Identification and external validation of a prognostic signature based on N6-methyladenosine- and tertiary lymphoid structures-related genes to evaluate survival prognosis and treatment efficacy in lung adenocarcinoma

Abstract

Background

The N6-Methyladenosine (m6A) RNA modification critically regulates cancer biology, and tertiary lymphoid structures (TLSs) shape antitumor immunity. However, their combined prognostic roles in lung adenocarcinoma (LUAD) remain unclear. This study applies an integrative multi-omics approach to construct an m6A- and TLS-related prognostic model and uncover underlying molecular mechanisms.

Methods

We identified m6A and TLS-related genes (MTGs) associated with LUAD and constructed a prognostic model using machine learning, which was validated with nomograms. Subsequent analyses included immune microenvironment profiling, tumor mutational burden (TMB), enrichment assays, drug sensitivity testing, and single-cell RNA sequencing (scRNA-seq). The expression of MTGs was detected using quantitative reverse transcription polymerase chain reaction (RT-qPCR).

Results

The risk model we developed demonstrated strong prognostic value, with areas under the curve (AUCs) exceeding 0.8 at 1, 3, and 5 years. The prognosis of the high-risk cohort (HRC) was significantly worse (P < 0.001). A nomogram incorporating this risk model (AUC = 0.825) outperformed one without it. TMB analysis revealed a higher TMB in the HRC, which is likely associated with a poorer prognosis. Drugs targeting the microtubule dynamics and apoptosis pathways showed increased efficacy in the HRC. Enrichment analysis indicated that the MTGs are primarily involved in cell adhesion, immune response, hematopoietic cell lineage, and cell cycle regulation. The scRNA-seq analysis further revealed that these 8 MTGs are predominantly expressed in fibroblasts and T/NK cell clusters, indicating their possible involvement in regulating local immune responses. RT-qPCR analysis confirmed the differential expression of MTGs.

Conclusions

This study demonstrates that the integrative multi-omics model reveals not only potential therapeutic targets but also new perspectives on LUAD immunogenetics.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-04318-w.

Introduction

Lung cancer, especially lung adenocarcinoma (LUAD), remains a leading cause of cancer mortality worldwide [1]. Despite advancements in targeted therapies and immunotherapy, prognosis remains poor due to tumor heterogeneity and complex immune regulation [2]. The tumor lymph node metastasis (TNM) classification is a standard tool for tumors prognosis, yet its predictive accuracy is limited [3, 4]. N6-methyladenosine (m6A), the most prevalent internal RNA modification, is known to regulate multiple aspects of cancer biology, including proliferation, apoptosis, and immune evasion [5, 6]. Recent studies have shown that m6A modification plays a critical role in shaping the immune microenvironment of lung cancer, influencing immune-cell activation, antigen presentation, and antitumor responses [7]. Meanwhile, tertiary lymphoid structures (TLSs), which form in non-lymphoid tissues under chronic inflammation, are increasingly recognized as crucial modulators of local immune responses and prognostic indicators in diverse cancers [8]. Despite the critical roles of both m6A regulation and TLS biology in LUAD, their combined prognostic value and potential interaction have not yet been elucidated.

Prior studies have largely focused on single cell omics or transcriptome-level analyses, which fail to capture the multidimensional regulation underlying tumor progression and immune crosstalk [9]. Integrative multi-omics strategies, combined with Mendelian randomization (MR), offer powerful frameworks to identify genes with potential causal effects on LUAD outcomes [10]. In parallel, single-cell RNA sequencing (scRNA-seq) provides fine-grained insights into the complexity of the immune landscape, while machine learning (ML) approaches deliver robust frameworks for developing clinically applicable prognostic models [11, 12]. Despite the availability of numerous m6A-related, TLS-related, and multi-omics-derived prognostic models in LUAD, few studies have systematically evaluated the combined prognostic contribution of m6A regulation and tertiary lymphoid structure biology [13]. To our knowledge, there are no published studies integrating m6A modifications with TLS-related genes to construct a unified prognostic framework and introducing MR to support gene selection.

In this study, we systematically integrated MR and ML to construct and validate a model based on m6A and TLS-related genes (MTGs) for LUAD. Downstream analyses, including pathway enrichment, tumor mutational burden (TMB), drug sensitivity, and tumor microenvironment (TME) profiling, were performed to explore the immunogenetic mechanisms underlying LUAD prognosis. The cellular localization and heterogeneity of key MTGs were revealed by scRNA-seq. Finally, we used reverse transcription quantitative polymerase chain reaction (RT-qPCR) in vitro to detect the expression of 8 MTGs in lung cancer and normal cell lines.

Materials and methods

Statistical workflow overview

To enhance the transparency and reproducibility of all analytical procedures performed in this study, we provide an overview of the complete statistical workflow used for data preprocessing, causal gene identification, ML-based model construction, external validation, and downstream functional analyses. The entire workflow consisted of the following major components: (1) data acquisition and preprocessing of TCGA and GEO cohorts; (2) normalization and batch-effect correction of external validation datasets; (3) differential expression analysis; (4) MR analysis for causal gene screening; (5) construction of the ceRNA regulatory network; (6) multi-algorithm machine learning modeling; (7) internal and external model validation; (8) characterization of the immune microenvironment and functional pathway; and (9) drug sensitivity prediction. All software tools, R packages, version numbers, key parameter settings, and validation strategies used in each analytical step are summarized in Table S1.

