Peripheral blood mononuclear cell DNA methylation signatures guide surgical decision-making in indeterminate pulmonary nodules

Abstract

Background

Distinguishing malignant from benign pulmonary nodules remained a significant clinical challenge. Given the involvement of DNA methylation in anti-tumor immunity, we aimed to investigated whether DNA methylation patterns in peripheral blood mononuclear cells (PBMCs) could serve as non-invasive biomarkers for pulmonary nodule classification.

Methods

Genome-wide DNA methylation profiling was performed using Methylome-seq on PBMCs from Discovery cohort, including patients with benign pulmonary nodules (BPN), minimally invasive adenocarcinoma (MIA), and early-stage invasive adenocarcinoma (eIAC). Differential analysis identified 56 candidates differentially methylated cytosines (DMCs) and regions (DMRs). Subsequently, 2 methylation features were validated using targeted bisulfite sequencing (TBS) in the Train cohort and Test cohort. We further developed 2 machine learning models integrating the DNA methylation features and routine blood biomarkers to enhance pulmonary nodule classification accuracy. SurgMalig Model was a surgical malignancy classifier used to differentiate between benign nodules and malignant nodules requiring surgical resection. And ResectGuide Model was a resection guide classifier used to subdivide malignant nodules into MIA and eIAC to inform surgical planning.

Results

The validated methylation features demonstrate significant discriminatory ability for pulmonary nodules properties compared to conventional blood biomarkers. SurgMalig Model achieves high performance in distinguishing benign from malignant nodules, with area under the curve (AUC) of 0.926 in the Train cohort and 0.806 in the Test cohort. ResectGuide Model effectively classifies MIA and eIAC subtypes, yielding AUC of 0.862 and 0.843 in the respective cohorts.

Conclusions

PBMC methylation signatures provide clinically actionable tools for surgical triage in indeterminate pulmonary nodules.

Plain language summary

It is hard to tell if small pulmonary nodules are harmless or cancerous and need surgery. Our study looked at whether a routine blood test could help. We examined special markers on DNA from blood immune cells of patients with pulmonary nodules. Then, we trained computer models to recognize patterns in these markers. The models were very good at telling which nodules were cancerous and needed surgery. They could also separate slow-growing cancers from faster-growing ones, helping doctors plan the right region for surgery. This means that in the future, a simple blood test might help avoid unnecessary operations.

Yin, Zhang, Qiu et al. profile genome‑wide DNA methylation in PBMCs and combine validated methylation features with routine blood biomarkers to build machine learning models for pulmonary nodule classification. The models accurately distinguish benign from malignant nodules and minimally invasive from early invasive adenocarcinoma, supporting PBMC methylation signatures as practical, non‑invasive tools to inform surgical decision‑making.

Introduction

Lung cancer remains the most commonly diagnosed malignancy worldwide¹. Despite significant advancements in treatment modalities in recent years, the overall survival rate of lung cancer patients remains low, primarily due to the fact that most patients are diagnosed at an advanced stage². Early diagnosis is critical for improving survival outcomes, as studies have shown that the 5-year survival rate for early-stage lung cancer patients can reach 70%–90%, compared to less than 20% for those diagnosed at advanced stages³. However, pulmonary nodules are frequently detected in the general population, with approximately 20%–50% of adults showing nodules on computed tomography (CT) scans, the majority of which are benign⁴. Therefore, the development of effective biomarkers for early lung cancer diagnosis and the differentiation of benign from malignant pulmonary nodules is of significant clinical importance.

Among malignant pulmonary nodules, minimally invasive adenocarcinoma (MIA) represents a distinct subtype with unique characteristics and clinical implications⁵. Unlike invasive adenocarcinoma (IAC), MIA exhibits limited invasion beyond the basement membrane, with minimal destruction of surrounding tissues and no distant metastasis. Consequently, MIA requires a less extensive surgical resection, does not necessitate lymph node dissection, and is associated with reduced patient trauma, faster recovery, and fewer postoperative complications⁶. In contrast, IAC involves more extensive invasion, requiring larger surgical resections and lymph node dissection or lymph node sampling, which increases patient trauma and complication rates^7,8. Currently, the differentiation between MIA and early-stage IAC relies heavily on intraoperative frozen section analysis, which is limited by its speed and accuracy⁹. Thus, the development of non-invasive methods to distinguish MIA from early-stage IAC is crucial for guiding surgical decision making.

