Identification of a novel ferroptosis-induced immunogenic cell death related signature based on a machine learning framework in colorectal cancer

Abstract

Background

Ferroptosis and immunogenic cell death play vital roles in colorectal cancer (CRC). The interplay between ferroptosis and immunogenic cell death (F-ICD) represents a promising frontier in cancer therapy. However, few studies have explored the combined regulatory effects of F-ICD in CRC.

Methods

In current study, we identified F-ICD related genes based on analysis of single-cell transcriptomics level and developed F-ICD related signature using 101 machine learning algorithms and WGCNA analysis. Differential analysis between normal and tumor samples was performed using DESeq2 (|logFC|>1, p. adj < 0.05). The RSF algorithm was chosen for further analysis due to its strong predictive performance, making it a robust tool for our study. An external validation was performed to access the expression level of seven key F-ICD related genes.

Results

By quantifying the expression levels of 44 genes related to F-ICD, we found that F-ICD activity was significantly elevated in NK cells, T cells, and some B cells. The module showed a significant correlation with the F-ICD score (r = 0.66). The predictive model had highly accurate AUCs in three datasets (0.99, 0.61, and 0.58 for the 3-years training sets), revealing the importance of F-ICD in different pathological stages and prognoses in CRC. Further results indicated that F-ICD was associated with pathways such as oxidative phosphorylation and NF-κB signaling. Patients with high F-ICD had significantly different mutation profiles and poorer prognoses.

Conclusion

This study developed a novel signature integrating ferroptosis and immunogenic cell death, creating a valuable model for predicting prognosis and the tumor immune environment in CRC. Furthermore, seven key genes emerged as promising targets for further investigation and therapeutic intervention, highlighting their potential role in ferroptosis and immunogenic cell death.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03147-1.

Introduction

Colorectal cancer (CRC) ranks as the third most commonly diagnosed malignancy and is the second leading cause of cancer-related mortality in 2022, representing approximately 9.6% of diagnosed cancer cases and 9.3% of cancer-related deaths [1]. Despite significant advancements in CRC treatment over the past decade, the 5-year overall survival (OS) rate is still less than 14% for patients with metastatic CRC [2]. With the emergence of CRC as a groundbreaking postsurgical treatment modality, combined with radiotherapy, chemotherapy, and targeted therapy, the management of CRC has transitioned into a new era [3]. Immune checkpoint inhibitors (ICIs) have emerged as the predominant immunotherapy strategy presently utilized in the fight against CRC [4]. While certain subtypes of CRC demonstrate favourable treatment responses, others such as MSS and MSI-L exhibit unsatisfactory efficacy [5, 6]. Investigating their specific mechanisms to devise novel combination strategies involving ICIs is imperative. Therefore, gaining deeper insights into the immunotherapy of patients with colorectal cancer remains imperative to improve the survival outcomes of patients afflicted with this malignancy.

Immunogenic cell death (ICD) is a form of cancer cell death that can be induced by specific chemotherapeutic agents, oncolytic viruses, various physical and chemical therapies [7]. The ICD possesses the remarkable capacity to activate adaptive immune responses within an immunocompetent milieu, leading to the stimulation of immune cells such as dendritic cells (DCs) and T cells [8]. Several studies have indicated that inducing ICD via methods such as chemotherapy can synergistically augment tumor immunogenicity, thereby fostering antitumour immune responses and consequently enhancing patient treatment outcomes [8–10]. Multiple predictive models based on ICD in CRC patients have been developed, demonstrating the potential for forecasting survival rates and predicting immunotherapy response outcomes [11–13]. The heterogeneity of CRC and the complexity of the tumor microenvironment can affect the accuracy and generalizability of these predictive models. Additionally, integrating ICD-based biomarkers with other molecular or clinical parameters remains a key area for further research.