Data downloads

The tumor data are primarily from the Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/, accessed on August 1, 2025). In addition, two external validation cohorts (GSE30219 and GSE50081) and one scRNA-seq cohort (GSE189357) were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/, accessed on August 1, 2025). To identify candidate MTGs for differential analysis, 7575 TLS-related genes (TRGs) and 1440 m6A-related genes (MRGs) were retrieved from the GeneCards database (https://www.genecards.org/, accessed on August 1, 2025) by conducting a targeted search for “tertiary lymphatic structure” and “N6-methyladenosine” and applying a score cut-off value of more than 1, which resulted in 9084 MTGs after deduplication. The expression profiles of these MTGs in the TCGA-LUAD dataset were analyzed using “limma”, and differentially expressed genes (DEGs) were selected according to strict criteria for further study: |log2 FC| was greater than 0.585 and the false discovery rate (FDR) was less than 0.05 [14].

In MR analysis, we used summary statistics for LUAD and blood-derived eQTLs obtained from the eQTLGen consortium through the IEU Genome-Wide Association Studies (GWAS) platform (https://gwas.mrcieu.ac.uk/) [15]. The eQTL dataset contained 19,942 genetic loci influencing gene expression, derived from blood samples of 31,684 healthy European individuals. Outcome data for LUAD were drawn from the same platform, encompassing 11,273 cases and 55,483 controls (study ID: ebi-a-GCST004744), comprising 7,849,324 single nucleotide polymorphisms (SNPs).

MR statistical assessment

To infer the causal effect of gene expression on LUAD risk, we performed MR using eQTL-derived instrumental variables (IVs). The MR framework relies on three core assumptions: (1) the genetic variants used as IVs are strongly associated with the exposure (gene expression); (2) the IVs are not associated with any confounders of the exposure–outcome relationship; (3) the IVs influence the outcome (LUAD) exclusively through the exposure and not through alternative biological pathways (no horizontal pleiotropy). The SNPs (P < 5 × 10⁻⁸) associated with gene expression were selected as IVs, based on eQTL summary statistics derived from GWAS datasets. To reduce confounding from strong linkage disequilibrium (LD), SNPs with pairwise r² > 0.001 within a 10,000 kb window of the lead SNPs were excluded. The intensity of each IV was subsequently assessed with the F statistic, and SNPs of F ≤ 10 were excluded to reduce weak instrument bias. The F-statistic was calculated as: F = (R² / K) / ((1 − R²) / (N − K−1)), where N is the sample size, K is the number of IVs, and R² represents the variance in gene expression explained by the IVs.

The R package “TwoSampleMR” was employed to perform five complementary MR methods for assessing the causal relationship between eQTLs and LUAD, with the inverse-variance weighted (IVW) approach used as the primary analysis and the MR-Egger, weighted median, simple mode, and weighted mode methods applied as supplementary analyses [16]. Several diagnostic methods were used to evaluate pleiotropy, heterogeneity, influential SNPs, and the robustness of the results. Funnel plots were generated to illustrate potential biases in causal estimation, and the impact of individual SNPs on the results was evaluated. The MR-Egger intercept was utilized to assess horizontal pleiotropy, with P-values under 0.05 indicating significant pleiotropy, hence undermining the causal inference. Only genes showing consistent causal estimates across MR methods, with no evidence of pleiotropy and sufficient instrument strength, were retained as “MR-supported causal genes” These genes were then intersected with differentially expressed m6A-related and TLS-related genes to generate a biologically and causally informed set of MTGs for downstream model construction.

Signature development via integrated machine learning framework

To ensure a robust and stable selection of MTGs, we implemented a comprehensive framework that combined 10 distinct ML algorithms across 117 configurations [17]. The prognostic modeling procedure entailed detecting the intersection between DEGs and genes obtained from MR analysis, facilitating the selection of MTGs with significant differential expression in LUAD. A total of 117 ML algorithms were applied to construct a risk model for these genes using Leave-One-Out Cross-Validation (LOOCV). The models were further validated through additional cross-validation using two independent GEO datasets (GSE30219 and GSE50081). To assess the performance of the models, the concordance index (C-index) was calculated for 117 models across these datasets, the risk model with the highest mean C-index was identified as the best model.

Validation model and nomogram

Risk scores for the prognostic model genes were calculated using the optimal algorithm, and tumor samples were stratified into high- and low-risk cohorts (HRC and LRC) based on the median risk threshold. Survival analysis was performed using the R packages “survival” and “survminer” to evaluate overall survival (OS) among groups [18]. Differences between the HRC and LRC were visualized using scatter plots of risk distribution, principal component analysis (PCA), and gene expression heatmaps. The K-M curves based on various clinical features were generated to investigate their viability in the model.

To validate the predictive significance of MTGs as independent markers in tumor populations, univariate and multivariate Cox regression analyses (UCRA and MCRA) were used to investigate the relationships between MTGs and pertinent clinical factors [19]. A nomogram was constructed based on the TCGA-LUAD cohort with the R packages “rms” and “replot”. Calibration plots, decision curve analysis (DCA), and receiver operating characteristic (ROC) curves were utilized to test the predictive accuracy and clinical applicability of the nomogram. In addition, we compared representative prognostic models published in recent years, and the MTG-based models achieved comparable or even better predictive performance at multiple time points, highlighting their robustness and clinical utility.

Enrichment analysis

The Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses were employed to investigate the biologically significant functions and pathways that distinguish the HRC from the LRC [20].

TMB analysis

Using the R package “maftools”, we computed the TMB scores and generated waterfall plots to visualize the mutation data from the TCGA-LUAD cohort [21]. In addition, we performed survival analysis to identify differences in OS between patients classified by different TMB levels.

