Epigenetic repression of oncogenic signaling defines immune-permissive tumor states associated with response to immune checkpoint inhibitors in advanced NSCLC

Abstract

Background

Immune checkpoint inhibitors (ICI) have transformed the treatment landscape of advanced non-small cell lung cancer (NSCLC), yet response rates remain below 40% and currently used biomarkers such as programmed cell death-ligand 1 (PD-L1) expression and tumor mutational burden (TMB) show limited predictive accuracy. We investigated whether tumor epigenetic programs define biological states associated with response to immunotherapy.

Methods

We performed comprehensive multiomic profiling of pretreatment tumor biopsies from 32 patients with advanced NSCLC receiving anti-programmed cell death protein 1/PD-L1 therapy. Genome-wide DNA methylation, RNA sequencing, locus-specific transposable element (TE) quantification, and whole genome/exome sequencing were integrated to identify molecular determinants of ICI response. Predictive models were evaluated using nested leave-one-out cross-validation. Baseline peripheral blood mononuclear cells (PBMC) immune profiling and functional studies using PI3K inhibition and DNA demethylation in NSCLC cell models provided orthogonal validation.

Results

Responders exhibited coordinated promoter hypermethylation and transcriptional repression of PI3K-Akt, Ras, MAPK, Wnt and NF-κB signaling pathways, independently of downstream genomic alterations. These changes were accompanied by enhanced antigen presentation, T cell receptor-major histocompatibility complex signaling, CD8+ T cell activation and other immune-related pathways, consistent with an immune-permissive tumor state. Independent baseline PBMC profiling provided orthogonal validation of these immune-related findings. Responders also displayed TE expression patterns linked to antiviral signaling and viral mimicry, while high TE load was associated with prolonged overall survival. Finally, we identified a 12-CpG hypermethylation signature associated with longer progression-free survival and developed an integrated multiomic random forest classifier with robust predictive performance under nested cross-validation (area under the curve=0.833), outperforming PD-L1 expression and TMB.

Conclusions

Our findings define an epigenetically orchestrated immune-permissive tumor state associated with clinical benefit from immune checkpoint blockade in advanced NSCLC. Coordinated repression of oncogenic signaling together with enhanced antigen presentation, antiviral signaling and adaptive immune activation provides a biological framework linking tumor epigenetic states with immunotherapy efficacy. These results support integrated epigenetic biomarkers for precision immunotherapy and provide a rationale for exploring epigenetic and PI3K-targeted combination strategies.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Epigenetic mechanisms are increasingly recognized as key regulators of tumor immunogenicity and response to immune checkpoint inhibitors, but how epigenetic reprogramming shapes tumor–immune interactions and establishes immune-permissive tumor states remains poorly understood.

WHAT THIS STUDY ADDS

• This study identifies an epigenetically defined tumor state associated with immunotherapy response, characterized by coordinated repression of oncogenic signaling, activation of viral mimicry programs through transposable element expression, and distinct systemic immune profiles, providing mechanistic insight into how epigenetic reprogramming shapes tumor-immune interactions in advanced non-small cell lung cancer.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• These findings provide a rationale for integrating epigenetic biomarkers into precision immuno-oncology and for evaluating therapeutic strategies combining epigenetic modulation or PI3K pathway inhibition with immune checkpoint blockade in future prospective studies.

Background

Lung cancer remains the leading cause of cancer-related mortality worldwide,¹ with non-small cell lung cancer (NSCLC) representing approximately 85% of total lung cancer cases.² During the last decade, immunotherapy based on immune checkpoint inhibitors (ICI) has emerged as an impactful approach to NSCLC treatment. It has been established as the first-line treatment for advanced NSCLC, either in monotherapy (programmed cell death-ligand 1 (PD-L1) staining ≥50%) or in combination with chemotherapy (PD-L1 staining <50%),³ and in combination with chemotherapy as neoadjuvant treatment in resectable stage III NSCLC.⁴ Unfortunately, the response rate to ICI remains below 40% for NSCLC,^{5 6} and a proportion of patients who respond, ultimately relapse with drug-resistant disease. In this scenario, the lack of a clear understanding of the mechanisms driving ICI response and the need for effective biomarkers to stratify patients at baseline represents key challenges in oncology practice.

As mentioned, PD-L1 positive staining by immunohistochemistry has been the criteria to undergo ICI therapy in advanced NSCLC,^{3 7} but it has demonstrated low specificity, heterogeneity and lack of standardization, which hampers efficient clinical implementation.^8–10 Although a high tumor mutational burden (TMB) can guide the use of ICI in patients with solid tumors,^{11 12} results in NSCLC have been inconsistent: it has shown to predict ICI benefit independent of PD-L1 staining in NSCLC but without benefits regarding overall survival (OS)¹³ and it has not outperformed PD-L1 staining in several trials.^{14 15} Moreover, discrepancies in aspects such as the methods to obtain the TMB (whole exome sequencing (WES) or targeted panels), its calculation and thresholds also prevent its clinical implementation in NSCLC.¹⁶

DNA methylation of cytosines in the context of CG dinucleotides (CpG sites) has been demonstrated to modulate cancer development and progression via epigenetic regulation of gene expression and changes in chromatin instability.^{17 18} It was reported that genome-wide DNA methylation loss correlated with immune evasion signatures independently of TMB in several solid tumors, including lung cancer.¹⁹ Specifically in NSCLC, it was described that tumors with high TMB presented a higher load of differentially methylated CpG sites with respect to normal tissue than low TMB tumors, and a higher rate of copy number variations.²⁰ In addition, transposable elements (TEs) (ie, human endogenous retroviruses (HERV and long interspersed nuclear elements-1 (L1)) are primarily repressed by DNA methylation and other epigenetic mechanisms,^21–23 and their aberrant reactivation has been observed in various cancers.^24–26 Reactivated TEs are transcribed into RNA that can form double-stranded structures (dsRNA) whose cytosolic sensing has been proposed to induce innate immune responses with interferon (IFN) and cytokine production in a process termed viral mimicry phenotype (VMP).²² Also, the presence of TE-derived neoantigens has been linked to the induction of inflammatory responses^{26 27} and correlated with enhanced ICI efficacy in several solid tumors^28–32 including NSCLC.³³

Therefore, we hypothesized that the tumor epigenetic status would affect the ICI response, representing a source of candidate predictive response biomarkers. Since antitumor response is a multifactorial process, this study aimed to integratively analyze transcriptomic, DNA methylation, and mutational profiles from a cohort of patients with NSCLC treated with ICI. Through this approach, we seek to deepen the understanding of the molecular mechanisms underlying response to immunotherapy and to identify novel biomarkers that enable improved patient stratification and, ultimately, to optimize therapeutic decision-making in precision oncology.

Methods

Patients and samples

Our cohort included 32 patients with advanced NSCLC treated with immune programmed cell death protein 1 (PD-1)/PD-L1 checkpoint blockade at Hospital Álvaro Cunqueiro (Complexo Hospitalario Universitario de Vigo, CHUVI, Spain). We obtained the somatic mutation (whole genome sequencing (WGS) and WES, n=27), epigenetic (DNA methylation, Illumina Infinium MethylationEPIC arrays, n=32) and transcriptomic (RNA sequencing, n=30) profiles from diagnostic formalin-fixed paraffin-embedded (FFPE) primary tumor biopsies at baseline, using the same paraffin block. Paired peripheral blood mononuclear cells (PBMCs) were collected and their WES data were used as germline controls for mutation profiles (figure 1A). Immunotherapy response was evaluated by CT scan 6 months after the first ICI dose. According to Response Evaluation Criteria in Solid Tumors (RECIST) V.1.1 criteria, patients with stable disease (SD), partial response (PR), and complete response (CR)were classified as responders, while non-evaluable patients (NE exitus) or patients with progressive disease (PD) at the time of evaluation were considered non-responders. Overall survival (OS) was defined as the time from diagnosis to the time of death, while progression-free survival (PFS) was defined as the time from the first ICI dose to the time of progression or death, whichever event occurred earlier.

