During perioperative care for non-small cell lung cancer (NSCLC) patients, clinical outcomes vary significantly. There is a critical need for more dependable biomarkers to identify high-risk individuals in the perioperative phase. This is essential for enhancing postoperative interventions and positively influencing clinical results.

We collected a tissue DNA methylation cohort of 73 stage I–III surgically treated patients as the discovery set for model development. The model was established using recurrence-free survival (RFS) as the primary endpoint. Subsequently, its prognostic value was validated in an independent cohort of 30 stage I–III surgical patients, and further confirmed across different patient subgroups.

We developed an Early to Mid-term NSCLC Recurrence LASSO score (EMRL) predictive model based on five differentially methylated regions (DMRs). The EMRL model was significantly associated with RFS in stage I-III surgically treated patients (RFS: log-rank P = 0.00032) and was confirmed as an independent prognostic factor in multivariate Cox regression analysis (HR = 0.35, 95% confidence interval 0.20–0.61, P < 0.001). Notably, EMRL not only identified high-risk patients within the same TNM stage but also demonstrated strong predictive performance in patient subgroups harboring EGFR-TKI-sensitive mutations and those with positive PD-L1 expression.

In this study, we developed a postoperative recurrence prediction model based on preoperative tissue methylation characteristics to identify individuals in I-III stage NSCLC patients following surgical resection who may have a higher risk of recurrence. This offers opportunities for early personalized treatment and follow-up strategy.

Non-small cell lung cancer (NSCLC) ranks as one of the most frequent cancers in China and the world, and is associated with extremely high morbidity and mortality [1, 2]. Approximately 35% of NSCLC patients have a potentially surgically resectable disease [3]. After the standard treatment, even for the same stage NSCLC patients with R0 resection, considerable heterogeneity of clinical outcomes still remains [4, 5]. This phenomenon suggests that there may be molecular mechanisms influencing NSCLC prognosis and efficacy. Identifying patients who are at risk of recurrence after surgery can improve the postoperative management and clinical outcomes of them. Therefore, there is a significant demand to identify biomarkers for risk stratification in the early phase of the clinical practice.

DNA methylation is an epigenetic modification that plays a crucial role in cell cycle, DNA repair, cell proliferation, and apoptosis, which are critical pathways involved in tumor development and progression [6, 7]. Abnormal DNA methylation patterns, including genome-wide hypomethylation of repetitive elements and CpG-poor regions, and regional hypermethylation of CpG islands at gene promoters, have been associated with various diseases [8,9,10]. Increasing evidence supports DNA methylation changes as reliable biomarkers for lung cancer risk assessment, early diagnosis, prognosis stratification, and treatment prediction [11,12,13,14,15,16]. Previous study have shown that LRRC3B methylation can serve as a powerful biomarker for predicting NSCLC immunotherapy response [17]. So far, some genes such as CDKN2A, CDH13 [18], KMT2C [19], LRRC3B [17] or methylation markers [20] have been identified, and their DNA methylation status appears to be associated with the prognosis of NSCLC patients and can be used to predict disease recurrence after surgery. However, most current prediction models are based on single-gene methylation levels, without considering that the differential methylation regions (DMR) are more functional parameter [21].

In the present study, we conducted a research on NSCLC cohorts using a previously developed methylation panel consisting of 80,672 CpG sites. Five candidate blocks were selected for the prognostic model using LASSO Cox regression, which demonstrated its effectiveness in predicting disease recurrence in NSCLC patients, both in the training and validation cohorts. This effectiveness was observed independently of variables such as TNM staging and histological subtypes.

