thyroid tissue, hierarchical clustering produced a more dispersed arrangement of hub genes (Fig. 6D; Supplementary Fig. 7), suggesting weaker functional cohesion. Interestingly, the overall darker color gradient in the normal tissue heatmap reflected higher absolute correlation values, pointing to tighter local gene–gene associations despite the lack of clear hub clustering. Corresponding co-expression networks (Fig. 6E, F) revealed a distinct configuration of nodes and edges compared to tumors, underscoring divergent transcriptional wiring between malignant and non-malignant thyroid tissue.

Fig. 6. Correlation heatmaps and co-expression networks in PTC tumor and normal thyroid tissues.

A Correlation heatmap of 37 epigenetic regulators in PTC tumor tissue, hierarchically clustered using complete linkage. Genes are color-coded by function: thyroid hub genes (blue), 5mC-related genes (green), 5hmC-related genes (purple), and 6mA-related genes (orange). Correlation coefficients are depicted in a blue-white-red gradient, with red for positive, blue for negative, and white for no correlation. Co-expression networks of tumor tissue genes at correlation thresholds of >0.6 (B) and >0.7 (C). Nodes represent genes (color-coded as in A), and edge thickness corresponds to the square of the absolute Pearson’s r value. D Correlation heatmap of 37 epigenetic regulators in normal thyroid tissue, hierarchically clustered using complete linkage. Genes are color-coded as in Fig. 6A. Correlation coefficients are depicted in a blue-white-red gradient. Co-expression networks of normal tissue genes at correlation thresholds of >0.6 (E) and >0.7 (F), using the same node and edge representations as in (B–C), highlighting stronger network connectivity compared with tumor tissue.

Table 3 summarizes the connectivity of co-expression networks in tumor and normal thyroid tissues across a range of correlation thresholds. In tumors, connectivity declined sharply as the threshold increased, from 36 nodes and 148 edges at r > 0.3 to only 7 nodes and 5 edges at r > 0.7, with complete loss of connectivity at r > 0.9. By contrast, normal thyroid tissue consistently exhibited greater network density, reaching 37 nodes and 228 edges at r > 0.3 and retaining 15 nodes and 19 edges even at r > 0.7. Notably, normal tissue maintained modest but persistent connectivity at the most stringent thresholds (r > 0.8 and r > 0.9), underscoring the resilience of its transcriptional regulatory networks. At the level of individual gene pairs, the most prominent disruption involved the immune-related module comprising SPI1, POU2F2, IRF5, and HOXA1. All four genes were tightly coordinated in normal tissue but became substantially decoupled in tumors, with the POU2F2–IRF5 correlation declining from 0.91 to 0.47. The 6mA-associated gene RENBP exhibited similarly striking changes, losing all its strong connections in tumors, while 5hmC-associated pairs such as ACTN1–NFATC1 and CEP170B–CASZ1 also lost their co-expression relationships. Thyroid cancer hub gene connections including PIK3CA–NF1, BRAF–NF1, and KRAS–NRAS persisted in both conditions but showed consistent attenuation, and HRAS maintained strong negative correlations with PIK3CA and NF1 across both tissue types. Conversely, several tumor-specific connections emerged that were absent in normal tissue, notably MAP2–CEP170B, MAP2–CELF4, and PXDN–ADAMTS12, indicating the establishment of new transcriptional relationships among 5hmC- and 6mA-associated genes during tumorigenesis. These contrasting patterns highlight fundamental differences in transcriptional architecture: tumor networks are fragmented and threshold-sensitive, whereas normal networks are robust and cohesive. Such divergence in epigenetic regulator interactions may underlie the destabilized transcriptional landscape of PTC and point toward therapeutic opportunities to restore regulatory coherence.

Table 3.

