Detection of lung cancer through a method of enzymatic methylation sequencing in cell-free DNA

Lung cancer stands as the predominant cause of cancer-related fatalities, posing a substantial public health concern^[1]. The majority of lung cancer (LC) diagnoses occur at advanced stages, exceeding 80%, which contributes significantly to the disease's high mortality rate^[2]. The prognosis of LC is heavily dependent on the stage at diagnosis, ranging from an optimistic 84% five-year survival rate for stage IA to a concerning 13% for stage IV^[3,4]. Therefore, early screening and diagnosis hold promise for improving the prognosis of LC and mitigating mortality rates^[5,6]. Presently, low-dose spiral computed tomography (LDCT) is recognized as the most effective screening method for LC. However, its propensity for false positives may lead to overdiagnosis, hindering its seamless integration into routine clinical practice^[7,8]. For early detection of LC, extensive investigations have been conducted on blood-based protein biomarkers. Tumor-associated antigens (TAAs), such as cytokeratin 19 fragment (CYFRA 21-1)^[9], carcinoembryonic antigen (CEA)^[10], neuron-specific enolase (NSE)^[11], and squamous cell carcinoma antigen (SCC-Ag)^[12], have shown promise, but their suboptimal sensitivity hinders their efficacy in early diagnosis. Tumor-associated autoantibodies (TAAbs) represent the immune system's responses to peptides shed by tumors or present on the surface of tumor cells^[13]. At present, the combination of TAAbs and low-dose computed tomography (LDCT) has been used for lung cancer screening. Although studies have shown that the specificity of TAAbs in the diagnosis of lung cancer is 91.60%, its low sensitivity limits its wide clinical application (only 56.53%)^[14]. Besides, alterations in tumor signaling pathways play a pivotal role from the very early stages of tumorigenesis. For example, dysregulation of the TGF-β signaling co-receptor TGFBR3 has been reported to promote malignant progression by influencing proliferation, apoptosis, and metastatic potential of lung cancer cells, highlighting the importance of tumor-derived molecular changes in disease initiation^[15]. Meanwhile, the rapid evolution of immunotherapy demonstrates that the immune system actively participates in shaping the tumor microenvironment throughout lung cancer development, further emphasizing the unmet clinical need for effective biomarkers capable of capturing tumor-specific molecular characteristics^[16]. Consequently, there is an urgent requirement for the development of a sensitive, robust, and minimally invasive approach to detecting early-stage LC.

Cell-free DNA (cfDNA) refers to extracellular DNA fragments circulating in plasma, released primarily through programmed cell death (apoptosis) or uncontrolled cell rupture (necrosis). A subset of these fragments, circulating tumor DNA (ctDNA), originates specifically from cancerous cells^[17,18], and contains a wealth of tumor-associated molecular alterations, such as genetic mutations^[19], methylation modifications^[20], end motifs^[21], fragment length profiles^[22], etc. The detection of these tumor-derived signals in ctDNA has significant implications for developing liquid biopsy-based diagnostic approaches in oncology^[23,24].

Among epigenetic modifications, CpG methylation—the addition of a methyl group to cytosine residues in CpG dinucleotides—is a fundamental regulatory mechanism in eukaryotic cells^[25]. This modification is crucial for normal cellular differentiation, tissue development, and organ function. Due to its widespread dysregulation across cancer types and its early appearance during tumorigenesis, aberrant DNA methylation in ctDNA represents a highly informative biomarker for cancer detection^[26].

For methylation analysis, bisulfite conversion remains the gold standard, as it differentiates methylated and unmethylated cytosines at single-base resolution by chemically converting unmethylated cytosines to uracil while leaving methylated cytosines intact^[27]. However, the harsh chemical treatment can cause significant DNA degradation, posing challenges for low-input cfDNA studies. In contrast, Enzymatic Methyl-seq (EM-seq) employs a gentler enzymatic approach to profile genome-wide methylation. Our previous work^[28] directly compared both methods on cfDNA and demonstrated that EM-seq minimizes DNA damage and, for the same input amount, achieves significantly higher sequencing depth and more uniform coverage compared to bisulfite conversion. This ability to generate robust methylation data from limited and fragmented cfDNA makes EM-seq particularly suited for liquid biopsy applications and motivated its selection for this study. Moreover, compared to other gentle conversion methods like TET-assisted pyridine borane sequencing (TAPS), EM-seq utilizes a single-tube enzymatic reaction that minimizes sample loss and handling errors, which is particularly beneficial when working with the limited quantities and fragile nature of cfDNA. This robustness in low-input conditions directly supports the high conversion efficiency and sequencing uniformity observed in our plasma cohort.