Figure 1. Clinical characteristics of the NSCLC cohort and their association with disease control 6 months after the first ICI dose. (A) Schematic overview of the cohort. Full transcriptome (RNA-seq) and DNA methylation (EPIC arrays) profiles were obtained from treatment-naive FFPE tumor biopsies of 32 patients with NSCLC treated with ICI. Also, the tumor mutational profile was derived by WGS/WES from the 27 patients with an available PBMC sample to be used as germline control. Created in https://BioRender.com. (B) Kaplan-Meier analysis for PFS with respect to disease control (responders (n=13) versus non-responders (n=19)). (C) Kaplan-Meier analysis for OS with respect to disease control. (D) Forest plot for the HR of disease control, sex, age, smoking status and PD-L1 IHC staining with respect to PFS, obtained by multivariate Cox regression. The error bars indicate 95% CI for the HR. (E) Forest plot for the HR of disease control, sex, age, smoking status and PD-L1 IHC staining with respect to OS, obtained by multivariate Cox regression. The error bars indicate 95% CI for the HR. FFPE, formalin-fixed paraffin-embedded; ICI, immune checkpoint inhibitor; IHC, immunohistochemistry; NSCLC, non-small cell lung cancer; OS, overall survival; PBMC, peripheral blood mononuclear cells; PD-1, programmed cell death protein 1; PD-L1, programmed cell death-ligand 1; PFS, progression-free survival; WES, whole exome sequencing; WGS, whole genome sequencing. *p value<0.05; ** p value<0.01; ***p value<0.005.

DNA and RNA extraction

DNA and RNA were extracted from the same FFPE block using the AllPrep FFPE DNA/RNA Mini Kit (Qiagen). Genomic DNA from PBMC was extracted with the QIAamp DNA Blood Mini Kit (Qiagen). DNA and RNA quantification was performed by Qubit fluorimetric assays (Thermo Fisher) while the quality was assessed using a TapeStation (Agilent).

Transcriptome profiling

For transcriptome profiling, RNA-seq libraries of the 30 FFPE tumor samples (12 responder and 18 non-responder patients) were constructed with the SureSelect RNA Direct protocol (Agilent) and paired-end data (100 bp × 2) were obtained within a NovaSeq 6000 system (Illumina). Quality control of raw reads was performed with fastQC.³⁴ Paired-end reads were aligned to the GRCh38/hg38 genome assembly (Genome Reference Consortium Human Build 38 patch release 14 (GRCh38.p14) version) with STAR³⁵ and transcript quantification (GENCODE.v47 GTF annotation) was performed with HTseq³⁶ with the intersection-non-empty mode. The locus-level specific activation of HERV and L1 elements was quantified with Telescope in average mode.³⁷ Transcriptional units for HERV and L1 were defined by combining RepeatMasker annotations from the same TE subfamily that are located in adjacent or nearby genomic regions, resulting in the locus-specific quantification of 14,968 HERV and 13, 545 L1 elements. Differentially expressed genes (DEGs) and TEs (DETEs) between responder and non-responder patients were called using DESeq2 Bioconductor package.³⁸ An adjusted p value <0.05 and an absolute log2-foldchange >1.5 were the cut-offs used to select DEGs. A p value <0.005 and an absolute log2-foldchange >1.5 were set for DETEs.

DNA methylation profiling

Tumor DNA of the 32 FFPE tumor samples (13 responder and 19 non-responder patients) was bisulfite-treated with the EZ-96 DNA Methylation Kit (Zymo Research), following the manufacturer’s recommendations for the EPIC array (Illumina). Subsequently, the bisulfite-modified DNA was treated with the InfiniumHD FFPE Restore Kit (Illumina) according to the manufacturer’s instructions and DNA methylation levels were quantified with the Infinium MethylationEPIC array (Illumina) following the manufacturer’s instructions. Samples were processed using EPIC array versions V.1.0 (866,836 probes; n=16) and V.2.0 (936,990 probes; n=16). For downstream analyses, only probes shared between both array versions were retained. Raw IDAT files were imported into R (V.4.4.1) and the minfi³⁹ Bioconductor package was used to extract the raw signal intensities and to calculate the β value for each probe as M/(M+U), where M and U refer to the mean raw methylated (M) and unmethylated (U) probe signal intensities, respectively. Comprehensive filtering was conducted using the wateRmelon⁴⁰ Bioconductor package, with the consequent removal of probes with detection of p value >0.01, probes with less than three beads in at least 5% of samples, non-CpG probes, cross-reactive probes, and probes targeting described single-nucleotide polymorphisms. Single-sample out-of-band normalization was applied to correct dye-bias and combat correction was used to account for slide batch effects. A total of 671,922 CpG loci were left after data curation. CpG probes were annotated to the GRCh38/hg38 reference genome with the IlluminaHumanMethylationEPICv2anno.20a1.hg38 Bioconductor package⁴¹ based on the Illumina EPIC V.2.0 manifest. Differentially methylated positions (DMPs) were identified with the limma⁴² Bioconductor package between responder and non-responder patients. A CpG was called differentially methylated if it showed a differential p value <0.001 and an absolute difference in the methylation levels of at least 10% between the two groups (absolute Δbeta-value >0.1). Differentially methylated regions (DMRs) of at least 350 bp were identified with the bumhunter method,⁴³ with the permutation approach to compute the p value and the family-wise error rate (FWER). Regions with at least two CpG sites, a p value <0.01, and a FWER ≤1 were defined as statistically significant DMRs. TE-body DNA methylation levels were obtained by averaging the β-values of the CpG sites within each of the 1,689 HERVs and 2,251 L1 loci extracted from RepeatMasker for TE quantification that were targeted by EPICv2 probes, then limma was applied to the locus-specific TE DNA methylation and a p value cut-off of 0.01 was used to call differentially methylated TEs (DMTEs).

Mutation profiling

We obtained the somatic mutational profile of the 27 patients from the cohort with an available PBMC sample to be used as a germline control. For 15 of the 27 patients, WGS data were obtained with the Truseq Nano DNA Library (Illumina) as detailed in our study by Brea-Iglesias et al,⁴⁴ while WES data were derived from the remaining 12 patients with the SureSelectV6 library protocol (Agilent). For all samples, paired-end sequencing (150 bp × 2) was performed on a NovaSeq6000 system (Illumina). Quality control of raw reads was performed with fastQC³⁴ and adapters were removed with the trimmomatic tool.⁴⁵ Sequencing reads were aligned to the GRCh38.p14 genome assembly with bwa-mem.⁴⁶

Prior to variant calling in both WGS and WES data, cross-sample contamination was estimated using GATK’s GetPileupSummaries and CalculateContamination together with the gnomAD resource from the GATK best-practices bundle.⁴⁷ Somatic single nucleotide variants and small insertions/deletions (INDELs) were detected with Mutect2 v0.7.17 (GATK)⁴⁸ in tumor-PBMC matched mode to account for germline variants. FilterMutectCalls was applied with default parameters, integrating contamination estimates and read-orientation modeling to identify FFPE-related artifacts. Variants flagged as potential artifacts in at least two patients were discarded from subsequent analyses. To further reduce low-quality calls and recurrent artifacts we also removed variants supported by <3 reads, with total depth <10, showing an allele frequency <10%, reads and the variants located in problematic genomic regions annotated by the UCSC Genome Browser: regions with common structural variants (DbVarConflict track),⁴⁹ repetitive regions with high mapping variability (ENCODE blacklist subtrack),⁵⁰ fragile chromosomal regions (HumCFS track),⁵¹ and regions systematically prone to sequencing errors and low coverage (giabCallConflict track).⁵² The remaining variants were annotated with the Ensembl Variant Effect Predictor (VEP) tool V.100.2.⁵³ Likely pathogenic variants were defined as those with HIGH or MODERATE predicted impact in VEP or classified as deleterious by SIFT⁵⁴ and PolyPhen⁵⁵ scores. Additionally, for WGS data, non-coding variants were removed, excluding intronic, intergenic, and regulatory region variants.

