Multiomic characterization of malignant pulmonary nodules and development of a methylation-based diagnostic Model

Abstract

Background

The molecular distinction between benign and malignant pulmonary nodules remains a significant diagnostic challenge. While genomic drivers are well studied, multiomic integration of the epigenetic–transcriptional landscape and its translation into noninvasive tools are lacking.

Methods

We performed a multiomic characterization (genomic, epigenomic, and transcriptomic) of 158 pulmonary nodules. Unsupervised factor analysis integrated these layers to identify core regulatory axes. A 9-gene cell-free DNA (cfDNA) methylation classifier was developed and validated in blood and tissue cohorts.

Results

Genomic profiling revealed EGFR mutations (exclusive to malignant nodules) and MYC amplification as fundamental initiators of malignancy. Multiomic factor analysis (Factor 1) revealed profound genetic‒epigenetic synergy, in which these alterations dictate a permissive methylome, leading to aberrant epigenetic programming of chromatin accessibility, as well as epigenetic-transcriptional effects: hypomethylation at the promoters of cell cycle genes that augments their expression, and hypermethylation at immune related pathways gene loci that silences their transcription. This effect orchestrates formation of proproliferative (E2F target/G2M checkpoint) and “immune-cold” malignant phenotype, characterized by elevated Treg/CD8+ ratios and fibroblast recruitment. Notably, we observed a gradual accumulation of methylation aberrations along the premalignant-to-invasive continuum (adenocarcinoma in situ [AIS]→minimally invasive adenocarcinoma [MIA]→adenocarcinoma [ADC]), identifying progressive epigenetic dysregulation as a hallmark of tumor aggressiveness. Global methylome remodeling drives ADC progression through hypermethylation-mediated silencing of tumor suppressors (RASA3 and PPARG) and hypomethylation-activated oncogenic axes, specifically the GDF15 axis, which independently predict poor survival in patients with lung ADC in the TCGA cohort. We translated these tissue-derived insights into a 9-gene cfDNA methylation classifier, which achieved exceptional diagnostic accuracy across independent cohorts (training AUC = 1.00; test AUC = 0.93; tissue AUC = 0.96). Rooted in the biological “ground truth” of tissue dysregulation, this classifier functions specifically as a functional readout of the core cell cycle and proliferative pathways, offering a robust, noninvasive tool for the biology-informed risk assessment of pulmonary nodules.

Conclusions

This study delineates an epigenetic–transcriptional regulatory network that drives nodule malignancy. Our findings provide a robust theoretical foundation and a high-performance liquid biopsy tool for the precise, noninvasive diagnosis of pulmonary nodules.

Supplementary information

The online version contains supplementary material available at 10.1186/s12967-026-08382-w.

Background

Lung cancer continues to be the foremost cause of cancer-related mortality worldwide [1], with lung adenocarcinoma (ADC) accounting for approximately 50% of cases [2]. The development of invasive ADC typically originates from atypical adenomatous hyperplasia (AAH), which may progress sequentially to adenocarcinoma in situ (AIS), then to minimally invasive adenocarcinoma (MIA), and ultimately to fully invasive ADC [3]. Patients diagnosed with AIS or MIA exhibit a nearly 100% five-year overall survival rate following surgical resection, which is considered curative [4–6]. In contrast, poorly differentiated invasive ADC is associated with aggressive clinical behavior, including lymph node metastasis and poor survival outcomes [7]. Current prognostic grading systems, such as the International Association for the Study of Lung Cancer (IASLC) classification, highlight tumor differentiation as a pivotal prognostic factor; however, the molecular mechanisms underlying differentiation remain inadequately characterized. Although previous investigations have elucidated genomic [8], epigenomic [9], and transcriptomic [10] alterations during ADC progression using tissue-based methodologies, substantial knowledge gaps persist.

Foremost among these is the urgent need for noninvasive biomarkers capable of distinguishing malignant from benign pulmonary nodules, which are detected in 20–30% of screening computed tomography (CT) scans, to mitigate overdiagnosis and facilitate timely clinical intervention. While CT imaging remains the clinical standard, concerns regarding radiation exposure and high false-positive rates underscore the necessity for alternative diagnostic approaches [11]. Circulating tumor DNA (ctDNA) analysis has emerged as a promising noninvasive modality for cancer detection [12], therapeutic guidance, and disease monitoring [13]. Nevertheless, conventional ctDNA assays are limited in sensitivity for early-stage disease because of low ctDNA abundance and technical challenges, including DNA degradation induced by bisulfite conversion and restricted convergence of methylation sites in targeted sequencing. Importantly, peripheral blood ctDNA circumvents key limitations inherent to tissue-based analyses, such as sampling bias and spatial heterogeneity, by providing comprehensive molecular profiles of tumor-derived markers in circulation. Notably, DNA methylation signatures are particularly well suited for ctDNA analysis because of their high specificity for early detection across diverse cancer types [14]. These attributes uniquely position methylation-based liquid biopsy as a powerful tool to elucidate epigenetic dysregulation during ADC progression.

In response to these challenges, the present study focused on a central translational question: Can coordinated molecular signatures that differentiate malignant from benign pulmonary nodules be elucidated and subsequently harnessed to develop a robust, noninvasive diagnostic tool? We hypothesized that malignant transformation involves integrated dysregulation across genomic, epigenomic, and transcriptomic dimensions and that these coordinated changes can be detected in peripheral blood through stable methylation signatures.

To evaluate this hypothesis, we employed whole-genome bisulfite-free methylation sequencing (wGM-seq) of 154 peripheral blood samples, along with multiomic profiling, including whole-exome sequencing (WES), RNA sequencing (RNA-seq), and wGM-seq, of 63 tissue samples. Our objectives were as follows: Phase 1 (tissue-based discovery): To comprehensively characterize the molecular landscape of pulmonary nodules and identify malignancy-associated alterations across the genomic, transcriptomic, and epigenomic layers. Phase 2 (liquid biopsy translation): To leverage the epigenomic layer for clinical translation, given that DNA methylation (i) is stable in circulation, (ii) can be detected in cell-free DNA (cfDNA) with high sensitivity, and (iii) reflects both tissue-of-origin and disease-associated changes. This integrative approach connects molecular insights derived from tissue analyses with noninvasive liquid biopsy applications, thereby providing a framework for precision diagnosis and biological stratification of pulmonary nodules.

Methods

Sample acquisition

To study the multiomic characteristics of benign and malignant nodules and construct a model for the discrimination of malignant nodules on the basis of methylation, we collected 217 samples from 158 patients with pulmonary nodules (106 malignant nodules and 52 benign nodules) from Beijing Hospital or Emergency General Hospital between 2022 and 2024, including 63 tissue samples (17 malignant nodules, 21 benign nodules, and 25 normal adjacent tissues [NATs]) and 154 peripheral blood samples (102 malignant nodules and 52 benign nodules). Thirty benign nodules were confirmed by definitive pathological diagnosis, and 12 benign nodules showed no evidence of growth confirmed by CT scan for more than 1 year. None of the patients received chemotherapy or radiotherapy before surgery. WES, RNA-seq, and wGM-seq were performed on the tissue samples. Peripheral blood samples were subjected to wGM-seq.