Previous studies have reported that ferroptosis plays a pivotal role in the reinvigoration of antitumour immunity [14]. Luo et al. reported that ferroptoic tumor cells can activate DCs and lymphocytes, and enhance the immunogenic milieu within the tumor, thereby promoting antitumour immune responses [15, 16]. Ferroptosis induces the release of DAMPs from tumor cells, precipitating endoplasmic reticulum stress and DNA damage, which significantly amplifies the effect of ICD in cancer immunotherapy [17]. Consequently, the interplay between ferroptosis-induced ICD and the immune system highlights the potential for novel therapeutic strategies that leverage ferroptosis to increase the efficacy of cancer immunotherapy. Recent studies have highlighted that inducing ferroptosis or ICD in tumor cells can enhance the efficacy of immunotherapy and chemotherapy, especially in tumors with poor immune infiltration. However, several challenges remain, including limited specificity, potential off-target effects, and the lack of reliable biomarkers for predicting therapeutic response. Moreover, the intricate crosstalk between ferroptosis and ICD within the tumor microenvironment is not yet fully understood.

In this study, single-cell sequencing analysis was conducted to identify ferroptosis-and immunogenic cell death (F-ICD)-related markers in CRC via machine learning algorithms. This molecular characteristic can further guide future prognosis and provide new insights for developing more effective therapeutic targets for CRC patients.

Materials and methods

Dataset acquisition and processing

RNA-seq data pertaining to colorectal adenocarcinoma (COAD), alongside its corresponding comprehensive clinical information, were thoroughly acquired from The Cancer Genome Atlas (TCGA) and two COAD cohorts (GSE39582, GSE17536). Additionally, samples exhibiting incomplete or absent clinicopathological data were meticulously excluded, culminating in a final cohort. The COAD single-cell RNA sequencing (scRNA-seq) dataset GSE161277 was acquired from the Gene Expression Omnibus (GEO) database. The RNA expression profiles obtained underwent logarithmic transformation using the Log2 (TPM + 1) method to achieve normalization to effectively reduce batch effects and improve the comparability of data across samples, which is crucial for the downstream analyses.

In order to delineate genes correlated with ferroptosis-immunogenic cell death (F-ICD), relevant gene sets were procured from the GeneCards database (https://www.genecards.org/). Furthermore, differentially expressed genes (DEGs) were discerned utilizing the TCGA database, with DEGs defined by fold change > 1.2 and p. adj < 0.05.

To compare gene expression levels between tumor and normal tissues across different datasets, we integrated RNA-seq data. Gene expression values were uniformly normalized to TPM and log2-transformed. To address potential technical variations arising from different data sources, we first assessed batch effects using PCA and hierarchical clustering. When batch effects were detected, we applied the ComBat algorithm from the R package to correct for non-biological variation. The success of batch correction was verified by re-assessing sample clustering patterns via PCA after adjustment. These procedures ensured comparability of gene expression profiles across datasets while preserving relevant biological signals.

Single-cell data analysis

Quality control of the scRNA-seq data was meticulously executed by employing the Seurat R package [18] to effectively eradicate potential doublets. The initial quality control procedures involved the retention of cells exhibiting mitochondrial gene content was less than 10%, alongside genes expressed in a minimum of three cells within an expression range from 500 to 25,000. Cell clusters were meticulously constructed via the “FindClusters” and “FindNeighbors” functions, followed by visualization via the “t-SNE” method for optimal depiction. Meticulous annotation was subsequently performed for each discerned cell cluster.

With the objective of precisely discerning the enrichment fraction of each individual cell comprising the scRNA-seq dataset, the “AUCell” package was employed. Cells were then annotated based on the highest AUCell score corresponding to a given cell-type signature. Additionally, to validate the annotations, we cross-referenced the assigned cell types with expression levels of canonical marker genes for each cell type. The “FindMarkers” function was used to analyze the DEGs between two distinct groups. The statistical significance of the (DEGs was meticulously assessed employing the Wilcoxon test (p. adj < 0.05). P-values for DEGs were indeed adjusted for multiple testing using the Benjamini-Hochberg method to ensure the robustness of our results and reduce the risk of type I errors. Genes exhibiting differential expression between cells characterized by high and low F-ICD scores at the single-cell transcriptome level were deemed to be implicated in F-ICD. The cell-to-cell communication analysis was also conducted by “CellChat” package.