We determined the TMB of each patient as the 10-base logarithm of the number of missense mutations present in coding regions divided by 10⁶.

Enrichment analyses

Functional enrichment over-representation analyses (ORA) of Gene Ontology (GO) terms and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways was performed for differentially methylated CpGs with the missMethyl⁵⁶ Bioconductor package, using the differentially methylated CpGs located in gene promoters and CpG islands with differential p value <0.005 and an absolute Δbeta-value >0.1 as input. ORA of GO terms and KEGG pathways was also performed on significantly upregulated/downregulated genes (p value <0.05 and absolute log2-foldchange >1.5) between responder and non-responder patients with the clusterProfiler⁵⁷ Bioconductor package. Gene set enrichment analyses (GSEAs) was performed on the raw log2-foldchange ratios derived from differential expression analyses with clusterProfiler. Also, enrichment scores of selected gene signatures were derived for each patient with the publicly available implementation of single-sample GSEA (ssGSEA)⁵⁸ on Rpubs by Pranali S (https://rpubs.com/pranali018/SSGSEA).

Evaluation of epigenetic silencing

To determine the genes that were repressed due to DNA methylation, we first defined cancer-specific promoter hypermethylation. For that, we averaged EPIC data from 137 normal lung samples downloaded from The Cancer Genome Atlas data portal to discard probes that showed a median DNA methylation higher than 20% in normal lung. Then, we selected the promoter-targeting CpGs as those probes located within 1,500 bp upstream the transcription starting site of each gene, according to the Illumina EPIC V.2.0 manifest annotation. For promoters targeted by multiple probes, the average β-value was used as the measure of promoter-level DNA methylation. For each promoter, samples were stratified as either methylated or unmethylated based on a DNA methylation cut-off of 30%. Epigenetically silenced genes were defined as those whose overall normalized expression (fragments per kilobase of transcript per million mapped reads, FPKM) was less in methylated than in unmethylated samples (one-sided Wilcoxon rank-sum test p value <0.05).

Tumor microenvironment composition and systemic immune profiling

The tumor microenvironment (TME) composition of each NSCLC tumor sample was assessed with the CIBERSORT deconvolution method, based on both gene expression (CIBERSORTx) with the LM22 signature matrix⁵⁹ and DNA methylation (methylCIBERSORT)⁶⁰ in absolute mode with 100 permutations.

To obtain an independent orthogonal validation of the immune-related findings derived from the tumor multiomic analyses, baseline PBMC immune profiling was performed in an independent cohort of 56 patients with advanced NSCLC treated with ICI (25 responders and 31 non-responders). None of these patients overlapped with the discovery cohort because viable PBMC samples were not available for the patients included in the tumor multiomic analyses. PBMC were stained with Fixable Viability Dye eFluor e450 (Thermo Fisher) and Fc‐γ receptor blocking was performed using 10% of Anti-Hu Fc Receptor (Thermo Fisher). Two complementary antibody panels were employed. The first panel characterized the major circulating immune populations using the following antibodies: CD3 AF700, CD4 APC-eFluor780, CD8 V500, CD14 APC, CD16 FITC, CD19 PECy7, CD56 PE-CF54, CD1c PE and HLA-DR BV605 (all from BioLegend), enabling the identification of T cells, B cells, natural killer (NK) cells, monocytes and dendritic cells. The second panel characterized T cell differentiation and activation status using antibodies against CD3 AF700, CD4 APC-Cy7, CD8 PerCP-Cy5.5, CD27 BV510, CD45RO APC, CD25 FITC, CD127 BV605, PD-1 PE-Cy7 and CD56 PE-CF54 (all from BioLegend). Data was obtained in a CytoFlex S Analyser (Beckman Coulter). After excluding debris, doublets and dead cells, cell populations were analyzed with CytExpert Software (Beckman Coulter). Results were expressed as percentage of positive cells and differences in immune-cell frequencies between responder and non-responder patients were evaluated using two-sided Wilcoxon rank-sum tests.

Evaluation of viral mimicry

We first calculated an “antiviral response score” for each patient as the arithmetic mean of the normalized expression (FPKM) of the 57 genes included in the GO biological process term “response_to_dsRNA” (GO:0043331),

$${\displaystyle {\displaystyle {\displaystyle \mathrm{A}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{r}\mathrm{a}\mathrm{l}\mathrm{R}\mathrm{e}\mathrm{s}\mathrm{p}\mathrm{o}\mathrm{n}\mathrm{s}\mathrm{e}\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}{\mathrm{e}}_{\mathrm{i}}=1/57\sum _{\mathrm{k}=1}^{57}\mathrm{F}\mathrm{P}\mathrm{K}{\mathrm{M}}_{\mathrm{i}\mathrm{k}}}}}$$

where i denotes an individual patient and k each of the 57 genes included in the gene set. Patients were subsequently classified into High or Low antiviral response according to the mean antiviral score of the cohort.

We evaluated the correlation between the antiviral response score and IFN-I, IFN-II and IFN-III FPKM levels, the expression of the JAK-STAT signaling pathway (average FPKM of genes included in KEGG hsa04630), and the expression of IFN-stimulated genes (ISGs) (average FPKM of genes included in REACTOME pathway R-HSA-1169410).

We then defined the Total TE Load for each patient as the sum of the normalized FPKM values of all L1 and HERV elements quantified by Telescope³⁷:

$${\displaystyle {\displaystyle {\displaystyle \mathrm{T}\mathrm{E}\mathrm{l}\mathrm{o}\mathrm{a}{\mathrm{d}}_{\mathrm{i}}=\sum _{\mathrm{j}=1}^{\mathrm{n}}\mathrm{F}\mathrm{P}\mathrm{K}{\mathrm{M}}_{\mathrm{i}\mathrm{j}}}}}$$

where j represents all quantified L1 and HERV loci (including HERVK, HERVL, HERV1, L1 ORF2, L1 full-length intact and L1 full-length truncated elements). Patients were classified into High or Low TE load according to the median TE load of the cohort. L1 load and HERV load were calculated analogously.

For correlation analyses, TE Load values and ssGSEA enrichment scores were standardized as z-scores before Pearson correlation analysis. Importantly, this normalization was performed exclusively for correlation analyses and was not used for patient stratification.

Finally, patients were stratified into four groups according to their TE load and antiviral response status: no-VMP (Low-TE/Low-antiviral), High-TE only (High-TE/Low-antiviral), High-antiviral only (Low-TE/High-antiviral), and VMP (High-TE/High-antiviral). Immunotherapy response, OS and PFS were subsequently evaluated across these four groups.

External validation

The identified molecular signatures associated with ICI outcome were validated in clinically relevant public datasets of patients with NSCLC treated with ICI. The expression levels of DEGs and the GSEA enrichment score of selected pathways were validated in the GSE161537 dataset,⁶¹ which includes the transcriptomic profile of 82 patients with NSCLC treated with ICI derived with the HTG EdgeSeq Oncology Biomarker Panel (expression levels of 2559 genes).

