Identifying ferroptosis-related genes in lung adenocarcinoma using random walk with restart in the PPI network

Abstract

Lung adenocarcinoma (LUAD), the most common non-small cell lung cancer subtype, often presents with subtle early symptoms leading to delayed diagnosis. Ferroptosis, a cell death process associated with iron metabolism dysregulation, has been linked to cancer onset, progression, and treatment resistance. Thus, identifying ferroptosis-related genes may offer novel insights for LUAD therapy. In this study, we employed the random walk with restart (RWR) algorithm on the LUAD protein-protein interaction (PPI) network to predict ferroptosis-related target genes. Gene set enrichment analysis (GSEA) explored the relationship between XBP1 and ferroptosis, while tumor microenvironment analysis evaluated the correlation between XBP1 expression and immune cell infiltration. External cohorts validation was performed using the GSE118370, GSE68465, and TCGA-LUAD datasets. Our analysis identified XBP1 as a potential ferroptosis-related gene in LUAD. GSEA confirmed a strong association between XBP1 and the ferroptosis process, along with its involvement in the tumor microenvironment, and external cohorts demonstrated its high expression and significant correlation with immune cell infiltration in LUAD tissues. These findings suggest that XBP1 plays a key role in LUAD development and progression, providing new perspectives for precision therapies targeting the ferroptosis pathway.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-14307-2.

Introduction

Lung adenocarcinoma (LUAD) is the most common subtype of non-small cell lung cancer (NSCLC), comprising approximately 40% of all lung cancer cases^1,2. Despite declining smoking rates, LUAD incidence has been steadily increasing worldwide, particularly among younger populations in smoke-free environments³. Due to its subtle early symptoms, patients are often diagnosed at later stages, resulting in poor treatment outcomes and a low five-year survival rate⁴. Consequently, in-depth studies on the etiology and pathogenic mechanisms of LUAD, as well as the identification of new molecular targets and biomarkers, have become a key focus in the field of chronic respiratory diseases. This research contributes to understanding the complex mechanisms underlying tumorigenesis and progression and provides a theoretical foundation and practical guidance for the early diagnosis, precise treatment, and clinical prognosis assessment of LUAD.

In recent years, ferroptosis, a newly discovered form of programmed cell death, has attracted significant attention^5–7. It is characterized by the accumulation of intracellular iron ions and lipid peroxides, leading to cell membrane damage and cell death. Growing evidence indicates that ferroptosis plays a crucial role in the initiation and progression of various cancers, particularly LUAD, where the ferroptosis mechanism has been found to be closely associated with tumor proliferation, metastasis, and drug resistance^8–11. Ferroptosis-associated key molecules and pathways, such as the system Xc- reverse transport channel, glutathione peroxidase 4 (GPX4), and iron metabolism regulatory factors, have become major research focuses for exploring LUAD’s underlying mechanisms^12–14. In particular, the interplay between dysregulated iron metabolism and the tumor microenvironment not only promotes tumor cell growth but also influences immune evasion mechanisms, offering new therapeutic targets for LUAD treatment¹⁵. Therefore, investigating the role of ferroptosis in LUAD helps uncover new treatment strategies while also providing a theoretical foundation for the development of ferroptosis-related drugs.

Although experimental methods have provided valuable data for tumor research, they are typically time-consuming and resource-intensive, particularly when constructing gene expression maps and screening potential pathogenic genes, which are expensive and difficult to perform efficiently^16,17. In order to address these challenges, machine learning (ML) algorithms have gained widespread application in cancer research, especially in genetic screening and biomarker discovery. ML techniques, such as random forest, elastic net regression, and random walk with restart (RWR), can efficiently process complex genomic data, identify potential pathogenic genes and pathways, and are not limited by the time and cost constraints of traditional methods^18–20. These techniques have been successfully applied to the study of genetic mechanisms in various diseases, including cancer, cardiovascular diseases, and chronic lung diseases^19,21,22. These methods can reveal the potential relationships between genes and diseases, offering new perspectives, particularly in identifying the pathogenic mechanisms of complex diseases such as LUAD, demonstrating significant potential.