Comparison of co-expression network connectivity in tumor and normal thyroid tissues across different correlation thresholds

	Network in tumor		Network in normal
	No. of nodes	No. of edges	No. of nodes	No. of edges
r > 0.3	36	148	37	228
r > 0.4	31	81	35	139
r > 0.5	22	46	27	77
r > 0.6	17	18	23	46
r > 0.7	7	5	15	19
r > 0.8	2	1	10	7
r > 0.9	0	0	3	2

Discussion

Papillary thyroid carcinoma (PTC), the most common endocrine malignancy, is primarily driven by genetic alterations in the MAPK pathway^2,37. However, the broader epigenomic landscape, particularly the coordinated dysregulation of multiple DNA modifications, remains poorly defined. To address this gap, we present the first combined analysis of 5mC, 5hmC, and 6mA in PTC tumorigenesis to the best of our knowledge. Our findings reveal modification-specific genomic distributions and functional partitioning, enhanced diagnostic value of combinatorial epigenetic signatures, and mutation-context dependencies in epigenetic-transcriptional coordination.

The distinct spatial distribution of DNA modifications suggests their specialized roles in pathogenesis. Tumor-specific 5mC hypermethylation was preferentially localized to promoter-proximal regions, aligning with the established role of promoter hypermethylation in silencing tumor suppressor genes³⁸. Furthermore, motif enrichment analysis of hypermethylated promoters revealed co-enrichment of IRF5 and ZFP57 binding sites, implicating two concurrent pathogenic mechanisms: impaired IRF5-dependent immune signaling and dysregulated ZFP57-mediated imprinting maintenance. Consistent with global 5hmC depletion reported in multiple cancers, including thyroid neoplasms^19,39, we observed a significant loss in 5hmC levels in PTC tumors. This depletion was preferentially localized within intronic and distal genomic regions enriched for TFBSs, which normally act as enhancer elements silenced by 5hmC⁴⁰. Such preferential loss suggests disruption of enhancer silencing and consequent activation of aberrant gene expression programs. Furthermore, the association of depleted regions with HOXA1 and RBM5 motifs implicates 5hmC loss in abnormal dedifferentiation and proliferation. The hybrid distribution of 6mA, enriched in both promoter-proximal and intronic regions, together with FEZF2/ZBTB6 motif associations, suggests dual roles in transcriptional regulation and stemness maintenance, consistent with observations in other malignancies. Our silhouette-guided feature selection demonstrated that combinatorial epigenetic signatures substantially enhance tumor-normal discrimination compared to individual modifications. The optimized 187-locus panel achieved near-equivalent stratification efficiency relative to the full differential methylation region set while reducing feature complexity by 38-fold, thereby enhancing clinical feasibility. Chromosomes 16 and 19 emerged as epigenetic hotspots with 2.1–7.9-fold enrichment across all modifications. These findings address a major limitation of prior single-modification studies in thyroid cancer and establish a framework for the development of biomarkers.

Integration with transcriptional and mutational data revealed context-specific epigenetic reprogramming. Whereas BRAF mutations drove strong activation of genes linked to epigenetic modifications, RAS mutations produced distinct patterns of dysregulation. This aligns with known MAPK pathway heterogeneity but newly implicates 5hmC and 6mA in modulating mutation-specific transcriptional programs. Co-expression network analysis further revealed fundamental architectural differences: normal thyroid tissue maintained robust connectivity even at high correlation thresholds, whereas tumor networks fragmented significantly, indicating epigenetically driven loss of transcriptional coordination. Several observations argue that this network fragmentation reflects targeted epigenetic deregulation rather than general genomic instability. PTC is a genomically stable tumor characterized by low somatic mutation burden and infrequent copy number alterations^4,41, making stochastic connectivity loss an unlikely explanation. Indeed, the disruption was markedly selective rather than uniform: the immune-related transcription factor module centered on SPI1, POU2F2, and IRF5 was among the most severely disrupted, consistent with the differential 5mC and 5hmC patterns identified at these loci^10,42, while the 6mA-associated gene RENBP lost nearly all of its strong connections in tumors. In contrast, core MAPK/PI3K hub gene connections persisted in both conditions, and HRAS maintained consistent negative correlations with PIK3CA and NF1 across both tissue types^2,37. Conversely, entirely new co-expression relationships emerged among 5hmC- and 6mA-associated genes such as MAP2, CEP170B, CELF4, and PXDN–ADAMTS12, indicating that epigenetic reprogramming in PTC does not merely dismantle existing regulatory networks but actively establishes new transcriptional relationships. This pattern of selective module disruption coupled with de novo module formation is inconsistent with random genomic instability but consistent with targeted epigenetic reprogramming, particularly given that the disrupted genes were directly identified through DMR-associated motif enrichment and DEG overlap in our epigenomic analysis. Together, these findings suggest that multiple DNA modifications converge to reshape the transcriptional architecture of PTC through the simultaneous disruption of normal regulatory circuits and the formation of tumor-specific co-expression modules.