WES

DNA was extracted using the FFPE kit (Promega) according to the manufacturer’s instructions. We constructed sequencing libraries from native DNA using the xGen® Exome Research Panel (Integrated DNA Technologies, Iowa, IA, USA) and the NEB Next Ultra DNA Library Prep Kit (Lot: NEB-0311611, NEB, UK) with KAPA polymerase (KapaBiosystems, Wilmington, MA, USA). WES was performed using a DNBSEQ-T7RS system (Geneplus-Beijing, Beijing, China), with 100-bp paired-end reads. The NAT served as a control to exclude germline mutations.

Fastp was used to filter out low-quality reads and reads that were too short and to cut off adapters from the raw reads to obtain clean reads. Clean reads were aligned to the human reference genome (hg19) using the Sentieon-genomics pipeline (version sentieon-genomics-201808). Read mapping was performed by bwa mem. The LocusCollector and Dedup algorithms were used to mark and remove duplicate reads. sQualCal was used in the base quality score recalibration stage. After alignment, sorting and deduping, realigner performed local realignment around the indels on the BAM file. DNAscope was used to call germline SNVs and InDels after only alignment and deduplication. ApplyVarCal was used in the variant quality score recalibration (VQSR) stage.

RNA next-generation sequencing

The NEBNext rRNA Depletion Kit (Human/Mouse/Rat) (NEB #Z1955E) was used to remove the targeted ribosomal RNA (rRNA). All RNA with a percentage of RNA fragments > 200 nucleotides (DV200) ≤50% skipped fragmentation was subjected to library preparation. After rRNA depletion and fragmentation, cDNA synthesis and NGS library preparation were performed using the NEBNext® Ultra™ II Directional RNA Library Prep Kit (NEB#E7760L). The library was quantitated using a Qubit 3.0 instrument (Life Invitrogen, USA), and its quality was assessed with a LabChip GX Touch instrument (PerkinElmer, USA). rRNA was discarded by aligning clean reads to the NCBI rRNA database using Bowtie2 (version: 2.2.8) after removing terminal adaptor sequences and low-quality data using fastp (version: 0.20.0). Clean reads without rRNA were aligned to the reference hg19 through STAR (version 2.6.1d). StringTie (version 2.0.6) was used for gene expression quantification, and the gene expression level was normalized by TPM (transcripts per million).

wGM-seq

Sample quality was assessed on the basis of the following DNA inclusion criteria: a total amount of 30 ng, as determined using a Qubit 3.0 fluorometer (Life Technologies), and fragment length distributions with a dominant peak at approximately 170 bp, as determined using the methods described above. DNA from each sample (50 ng) was used for library construction. The samples underwent two enzymatic steps, followed by the KAPA Hyper Library construction protocol (KAPA Biosystems, Wilmington, MA). The wGM-seq procedure was carried out following previously published protocols in cooperation with the Gene plus company [13]. We utilized the Hieff NGS® Ultima Pro DNA Library Prep Kit for Illumina (Yeason, cat. on. 12201ES96) for end repair, A-tailing, and adaptor ligation. After the magnetic beads were purified, 5-methylcytosine was oxidized to 5-acylcytosine by the TET enzyme, and then 5fC or 5caC was treated with pyridine borane and reduced to dihydrouracil (DHU). Finally, PCR amplification was performed, and a barcode was introduced to obtain the sequencing library. After this series of processes was completed, the methylated cytosine was identified as thymine (T) by the DNBSEQ-T7RS sequencing platform.

Somatic mutation calling

Somatic single-nucleotide variants (SNVs) and small insertions and deletions (InDels) were called and annotated by GATK Mutect2 (version 4.1.4.1) and VEP software, respectively. We retained mutations for which the variant allele frequency (VAF) was ≥ 0.01, the alternative allele depth was ≥ 5, and the FILTER field was equal to ‘PASS’. Microsatellite instability (MSI) was calculated using a published MSI sensor tool (v0.2).

Copy number variations (CNVs)

Somatic CNVs were identified with GATK (v0.5.11). GISTIC analysis was performed to identify significant arm-level and focal-level CNVs.

Mutational signatures

We used the “sigminer” R package to extract mutational signatures based on the COSMIC signature V3 (https://cancer.sanger.ac.uk/signatures/). Specifically, first, we used the read_maf method to load somatic mutations and tallied components in each sample, generating a sample-by-component matrix using the sig_tally method. Then, we used the sig_auto_extract method to automatically obtain the optimal signature number and extract signatures on the basis of Bayesian nonnegative matrix factorization (NMF). After the signatures were extracted, get_sig_similarity was performed to compare the identified signatures with the COSMIC signatures (version 3) to obtain their etiologies. We could also use the sig_fit method to compute the absolute and relative exposure of each identified signature for each sample.

T-cell–inflamed gene expression profile (GEP) scores and immune signatures

The GEP score was calculated as the weighted sum of the standardized expression values of 18 genes associated with IFN-γ signaling, cytotoxic effector molecules, antigen presentation, and active T-cell cytokines [15]. The relative abundance of immune cells in tissue transcriptomic data in pulmonary nodules was estimated by ImSig [16].

Immune cell infiltration and immunity

We used the single-sample gene set enrichment analysis (ssGSEA) method of the “GSVA” R package to quantify the relative scores of immune cells and tumor microenvironment (TME)-related signatures in the TME. Genes for immune cells and TME signatures were obtained from a previous study [17].

Differentially expressed genes and pathway enrichment

We performed differential expression gene analysis using the “DESeq2” R package. Differentially expressed genes were identified using a dual threshold: adjusted p < 0.05 and | log₂(fold change) |>1. The gene set enrichment analysis (GSEA) and visualization results were analyzed by the “fgsea” R package on the basis of genes preranked by log₂FC between groups. Pathways whose adjusted P was < 0.05 and whose absolute normalized enrichment score was greater than 1 were considered to be significantly enriched.

Identification and functional annotation of differentially methylated CpG sites (DMCs)

DMCs among the malignant, NAT, and benign pulmonary nodule groups were identified using the “DSS” R package, a statistical framework optimized for methylation sequencing data by decreasing dispersion estimates to increase the power of differential methylation detection. DMCs were identified with the following parameters set in DSS: q value = 0.05 and delta = 0.1.

The identified DMCs were subsequently subjected to genomic annotation using the “annotatr” R package, which maps CpG sites to gene features (e.g., promoters, exons, introns, intergenic regions) on the basis of the hg19 human genome assembly. For functional enrichment analysis, genes annotated to DMCs were subjected to pathway enrichment using the “enricher” function in the “clusterProfiler” R package. Enrichment analysis was conducted against the Kyoto Encyclopedia of Genes and Genomes (KEGG) database, with P adjusted for multiple testing using the Benjamini–Hochberg (BH) method. Only pathways with an adjusted p < 0.05 were considered to be significantly different.

Methylation block construction

First, we used “wgbstools segment” to divide blocks (at least 3 CpGs) in advance according to the GM-seq genome-wide methylation data of tissue samples, and the maximum allowed block size was 1000 bp. Second, we used “wgbstools homog” to calculate the extreme fragment ratio for each sample (the number of fragments with a fragment methylation rate greater than or equal to 0.8 as a proportion of the total number of fragments).