The cohort experimental design of our study is as shown in the SFigure 1. From December 2017 to July 2019, a cohort of 73 patients diagnosed with resectable non-small cell lung cancer (NSCLC) stages IA-IIIB, including adenocarcinoma (LUAD), squamous cell carcinoma (LUSC), and rare subtypes (Others), such as large cell carcinoma, sarcomatoid carcinoma, was retrospectively enrolled in a discovery study at Fudan University Shanghai Cancer Center. All patients underwent curative surgery, and their median follow-up duration was 42 months. Baseline tissue samples from all 73 patients were collected and analyzed for DNA methylation. Among these, 53 samples underwent targeted sequencing of 168-gene panel. Additionally, baseline blood samples from 60 patients were collected and tested for DNA methylation. From January 2019 to April 2020, an additional cohort of 30 patients was collected at the same center for the validation set. Similarly, all patients underwent curative surgery, with a median follow-up time of 45 months. Baseline tissue samples were collected from all patient and analyzed for DNA methylation, with 22 of these tissue samples undergoing targeted sequencing of 168-gene panel. Furthermore, blood samples from 108 healthy individuals and adjacent non-tumor tissues from 59 lung cancer patients (located at least 5 cm away from the tumor) were collected for methylation analysis. To ensure a diverse participant population, the healthy control group was recruited from different physical examination centers. They were individuals without a history of malignancy or critical illnesses such as hepatitis, liver cirrhosis, chronic obstructive pulmonary disease, and colorectal disease. The participants underwent routine health checkups, including blood tests, urinalysis, blood biochemistry tests, electrocardiograms, low-dose chest CT scans, and abdominal ultrasounds. Only participants with normal test results were included in the study (sFig 1 A).

After surgery, the follow-up assessments involved contrast chest computed tomography (CT), brain magnetic resonance imaging, bone emission CT, and abdominal ultrasound to detect local recurrence or distant metastases. Patients were contacted every 3 months after the date of surgery in clinic or by telephone about the information of disease recurrence and survival. The ethical approval for this study was obtained from the ethics committee at Fudan University Shanghai Cancer Center (No: 2017-151-1354).

Ten microliters (µL) of peripheral blood were collected and stored at 4 °C. Subsequently, the collected blood samples were centrifuged at 2000 g for 10 min at 4 °C. The resulting supernatant was carefully transferred to a new centrifuge tube and subjected to a second centrifugation at 16,000 g for 10 min at 4 °C. The obtained supernatant was then stored at -80 °C for subsequent analysis. cfDNA extraction was performed using the QIAamp Circulating Nucleic Acid Kit (Qiagen, Hilden, Germany). Collected samples were subjected to DNA isolation using the QIAamp DNA FFPE Tissue Kit (Qiagen, Valencia, CA, USA) according to the manufacturer’s instructions. Quantification of tissue DNA and was performed using the Qubit 2.0 Fluorimeter with dsDNA HS assay kits (Life Technologies, Carlsbad, CA, USA). A total of 20 to 80 ng of tissue DNA or cfDNA was required for NGS library construction.

Bisulfate sequencing library preparation for DNA methylation profiling was performed using ELSA-seq with a panel consisting of 80,672 CpG sites, spanning 1.05 Mb of the human genome (Burning Rock Biotech, Guangzhou, China). Sequencing of target libraries was performed on a NovaSeq 6000 sequencer (Illumina, San Diego, CA, USA), with an average sequencing depth of 1000×. DNA methylation levels were analyzed using an optimized bioinformatics pipeline based on the alignment of sequencing reads to C to T- and G to A-transformed hg19 genome, as described previously [14, 22, 23]. As previously described [24], methylation blocks were defined as genomic regions between neighboring CpG sites, where the r2 value was calculated using a modified correlation matrix. A total of 80,672 CpG sites were categorized into 8,312 methylation blocks using linkage disequilibrium and statistical modeling. Among these methylation blocks, 84% were annotated within genes, with 59%, 18%, and 7% located in promoter regions, introns, and exons, respectively.

Targeted sequencing was carried out using a panel consisting of 168 cancer-related genes (Burning Rock Biotech, Guangzhou, China). Indexed samples were sequenced on a Nextseq500 sequencer (Illumina, San Diego, CA, USA) with paired-end reads, with an average sequencing depth of 1000× for tissue samples. Somatic variants were called using optimized bioinformatics pipelines that can accurately report various cancer-related genetic alterations, including single-nucleotide variants (SNVs), insertion and deletion variants (indels), copy number variants (CNVs), and genomic rearrangements, as described previously [14, 25].