In this research, our objective was to discern lung cancer-specific differentially methylated positions (DMPs) and assess the diagnostic ability of these cfDNA methylation biomarkers to distinguish early-stage LC from benign nodules. First, we commenced by contrasting the methylation signatures of LC tissues and paraneoplastic tissues from The Cancer Genome Atlas (TCGA) to pinpoint LC-specific DMPs. Then we designed a lung cancer detection panel of LC-specific DMPs. To develop a high-performance LC screening model, methylation analysis data were obtained from plasma samples of 127 individuals with early-stage lung cancer or non-lung cancer through targeted EM-seq analysis.

Lung cancer patients and non-lung cancer (NL) controls were recruited at Zhongshan Hospital (Xiamen) of Fudan University. Tumor staging was assigned according to the 8^th edition of the IASLC Lung Cancer Staging System. The study was conducted in compliance with the Declaration of Helsinki and approved by the institutional review board. Written informed consent was provided by each participant or their legally authorized representative.

Blood sample preparation and cfDNA extraction

For each subject, 10 mL of whole blood was collected into cell-free DNA preservation tubes (Vangenes). Plasma was separated within 2 h through a two-step centrifugation procedure, as described in our previous report^[28], and stored at −80 °C. Circulating cell-free DNA (cfDNA) was extracted from 2 mL of thawed plasma using the Concert Plasma Cell-Free DNA Extraction Kit and quantified with the Qubit dsDNA HS Assay Kit (Thermo Fisher). Although no dedicated fragment analysis was performed upfront, post-sequencing insert size metrics displayed the canonical cfDNA profile with a median length of approximately 165 bp and minimal high-molecular-weight contamination, confirming adequate sample quality for methylation analysis (Supplementary Fig. S1).

EM-seq library construction

EM-seq libraries were prepared following an adapted version of the protocol by Guo et al.^[28]. In brief, cfDNA inputs underwent end repair and adapter ligation with the VAHTS Universal DNA Library Prep Kit for Illumina V3 (Vazyme). The resulting fragments were treated with the EpiArt DNA Methylation Kit (Vazyme) to convert unmethylated cytosines. Library amplification was carried out with KAPA HiFi HotStart Uracil+ ReadyMix (Roche) and indexed using NEBNext Unique Dual Index Primers (NEB). To monitor conversion efficiency, two spike-in controls were added to every sample: unmethylated λ DNA and hypermethylated pUC-19 DNA. The proportion of non-converted λ DNA (theoretically 0%) and pUC-19 DNA (theoretically 100%) was assessed for each library. Across all 127 analyzed samples, the median λ non-conversion rate was 0.02% (range 0.00%−0.31%), and the median pUC-19 non-conversion rate was 98.55% (range 96.05%−99.58%), indicating robust conversion performance. Detailed per-sample quality metrics are provided in Supplementary Table S1.

Targeted capture and sequencing

For target enrichment, we pooled up to 24 barcoded libraries and hybridized them to a custom NAD oligonucleotide panel (Nanodigmbio) using NADprep Hybrid Capture Reagents. Following stringent washes, captured molecules were amplified on-bead with VAHTS HiFi Amplification Mix (Vazyme) and sequenced on an Illumina NovaSeq 6000 system (2 × 150 bp paired-end). To avoid capture bias related to the original methylation status, a five-probe design was employed for each CpG site^[28], the probes targeted: (1) the unconverted forward strand; (2) the methylated forward strand after C-to-T conversion; (3) the methylated reverse strand; (4) the unmethylated forward strand; and (5) the unmethylated reverse strand.