Single‑sample gene set enrichment analysis (ssGSEA) and functional analyses

Single sample gene set enrichment analysis (ssGSEA) was employed to quantify activity profiles derived from the original gene expression data. Each ssGSEA score signifies the extent to which the genes within a specific gene set are collectively upregulated or downregulated within a given sample. The TICD scores were computed via the “gsva” function from the R package “GSVA” for the TCGA-COAD samples [19]. To obtain biological processes associated with individual signatures, we tested the GSVA scores across 50 hallmark pathways. Pathways exhibiting significant differences between the high- and low-risk groups were subsequently analyzed via the “limma” package [20]. To identify biological processes potentially involved in different risk subgroups, Gene Ontology (GO) gene sets were generated using the recommended criteria of FDR < 0.25 and |NES| > 1, with the assistance of the “clusterProfiler” package.

Identification of ferroptosis-immunogenic cell death-related signature genes via WGCNA

In this study, WGCNA of TCGA-COAD bulk RNA-seq data was performed via the “WGCNA” package. To maintain adherence to the scale-free assumption of the constructed biological networks, the soft thresholding parameter value was selected following the guidelines provided by WGCNA. The weighted neighbor-joining matrix was subsequently transformed into a topological overlap matrix (TOM), followed by the creation of a TOM dissimilarity matrix. To cluster the cell types within our gene-set, we utilized the dynamic cut tree algorithm, which employs an innovative dynamic branch cutting method to identify clusters within a dendrogram. Ultimately, the modules exhibiting the strongest correlation with the TICD were identified for further analysis.

Machine learning to build predictive models

To construct highly predictive model, differential analysis between normal and tumor samples in the TCGA-COAD bulk RNA-seq dataset was conducted via the DESeq2 package (|log FC| > 1 and p. adj < 0.05). Subsequently, we intersected DEGs identified at the bulk RNA-seq level with genes present in the F-ICD-associated modules identified through WGCNA. The genes resulting from this intersection were deemed to play a role in ferroptosis and immunogenic cell death across both bulk and single-cell transcriptome analyses. Consequently, we designated these genes as immunogenic cell death-related genes (F-ICD-related genes). To develop a robust prognostic model, we employed a machine learning approach based on the LASSO (Least Absolute Shrinkage and Selection Operator) Cox regression algorithm. Initially, univariate Cox regression analysis was performed to identify genes significantly associated with overall survival (p < 0.05). The candidate genes were then subjected to LASSO Cox regression using the “glmnet” R package to prevent overfitting and select the most informative features.

Construction of the prognostic model and predictive nomogram

The TCGA-COAD, GSE39582, and GSE17536 cohorts were stratified into high and low-risk groups on the basis of the median risk score. Kaplan-Meier survival curve analysis was conducted via the “survminer” package to assess potential differences in OS between the high- and low-risk groups, with statistical significance determined via the log-rank test (p < 0.05). K-fold cross-validation was used to ensure the model’s generalizability and to prevent overfitting. ROC curve analysis was conducted via the “timeROC” package to evaluate the sensitivity and specificity of the F-ICD-related signature (F-ICDRS) in predicting OS in CRC patients. Furthermore, we compared the area under the curve (AUC) of the F-ICDRS with that of other clinical characteristics. Univariate and multivariate Cox regression analyses were conducted on the TCGA-CRC, GSE39582, and GSE17536 datasets to ascertain whether the F-ICDRS serves as an independent prognostic factor in CRC patients. We subsequently devised a nomogram that integrates the clinical characteristics of F-ICDRSs and the expected survival of CRC patients to improve the prognostic accuracy and predictive capability of our model. ROC curves and C-indexes were used to evaluate the accuracy and precision of the nomogram.