This approach is particularly well-suited for large-scale cohort analysis encompassing hundreds of samples. The strong discriminatory power of the panel revealed in this study, combined with the scalability of the underlying technology, underscores the significant potential of such combinatorial epigenetic signatures in future clinical diagnostics and biomarker screening applications. From a translational standpoint, the 187-locus panel is amenable to clinical implementation through multiple avenues. Circulating tumor DNA methylation profiling has shown promise for non-invasive thyroid cancer detection, with methylation signatures detectable in plasma capable of distinguishing PTC from healthy controls⁴³. The panel loci could also be converted into targeted methylation-specific PCR assays suitable for fine-needle aspiration biopsy specimens⁴⁴. Notably, Oxford Nanopore sequencing enables simultaneous, direct detection of 5mC, 5hmC, and 6mA from native DNA without bisulfite conversion⁴⁵, aligning precisely with the multi-modification architecture of our panel and offering a streamlined path toward clinical deployment. Study limitations include modest sample size and unresolved causality between epigenetic alterations and tumor initiation. Future research should investigate the dynamic interplay of DNA modifications during progression, functionally validate hub genes using CRISPR-based approaches, assess the conservation of chromosomal hotspots across malignancies, construct a pan-epigenome atlas across expanded PTC cohorts to systematically resolve epigenetic regulatory hierarchies and tumor heterogeneity, and validate the diagnostic and early detection potential of this optimized epigenetic signature in circulating cell-free DNA (cfDNA). Collectively, this multi-omics analysis establishes PTC as a paradigm of coordinated epigenetic dysregulation. By delineating modification-specific logic, diagnostic synergies, and mutation-context dependencies, our findings provide a roadmap for advancing epigenetic targeting strategies in thyroid oncology.

Author contributions

Q.L. and J.Y. contributed equally to this work. F.F., M.Y., and W.X. conceived and supervised the study. Y.W. contributed to data analysis. Q.L. and J.Y. performed the experiments and bioinformatic analyses. All authors contributed to experimental design, data interpretation, and manuscript preparation.

Peer review

Peer review information

. Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Dr. Nilanjan Banerjee.

Funding

The study was supported by the National Natural Science Foundation of China (62402343); the Key University Science Research Project of Anhui Province (2025AHGXZK30886, 2025AHGXZK40699, 2023AH040267); Basic Research Center for Agricultural Frontiers and Interdisciplinary Sciences (BRC-AFIS), Innovation Program of Chinese Academy of Agricultural Sciences (CAAS-BRC-AFIS-2025-02); the Health Science Research Project in Anhui Province (AHWJ2023A10147).

Data availability

The raw sequence data reported in this paper have been deposited in the Genome Sequence Archive²⁴ at the National Genomics Data Center²⁵, China National Center for Bioinformation / Beijing Institute of Genomics, Chinese Academy of Sciences (GSA-Human: HRA013138), accessible at https://ngdc.cncb.ac.cn/gsa-human. Source data underlying the graphs and charts in the main figures are provided as Supplementary Data. All other data supporting the conclusions of this study are available from the corresponding authors upon reasonable request.

Code availability

No custom code was developed for this study. All analyses were performed using publicly available software packages, including Trim Galore (v0.6.10), Bowtie2 (v2.5.1), SAMtools (v1.6), Picard (v2.27.5), MACS2 (v2.2.7.1), deepTools (v3.5.1), DiffBind (v3.10.1), ChIPseeker (v1.36.0), HOMER (v5.1), rGREAT (v2.2.0), and R (v4.3.3).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s42003-026-10495-3.