Marker selection and malignant/benign classification model construction

DNA methylation sequencing was performed on samples from patients with pulmonary nodules, including both blood and tissue samples. First, the extreme fragment ratio of each block from tissues for each sample was calculated; second, gene-level average methylation was computed by averaging the imputed values of all blocks annotated to each gene per sample; third, the blood cohort (n = 154, 102 malignant vs. 52 benign) was randomly split into a training cohort (70%, n = 108) and an internal test cohort (30%, n = 46). Feature selection was conducted solely in the training cohort, genes whose average methylation difference was > 5% between the malignant and benign groups were retained. Univariate receiver operating characteristic (ROC) curve analysis was subsequently performed, and genes whose area under the curve (AUC) was > 0.65 or <0.35 were selected for model construction. Missing methylation values were excluded; no imputation was applied. An independent tissue cohort (n = 33, 14 malignant vs. 19 benign) was used as an external validation set to verify the generalizability of the model. This cohort was collected from the same clinical center but with distinct sample types (tissue vs. blood) and independent pathological confirmation, ensuring that it was not involved in any step of model training or parameter optimization. The performance of the model in this independent cohort was evaluated using the same risk score formula derived from the blood training cohort, without any retraining or adjustment.

Finally, GSEA and gene set variation analysis (GSVA) were employed for DNA methylation data sequencing to elucidate the biological mechanisms underlying the predictive efficacy of the model. Spearman correlation analysis was performed for the model-predicted score to determine the relationships between the final predicted scores and gene expression or GSVA scores. GSEA and GSVA were performed using the “clusterProfiler” and “GSVA” packages, respectively, and information on the gene sets in “HALLMARK”, “KEGG” and “Gene Ontology (GO)” was obtained from http://www.gsea-msigdb.org/gsea/msigdb/index.jsp. A total of 1350 gene sets associated with tumors and the TME were analyzed as a point of importance. We summarized the 1350 gene sets into six classes according to the information of these gene sets provided in the molecular signatures database (MSigDB), including 421 “immunity”, 127 “cell cycle”, 315 “metabolism & energy”, 286 “genetic and epigenetic information”, 129 “extracellular matrix (ECM) & metastasis” and 72 “cell death” gene sets. The gene sets for which the raw P value was < 0.05 and the FDR was < 0.25 were considered statistically significant.

Motif enrichment analysis

Differential methylation may influence transcription factor (TF) binding. To identify overrepresented known TF motifs due to differential methylation sites after transformation, “findMotifsGenome.pl” from HOMER was applied to DMCs (±50 bp) overlapping with gene promoter regions. DMC regions with hypermethylation and hypomethylation were explored separately to determine the effects of different methylation statuses. The significantly enriched TFs were defined as those whose q value was ≤ 0.1.

Analysis of differential methylation blocks (DMBs) in blood samples

DMB analysis between benign nodules and AIS/MIA/well-differentiated ADC/moderately differentiated ADC/poorly differentiated ADC was performed by t test and the Mann‒Whitney test. Hyper-DMBs were defined by methylation values < 0.3 in the benign samples and a mean value across tumors that was at least 0.2 above the mean of benign samples. Hypo-DMBs were defined by methylation values ≥ 0.5 in the benign samples and a mean value across benign samples that was at least 0.2 below the mean for tumors.

The increasing and decreasing DMBs were performed using the “lmFit” function and the “ebayes” function (trend = TRUE) using “limma” by fitting to the following linear model: DMB~benign/AIS or MIA/well-differentiated ADC/moderately differentiated ADC/poorly differentiated ADC

We obtained the P and β coefficients in the linear model for a DMB. For increasing methylation from benign nodules (baseline) to AIS/MIA, well-differentiated ADC, moderately differentiated ADC, and poorly differentiated ADC: P_AIS/MIA < 0.05, P_{well-differentiated ADC} < 0.05, P_{moderately differentiated ADC} < 0.05 and P_{poorly differentiated ADC} < 0.05; and β_{poorly differentiated ADC} > β_{moderately differentiated ADC} > β_{well-differentiated ADC} > β_AIS/MIA > 0. For decreasing methylation from benign nodules (baseline) to AIS/MIA, MIA, well-differentiated ADC, moderately differentiated ADC, and poorly differentiated ADC: P_AIS/MIA < 0.05, P_{well-differentiated ADC} < 0.05, P_{moderately differentiated ADC} < 0.05 and P_{poorly differentiated ADC} < 0.05; and β_{poorly differentiated ADC} < β_{moderately differentiated ADC} < β_{well-differentiated ADC} < β_AIS/MIA < 0.

Deconvolution of tumor composition using DNA methylation

We used “MethylCIBERSORT” [18] for tumor cell fraction deconvolution. Among the results, CD4⁺ represents Tregs and conventional T helper (Th) cells, CD14⁺ represents monocytes and macrophages, CD19⁺ represents B cells, and CD56⁺ represents natural killer (NK) cells.

Virtual knockout of GDF15

Simulated knockout analysis of the GDF15 gene was performed on 10 early-stage ADC sample datasets derived from publicly available single-cell RNA sequencing (scRNA-seq) data (GSE131907) using “scTenifoldKnk” [19]. Briefly, “scTenifoldKnk” was employed to computationally mimic the genetic perturbation of GDF15, generating gene expression perturbation profiles for each sample. The genes were subsequently ranked by calculating their Z score (a metric reflecting the magnitude of expression change induced by GDF15 knockout), resulting in a sorted ranked gene list. This ranked list was used for GSEA to identify biological pathways associated with GDF15 perturbation. The normalized enrichment scores (NESs) of pathways were visualized using “pheatmap” in R, and statistically significant pathways with raw p < 0.05 and FDR < 0.25 were marked by asterisks.

Multiomic dataset integration

Multiomic data integration was performed using Multi-Omics Factor Analysis v2 (MOFA2) [20]. For each omic layer, features were preprocessed and filtered as follows: (1) Somatic mutations and CNVs were restricted to known cancer driver genes. (2) For mRNA expression, we retained genes that were differentially expressed between malignant nodules and adjacent normal tissue or between benign nodules and adjacent normal tissue (union of two differential gene sets) and normalized expression values using log2(exp+1) transformation. (3) For DNA methylation, we selected CpG loci at gene promoters that were differentially methylated between malignant nodules and adjacent normal tissue or between benign nodules and adjacent normal tissue (union of two differential methylation sets). The filtered multiomic datasets were input into MOFA2, with the following model parameters: maximum number of latent factors = 15, convergence mode = medium, and maximum number of iterations = 500. Variance partitioning was used to quantify the contribution of each factor to variation across omic layers. Pearson correlation was used to link factor scores to clinical covariates (malignancy status, histological subtype, and level of differentiation). Features with absolute loadings > 0.2 were prioritized to define factor-associated signatures, which were visualized via heatmaps and scatter plots. Functional enrichment of gene signatures was performed using clusterProfiler (Hallmark/Reactome/KEGG pathways).

Statistical analysis

All the statistical analyses were performed using R (version 4.2.2). The Wilcoxon rank-sum test (nonparametric test) was used to assess distribution differences between the malignant and benign groups. Fisher’s exact test was employed to evaluate differences in group composition ratios. The Kaplan‒Meier method was used for survival analysis. Statistical significance was defined as α = 0.05 for two-sided tests.

Results

Patient cohort description

This study included samples from a total of 158 patients who presented with pulmonary nodules, of which 52 were classified as benign and 106 as malignant (Table S1). The benign lesions included conditions such as chronic pulmonary inflammation, pulmonary tuberculosis, granuloma, bronchiolar adenoma, and hamartoma, among others. The malignant lesions included 4 AIS, 20 MIA, 2 small cell lung cancer (SCLC), 3 invasive squamous cell carcinoma (SCC), and 71 invasive ADC samples. Within the invasive ADC subgroup, tumor differentiation was further stratified into well-differentiated (n = 10), moderately differentiated (n = 42), and poorly differentiated (n = 17) categories (Table 1). Statistical comparisons revealed that sex, tobacco exposure, and drinking history were well balanced between malignant and benign groups. Malignant group had a significantly higher mean age than the benign group (Table 1, p < 0.001, Mann-Whitney U test). This observation aligns with the known epidemiology of lung cancer risk increasing with age.