Statistical analyses

All statistical analyses were conducted within the R environment (V.4.4.1). We performed survival analyses for OS and PFS with censoring at the date of the last follow-up if a patient was alive and without disease progression. Multivariate Cox proportional hazards ratio models were fit to the clinical variables disease control, sex, age, smoking status and PD-L1 staining for both OS and PFS as response variables, using the coxph function from the survival R package. Survival curves were estimated using the Kaplan-Meier method and were compared with the log-rank test. Wilcoxon rank-sum test was used for two-group comparisons of continuous variables. All correlation analyses were carried out with the Spearman method. Euclidean distances and the Ward clustering method⁶² were used to compute the dendrograms within the heatmaps. Two-sided Fisher’s exact tests were used to evaluate the relation between categorical variables. False discovery rate (FDR) was inspected with the Benjamini-Hochberg correction. When FDR 5% cut-off of differential analyses did not yield significant results, the distribution of raw p values was evaluated using quantile-quantile (Q-Q) plots comparing the observed and expected p value distributions under the null hypothesis, together with p value histograms. The number of nominally significant features was also compared with the number expected under the global null hypothesis at predefined significance thresholds as follows: E[N] = total features × nominal p value, to assess potential inflation or deflation of the test statistics. For DNA methylation and TE expression analyses, the distribution of nominal p values was additionally inspected using histograms and Q-Q plots to evaluate the calibration of the statistical tests and to assess potential p value inflation (online supplemental figure S1).

To evaluate the robustness of the molecular associations, additional sensitivity analyses were performed by repeating the differential DNA methylation, differential gene expression and TE analyses while adjusting separately for histological subtype (adenocarcinoma vs non-adenocarcinoma), treatment line (first-line vs later-line) and treatment regimen (ICI monotherapy vs ICI plus chemotherapy). Concordance between the original and adjusted models was assessed by calculating Spearman correlations of effect sizes and the proportion of features preserving the direction of effect.

Fisher’s exact tests were also used to evaluate the enrichment of DMPs across chromosomes (1-22, X, and Y) and genomic regions (CpG islands, promoters, and opensea). We conducted hierarchical testing: first at the chromosome level for hypomethylated and hypermethylated DMPs, then within specific genomic regions (CpG islands, promoters, and opensea) for each chromosome. Enrichment p values were adjusted for multiple testing using the Benjamini-Hochberg correction (adjusted p value <0.05).

To build a multimarker machine learning classifier for long-term response prediction, a random forest model was fit to a set of variables considered potential biomarkers: (1) the log10-transformed FPKM values of the DEGs together with PD-L1 (CD274), (2) the DNA methylation values of the CpG sites associated with PFS, (3) the DNA methylation value of DMRs, (4) the z-score of the single-sample enrichment scores of pathways differentially expressed between responders and non-responder patients, (5) TME composition scores derived by gene expression of M1 macrophages, activated mast cells, and resting NK cells, together with the scores for CD8+ T cells derived by both gene expression and DNA methylation data, (6) the z-score value of the TMB, and (7) clinical variables such as age, sex, smoking status, the circulating neutrophil-to-lymphocyte ratio (NLR), and the blood levels of lactate dehydrogenase (LDH, U/mL). To assess the performance and generalizability of the random forest classifier, a nested leave-one-out cross-validation (nested LOOCV) framework was implemented. Within each outer LOOCV iteration, feature selection was performed exclusively on the training set using the predefined filtering criteria, thereby preventing information leakage from the test sample. A random forest model was subsequently trained on the selected features and evaluated on the held-out sample. This procedure was repeated for every sample in the cohort to generate the receiver operating characteristic curve and the corresponding area under the curve (AUC). The optimal classification threshold was determined using the Youden Index. Feature importance was calculated independently in each training iteration, and variables were ranked according to their mean importance across all nested LOOCV iterations. The 30 highest-ranked features were retained to define the final predictive signature.

Results

Cohort description

Treatment-naive FFPE primary tumor biopsies were obtained from 32 patients with advanced NSCLC (median age of 64.5 years) that were treated with immune checkpoint blockade at first (n=27), second (n=4) or third line (n=1) (table 1). Most patients received single-agent ICI therapy (anti-PD1 pembrolizumab, n=24 or anti-PD-L1 atezolizumab, n=2), while 6 of them were treated with ICI in combination with chemotherapy. Responder patients at 6 months (n=13) showed longer PFS and OS (log-rank p value <0.0001 and p value=0.00011, respectively, figure 1B,C) than non-responder patients (n=19). Multivariate HR analyses were conducted to evaluate the association between clinical variables and improved survival. In agreement with the previous result, response to ICI showed a statistically significant impact on PFS and OS (p value <0.001). Also, male sex showed an unfavorable association with both PFS and OS (p value=0.043 and p value=0.003, respectively). With respect to smoking, a former smoking history had a negative impact on PFS (p value=0.031), while OS was not affected by the smoking history of the patient. Interestingly, analysis of age (stratified based on the cohort median of 64.5 years old) and PD-L1 immunohistochemical staining (stratified as tumor proportion score high vs low at the cohort median of 65%) revealed no statistically significant association with PFS or OS in this cohort (figure 1D,E).

Table 1. Clinical characteristics of the NSCLC cohort.

	Disease control at 6 months
	All (n=32)	R (n=13)	NR (n=19)
Sex
Female	16	7	9
Male	16	6	10
Age (median) (range)	64.5 (43–83)	66 (46–77)	64 (43–83)
Smoking status
Person who smokes	16	6	10
Person who previously smoked	13	7	6
Person who has never smoked	3	0	3
Tumor type
ADC	23	9	14
SCC	7	4	3
Undifferentiated	1	0	1
Other	1	0	1
Tumor stage at diagnosis
IIB	1	1	0
IIIA	2	1	1
IIIB	2	1	1
IIIC	3	1	2
IVA	11	6	5
IVB	11	2	9
IVC	2	1	1

Tumor staging at diagnosis was based on the eight edition of the TNM Classification for Lung Cancer.

ADC, adenocarcinoma; NR, non-responder; NSCLC, non-small cell lung cancer; R, responder; SCC, squamous cell carcinoma; TNM, tumor, node, metastasis.

Differential expression and methylation profiles associated with ICI response

We first performed a differential expression analysis between responder and non-responder patients. We identified 70 DEGs (figure 2A, online supplemental table S1), most of them (69) downregulated in responder patients, while only 1 gene was significantly upregulated. From the differential methylation analysis, we identified 156 DMPs, from which 122 were hypermethylated versus 34 hypomethylated in responder patients (figure 2B). 50% of the hypermethylated CpGs were present in promoter regions while most hypomethylated CpGs were in intergenic or opensea regions (76%) (online supplemental table S2). Enrichment analyses were performed to assess chromosome-specific and region-specific differential methylation patterns. A chromosome-based distribution analysis revealed a significant enrichment of hypomethylated DMPs exclusively on chromosome X, both at the chromosome-wide level and specifically within CpG islands and opensea regions (Fisher’s exact test adjusted p value <0.05), while no significant enrichment was reported for autosomes or hypermethylated DMPs (online supplemental table S3).