A 5-µm section was cut from each FFPE block of the tumor and stained using the Dako 22C3 mouse monoclonal antibody on the Dako Autostainer Link-48 platform, following the manufacturer’s instructions. Cores with a neoplastic component ≥ 30% were deemed sufficient. PD-L1 results was evaluated independently by two pathologists, and any expression in tumor cells or immune cells was considered PD-L1 positive.

Statistical analysis was performed using R version 3.3.3 software. To evaluate group differences, various statistical tests were employed. Fisher’s exact test and the chi-square test were utilized for categorical variables, while the Mann-Whitney test was used for continuous variables. Survival variables, including overall survival (OS) and recurrence-free survival (RFS), were assessed using the Kaplan-Meier (KM) method, the log-rank method, and Cox regression (hazard ratio [HR] and 95% confidence interval [CI]). All p-values were calculated with two-tailed tests and considered significant when p < 0.05.

In the discovery cohort, we performed unsupervised clustering of methylation data using the K-means method (R package, heatmap). Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses were executed using R (R package, clusterProfiler) [26,27,28].For the identification of differentiated methylation regions (DMRs) between tumors and adjacent non-tumor tissues, both the Mann-Whitney test and the Kolmogorov-Smirnov test were applied. Principal component analysis on DMRs was conducted using PCA (R package, FactoMineR). The 5-DMR prognostic model was built via the least absolute shrinkage and selection operator (LASSO).

The clinical characteristics of the discovery cohort (n = 73) and the validation cohort (n = 30) are presented in Table 1. In the discovery cohort, the median age was 59.5 years (range: 52.5–63 years), and the proportion of patients with lung adenocarcinoma (LUAD) was 78.1% (n = 57). Additionally, 31.5% of patients (n = 23) were in stage I, and 24.7% (n = 18) were in stage II. Similarly, in the validation cohort, the median age was 61 years (range: 56–67 years), with 83.3% (n = 25) having LUAD. In the validation cohort, 36.7% of patients (n = 11) were in stage I, and 26.7% (n = 8) were in stage II. The median recurrence-free survival (RFS) in the discovery cohort was 28 months (95% CI, 22 to 38.2, sFig. 1B), while in the validation cohort, it was 59.9 months (95% CI, 29 to not estimable, sFig. 1 C). The two cohorts demonstrated a high degree of similarity and balance in key features and distributions, enhancing the reliability of comparisons and inferences between them.

To observe the global DNA methylation features among different sample types, we conducted unsupervised clustering on all samples (Fig. 1). We found clear differences in methylation patterns between normal and tumor samples, regardless of whether the samples originated from blood or tissue. Furthermore, compared to blood samples, the differences between tumor and adjacent non-tumor tissues were even more pronounced. This underscores the distinct methylation patterns in tissue samples, emphasizing the heightened significance of these variations within a tissue context. Furthermore, there was no clear seperation between LUSC and Others subtypes. This suggests that the global DNA methylation patterns are relatively similar among different subtypes.

Global DNA methylation features among different sample types

Full size image

To characterize the specific epigenetic features related to NSCLC, we compared DNA methylation levels between tumor and adjacent non-tumor tissues, identifying 1279 significantly differentially methylated regions (DMRs) with a β-value difference > 0.1. (Fig. 2A). These DMRs demonstrated statistical significance in Mann-Whitney testing (FDR < 0.0001) and Kolmogorov-Smirnov testing (P value < 0.0001). Among the identified DMRs, 24 were characterized as hypomethylation, while 1255 displayed hypermethylation, indicating a higher prevalence of hypermethylation patterns in NSCLC. In contrast, only 55 hypomethylated DMRs were identified in blood samples from patients and healthy controls (sFig. 2). These findings further emphasize the suitability of tissue-based DNA methylation data for the identification of NSCLC-specific methylation sites.

