This study was conducted at Zhoupu Hospital, Shanghai, China, between August 2023 and January 2025. A total of 120 Han Chinese individuals were enrolled, including 60 patients diagnosed with GD and 60 age- and sex-matched healthy controls. Participants were recruited from the Department of Endocrinology and the Physical Examination Center. Baseline demographic and clinical characteristics are summarized in Table 1 .

To investigate disease phenotypes, GD patients were further stratified into two subgroups: (1) NGD, defined as patients with first-time diagnosis of hyperthyroidism and no prior treatment; and (2) RGD, defined as patients with a disease duration >2 years, persistent or recurrent hyperthyroid symptoms, and positive TRAb levels (≥1.5 IU/L) despite standard treatment. Clinical data collected included age at onset, thyroid size, presence of ophthalmopathy, smoking status, and family history of hyperthyroidism (defined as having at least one first-degree relative affected). Drug history was recorded for methimazole (MMI), and patients were categorized according to their daily dosage (≥5 mg/day or <5 mg/day).The study was approved by the Ethics Committee of Shanghai Health Medical College (Approval No. 2023-C-123-E01), and written informed consent was obtained from all participants.

GD diagnosis was based on clinical symptoms, laboratory findings, and thyroid ultrasonography. Fasting venous blood samples (10 mL, non-anticoagulated) were collected after ≥8 hours of fasting. Thyroid function was assessed using electrochemiluminescence (ECL) assays. Inclusion criteria for GD comprised: clinical manifestations of hyperthyroidism, suppressed serum TSH levels, positive TRAb, and ultrasonographic evidence of diffuse goiter.

Healthy controls were selected from individuals with normal physical and biochemical examination results and no history of chronic diseases. Control participants also provided fasting venous blood samples, and their thyroid peroxidase antibody (TPOAb) levels were assessed using ECL. Inclusion criteria for controls included TPOAb negativity and absence of a family history of autoimmune thyroid diseases (AITDs). Exclusion criteria for both groups included the presence of diffuse goiter or thyroid nodules >5 mm detected by ultrasonography, as well as a history or diagnosis of other autoimmune diseases.

PBMCs were isolated from 5.0 mL of venous blood collected in EDTA-K2 tubes after overnight fasting. Genomic DNA was extracted using the Relaxgene DNA Isolation Kit (Tiangen Biotech, China) following the manufacturer’s instructions. DNA concentration and purity were assessed using a NanoDrop 2000 spectrophotometer (Thermo Scientific, USA). Only samples meeting the following quality criteria were included: concentration ≥20 ng/μL, OD260/280 ratio between 1.7–1.9, and OD260/230 ≥2.0.

DNA methylation levels at the promoter regions of IL17, IL21, and IL22 genes were analyzed using the MethylTarget™ assay (Genesky Corporation, Shanghai, China). CpG-rich promoter regions (GC content >20%, length >200 bp) were selected, and two rounds of PCR were conducted per target gene. Methylation at each CpG site was quantified as the ratio of methylated to total cytosines. CpG site annotation was based on genomic coordinates. Primer sequences and PCR conditions are provided in Table 2 and Supplementary Table S1 .

To identify potential diagnostic biomarkers, four supervised machine learning algorithms were applied: Random Forest (RF), Support Vector Machine (SVM), Generalized Linear Model (GLM), and eXtreme Gradient Boosting (XGBoost). All models were implemented in R using the caret package (v7.0-1) with standardized preprocessing including mean-centering and near-zero variance filtering (23). Model performance was evaluated using repeated stratified 5-fold cross-validation (3 repetitions), with performance metrics averaged across all folds.

Model interpretability was assessed using Shapley additive explanations (SHAP), implemented via the DALEX R package (v2.4.3), to evaluate the relative importance of each CpG site based on cooperative game theory (24). Diagnostic performance was quantified using receiver operating characteristic (ROC) curves generated by the pROC package (v1.18.5), and the area under the curve (AUC) was interpreted as follows: <0.6, non-informative; 0.6–0.7, suboptimal; 0.7–0.9, moderate to strong; >0.9, excellent discrimination (25).

To explore the relationship between methylation profiles and clinical features of GD, patients were stratified by age, age at onset, sex, disease subtype, family history, presence of ophthalmopathy, smoking status, and thyroid function indices. FT3 and FT4 levels were dichotomized at their median values. TSH was categorized as ≤0.001 vs. >0.001 mIU/L, and TRAb as ≤1.5 vs. >1.5 IU/L. Variables with >50% missing data (e.g., T3, T4, Tg) were excluded.

Expression data were log2-transformed using the formula log2(X+1). Univariate binary logistic regression was performed using the glm function in R to identify significant predictors. Variables with p < 0.1 were included in multivariate logistic regression to identify independent associations, with significance set at p < 0.05. Statistical modeling was conducted using the rms (v6.4.0) and ResourceSelection (v0.3.5) R packages.