Figure 2. Baseline epigenomic and transcriptomic profiles associated with disease control after ICI immunotherapy in NSCLC. (A) Volcano plot of differential expression analysis, the 70 differentially expressed genes (BH adjusted p value <0.05 red dashed line, and absolute log2-fold change > 1.5 black dashed lines) are highlighted in blue (down in responders) or red (up in responders). (B) Heatmap showing the DNA methylation levels of the 156 differentially methylated positions. Euclidean distances and the Ward clustering method were used to compute the dendrogram. (C) Heatmap showing the DNA methylation levels of the 18 DMRs located in gene promoters. Euclidean distances and the Ward clustering method were used to compute the dendrogram. (D) NES of the GSEA of KEGG pathways and GO biological process terms based on the log2-fold changes of differential expression. A positive NES implies an upregulation in responder patients, while a negative value means an upregulation in non-responders. (E) Over-representation analysis of KEGG pathways and GO biological process terms based on the differentially methylated CpG sites located in gene promoters that were hypermethylated in responders. (F) Over-representation analysis of KEGG pathways and GO biological process terms based on the differentially methylated CpG sites located in gene promoters that were hypomethylated in responders. ADC, adenocarcinoma; BH, Benjamini-Hochberg; CR, complete response; DMR, differentially methylated region; ECM, extracellular matrix; GO, Gene Ontology; GSEA, gene set enrichment analysis; ICI, immune checkpoint inhibitor; IL, interleukin; KEGG, Kyoto Encyclopedia of Genes and Genomes; MHC, major histocompatibility complex; NE, non-evaluable patients; NSCLC, non-small cell lung cancer; PD, progressive disease; PD-1, programmed cell death protein 1; PD-L1, programmed cell death-ligand 1; PR, partial response; SCC, squamous cell carcinoma; SD, stable disease; TCR, T cell receptor; Th1, T helper cell 1; TNF, tumor necrosis factor; NES: normalized enrichment score.

We identified 38 DMRs between responder and non-responder patients. Of these, 18 DMRs overlapped with gene promoters (promoter-DMRs) (figure 2C), 9 were located in gene bodies and 12 in intergenic regions (online supplemental table S4). Among the 18 promoter-DMRs, 15 exhibited hypermethylation in responder patients, while 3 showed hypomethylation. To determine whether promoter methylation status influenced gene expression, we performed correlation analysis between the average methylation level of each promoter-DMR and the corresponding FPKM expression (online supplemental table S4). Four genes showed a significant negative correlation between promoter methylation and their expression: H2AW, GORASP2, TNFSF11, and MIR193B (Spearman rho <−0.3 and p value <0.05, online supplemental figure S2A). To validate these correlations, we compared their expression levels between responder and non-responder patients. Consistent with the presence of hypermethylated promoter-DMRs, both GORASP2 and TNFSF11 showed significantly lower expression in responder patients (p value <0.05, online supplemental figure S2B). An overall lower expression in responders was also observed for MIR193B, although not statistically significant (p value=0.54, online supplemental figure S2B). Conversely, H2AW, which harbored a hypomethylated promoter-DMR in responders, exhibited higher expression in this group, although this difference did not reach statistical significance (p value=0.14, online supplemental figure S2B). To evaluate the robustness of these findings, differential DNA methylation, gene expression and TE analyses were repeated after adjustment for histological subtype, treatment line and treatment regimen. The adjusted analyses showed a high degree of concordance with the original models across all molecular layers, supporting the robustness of the identified molecular signatures (online supplemental table S5).

Combined functional enrichment

We performed enrichment analyses based on differential expression and methylation profiles. On one hand, GSEA and ORA enrichment of gene expression reported a downregulation of proliferative and antiapoptotic pathways such as Ras, NF-kB, tumor necrosis factor (TNF) or Wnt signaling pathways in responder patients (figure 2D, online supplemental figure S2C), which were also associated with hypermethylated promoters (figure 2E). Moreover, CpG sites hypermethylated in responder patients were enriched in the PI3K-Akt and mTOR signaling pathways, as well as in the ErbB signaling pathway, an upstream regulator of PI3K-Akt (figure 2E). Moreover, their downregulation in ICI response was also confirmed by GSEA of gene expression (figure 2D, online supplemental figure S2C). On the other hand, we found an upregulation of the Notch signaling pathway and of the Polycomb repressive complex in responder patients (figure 2D), being also concordantly enriched in hypomethylated promoter GpG sites (figure 2F). A set of terms related to the physical characteristics of the TME, such as the extracellular matrix (ECM)-receptor interaction, tight junctions, and focal adhesions showed a downregulation in responders and were also associated with promoter hypermethylation (figure 2D,E, online supplemental figure S2D).

We also identified an enrichment in several immune-related mechanisms. Terms associated with an immunosuppressive TME, such as interleukin (IL)-17A mediated signaling pathway and the PD-1/PD-L1 checkpoint pathway appeared to be downregulated in responder patients (figure 2D), being the latter also enriched in hypermethylated promoters (figure 2E). Also, responder patients exhibited upregulation of multiple processes associated with T cell activity, including antigen processing and presentation, T cell activation via T cell receptor (TCR)-major histocompatibility complex (MHC) interaction, T cell-mediated antitumor response, and immune-related leukocyte migration. Specifically, the upregulation of CD8+ αβ-T cell differentiation (figure 2D), and its association with promoter hypomethylation (figure 2F), may indicate an enhanced CD8+ T cell cytotoxic activity in responders. In contrast, CD4+ T cells displayed a complex regulation in responders: while the antigen processing and presentation of peptide antigens by MHC-II was upregulated together with a suppressed negative regulation of CD4+ αβ-T cell differentiation (figure 2D,F), CD4+ αβ-T cell costimulation and its cytokine production were associated with promoter hypermethylation (figure 2E), indicating an activation of CD4+ T cell activity and suggesting a potential inhibition of these CD4+ T cell functions. B cell-related processes also showed mixed patterns in responders, as B cell activation and its regulation showed downregulation despite upregulation of B cell receptor signaling (figure 2D,E). Additionally, GSEA indicated reduced NK cell cytotoxicity and decreased positive regulation of IL-16 production, Th1 cytokine production, and response to IL-2 in responders (figure 2D). Notably, negative regulation of IL-2 production was downregulated and enriched in hypermethylated promoters (figure 2D,E), while IL-16-mediated signaling was upregulated and associated with promoter hypermethylation (figure 2D,E).

Validation in an independent ICI-treated NSCLC cohort

To assess the generalizability of our findings, we performed a validation analysis using the GSE161537 dataset, an independent cohort of 82 patients with advanced NSCLC (median age of 67 years; 36% female, 64% male) treated with anti-PD1 ICI at first (n=2), second (n=47) or third line (n=33).⁶³ This external validation successfully confirmed key signatures identified in our discovery cohort. Specifically, responder patients showed consistent downregulation of the PI3K-Akt, mTOR, Ras, and TNF signaling pathways, as well as reduced expression of ECM-receptor interaction, tight junction, and focal adhesion pathways. Furthermore, the upregulation of ErbB signaling, T cell-mediated immune response to tumor cells, and antigen processing and presentation signatures observed in our cohort were also replicated in this independent dataset (online supplemental figure S2E). This cross-cohort validation demonstrates the robustness of these molecular signatures and supports their potential clinical relevance of these molecular signatures as predictors of ICI response in NSCLC.

Tumor-specific epigenetic silencing of gene expression

We performed the identification of epigenetically silenced genes by integrating promoter DNA methylation and expression profiles of responder and non-responder patients separately. For each gene, we stratified samples as methylated/unmethylated based on the DNA methylation levels of its promoter. Then, we evaluated if the overall gene expression was lower in the methylated samples. Thus, we detected 87 genes that were subjected to epigenetic silencing by DNA methylation exclusively in responder patients (online supplemental table S6, p value <0.05) and 112 genes that were epigenetically silenced by DNA methylation in non-responders, while 12 genes were silenced by DNA methylation in both groups (online supplemental table S6). It is noteworthy that between the two sets of genes we found targets of the Polycomb repressive complex 2 (PRC2): 15 and 13 PRC2 targets were subjected to epigenetic silencing in responders and non-responders, respectively. Also, eight and five genes implicated in chromatin remodeling but independent of Polycomb activity were silenced by DNA methylation in responders and non-responders, respectively. This might suggest a role of the PRC2 activity and other levels of epigenetic regulation in the response to immune checkpoint blockade. Interestingly, three genes that regulate the PI3K-Akt signaling pathway (ITGA4, FGF2, and IGF2) were also silenced by DNA methylation exclusively in non-responders, being ITGA4 also a target of the Polycomb repressive complex and IGF2 regulated by chromatin remodeling independently of Polycomb. The genes SLIT2, EDN3, and BST1 were also silenced by DNA methylation exclusively in non-responders and, together with ITGA4, are implicated in leukocyte migration.