Overview of DMRs in NSCLC. (A) Volcano plot illustrating the P value and β value difference of methylation regions between tumor and adjacent non-tumor tissues. (B) The distribution of DMRs in different genomic regions. (C) PCA analysis of DMRs in different histological subtypes. (D) The β value difference of the hypermethylation DMRs and their inner-correlations among LUAD, LUSC and Others. (E) The KEGG and GO enrichment pathways for genes associated with hypermethylated DMR

Full size image

These 1279 DMRs related to NSCLC mostly are located in promoter regions(27.36%), introns(28.02%), and intergenic regions(26.29%) (Fig. 2B). Principal Component Analysis (PCA) based on DMRs showed no clear separation between LUAD, LUSC, and Others subtypes (Fig. 2C). Notably, a significant positive correlation was observed in the β-value variances of the 1255 hypermethylated DMRs among LUAD, LUSC, and Others subtypes(Fig. 2D). These findings further substantiate the notion of consistent DMR patterns across various histological classifications in NSCLC.

According to KEGG analysis, genes in hypermethylated regions are associated with neuroactive ligand-receptor interaction and cAMP signaling pathway (Fig. 2E). Additionally, GO analysis revealed that these genes functioned as DNA-binding transcription activator activity and DNA-binding transcription activator activity specific to RNA polymerase II (Fig. 2E).

The flowchart of this study is illustrated in Fig. 3. To specifically identify recurrence-related DMRs, a univariate Cox analysis conducted in the discovery cohort revealed 21 DMRs significantly associated with RFS (recurrence free survival). To refine our DMR selection further, we conducted a multivariate Cox analysis, considering clinical variables such as age, gender, subtype and staging. Finally, LASSO regression was employed to construct an EMRL score (Early to Mid-term NSCLC Recurrence LASSO score) based on 5 DMRs (optimizing the λ parameter and considering the distribution of LASSO coefficients are shown in sFig. 3B and sFig. 3 C, respectively). The median value from the discovery cohort was set as the cutoff, and we observed that this EMRL score effectively predicted the risk of patient recurrence in both the discovery cohort (RFS: log-rank P = 0.00032, Fig. 4A; methylation data are shown in Fig. 4E) and the validation cohort (RFS: log-rank P = 0.0072, Fig. 4B). Although there was a consistent trend in overall survival (OS), it did not reach statistical significance (sFig. 3 C-D). Furthermore, multivariate Cox analysis showed the EMRL score was an independent prognostic factor for predicting recurrence-free survival (HR = 0.35, 95% CI 0.18–0.67, P = 0.001, Fig. 4C; Table 2). Combining the discovery and validation cohorts and conducting multivariate Cox analysis produced consistent results (HR = 0.35, 95% confidence interval 0.20–0.61, P < 0.001, Fig. 4D; Table 3). These findings suggest that this model exhibits strong and effective performance in predicting patients’ recurrence, and independent of other critical clinical factors, including TNM staging and histological subtype.

Overall flowchart of this study

Full size image

Construction and validation of recurrence prediction model based on DMRs. Association between the EMRL score and RFS in the discovery cohort (A) and the validation cohort (B). (C) Multivariable Cox analyse of RFS in the discovery cohort. (D) Multivariable Cox analyse of RFS in the the combined discovery and validation cohorts. (E) Heatmap illustrating the methylation data and features of the 5 DMRs involved in the LASSO model