Recent advances in epigenetics have highlighted the potential of DNA methylation as a valuable tool for disease diagnosis, prognosis, and treatment^10–12. Emerging studies suggest that DNA methylation patterns in PBMCs may serve as a promising avenue for early cancer detection¹³. For instance, an 850 K chip analysis identified potential DNA methylation differences between healthy individuals and breast cancer patients, leading to the development of a diagnostic model based on four methylation sites using pyrosequencing and targeted bisulfite sequencing (TBS)¹⁴. This model effectively differentiated benign from malignant breast nodules. Similarly, another study identified methylation differences in PBMCs between healthy individuals and colorectal cancer patients, constructing a diagnostic model based on three methylation sites using TBS and multiplex qPCR¹⁵. A recent prospective cohort study further demonstrated that DNA methylation in PBMCs could distinguish benign thyroid nodules from thyroid cancer, achieving an AUC of 0.863¹⁶. These findings collectively suggest that DNA methylation in PBMCs holds promise for differentiating benign and malignant nodules in breast and thyroid cancers. However, its potential role in distinguishing pulmonary nodules remains unexplored.

This study employs methylome-seq and TBS to profile DNA methylation patterns in PBMCs from patients with pulmonary nodules, identifying differentially methylated cytosines (DMCs) and differentially methylated regions (DMRs) as candidate diagnostic biomarkers in benign and malignant pulmonary nodules. Subsequently, we construct 2 models based on these PBMCs methylation signatures combined with conventional blood biomarkers by machine learning. The SurgMalig model serves as a gatekeeper to identify nodules warranting resection, while the ResectGuide model provides guidance on the appropriate surgical extent for identified malignancies.

Methods

Human samples

This retrospective, cross-sectional case-control study was approved by the Ethical Committee of Shenzhen People’s Hospital (Approval No. LL-KY-2024189-02) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from the individuals for the publication of any potentially identifiable images or data included in this article. Peripheral blood samples (n = 250) were consecutively collected from treatment-naïve patients diagnosed with pulmonary nodules at the Department of Laboratory Medicine, Shenzhen People’s Hospital, between January 2024 and June 2025. All enrolled patients were deemed at high risk for lung cancer based on clinical and radiological assessment and subsequently underwent surgical resection with curative intent. Eligible participants provided 2 mL of venous blood prior to any surgical intervention. Following surgical resection, pulmonary nodules were classified and staged definitively based on comprehensive histopathological examination according to the WHO Classification of Thoracic Tumors (5th edition). Benign pulmonary nodules (BPN) were defined as nodules exhibiting histological features of hamartoma, granulomatous inflammation, or chronic inflammation. MIA and early-stage invasive adenocarcinoma (eIAC) were diagnosed based on established pathological criteria. Exclusion criteria applied to all groups included prior history of malignancy, receipt of chemotherapy or radiotherapy, insufficient DNA yield after extraction, and concurrent diagnoses known to significantly alter PBMC DNA methylation profiles, such as active autoimmune diseases or acute systemic inflammatory conditions. The clinical and biological characteristics of the participants are detailed in Supplementary Tables 1–3.

PBMCs separation

Five milliliters centrifuge tube was prepared with 1 mL of Ficoll solution. Subsequently, 1 mL of whole blood was carefully layered on top of the Ficoll solution. The sample was centrifuged at 800 rpm for 20 min. Following centrifugation, the PBMCs layer, located at the interface, was carefully transferred to a 1.5 mL microcentrifuge tube. The PBMCs were pelleted by centrifugation at 3000 rpm for 5 min, and the supernatant was discarded. To remove residual red blood cells, the pellet was treated with red blood cell lysis buffer. The cells were then washed three times with phosphate buffered saline (PBS), with centrifugation at 3000 rpm for 5 min after each wash. The final pellet, consisting of purified PBMCs, was retained for downstream applications.

Methylome-seq

Whole-genome bisulfite sequencing libraries were constructed using the Acegen Methylome-seq Library Preparation Kit (Acegen, Cat. No. AG1209) following the manufacturer’s protocol. Briefly, PBMCs genomic DNA spiked with 1‰ unmethylated Lambda DNA was sheared by sonication to an average size of 100–500 bp and treated with bisulfite. The heat-denatured DNA was then ligated to 3′-dA-tailed and single-stranded adapters. Subsequently, the single-stranded DNA connected to the 3′ adapter was extended to double-stranded DNA, and a 5-methylcytosine-converted adapter was ligated. PCR amplification was performed for approximately 10 cycles using Illumina 8 bp dual-index primers. The libraries were then enriched using the Acegen MethylDesign panel, followed by additional PCR amplification. The constructed libraries were analyzed using an Agilent 2100 Bioanalyzer and finally sequenced using a 150×2 paired-end sequencing protocol. Methylome-seq data from the Discovery cohort was publicly available in the NCBI BioProject with accession number PRJNA1327072.