Building on the above background, this study integrates ML algorithm with protein-protein interaction (PPI) networks and employs the RWR algorithm to identify potential ferroptosis-related target genes in LUAD. By selecting the subgraph with the highest interaction density within the PPI network as the seed nodes, the RWR algorithm predicts multiple candidate genes associated with ferroptosis. Further cross-referencing with differentially expressed genes from the GSE118370 dataset led to the identification of XBP1 as a promising candidate gene. Gene set enrichment analysis (GSEA) revealed a significant association between XBP1 and ferroptosis-related pathways in LUAD, suggesting its potential role in the onset and progression of the disease. Additionally, tumor microenvironment analysis highlighted the potential regulatory role of XBP1 in immune responses, supporting its potential as a biomarker for LUAD. To validate these predictions, we confirmed the high expression of XBP1 in tumor cells through three independent external datasets. These findings provide new molecular biomarkers for the early diagnosis of LUAD and offer a theoretical basis for developing targeted therapeutic strategies for ferroptosis.

Results

Collection of LUAD-related targets and ferroptosis-associated genes

Based on the comprehensive set of known associations between human diseases and genes, 10,902 disease targets were identified in the GeneCards database, with a relevance score of ≥ 1 as the screening criterion. To address the potential issue of non-specificity associated with GeneCards, we further integrated disease targets from the DisGeNET (688 genes) and OMIM (85 genes) databases, both of which provide curated, high-confidence gene-disease associations. Targets from all three databases were merged, and duplicates were removed, resulting in 11,514 LUAD-related targets. In the FerrDb V2 database, 484 ferroptosis-related genes were retrieved.

Construction of the LUAD PPI network and subnetwork identification

The LUAD-related PPI network was constructed using the STRING plugin in Cytoscape. After removing isolated nodes, the PPI network contained 8778 nodes and 296,670 edges (Supplementary Table S1). To further simplify the complexity of the PPI network and reduce redundant information, the network analysis tool MCODE was employed to cluster the dense PPI network, selecting the top 1 cluster for subsequent RWR analysis. The MCODE algorithm evaluates the degree of node aggregation by calculating factors such as node density, K-core value, and adjacent node subgraphs. The score value of each node reflects the density of that node and its surrounding nodes. The larger the score value, the more significant the node in the network. Based on these score values, MCODE begins with the densest node and uses the getClusterCore method to expand from this seed node, sequentially adding adjacent nodes that satisfy the criteria, ultimately forming a functional module. This approach effectively reduces network complexity and helps identify key proteins associated with the disease, providing biologically meaningful candidate genes for further research. The top 1 cluster contains 142 nodes and 8944 edges, as shown in (Fig. 1).

Fig. 1.

PPI network constructed by MCODE clustering of the top 1 cluster.

Prediction of ferroptosis-associated genes Using RWR

A total of 142 LUAD-related target genes were intersected with 484 ferroptosis-related genes, resulting in the identification of ATF4, which was subsequently used as the seed node for the RWR algorithm. Given its biological significance in ferroptosis and LUAD progression^23,24and its centrality within the top cluster identified by MCODE, ATF4 was considered a suitable starting point for the RWR algorithm. RWR with a restart probability of 0.7 was then conducted based on the top 1 cluster identified by the MCODE to identify potential key genes associated with LUAD ferroptosis. Through iterative RWR analysis, genes were prioritized based on their functional proximity and topological relevance to ATF4 within the LUAD-specific PPI network, resulting in a top 10 list including TP53, GAPDH, XBP1, CREB1, DDIT3, EIF2S1, JUN, ACTB, MAPK3, and AKT1. As the RWR was performed on a LUAD-derived PPI network, the resulting genes are functionally relevant to LUAD and may represent ferroptosis-associated LUAD regulators. The complete gene score ranking is available in Supplementary Table S2. Figure 2 illustrates the pipeline used for RWR algorithm-based prediction of LUAD ferroptosis-related genes.

Fig. 2.

Pipeline of RWR algorithm-based prediction of LUAD-related genes.