LC-specific methylation markers selection and panel design

Methylation data from 69 paired lung adenocarcinoma and adjacent normal tissues were retrieved from The Cancer Genome Atlas (TCGA) (Supplementary Table S2). We performed differential methylation analysis on all autosomal CpGs using the limma R package, applying a threshold of |log₂(fold change)| > 0.35 and p < 10⁻¹⁵. The |log₂(fold change)| > 0.35 threshold corresponds to approximately 1.28-fold change in methylation level, representing a commonly accepted standard for TCGA 450K array differential methylation analysis that balances sensitivity and specificity for biomarker discovery. The stringent cutoff of p < 10e-15 was selected to ensure extremely high statistical confidence. This threshold is well justified given the analysis of 69 paired samples with substantial statistical power, and it effectively minimizes false positives when screening across > 480,000 CpG sites. Together, these thresholds yielded 2,680 DMPs (~0.56% selection rate), consistent with the expected proportion of true differentially methylated sites in cancer vs normal tissue comparisons. To maximize the panel's clinical relevance, we supplemented these DMPs with additional CpG loci that had previously shown value for lung cancer detection^[29−31].

Data processing

Raw reads were processed with an updated version of the pipeline described in Guo et al.^[28]. Adapter trimming and quality filtering were performed using Fastp (v0.23.2)^[32]. Clean reads were aligned to the human reference genome (hg38) with Bismark (v0.23.0)^[33] and deduplicated. To exclude incompletely converted fragments, any read containing > 3 methylated cytosines in a non-CpG context (CH) was discarded^[28]. Methylation counts at each CpG were extracted with bismark_methylation_extractor using the flags --bedGraph --no_overlap --comprehensive --counts.

Development of LC screening model

We first estimated the individual discriminatory power of each CpG in the training set by calculating its area under the receiver operating characteristic curve (AUC). Markers with an AUC > 0.7 or < 0.3 were carried forward. Feature selection was then refined using Random Forest and least absolute shrinkage and selection operator (Lasso) regression. The resulting marker set was used to train a logistic regression classifier. Performance on the held-out test set was evaluated using the AUC metric, implemented with the roc_auc_score function from scikit-learn^[34].

Data and code availability

The beta-value matrix for the 22 diagnostic CpGs across all 127 plasma specimens is available in Supplementary Table S3. The custom R script for feature selection and model training is publicly available at https://github.com/liaohongyue/Methylation-predicts-lung-cancer Additional data and materials can be requested from the corresponding author.

To identify methylation sites pertinent to LC, the DNA methylation data of 69 pairs of cancerous and paracancerous healthy tissues were acquired from the TCGA database and analyzed. The specific information is provided in Supplementary Table S2. Differential methylation analysis revealed the presence of 2,680 LC-specific DMPs (Fig. 1a; Supplementary Table S4). They locate on 1,555 genes, among these DMPs, 1,417 (53%) exhibited elevated methylation levels in tumor tissues, while the remaining 1,263 (47%) demonstrated reduced methylation levels in tumor tissues (Fig. 1a). Enrichment analysis unveiled associations of these DMPs with various biological processes, including the processes of embryonic organogenesis and the establishment of anterior/posterior positional patterns, cellular components such as neuronal cell body, transporter complexes, and ion channel complexes, as well as molecular functions such as transcriptional modulation of DNA-binding activator function, characteristic of RNA polymerase II, and monoatomic ion channel activity (Supplementary Fig. S2). A predominance of LC-specific DMPs exhibited heightened methylation ratios within CpG islands (27.5%) and CpG shores (25.2%) (Fig. 1b). Furthermore, the heatmap representation of LC-specific DMPs underscores distinct methylation patterns observed between lung cancer and paracancerous tissues (Fig. 1c).

Figure 1. 

Selection of methylation markers. (a) Display of the pronounced differences in methylation levels between lung cancer (LC) samples and corresponding healthy paraneoplastic tissues. (b) The distribution of the LC-specific methylation sites. (c) Heatmap visualization depicting the prevalence of hypomethylated and hypermethylated DMPs contrasting lung cancerous with paraneoplastic healthy tissues.

Cohort characteristics

There were 127 plasma samples collected, including 50 from LC patients and 77 from non-lung-cancer individuals. The workflow is delineated in Fig. 2a and b. The lung cancer (LC) cohort consisted of 50 treatment-naïve patients with stage I disease, predominantly comprising adenocarcinoma (86.7%, n = 43), with a smaller subset of squamous cell carcinoma (13.3%, n = 7). The non-lung cancer (NL) control group consisted of 77 individuals with various benign pulmonary conditions or non-pulmonary diseases. The detailed clinical characteristics of this cohort are summarized in Supplementary Table S5. This cohort included a spectrum of patients with benign pulmonary nodules (e.g., infectious granulomas, inflammatory nodules, and benign tumors), non-nodular pulmonary diseases (e.g., pneumonia, interstitial lung disease), and a few individuals with non-pulmonary conditions. The inclusion of such a heterogeneous control group, particularly those with conditions that mimic lung cancer radiologically, was designed to rigorously test the specificity of our diagnostic model in a clinically relevant setting.