TBS

TBS libraries were prepared using the Acegen Targeted Methyl Panel Kit (Acegen, Cat. No. AG0508) in accordance with the manufacturer’s instructions. Briefly, bisulfite sequencing PCR primers were designed using Acegen’s proprietary script. PBMCs Genomic DNA was subjected to bisulfite conversion using the ZYMO EZ DNA Methylation Gold Kit (Zymo Research, Irvine, CA, USA). The converted DNA served as the template for PCR amplification, which was carried out for 10 cycles. For each sample, amplicons corresponding to multiple genes were subjected to end repair, 3′-adenylation, and ligation to adapters containing 5-methylcytosine modifications. The resulting libraries were quality checked using an Agilent 2100 Bioanalyzer and subsequently sequenced on an Illumina platform with a 150×2 paired-end configuration. TBS data from the Train and Test cohorts were publicly available in the NCBI BioProject with accession number PRJNA1436047.

Bioinformatics analysis

Based on the DMCs and DMRs annotation results, genes overlapping with DMCs and DMRs in the genebody and its flanking regions (upstream 2 kb, genebody, and downstream 2 kb) were subjected to functional enrichment analysis. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis was conducted to identify significant pathways enriched among the DMR-associated genes and DMC-associated genes.

Feature selection through machine learning based ensembles

Machine learning framework comprised twelve distinct algorithms, encompassing regularization methods (Lasso, Ridge, ElasticNet), generalized linear model approaches (Stepglm, glmBoost, plsRglm), ensemble learning techniques (Random Forest, GBM, XGBoost), and pattern recognition models (SVM, LDA, Naïve Bayes). By combining both linear regression frameworks and advanced nonlinear learning structures, the design ensured broad and robust analytical coverage. The computational pipeline adopted a dual algorithm architecture: feature ranking through recursive feature elimination (RFE) for predictor screening, and predictive model construction via stacked generalization. Implemented under stratified 10-fold cross-validation schema, this ensemble strategy generated 113 unique algorithmic permutations through systematic hyperparameter optimization.

Model evaluation and biomarker identification

All trained models were evaluated in both the training and validation cohorts. Model performance was primarily assessed using the area under the receiver operating characteristic curve (AUC), with mean AUC values across bootstrap iterations serving as the central performance criterion. Furthermore, to provide a robust estimate of model stability and mitigate sampling bias, we incorporated a bootstrap resampling procedure (nboot = 2000). This allowed the calculation of 95% confidence intervals (CIs) for key performance metrics, including AUC and PRAUC, thereby offering a more rigorous assessment of model variability. The model demonstrating the robust performance in the Train and Test cohorts was designated as the optimal model. From this best performing model, selected genes were extracted. Local model explanation by the SHAP method.

Statistics and reproducibility

For the confirmation of each DMC or DMR in independent PBMC samples, assuming an estimated AUC of 0.80 or greater was deemed the lower bound of acceptable accuracy from a null value of 0.5 with 90% power. A minimum of 17 cases and BPN were required with a case/control ratio of 1:1 to achieve this difference at a two-sided significance level of 0.05, and participants enrolled in the Train cohort (30 BPN, 37 MIA, 33 eIAC) and the Test cohort (24 BPN, 34 MIA, 35 eIAC) were sufficient. Statistical analyses were performed with the IBM SPSS Statistics 25.0 and R version 3.4.0. The data variability, which was presented as the SD (mean ± SD), was analyzed via Mann–Whitney U test between two groups for abnormally distributed data. Receiver operating characteristics (ROC) curves were plotted to assess AUC, sensitivity, and specificity. We considered the value with the maximization of the Yuden index as the cut-off value. During the identification of pulmonary nodule subtype-specific DMCs and DMRs via Methylome-seq, the p values were corrected for multiple hypothesis testing using the Benjamini-Hochberg (B-H) procedure to obtain a q value. q < 0.05 was used as the significance cut-off for DMRs. Although q < 0.05 was initially used as the significance threshold for DMCs, the large number of DMCs identified led us to subsequently filter for more significantly differential methylation by applying a stricter threshold of q < 0.001.

Results

Clinical cohorts and participant characteristics

We established three well characterized cohorts to discover and validate PBMC DNA methylation biomarkers for the non-invasive classification of indeterminate pulmonary nodules (Fig. 1). 10 subjects were excluded due to inadequate DNA after extraction. Consequently, 47 samples (12 BPN, 15 MIA, 20 eIAC) were used as the Discovery cohort for Methylome-seq detection and the remaining 100 samples (30 BPN, 37 MIA, 33 eIAC) were used as the Train cohort for TBS detection. Moreover, 93 samples (24 BPN, 34 MIA, 35 eIAC) was Test cohort for evaluating the performance of models constructed by machine learning. Critically, the three diagnostic groups within each cohort were well-matched for age, gender, and smoking history, minimizing potential confounding effects (Supplementary Tables 1–3). Nodule size distribution was also comparable across groups in the Discovery and Train cohorts. While nodule type showed a significant difference between MIA and BPN in the Train cohort, this reflects the inherent radiological spectrum of these entities and underscores the need for improved diagnostic tools beyond imaging. All patients underwent surgical resection, enabling definitive histopathological classification. The vast majority of patients in these cohort eIAC groups were stage IA patients, who are difficult to distinguish clinically from MIA and BPN patients.