Differential gene expression analysis in LUAD microarray data

Differential expression analysis identified 971 upregulated genes and 1419 downregulated genes (Supplementary Table S3). PCA analysis revealed a clear separation between LUAD samples (red) and normal lung tissue samples (blue) in the two-dimensional principal component space. The first principal component (Dim1) and the second principal component (Dim2) accounted for 20.3% and 2.26% of the total variance, respectively. This significant separation indicates a distinct gene expression difference between LUAD and normal samples, as shown in (Fig. 3A). Figure 3B presents the heatmap of the top 100 differentially expressed genes. Figure 3C shows the volcano plot of upregulated and downregulated genes under stricter selection criteria |logFC| > 1.5 and p < 0.0001. After conducting an intersection analysis with the top 10 genes predicted by the RWR algorithm, we found that XBP1 prominently featured in both the differentially expressed genes and RWR-predicted genes, indicating that it may be a key gene involved in LUAD progression.

Fig. 3.

Identification of differentially expressed genes in LUAD microarray data.

Ferroptosis and tumor signaling pathway analysis of key genes

Through GSEA of the target gene XBP1, several BP and KEGG signaling pathways closely associated with ferroptosis and immune regulation were identified, as shown in (Fig. 4).

Fig. 4.

GSEA of the candidate target XBP1. KEGG pathway image adapted from the KEGG database²⁵© Kanehisa Laboratories.

Ferroptosis, an iron-dependent form of programmed cell death, is closely linked to tumor metabolism and microenvironment regulation. In this study, the positive regulation of oxidative stress-induced cell death was identified, which closely matches the mechanism of ferroptosis. LUAD cells often evade ferroptosis by regulating glutathione metabolism and reactive oxygen species (ROS) levels, but under certain conditions, the accumulation of oxidative stress can induce ferroptosis in tumor cells, thereby inhibiting tumor progression. Additionally, the identification of oxaloacetate metabolic process and glycerophospholipid metabolism further supports the involvement of ferroptosis in LUAD. Oxaloacetate, a key intermediate in the tricarboxylic acid cycle, may influence mitochondrial energy metabolism regulation during ferroptosis. Glycerophospholipid metabolism directly participates in lipid peroxidation, a hallmark of ferroptosis. Abnormalities in these metabolic pathways suggest that LUAD cells might avoid ferroptosis through metabolic reprogramming to maintain their proliferative capacity.

This study also revealed several signaling pathways and biological processes closely tied to tumor immune escape and immune microenvironment remodeling. Natural killer (NK) cell-mediated cytotoxicity and the B cell receptor signaling pathway indicate significant alterations in both innate and adaptive immunity in LUAD. NK cells play a crucial role in tumor immune surveillance, but immune suppressive factors in the tumor microenvironment may weaken NK cell cytotoxicity and promote tumor immune escape. Moreover, regulation of sprouting angiogenesis is linked not only to tumor neovascularization but also to the dynamic equilibrium between angiogenesis and immune cell recruitment, which may impact immune microenvironment remodeling.

Additionally, the mTOR signaling pathway and glutathione metabolism were found to play dual roles in ferroptosis and immune regulation. The mTOR signaling pathway, a critical hub in cell growth and metabolic regulation, is closely related to tumor immune suppression, metabolic reprogramming, and ferroptosis. Overactivation of mTOR may enhance the metabolic adaptability and immune evasion of LUAD cells, while inhibiting mTOR signaling could increase their sensitivity to ferroptosis. Additionally, glutathione metabolism plays a protective role in regulating ferroptosis and may also influence the function and anti-tumor activity of immune cells by maintaining the redox balance in the tumor microenvironment.

Finally, GSEA enriched the graft-versus-host disease pathway, which is involved in immune regulation and inflammatory responses, suggesting potential dysregulation of the immune-inflammatory network in LUAD. Abnormal activation of this pathway could exacerbate the development of an immune suppressive microenvironment, further promoting tumor progression.

Tumor microenvironment analysis

To further explore the role of XBP1 in the TME of LUAD, we analyzed the relationship between XBP1 and six types of tumor-infiltrating immune cells (B cells, CD8 + T cells, CD4 + T cells, macrophages, neutrophils, and dendritic cells) using data from the TIMER database. As shown in Fig. 5, there is a weak correlation between XBP1 expression levels and tumor purity. This phenomenon likely indicates that XBP1 expression is more strongly influenced by specific microenvironmental factors or molecular signaling rather than by the abundance of tumor cells alone. Moreover, this low correlation may imply that XBP1 plays a cell-type-specific role in regulating the tumor microenvironment, rather than uniformly affecting the overall proportion of tumor cells or stromal cells.