Methylation data are presented as mean±standard error of the mean (SEM). Normally distributed continuous variables are shown as mean±standard deviation (SD), while non-normally distributed data are expressed as median with interquartile range (IQR). Group comparisons were conducted using independent samples t-tests or Mann–Whitney U tests, as appropriate. Correlations between methylation levels and clinical variables were evaluated using Spearman’s rank correlation coefficient. Categorical variables were analyzed using Chi-square or Fisher’s exact tests.

Univariate binary logistic regression was performed to evaluate associations between clinical features and methylation subgroups, and odds ratios (ORs) with 95% confidence intervals (CIs) were calculated. Variables with p < 0.1 in univariate analysis were included in multivariate logistic regression to identify independent predictors (p < 0.05). All statistical analyses, including data preprocessing, descriptive statistics, and regression modeling, were conducted using SPSS version 26.0 (IBM Corp., Armonk, NY, USA) and R version 4.3.3 (R Core Team, Vienna, Austria).

To investigate DNA methylation alterations across different stages of Graves’ disease (GD), we enrolled 35 NGD patients, 25 RGD patients, and 60 NCs. PBMCs were isolated and subjected to DNA extraction. Following stringent quality control, promoter regions of the Th17-related cytokines IL17, IL21, and IL22 with GC content >20% were selected for methylation analysis.

Compared to NCs, the majority of CpG sites in GD patients exhibited significantly reduced methylation levels, with high intra-group consistency ( Supplementary Figures S1A-F ). Principal component analysis (PCA) of methylation profiles clearly distinguished GD patients from controls ( Figures 1A–C ), and revealed a significant overall decrease in methylation levels in GD patients (P=2.5 × 10^-7, Figure 1D ). Both NGD (P=1.4 × 10^-7) and RGD (P=6.6 × 10^-7) groups showed significantly lower methylation compared to NCs, with NGD exhibiting the most pronounced hypomethylation, suggesting stage-specific methylation dynamics ( Figures 1E–F ).

We further assessed gene-specific methylation of the IL17, IL21, and IL22 promoter regions. All three genes exhibited significantly lower promoter methylation levels in GD patients relative to NCs (p < 0.05, Figure 2A ). Similar reductions were observed in both NGD and RGD subgroups (p < 0.05, Figures 2B, C ).

To delineate CpG site-specific methylation patterns, we analyzed promoter regions >200 bp in length, containing 16 CpG sites for IL17, 10 for IL21, and 17 for IL22 ( Supplementary Table S2 ). The methylation levels of CpG sites for each gene are shown in Figures 3A–C . Site-level methylation analyses demonstrated significant hypomethylation in GD patients at 12 CpG sites in IL17 (p < 0.05, Figure 4A ), 9 sites in IL21 ( Figure 4B ), and 13 sites in IL22 ( Figure 4C ).

After excluding CpG sites without statistically significant methylation differences, the remaining CpG sites in the promoter regions of IL17, IL21, and IL22 were used to develop predictive models. Four machine learning algorithms were applied: RF, SVM, GLM, and XGB. Among them, the RF, SVM, and XGB models demonstrated superior performance in terms of lower residual distributions and higher classification accuracy based on ROC curve analysis ( Figures 5A–C ). Thus, these three models were selected for downstream analyses.

Four CpG sites—chr4_123542199_R, chr4_123542549_R, chr12_68647247_R, and chr12_68647735_R—were consistently identified by the three models and prioritized for further evaluation ( Figures 5D, E ).

The four CpG sites identified through model-based feature importance were subjected to ROC curve analysis to assess their diagnostic utility ( Figure 6 ). All four CpG sites demonstrated AUC values >0.7, indicating moderate diagnostic performance. Specifically, the AUC values for chr4_123542199_R, chr4_123542549_R, chr12_68647247_R, and chr12_68647735_R were 0.764, 0.754, 0.752, and 0.738, respectively ( Figures 6A–D ). These results suggest that hypomethylation at these sites may serve as potential biomarkers for GD diagnosis.

Clinical variable distributions are presented in Supplementary Table S4 . To investigate the clinical relevance of the four identified CpG sites, univariate logistic regression analyses were performed ( Supplementary Figure S2 ). Hypomethylation at chr4_123542549_R was significantly associated with elevated TRAb levels (OR=4.00, 95% CI: 1.30–12.33, p=0.02), and this association remained significant in multivariate analysis ( Figure 7 ; Supplementary Table S3 ). These findings suggest that demethylation at this site may influence TRAb production or activity, potentially contributing to disease pathogenesis and reflecting disease severity.

Furthermore, both univariate and multivariate analyses revealed a significant association between hypomethylation at chr12_68647247_R and increased FT4 levels (OR=0.29, 95% CI: 0.10–0.84, p=0.02, Figure 7 ; Supplementary Table S3 ). This indicates a possible regulatory role for this site in thyroid hormone synthesis or metabolism, suggesting its utility as a marker of thyroid functional status in GD.

No significant associations were observed between methylation levels at chr4_123542199_R or chr12_68647735_R and clinical parameters (p > 0.05, Supplementary Figure S2 ).