Tumor microenvironment composition and systemic immune profiling

The estimation of cell type abundances of the TME was assessed with the CIBERSORT framework, from both bulk gene expression (CIBERSORTx) and DNA methylation (methylCIBERSORT) data. TME composition based on gene expression showed that responder patients presented a higher proportion of activated mast cells and lower proportions of resting NK cells than non-responder patients (p value <0.05, online supplemental figure S3A). Although no other TME-infiltrating immune cell populations identified by in silico deconvolution of gene expression and DNA methylation profiles showed a statistically significant association with ICI response, several consistent trends were noted. On one hand, both DNA methylation and gene expression-based deconvolution showed a higher presence of CD8+ T cells in responders (online supplemental figure S3A,B), and the proportion of CD8+ T cells determined by both DNA methylation and gene expression data deconvolution was positively correlated (online supplemental figure S3C). On the other hand, responder patients presented higher fractions of macrophages, especially M1 macrophages (online supplemental figure S3A) concordant with a higher proportion of CD14+ cells reported by DNA methylation deconvolution (online supplemental figure S3B), and a lower proportion of M2 macrophages, activated myeloid dendritic cells and resting mast cells (online supplemental figure S3A) as determined by transcriptomic deconvolution, and a lower proportion of cancer-associated fibroblasts (online supplemental figure S3B) as determined by methylation-based deconvolution. We did not report any association between the TMB and the TME composition (online supplemental figure S3D,E).

To obtain an orthogonal assessment of baseline immune status, baseline PBMC immune profiling was performed in an independent cohort of 56 patients with advanced NSCLC (25 responders and 31 non-responders; the clinical characteristics of the validation cohort are summarized in online supplemental table S7). This cohort did not overlap with the tumor discovery cohort, providing an independent validation of the immune-related observations. The overall frequencies of circulating B and T cells were comparable between responder and non-responder patients. However, responders tended to display increased proportions of dendritic cells and CD4+ T cells, whereas non-responders showed higher frequencies of NK cells and monocytes, particularly classical monocytes (online supplemental figure S3F).

Analysis of T cell subsets revealed distinct immune profiles associated with clinical response. Responders exhibited a significantly lower proportion of circulating CD8+ T cells (p value<0.05), together with higher CD4+ T cell frequencies, resulting in an increased CD4/CD8 ratio (online supplemental figure S3F,G). Within the CD4 compartment, responders showed a significantly higher proportion of naïve CD4+ T cells (p value<0.05), whereas effector and effector-memory CD4+ subsets were more abundant in non-responders (online supplemental figure S3G). Likewise, responders displayed higher frequencies of naïve and central-memory CD8+ T cells, whereas non-responders were enriched in effector and effector-memory CD8+ subsets (online supplemental figure S3G).

Evaluation of PD-1 expression further supported these differences in baseline immune status. Non-responders exhibited significantly higher frequencies of circulating PD-1+CD4+ T cells (p value<0.05), whereas PD-1 expression on CD8+ T cells was comparable between groups. Consequently, the PD-1+CD4+/CD8+ ratio was significantly increased in non-responders (p value<0.01) (online supplemental figure S3H).

Transposable elements activity suggests the presence of viral mimicry and impacts ICI outcome

We evaluated possible differences in TE activation in ICI response. At the family level, response was not significantly statistically associated neither to L1 load nor HERV load (online supplemental figure S4A). Then, we explored TE activity and DNA methylation at locus level, finding 56 out of 28,516 elements differentially expressed between responder and non-responder patients (p value <0.005 and absolute log2-foldchange >1.5) (figure 3A, online supplemental table S8). Of those, 49 were downexpressed and 7 overexpressed in responders. We also identified 33 TE showing differential methylation levels (p value <0.01), from which 27 were hypermethylated in responders while the remaining 6 were hypomethylated (figure 3B, online supplemental table S9). These differentially expressed and methylated TE did not show overlapping, and when we correlate the DNA methylation and expression levels of all the TE, only L1 ORF2 elements displayed a weak inverse correlation between its activation and methylation (p value <0.05) (online supplemental figure S4B).

Figure 3. VMP evaluation. (A) Volcano plot of differential TE expression analysis, the 56 differentially active TEs (p value <0.005 red dashed lines and absolute log2-foldchange >1.5 black dashed lines) are highlighted in blue (down in responders) or red (up in responders). (B) Heatmap showing the body DNA methylation levels of the 33 differentially methylated TEs. Euclidean distances and the Ward clustering method were used to compute the dendrogram. (C) Boxplots of the TE activity (L1 ORF2, hERV1 family and total TE load) with respect to antiviral response (High/Low stratification was based on the mean antiviral score). The central mark represents the median, with 25th and 75th percentiles at the box, 5th and 95th percentiles at the whiskers. Differential p value was computed by the Wilcoxon rank sum test. (D) Scatterplot showing the Spearman coefficient and p-value of the correlation between the antiviral score and the expression of IFN-II, JAK/STAT pathway and ISGs. (E) Scatterplot showing the Spearman coefficient and p value of the correlation between the TE load (z-score) and the ssGSEA enrichment (z-scores) of the response to dsRNA signature, response to IFN-II and JAK/STAT signaling pathway. (F) Kaplan-Meier survival curves for OS of the VMP stratification: High antiviral only, High TE Only, Low-antiviral/Low-TE (no-VMP) and High-antiviral/High-TE load (VMP). Log-rank test was used to derive the p value. (G) Kaplan-Meier survival curves for OS with respect to TE activity stratification response (High/Low TE-activity was defined based on the mean TE load). Log-rank test was used to derive the p value. CR, complete response; dsRNA, double-stranded RNA; FPKM, fragments per kilobase of transcript per million mapped reads; HERV, human endogenous retrovirus; IFN, interferon; ICI, immune checkpoint inhibitor; ISGs, interferon-stimulated genes; L1, long interspersed nuclear elements-1; NE, non-evaluable patients; OS, overall survival; PD, progressive disease; PR, partial response; SD, stable disease; ssGSEA, single-sample gene set enrichment analysis; TE, transposable element; VMP, viral mimicry phenotype.

Then, we evaluated the presence of a possible VMP using the expression of the genes annotated to the GO term “response to dsRNA” as an antiviral score. We found that patients with increased levels of L1 ORF2 and HERV1 showed higher antiviral responses (p value <0.05) (figure 3C). This association was also shown for the total tumor TE load (figure 3C). The antiviral response score also correlated with higher IFN-II expression, the upregulation of the JAK/STAT signaling pathway and elevated expression of ISGs (figure 3D). To test whether TE levels correlated with the presence of VMP and whether such a state was associated with disease control and longer survival, we stratified patients into four groups according to their TE load (High/Low-TE load) and to their antiviral response (High/Low-antiviral score). Only co-occurring high-TE load and High-antiviral response were considered as VMP. In our cohort, the total load of TEs and precisely the total activity of L1 elements was positively correlated with the expression of the antiviral response signature (p value <0.001, rho=0.81 and 0.86 respectively), with the enrichment in the response to IFN-II signature (p value <0.05, rho=0.36 and 0.44, respectively) and with the enrichment in the JAK/STAT signaling pathway (p value <0.001, rho=0.72 and 0.78, respectively) (figure 3E). Total hERV levels did not report an association with the evaluated antiviral response.