Fig. 5.

The correlation between XBP1 and tumor-infiltrating immune cells in LUAD.

For the other six immune cells, a general negative correlation between XBP1 expression and immune cell infiltration levels was observed. The significant negative correlation with B cell infiltration suggests that XBP1 may regulate immune cell differentiation and function during endoplasmic reticulum stress, further supporting its potential role in antigen presentation and antibody production. Regarding CD8 + T cell infiltration, there was no significant correlation with XBP1 expression, while the weak negative correlation with CD4 + T cells, though statistically significant, suggests that XBP1 may indirectly influence CD4 + T cell function by modulating immune-suppressive signals within the tumor microenvironment, possibly playing a role in T cell differentiation or immune tolerance. Additionally, the negative correlation with macrophages, neutrophils, and dendritic cells suggests that XBP1 could have a broader regulatory effect on the tumor immune microenvironment. Macrophages and neutrophils are crucial in tumor inflammation, and their reduced infiltration may imply that XBP1 influences their recruitment or activation by modulating pro-inflammatory or anti-inflammatory signals. Notably, the significant negative correlation with dendritic cells suggests that XBP1 may play a key role in tumor immune evasion by affecting the infiltration of these critical antigen-presenting cells, thereby potentially compromising the tumor’s antigen presentation capability.

External cohorts validation of XBP1 expression

To further validate the high expression of XBP1 in LUAD tumor tissues, we analyzed three external cohorts, including two GEO database cohorts (GSE10072 and GSE140797) and the TCGA-LUAD dataset. These datasets encompass different research cohorts and sample sources, offering high representativeness and reliability. The analysis revealed that XBP1 expression was significantly higher in tumor tissues than in adjacent normal tissues across all datasets. Notably, the GSE10072 and TCGA-LUAD datasets showed highly significant differences (p < 0.0001), while the GSE140797 dataset, though with a smaller sample size, also exhibited statistical significance (p < 0.05), as shown in (Fig. 6). The large sample size of the TCGA dataset further confirmed the robustness of these findings, strengthening the statistical power of the analysis and the generalizability of the conclusion.

Fig. 6.

Validation of the high expression of XBP1 in LUAD tumor tissues using external cohorts.

Discussion

LUAD is the most prevalent form of lung cancer, and the identification and screening of relevant targets are critical for elucidating the molecular mechanisms of the tumor and developing novel therapeutic strategies²⁶. Recently, ferroptosis, a newly recognized form of iron-dependent cell death, has been increasingly acknowledged for its pivotal role in the initiation and progression of LUAD^27,28. Ferroptosis involves the accumulation of intracellular iron and the generation of ROS, leading to lipid peroxidation of the cell membrane and eventual cell death²⁹. In particular, ferroptosis is closely associated with increased oxidative stress during tumor cell metabolic reprogramming³⁰ significantly affecting tumor cell survival and proliferation. Therefore, exploring ferroptosis-related targets aids in understanding the pathological mechanisms of LUAD and provides new insights for targeted therapy.

ML algorithms, when applied to gene expression data, are capable of detecting underlying patterns and associations by thoroughly analyzing large biological datasets. In comparison to traditional experimental approaches, ML offers the advantage of efficiently identifying key genes and providing more precise predictions in a shorter timeframe. This method overcomes the limitations of conventional experimental techniques in terms of speed, accuracy, and comprehensiveness, providing a more effective approach to the discovery of LUAD targets and the optimization of therapeutic strategies.

In this study, LUAD-associated targets were identified using several databases, including GeneCards, DisGeNET, and OMIM, and a PPI network was constructed. The MCODE clustering method was employed to select the top 1 cluster, with genes intersecting with ferroptosis-related genes used as seed nodes. The RWR algorithm was then applied to predict LUAD-related targets. Cross-analysis with differentially expressed genes from LUAD microarray data revealed XBP1 as a key candidate target gene. The high expression of XBP1 in LUAD and its essential role in ferroptosis have been supported by several studies^31–33.