Fig. 1. Study design and clinical implementation of PBMC methylation-based classifiers for pulmonary nodule management.

a Schematic workflow outlining the multi-cohort study design: Discovery cohort for Methylome-seq; Train cohort and Test cohort for TBS; Development and validation of machine learning models (SurgMalig, ResectGuide). b Identification of pulmonary nodule subtype-specific DMCs and DMRs via methylome-seq. c Validation of discriminatory methylation biomarkers via TBS. d Proposed clinical integration pathway based on models integrating PBMC methylation signatures using machine learning: The SurgMalig model guided surgical resection necessity. The ResectGuide model informed surgical extent for malignant nodules.

Genome wide methylation profiling reveals PBMCs methylation landscapes of distinct pulmonary nodules

Study design and methodology are summarized in Fig. 1. To identify specific DNA methylation biomarkers in PBMC from pulmonary nodule patients, we performed Methylome-seq in a Discovery cohort consisting of PBMC samples from 12 BPN patients, 15 MIA patients, and 20 eIAC patients. All detected CpG sites underwent sequencing depth filtering (depth ≥5) to ensure methylation level accuracy. Following filtering, we detected a mean of 3,200,000 CpG sites per sample (totaling 4,045,187 sites), with methylation distributions per sample detailed in Supplementary Table 4. These sites were preprocessed into a unified methylation matrix for downstream analysis. While the density distribution of CpG methylation levels was similar across BPN, MIA, and eIAC groups, the eIAC group exhibited a statistically significant increase in mean global PBMC methylation compared to both BPN and MIA groups (Supplementary Fig. 1).

Identification of pulmonary nodule subtype-specific DMCs and DMRs via Methylome-seq

DMC was defined as CpG site meeting all criteria: (i) pairwise comparison q < 0.05, (ii) inter-group methylation difference ≥10%, and (iii) coefficient of variation (CV) ≥ 0.7. Volcano plots and heatmaps revealed 52,812 DMCs in MIA versus BPN, 77,759 DMCs in MIA versus eIAC, and 48,058 DMCs in BPN versus eIAC comparisons (Fig. 2a–d; Supplementary Fig. 2A, B). Principal component analysis (PCA) based on these DMCs demonstrated clear separation of the MIA from both BPN and eIAC, highlighting unique PBMC methylation signatures associated with MIA (Fig. 2e). Genomic distribution analysis showed predominant localization of DMCs in gene bodies, while regulatory regions contained fewer sites (Fig. 2f–h). KEGG pathway analysis of DMC-modified genes revealed significant enrichment in pathways critical for development and signaling, including neuroactive ligand receptor interaction, cAMP signaling, and hormone regulation (Fig. 2i, j and Supplementary Fig. 2C).

Fig. 2. Identification of DMCs discriminating pulmonary nodule subtypes via methylome-seq.

a Volcano plots presented DMCs in MIA versus BPN comparisons. b Heatmaps presented DMCs in MIA versus BPN comparisons. c Volcano plots presented DMCs in eIAC versus MIA comparisons. d Heatmaps presented DMCs in eIAC versus MIA comparisons. e PCA performed using all DMCs. Samples from MIA, BPN and eIAC were shown in different colors, with blue dots indicating MIA, red dots indicating BPN, and green dots indicating eIAC. f Genomic distribution of DMCs identified in MIA versus BPN comparisons. g Genomic distribution of DMCs identified in eIAC versus MIA comparisons. h Genomic distribution of DMCs identified in eIAC versus BPN comparisons. i KEGG analysis of the genes that were modified by DMCs identified in MIA versus BPN comparisons. j KEGG analysis of the genes that were modified by DMCs identified in eIAC versus MIA comparisons.

