Development of a novel disulfidptosis-correlated m6A/m1A/m5C/m7G gene signature to predict prognosis and therapeutic response for lung adenocarcinoma patients by integrated machine-learning

Abstract

Background

Lung adenocarcinoma (LUAD) represents a significant global health burden, necessitating advanced prognostic tools for improved patient management. RNA modifications (m6A, m1A, m5C, m7G), and disulfidptosis, a novel cell death mechanism, have emerged as promising biomarkers and therapeutic targets in cancer.

Methods

We systematically compiled disulfidptosis-correlated genes and RNA modification-related genes from existing literature. A novel disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore was computed using integrated machine-learning algorithms. Transcriptomic data from TCGA and GEO databases were downloaded analyzed. Single-cell RNA-sequencing data from the TISCH database was processed using the Seurat package. Genes’ protein–protein interaction network was constructed using the String database. Functional phenotype analysis was performed using GSVA, ClusterProfiler, and IOBR packages. Consensus clustering divided patients into two distinct groups. Drug sensitivity predictions were obtained from the GDSC1 database and predicted using the Oncopredict package.

Results

The disulfidptosis-correlated m6A/m1A/m5C/m7G risk score effectively stratified LUAD patients into prognostically distinct groups, demonstrating superior predictive accuracy compared to conventional clinical parameters. Patients in different risk groups exhibited significant molecular and clinical differences. Subsequent analyses identified two molecular subtypes associated with RNA modification and disulfidptosis, revealing differences in immune infiltration and prognosis. Functional enrichment analyses highlighted pathways involving RNA modification and disulfidptosis, underscoring their roles in LUAD pathogenesis. Single-cell analysis revealed distinct features between high- and low-risk status cells.

Conclusion

This study introduces a novel disulfidptosis-correlated m6A/m1A/m5C/m7G risk score as a robust prognostic tool for LUAD, integrating insights from RNA modifications and cell death mechanisms. The risk score enhances prognostic stratification and identifies potential targets for personalized therapeutic strategies in LUAD. This comprehensive approach emphasizes the critical roles of RNA modifications and disulfidptosis in LUAD biology, paving the way for future research and clinical applications aimed at improving patient outcomes.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-024-01530-y.

Introduction

Lung adenocarcinoma (LUAD) is the most prevalent subtype among all types of lung cancer, contributing to both mortality and morbidity globally [1, 2]. Due to LUAD's aggressive nature, fewer than 20% of patients survive more than 5 years after diagnosis [3]. LUAD is currently being treated with targeted therapies and immunotherapies, which are more effective and less harmful than traditional chemotherapy and radiotherapy [4]. Despite immunotherapy and targeted treatments offering advantages to a small percentage of patients, survival rates remain low. Consequently, finding reliable LUAD prognosticators is imperative.

RNA modification is a critical avenue of epigenetic control that significantly influences oncogenesis, progression, and prognosis. With the continuous evolution of RNA sequencing technologies, a multitude of RNA modifications have been unearthed. These include but are not limited to N6-methyladenosine (m6A), 5-methylcytosine (m5C), N1-methyladenosine (m1A), and N7-methylguanosine (m7G) modifications [5]. N6-methyladenosine (m6A), often referred to as the methyl group attached to the sixth nitrogen atom of adenine, represents the most prevalent chemical modification found in RNA transcripts. Various components involved in m6A regulation, including writers, readers, and erasers, have been associated with cancer and are being explored as potential targets for therapeutic interventions [6, 7]. m1A is a significant post-transcriptional RNA modification catalyzed by methyltransferases. Unlike m6A, m1A involves methylation at the N1 position of adenylate. In various cancer cell lines, regulators of m1A have been observed to demethylate tRNAs and generate short RNA fragments derived from tRNAs, thereby promoting enhanced growth of cancer cells [8]. m5C refers to the methylation of the fifth carbon atom of cytosine in RNA. This modification plays a crucial role in regulating mRNA stability, expression, and translation. It is pivotal for processes such as cancer cell proliferation, metastasis, and the development of tumor stem cells [9–11]. The methylation modification m7G occurs on tRNA, contributing significantly to the stability maintenance of tRNA molecules [12, 13]. Recent studies indicate that the regulation of gene expression levels by m6A, m1A, m5C, and m7G modifications is closely linked to tumor progression [14–19]. Recent studies indicate that the global levels of m6A and the expression of its regulatory proteins—comprising writers, erasers, and readers—are frequently altered in various cancers. These dysregulations play a crucial role in cancer initiation, progression, metastasis, and contribute to drug resistance and relapse [20]. Other intracellular processes, including mutations that result in the gain or loss of m6A sites, as well as various extracellular signals, can influence cellular m6A modifications. These alterations are therefore linked to the development of human cancers.

In early 2023, researchers identified a novel type of cell death distinct from ferroptosis and apoptosis, termed disulfidptosis. This phenomenon involves the abnormal buildup of disulfides within cells, triggering a potent stress response that proves highly toxic to the affected cells [21, 22]. The authors revealed that the reduced form of NADPH is crucial for cell survival by mitigating disulfide bonds. Their findings demonstrated that disulfide stress precipitates disulfidptosis, prompting them to propose a novel strategy aimed at targeting disulfides in cancer treatment [23]. Based on the findings, m6A/m1A/m5C/m7G modifications and disulfidptosis play critical roles in both tumor growth and anti-tumor processes. However, their specific roles in LUAD remain underexplored, with limited research examining the correlation between m6A/m1A/m5C/m7G regulators and genes related to disulfidptosis. No study has yet integrated these elements to construct prognostic gene signatures for LUAD.