Full size image

Moreover, we conducted an additional subgroup analysis to confirm the models ability in specific clinical scenario. Within TNM-staged subgroups, the EMRL score consistently exhibited predictive value for RFS (Stage I: log-rank P = 0.014, Stage II: log-rank P = 0.066, Stage III: log-rank P = 0.016, Fig. 5A-C). These results emphasize the EMRL score ability to provide valuable prognostic insights across various disease stages. Nevertheless, the OS did not reach a significant difference (sFig 4A-C). When grouped by EGFR L858R/19del mutation status, the EMRL score also demonstrated predictive value for RFS (EGFR L858R/19del: log-rank P = 0.0042, non-EGFR L858R/19del: log-rank P = 0.0017, Fig. 5D-E). Although OS did not reach statistical significance, a consistent trend was consistent(sFig 4D-E). Notably, in the PD-L1-stratified analysis, EMRL score were significantly associated with RFS (log-rank P < 0.0001, Fig. 5G) and OS (log-rank P = 0.048, sFig 4G) in the PD-L1 positive subgroup. Conversely, no similar associations were observed in the PD-L1 negative subgroup (Fig. 5F, sFig 4F). This findings strongly suggested that EMRL score might have a valuable impact on recurrence and prognosis, specifically within PD-L1 positive patients.

This figure illustrates the association between EMRL Score and RFS in different subgroups: (A) stage I, (B) stage II, (C) stage III, (D) EGFR L858R&19del, (E) non-EGFR L858R&19del, (F) PD-L1 negative, (G) PD-L1 positive

Full size image

Additionally, we conducted DNA panel sequencing from patients with available samples. We observed that patients in EMRL-Low group had a higher percentage of TP53 (74.36% vs. 54.29%) and RB1 (23.08% vs. 8.57%) mutations (sFig. 5 A). Conversely, patients in EMRL-High group exhibited a higher percentage of ALK (14.29% vs. 2.56%) and SMAD4 (8.57% vs. 2.56%) mutations (sFig. 5B). Correlation analysis with clinical characteristics revealed that the EMRL-High group had a significantly higher proportion of patients in stage III (P = 0.018, sFig. 5 C). Furthermore, there was a tendency toward a higher representation of patients with rare cancer subtypes (P = 0.076, sFig. 5D) and those showing moderate to low levels of differentiation (P = 0.178, sFig. 5E). Of particular interest, the EMRL-High group had a significantly higher percentage of patients with PD-L1 positive expression in their tumor tissue (P = 0.028, sFig. 5 F).

In addition to the value of gene mutation information in guiding therapeutic management for lung cancer patients, there is growing evidence supporting the utility of DNA methylation profiling as valuable biomarkers for diagnosis, prognosis, and prediction [13, 14, 29, 30]. Numerous studies have demonstrated that DNA methylation alterations in lung cancer were commonly observed at the early stages of tumor development, often preceding any of detectable morphological changes [7, 31, 32].

In this study on Chinese NSCLC patients, we performed parallel somatic mutation and methylation profiling using blood and tumor tissue samples. Overall, DNA methylation levels effectively distinguish normal and tumor samples, with tissue-based data better suited for identifying NSCLC-specific methylation sites.

Leveraging these differences, we identified a set of differentially methylated regions (DMRs) and constructed a prognostic model composed of five methylation blocks, referred to as the EMRL model, to stratify patients by postoperative recurrence risk. The model demonstrated strong prognostic performance, and its clinical relevance was further validated in an independent external cohort of stage I–III surgically resected NSCLC patients.

Importantly, multivariate Cox regression analysis confirmed the EMRL model as an independent prognostic factor for recurrence-free survival in both the discovery and combined datasets, even after adjusting for conventional clinical variables such as TNM stage and EGFR mutation status. These findings underscore the potential of integrating tissue-based methylation markers into clinical workflows for individualized postoperative risk assessment in early- to mid-stage NSCLC.

This result highlights the ability of our methylation-based risk score to further stratify NSCLC patients who share the same TNM stage, thereby identifying a subgroup with significantly higher recurrence risk that would not be distinguishable by conventional staging alone. By uncovering this hidden heterogeneity, the model provides an opportunity for clinicians to deliver more tailored surveillance strategies and consider additional adjuvant interventions for those at elevated risk, even if they present with early- or intermediate-stage disease.