DMR, defined as contiguous genomic segments showing consistent methylation differences between groups, provide enhanced statistical power and biological interpretability compared to single site DMC. We partitioned the genome into 50 bp bins and performed pairwise comparisons across groups, designating bins meeting all criteria as DMRs: (i) pairwise comparison q < 0.05, (ii) inter-group methylation difference ≥10%, and (iii) CV ≥ 0.7. Volcano plots and heatmaps identified 1373 DMRs in MIA versus BPN, 2514 DMRs in MIA versus eIAC, and 1090 DMRs in BPN versus eIAC comparisons (Fig. 3a–d; Supplementary Fig. 3A, B). PCA using DMRs profiles also effectively segregated the MIA group from BPN and eIAC (Fig. 3e). DMRs were similarly enriched in gene bodies and implicated analogous developmental pathways through KEGG analysis (Fig. 3f–i and Supplementary Fig. 3C).

Fig. 3. Identification of DMRs discriminating pulmonary nodule subtypes via Methylome-seq.

a Volcano plots presented DMRs in MIA versus BPN comparisons. b Heatmaps presented DMRs in MIA versus BPN comparisons. c Volcano plots presented DMRs in eIAC versus MIA comparisons. d Heatmaps presented DMRs in eIAC versus MIA comparisons. e PCA performed using all DMRs. Samples from MIA, BPN and eIAC were shown in different colors, with blue dots indicating MIA, red dots indicating BPN, and green dots indicating eIAC. f Genomic distribution of DMRs identified in MIA versus BPN comparisons. g Genomic distribution of DMRs identified in eIAC versus MIA comparisons. h KEGG analysis of the genes that were modified by DMRs identified in MIA versus BPN comparisons. i KEGG analysis of the genes that were modified by DMRs identified in eIAC versus MIA comparisons.

Given the large number of candidates DMCs, we tightened the significance threshold to q < 0.001 in pairwise comparisons. Subsequently, we identified 21 DMCs common to all pairwise comparisons (MIA versus BPN, MIA versus eIAC, BPN versus eIAC) that collectively discriminated among the three groups (Fig. 4a). Heatmaps demonstrated distinct methylation patterns of these DMCs, indicating that they were significantly different in both the MIA group, the BPN group and the eIAC group (Fig. 4b; Supplementary Table 5). Similarly, 35 DMRs shared across pairwise comparisons distinguished the three groups, with heatmaps visualizing their group specific methylation profiles (Fig. 4c, d; Supplementary Table 6).

Fig. 4. Common DMCs and DMRs discriminate all three pulmonary nodule subtypes.

a Venn diagram of 21 overlapping DMCs from MIA versus BPN, MIA versus eIAC, and BPN versus eIAC comparisons. b Heatmap of methylation levels for the 21 common DMCs in each group. c Venn diagram of 35 overlapping DMRs from all pairwise comparisons. d Heatmap of methylation levels for the 35 common DMRs in each group.

Robust validation of discriminatory methylation biomarkers via TBS

TBS was a PCR based deep sequencing method for precise DNA methylation quantification, which was employed to validate differential methylation (Fig. 5a). Bisulfite sequencing primers were successfully designed for 20 out of 21 DMCs and 29 out of 35 DMRs and were subsequently used for TBS analysis in the Train cohort (Supplementary Table 7). The results showed that 8 DMCs and 4 DMRs were significantly different in MIA versus eIAC groups, 5 DMCs and 7 DMRs were significantly different in MIA versus BPN groups, and 4 DMCs and 5 DMRs were significantly different in BPN versus eIAC groups (Fig. 5b). Crucially, only two features, DMC-16 and DMR-13, demonstrated statistically significant differential methylation across all three pairwise comparisons (Fig. 5b). Both the methylome-seq results in the Discovery cohort and the TBS results in the Train cohort suggested that DMC-16 and DMR-13 were significantly different among the three groups (Fig. 5c–f). Additionally, we obtained consistent results in a Test cohort using TBS to detect DMC-16 and DMR-13 (Fig. 5g, h). Moreover, ROC curve results demonstrating the ability of DMC-16, DMR-13, and their combination score to distinguish patients with BPN who did not require surgical resection from patients with MIA and eIAC who did require surgical resection, yet conventional serum tumor biomarkers were unavailable (Fig. 5i, j; Supplementary Fig. 4A, B). Besides, ROC curve results demonstrating DMC-16, DMR-13 and their combination score were also able to differentiate to some extent between patients with MIA requiring only sublobectomy and patients with eIAC requiring lobectomy, and similarly conventional serum tumor biomarkers did not have the ability to differentiate (Fig. 5k, l; Supplementary Fig. 4c, d). The relevant metrics of these ROC curves were shown in Supplementary Tables 8 and 9.