Tumor microenvironment analysis

Using the R package “ESTIMATE”, we assessed the differences in ESTIMATE, matrix, and immune scores between the HRC and LRC. In addition, we used various algorithms, such as XCELL, TIMER, QUANTISEQ, and MCPCOUNTER, for immune scoring [22]. We conducted immune infiltration analyses to assess the proportion of different immune cell populations in various cell lines. Single-sample gene set enrichment analysis (ssGSEA) was performed to calculate immune function scores for each sample, and differences in these scores between the HRC and LRC were compared.

Immunotherapy and chemotherapy response in LUAD patients

To investigate immune dysfunction and rejection differences across the HRC and LRC, TIDE scores were obtained via the Tumor Immune Dysfunction and Exclusion (TIDE) platform (http://tide.dfci.harvard.edu/login/) [23]. TIDE scores were compared across the HRC and LRC to assess differences in immunotherapy responses. Using the “oncoPredict” R package, we calculated the half-maximal inhibitory concentration (IC50) of selected antitumor drugs [24]. We assessed the difference in drug response to drugs across the HRC and LRC depending on IC50 values.

ScRNA-seq analysis of MTGs

The scRNA-seq dataset GSE189357 used in our study consisted of 9 LUAD samples. The R package “Seurat” was used to preprocess the scRNA-seq data by converting it into Seurat objects. Cells with fewer than 500 detected genes or fewer than 3 detected genes were excluded. The mitochondrial gene proportion was calculated using the “PercentageFeatureSet” function, and violin plots were generated with the “VlnPlot” function to visualize their distribution. Cells with mitochondrial gene percentages exceeding 25% were removed as outliers. For dimensionality reduction and batch effect correction, PCA and Harmony were applied. PCA reduced the data to a lower-dimensional space, and Harmony was applied to correct batch effects. Uniform Manifold Approximation and Projection (UMAP) was employed to visualize cell clustering. Intercellular distances were computed using the “FindNeighbors” function, and clustering was performed based on the resolution settings in the “FindClusters” function. Finally, cell type annotation was performed using the “SingleR” R package [25].

CeRNA signaling network

The miRanda (http://www.microrna.org/) and miRDB (http://www.mirdb.org) databases were used to identify miRNAs associated with 8 MTGs, while the SpongeScan database (http://www.targetscan.org/) was used to identify miRNA-lncRNA interactions [26]. Cytoscope software (v3.8.0) was used to create and visualize lncRNA-mRNA-miRNA competing endogenous RNA (CeRNA) regulatory networks.

Verify protein and mRNA expression of MTGs

We examined the protein expression levels of MTGs across different tissues using the Human Protein Atlas (HPA) website (https://www.proteinatlas.org/). Furthermore, we obtained a human lung fibroblast cell line ( BEAS-2B), and three human lung adenocarcinoma cell lines (PC-9, A549 and H1975) from Sunncell Biotechnology Co., Ltd. Total RNA was extracted from these cell lines using TRIzol Reagent and reverse transcribed with the Prime Script RT Kit (VWR International). Gene expression levels were analyzed by RT-qPCR [27]. Table S2 lists the primers for 8 MTGs.

Results

Causal relationship between MTGs and LUAD

Figure 1 presents the specific mechanisms of TLSs and m6A, while Fig. 2 illustrates the methodological roadmap of this study. MR analysis identified 254 gene symbols causally associated with LUAD, supported by IVs with F-statistics > 10 and consistent results across five MR methods, demonstrating robust causal estimates (Table S3). We obtained the TCGA-LUAD data, including 503 tumor patients and 59 normal patients (Table 1). A total of 9,084 MTGs were identified from the TCGA-LUAD cohort, and 2,221 DEGs were obtained after filtering (Fig. 3A). Among the 254 MR-supported causal genes, we intersected these with the DEGs derived from m6A- and TLS-related gene sets, yielding 36 genes that were not only differentially expressed in LUAD but also demonstrated causal evidence (Table S4, Fig. 3B). Cochran’s Q test and the MR-Egger intercept results showed no significant heterogeneity or pleiotropy within the dataset (Table S5).

Fig. 1.

Mechanistic overview of TLS formation and m6A RNA modification. Schematic representation illustrating the biological functions and regulatory pathways of TLSs and m6A methylation and their potential interaction in modulating the TME

Fig. 2.

Study workflow. Flowchart summarizing the overall analytical pipeline, including data preprocessing, MR-based causal gene screening, ML model construction, validation, immune analyses, and experimental verification

Table 1.

Clinical information of the patients in the TCGA and two GEO groups

Characteristics	TCGA(n = 503)		GSE30219(n = 85)		GSE50081(n = 127)
	n	%	n	%	n	%
Age
<=65	238	47.32	60	70.59	42	33.07
> 65	255	50.70	25	29.41	85	66.93
Unknow	10	1.99	-	-	-	-
Status
Alive	321	63.82	40	47.06	76	59.84
Dead	182	36.18	45	52.94	51	40.16
Gender
Female	272	54.08	19	22.35	62	48.82
Male	231	45.92	66	77.65	65	51.18
Stage
Stage I	270	53.68	-	-	92	72.44
Stage II	120	23.86	-	-	35	27.56
Stage III	80	15.90	-	-	-	-
Stage IV	25	4.97	-	-	-	-
Unknow	8	1.59	-	-	-	-
T stage
T1	168	33.40	71	83.53	43	33.86
T2	269	53.48	12	14.12	82	64.57
T3	45	8.95	2	2.35	2	1.57
T4	18	3.58	-	-	-	-
Unknow	3	0.60	-	-	-	-
M stage
M0	335	66.60	85	100	127	100
M1	24	4.77	-	-	-	-
Unknow	144	28.63	-	-	-	-
N stage
N0	326	64.81	82	96.47	94	74.02
N1	95	18.89	3	3.53	33	25.98
N2	69	13.72	-	-	-	-
N3	2	0.40	-	-	-	-
Unknow	11	2.19	-	-	-	-

GEO Gene Expression Omnibus, TCGA The Cancer Genome Atlas, T stage Tumor stage, N stage Node stage, M stage metastasis stage

Fig. 3.