Table 1.

Baseline clinical characteristics of patients with malignant and benign nodules

	Benign (N = 52)	Malignant (N = 106)	P
Sex			0.396
Female	27 (51.9%)	63 (59.4%)
Male	25 (48.1%)	43 (40.6%)
Age			<0.001
Mean (SD)	56.1 (15.0)	65.0 (10.3)
Median [min, max]	58.5 [22.0, 81.0]	67.0 [31.0, 81.0]
Smoking history			0.837
No	40 (76.9%)	84 (79.2%)
Yes	12 (23.1%)	22 (20.8%)
Drinking history			0.458
No	44 (84.6%)	94 (88.7%)
Yes	8 (15.4%)	12 (11.3%)
Stage
I	0 (0%)	98 (92.5%)
II	0 (0%)	5 (4.7%)
III	0 (0%)	3 (2.8%)
Missing	52 (100%)	0 (0%)
Pathological type
AIS	0 (0%)	4 (3.8%)
MIA	0 (0%)	20 (18.9%)
Invasive ADC	0 (0%)	71 (67.0%)
Invasive SCC	0 (0%)	3 (2.8%)
Missing	52 (100%)	7 (6.9%)
Differentiation
Poor	0 (0%)	17 (16.0%)
Moderate	0 (0%)	43 (40.6%)
Well	0 (0%)	10 (9.4%)
Missing	0 (0%)	2 (1.9%)

Differences in the genomic landscape between malignant and benign pulmonary nodules

To clarify the fundamental molecular disparities between malignant and benign pulmonary nodules, we initially examined the distribution characteristics of mutation burdens across various signaling pathways. The receptor tyrosine kinase (RTK) signaling pathway exhibited the greatest mutation burden, followed sequentially by transcription factors, the PI3K signaling pathway, other signaling cascades, and RNA abundance-related pathways. These genetic alterations were predominantly enriched in malignant nodules, indicating their critical involvement in tumorigenesis and malignant progression. Notably, mutations in the epidermal growth factor receptor (EGFR) gene emerged as a principal molecular driver in malignant cases and were present in 78.6% (11 out of 14) of such samples.

Analysis of mutational signatures revealed that the predominant COSMIC signatures were SBS6, associated with defective DNA mismatch repair, and SBS13, linked to the activity of the APOBEC family of cytidine deaminases, with SBS13 being more prevalent in malignant nodules. Metrics indicative of genomic instability, including homologous recombination deficiency (HRD), loss of heterozygosity (LOH), telomeric allelic imbalance (TAI), and large-scale state transitions (LSTs), were exclusively observed in malignant samples (Fig. 1A). Interestingly, two nodules with high microsatellite instability (MSI-H) were identified among benign nodules. Further pathway enrichment analysis demonstrated significant mutational divergence in the RTK–RAS pathway between malignant and benign nodules (p < 0.001, Fisher’s exact test; Fig. 1B), primarily driven by alterations in EGFR, MTOR, PIK3CA, MAP2K1, PTEN, NF1, and TSC2 (Fig. 1C). CNV analysis revealed malignant-specific amplification of oncogenes, including EGFR, MYC, HDAC1, LCK, CSF3R, and AGO1. EGFR functions as a canonical upstream driver of the RAS–ERK/MAPK signaling pathway, and its amplification leads to sustained pathway activation, thereby promoting malignant cellular phenotypes such as enhanced proliferation and invasion. Concurrent amplification of MYC, a downstream target of the EGFR pathway, further amplifies the oncogenic potential of this signaling cascade. Notably, MYC amplification was exclusively observed in moderately differentiated invasive ADC (Fig. 1D), which suggests a strong association with the tumor malignancy grade. These findings underscore the malignant-specific amplification of key oncogenes as a hallmark of pulmonary nodule malignancy.

Fig. 1.

Genomic profiling distinctions between malignant and benign pulmonary nodules. (A) overview of the most frequently occurring somatic mutations in malignant compared with benign pulmonary nodules. (B) comparative distribution of cancer-associated signaling pathways in malignant versus benign nodules. (C) mutation frequencies within genes of the RTK–RAS signaling pathway, with the left panel representing benign samples and the right panel representing malignant samples. (D) frequencies of significant chromosomal arm-level alterations as identified by GISTIC2 analysis

Transcriptomic profiles and the TME

To understand the functional consequences of the genomic alterations identified above, we next performed a transcriptomic profiling analysis across these two groups. The upregulation of genes within the HOX family (notably HOXB9 and HOXC10/11/13) (Tables S2-4), which are implicated in developmental processes and oncogenesis, as well as genes associated with cellular stemness (MUC21, NELL1, and GNGT1) (Fig. 2A), was detected in malignant nodules. Conversely, genes whose expression was downregulated in malignant nodules included those involved in metabolic regulation (ATP2B2 and MAS1), differentiation (DAZL, NDP, and TCF23), immune and stress responses (CHIT1, MMP9, and EPYC), cytoskeletal structure (MYBPC1 and OCSTAMP), and epigenetic modification (FBXO40) (Fig. 2A). Consistent with the genomic findings of RTK–RAS pathway alterations, GSEA revealed that malignant nodules were enriched in pathways related to cell cycle progression and oncogenic signaling, including E2F targets, the G2M checkpoint, MYC targets V1/V2, KRAS signaling downregulation, and glycolysis (Fig. 2B, C). In contrast, benign nodules were characterized by the activation of immune-related pathways, such as the interferon-γ response, inflammatory response, and IL6-JAK-STAT3 signaling pathways (Fig. 2B, C).

Fig. 2.

Differences in transcriptome profiles between malignant and benign pulmonary nodules. (A) volcano plot illustrating differentially expressed genes between malignant and benign nodules. (B) GSEA results for malignant and benign pulmonary nodules. (C) GSEA results for malignant versus NAT and benign versus NAT. (D) pathway enrichment analysis results for DEGs between invasive ADC and MIA. (E) assessment of T-cell–inflamed GEP scores and T-cell infiltration (ImSig) across malignant, benign, and NAT samples. (F) ssGSEA scores for immune and stromal cell populations in malignant, benign, and NAT samples. (G) ssGSEA scores for tumor microenvironment-related gene signatures across malignant, benign, and NAT samples

When MIA was compared with invasive ADC, the latter showed enrichment in interferon-α/γ signaling, cell cycle pathways, and E2F targets, whereas MIA exhibited upregulation of inflammatory pathways and the KRAS, TGF-β, and WNT signaling cascades (Fig. 2D, Table S5). The T-cell–inflamed GEP score and relative T-cell abundance were significantly reduced in malignant tissues, particularly in poorly differentiated tumors (Fig. 2E). Furthermore, immune infiltration scores for NK cells and effector T cells were higher in benign nodules and matched NATs than in malignant tissues.