Fig. 5. Validation of DMCs and DMRs discriminating pulmonary nodule subtypes via TBS.

a Schematic diagram of TBS detection process. b Venn diagram of 2 overlapping DMCs and DMRs from MIA versus BPN, MIA versus eIAC, and BPN versus eIAC comparisons via TBS. Methylome-seq results of DMC-16 (c) and DMR-13 (d) in the Discovery cohort. TBS results of DMC-16 (e) and DMR-13 (f) in the Train cohort. TBS results of DMC-16 (g) and DMR-13 (h) in the Test cohort. i ROC curves to identify MIA from BPN and eIAC patients using PBMCs DNA methylation signatures in the Train cohort. j ROC curves to identify MIA from BPN and eIAC patients using conventional serum tumor biomarkers in the Train cohort. k ROC curves to identify MIA from BPN patients using PBMCs DNA methylation signatures in the Train cohort. l ROC curves to identify MIA from eIAC patients using conventional serum tumor biomarkers in the Train cohort. Data are shown as mean ± SD; * P < 0.05, ** P < 0.01, *** P < 0.001. P values were determined by a two-sided Mann–Whitney U test.

Construction of surgical malignancy classification model and resection guide classification model based on PBMC methylation signatures by machine learning

However, the discriminatory power of the methylation signature alone for indeterminate pulmonary nodules required further enhancement. Given that immune system methylation reflected not only methylation status but also dynamic changes in circulating immune cell composition, we integrated the identified methylation features with 29 routine hematologic parameters and 5 tumor blood biomarkers using 113 machine learning models.

First, we developed the SurgMalig Model as a surgical malignancy classifier to discriminate between malignant nodules requiring surgery (MIA and eIAC) and BPN (Fig. 6). A heatmap demonstrated that the Lasso + Stepglm[forward] algorithm yielded the best AUC performance, identifying it as the most suitable algorithm for the SurgMalig Model (Fig. 6a). The model exhibited high performance in distinguishing benign from malignant nodules, achieving an AUC of 0.926 in the Train cohort and 0.806 in the Test cohort, along with a PRAUC of 0.970 and 0.933, respectively (Fig. 6b). Model performance was further rigorously assessed using confusion matrices, with overall accuracy reaching 0.860 in the Train cohort and 0.742 in the Test cohort (Fig. 6c, d). The precision of the SurgMalig Model was 0.952 and 0.925, indicating that the majority of cases predicted as positive were correctly classified (Supplementary Table 10). The SurgMalig Model incorporates 9 features, including Basophil, CA19-9, Cyfra21-1, DMC-16, DMR-13, MCH, MCHC, PLT, and Smoking History. Among individual features, DMC-16 and DMR-13 outperformed others when used independently in both the Train and Test cohorts (Fig. 6e, f and Supplementary Fig. 5A, B). A bar chart of SHAP values was employed to quantify and visualize the contribution of each feature to the predictions made by the SurgMalig Model (Fig. 6g). SHAP values reflect the extent to which each feature influences the final prediction under varying feature combinations. High SHAP values for DMC-16 and DMR-13 indicated that these features had a substantial impact on the model’s predictions (Supplementary Fig. 5C–G). A bee swarm plot illustrated the distribution of SHAP values for each feature, offering a visual representation of how different value ranges of each feature affect the model’s output (Fig. 6h). Among all features, DMC-16 and DMR-13 exhibited the highest influence, with higher feature values corresponding to increased SHAP values, suggesting that elevated levels of DMC-16 and DMR-13 significantly enhance the model’s predictive tendency toward malignancy.

Fig. 6. Development and performance of the SurgMalig machine learning model for surgical malignancy classification.

a Comparison of AUC performance among multiple models for distinguishing benign nodules (BPN, no surgery) from malignant nodules (MIA + eIAC, surgery required). b ROC and PR curves of the SurgMalig model in the Train cohort and Test cohort. SurgMalig model’s confusion matrices in the Train cohort (c) and Test cohort (d). Performance of SurgMalig model with different features in the Train cohort (e) and Test cohort (f). g Bar chart of the SHAP value indicates the importance of top 5 feature in the SurgMalig model prediction. h Bee Swarm plot showed the SHAP value distribution of each feature in the SurgMalig model.

Subsequently, we constructed the ResectGuide Model as a resection guide classifier to further subclassify malignant nodules into MIA and eIAC, thereby assisting surgical decision making (Fig. 7). The heatmap indicated that the Enet [alpha = 0.2] algorithm achieved the optimal AUC performance, confirming its suitability for the ResectGuide Model (Fig. 7a). This model demonstrated strong performance in differentiating MIA from eIAC, with an AUC of 0.862 in the Train cohort and 0.843 in the Test cohort, and a PRAUC of 0.881 and 0.962, respectively (Fig. 7b). Evaluation via confusion matrices showed an overall accuracy of 0.814 in the Train cohort and 0.783 in the Test cohort (Fig. 7c, d). The precision of the ResectGuide Model was 0.955 and 0.917, implying that most predictions labeled as positive were accurate (Supplementary Table 10). The ResectGuide Model comprised 11 features, DMC-16, DMR-13, Diameter, Monocyte, Monocyte Percentage, CEA, PDW, Basophil, RDW-CV, Eosinophil Percentage, and PLT. When evaluated individually, these features demonstrated comparable performance in predicting the subclassification (Fig. 7e, f; Supplementary Fig. 6a, b). SHAP analysis revealed that DMC-16 consistently had high SHAP values, underscoring its prominent role in the predictions of the ResectGuide Model (Fig. 7g, h; Supplementary Fig. 6c–g).