Given the critical roles of disulfidoptosis and RNA methylation in LUAD, our study focused on developing a genetic signature involving twelve genes that correlate with both disulfidptosis and m6A/m1A/m5C/m7G modifications. Using these genes, we devised a disulfidoptosis-correlated m6A/m1A/m5C/m7G risk score and identified two distinct patient clusters. Significant differences were observed across risk categories and clusters in terms of immune cell infiltration, clinical characteristics, prognosis, and therapeutic responsiveness. These findings suggest that RNA methylation regulators associated with disulfidptosis could be pivotal in shaping therapeutic strategies for LUAD patients. This study provides novel insights and a basis for further exploration into the roles of disulfidoptosis and RNA methylation in LUAD progression, offering potential avenues for future research and treatment advancements.

Methods and materials

Bulk-RNA sequencing data acquired

The TCGA provided bulk tumor transcriptomic data for 497 LUAD patients, along with their clinical information, transcriptomic profiles, and CNV and SNV data, accessible via the TCGA database (https://portal.gdc.cancer.gov/) [24]. External validation was performed using datasets GSE72094, GSE31210, GSE50081, GSE41271, GSE42127, GSE30219, and GSE29016 obtained from GEO (https://www.ncbi.nlm.nih.gov/geo/) [25]. We selected GEO datasets based on the following criteria: 1) The dataset includes complete prognostic information for NSCLC patients. 2) The sample size is sufficiently large to ensure statistical power. 3) The data is from well-characterized cohorts with consistent inclusion criteria. 4) The study provides detailed molecular and clinical data relevant to our analysis. 5) The dataset has been subject to rigorous quality control measures. The detailed information for these datasets (Table S12) and the patients' clinical information (Table S13) are provided in the supplementary tables. Differential expression analysis between patients at different risk group was conducted using the Limma package, defining differentially expressed genes (DEGs) based on criteria of adjusted p-value (Padj) < 0.01 and absolute log2(Fold-change) > 1.

Single-cell RNA-sequencing data acquired

The single-cell RNA-sequencing dataset GSE148071 in h5 format was acquired from the TISCH [26] database and processed by the Seurat package. The raw matrix of scRNA-seq data underwent rigorous filtering using three criteria to ensure high data quality: genes expressed in at least three cells were retained, cells with fewer than 200 or more than 2500 genes were excluded, and cells with over 5% mitochondrial gene content were removed. Subsequently, Seurat's "NormalizeData" function was employed with the normalization method set to "LogNormalize". A Seurat object was then generated from the normalized scRNA-seq data, followed by the identification of the top 2,000 highly variable genes using the "FindVariableFeatures" function. The dataset was further processed by performing principal component analysis (PCA) using the "RunPCA" function to reduce dimensionality. Cell clustering was performed using the "FindNeighbors" and "FindClusters" functions. Subsequently, UMAP was applied, and the resulting UMAP coordinates (UMAP_1 and UMAP_2) were used to visualize cell clustering. Lastly, the AUCell package was utilized to assess the activity of gene sets within the single-cell dataset [27].

Acquisition of disulfidptosis and m6A/m1A/m5C/m7G associated genes

The Writer, Reader, and Eraser genes responsible for m6A, m1A, m5C, and m7G modifications were sourced from literature references [28–31]. Genes regulated by m6A, m1A, m5C, and m7G were compiled and deduplicated (Table S1). Disulfidoptosis-related genes were identified based on Liu et al.'s study (Table S2) [23].

Building a network of interactions between proteins (protein—protein interaction)

We examined the protein–protein interaction network of the 109 m6A/m1A/m5C/m7G regulators associated with disulfidptosis from the STRING database [32].

Development of a disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore

Following the completion of univariate Cox regression analysis with a threshold set at 0.01, we identified 13 disulfidoptosis-correlated m6A/m1A/m5C/m7G genes associated with patient prognosis in LUAD. Utilizing ten machine learning algorithms (GBM, RSF, SuperPC, Survival-SVM, Lasso, Stepwise Cox, Ridge, Enet, CoxBoost, and plsRcox), we developed an integrated disulfidoptosis-correlated m6A/m1A/m5C/m7G riskscore using the RSF + Enet (α = 0.1) combination. After evaluating 113 different permutations, we selected this approach, consistent with our previous method [33]. Subsequently, coefficients for each genes were obtained. Using these gene expressions and coefficients, we computed the disulfidoptosis-correlated m6A/m1A/m5C/m7G riskscore for LUAD patients, categorizing them into high-risk and low-risk groups based on the optimal cut_off value generated by the surv_cutpoint function from the survminer package.

The riskscore is calculated as follows:

$$\text{Riskscore}=\sum \left(\text{coefi}\times exp\text{i}\right)$$

where coefi represents the Elastic Net regression coefficient and expi denotes the expression level of disulfidptosis-related m6A/m1A/m5C/m7G genes.

PCA analysis was conducted using the 'status' R package, while time-dependent ROC curves were generated using 'survminer' and 'timeROC' packages. Additionally, we constructed a nomogram using the 'rms' R package, integrating the riskscore with clinical factors.