Single-gene survival and expression analysis

To assess the expression level of F-ICD-related genes, we assessed the mean normalized expression level of each gene in the TCGA-CRC samples. To examine the prognostic significance of individual F-ICD-related genes in CRC patients, Kaplan-Meier plotter (kmplot.com) was used. To determine the expression levels of F-ICD-related genes in CRC cells, we downloaded gene expression profiles from CCLE datasets (portals.broadinstitute.org/ccle).

Analysis of mutant allele tumor heterogeneity (MATH)

To quantify tumor heterogeneity, mutant-allele tumor heterogeneity (MATH) was established on the basis of the variation in a tumor. The MATH score is derived from whole-exome sequencing analysis of DNA obtained from tumors and corresponding normal tissues [21]. Using previously established techniques, we calculated the MATH score for each CRC patient. Subsequently, survival analysis was performed using these individual MATH scores. The “maftools” R package was employed to generate waterfall plots, which visually depict the mutation landscape of CRC patients categorized into high- and low-risk groups. This analysis allowed us to investigate somatic mutations associated with F-ICDRS.

Analysis of copy number variation (CNV)

Specifically, CNV analysis was performed using the GISTIC2.0 algorithm applied to segmented copy number data downloaded from the TCGA database. GISTIC2.0 identifies regions of significant focal and arm-level copy number amplifications and deletions by calculating a G-score and estimating false discovery rates (q-values). Default parameters were used, including a confidence level of 0.99 and a broad length cutoff of 0.7. The output was further processed to determine the frequency and significance of CNV events in the cohort. This description has been incorporated into the Methods section to improve reproducibility and transparency.

Statistical analysis

All the statistical analyses were conducted via R software (version 4.2.2). To compare groups, t-tests or the Mann-Whitney U test was used employed to compare continuous variables between groups. Survival analyses were conducted via the Kaplan-Meier method, and the predictive performance of the risk model was evaluated via the “survivalROC” R package, which uses time-dependent receiver operating characteristic (ROC) curves to assess subject work characteristics. Univariate and multivariate Cox regression analyses were conducted to explore independent prognostic factors. Unless specified otherwise, statistical significance was defined as p < 0.05.

Results

Ferroptosis-immunogenic cell death characterization of the ScRNA samples

Data from three CRC patients were extracted from the GSE161277 single-cell dataset (Fig. 1). The cells were initially screened on the basis of gene expression levels and mitochondrial content within individual cells (Fig. 1A, B). We utilized principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) for dimensionality reduction analysis. With respect to marker genes, we obtained six major clusters, including tumor cells, CD4⁺ T cells, CD8⁺ T cells, monocytes, natural killer (NK) cells and B cells (Fig. 1C). A clustered heatmap was constructed to show the top five marker genes for each cluster (Fig. 1D). Forty-four F-ICD-related genes were obtained from the GeneCards database (Supplementary Table 1). Next, we quantified the expression levels of a set of 44 F-ICD-related genes in all cells to assess the activity of F-ICDs across different cell types (Fig. 1E). Among the six cell types, we observed significantly higher F-ICD activity in NK cells, T cells, and a subset of B cells (Fig. 1F). On the basis of F-ICD activity, we identified differentially expressed genes (DEGs) for model building (Supplementary Table 2).

Fig. 1.