Figure 2. 

Schematic illustration of the study. (a) Study design. A total of 127 participants were included in this study. Researchers performed methylation sequencing of cfDNA in plasma. Eighty-six participants (35 LC, 51 NL) were assigned to training to build a machine learning model based on a logistic regression algorithm. Forty-one participants (15 LC, 26 NL) were assigned to testing to confirm the performance of the model. (b) Schematic diagram of the plasma cell free DNA testing process.

LC detection using plasma samples

All plasma samples underwent targeted enzymatic methyl sequencing, yielding an average of 21 million reads and an average unique read depth of 852× (Supplementary Table S1). Using 35 LC and 51 NL samples for training, we established a diagnostic model based on 22 optimally selected methylation features (Supplementary Table S6). This model demonstrated strong performance in the training set, with 88.57% sensitivity, 78.43% specificity, and an AUC of 0.8913 (Fig. 3a). When tested on the remaining 15 LC and 26 NL samples, it achieved 60% sensitivity, 76.92% specificity, and an AUC of 0.7025 (Fig. 3b).

Figure 3. 

Verification and construction of the LC screening model. (a) ROC curve for the model in the training dataset. (b) ROC curve for the model in the test dataset. (c) Gene ontology analyses applied to the genes characterized by the model features.

Functional enrichment analysis of the 16 genes harboring these 22 features revealed several significantly enriched KEGG pathways (p.adjust < 0.01), including cellular senescence, cell cycle, oocyte meiosis, and the PPAR signaling pathway. Additional pathways related to various cancers (e.g., gastric cancer, ductal carcinoma in situ) were also enriched (p.adjust < 0.05) (Fig. 3c).

Gene-level annotation of the 22 methylation features revealed that they are distributed across 16 genes, including known tumor suppressor genes (BRCA1, DLC1, MGMT, TFPI2) and oncogenes (CCND2, TLX3), as well as genes involved in transcriptional regulation, signal transduction, and metabolic processes (SDCCAG8, RASGRF1, ZNF536, etc.) (Supplementary Table S6). Notably, five of these genes (MGMT, CCND2, BRCA1, SYT5, DLC1) have been previously reported as methylation markers in lung cancer, supporting the biological relevance of our panel^[31,35−39]. The remaining genes represent novel candidates that may enhance the specificity of early-stage lung cancer detection.

No significant differences in methylation levels (beta values) at 22 features were observed across plasma sample subgroups stratified by patient characteristics (Supplementary Figs. S3−S6).

Early diagnosis of LC will significantly improve the 5-year survival rate. Yet, current screening methods, notably LDCT, exhibit limitations including overdiagnosis, cost-effectiveness issues, and radiation exposure^[8,40,41]. The invention of cfDNA-based early screening techniques holds promise as a revolutionary alternative^[42]. In this study, LC-specific DMPs from plasma cfDNA were identified from the TCGA database and used to construct an early diagnostic model with 60% sensitivity and 76.92% specificity. The biological rationale for LC detection via plasma cfDNA was indirectly corroborated by the enrichment of methylation markers in pathways related to transcriptional regulation, organismal development, and intracellular signaling.

A key methodological strength of our approach is the use of enzymatic methyl-sequencing (EM-seq) for methylation conversion. Compared to conventional bisulfite treatment, EM-seq minimizes DNA damage, reduces conversion errors, and provides higher coverage, particularly for low-input cfDNA samples^[28,43]. These features are critical for early-stage cancer detection, where the tumor-derived cfDNA fraction is often extremely low. Recent advances have demonstrated that enzymatic conversion better preserves nucleic acid integrity and enhances detection sensitivity by reducing fragmentation and base-conversion artifacts^[28,43], directly supporting the superior performance of our assay.