GSEA revealed that XBP1 is closely associated with ferroptosis in LUAD and is involved in immune regulation and metabolic reprogramming. XBP1 promotes ferroptosis by positively regulating oxidative stress-induced cell death, thereby enhancing the tumor cells’ response to oxidative damage and facilitating cell death³⁴. Furthermore, XBP1 may allow LUAD cells to evade ferroptosis by modulating oxaloacetate and glycerophospholipid metabolism, thus sustaining tumor cell proliferation and growth³⁵. This finding underscores the critical role of ferroptosis in LUAD progression, with metabolic reprogramming emerging as a key mechanism through which tumor cells may avoid ferroptosis.

Beyond ferroptosis, XBP1 also impacts the TME by regulating the mTOR signaling pathway and glutathione metabolism, promoting immune evasion. TME analysis showed a negative correlation between XBP1 expression and immune cell infiltration, particularly with dendritic cells, B cells, CD4 + T cells, and macrophages. This suggests that XBP1 may suppress immune cell function and infiltration by modulating the endoplasmic reticulum stress response and immune-suppressive signals. This mechanism likely contributes to the tumor’s ability to evade immune surveillance, particularly in the case of dendritic cells, where XBP1 may impair antigen presentation, thus reducing the immune system’s ability to recognize and eliminate tumor cells. Multi-dataset validation (GSE118370, GSE68465, and TCGA-LUAD) further confirmed the high expression of XBP1 in LUAD and its potential involvement in ferroptosis and TME.

Admittedly, experimental validation is necessary to confirm the role of XBP1 in LUAD-related ferroptosis and immune modulation, which represents a limitation of this study. Nevertheless, our use of integrated bioinformatics and machine learning provides a robust framework for identifying high-confidence targets. These results offer a valuable foundation for future mechanistic and clinical investigations into ferroptosis in LUAD.

Conclusions

This study integrates ML with the PPI network to predict XBP1 as a potential target gene related to ferroptosis in LUAD through the RWR algorithm. Both GSEA and TME analyses demonstrate that XBP1 is closely linked to ferroptosis-related pathways and plays a crucial role in the immune microenvironment of LUAD, with particular emphasis on its unique impact on immune cell infiltration. Validation across external cohorts further confirmed the elevated expression of XBP1 in LUAD tissues and its strong correlation with immune cell infiltration. These findings highlight XBP1 as a promising biomarker for the early diagnosis of LUAD and provide a solid theoretical basis for developing precision therapies targeting ferroptosis mechanisms.

Materials and methods

LUAD and ferroptosis-associated genes

Using “Lung Adenocarcinoma” and “LUAD” as keywords, a search was conducted across several disease-related databases, including GeneCards (https://www.genecards.org/), DisGeNET (https://disgenet.com/), and OMIM (https://www.omim.org/). Only gene data from Homo sapiens were considered. After removing duplicates, target genes associated with LUAD were identified.

Ferroptosis-related genes were sourced from FerrDb V2³⁶ (http://www.zhounan.org/ferrdb/current/), the first manually curated database dedicated to ferroptosis. This database is used for the management and identification of biomarkers, regulators, and diseases related to ferroptosis.

Construction of the protein-protein interaction network

The RWR algorithm was applied to identify LUAD-related target proteins within the PPI network, which is one of the most essential tools for identifying and predicting key genes. The LUAD-related target PPI network was constructed using the STRING plugin in Cytoscape (Version 3.10.1). Each protein-protein interaction was assigned a confidence score, derived from multiple sub-scores, including gene co-expression, cross-species co-occurrence, database annotations, experimental validation, gene fusion events, genomic proximity, cross-species interactions, and text mining. By evaluating protein associations based on various factors such as gene expression pattern similarity, evolutionary conservation, functional relevance, and experimental validation, a final score was computed by integrating these sources of evidence, which quantifies the confidence in the protein-protein interactions.

Random walk with restart algorithm for gene prioritization

RWR is a graph theory algorithm widely used for network analysis and node ranking^37–40. Its fundamental concept is to model a random walk process starting from specific nodes in the graph, known as seed nodes, while employing a restart mechanism that allows the walker a certain probability of returning to these seed nodes.