Identification of MTGs and construction of the prognostic gene set. A Volcano plot of differentially expressed MTGs between tumor and normal tissues. B Venn diagram showing the intersection of MR-derived causal genes and DEGs. C, D LASSO regression for feature selection: coefficient profiles C and penalty parameter tuning D. E K-M curves identifying eight prognosis-related MTGs. F Circos plot showing chromosomal locations and expression distribution of the eight MTGs

Establish and validate the risk model

Using a comprehensive ML framework integrating the TCGA-LUAD and GEO datasets, we evaluated 117 predictive models and computed their corresponding C-indices (Figure S1A). The combination of LASSO (Least Absolute Shrinkage and Selection Operator) and RSF (Random Survival Forest) achieved the highest C-index (Fig. 3C-E). Accordingly, this algorithm was selected for model construction, ultimately identifying the top eight most important MTGs: LOXL2, CBFA2T3, CD33, IER3, RUNX2, CENPK, CCDC50, and SPTBN1. To clarify the biological basis of MTGs, we further mapped the final eight model genes back to the original MRG and TRG lists retrieved from GeneCards. According to the keyword-based relevance scoring system (> 1), LOXL2, RUNX2, IER3, CBFA2T3, CD33, SPTBN1, and CCDC50 showed documented associations with TLS formation, immune cell chemotaxis, or stromal remodeling. In contrast, SPTBN1 and CENPK demonstrated significant functional links to m6A machinery-related pathways or m6A-regulated immune processes. Notably, SPTBN1 displayed evidence supporting both TLS-related and m6A-related biological roles, and thus was considered a dual-related MTG. This classification supports the mechanistic specificity of the MTGs incorporated in our prognostic model.

The chromosomal position of each gene is shown in Fig. 3F. Forest, funnel, and scatter plots of the MR results indicated low heterogeneity and stable causal estimates for each of the 10 model genes (Figs. S2-4). Risk scores for individual LUAD patients were generated via the predict function of the LASSO + RSF model, and patients were stratified into HRC and LRC. Table 2 summarizes the clinical information for the HRC and LRC populations. K-M curves revealed significantly longer OS in the LRC than in the HRC across all cohorts, with P-values for the TCGA, GSE30219, and GSE50081 cohorts all below 0.001. Consistently, time-dependent ROC curves demonstrated strong predictive performance, with 1-, 3-, 5-year Areas Under the Curve (AUCs) of 0.821/0.815/0.823 (TCGA), 0.728/0.791/0.757 (GSE30219) and 0.673/0.649/0.682 (GSE50081) (Figure S1B-D). The 1, 3, and 5-year AUC values of the MTG-related prognostic model and the existing LUAD prognostic characteristics can be seen in Table S6. Scatter plots of risk scores, PCA results, and gene expression heatmaps further distinguished HRC and LRC across cohorts (Fig. S5A-C).

Table 2.

Clinical information for 503 TCGA patients in different risk categories

Characteristics	High-risk group (n = 251)		Low-risk group (n = 252)
	n	%	n	%
Age
<=65	134	53.39	104	41.27
>65	114	45.42	141	55.95
Unknow	3	1.20	7	2.78
Status
Alive	124	49.40	197	78.17
Dead	127	50.60	55	21.83
Gender
Female	127	50.60	145	57.54
Male	124	49.40	107	42.46
Stage
Stage I	108	43.03	162	64.29
Stage II	70	27.89	50	19.84
Stage III	56	22.31	24	9.52
Stage IV	14	5.58	11	4.37
Unknow	3	1.20	5	1.98
T stage
T1	67	26.69	101	40.08
T2	143	56.97	126	50.00
T3	30	11.95	15	5.95
T4	10	3.98	8	3.17
Unknow	1	0.40	2	0.79
M stage
M0	169	67.33	166	65.87
M1	14	5.58	10	3.97
Unknow	68	27.09	76	30.16
N stage
N0	140	55.78	186	73.81
N1	59	23.51	36	14.29
N2	48	19.12	21	8.33
N3	1	0.40	1	0.40
Unknow	3	1.20	8	3.17

TCGA The Cancer Genome Atlas, T stage Tumor stage, N stage Node stage, M stage metastasis stage

Prediction nomogram for this model

To evaluate the independent prognostic relevance of MTGs, we performed UCRA and MCRA. UCRA identified MTGs as significant risk factors (HR = 2.664, 95% CI: 2.300-3.085, P < 0.001) (Fig. 4A), and MCRA further confirmed their independent prognostic value (HR = 2.524, 95% CI: 2.171–2.936, P < 0.001) (Fig. 4B). A nomogram incorporating clinical features was constructed to generate individualized survival predictions for patients with LUAD (Fig. 4C). ROC and DCA analyses indicated that our model exhibited superior predictive performance compared with clinical traits alone (Fig. 4D-E). The calibration curve supported the model’s reliability, as the C-index of the nomogram incorporating this risk model exceeded that of the nomogram without it (Fig. 4F-G). Scatter plots illustrated associations between clinical traits and risk scores (Fig. S6A-C). A heatmap integrated the expression patterns of the 8 MTGs with clinical data and risk groups (Fig. S6D). Stratified survival analyses across clinical subgroups consistently validated the predictive value of the model (Fig. S7A-L).