Furthermore, in the subgroup of patients harboring EGFR-TKI sensitive mutations—a population that typically benefits from targeted therapy—the risk score retained its predictive power for RFS. This suggests its potential utility in refining post-surgical management plans, such as optimizing the timing, duration, or intensity of EGFR-TKI administration to prevent recurrence more effectively.

In addition, the score demonstrated significant prognostic value among patients with positive PD-L1 expression. These findings indicate that the score may aid in guiding the use of immune checkpoint inhibitors or in designing risk-adapted follow-up protocols. Collectively, these results support the broad applicability of the model across molecular and immunological subtypes of NSCLC, underscoring its potential as a clinically relevant tool to inform personalized postoperative care.

This study has several limitations that warrant consideration. First, the patient cohorts included in the analysis were relatively small in size and derived from a single clinical center. Due to the high technical demands of the methylation profiling methods used, the study was only feasible in large, well-equipped clinical institutions. To address this limitation and enhance the generalizability of our findings, we plan to conduct a multi-center, prospective clinical study in the next phase of our research.

Second, the predictive accuracy of the EMRL model for OS remains to be improved. In the current cohorts, only 35.6% of patients in the discovery set and 20% in the validation set reached OS endpoints, resulting in a limited number of events. This low event rate may have reduced the statistical power of the OS analysis. In future studies, we aim to increase the sample size and extend the follow-up duration to more robustly assess the model’s performance in predicting OS.

Third, a substantial proportion of patients were non-local residents who chose to seek follow-up treatment at medical centers outside Shanghai after disease recurrence. This introduced heterogeneity in post-recurrence treatment regimens, making it difficult to control for treatment-related confounders and likely contributing to the decreased predictive capacity of the model for OS outcomes.

Moreover, while our findings offer meaningful clinical insights, the underlying biological mechanisms linking the identified DMRs to lung cancer progression remain incompletely understood. Due to current limitations in experimental conditions, functional validation has not yet been performed. In future work, we intend to investigate how methylation changes in these regions affect gene expression and signaling pathways involved in NSCLC pathogenesis. We will also explore their biological roles through in vitro and in vivo studies to better elucidate their functional relevance.

Finally, although the EMRL model successfully stratifies patients with the same TNM stage—including those harboring EGFR-TKI sensitive mutations and those with positive PD-L1 expression—into high- and low-risk groups for recurrence, its utility in guiding postoperative adjuvant therapy decisions requires further validation. Large-scale clinical trials will be essential to confirm the clinical benefit of applying this model in tailoring personalized treatment strategies for NSCLC patients after surgery.

In summary, this study presents a novel prognostic model based on preoperative tissue DNA methylation patterns to predict postoperative recurrence risk in stage I–III NSCLC patients. By identifying individuals at higher risk of recurrence following surgical resection, the model holds promise for enabling earlier, more personalized interventions and follow-up strategies. To further establish its clinical utility and robustness, additional validation in larger, multi-center cohorts will be essential. In the future, integrating this model with other molecular and clinical biomarkers, and exploring its role in guiding adjuvant therapeutic decisions, may further enhance its translational value and contribute to precision oncology in early- to mid-stage lung cancer.

Download references

Author notes

Hang Li and Hong Hu were in charge of the research design. Yi lu, Jing Wang and Bing Li were in charge of the experimental methodology. Haiqing Chen and Tong Li were in charge the specimen collection, also the data verification. Hang Li and Fangqiu Fu wrote the main manuscript text and Yi Lu, Jing Wang prepared figures. All authors reviewed the manuscript.

Correspondence to Hong Hu.

The ethical approval for this study was obtained from the ethics committee at Fudan University Shanghai Cancer Center (No: 2017-151-1354). Informed consent to participate was obtained from all of the participants in the study.

There were no identifying images or other personal or clinical details of participants that compromise their anonymity in this study. The consents for publication were also collected from all the authors.

The authors declare no competing interests.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Below is the link to the electronic supplementary material.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions