2.1 Patient samples and characteristics

We collected 432 archived Formalin-fixed paraffin-embedded (FFPE) tissue samples from lung nodules. Samples with insufficient Genomic DNA (gDNA) amounts, low gDNA quality, low library yields, or failed sequencing quality control metrics were excluded. After thorough analysis, we obtained a final set of 422 samples, which were classified into three groups: benign (n = 124), preinvasive (AAH+AIS, n = 32), and invasive (MIA+IA, n = 266). The average age of the benign group was 50 years, younger than the preinvasive group (55 years) and the invasive group (57 years) (p < 0.05). In terms of gender, 59.4% of the preinvasive and invasive groups were female, while 62.1% of the benign group were male. Specifically, 78.2%, 90.6%, and 81.2% of the benign, preinvasive, and invasive groups were nonsmokers, respectively. Regarding nodule density, 51.9% of all nodules were solitary nodules (SNs), 37.9% were pure ground glass nodules (pGGNs), and 9.2% were part-solid nodules (PSNs). Remarkably, 62.5% of preinvasive and 49.2% of invasive nodules were pGGNs, while only 7.3% of benign nodules were pGGNs. Additionally, 87.9% of benign patients had SNs, but this percentage decreased to 41.0% among the invasive groups. Among the benign nodules, 29.8% were inflammation, while 17.7% were tuberculosis. It is worth noting that 97.7% of invasive nodules were very early lung cancers, classified as American Joint Committee on Cancer (AJCC) stage I (Table 1).

2.2 Somatic mutation landscape of lung nodules

In the WES study, we analyzed a total of 422 lung nodule tissues (124 benign, 32 AAH and AIS, 266 MIA and IA) and their matched white blood cells (WBCs). Twenty-two significantly mutated genes were identified in the invasive samples (Figure 1A), among which the two most frequently mutated genes were EGFR (61%) and TP53 (24%), consistent with previous reports.¹¹

The frequency of EGFR mutations increased from benign (3%), to preinvasive (AAH+AIS, 16%), MIA (35%), and IA (73%) (Figure 1A). The EGFR mutation frequencies were statistically higher in females (vs. male, p = 0.0016) and nonsmokers (vs. smokers, p = 0.0004) (Figures 1A and S1). The most prevalent types of EGFR mutations identified were exon 21 p.L858R (39%) and exon19 deletions (36%) (Figure S2A), and these mutations were highly prevalent in MIA (81.25%, 50% exon 21 p. L858R, 31.25% exon 19 deletions) and IA (78.91%, 39.46% exon 21 p. L858R, 39.46% exon 19 deletions) (Figure S2B). These hotspots EGFR mutations observed in our early-stage lung cancer cohort were consistent with those found in late-stage lung cancers in another Chinese lung cancer cohort.^{15, 16} Other EGFR mutations, such as exon 20 insertions, exon 18 p.G719S/A, exon 18 p.E709A, exon 21 p.L861Q/R, and exon 20 p.T790M, were also detected (Figure S2A and B). Notably, EGFR mutations in benign nodules were infrequent and inactive, as demonstrated by functional annotation of mutated genes (Figure S6).

TP53 mutations were significantly more prevalent in IA (35.43%) compared to MIA (2.2%) (Figure 1A) and were scarcely detected in preinvasive and benign samples. This suggests that TP53 mutations might not influence nodule growth but could play an important role in promoting invasiveness development, which is in line with a recent finding.¹⁷ Patients harboring TP53 mutations were more likely to be smokers, strongly associated with older age (≥ 55 years), male gender, solid nodules, and higher TMB (≥ 1.39/Mb) (Figure S1).

The mean TMB of this cohort was 3.34/Mb, which was lower than the studies of late-stage lung cancers.¹⁸ The average TMB in invasive samples (MIA, IA) was 2.94/Mb. When comparing within nodule types, the TMB of solid nodules (S) was significantly higher than that of pure GGN (G) (p < 0.001) (Figure 1B). Furthermore, when TMB was compared within different pathological subtypes, we observed that it was higher in invasive stage compared to benign and preinvasive stage (Figure 1C). Interestingly, within the IA groups, the TMBs of pure GGN, part-solid, and solid nodules gradually increased as the nodule density increased (Figure 1C). When intersecting pathological subtypes with EGFR mutation status, the median TMB in the EGFR mutation (MUT) group was significantly lower compared to the EGFR wild type (WT) group in IA (p < 0.01), consistent with a previous study.¹⁹ However, this trend was reversed among benign nodules, possibly due to the very low proportion of the EGFR MUT group in benign group (Figure 1D). We further compared EGFR mutation status among nodule types based on TMB and found that the median TMB in the EGFR mutation group was statistically higher than that in the EGFR WT group with pure GGNs (G) and solid nodules (S) (Figure 1E) (p < 0.001). Considering the impact of TP53 mutation on TMB in IA nodules, a group of TP53 MUT & EGFR WT samples appeared to have the highest TMB compared with other combinations of TP53 WT/MUT and EGFR WT/MUT (Figure S3), which is in agreement with other studies^{16, 20} showing that TMB was higher in tumors with TP53 MUT or EGFR WT.

To assess the functional significance of mutated splicing genes, nodule samples were divided into splicing gene MUT and WT groups, and alternative splicing events from RNA-Seq data were analyzed (Figure 2A). In invasive samples, a significantly higher number of splicing events were detected in samples carrying splicing gene mutation (Figure 2B). A total of 417 genes were identified with splicing events (splicing events ≥10). Pathway enrichment analysis revealed that these genes with splicing events were primarily involved in the cell cycle and other cancer-related pathways, such as KRAS signaling and heme metabolism (Figure S4A). RBM10 (12%) and RBMX (RNA Binding Motif Protein X-Linked) (7%) were the most common splicing mutation genes. However, no significant difference in alternative splicing events was observed across pathological subtypes or between MUT and WT groups (excluding any splicing mutation) (Figure S4B and C).

Furthermore, we analyzed the detected CNVs by gene annotation based on the significance of differential genes between malignant and benign nodules. However, no lung cancer-related gene was identified (data not shown).

2.3 The transcriptomic features and immune microenvironment of lung nodules

To elucidate the transcriptomic landscape of lung nodules, RNA-Seq was conducted on 226 lung nodules samples, comprising 52 benign, 29 AAH+AIS, and 145 MIA+IA cases. Through the integration of WES and RNA-Seq data, four distinct molecular subgroups were identified based on unsupervised hierarchical clustering analysis of the gene expression profiles (Figure 3A).

Group 1, classified as the EGFR-driven group, was predominantly composed of invasive nodules harboring EGFR mutations. Markedly increased expression of early and late estrogen response (ER) and androgen response (AR) genes was observed. Estrogen and its receptors have been postulated to play a role in the development and progression of lung cancer.²¹ Group 2, referred to as the EGFR/TP53 co-mutation group, was enriched for IA nodules carrying EGFR/TP53 co-mutations. These nodules exhibited upregulation of genes associated with mitotic spindles, the G2/M checkpoint, and E2F targets. Most of these genes are linked to cell cycle regulation and tumor proliferation. Notably, the glycolysis pathway was also upregulated in Group 2, aligning with the Warburg effect in tumor metabolism and supporting the idea that tumors generate energy through glycolysis to fuel proliferation and metastasis via metabolic reprogramming.²² Group 3, designated as the “inflamed group,” was enriched with benign nodules and included some MIA and IA samples. Increased expression of numerous inflammation-related genes, such as IL-2 (Interleukin-2), IL-6 (Interleukin-6), and IFN (Interferons), was observed. Group 4 represented the stem-like group, which comprised the majority of preinvasive (AAH, AIS) nodules and some MIA nodules. This group exhibited upregulation of stem cell pathways, including NOTCH, Wnt/β-catenin, and hedgehog signaling, implicating the important role of stem cell pathways in tumor initiation (Figure 3A).

To explore the immune infiltration in lung nodules, we computed the immunophenotype score and immune cell infiltration score using single-sample Gene Set Enrichment Analysis based on the RNA-seq data, similar to a previous study.²³ Notably, immune cell infiltration seemed more pronounced in the inflamed (Group 3) and stem-like (Group 4) groups, where benign and preinvasive nodules tend to cluster. In contrast, EGFR-driven (Group 1) and EGFR/TP53 co-mutation (Group 2) groups appeared relatively “cold” with fewer immune infiltrates (Figure 3A and B). The EGFR-driven group (Group 1) exhibits a lack of T cell infiltration, especially CD8 T cell infiltration, which is consistent with the clinical observation that lung cancer patients with EGFR mutation tend to have an unfavorable response to PD-1/PD-L1 checkpoint inhibitors.²⁴ In the EGFR/TP53 co-mutation group (Group 2), CD8 T cell infiltration was also low. Conversely, Group 2 showed higher levels of Th2 and Treg cell infiltration. A recent study revealed that the loss of TP53 function in tumor cells results in the accumulation of Treg cells.²⁵ Due to the inflammatory characteristics of benign nodules, immune cell infiltration in the inflamed group was significant, which aligns with the observation that this group demonstrated activation of numerous inflammatory pathways, such as IL-2/STAT5 and IL-6/JAK/STAT3. Interestingly, the stem-like group (Group 4) exhibited very high natural killer (NK) cell infiltration (Figure 3B). We also identified a strong correlation between the WNT/β-catenin pathway (Pearson's correlation coefficient (PCC) = 0.8) and a moderate correlation of the NOTCH pathway (PCC = 0.48) in relation to NK cells (Figure 4A and C). NK cells are a crucial component of the innate immune system, contributing to immunosurveillance.²⁶ The Wnt/β-catenin signaling pathway is involved in NK cell development and activates the development and function of natural killer T (NKT) cells.²⁷ We postulated that NK cells infiltrating preinvasive nodules were part of the host's innate immunosurveillance mechanism to prevent invasion. To evaluate whether this association is specific to preinvasive nodules, we examined it using TCGA data. The results indicated that the correlation was moderate for both the WNT/β-catenin (PCC = 0.37) and NOTCH (PCC = 0.56) signaling pathways (Figure 4B and D).

2.4 The DNA methylation characterization of lung nodules

To investigate the DNA methylation patterns in lung cancer, 83 DNA samples (comprising 18 benign, 19 MIA, 46 IA) were subjected to targeted DNA methylation sequencing. The data from a custom panel, containing 12,899 preselected probes, were clustered into four groups (M1, M2, M3, and M4) using unsupervised hierarchical clustering. Consequently, the M2 and M4 groups exhibited distinct methylation patterns across pathology subtypes. In the M2 group, markers were hypomethylated in benign and MIA nodules, while they were hypermethylated in IA nodules. In contrast, markers in the M4 group were hypermethylated in benign and MIA nodules but hypomethylated in IA nodules (Figure 5A).

To investigate gene functions, a hallmark gene set was utilized to conduct pathway enrichment analysis (Figure 5B). Genes within the M2 group were enriched in myogenesis, epithelial mesenchymal transition (EMT), IL2/STAT5 signaling, and the upregulation of KRAS signaling. EMT plays a role in tumor invasion and metastasis.²⁸ IL2/STAT5 signaling regulates Treg cell differentiation and maturation.²⁹ KRAS signaling was related to cell proliferation. Genes in the M4 group were enriched in ER and AR pathways, TGFβ signaling, and the downregulation of KRAS signaling. TGFβ signaling modulates tumor invasion and metastasis through EMT, angiogenesis, and immunosuppression.³⁰ The enrichment of ER and AR pathways in M4 was consistent with the transcriptomics analysis of EGFR-driven group (Group 1) in this study, further supporting the involvement of hormonal pathways in lung cancer development (Figure 3A).

2.5 The integration of gene expression and DNA methylation

There is evidence of both strong positive and strong negative correlations between gene methylation and gene expression.³¹ To pinpoint potentially relevant DNA methylation and gene expression alterations during the progression of lung nodules, an integrated analysis of DNA methylation and gene expression was conducted. Pearson's correlations were calculated to identify statistically significant inverse correlations between DNA methylation and gene expression of enriched genes from pathway analysis, yielding fourteen genes with DNA methylation alterations situated at the promoter regions. In the M2 group, five genes, including BDNF (brain-derived neurotrophic factor), CD48 (CD48 molecule), COX7A1 (cytochrome c oxidase subunit 7A1), IKZF1 (IKAROS family zinc finger 1), and TNFRSF1B (TNF receptor superfamily member 1B), displayed progressively increasing methylation levels, accompanied by corresponding decreases in gene expression levels from benign, MIA, to IA (Figure 6). The highest inverse correlation values in the M2 group were observed for CD48 (−0.7), COX7A1 (−0.65), and TNFRSF1B (−0.61). Our findings regarding COX7A1 were consistent with recent studies showing that COX7A1 had higher methylation and lower expression in the NSCLC cell line,³² which suppressed cell viability by downregulating tumor autophagy in vitro.³³ Previous studies have demonstrated that TNFRSF1B, a gene associated with apoptosis, had mutations and lower expression in NSCLC patients,^{34, 35} which is supported by our findings. However, it has been reported that BDNF had a high expression level in lung squamous cell carcinoma and NSCLC, linked to tumor proliferation and invasion.³⁶ BDNF can bind to TrkB receptors to activate ERK/CREB, PI3K/AKT, and NF-kB pathways, ultimately enhancing tumor survival, proliferation, and migration.³⁷ In contrast, our study revealed that BDNF had higher methylation and lower expression in IA, suggesting that BDNF may have distinct functions in early-stage versus late-stage lung cancers.

Conversely, in the M4 group, nine genes, including CDC6 (cell division cycle 6), CLDN4 (claudin 4), EGF (epidermal growth factor), EGFR, ESRP2 (epithelial splicing regulatory protein 2), GRB7 (growth factor receptor bound protein 7), KRT8 (keratin 8), SFN (stratifin), and SLC22A18 (solute carrier family 22 member 18), exhibited progressively decreasing methylation levels and increasing levels of gene expression from benign, MIA, to IA (Figure 7). The highest inverse correlation values were observed for EGF (−0.78), ESRP2 (−0.69), and CDC6 (−0.63). EGF and EGFR had lower DNA methylation and higher expression in MIA and IA. Studies have suggested that the stimulation of EGFR by its ligand, EGF, may have pro-tumorigenic effects, such as enhancing cell migration, with clinical implications.³⁸ Upregulation of EGFR leads to activation of EGFR signaling and promotes tumor progression through EGFR downstream pathways, including PI3K/Akt/mTOR and RAS/MEK/ERK pathways.^{39, 40} Our findings emphasize the importance of the EGFR pathway and uncover a novel mechanism regulating this pathway during the neoplastic transformation from benign to IA. A previous study reported that NSCLC patients with strong overexpressing of CDC6 had a significantly poorer prognosis compared to those with normal expression levels,⁴¹ consistent with our research. KRT8, a type II intermediate filament protein, is over-expressed in various cancers, including lung cancer.⁴² While overexpression of KRT8 has been attributed to DNA copy number gains,⁴³ our results suggested that DNA hypomethylation may represent another mechanism for the upregulation of KRT8. Collectively, these genes hold promise as potential biomarkers for diagnostics and drug targets for the early interception of lung cancer.

4.1 Participating patients and sample collection

This study enrolled patients with lung nodules (5–30 mm in diameter) detected by CT/LDCT scan at The First Affiliated Hospital of Guangzhou Medical University in China between 2018 and 2019. None of the patients had received any preoperative cancer treatments. FFPE tissue samples were acquired from subsequent surgical resections, and matched WBCs were also collected to subtract germline mutations. Both genders were included in the study, and smoking history information was gathered. The pathological results of all samples were determined based on surgically resected tissue sections according to the 2015 WHO Histological Classification of Lung Cancer.⁶⁸ Written informed consents was obtained from all participants.

Sections of 5−10 µm in thickness were cut from the FFPE tissue samples of 432 patients with lung nodules. Approximately 8−10 mL of whole blood sample was drawn from each patient using a Cell-Free DNA BCT blood collection tube (Streck, Inc. Cat# 218962). WBCs were separated immediately upon receipt of the whole blood samples according to the standard protocol and stored at −80°C until DNA isolation. Purity of tumor cells were estimated by two independent pathologists by reviewing Hematoxylin and Eosin (H&E) staining slides, with a minimum tumor cell content of 10% required for subsequent sequencing.

4.2 Nucleic acid isolations

gDNA and total RNA were simultaneously extracted from FFPE tissue samples using AllPrep FFPE DNA/RNA Kit (Qiagen, Cat# 80234) according to the manufacturer's instructions. gDNA from WBC samples was isolated using the DNeasy Blood & Tissue Kit (Qiagen, Cat# 69504). The concentrations of DNA and RNA were measured using the Qubit dsDNA HS Assay Kit (Thermo Fisher Scientific, Cat# Q32854) and Qubit RNA HS Assay Kit (Thermo Fisher Scientific, Cat# Q32855), respectively. The quality of DNA and RNA were assessed using the Agilent High Sensitivity DNA Kit (Cat# 5067-4626) and the Agilent RNA 6000 Pico Kit (Cat# 5067-1513) on a 2100 Bioanalyzer Instrument (Agilent Technologies, Inc.), respectively. Only samples with DV200 values (the percentage of fragments > 200 nucleotides) ≥ 30% were selected for RNA-Seq.

4.3 Whole exome sequencing and RNA-Seq

WES and RNA-Seq were carried out by Mingma Technologies Co., Ltd. (Shanghai, China). In brief, 300 ng of gDNA from FFPE tissue or WBC samples were fragmented into 150−200 bp pieces (mean size) using a LE220-plus Focused-ultrasonicator (Covaris, Inc.). DNA libraries were prepared using SureSelectXT Reagent Kit (Agilent, Cat# G9611A) following the manufacturer's instructions. Whole exome enrichment was conducted using SureSelectXT Human All Exon V6 Kit (Agilent, Cat# 5190−8864). For RNA from FFPE tissue, 200 ng was converted into sequencing libraries using the TruSeq RNA Exome Kits (Illumina, Cat# 20020189, 20020490, 20020492/20020493, 20024145). The Resulting libraries were quantified using the Qubit dsDNA HS Assay Kit, and their size distributions were analyzed by Agilent DNA 1000 Kit (Cat# 5067-1504) on a 4200 TapeStation System (Agilent Technologies, Inc.). Sequencing was performed in the 2×150-bp paired-end mode on the HiSeq X Ten or NovaSeq 6000 platform (Illumina, Inc.).

4.4 Targeted DNA methylation assay

Full details of DNA fragmentation, bisulfite conversion, and targeted methylation sequencing were described previously.⁶⁹ Briefly, gDNA from 83 selected samples were fragmented into approximately 200 bp (peak size) using an M220 Focused-ultrasonicator (Covaris, Inc.) following the manufacturer's instructions. 50 ng of purified, fragmented gDNA was used for the subsequent bisulfite conversion step. Bisulfite conversion was carried out using the EZ DNA Methylation-Lightning Kit (Zymo Research, Cat# D5031) according to the manufacturer's protocol. The AnchorIRIS library preparation technology was employed for targeted methylation analysis. AnchorIRIS prehybridization library construction was performed using the AnchorDx EpiVisio Methylation Library Prep Kit (AnchorDx, Cat# A0UX00019) and AnchorDx EpiVisio Indexing PCR Kit (AnchorDx, Cat# A2DX00025). Target enrichment was performed using AnchorDx EpiVisio Target Enrichment Kit (AnchorDx, Cat# A0UX00031). A custom-made lung cancer methylation panel (LC Panel), which includes 12,899 preselected probes enriched for lung cancer-specific methylation regions, was utilized in the present study.⁷⁰ After probe hybridization, the resulting libraries were sequenced in the 2 × 150-bp paired-end mode on the NovaSeq 6000 platform.

4.5 Data quality control

We used the default parameters of the FastQC (version 0.11.2) (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/) software to perform quality control (QC) on all samples. Skewer (v0.2.2)⁷¹ software was applied to remove splice sequences from the sequencing platform raw data (Fastq). Sequences with a length of at least 75 bp were filtered by the splice removal process, and the resulting filtered data (clean reads) were used for downstream bioinformatic analyses. The exome capture efficiency and targeted region coverage were evaluated across the entire sample set. Lung nodules were sequenced at a mean targeted region coverage of 417×, with 97.7% coverage of ≥20×; WBCs were sequenced at a mean targeted region coverage of 215×, with 96.7% coverage of ≥20×, qualifying for subsequent variant analysis. The quality scores of 30 (%) for RNAseq was greater than 93.5. The average sequencing data amounts were 62.3 M, the average clean reads obtained after removing splices and low quality were 62.2 M, the average clean rate was 99.2%, and the average mapping rate was greater than 95%. For DNA methylation sequencing, the bisulfite conversion efficiency was greater than 99.2%, and the mean targeted region coverage was 785×, with 99% coverage of > 30×.

Three samples were excluded from further analysis due to low coverage in WES, two samples due to low coverage in RNA-seq, and five samples due to a mismatch between tumor and normal samples. Therefore, 422 patients were included in the final analysis. Among them, 422 had qualified WES data, 228 had qualified RNA-seq data, and 226 had both WES and RNA-seq data (paired). For DNA methylation sequencing, 20 benign and 80 malignant samples (including 20 with TP53 mutations, 20 with EGFR mutations, 20 with TP53 & EGFR co-mutations, and 20 without mutations) were included. Seventeen samples were excluded due to failing sequencing QC. Finally, 83 samples had qualified DNA methylation sequencing data.

4.6 Somatic variant identification

After removing adapters and low-quality reads, the Sentieon (version 201911)⁶⁹ pipeline with default parameters was implemented to process the following steps sequentially: reads alignment to the human genome reference assembly hg19 using the Burrows–Wheeler Aligner (BWA) algorithm and sorting, removal of PCR duplicates, InDel realign, base quality score recalibration, somatic mutation calling including single nucleotide variations (SNVs) and short insertion/deletions (indels), and variant quality score recalibration. During the mutation calling stage, the reads from the tumor sample were compared with the paired blood sample from the same patient to generate the somatic mutation list. Somatic mutations were then filtered and retained only with the variant allele frequency (VAF) ≥ 0.05 and supported by at least three reads. They were further annotated using the Variant Effect Predictor (VEP) package.⁷⁰

4.7 Tumor mutation burden (TMB)

TMB was defined as the total number of somatic nonsynonymous mutations (SNVs or indels) in the tumor exome for each patient. The number was divided by the total size of the targeted regions to calculate the TMB score in counts/Mb. The Agilent SureSelect Human All Exon V6 kit was used, and its estimated total targeting size (exome) is 38 Mb. TMB was regressed on each gene to test for a statistical association between genes and TMB. Genes were identified according to the following criteria: correlation |r| > 0.4 and false discovery rate (FDR) < 0.05.

4.8 Copy number variation (CNV)

CNVkit v0.8.5 (https://cnvkit.readthedocs.io/en/stable/index.html) was used to analyze the copy number variation. In particular, for tumor/normal pairwise analysis, all normal samples were pooled together and then each tumor sample was compared with the normal pool to obtain tumor-specific CNV, with filters of |log2ratio| ≥0.4 and bin number ≥ 5.

4.9 Splicing event counting

After assigning transcripts, alternative splicing events were identified using vast-tools. Percent Splice In (PSIs) calculated by vast-tools was then filtered using the cutoff psi > 80 (for IR, Alt5, Alt3) or psi < 20 (for EX) to count event number for each splicing type. Only high-frequency splicing events (occurring in ≥ 10 samples) were preserved. The total splicing event number of a sample is the sum of the event number for each splicing type. The splicing gene list was curated based on the mRNA splicing gene set from Reactome.⁷²

4.10 Differential gene expression analysis

RNA-seq reads were mapped to human genome reference assembly hg19 using STAR (version 020201).⁷³ The raw read counts were then normalized by log2-counts per million normalization. The list of differentially expressed genes between groups was generated using DESeq2.⁷⁴ Biological pathway enrichment was performed using the Kyoto Encyclopedia of Genes and Genomes (KEGG) and MSigDB hallmark gene sets. The FDR was calculated using the Benjamini and Hochberg (BH) false discovery rate (FDR) algorithm. We used an FDR ≤ 0.05 as the criterion to select significantly enriched KEGG pathways.

4.11 KEGG pathway enrichment analysis

The KEGG pathway database was used to perform pathway enrichment analysis. ClusterProfiler was used to obtain the latest online database and conduct an over-representation test with the “enrichKEGG” function.⁷⁵

4.12 Clustering analysis of ssGSEA scores based on hallmark gene set

The single sample gene set enrichment analysis (ssGSEA) scores of each sample based on Hallmark 50 gene sets were calculated, and then unsupervised hierarchic clustering of samples was performed based on the Hallmark ssGSEA scores to generate subgroups using the seaborns (version 0.11.2) package. The clustering result was then visualized by matplotlib in Python, and the immune cell ssGSEA scores corresponding to each sample were also plotted in the clustering heatmap. The Hallmark gene set used here was downloaded from the Gene set enrichment analysis (GSEA) molecular signatures database (gsea-msigdb.org).

4.13 Immune cell signature score

RNA-seq reads were assembled using StringTie2 (version 1.3.5), and FPKM and TPM data matrices were generated. TPM data were used to calculate the immunopheno score using the immunophenogram R package.⁷⁶ FPKM data were used to calculate the immune infiltration score using the ssGSEA method.²³ Infiltration cell types were clustered by heatmap analysis with the seaborns package and visualized by matplotlib in Python.

4.14 DNA methylation analysis

Sequencing adapters and three prime low-quality bases were trimmed from raw sequencing reads using fastp (v0.19.6) software.⁷⁷ In sillico converted hg19 was used as reference genome and cleaned reads were mapped to reference genome using Bismark (v0.18.2) software.⁷⁸ Aligned reads were evaluated by Picard (v2.5.0) for metrics that measured the performance of target-capture based bisulfite sequencing assays (http://broadinstitute.github.io/picard). The “bismark_methylation_extractor” tool was used to calculate methylation level for each targeted CpG site.

4.15 Differential methylation analysis

Differential methylation analysis was conducted using R package DSS (version 2.14.0).⁷⁹ Differentially methylated CpGs (DMs) were first identified (criteria: FDR < 0.05, delta > 0.05), and adjacent DMs were further merged into differentially methylated regions (DMRs). DMRs were intersected with protein-coding genes (hg19 Ensembl (v75), n = 20,232) by using annovar.⁸⁰

4.16 Unsupervised hierarchical clustering of DNA methylation datasets

The preselected 12,899 probes were used for unsupervised hierarchical clustering. The methylation level of each targeted regions was calculated as the ratio of the methylated CpGs and the total sequenced CpGs (sum of methylated and unmethylated CpGs). Before clustering, the methylation levels of each targeted region were Z-score normalized. The R function “hclust” was used to perform hierarchical clustering with “ward.D2” as the clustering algorithm. GSEA and MSigDB with BioCarta gene set, oncogenic signature gene sets and Hallmark gene sets were also used for biological pathway enrichment analysis.

4.17 Statistics analysis

Unless specified otherwise, Pearson's chi-square test was used for p-value calculations between two categorical variables. The two-sided Mann–Whitney U test was used for comparisons between two continuous variables. The most frequently mutated genes and their distribution across different clinical features, the significantly different mutation genes between groups, and the mutually exclusive or co-occurring mutated genes were calculated and visualized using the MAFtools package in R.⁸¹ GSEA was performed using the GSEA software.⁸² Pearson's correlation analysis was employed to calculate the coefficient between gene expression and methylation data. The p-value < 0.05 was determined to be statistically significant for all analyses.