Although patients were evenly distributed among the four defined TE/antiviral score groups and the response (Fisher’s exact test p value=0.815, online supplemental figure S4C), we found a statistical significance in OS across the four groups (log-rank test p value=0.039, figure 3F), with no-VMP patients and patients with High-antiviral response/Low-TE showing the lowest survival rates. The same tendency was reported for PFS although non-significant (online supplemental figure S4D). Finally, patients were stratified into High/Low based on TE load only. Patients with high TE activity reported statistically significant longer OS (log-rank test p value=0.016, figure 3G), while the TE load effect on longer PFS was not significant (log-rank test p value >0.05, online supplemental figure S4E).

Evaluation of the best baseline biomarkers for predicting ICI response

We first evaluated the association between ICI response and patient survival and the currently accepted biomarkers: TMB and PD-L1 expression. Interestingly, responder patients showed a lower mean TMB and the association between baseline TMB and response was not significant (figure 4A). Concordantly, these biomarkers did not correlate with longer survival (online supplemental figure S5).

Figure 4. Multimarker signature to predict ICI response at baseline. (A) Boxplots of the TMB (derived by WES and WGS data: non-synonymous variants located in coding regions), PD-L1 expression (z-score of CD274 gene FPKM normalized expression) with respect to response to ICI. The central mark represents the median, with 25th and 75th percentiles at the box, 5th and 95th percentiles at the whiskers. Differential p value was computed by the Wilcoxon rank-sum test. (B) Kaplan-Meier survival curves for PFS based on the average methylation of the 12-CpG sites associated with PFS patients were stratified as High-methylation (≥50%) or Low-methylation (<50%). (C) Average variable importance of candidate biomarkers determined by averaging, in each nested loop, the median decrease in accuracy of the RF classifier if the respective variable is eliminated from the model. (D) ROC curve analysis and AUC for the RF models built for long-term response prediction, derived by nested leave-one-out cross-validation. The cut-off to determine sensitivity and specificity values was based on the Youden index method. AUC, area under the curve; FPKM, fragments per kilobase of transcript per million mapped reads; ICI, immune checkpoint inhibitor; LDH, lactate dehydrogenase; PD-L1, programmed cell death-ligand 1; PFS, progression-free survival; RF, random forest; ROC, receiver operating characteristic; TMB, tumor mutational burden; WES, whole exome sequencing; WGS, whole genome sequencing.

With the aim to define a CpG signature that predicts longer survival, we applied a Cox regression with Least Absolute Shrinkage and Selection Operator (LASSO) penalty to the methylation levels of the 156 DMPs against PFS. The resulting signature consisted of 12 CpG sites with non-zero coefficients at the regularization parameter value that results in the lowest cross-validated mean squared error (online supplemental table S2). Patients with a high-methylated signature (12 CpG-average methylation ≥50%) showed significantly longer PFS (p value <0.00025) than patients with low methylation (figure 4B).

Moreover, aiming to prioritize potential candidate biomarkers for ICI response prediction at baseline, a random forest classifier was fit to the 70 DEGs, the 12 CpG sites associated with PFS, the 38 DMRs, the TMB value, the single-sample enrichment scores of differentially regulated pathways, the TME composition (cell-type scores derived by gene expression and DNA methylation) and the clinical variables age, sex, smoking status, the circulating NLR, and the LDH blood levels. To obtain a robust estimate of classifier performance while minimizing information leakage, a nested leave-one-out cross-validation framework was implemented. These candidate biomarkers were ranked according to their mean importance across the nested leave-one-out cross-validation iterations, quantified as the decrease in classification accuracy when each variable was omitted. The complete ranking of all evaluated variables is provided in online supplemental table S10, while the highest-ranked features are shown in figure 4C. The TFPI2 gene is the strongest predictor and its high expression is negatively associated with response, as well as high blood LDH. A second random forest was fit to the top-30 retained after the feature selection and the model performance was evaluated by leave-one-out cross-validation. The obtained multi-marker signature achieves a cross-validated AUC of 0.833 with a sensitivity of 81.8% and a specificity of 73.3% at the optimal Youden threshold (figure 4D).

Discussion

In this study, we performed a comprehensive profiling of baseline primary tumor biopsies of patients with advanced NSCLC treated with ICI, to identify epigenetic and transcriptomic determinants of ICI response in advanced NSCLC and to propose a clinically relevant biomarker signature for outcome prediction. We identified a total of 156 differentially methylated CpG positions and 38 DMRs between responder and non-responder patients, being most of them (122 of the CpG positions and 32 regions) hypomethylated in non-responder patients, in line with previous evidence in NSCLC ICI-treated patients.^{19 64 65} We conducted parallel whole transcriptomic profiling of the same samples that allowed the identification of a set of 70 DEGs. Also, the integration of the transcriptomic and DNA methylation profiles enabled the identification of 87 genes that were silenced by DNA methylation exclusively in responders, while 122 were exclusively silenced in non-responders. These findings indicate that responders and non-responders are characterized by distinct epigenetic programs of gene regulation, suggesting that DNA methylation contributes to shaping tumor states associated with differential sensitivity to immune checkpoint blockade.

Our enrichment analysis supported a functional link between the promoter CpG island methylation status and the expression levels of several biological processes. Interestingly, responder patients showed a global downregulation of the PI3K-Akt signaling pathway. This finding agrees with previous studies showing that PI3K signaling promoted an immunosuppressive TME by recruiting myeloid cells that inhibit CD8+ T cell proliferation in mice,⁶⁶ correlated with the inactivation of innate immunity and immunoescape in triple-negative breast cancer.⁶⁷ Moreover, its inhibition synergized with PD1 blockade improving effector T cell infiltration in preclinical studies with patients with sarcoma cell lines,⁶⁸ colon,⁶⁹ and bladder cancer.⁷⁰ Concordantly, we found that responder patients showed a lower proportion of intratumor activated myeloid dendritic cells while higher levels of CD8+ T cells and M1 macrophages. In addition, PIK3CA expression positively correlated with the expression of immunosuppressive cytokines. Notably, loss-of-function mutations in PTEN have been linked to sustained activation of PI3K signaling, promoting an immunosuppressive and poorly infiltrated TME, ultimately leading to resistance to ICI.⁷¹ Our findings are consistent with this framework, as responder patients displayed epigenetically mediated downregulation of PI3K-Akt signaling, suggesting that suppression of this pathway may favor immune activation and treatment efficacy. Importantly, our data indicate that this modulation occurs, at least in part, through epigenetic mechanisms rather than genomic alterations alone. Our observation of methylation-associated repression of HER2 signaling further supports the notion that epigenetic regulation of oncogenic signaling networks may represent a convergent mechanism underlying resistance to immunotherapy.