The enrichment of specific cellular subsets, including IL-1 cancer-associated fibroblasts (CAFs) [21], B cells, and myofibroblastic CAFs (myCAFs), with functional roles in focal adhesion, ECM organization, and collagen formation [22], as well as the C10_COMP [23] fibroblast subset, characterized by the expression of metalloproteinases, TGF-β signaling pathway components, and ECM-related genes, was detected in malignant tissues (Fig. 2F). The regulatory T-cell (Treg) suppression signature was significantly elevated in malignant tissues compared with that in NATs, whereas signatures associated with immune checkpoint molecules, complement activation, epithelial–mesenchymal transition (EMT), and M1 macrophage polarization were more prominent in benign and NAT samples (Fig. 2G).

The expression of immune-related genes such as PDCD1LG2, CCL5, IL10, and GZMA was significantly lower in malignant nodules than in benign nodules and NATs, whereas VTCN1 expression was markedly higher in malignant nodules (Figure S1A). No significant differences in immune-related gene expression were observed between ADC and MIA; however, TNFSF9 expression was elevated in moderately differentiated ADC relative to poorly differentiated ADC, and IL2RA expression was higher in poorly differentiated ADC than in moderately differentiated ADC (Figure S1A).

Alterations in methylation profiles during neoplasia

Given the central role of epigenetic regulation in cancer, we sought to characterize the methylation landscape and its relationship with the transcriptional phenotypes observed. We initially conducted hierarchical clustering analysis on highly variable CpG sites. Clustering of the top 1% most variable CpG sites effectively differentiated malignant nodules from benign nodules and grouped nodules with their corresponding NATs, suggesting individual-specific methylation patterns (Fig. 3A). Compared with benign nodules and NAT samples, malignant nodules demonstrated global hypomethylation across various genomic regions, including those containing CpG shores and exons (Fig. 3B–C, Figure S2A). Previous studies have highlighted the critical role of epigenetic modifications in the early stages of ADC carcinogenesis [9].

Fig. 3.

Differences in methylation profiles between malignant and benign pulmonary nodules. (A) hierarchical clustering based on the top 1% most variable mCpgs across malignant, benign, and NAT samples. (B) comparative analysis of mean methylation levels across distinct genomic regions in malignant, benign, and NAT samples. (C) global methylation levels of mCs and mCpgs in malignant, benign, and NAT samples. (D) pathway enrichment analysis of differentially methylated CpGs located within gene promoter regions. (E) distribution patterns of DNA sequence motifs associated with hyper- and hypomethylated CpGs. (F) inference of immune and stromal cell infiltration derived from mCpG methylation profiles in malignant, benign, and NAT samples

We further assessed global methylation levels in peripheral blood. Interestingly, plasma cfDNA from patients with malignant lesions was hypermethylated (Figure S2B), which may be attributed to the predominance of cfDNA originating from leucocytes, such as neutrophils and lymphocytes [24, 25]. No significant differences were observed in the overall methylated cytosine content in peripheral blood across different pathological subtypes (Figure S2C). However, the methylation levels of methylated CpGs (mCpGs) varied significantly among AIS, well-differentiated invasive ADC, moderately differentiated invasive ADC, and poorly differentiated invasive ADC (Figure S2C).

DMCs identified among malignant, benign, and NAT samples were enriched in cancer-related pathways, including the ERBB signaling, focal adhesion, and VEGF/MAPK/calcium signaling pathways (Fig. 3D). Notably, mutations in EGFR, a key genomic driver in pulmonary nodules, are known to constitutively activate the MAPK pathway. This observation, in conjunction with the enrichment of methylation within MAPK signaling identified in the present study, highlights the integrated contribution of both genetic alterations and epigenetic modifications to the regulation of tumorigenesis. Motif enrichment analysis of DMCs distinguishing malignant from benign tissues revealed hypermethylated sites associated with IRF3 and IRF8 transcription factors, whereas hypomethylated sites were linked to members of the FOX family and TEAD2 transcription factors (Fig. 3E, Tables S6-8).

Additionally, immune cell infiltration was inferred from the DNA methylation data, which revealed significantly greater numbers of effector CD4⁺ T cells, Tregs, and fibroblasts in malignant tissues than in benign tissues (Fig. 3F). The ratio of Tregs to CD8⁺ T cells, a metric previously recognized as prognostically relevant in cancer [24, 25], was also increased in malignant tissues. These observations align with tumor-infiltrating lymphocyte (TIL) profiles derived from gene expression analysis, indicating that malignant nodules are predominantly enriched with fibroblasts and Tregs, which may facilitate tumor immune evasion.

Integration of multiomic data reveals a phenotype-driven molecular architecture in pulmonary nodules

Comprehensive multiomic integration was conducted to elucidate the principal molecular drivers underlying the difference between malignant and benign pulmonary nodules, with the aim of identifying robust targets to enhance noninvasive methylation-based diagnostic approaches. Unsupervised factor analysis was applied to the combined genetic variation, mRNA expression and DNA methylation datasets (Figure S3A), revealing Factor 1 as the predominant component capturing coordinated variability across the three omic layers (Figure S3B). Notably, Factor 1 and Factor 9 demonstrated strong discriminatory power in distinguishing malignant lesions from benign lesions (Figs. 4A-B, S3C).

Fig. 4.

Multiomic factors elucidate the differences between malignant and benign lesions. (A) PCA of multiomic factors in malignant and benign lesions. (B) distribution of Factor 1 and Factor 9 scores within malignant and benign groups. (C) rank-weighted gene importance plot, highlighting the top genes driving transcriptome differences between malignant and benign lesions. (D) gene expression of the top-ranked genes in Factor 1 between malignant and benign samples. (E) enriched biological pathways for genes positively and negatively correlated with Factor 1 and Factor 9. (F) enriched biological pathways corresponding to promoter-region methylation sites positively and negatively correlated with Factor 1 and Factor 9

Mechanistic investigations revealed that Factor 1 serves as a central molecular axis that integrates oncogenic activation events, such as MYC amplification and EGFR mutation, with transcriptional regulation of genes involved in cell cycle progression and EMT, in addition to epigenetic silencing mechanisms (Figs. 4C, S3D). Notably, mutant EGFR/MYC amplifications are linked to promoter hypomethylation of key epigenetic regulators (Fig. 4D), including SMARCA4 (chromatin remodeling), ERCC3/GTF2H4 (DNA repair), and KAT8/SGF29 (epigenetic modification). This focal epigenetic loss drives aberrant gene overexpression, fueling the proproliferative and genomic instability phenotypes that characterize the malignant transition of pulmonary nodules.

Scatter plots (Figure S3D, E) confirmed that the expression of genes positively correlated with EGFR mutation and MYC amplification along Factor 1 (e.g., CRYM, CLDN4, SMIM22, CLDN3, and ELF3) was increased in malignant samples, which is consistent with a proproliferative epithelial phenotype. Conversely, genes negatively correlated with the EGFR/MYC status (e.g., SLAMF7, LILRB4, and CHI3L1), which are associated with immune activation, exhibited decreased expression in malignant nodules, reflecting an immunosuppressive microenvironment. These findings align with the characteristic “noninflamed” tumor microenvironment of EGFR-mutant lung cancer, characterized by low CD8⁺ T-cell infiltration [26]. Concurrently, CpG sites demonstrated coordinated epigenetic alterations along Factor 1 (Figure S3D, F): Loci negatively correlated with EGFR/MYC amplification were characterized by hypomethylation of promoters linked to cell proliferation, invasion, and metabolic processes in malignant lesions, whereas loci positively correlated exhibited hypermethylation of promoters associated with immune response and cell death pathways. This promoter-specific epigenetic modulation highlights the role of Factor 1 in orchestrating oncogenic transformation via bidirectional regulation: hypomethylation facilitates the overexpression of proliferation-related genes, while hypermethylation suppresses immune-related gene expression, collectively contributing to the establishment of an immunologically “cold” TME.

Functional enrichment analyses using GSEA further substantiated these findings, revealing that Factors 1 and 9 are positively associated with pathways governing cell cycle progression and DNA replication but inversely correlated with immune response pathways (Fig. 4E). In contrast, methylation enrichment analysis revealed an opposing regulatory pattern, directly implicating promoter methylation as a negative modulator of oncogenic and EMT-related gene expression (Fig. 4F). Collectively, these results delineate a core epigenetic–transcriptional regulatory network that drives malignant pulmonary nodule pathogenesis and identify actionable methylation targets that hold promise for refining the precision of noninvasive diagnostic models.

Prognostic biomarkers for early malignant pulmonary nodules

To identify methylation-based prognostic biomarkers for early-stage malignant pulmonary nodules, thereby enhancing the prognostic utility of a noninvasive diagnostic model, we initially characterized methylation blocks across malignant nodules, benign nodules, and NATs. Utilizing methylation data derived from malignant pulmonary nodules, benign pulmonary nodules, and NAT samples obtained from patients with benign nodules, we identified a total of 2,611,037 methylation blocks, as detailed in the Methods and Materials section. Principal component analysis (PCA) of these methylation blocks effectively distinguished malignant nodules from benign nodules and NATs; however, it did not distinguish between benign nodules and NATs (Fig. 5A).

Fig. 5.

Prognostic biomarkers in malignant pulmonary nodules. (A) PCA of methylation blocks across malignant, benign, and NAT samples. (B) quantification of DMBs comparing AIS, MIA, well-differentiated ADC, moderately differentiated ADC, and poorly differentiated ADC with benign samples (blood samples). (C) pathway enrichment analysis of promoter-region DMBs in AIS, MIA, and variously differentiated ADC relative to benign samples. (D) average methylation levels of upregulated and downregulated DMBs across AIS, MIA, and ADC subgroups. (E) heatmap depicting specific DMBs in malignant, benign, and NAT samples, including their genomic and CpG region distributions. (F) pathway enrichment of specific DMBs contrasting malignant with nonmalignant (benign plus NAT) samples. (G) correlation analysis between GDF15 methylation levels and TPM. (H) boxplot illustrating GDF15 TPM across malignant, benign, and NAT samples. (I) Kaplan‒Meier survival analysis evaluating the association between GDF15 TPM and disease-free survival in stage I lung ADC from the TCGA dataset. (J) boxplot comparing mean GDF15 methylation levels in tumor tissues and paired NAT samples from the TCGA lung ADC dataset. (K) scatter plot depicting the correlation between GDF15 methylation and mRNA expression. (L) Kaplan‒Meier survival curve stratified by high versus low GDF15 methylation levels

To verify the detectability of tissue-assembled methylation blocks in peripheral blood, as well as to elaborate on methylation profiles associated with distinct pathological subtypes, we quantified the extreme fragment ratio of tissue-derived cfDNA in peripheral blood by employing tissue-assembled methylation blocks. The number of hypermethylated CpG sites, relative to that in benign cases, progressively increased from AIS to poorly differentiated ADC. Conversely, the number of hypomethylated CpG sites increased from AIS to moderately differentiated ADC but subsequently decreased from moderately to poorly differentiated ADC (Fig. 5B). DMBs across pathological types were predominantly localized within gene introns, intergenic regions, and CpG intergenic regions (Figure S2D). Notably, the DMBs identified in AIS samples were enriched in pathways related to p53 signaling, oxidative phosphorylation, fatty acid metabolism, and the interferon gamma response (Fig. 5C). Both MIA and AIS samples were enriched in pathways related to reactive oxygen species and the apical surface (Fig. 5C). Notably, adipogenesis and protein secretion pathways were specifically enriched in the DMBs of the moderately differentiated ADC group, whereas E2F target genes and complement pathways were uniquely enriched in the DMBs in the poorly differentiated ADC group (Fig. 5C). We identified 20 DMBs with decreased methylation and 622 DMBs with increased methylation when comparing benign nodules with poorly differentiated ADC nodules (Fig. 5D). The methylation levels of genes involved in the regulation of myogenesis, hypoxia (e.g., PPP1R3C), and the G2M checkpoint (e.g., CASP8AP2) progressively decreased with increasing tumor invasiveness, whereas the methylation levels of genes such as MGLL (associated with fatty acid metabolism) and PPARG (involved in cholesterol homeostasis) gradually increased (Figure S2E). Promoter hypermethylation of tumor suppressor and immune-related genes (including RASA3, MIR519E, PPARG, and FCRL6) alongside promoter hypomethylation of oncogenes and immunosuppressive genes (such as STEAP1B, ERMP1, and PPP1R3C) may contribute to tumor progression. We further identified 15,464 malignant-specific DMBs predominantly located within intergenic and intronic regions (Fig. 5E), which were enriched for genes involved in cell adhesion and the PI3K–AKT signaling pathway (Fig. 5F). Integrative analysis of DNA methylation and gene expression revealed a significant inverse correlation between the methylation level and expression level of GDF15 (r = −0.65, p = 0.011), with GDF15 being hypomethylated and overexpressed in malignant samples (Fig. 5G, H). Moreover, elevated GDF15 expression was associated with poor disease-free survival (DFS) in early-stage (stage I) adenocarcinoma patients in The Cancer Genome Atlas (TCGA) database, with a median DFS of 23.08 months in the low-expression group versus 17.75 months in the high-expression group (p = 0.0073, Fig. 5I). Validation in the TCGA lung ADC cohort confirmed significantly lower GDF15 methylation in tumor tissues than in NATs (p < 0.001, Fig. 5J), a negative correlation between GDF15 methylation and mRNA expression (r = −0.39, p = 0.053, Fig. 5K), and an association between higher GDF15 methylation levels and improved overall survival, even in early-stage lung ADC (log-rank p = 0.0047, Fig. 5L). Functional insights derived from virtual knockout assays of GDF15 demonstrated substantial perturbations in the expression of metabolism and energy- and immunity-related genes following GDF15 ablation, a regulatory relationship that was consistently observed across the majority of lung ADC samples (Figure S2F).

Methylation-based diagnostic model

On the basis of our finding that epigenetic alterations are a hallmark of malignant nodules and are detectable in blood, we sought to develop a plasma cfDNA methylation classifier that could translate this biological insight into a noninvasive diagnostic tool. A total of 187 well-classified pulmonary nodule samples were subjected to wGM-seq, comprising 154 blood samples (102 malignant, 52 benign) and 33 tissue samples (14 malignant, 19 benign). Among genes in these samples, 64 genes whose average methylation difference exceeded 5% between the malignant and benign groups were selected for ROC curve analysis. Nine genes (DISC2, MIR4698, MIR4709, MIR891B, OR2T8, OR7D4, OR8D2, TAS2R50, and TMED7) were subsequently incorporated into the predictive model.

The model demonstrated excellent discriminatory performance, with AUC values of 1.00 (95% CI: 1.00–1.00) in the training cohort, 0.93 (95% CI: 0.85–1.00) in the test cohort, and 0.96 (95% CI: 0.89–1.00) in the tissue cohort (Fig. 5A). When the blood cohorts (training and test cohorts) were combined, the model achieved a sensitivity of 0.93 and a specificity of 0.88, yielding an overall accuracy of 0.92 (Fig. 5B). The negative predictive value (NPV) and positive predictive value (PPV) in the blood cohort were 0.87 and 0.94, respectively. Comparable performance metrics were observed in the tissue cohort, with a sensitivity of 0.93, specificity of 0.89, accuracy of 0.91, NPV of 0.94, and PPV of 0.87 (Fig. 5B).