Initially, each node in the graph is associated with a vector representing the node’s visit probability. For seed nodes, their initial probability is assigned equally, while the initial probability for all non-seed nodes is set to 0. During each random walk step, the walker selects an adjacent node to move forward. Unlike a traditional random walk, the RWR algorithm incorporates a restart mechanism that gives the walker a certain probability of returning to the seed node at each step. This restart probability is typically set as a constant value between 0 and 1. Each time the node probability vector is updated, the probabilities of the nodes are weighted according to the transition results from the random walk. The update formula is as follows:

1

In this context, denotes the ranking vector, W represents the network adjacency matrix, 1-c is the restart probability, and e represents the starting vector of initial probabilities. The resulting probability vector reflects the correlation or importance of each node in relation to the seed nodes. A higher probability value for a node indicates a stronger association between that node and the seed nodes.

Differential expression analysis identifies candidate LUAD genes

LUAD microarray data were retrieved from the GEO database (https://www.ncbi.nlm.nih.gov/geo/), utilizing the GPL570 (Affymetrix Human Genome U133 Plus 2.0 Array) platform, with accession number GSE118370. The cohort for this study consisted of 12 samples, including 6 invasive LUAD tissue samples and 6 normal lung tissue control samples.

Differential gene expression analysis in LUAD was performed using the limma package (Version 3.60.6) in R Studio (Version 4.4.1). Genes were considered differentially expressed if |logFC| > 1 and p < 0.05 ^41,42. A volcano plot was generated for genes with |logFC| > 1.5 and p < 0.0001, and a heatmap was created for the top 50 upregulated and downregulated genes. The differentially expressed genes were intersected with the top 10 genes predicted by the RWR algorithm, and the candidate gene XBP1 was identified as a key ferroptosis-related gene in LUAD.

Gene set enrichment analysis

GSEA (http://software.broadinstitute.org/gsea/index.jsp) was conducted to assess functional dynamics and pathway alterations in the samples. Based on XBP1 expression levels, the samples were classified into high-expression (≥ 50%) and low-expression (< 50%) groups, using the median expression value as the cutoff, which is a commonly accepted strategy in single-gene GSEA analyses^43,44. Two gene sets, c5.go.bp.v7.4.symbols.gmt and c2.cp.kegg.v7.4.symbols.gmt, were downloaded from the Molecular Signatures Database (http://www.gsea-msigdb.org/gsea/downloads.jsp) and used to evaluate the enrichment of Gene Ontology (GO) biological processes (BP) and KEGG pathways, respectively. The analysis was conducted with the following parameters: a minimum gene set size of 5, a maximum gene set size of 5000, and 1000 resamplings. To ensure the reliability of the results, statistical significance was defined by p < 0.05 and FDR < 0.25.

Tumor microenvironment analysis

TIMER⁴⁵ (http://timer.cistrome.org/) is a widely used tool for integrated analysis of tumor immune infiltration across various cancer types, aimed at exploring the relationship between gene expression and immune cell infiltration. The TIMER system was applied to examine the correlation between the candidate gene XBP1 in LUAD and several tumor-infiltrating immune cells, including B Cells, CD8 + T Cells, CD4 + T Cells, Macrophages, Neutrophils, and Dendritic Cells.

Validation of predicted gene expression with external cohorts

To validate the candidate genes predicted by the RWR algorithm, this study employed three independent external cohorts: GSE118370, GSE68465, and TCGA-LUAD. Gene expression data from LUAD samples in these datasets were extracted and compared with the candidate genes identified in this study. To maintain consistency and comparability across datasets, all data were normalized, and batch effects were corrected to mitigate inter-batch variability.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

C.L. performed the experiments and prepared the manuscript. P.H. collected experimental data and contributed to the manuscript preparation. R.Q. performed data processing and analysis. X.Y. and C.D. collected provided resources and performed software-related tasks. F.H. designed and directed the project, conceptualized the study, and provided methodology and supervision. All authors agreed on the final version of the manuscript.

Data availability

The publicly available datasets used in this study were obtained from GSE118370 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE118370), GSE68465 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE68465), and TCGA-LUAD (https://portal.gdc.cancer.gov/projects/tcga-luad).

Declarations

Competing interests

The authors declare no competing interests.