Several proliferative and survival pathways, including Ras, MAPK, Wnt, NF-κB, and mTOR showed promoter hypermethylation and appeared down-regulated in responders. Interestingly, these pathways have been reported to exert immunosuppressive effects acting on the PD1/PD-L1 axis: on one hand, an upregulation in RAS signaling via MEK increases PD-L1 expression by the stabilization of its messenger RNA in lung cancer cell lines⁷² and lung adenocarcinomas.⁷³ Increased MAPK and PI3K/mTOR signaling also suppress immune responses by increasing PD-L1 expression in melanoma cells.⁷⁴ Previous evidence on patients with NSCLC treated with ICI reported a heightened transcriptional activity in the NF-kB in non-responders.⁷⁵ Also, enhanced NF-κB signaling can upregulate PD-L1 expression at transcription and post-transcriptional levels⁷⁶ and has been reported to induce immune suppression in ovarian cancer through PD-L1 upregulation.⁷⁷

Classical immune-related pathways that are crucial for a successful response to immune checkpoint inhibition, including antigen processing and presentation, TCR-MHC signaling, CD8+ T cell activation, T cell antitumor response and positive leukocyte migration, were positively enriched in responder patients. In silico TME deconvolution also revealed immunological features associated with ICI benefit, including increased CD8+ T cell infiltration and M1 macrophage abundance. Importantly, these findings were complemented by orthogonal baseline immune profiling of PBMC, which identified distinct systemic immune states between responders and non-responders, characterized by higher frequencies of naïve CD4+ T cells and lower PD-1+CD4+ T cells in responders, supporting the existence of a more immune-competent baseline immune landscape. It is well recognized that, in NSCLC, favorable prognosis and clinical benefit from ICI are associated with increased CD8+ T cell infiltration^78–81 and proinflammatory M1 macrophages.^{80 82 83} We also observed lower fractions of immunosuppressive cell populations, including M2 macrophages, which have been associated with poor prognosis^{84 85} and resistance to ICI,⁸³ together with reduced fibroblasts and activated myeloid dendritic cells, both of which contribute to immune tolerance through the secretion of immunomodulatory cytokines and promotion of T cell exclusion.^86–90 Collectively, these observations support the existence of an immune-permissive tumor state characterized by coordinated epigenetic repression of oncogenic signaling together with enhanced antigen presentation and adaptive immune activation. Although our study was not designed to establish causality, the remarkable concordance across DNA methylation, transcriptomics, tumor deconvolution and an independent PBMC validation cohort strongly supports the biological relevance of this coordinated immune-permissive state.

Increased TE activity has previously been associated with the induction of antiviral response signatures, leading to the so-called VMP, which has been proposed to enhance tumor immunogenicity and potentially improve responsiveness to ICI.^{22 30} In fact, we have recently reported that patients with NSCLC and muscle-invasive bladder cancer responding to ICI showed a higher rate of L1 active transductions.⁹¹ In our real-world NSCLC cohort, we evaluated the presence of VMP and the contribution of TE activity to ICI outcome. Locus-specific analyses of TE revealed distinct differences between responders and non-responders, particularly involving L1 elements that were hypomethylated in non-responding patients. This observation aligns with prior reports linking global L1 hypomethylation, as a surrogate of global DNA methylation, to poor prognosis and immune evasion in patients with NSCLC treated with ICI.^{19 65 92} Previous studies have also linked an improved ICI response in NSCLC with HERV overexpression.³³ Also, a HERV-driven antitumor immune response was also reported in small-cell lung cancer.⁹³ In this sense, a novel aspect of our findings was the demonstration that, in our NSCLC cohort, total L1 load rather than HERV was strongly associated with antiviral response signatures, including activation of the dsRNA-sensing pathway, type II IFN signaling, JAK/STAT pathway activation, and induction of ISGs, suggesting a central role for L1 activity in driving the VMP state in this setting. Although overall TE load was not predictive of improved ICI response, we observed that patients with higher TE load experienced longer OS irrespective of their antiviral response status. Together, these observations suggest that TE activity, particularly L1 activation, contributes to tumor immunogenicity and may influence long-term benefit from immune checkpoint blockade, even if it does not directly predict short-term clinical response.

Emerging evidence suggests that the ECM is a key factor that modulates the TME and influences antitumor immune responses.^{94 95} Our data showed that DNA methylation also influences the ECM organization in NSCLC, as the ECM-receptor interaction, focal adhesion and tight junctions signatures appeared hypermethylated and showed a concordant downregulation in responder patients, which aligns with the fact that a stiffer, denser, and fibrotic ECM hampers tumor T cell infiltration and decreases the efficacy of ICI.^96–98 Recently, collagen-producing macrophages have been shown to promote a stiff, fibrotic TME that may limit CD8+ T cell activity.^{99 100} This relevance of ECM in ICI response was further supported by liquid biopsy studies, where serum markers of increased collagen formation predicted poor outcomes in ICI-treated patients with melanoma.¹⁰¹

Unexpectedly, PRC2-related signatures were enriched in responders, contrasting with previous reports linking PRC2 activity to immune evasion and resistance to immunotherapy.^102–104 These findings suggest that the immunological role of PRC2 is context dependent and likely influenced by its interaction with additional epigenetic mechanisms regulating chromatin accessibility and gene expression.¹⁰⁵

Finally, in our real-world cohort neither PD-L1 staining nor PD-L1 expression showed an association with PFS nor OS, adding evidence to the fact that PD-L1 alone may not be predictive of ICI benefit.^106–108 A high TMB was also proposed to guide the use of ICI such as pembrolizumab in patients with solid tumors.^{11 12} However, our NSCLC responder patients showed a lower mean TMB than non-responders, and the association between baseline TMB and response to ICI was not significant, once again evidencing the lack of predictive power of currently used biomarkers, and highlighting the need to consider other factors as complementary dimensions of ICI responsiveness. In this sense, our integrative analyses enabled the identification of novel potential predictors of ICI response in NSCLC. On one hand, we identified a 12-CpG hypermethylation signature that was associated with longer PFS with high accuracy. This signature aligns with the evidence that DNA hypermethylation markers may predict a favorable ICI outcome in NSCLC as reported by several studies on tissue^{19 64 65} and liquid biopsies.⁹² Moreover, we built a multimarker signature to predict ICI response, including transcriptomic and DNA methylation markers, signature-enrichment scores and clinical variables that achieved robust predictive performance under nested cross-validation (cross-validated AUC 0.833). These results challenge the reliance on single biomarkers such as PD-L1 staining or TMB, instead moving towards multiomics signatures that capture multiple factors/layers influencing immune response.

Despite the strengths of the multiomic integration and the consistency of the observed biological patterns, this study has several limitations. The relatively modest cohort size warrants validation in larger, independent prospective studies. Nevertheless, the deliberate inclusion of treatment-naïve patients receiving first-line ICI reduced treatment-related heterogeneity, while the high concordance observed in sensitivity analyses adjusting for histological subtype, treatment line and treatment regimen, together with the validation of key findings in an independent external dataset, supports the robustness and biological relevance of the identified molecular signatures. Finally, although our findings strongly associate coordinated epigenetic repression of oncogenic pathways with clinical benefit, the mechanistic contribution of these pathways to immune modulation requires dedicated functional studies beyond the scope of this patient-derived multiomic analysis.

Together, our findings define an epigenetically orchestrated immune-permissive tumor state associated with clinical benefit from immune checkpoint blockade. We identified a hypermethylation signature associated with longer PFS, developed a multimarker classifier with robust predictive performance, and demonstrated that coordinated repression of oncogenic pathways, including PI3K-Akt/mTOR signaling and ECM remodeling, together with L1-driven viral mimicry, characterizes tumors with improved immunotherapy outcomes.

From a therapeutic perspective, our findings provide a rationale for targeting epigenetically regulated oncogenic pathways to enhance immunotherapy efficacy. In particular, the epigenetic downregulation of PI3K-Akt signaling observed in responder tumors, together with previous evidence showing that PI3K inhibition enhances the efficacy of immune checkpoint blockade by reducing myeloid-mediated immunosuppression and promoting T cell infiltration, supports further evaluation of PI3K-targeted combination strategies in advanced NSCLC. More broadly, our findings highlight epigenetic regulation as a promising therapeutic avenue to overcome resistance and improve clinical benefit from ICI.

Data availability statement

Data are available upon reasonable request.