Consensus clustering identified m6A/m1A/m5C/m7G related subgroups

After employing the robust ConsensusClusterPlus package [34] for clustering analysis, we successfully delineated two distinct subgroups among LUAD patients, characterized by the genetic profiles of 12 prognostic regulators of m6A, m1A, m5C, and m7G modifications. These subgroups exhibited notable disparities in clinical features and prognostic outcomes.

Functional enrichment analysis

We conducted enrichment analyses, including GO, KEGG, and GSEA, using the R package ClusterProfiler [35] to elucidate the biological pathways associated with the riskscore. Furthermore, GSVA analysis was performed using the GSVA package [36] to delve deeper into the underlying mechanisms implicated.

Quantifying patients' immune infiltration level

We employed seven distinct algorithms to evaluate immune cell infiltration levels in LUAD patients using the TCGA dataset. These algorithms, namely quantTIseq, TIMER, EPIC, MCP-counter, ESTIMATE, and xCell, were implemented using the R package 'IOBR' [37]. Additionally, ssGSEA from the GSVA package was utilized. Furthermore, we investigated the correlation between the expression of immune-related molecules and the riskscore.

Cellchat analysis

With the 'CellChat' package [38], we examined intercellular interactions within the tumor microenvironment (TME), identifying diverse ligand–receptor pairs that mediate communication between different cell types.

Finding potential drugs targeting the riskscore

We integrated data from the GDSC1 database (https://www.cancerrxgene.org/ [39]) with analyses using the 'oncoPredict' package [40] to assess the responsiveness of LUAD samples to various therapeutics, based on their IC50 values. A lower IC50 value indicates higher sensitivity to the treatment.

Statistic analysis

R (version 4.1.1) was utilized for all statistical analyses. To compare two groups, either the Wilcoxon test or t-test was employed as appropriate. Spearman correlation coefficient was used for correlation analyses. Kaplan–Meier analysis was employed to assess differences in overall survival between low- and high-risk categories. We evaluated the predictive significance of the disulfidoptosis-correlated m6A/m1A/m5C/m7G riskscore and clinicopathological features using multivariate and univariate Cox regression analyses. Statistical significance was denoted as * for p < 0.05, ** for p < 0.01, and *** for p < 0.001; Ns indicated p ≥ 0.05.

Results

Identification of hub disulfidptosis-correlated m6A/m1A/m5C/m7G genes

To initiate our study, we performed correlation analyses on selected m6A/m1A/m5C/m7G genes that are associated with disulfidptosis. Our selection criteria stipulated that these genes must exhibit a significant relationship (|r|> 0.1 and P < 0.05) with at least one disulfidptosis-related gene. Under these criteria, we identified 109 disulfidptosis-related m6A/m1A/m5C/m7G genes, which were subsequently chosen for further investigation (Fig. 1A, Table S3). The protein–protein interaction (PPI) network analysis revealed a closely interconnected relationship among these 109 genes (Fig. 1B). Furthermore, Gene Ontology (GO) analysis highlighted that these 109 genes are involved in functions closely associated with RNA modification, including 'translational initiation', 'translation regulator activity', 'nucleic acid binding', 'RNA methyltransferase activity', and 'RNA methylation' (Fig. 1C).

Fig. 1.

Identification of hub disulfidptosis-correlated m6A/m1A/m5C/m7G genes. A Correlation analysis to selected disulfidptosis-correlated m6A/m1A/m5C/m7G genes. B 109 genes’ PPI network. C GO analysis. D 13 genes’ differential expression between normal and tumor tissues. E 13 hub m6A/m1A/m5C/m7G genes significantly associated with disulfidptosis. F 13 genes’ gene mutation frequency. G 13 genes’ copy number variation. H Disulfidptosis-related genes’ relationship with immune cells. I 13 disulfidptosis-correlated m6A/m1A/m5C/m7G genes’ relationship with immune cells

To identify hub genes among these 109 candidates, we conducted univariate Cox regression analysis using the TCGA dataset. Genes with a p-value less than 0.01 were designated as hub genes, resulting in the identification of 13 genes that exhibited consistent prognostic significance. This 13 genes all exhibited strong power to stratify LUAD patients into different prognostic groups (Supplementary Fig. 1). Apart from IGFBP1, these 13 genes showed significant differential expression between normal and tumor tissues (Fig. 1D). Moreover, correlation analysis demonstrated that all 13 m6A/m1A/m5C/m7G genes were tightly associated with disulfidptosis-related genes, underscoring our success in identifying disulfidptosis-correlated m6A/m1A/m5C/m7G genes.

In the TCGA cohort, these 13 hub genes exhibited varying levels of mutational activity, primarily characterized by missense mutations, with an overall mutation frequency of 14.18% (Fig. 1F). Notably, FLI1 displayed the highest mutation frequency. Additionally, we observed varying degrees of DNA copy number variation (CNV) among these 13 hub genes, with most genes showing diverse CNV patterns, whereas EIF4G1 exhibited the highest CNV amplifications (Fig. 1G).

Furthermore, we analyzed the relationship of these genes with immune cell infiltration. Our findings demonstrated significant associations between disulfidptosis-related genes (Fig. 1H), these 13 hub genes (Fig. 1I), and various types of immune cells, suggesting their close involvement in the tumor microenvironment.

Developing the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore by machine-learning

We initiated the development of a disulfidptosis-correlated m6A/m1A/m5C/m7G risk score using a machine learning-driven approach based on 13 prognostic genes. Utilizing the TCGA dataset as the training set, we constructed 113 prediction models and evaluated their performance across seven independent validation sets (GSE72094, GSE50081, GSE41271, GSE42127, GSE30219, GSE31210, and GSE29016). To ensure consistent predictive accuracy across all datasets, we opted for the 'RSF + Enet (α = 0.1)' composition, which yielded a model with an average C-Index of 0.633 across all eight datasets (Fig. 2A). The Random Survival Forests (RSF) algorithm ranked the relative importance of these 13 genes (Supplementary Fig. 2A). Subsequently, the Elastic Net (Enet) algorithm reduced the gene count from 13 to 12 (Supplementary Fig. 2B), excluding the gene 'GFI1B,' and assigned coefficients to each of the remaining genes in the final model (Supplementary Fig. 2C). This Enet model comprised the selected 12 genes, integrating their respective coefficients to enhance predictive performance.

$$\begin{array}{cc}\hfill \mathrm{Riskscore}& =\mathrm{CCNB}1_{}^{\ast }0.11920576+\mathrm{EIF}4\mathrm{G}1_{}^{\ast }0.07908695\hfill \\ \hfill & -\mathrm{FLI}1_{}^{\ast }0.00998827+\mathrm{FOXM}1_{}^{\ast }0.02352144\hfill \\ \hfill & -\mathrm{GATA}2_{}^{\ast }0.07692774+\mathrm{HNRNPA}2\mathrm{B}1_{}^{\ast }0.05213371\hfill \\ \hfill & +\mathrm{IGF}2\mathrm{BP}3_{}^{\ast }0.04259547+\mathrm{IGFBP}1_{}^{\ast }0.14250547\hfill \\ \hfill & -\mathrm{NSUN}6_{}^{\ast }0.07173466-\mathrm{TRDMT}1_{}^{\ast }0.09440691\hfill \\ \hfill & +\mathrm{WDR}4_{}^{\ast }0.05823039-\mathrm{YTHDF}2_{}^{\ast }0.03467579\hfill \\ \hfill \end{array}$$

Fig. 2.

Developing the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore by integrated machine-learning. A A comprehensive evaluation was conducted across 8 independent cohorts, resulting in the construction of 113 distinct models based on C-Index assessments. B–I Comparative analyses of prognostic variations, principal component analysis (PCA), and time-ROC analysis were performed to distinguish between high- and low-risk groups in the following datasets: TCGA (B), GSE72094 (C), GSE50081 (D), GSE41271 (E), GSE42127 (F), GSE30219 (G), GSE31210 (H), and GSE29016 (I)

The optimal cut_off value generated by the surv_cutpoint function from the survminer package was employed to stratify patients into two distinct groups based on their prognostic outlook. Patients classified in the high-risk group demonstrated significantly poorer prognoses compared to those in the low-risk group, not only within the TCGA training set (Fig. 2B) but also across seven external validation cohorts: GSE72094 (Fig. 2C), GSE50081 (Fig. 2D), GSE42127 (Fig. 2E), GSE41271 (Fig. 2F), GSE31210 (Fig. 2G), GSE30219 (Fig. 2H), and GSE29016 (Fig. 2I). Additionally, PCA highlighted discernible distinctions between individuals in high- and low-risk groups across all datasets. Time-ROC curves further underscored the robust predictive capabilities of the riskscore in forecasting patients' prognosis, achieving high AUC values consistently across Fig. 2B-I.

Dividing LUAD patients into two subgroups based on 12 model genes

We utilized unsupervised clustering of the 12 disulfidptosis-correlated m6A/m1A/m5C/m7G model genes to unveil previously unidentified subtypes related to disulfidptosis and RNA-methy;ation in LUAD. Optimal clustering with k = 2 revealed distinct groups, clearly delineating LUAD patients into two categories (Fig. 3A). Within the TCGA cohort, these 12 genes showed differential expression patterns across the identified clusters (Fig. 3B). Principal component analysis (PCA) further emphasized significant distinctions between these clusters (Fig. 3C).

Fig. 3.

Consensus clustering identified two clusters based on 12 prognostic related genes. A Based on the consensus clustering matrix (k = 2), LUAD patients were divided into two clusters. B Expression of disulfidptosis-correlated m6A/m1A/m5C/m7G genes between cluster1 and cluster2. C PCA analysis. D Patients in cluster II had worse prognosis. E Patients in cluster I had more benign clinical features. F–I Comparison of the immune infiltration level between cluster I and cluster II based on the ssGSEA (F), EPIC (G), TIMER (H), and ESTIMATE (I) algorithms

Importantly, patients in Cluster I exhibited significantly better prognoses compared to those in Cluster II (Fig. 3D), whereas Cluster II patients displayed more advanced clinical characteristics (Fig. 3E). Analysis using Single-sample Gene Set Enrichment Analysis (ssGSEA, Fig. 3F), EPIC (Fig. 3G), TIMER (Fig. 3H), and ESTIMATE (Fig. 3I) underscored that patients in Cluster I had higher levels of immune cell infiltration and immune-related scores, indicative of an immune-active status.

The riskscore significantly associated with LUAD patients’ clinical characteristics

The heatmap illustrates the transcriptional profiles of the 12 disulfidptosis-correlated m6A/m1A/m5C/m7G model genes within the TCGA cohort (Fig. 4A). We observed significant differences in the expression of 11 genes between high- and low-risk groups, with YTHDF2 being the exception (Fig. 4B). Furthermore, we explored the prognostic significance of the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore across different clinical subgroups. Our analysis revealed that the riskscore consistently stratified LUAD patients into distinct prognostic groups in different clinical subgroups across various cohorts (Fig. 4C). Additionally, the riskscore showed significant associations with clinical parameters such as clinical stage, T stage, N stage, and disease relapse across these cohorts (Fig. 4D). These findings highlight the potential of the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore as a robust prognostic biomarker in LUAD.

Fig. 4.

The association between clinical features and the riskscore. (A, B) 12 model genes’ different expression among high- and low- risk groups illustrated by the heapmap (A) and the group comparison map (B). (C) The disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore exhibited prognostic value at different subgroups across various groups. (D) The disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore significantly associated LUAD patients’ clinical features

Nomogram’s development and validation

Considering the importance of age, gender, and clinical stage as key indicators for assessing prognosis in NSCLC patients, we hypothesized that integrating these factors with the riskscore could improve prognostic predictions for LUAD patients. To achieve this, we developed a nomogram that combines the riskscore with clinical variables such as age, gender, and TNM stage. In the TCGA cohort, the nomogram achieved a C-Index of 0.704, indicating its strong predictive accuracy. Additionally, calibration plot analysis (Fig. 5B) and decision curve analysis (DCA) (Fig. 5C) confirmed the reliability of the nomogram in predicting patients' prognosis. Importantly, ROC plots demonstrated that the nomogram outperformed individual predictors such as the riskscore, age, gender, and TNM stage in predicting survival outcomes for LUAD patients (Fig. 5D-F). These findings underscore the potential of the integrated nomogram as a robust tool for clinical decision-making and prognostic assessment in LUAD.

Fig. 5.

Construction of a nomogram. A–D The nomogram model (A) and the calibration curve (B), the DCA curve (C), and the ROC curve (D–F) to evaluate its efficacy

Comparison the predictive efficacy of the riskscore with existing clinical features

Cox regression analyses were conducted to assess whether the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore serves as a standalone prognostic indicator for patients with LUAD. Independently of other clinical factors, the riskscore demonstrated significant prognostic value across multiple cohorts, including TCGA (Fig. 6A, B), GSE72094 (Table S5), GSE31210 (Table S6), GSE30219 (Table S7), GSE41271 (Table S8), GSE42127 (Table S9), GSE50081 (Table S10), and GSE29016 (Table S1).These findings establish the riskscore as a robust and independent predictor for individuals diagnosed with LUAD.

Fig. 6.

The disulfidptosis-correlated m6A/m1A/m5C/m7G riskscor is more powerful than existing clinical markers. A, B Uni- (A) and multi- (B) variable cox regression performed in the TCGA cohort. C, D Comparing the AUC value (C) and the C-Index (D) of the riskscore with existing clinical characteristics

To evaluate the predictive effectiveness of the riskscore relative to traditional clinical variables in LUAD patients, we conducted an analysis comparing AUC values (Fig. 6B) and the C-Index (Fig. 6C) for each factor. Remarkably, both analyses consistently demonstrated that the riskscore outperformed most clinical markers in terms of predictive performance. However, in some cohorts, it was slightly less powerful than the TNM Stage, underscoring its overall excellent efficiency.

Functional enrichment analysis

To elucidate the underlying mechanisms contributing to the excellent predictive ability of the riskscore, we conducted further studies. Initially, using the limma package, we identified genes with differential expression between high- and low- risk groups (Fig. 7A, Table S4). Among these, MYBL2 and UBE2C emerged as the most highly expressed genes in the high-risk group, correlating with poorer survival outcomes in LUAD patients (Fig. 7B). Conversely, SCGB3A2 and SFTPB, which were highly expressed in the low-risk group, were associated with better survival outcomes (Fig. 7C). This differential gene expression pattern indicates that high-risk factors for LUAD are characterized by elevated expression in the high-risk group, whereas protective factors exhibit heightened expression in the low-risk group, reaffirming the riskscore's reliability as a prognostic indicator for LUAD.

Fig. 7.

Finding the potential mechanism behind the riskscore’s precision predict ability. A Identification of differentially expressed genes between high- and low-risk groups. B, C Genes highly expressed in the high-risk group as risk factors for LUAD (B), and genes highly expressed in the low-risk group as protective factors for LUAD (C). D, E GSVA analysis highlighting gene sets exhibiting elevated activity in the high-risk (D) and low-risk (E) groups. F, G GSEA analysis revealing enrichment of malignant phenotypes in the high-risk group (F) and enrichment of benign phenotypes in the low-risk group (G)

Furthermore, GSVA analysis corroborated these findings by revealing that gene sets linked to aggressive features such as 'lung cancer poor survival', 'metastasis up', and 'epithelial mesenchymal transition up' exhibited increased activity in the high-risk group (Fig. 7F). In contrast, gene sets associated with favorable phenotypes such as 'lung cancer good survival', 'metastasis down', and 'epithelial mesenchymal transition down' showed higher activity in the low-risk group (Fig. 7E). Additionally, results from GSEA analysis supported these observations, demonstrating enrichment of gene sets representing malignant phenotypes in the high-risk group (Fig. 7F) and enrichment of gene sets representing benign phenotypes in the low-risk group (Fig. 7G).

These comprehensive analyses underscore the biological relevance of the riskscore in predicting LUAD prognosis, highlighting its association with critical molecular pathways and phenotypic characteristics of the disease.

The disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore affects LUAD patients’ tumor immune microenvironment (TME)

In LUAD patients across different risk categories, notable differences in immunocyte infiltration were observed (Fig. 8A). Additionally, a variety of immune regulators, including chemokines, receptors, immunostimulators, immunoinhibitors, and MHC molecules, exhibited significant differential expression between high- and low-risk groups (Fig. 8B). Notably, the riskscore demonstrated predominantly negative associations with most immune cells and immune-related molecules (Fig. 8C). Interestingly, the riskscore showed significant positive correlations with both Tumor Mutation Burden (TMB, Fig. 8D) and MicroSatellite Instability (MSI, Fig. 8E), suggesting its association with malignancy.

Fig. 8.

Immune infiltration analysis to explore the riskscore’s relationship with the TME. A, B Heatmaps were plotted to show the differential expression of immune cells (A) and immune-related molecules’ expression (B) between high- and low- risk groups. C The riskscore negatively correlated with most kinds of immunocytes and immune-related molecules. D, E The riskscore positively correlated with TMB (D) and MSI (E)

Single-cell RNA-sequencing analysis

In addition to analyzing bulk-RNA sequencing data, we investigated single-cell sequencing data from GSE148071 to gain insights into the distribution and significance of the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore at a cellular level. From this dataset, we identified 35 distinct cell clusters (Fig. 9A) and classified them into 10 cell populations using meta-data from the TISCH database (Fig. 9B). The expression profiles of the 12 disulfidptosis-correlated m6A/m1A/m5C/m7G genes were depicted in Fig. 9C, and we calculated the riskscore for each cell in the GSE148071 dataset using the same methodology as in the bulk dataset. The distribution of riskscore across different cell types revealed that proliferating T cells and malignant cells exhibited the highest riskscore (Fig. 9D). Subsequently, we stratified each cell population into High-riskscore and Low-riskscore groups based on their mean riskscore (Fig. 9E). We observed significant differences in the proportion of cells between these groups, with the high-riskscore cluster showing a notable increase in malignant cell proportion (Fig. 9F). Analysis of cell–cell communication patterns indicated that cells with high riskscore exhibited more frequent and stronger interactions compared to those with low riskscore (Fig. 9G, H). Furthermore, most signaling pathways showed higher activity in the high-riskscore group, except for four specific pathways (Fig. 9I).

Fig. 9.

Single-cell RNA-sequencing analysis. A Identification of 35 cell clusters. B Annotating 10 types of cells. C 2 disulfidptosis-correlated m6A/m1A/m5C/m7G genes’ expression at single-cell level. D The riskscore’s distribution across different cells. E Split all cells into high- and low- risk groups. F High-risk group had higher malignant cells’ proportion. G, H Comparison of the number and intensity of interaction between high- and low- riskscores’ cells using the bar plot (G) and the network diagram (H). (I) Most signaling had higher intensity and activity in high-risk group

In summary, exploring the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore at the single-cell level provided deeper insights into its biological significance.

Comparing the genomics difference between high- and low- risk groups

We also conducted comprehensive genomics analyses to compare the genetic landscape between LUAD patients classified into high-risk and low-risk groups based on the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore. Initially, our focus was on the top 20 genes with the highest mutation frequencies, visually illustrating the disparities between these groups (Fig. 10A, B). Subsequent examination revealed that the top 22 genes with the most significant differences in mutation frequency between high- and low-risk groups exhibited higher frequencies in the former (Fig. 10C). Furthermore, we identified co-mutation relationships among these genes (Fig. 10D). Additionally, the riskscore demonstrated a strong positive correlation with various types of gene mutations, including silent mutations, nonsilent mutations, SNV neoantigens, and the number of genomic segments affected (Fig. 10E).

Fig. 10.

Comparing the genomics difference between high- and low- risk groups. A, B Comparative analysis of somatic mutation frequencies in high-risk (A) and low-risk (B) patient cohorts. C, D Identification of the top 22 genes with significant differences in mutation frequencies between the high- and low-risk groups (C), along with their co-occurrence relationships (D). E Positive correlation of the riskscore with various types of mutations. F, G Analysis of the top 12 copy number variations (CNVs) in the high-risk (F) and low-risk (G) groups. H, I Visualization of patients' G-scores using chromosomal plots in the high-risk (H) and low-risk (I) groups

Moreover, our analysis uncovered notable differences in copy number variation (CNV) events between the two risk groups (Fig. 10F, G). Patients in the high-risk group displayed a higher frequency and more complex array of CNV events, whereas those in the low-risk group exhibited fewer and less intricate CNV events. Chromosomal plots further illustrated that patients in the high-risk group had higher G-scores compared to those in the low-risk group (Fig. 10H, I), indicating a predisposition towards malignant features among high-risk LUAD patients.

In summary, these genomic analyses provide detailed insights into the molecular differences associated with the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore, highlighting its relevance in understanding the genetic landscape and potential clinical implications in LUAD.

Finding potential drugs targeting the riskscore

Based on drug information from GDSC1, we investigated the association between the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore and sensitivity to targeted and chemotherapy agents. Our analysis revealed that the IC50 values of 14 commonly used therapeutic drugs in clinical practice were lower in patients with high riskscore, suggesting heightened sensitivity to these drugs. The therapeutic agents included Pemetrexed, Paclitaxel, Vinorelbine, Etoposide, Cetuximab, Axitinib, Bosutinib, Nilotinib, Vinblastine, Gemcitabine, Cisplatin, Docetaxel, Methotrexate, and Gefitinib (Fig. 11A).

Fig. 11.

Patients in high-risk group may more sensitive to chemotherapy and targeted therapy drugs. A 14 drugs’ IC50 value between high- and low- risk groups. B The riskscore’s relationship with drugs’ IC50 value

Furthermore, correlation analysis indicated a negative correlation between the riskscore and the IC50 values of these drugs (Fig. 11B). These findings provide a foundation for tailoring clinical medication strategies based on the riskscore of individual patients.

Discussion

Lung adenocarcinoma (LUAD) is a formidable malignancy with significant global impact on public health. Despite extensive efforts, effective treatment of LUAD remains a considerable challenge. Recent advancements in high-throughput sequencing technology have led to the discovery of numerous prognostic markers. Among these, RNA methylations represent crucial RNA post-transcriptional modifications intricately linked to the initiation and progression of various cancers [41, 42]. METTL3 acts as an oncogene in lung cancer through various mechanisms. Lin et al. discovered that METTL3 enhances RNA translation independently of methyltransferase and reader protein activity. It boosts RNA translation by directly attracting translation initiation factors. Additionally, METTL3 promotes the growth, survival, and invasion of lung adenocarcinoma cells by upregulating EGFR and TAZ [43]. Li et al. demonstrated that FTO enhances the proliferation of NSCLC by elevating USP7 expression (44). Liu et al. further showed that FTO overexpression downregulates m6A levels in MZF1 mRNA transcripts, which stabilizes the mRNA and increases MZF1 expression, thereby promoting the proliferation and invasion of lung squamous cell carcinoma cells [45]. Additionally, the m6A demethylase ALKBH5 has been shown to inhibit tumor growth and metastasis in NSCLC patients by decreasing YTHDFs-mediated YAP expression, although some studies suggest that ALKBH5 may also facilitate the progression of NSCLC [46, 47]. Programmed cell death mechanisms, such as apoptosis and autophagy, facilitate coordinated cellular demise, thereby contributing to the overall benefit of living organisms [48, 49]. Recently, Liu et al. introduced a novel type of cell death termed disulfidptosis. The primary mechanism of disulfidptosis involves the accumulation of intracellular disulfides due to glucose starvation in SLC7A11high cells, which subsequently bond with actin cytoskeleton proteins [23, 50]. The connection between disulfidptosis and RNA methylation, as well as their specific roles in LUAD, remains largely unexplored in current research. To address this gap, our study aims to comprehensively investigate disulfidptosis-correlated m6A/m1A/m5C/m7G genes in LUAD. We will begin by identifying and categorizing genes that are associated with both disulfidptosis and RNA methylation modifications. Subsequently, we will explore the interactions among these genes and their biological functions within the context of LUAD. This research seeks to establish a novel understanding of how these molecular processes intersect and potentially contribute to the prognosis and treatment of LUAD.

The disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore, derived through integrated machine-learning methods (Final selection: RSF + Enet (α = 0.1)), has shown significant efficacy in predicting outcomes for LUAD patients. Notably, PCA analysis highlighted distinct differences between high and low-risk groups. Subsequently, patients were stratified into two clusters based on 12 prognostic-related genes used in constructing the riskscore, revealing varied gene expression patterns, prognostic outcomes, and immunoinfiltration landscapes. Further investigation explored the prognostic mechanisms of the riskscore, including differences in immune cell infiltration, expression of immune-related molecules, Tumor Mutation Burden (TMB), and MicroSatellite Instability (MSI) among high and low riskscore patients. Nomogram construction illustrated enhanced prognostic accuracy when integrating the riskscore with clinical parameters, offering a novel predictive approach for LUAD prognosis. Additionally, analysis of single-cell sequencing data identified predominant expression of the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore in proliferating T cells and malignant cells, with the high-risk group exhibiting a higher proportion of malignant cells and lower immune cell presence. Cross-validation between bulk-RNA and single-cell RNA sequencing confirmed the riskscore as a significant risk factor in LUAD, providing deeper insights. Exploration of GDSC1 database data on targeted therapy drugs revealed lower IC50 values for most targeted therapy and chemotherapy drugs in the high-risk group, suggesting potential benefits from these treatments for patients with high riskscore in LUAD.

The disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore comprises 12 genes: CCNB1, EIF4G1, FLI1, FOXM1, GATA2, HNRNPA2B1, IGF2BP3, IGFBP1, NSUN6, TRDMT1, WDR4, and YTHDF2. Among the 13 genes, CCNB1, EIF4G1, FOXM1, HNRNPA2B1, IGF2BP3, and WDR4 show significantly elevated expression in LUAD tissue compared to normal lung tissue, and the high expression of these genes is associated with poorer prognosis in lung adenocarcinoma patients. This suggests that these genes may be involved in the development of lung adenocarcinoma and influence its progression, leading to worse outcomes for LUAD patients. Conversely, FLI1, TRDM1, and GATA2 exhibit significantly decreased expression in LUAD tissue compared to normal lung tissue, and reduced expression of these genes correlates with worse prognosis in lung adenocarcinoma patients. This implies that increasing the expression of these genes could potentially inhibit the onset of lung adenocarcinoma and improve patient prognosis, indicating their protective role. For IGFBP1, there is no significant difference in expression between normal lung tissue and LUAD tissue, but its high expression is associated with poorer prognosis in LUAD patients, suggesting it may not play a role in the onset of LUAD but significantly impacts its progression. Interestingly, NSUN6 and YTHDF2 show increased expression in LUAD tissue, yet high expression of these two genes appears to improve prognosis for LUAD patients. This indicates their complex pathogenic mechanisms, warranting further in-depth research. CCNB1 has been identified as a risk factor in LUAD. Previous studies have demonstrated that mir-139-5p, which targets CCNB1, inhibits malignant phenotypes such as migration and invasion in LUAD cells [51]. EIF4G1, a crucial component of the EIF4F transcription initiation factor complex, is notably upregulated in lung cancer. Its elevated expression has been linked to low differentiation of lung cancer cells and poor overall survival among patients with non-small cell lung cancer. Functionally, EIF4G1 promotes the G1/S transition of the cell cycle and enhances tumor cell proliferation in NSCLC. Mechanistically, EIF4G1 regulates both the expression and phosphorylation of mTOR, thereby facilitating its role in promoting tumorigenesis [52]. Numerous studies have highlighted FOXM1 as a pivotal factor in lung cancer, emphasizing its detrimental impact on patient outcomes [53, 54]. Additionally, Xiu et al. demonstrated that FOXM1 contributes to radioresistance in lung cancer by upregulating KIF20A [55], while Madhi et al. indicated that inhibiting FOXM1 could enhance the efficacy of immunotherapy in lung cancer [56], underscoring FOXM1 as a promising therapeutic target. Conversely, GATA2 emerged from our research as a protective factor in LUAD (HR < 1), consistent with previous findings showing its role in suppressing cell proliferation, invasion, migration, and epithelial-mesenchymal transition (EMT) in lung cancer cell lines like A549 and H460 [57]. On the other hand, HNRNPA2B1 was identified as a risk factor, with multiple studies confirming its promotion of malignant phenotypes in LUAD cells [58, 59]. Research by Li et al. revealed that HNRNPA2B1 modulates the m6A modification of the lncRNA MEG3, suggesting its significant involvement in RNA modifications that impact tumorigenesis [60]. IGF2BP3, another risk factor for LUAD, has been found to confer resistance to ferroptosis—a regulated cell death process—through its m6A reading domain, binding to m6A-methylated mRNAs encoding anti-ferroptotic factors [61]. NSUN6, however, acts as a suppressor of lung cancer progression by regulating NM23-H1 expression via m5C modification in the 3'-UTR of NM23-H1 mRNA, thereby inhibiting cancer cell proliferation, migration, and EMT [62]. These results suggesting RNA modification is an important mechanism which affects tumor progression. WDR4 has been implicated in promoting lung cancer progression through m7G tRNA modifications and enhancing mRNA translation efficiency [13]. YTHDF2, known for its role in regulating the immune microenvironment, also contributes to lung cancer development [63, 64]. In contrast, the specific role of TRDMT1 in lung adenocarcinoma remains largely unexplored, presenting a gap in current knowledge that warrants further investigation..

With the advent of cancer immunotherapy, there is newfound optimism for individuals battling cancer, despite the persistent challenge of evading the immune system (65, 66). Our research reveals significant differences in immune cell infiltration levels and the expression of various regulators within the tumor microenvironment (TME) across different risk categories. Notably, most immune cells and TME regulators show a negative correlation with the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore, with higher expression or infiltration observed in the low-risk group. Individuals with lower riskscores exhibit an 'immune hot' phenotype, characterized by robust immune cell infiltration. This finding underscores the potential of the riskscore in predicting patient prognosis and suggests avenues for developing targeted therapies for lung cancer. Understanding these relationships sheds light on the riskscore's role in tumor immune evasion mechanisms and provides a foundation for future therapeutic strategies aimed at enhancing immune responses against lung cancer.

While our study demonstrates the promising prognostic and predictive capabilities of the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore in LUAD patients, there are several limitations that warrant consideration. Firstly, all data utilized in this research were sourced from publicly available databases, and the riskscore's effectiveness has not been validated using our own cohort of subjects. Secondly, the genes comprising the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore are established RNA methylation-related genes, and no novel genes related to RNA methylation or disulfidptosis were discovered. Furthermore, not all 12 genes included in the riskscore have been definitively linked to disulfidptosis, necessitating further investigation to confirm their involvement.

In our future studies, we aim to validate the efficacy of the disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore in larger patient cohorts. Additionally, we plan to explore the interplay between the riskscore, RNA methylation regulators, and the disulfidptosis process.

In summary, our research introduces a novel disulfidptosis-correlated m6A/m1A/m5C/m7G riskscore that effectively predicts prognosis and therapy response in LUAD patients. It opens avenues for future investigations into genes associated with disulfidptosis and RNA methylation modifications, offering insights into integrating phenotype-related genes for constructing robust prognostic models.

Author contributions

FC Lai conceived, designed and supervised the study. BL Xu and LY Zhang collected, analyzed the data and wrote the manuscript. LJ Lin and YF Lin revised and reviewed the manuscript. All authors read and approved the final manuscript and consent for publication.

Data availability

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/supplementary material.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the First Affiliated Hospital of Fujian Medical University, and this article does not contain any studies with human participants or animals.

Consent for publication

The content of this manuscript has not been previously published and is not under consideration for publication elsewhere. All authors consent for publication in this journal.

Competing interests

The authors declare no competing interests.