Fig. 4.

Development and validation of the MTG-based prognostic signature. A, B UCRA and MCRA evaluating clinical variables and risk score in TCGA-LUAD. C Nomogram integrating clinical features and MTG risk score. D–E ROC curves and DCA assessing predictive performance. F–G Calibration curves comparing predicted and observed 1-, 3-, and 5-year OS probabilities with and without incorporating the risk model

Enrichment analysis

We conducted GO and KEGG analyses to explore the functional landscape of MTGs. The top six biological processes (BPs) in the GO analysis were primarily associated with cell adhesion regulation, chemotaxis, wound healing, and leukocyte migration, emphasizing the involvement of MTGs in immune cell recruitment and tissue repair. In the Cellular Component (CC) category, MTGs were primarily enriched on the external side of the plasma membrane and within the cytoplasmic vesicle lumen, suggesting that their functions are largely localized to membrane-associated or extracellular compartments. In the Molecular Function (MF) category, MTGs were primarily enriched in integrin binding, growth factor binding, and cytokine receptor interactions, highlighting their roles to intercellular signaling and immune modulation (Fig. 5A, Table S7). KEGG pathway analysis revealed that the top six enriched pathways were hematopoietic cell lineage, cell cycle, complement and coagulation cascades, phagosome, PI3K-Akt signaling, and cell adhesion molecules, collectively indicating that MTGs are involved in hematopoiesis, immune activation, and proliferative regulation (Fig. 5B, Table S8).

Fig. 5.

GO and KEGG enrichment analysis. A GO enrichment circle plot. B KEGG pathway enrichment bar plot

Tumor mutational burden

We calculated the TMB value for each patient. Waterfall plots were generated to visualize the mutational landscape of the top mutated genes in both groups. In the LRC, approximately 87.3% of specimens harbored genetic mutations (Figure S8A), whereas in the HRC, mutations were detected in approximately 95.6% of specimens (Fig. S8B). Significant differences in TMB were observed between the HRC and LRC (P < 0.0001) (Fig. 8C). Furthermore, the samples were stratified into categories of high and low mutation levels, revealing notable discrepancy in survival outcomes (P < 0.001) (Fig. S8D).

Immune microenvironment analysis

According to immune infiltration analysis, the HRC exhibited higher proportions of resting NK cells, activated memory CD4⁺ T cells, and M0 macrophages. In contrast, the LRC showed higher levels of plasma cells, activated NK cells, eosinophils, monocytes, resting dendritic cells, resting memory CD4 T cells, resting and activated mast cells (Fig. 6A). Immune function score analysis revealed that, compared with the LRC, the HRC exhibited significantly lower scores for APC co-inhibition, parainflammation, inflammation-promoting activity, cytolytic activity, checkpoint pathways, HLA function, APC co-stimulation, CCR signaling, T-cell co-inhibition, T-cell co-stimulation, and the type II IFN response (Fig. 6B). In addition, ESTIMATE score (Fig. 6C), the immune score (Fig. 6D), and stromal score (Fig. 6E) were significantly higher in the LRC. Conversely, tumor purity score was lower in the LRC than in the HRC (Fig. 6F). Multiple algorithms were employed to further investigate the link between immune cell types and risk scores (Figs. 6G, S9). Taken together, these results indicate that the HRC is characterized by a less immunoreactive and more suppressive tumor microenvironment, whereas the LRC exhibits a more active and immunostimulatory immune landscape. Such differences in immune infiltration and ESTIMATE-derived stromal/immune scores further support the immunological relevance of the MTG-based risk signature and align with its prognostic stratification performance.

Fig. 6.

TME characteristics between HRC and LRC. A Relative abundance of 22 immune-cell types estimated by CIBERSORT. B ssGSEA-based immune-function scores. C–F Comparison of ESTIMATE score, immune score, stromal score, and tumor purity between HRC and LRC. G Correlation between risk score and immune-cell infiltration across six algorithms. [*Statistical significance: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001]

Immune targets and chemotherapy

TIDE analysis revealed that the HRC characterized by MTGs exhibited an elevated T-cell exclusion score (Fig. S10B), whereas T-cell dysfunction scores (Fig. S10A) and TIDE scores (Fig. S10C) were lower compared with the LRC. These results suggest that MTGs in the HRC are associated with enhanced T-cell exclusion and reduced T-cell dysfunction and TIDE scores, indicating a stronger immune evasion mechanism in this subgroup. Furthermore, we calculated the IC50 value of antitumor drugs for both the HRC and LRC. A total of 106 chemotherapeutic agents were screened, revealing significant differences in estimated IC50 values between the two subgroups (Table S9). The HRC showed greater sensitivity to chemotherapeutic agents targeting microtubule dynamics and apoptosis pathways, such as Docetaxel_1007 and Sepantronium.bromide_1941. In contrast, the LRC was more sensitive to chemotherapeutic agents targeting the PI3K-AKT-mTOR and transcription-regulation pathways, such as Dactolisib_1057 and Dactinomycin_1811. The IC50 values of these drugs can help guide the selection of appropriate treatments.