Fig. 7. Development and performance of the ResectGuide machine learning model for surgical resection guidance.

a Comparison of AUC performance among multiple models for subdividing malignant nodules into MIA and eIAC. b ROC and PR curves of the ResectGuide model in the Train cohort and Test cohort. ResectGuide model’s confusion matrices in the Train cohort (c) and Test cohort (d). Performance of ResectGuide model with different features in the Train cohort (e) and Test cohort (f). g Bar chart of the SHAP value indicates the importance of top 5 feature in the ResectGuide model prediction. h Bee Swarm plot showed the SHAP value distribution of each feature in the ResectGuide model.

Discussion

This study uniquely focuses on PBMC DNA methylation as a surrogate for the host systemic immune response to early pulmonary neoplasia. Multiple high-risk factors for lung cancer, including smoking, air pollution, and chronic inflammation, are closely associated with immune responses¹⁷. Immune cells dynamically react to tumorigenic changes, playing a pivotal role in protecting the host from tumor progression and influencing the development of lung cancer¹⁸. PBMCs, which consist of circulating monocytes, T cells, B cells, and natural killer cells, can be easily obtained from routine blood samples and closely reflect the state of immune responses^19,20. Consequently, PBMCs have been identified as surrogate biomarkers for aging clock and chronic diseases such as type 2 diabetes and ischemic heart disease through the use of the MethylationEPIC Array^21–23. Additionally, RNA transcriptome analysis of PBMCs has revealed a 23 gene panel that can be used for the diagnosis and prognostic monitoring of colorectal cancer²⁴. Compared to ctDNA, PBMCs-derived DNA is more accessible and available in higher concentrations. Recent studies have further demonstrated that DNA methylation in PBMCs may serve as a complementary source of tumor biomarkers. Cancer specific DNA methylation alterations have been identified in peripheral immune cells of patients with hepatocellular carcinoma, colorectal cancer, breast cancer, and thyroid cancer^13–16,25. This study further reveals that DNA methylation in PBMCs can also be utilized to monitor the transition from benign pulmonary nodules to lung cancer. There was a well-established association between patient age and increased risk of immune aging and cancer, and there is a recorded association between age and blood cell DNA methylation²⁶. At the time this study was designed, the relationship between PBMC DNA methylation and lung cancer was unclear. To prevent potential confounding by age given our limited sample size, the three diagnostic groups within each cohort were intentionally well-matched for age. Now, based on the findings of this study, which confirm an association between PBMC methylation and lung cancer, future research with larger sample sizes could directly build upon our conclusions to investigate the specific influence of age on the model’s performance.

It was indeed possible that the observed methylation alterations could stem from differences in PBMC cell-type proportions. To address this, we analyzed routine blood test data from the MIA, BPN, and eIAC groups in both the Train and Test sets. Our analysis showed no significant differences in the major PBMC components, monocytes and lymphocytes, across these groups (Supplementary Tables 2 and 3). These findings largely ruled out cell-type proportion differences as a primary driver of the methylation changes. However, our results did not yet pinpoint in which specific cellular subpopulations these methylation alterations occur. Recently, research teams have begun developing single-cell methylome profiling technologies^27,28. Building on our preliminary conclusions, future studies could apply single-cell methylomics to further identify the specific cell subsets harboring these methylation changes.

The robustness of our approach is underscored by the stringent multi-stage validation process. Unbiased Methylome-seq revealed numerous differential methylation features (Figs. 2–4). Notably, only two features, DMC-16 and DMR-13, were consistently validated across all three pairwise comparisons in both the Train and Test cohort using TBS (Fig. 5a–h). DMC-16 was located in the promoter region of the ASIC2 gene, whereas DMR-13 was situated within the gene body of SPATS2. ASIC2 was a member of the acid-sensing ion channel family²⁹. Monocytes sensed extracellular acidification via ASIC2, and reduced expression of ASIC2 impairs their chemotactic migration and polarization, thereby attenuating its antitumor effects³⁰. Single-cell sequencing data revealed that SPATS2 was significantly downregulated in infiltrating monocytes/macrophages within lung cancer tissues³¹. Given the established role of SPATS2 in regulating sperm cell differentiation, we speculated that it might also participate in macrophage polarization³². However, relevant studies remain limited, and further experimental investigation is warranted to explore this potential function.