Logistic regression analysis revealed that both patient age and the methylation-based prediction score were significant predictors of benign versus malignant classification (all p < 0.001, Table S9). Notably, the prediction score exhibited a particularly strong association with malignant status, underscoring its critical role in differentiating between benign and malignant nodules within this model.

GSEA revealed that 79.4% (100/126) of cell cycle-related pathways were significantly enriched in the low-risk score group (Fig. 5B). Similarly, GSVA revealed that 81.0% (102/126) of cell cycle-related pathways were significantly enriched in the low-risk score group, whereas ECM- and metastasis-associated pathways were predominantly enriched in the high-risk group (Fig. 5C). These results are consistent with the established inhibitory role of DNA methylation on gene expression, suggesting that lower risk scores correspond to decreased cell cycle activity, which aligns with the increased proliferative capacity characteristic of malignant nodules. This mechanistic concordance reinforces the validity of the model in capturing authentic molecular distinctions between benign and malignant pulmonary nodules.

Discussion

This study establishes a comprehensive multiomic framework that uncovers the fundamental molecular disparities separating benign and malignant pulmonary nodules, delineates the molecular progression from AIS to invasive malignancy. By integrating genomic drivers, epigenetic remodeling, and microenvironmental changes, we identified a core regulatory axis—Factor 1—that orchestrates the formation of a proproliferative and “immune-cold” phenotype, providing a biological rationale for the high performance of our 9-gene cfDNA methylation classifier.

On CT scans, certain benign pulmonary nodules may display features mimicking malignant lesions, including classic signs like speculation, pleural retraction, and lobulation [27]. This significant overlap in imaging phenotypes creates a major bottleneck for accurate discrimination of nodule benignancy and malignancy using CT imaging alone. Our study identified that malignant nodules were characterized by a high frequency of EGFR mutations (78.6%) and MYC amplification. These genetic alterations function as core oncogenic drivers that not only sustain tumor cell proliferation and viability [9] but also exacerbate metabolic reprogramming and genomic instability [28–30]—distinct molecular signatures that are largely undetectable in their benign counterparts. Notably, MYC amplification was uniquely enriched within moderately differentiated ADC, suggesting that this genomic alteration likely emerges as a late-stage event during tumorigenesis. Our analysis suggests that this genetic background is closely coupled with an epigenetic landscape characterized by global hypomethylation (decreased mC and mCpG levels) in malignant tissues [31]. Notably, this hypomethylation is not random but enriched in motifs for pioneer transcription factors such as FOXA1 [32, 33] and FOXM1 [34], suggesting that epigenetic remodeling provides the necessary promoter modification for oncogenic programs to modulate cell cycle progression, epithelial cell differentiation.

Consistent with previous reports [35, 36], our findings confirm that malignant nodules are characterized by heightened proliferative activity, metabolic reprogramming, and an immunosuppressive landscape; conversely, benign nodules exhibit a distinct phenotype defined by robust immune activation. Specifically, benign lesions are characterized by “immune-active” signatures, including elevated interferon-γ responses and increased effector T-cell and NK-cell infiltration. In contrast, malignant nodules exhibit an “immune-cold” or immunosuppressive phenotype, marked by reduced T-cell–inflamed GEP scores and the recruitment of Tregs and CAFs, which could drive ECM remodeling and TGF-β signaling, likely facilitating tumor invasion and metastasis [37–40]. DNA methylation data provide additional support for these immunosuppressive phenotypes, as evidenced by the enrichment of Tregs and CAFs in malignant nodules.

Compared with previous studies that focused on single-layer omic analysis [8–10], our research offers a more comprehensive view of malignant nodule formation. While earlier reports highlighted EGFR as a driver [41], our study found a multiomic Factor 1 linked its mutation and MYC amplification to the hypomethylation of the promoters of essential epigenetic regulators, including members of the SWI/SNF chromatin remodeling complex, histone acetyltransferases and transcriptional regulator. This genetic–epigenetic crosstalk reinforces a permissive chromatin state that facilitates the activation of growth-promoting transcriptional programs. Consequently, these findings underscore how genomic mutations actively shape a dysregulated methylome to drive the malignant evolution of pulmonary nodules. Intriguingly, Factor 1 revealed that alterations in EGFR and MYC are intrinsically linked to the transcriptomic upregulation of EMT related genes, including CLDN3/4 and ELF3 [42]. Previous studies have reported that ELF3 as a transcription factor to regulate CLDN4 to promote a more aggressive cancer phenotype [43]. EGFR mutation and MYC amplification were also negatively associated with immune-modulatory, offering indirect explanation for the limited clinical benefit of immunotherapy in patients with EGFR mutations. Genetic-transcription axis integrates epithelial phenotype regulation，metabolic reprogramming, and immunosuppressive regulation. In addition to their utility as potential discriminative biomarkers for benign versus malignant nodule stratification, they offer novel therapeutic targets for combination strategies targeting the MYC/EGFR pathway. Furthermore, Factor 1 reveals that this immunosuppression is also epigenetically regulated; hypermethylation of immune-response promoters in malignant tissues suppresses antitumor immunity, suggesting that epigenetic interventions might be necessary to “prime” these early-stage tumors for future immunotherapy. Taken together, these observations underscore the sophisticated interplay among genetic variations, epigenetic remodeling (predominantly DNA methylation), and transcriptional regulation—an interaction that modulates the proproliferative phenotype and immune microenvironment of pulmonary nodules, thereby establishing a definitive mechanistic link between epigenetic perturbations and the functional divergence of benign versus malignant lesions.

Additionally, our identification of global hypomethylation in tissue, as opposed to hypermethylation in plasma cfDNA, adds a layer of complexity to the existing understanding of liquid biopsy signals. The observed plasma hypermethylation in malignant cases, potentially driven by the contribution of leukocyte-derived cfDNA and broader peripheral biological backgrounds, rather than being a direct mirror of the landscape of the primary tumor [24, 25], contrasts with direct tissue-based hypomethylation, highlighting the need for sophisticated classifiers that account for systemic biological backgrounds. Accordingly, this study incorporated DNA methylation data from pulmonary nodule tissues, with the specific aim of constructing tumor tissue-contextualized methylation blocks and eliminating confounding factors derived from peripheral blood. By leveraging tissue-specific epigenomic landscapes as a reference, we were able to circumvent systemic biological noise—most notably leukocyte-derived cfDNA signals—that often compromise the specificity of liquid biopsies. This “tissue-first” strategy ensures that the identified molecular markers reflect intrinsic epigenetic remodeling within the nodule rather than peripheral biological backgrounds.