ScRNA of MTGs

To explore the relationship between the eight MTGs and immune cells, we performed single-cell profiling. After quality control using the “Seurat” R package, 23,451 genes and 118,602 cells were retained for downstream analysis (nFeature: 300-5,000; nCount: 5,000–20,000; percent.mt: 1%-10%) (Fig. S11A-C). The percentage of mitochondrial genes (percent.mt) exhibited a negative correlation with nCount (r = -0.11) (Fig. S11D), while nFeature exhibited a strong positive correlation with nCount (r = 0.9) (Fig. S11E). PCA revealed minimal batch effects, with samples being uniformly distributed. Using the Harmony ElbowPlot to identify the inflection point, the first 10 principal components (PCs) were selected for further analysis, as they captured the majority of the signal (Fig S11F-G). UMAP dimensionality reduction and clustering at a resolution of 0.1 identified 11 distinct cell subclusters (Figs. S11H, 7A), annotated as B cells, endothelial cells, epithelial cells, fibroblasts, mast cells, myeloid cells, plasma cells, proliferative cells, and T/NK cells (Fig. 7B). The eight key MTGs were predominantly expressed in fibroblasts and T/NK cells (Fig. 7C), highlighting their potential roles in regulating immune dynamics within the TME.

Fig. 7.

Single-cell expression profiles of MTGs in LUAD. A, B Clustering and annotation of cell subpopulations at different resolutions. C Expression distribution of the eight MTGs across annotated cell types

CeRNA network of MTGs

We established a ceRNA network containing lncRNA-miRNA-mRNA interactions, demonstrating that multiple lncRNAs and miRNAs may interact with or be regulated by MTGs. This network provides mechanistic insights into the post-transcriptional regulation of MTGs in LUAD and suggests potential molecular axes for therapeutic targeting (Fig. S14) (Table S10).

Verify protein and mRNA levels of MTGs

Protein expression levels of MTGs were assessed in both lung cancer and normal tissues (Fig. 8A). RT-qPCR analysis revealed that CBFA2T3, CD33, CCDC50 and SPTBN1 were predominantly expressed in normal cell lines, whereas higher levels of LOXL2, IER3, RUNX2, and CENPK were detected in LUAD cell lines (Fig. 8B).

Fig. 8.

Protein and mRNA expression levels of MTGs. A HPA immunohistochemistry images comparing MTG protein levels in LUAD versus normal lung tissues. B RT-qPCR validation of MTG expression in LUAD and normal lung epithelial cell lines

Discussion

LUAD, as one of the most prevalent and deadly forms of non-small cell lung cancer (NSCLC), presents a major clinical challenge [1, 25]. The widely used TNM staging system has significant limitations in predicting outcomes due to LUAD’s molecular heterogeneity and diverse immune responses within tumors [28]. To address these limitations, immune-related gene signatures have been developed and shown to improve the prediction of prognosis and immunotherapy responsiveness in LUAD, highlighting the need to incorporate both epigenetic and immune determinants into prognostic modeling [29, 30]. Although the advent of targeted therapies and immunotherapy has significantly improved clinical outcomes in subsets of patients, many individuals still experience poor prognosis due to intrinsic or acquired resistance to these therapies [31]. Given this complexity, there is a growing need to identify novel prognostic biomarkers and predictive models that can enhance the precision of treatment decisions and patient management. This study addresses this need by integrating multi-omics approaches, including m6A RNA modification and TRG expression, to develop a comprehensive prognostic model. The risk model constructed in this study, which incorporates ML algorithms and MR, serves as an effective tool for forecasting the prognosis of LUAD patients with high accuracy and potential clinical utility [32]. Our findings demonstrate that the eight MTGs identified in this model can effectively predict the survival of patients with LUAD, contributing to the advancement of personalized medicine.

The model utilizing eight MTGs proves to be highly valuable in predicting the OS of LUAD patients. The AUCs of the risk model were all greater than 0.8 at 1, 3, and 5 years, indicating robust predictive power across various time points. This model divides LUAD patients into HRC and LRC, with significant differences in survival outcomes, providing a more granular approach than traditional TNM staging. The HRC exhibited significantly worse survival outcomes, highlighting the importance of this model in identifying patients who could benefit from more intensive treatment approaches [33]. Nomograms combined with this risk model outperform traditional nomograms containing clinical variables, achieving higher AUC values. This suggests that incorporating MTGs into the clinical decision-making process can enhance the accuracy of prognosis, providing better guidance for individualized treatment planning [34]. These findings align with recent studies showing that multi-omics-based models outperform traditional methods in predicting cancer prognosis and guiding therapeutic decisions [35, 36].

The interaction between m6A regulation and TLS-associated biology may synergistically shape the LUAD immune landscape by modulating RNA stability, stromal organization, and immune cell recruitment, thereby influencing tumor progression and clinical outcomes. The eight MTGs identified in this study—LOXL2, RUNX2, IER3, CBFA2T3, CD33, CCDC50, CENPK, and SPTBN1—play crucial roles in the progression and immune landscape of LUAD. Specifically, LOXL2, RUNX2, IER3, CBFA2T3, CD33, and CCDC50 are associated with TLS formation, immune-cell chemotaxis, or stromal remodeling; CENPK is functionally linked to m6A machinery–related pathways or m6A-regulated immune processes; and SPTBN1 is mechanistically involved in both TLS biology and m6A-regulated immune pathways. For example, CENPK has been shown to undergo m6A modification on its mRNA, which plays a role in promoting cancer stemness and treatment resistance [37]. IER3 undergoes m6A-dependent post-transcriptional regulation, which may influence its function in apoptosis and inflammatory signaling [38]. SPTBN1 inhibits CXCL1-driven recruitment of tumor-associated macrophages and M2 polarization, providing further mechanistic support for its role in shaping the tumor immune microenvironment and inflammatory signaling [39]. CCDC50 mediates selective lysosomes function to prevent reactive oxygen species (ROS)-induced cell death [40]. LOXL2 has been shown to contribute to the formation and remodeling of the extracellular matrix within TLS, thereby enhancing the invasiveness of cancer cells and promoting a pro-tumor immune environment [41]. RUNX2, a gene known for its role in osteogenesis, drives tumor growth by regulating the cell cycle [42]. CBFA2T3 regulates immune cell differentiation and migration, suggesting a role in shaping the immune response within the TME [43]. CD33, a surface marker predominantly expressed on myeloid cells, has been linked to immune suppression in cancers, potentially promoting an immunosuppressive microenvironment by inhibiting the activation of T cells and other immune effector cells [44]. Together, these findings suggest that the eight MTGs participate in interconnected regulatory networks involving m6A-dependent RNA modification, lysosomal homeostasis, stromal remodeling, myeloid cell regulation, and immune cell recruitment. Through these pathways, the MTGs collectively influence TLS organization, macrophage polarization, oxidative stress signaling, and cytotoxic immune activation, providing a mechanistic basis for their association with LUAD immune heterogeneity and patient prognosis.

Moreover, scRNA-seq analysis revealed that these genes are predominantly expressed in fibroblasts and T/NK cells. This finding suggests that the identified MTGs are involved in regulating the immune environment and tumor progression by interacting with both cancer and immune cells. In fibroblasts, MTG expression may modulate extracellular matrix remodeling, stromal organization, and fibroblast-derived chemokine signaling, processes known to influence immune-cell recruitment and activation [45]. Such alterations could contribute to the enrichment of resting NK cells and M0 macrophages observed in the HRC, reflecting a less immunostimulatory and more suppressive microenvironment. Specifically, fibroblasts serve a critical function in creating a supportive stromal environment, while T/NK cells are essential for immune surveillance and response [46–48]. MTG expression in T/NK cells may also influence cytotoxic activity, exhaustion status, and cytokine responsiveness, thereby helping explain why the LRC exhibited higher proportions of activated NK cells, dendritic cells, and monocytes [49]. These interactions may alter immune responses within the tumor, affecting tumor growth, immune evasion, and treatment resistance. Collectively, the LRC possesses a more immunoreactive and supportive TME, whereas the HRC is characterized by an immunosuppressive landscape with reduced cytolytic activity and impaired antigen-presentation capacity. Such differences between high- and low-risk groups further strengthen the biological relevance of our MTG-based signature and highlight its close association with TME regulation in LUAD.

The predictive power of the risk model extends beyond survival prognosis and into clinical treatment guidance. By evaluating the drug sensitivity profiles of patients based on their risk scores, the model offers valuable insights into potential therapeutic strategies. For instance, the HRC showed increased sensitivity to chemotherapy drugs targeting microtubule dynamics and apoptosis pathways, such as Docetaxel and Sepantronium. Docetaxel, a taxane-based chemotherapy agent, disrupts microtubule function, leading to cell cycle arrest and apoptosis [50, 51]. This action is crucial for targeting rapidly dividing tumor cells in the HRC. Similarly, Sepantronium, which targets apoptosis mechanisms and inhibits mitochondrial function, is particularly effective in this cohort due to the elevated expression of genes involved in cellular stress and survival mechanisms [52]. In contrast, the LRC exhibited greater responsiveness to chemotherapeutic agents targeting the PI3K-AKT-mTOR signaling and transcriptional regulation pathways, such as Dactolisib and Dactinomycin. The PI3K–AKT–mTOR pathway plays a central role in regulating cell growth, metabolism, and survival, and is frequently activated in cancer to drive uncontrolled cell proliferation [53]. Targeting this pathway with Dactolisib, a dual PI3K/mTOR inhibitor, offers a promising treatment strategy for the LRC by disrupting the oncogenic signaling that drives tumor growth [54]. These findings underscore the critical importance of tailoring therapies based on individual molecular profiles, optimizing therapeutic efficacy, and minimizing adverse effects by targeting specific pathways involved in each patient’s disease [55]. These results further highlight the value of incorporating genetic signatures into treatment planning, as such molecular profiles enable clinicians to select therapies most likely to benefit individual patients.

This study introduces several innovative aspects to LUAD prognosis and treatment. First, it integrates multi-omics data, including m6A RNA modification and TRG expression, into a comprehensive risk model, offering a more nuanced understanding of LUAD biology compared to traditional single-omics studies. The incorporation of ML algorithms enhances the robustness and predictive performance of the model, while the inclusion of MR analysis facilitates the identification of genes with potential causal relevance to LUAD outcomes. The scRNA-seq analysis provides a high-resolution characterization of the immune landscape, revealing cell type–specific expression patterns and immunological functions of MTGs. To more clearly demonstrate the novelty of our model, we incorporated a systematic comparison with previously published m6A-related, TLS-related, and immune-based prognostic signatures in LUAD. This comparison highlights that our integrated MTG model not only provides superior or comparable predictive performance but also captures a broader spectrum of immunogenetic mechanisms than existing models. However, our study has certain limitations. The primary limitation is the reliance on publicly available datasets, which may introduce biases related to sample size, cohort composition, or data quality. Although RT-qPCR confirmed the mRNA expression differences of the MTGs, protein-level and functional experiments such as IHC validation, gene perturbation assays, and immune cell co-culture models were not performed due to experimental constraints. These validations will be prioritized in future work to further substantiate the mechanistic roles of MTGs in LUAD.

Conclusion

The prognostic model based on the eight MTGs that we constructed and validated demonstrated strong predictive performance, providing an improved framework for individualized management of patients with LUAD. The m6A modification and TLS interact to critically influence the immune landscape and tumor progression. The model also revealed differential drug sensitivities between the HRC and LRC, underscoring the importance of tailoring therapies based on genetic characteristics. This study introduces innovative multi-omics and ML approaches for LUAD prognosis and treatment, but is limited by reliance on publicly available datasets and requires further experimental validation.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

Area under the curve

BP

Biological process

CC

Cellular component

CeRNA

Competing endogenous RNA

CI

Confidence interval

C-index

concordance index

DCA

Decision curve analysis

DEGs

Differentially expressed genes

FDR

False discovery rate

GEO

Gene expression omnibus

GO

Gene ontology

GWAS

Genome-wide association studies

HPA

Human protein atlas

HR

Hazard ratio

HRC

High-risk cohort

IC50

Half maximal inhibitory concentration

IVW

Inverse variance weighted

IVs

Instrumental variables

KEGG

Kyoto encyclopedia of genes and genomes

K-M

Kaplan-Meier

LASSO

Least absolute shrinkage and selection operator

LD

Linkage disequilibrium

LOOCV

Leave-one-out cross-validation

LRC

Low-risk cohort

LUAD

Lung adenocarcinoma

M6A

N6-methyladenosine

MCRA

Multivariate Cox regression analysis

MF

Molecular function

ML

Machine learning

MR

Mendelian randomization

MRGs

m6A-related genes

MTGs

m6A- and TLS-related genes

NSCLC

Non-small cell lung cancer

OS

Overall survival

OR

Odds ratio

PCA

Principal Component Analysis

PCs

Principal components

percent.mt

Percentage of mitochondrial genes

ROC

Receiver operating characteristic

RSF

Random survival forest

RT-qPCR

Reverse transcription quantitative polymerase chain reaction

ScRNA-seq

Single-cell RNA sequencing

SNP

Single nucleotide polymorphism

ssGSEA

single-sample gene set enrichment analysis

TCGA

The cancer genome atlas

TIDE

Tumor immune dysfunction and exclusion

TLS

Tertiary lymphoid structures

TRGs

TLS-related genes

TMB

Tumor mutational burden

TME

Tumor microenvironment

TNM

Tumor lymph node metastasis

UCRA

Univariate Cox regression analysis

UMAP

Uniform manifold approximation and projection

Author contributions

Shuoming Wu had full access to all the data in the manuscript and takes responsibility for the integrity of the data and the accuracy of the data analysis. All authors read and approved the final manuscript.Conceptualization: Xiangbao Yang, Chengwen Zheng, Yinpeng Pan, and Shuoming Wu.Methodology: Xiangbao Yang, and Shuoming Wu.Software: Xiangbao Yang, and Shuoming Wu.Validation: Xiangbao Yang, and Shuoming Wu.Formal analysis: Xiangbao Yang, Chengwen Zheng, Yinpeng Pan, and Shuoming Wu.Investigation: Xiangbao Yang, and Shuoming Wu.Resources: Xiangbao Yang, and Shuoming Wu.Data Curation: Xiangbao Yang, Chengwen Zheng, Yinpeng Pan, and Shuoming Wu.Writing - Original Draft: Xiangbao Yang, Chengwen Zheng, Yinpeng Pan, and Shuoming Wu.Writing - Review & Editing: Xiangbao Yang, and Shuoming Wu.Visualization: Xiangbao Yang, and Shuoming Wu.Supervision: Xiangbao Yang, and Shuoming Wu.Project administration: Xiangbao Yang, and Shuoming Wu.Funding acquisition: Shuoming Wu.

Funding

This study was supported by Health Science and Technology Project of Lianyungang City (Grant number: 202308). Role of the Funding: The funding had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Data availability

The data supporting the findings of this study are available within the manuscript and its supplementary files. The datasets used and analyzed in this study are available from the following repositories: TCGA (The Cancer Genome Atlas) database: https://portal.gdc.cancer.gov/, accessed on August 1, 2025 (TCGA-LUAD).GEO (Gene Expression Omnibus) database: https://www.ncbi.nlm.nih.gov/geo/, accessed on August 1, 2025 (GSE30219, GSE50081, and GSE189357).Genecards database: https://www.genecards.org/, accessed on August 1, 2025.GWAS (Genome-Wide Association Studies) platform: https://gwas.mrcieu.ac.uk/.TIDE (Tumor Immune Dysfunction and Exclusion) platform: http://tide.dfci.harvard.edu/login/miRanda database: http://www.microrna.org/.miRDB database: http://www.mirdb.org/.SpongeScan database: http://www.targetscan.org/.HPA (Human Protein Atlas) database: https://www.proteinatlas.org/.

Declarations

Ethics approval and consent to participate

This article does not contain any studies with human participants or animals performed by any of the authors.

Informed consent

For this type of study formal consent is not required.

Competing interests

The authors declare no competing interests.