DMC-16 and DMR-13 combined performance in distinguishing surgical necessity and informing surgical extent significantly surpassed conventional serum biomarkers (Fig. 5i–l). This targeted validation strategy prioritizes markers with high translational potential. Furthermore, our findings contrast with a contemporaneous study reporting different PBMC methylation markers for lung cancer diagnosis³³. We validated their three DMCs using TBS and found no significant differences in malignant pulmonary nodule patients (Supplementary Fig. 7). This discrepancy likely stems from their use of healthy controls without pulmonary nodules versus our clinically relevant BPN control, highlighting the critical importance of appropriate comparator groups when developing diagnostic tools for indeterminate nodules. Our model’s specificity is intrinsically tailored to this challenging diagnostic context. The integration of these models into a proposed clinical pathway illustrates a practical roadmap for implementation (Fig. 1d).

In recent years, increasing attention has been focused on the critical role of RNA methylation in tumor progression^34,35. Our previous studies have revealed that, similar to DNA methylation in PBMCs, RNA methylation in peripheral blood leukocytes also serves as a tumor biomarker. The m⁶A modification levels in peripheral blood leukocytes were significantly upregulated in CRC patients and correlated with tumor staging³⁶. Similarly, elevated m⁶A modification levels were observed in peripheral blood leukocytes from lung cancer patients, with enrichment in CD4⁺ T cells and strong associations with key m⁶A erasers, FTO and ALKBH5³⁷. However, our previous research primarily focused on distinguishing lung cancer patients from healthy individuals. Whether RNA methylation can further stratify nodule types remains to be explored. Given the distinct regulatory mechanisms underlying RNA methylation and DNA methylation, the combined analysis of both epigenetic modifications in PBMCs may significantly improve the classification accuracy of pulmonary nodules.

Despite the promising findings of this study, several limitations warrant acknowledgment. First, the sample size was derived from a single center, which may limit the statistical robustness of our model. Future work will validate these results through multicenter cohorts with expanded participant recruitment. Second, functional validation of the identified methylation-driven targets is required to elucidate the molecular mechanisms underlying carcinogenesis and progression. Furthermore, while methylation biomarkers demonstrate diagnostic potential, the sole reliance on epigenetic signatures may incompletely capture the complexity of lung cancer pathogenesis, which involves dynamic interplay among genetic, epigenetic, and proteomic alterations³⁸. To address this, we plan to integrate multi-omics approaches, including radiomics, methylation profiling, and proteomic biomarkers, to develop a holistic framework for early diagnosis and risk stratification.

In conclusion, this exploratory study establishes DNA methylation signatures in PBMCs as a powerful class of non-invasive biomarkers for the classification of indeterminate pulmonary nodules. By translating these signatures into the SurgMalig and ResectGuide machine learning models, we provide clinicians with clinically actionable tools to guide two pivotal surgical decisions: whether resection is necessary, and if so, what the optimal extent should be. Our approach overcomes key limitations of current diagnostic methods by focusing on the accessible systemic immune compartment, requiring minimal blood volume, and delivering high accuracy specifically within the challenging indeterminate nodule population. This work not only fills a critical gap in the early detection and management of lung cancer but also opens avenues for understanding the systemic immune epigenetic response to early neoplasia. Multi-cohort validation and integration into clinical pathways have the potential to significantly reduce diagnostic uncertainty, minimize unnecessary procedures, optimize surgical management, and ultimately improve outcomes for patients with pulmonary nodules.

Author contributions

Conceptualization, Y.H.F., H.Z.J., and Y.X.P.; Methodology, Y.H.F. and Z.N.F.; Data collection, Q.L.M., Z.J.J., and D.J.H.; Statistical analysis, H.Z.J. and Y.H.F.; Funding acquisition, Y.H.F. and Y.X.P.; Study supervision, X.J.N. and Z.N.F. All authors reviewed the manuscript and approved the final revision.

Peer review

Peer review information

Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The data supporting the findings of this study have been deposited in the Sequence Read Archive (SRA) under accession numbers PRJNA1327072 and PRJNA1436047. After we recently released the data from the website, these data are now freely available for download. The source data for Figs. 1–4 are available under PRJNA1327072. The source data for Figs. 5–7 are available under PRJNA1436047 and in Supplementary Data 1.

Code availability

The codes used to construct machine learning models in this study can be accessed at 10.5281/zenodo.19446997³⁹.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01616-4.