Building upon this robust biological foundation, we conducted a 9-gene cfDNA methylation classifier, which achieved exceptional diagnostic performance, with an AUC of 0.93–0.96 across independent cohorts. Unlike previous models that rely solely on clinical features (e.g., the Mayo Clinic or VA model), our model is based on the biological “ground truth” of tissue-derived epigenetic dysregulation. This biological grounding is further evidenced by the results of GSVA and GSEA, which revealed that the risk scores of the model directly parallel the transcriptomic enrichment of E2F targets and G2M checkpoints observed in malignant tissues, effectively serving as a functional readout of the underlying tumor biology. Therefore, the cfDNA methylation classifier not only serves as a diagnostic tool but also reflects the underlying epigenetic dysregulation observed in malignant tissues. The enrichment of cell cycle pathways in the low-risk score group aligns with the proliferative transcriptional phenotype seen in tissue, suggesting that the model captures biologically relevant epigenetic silencing of growth-promoting genes. The diagnostic classifier serves as a functional readout of the epigenetic dysregulation observed in tissue: by capturing the hypermethylation of genes associated with cell cycle control, it reflects the same proliferative phenotype that distinguishes malignant from benign nodules at the transcriptional level. While our model demonstrated high accuracy in both an internal hold-out blood cohort and an independent tissue cohort in this study, the most definitive validation requires application to a large, prospective, and multicenter cohort of plasma samples. Future studies should focus on such external validation to confirm generalizability across diverse populations and clinical settings before clinical implementation. While our diagnostic model demonstrated high accuracy, we acknowledge that cfDNA methylation profiles can be influenced by nonmalignant conditions, such as age and systemic inflammation. Although our logistic regression analysis indicated that the methylation signature contributed independently of age, future studies incorporating larger, prospectively collected cohorts should systematically evaluate and adjust for such clinical and biological confounders to further refine the specificity of the model.

Beyond establishing a classifier to distinguish malignant nodules, our study further uncovered GDF15 as a potential prognostic biomarker that predict the clinical outcomes of malignant nodules. It supported by its epigenetic dysregulation, overexpression in tumors, and association with poor survival in public data. Consistent with the observations from the prior study [44] that highlighted GDF15 as a stress-responsive cytokine linked to adverse clinical outcomes in tumor patients, our cohort extends these findings by validating GDF15 as a potential prognostic biomarker specifically in malignant pulmonary nodules—a critical gap in existing literature, which has primarily focused on GDF15’s prognostic role in advanced cancers (e.g., hepatocellular carcinoma, colorectal cancer) rather than early-stage pulmonary lesions. While the causal mechanisms remain to be experimentally elucidated, in silico network analysis suggests that GDF15 knockdown perturbs key pathways, such as the metabolism and energy. Although these findings are correlative, they provide a strong rationale for prioritizing GDF15 in future functional and clinical validation studies.

Despite the insights gained, several limitations warrant consideration. First, our cross-sectional design lacks longitudinal tracking, which restricts a comprehensive understanding of the epigenetic transition along the premalignant–invasive–metastatic continuum. While our multiomic analysis identified key factors, the absence of spatial transcriptomic and proteomic data limits our ability to fully map the regulatory cascades between the epigenome and the translational landscape. Second, while GDF15 emerged as a critical epigenetic driver linked to a poor prognosis, its direct causal role in tumor progression and immune exclusion requires further functional validation using CRISPR-based modulation or patient-derived organoids. Finally, for clinical translation, the current 9-gene cfDNA classifier should be optimized for cost efficiency through targeted sequencing and refined to increase sensitivity in patients with low ctDNA abundance. Notably, owing to the retrospective nature of this study and the availability of clinical samples with matched methylation data, we did not obtain an additional multicenter external validation cohort. This limitation should be acknowledged: the current validation relies on internal test (blood-derived) and independent tissue-derived cohorts from a single center, which may restrict the generalizability of the model to diverse patient populations (e.g., populations of different ethnicities or clinical backgrounds or samples analyzed by different detection platforms). Future prospective studies with multicenter, large-scale sample cohorts are needed to further validate the robustness and clinical applicability of the model. Future studies should aim for a multimodal diagnostic framework on the basis of multicenter data by integrating methylation scores with radiomics and exosome-based biopsies to further improve diagnostic precision for early-stage lung cancer.

Conclusions

By leveraging multiomic insights, we developed a 9-gene cfDNA methylation classifier that transforms complex tissue-based molecular landscapes into a robust clinical tool. The predictive power of the classifier stems from its ability to reflect the divergence between malignant (proliferative and immune-cold) and benign (immune-active) trajectories, providing a noninvasive, biology-informed framework for the early stratification of pulmonary nodules (Fig. 6E).

Fig. 6.

Multiomic characterization, diagnostic model development, and clinical validation for pulmonary nodule malignancy risk stratification. (A) AUC values of the risk score in the training, testing, and tissue validation cohorts. (B) performance metrics of the methylation-based diagnostic model across these cohorts. (C) radar chart illustrating pathways enriched in the low-risk group as determined by GSEA. (D) radar chart of pathways enriched in the low-risk score group based on GSVA. E) multiomic characterization and translational model: from tissue-based biological insights to clinical application of liquid biopsy

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

wGM-seq

Whole-genome methylation sequencing

cfDNA

Circulating free DNA

EGFR

Epidermal growth factor receptor

ADC

Lung adenocarcinoma

AAH

Atypical adenomatous hyperplasia

AIS

Lung adenocarcinoma in situ

MIA

Minimally invasive ADC

IASLC

International Association for the Study of Lung Cancer

CT

Computed tomography

ctDNA

Circulating tumor DNA

WES

Whole-exome sequencing

VQSR

Variant quality score recalibration

rRNA

Ribosomal RNA

SNVs

Single-nucleotide variants

InDels

Insertions and deletions

VAF

Variant allele frequency

MSI

Microsatellite instability

CNV

Copy number variation

NMF

Nonnegative matrix factorization

WGD

Whole-genome doubling

ssGSEA

Single-sample Gene Set Enrichment Analysis

TME

Tumor microenvironment

GSEA

Gene set enrichment analysis

TF

Transcription factor

DMBs

Differential methylation blocks

Th

cell T helper cell

SCLC

Small cell lung cancer

HRD

Homologous recombination deficiency

LOH

Loss of heterozygosity

TAI

Telomeric allelic imbalance

LST

Large-scale state transition

GEP

Gene expression profile

NK

cell Natural killer

NAT

Normal adjacent tissue

CAF

Cancer-associated fibroblast

Treg

Regulatory T-cell

EMT

Epithelial–mesenchymal transition

DMCs

Differentially methylated CpGs

TILs

Tumor-infiltrating lymphocytes

PCA

Principal component analysis

ROS

Reactive oxygen species

DFS

Disease-free survival

TCGA

The Cancer Genome Atlas

ROC

Receiver operating characteristic

AUC

Area under the curve

NPV

Negative predictive value

PPV

Positive predictive value

GSVA

Gene set variation analysis

ECM

Extracellular matrix

Authors’ contributions

WZ, TX, QW, and MS had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. WZ, TX, QW, and XG conceived and designed the study. WZ, TX, JW, JZ, and MS undertook the statistical analyses. WZ, TX, and MS wrote the first draft of the manuscript. YZ and YL are the study guarantors. WZ, TX, QW, JW, MZ, ZW, YZ, and YL interpreted the data, reviewed the paper, and critically revised the manuscript for important intellectual content. All the authors read and approved the final version of the manuscript.

Funding

This work was supported by the National High-Level Hospital Clinical Research Funding [BJ-2021-232].

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the institutional review board of Beijing Hospital and registered at ClinicalTrials.gov (Approval No.: NCT05415670; Approval Date: 2022–06-08). We have complied with all relevant ethical regulations for work with human participants in accordance with the Declaration of Helsinki, and written informed consent was obtained.

Consent for publication

Not applicable.

Competing interests

All the authors declare that they have no competing interests.