The biological plausibility of cfDNA methylation as an early diagnostic biomarker is reinforced by the underlying molecular events in tumorigenesis. Oxidative stress-induced apoptosis, characterized by the dysregulation of Bax/Bcl-2 and caspase cascades^[44], is a pivotal driver of lung cancer progression and is closely linked to aberrant epigenetic regulation. Such molecular perturbations may give rise to methylation alterations that are shed into the circulation, providing a detectable signal in cfDNA. Thus, the methylation signatures captured by our model may reflect early tumorigenic changes, strengthening the rationale for their use in non-invasive detection.

Compared with previous cfDNA methylation studies, our model demonstrated a relatively higher sensitivity for stage I lung cancer. For instance, Liu et al. reported a sensitivity of only 23% for stage I LC using cfDNA methylation testing^[45], whereas our EM-seq-based classifier achieved 60%. It is noteworthy that the effectiveness of cfDNA methylation screening can vary by histological subtype; studies have shown higher detection rates in squamous cell carcinoma than in adenocarcinoma, possibly due to differences in necrosis levels^[46,47]. In our cohort, 86.67% of stage I patients had adenocarcinoma, and the observed 60% sensitivity is therefore largely representative of this subtype. Recent work has also reported promising sensitivities for stage I lung adenocarcinoma, underscoring the clinical value of early detection in this population^[48]. Additionally, conventional serum protein biomarkers such as CEA and CYFRA 21-1 have insufficient sensitivity for early-stage disease (e.g., CYFRA 21-1 sensitivity as low as 5% at a common cutoff)^[9]. In contrast, our methylation-based classifier, which directly detects tumor-derived epigenetic alterations, offers improved sensitivity for early-stage lung adenocarcinoma and holds promise as a next-generation non-invasive biomarker strategy.

Despite these encouraging results, several limitations must be acknowledged. First and foremost, our study lacks an independent external validation cohort, which is the gold standard for confirming the robustness and generalizability of a diagnostic model. This absence directly limits the translational confidence of our classifier, as model performance can decline substantially when applied to populations with different demographic or clinical characteristics. We are actively collecting samples for a large-scale independent validation to address this critical gap. Second, the overall sample size was modest, which increases the risk of overfitting, as indicated by the drop in area under the curve (AUC) from the training set (0.8913) to the test set (0.7025). Third, the lung cancer cohort was predominantly composed of adenocarcinoma cases (86.67%), with only seven squamous cell carcinoma samples. Therefore, the reported sensitivity primarily reflects the detection of lung adenocarcinoma, and broad applicability across all non-small cell lung cancer subtypes remains to be validated in a histologically balanced cohort. Fourth, most participants were symptomatically diagnosed rather than identified through screening, which may overestimate the classifier's performance in an asymptomatic screening setting. Finally, biological and technical variables—such as inter-individual differences in cfDNA tumor fraction (often < 0.1% in early-stage disease)^[49], smoking history, sequencing depth, and pre-analytical handling—can influence methylation signals and may have contributed to the observed variability. Future assay iterations could integrate fragmentomic features or methylation haplotype burden as internal calibrators to account for variations in ctDNA shedding^[47].

Regarding future directions, prospective evaluation in LDCT-screened populations is essential. LDCT remains the cornerstone of lung cancer screening but suffers from a high false-positive rate (> 90%), leading to unnecessary invasive procedures^[50]. Combining cfDNA methylation biomarkers with CT imaging features has been shown to improve pulmonary nodule risk stratification^[51]. We therefore plan to investigate the integration of our methylation panel with radiological assessment to enhance diagnostic accuracy for indeterminate nodules. Furthermore, expanding the probe panel to include additional lung cancer-related markers and performing cross-validation in more diverse cohorts will be crucial to improve both sensitivity and generalizability.

In conclusion, this study demonstrates that a plasma cfDNA methylation signature, detected by the EM-seq technique, holds significant promise as a non-invasive tool for the early detection of lung cancer, particularly for adenocarcinoma, which represented the dominant histological type in our stage I cohort. Our diagnostic model, comprising 22 lung cancer-specific DMPs, achieved clinically informative sensitivity in this predominantly adenocarcinoma stage I population, a group for which current diagnostic options remain limited. The biological relevance of the selected markers is reinforced by their enrichment in cancer-associated pathways. Given the limited representation of squamous cell carcinoma and other subtypes, further validation in a more histologically diverse cohort is necessary to assess the generalizability of this cfDNA-based approach for early lung cancer diagnosis.