The data analyzing process of the study. The processing procedure of scRNA-seq data and the ferroptosis-immunogenic cell death (F-ICD) characteristic in the scRNA-seq data. A The quality of single-cell data before processing. B The quality of single-cell data after processing. C The correlation of quality control data in samples after quality control. D The results of hierarchy cluster analysis. E The t-SNE diagram illustrating the diversity of cell types as discerned by marker genes. F The heatmap visually depicts the foremost five marker genes within each cellular cluster. G The activity score of F-ICD in each individual cell is delineated. H Violin plot showing F-ICD results in different cell types

Identification of key modules correlated with ferroptosis-immunogenic cell death in CRC

The TCGA-CRC samples were subjected to the ssGSEA algorithm to calculate the F-ICD activity score. This scoring enabled the assessment of F-ICD-related gene expression patterns and subsequent WGCNA. WGCNA was performed to identify modules of highly correlated genes (Fig. 2A). Six modules were identified by setting the minimum module size of 60 genes and the module dissimilarity threshold to 0.25 (Fig. 2B, C). The findings demonstrated that the blue module was significantly correlated with the F-ICD score. (r = 0.66). This moderate-to-strong positive correlation suggests a meaningful relationship between ferroptosis-induced immunogenic cell death in CRC, which could imply the potential biological of CRC cells. The volcano plot depicted the DEGs between tumor and normal tissues in the TCGA-CRC dataset (Fig. 2D). We identified 62 hub genes that were considered F-ICD-related genes (F-ICDs) by intersecting the F-ICD feature genes with the DEGs (Fig. 2E). Functional enrichment analysis of F-ICDs (Fig. 2F) revealed that F-ICDs were enriched in functions such as cadherin binding (MF), membrane microdomain (CC), response to oxidative stress and lymphocyte proliferation (BP). Through univariate Cox regression analysis, model construction, and validation, we identified a total of 9 shared genes. We explored the correlations among the 9 genes (Fig. 2G). These analyses allowed us to uncover potential interactions among the identified genes, providing further insights into their collective roles in CRC progression and prognosis. The results of the CNV analysis indicated that no significant CNV was observed in these genes (Fig. 2H).

Fig. 2.

Identification of the ferroptosis-immunogenic cell death-related genes (F-ICD genes). A The dendrogram illustrates the hierarchical clustering of TCGA-COAD samples, while the heatmap at the bottom delineates the calculated F-ICD scores for each sample via the ssGSEA algorithm. B The dendrogram displays the clusters resulting from the WGCNA analysis. C The heatmap representing module-trait relationships illustrates a close association between the blue module and the F-ICD trait. D The volcano plot displays the outcomes of the differential analysis conducted on TCGA-COAD tumor samples versus normal samples, with distinct markings for the top 5 genes. E The Venn diagram illustrates the shared genes between the blue module and the differentially expressed genes (DEGs) identified in the bulk RNA-seq analysis. F. Gene Ontology (GO) enrichment analysis was performed on the genes associated with F-ICD. G The co-relationship analysis for the F-ICD genes. H The frequency of copy number variation (CNV) between the F-ICD genes is indicated

Construction and evaluation of the machine learning-based prediction risk model

To build an accurate prediction model based on F-ICDs, we employed 101 machine-learning algorithms to compute the C-index in both the training and validation sets (Fig. 3A). We selected the RSF algorithm for further analysis because of its high predictive efficacy, thereby making it a reliable choice for our research endeavors. To further validate this model, we analysed its predictive ability and authenticity. Survival analysis of the three datasets revealed significant differences between the high-risk and low-risk groups (Fig. 3B–D). These findings indicates that the risk stratification is robust and effectively distinguishes between different prognostic outcomes in the studied populations. The results indicated that the area under the curve (AUC) values of the F-ICDs were 0.99, 0.61, and 0.58 for the 3year training sets in the 3 datasets, respectively (Fig. 3E–G).

Fig. 3.

A consensus ferroptosis-immunogenic cell death-related signature (F-ICDRS) was meticulously devised and rigorously validated through an innovative integrative procedure grounded in machine learning methodologies. A 101 distinct prediction models are developed employing a rigorous tenfold cross-validation framework, followed by meticulous computation of the C index for each model across all validation datasets. B–D Kaplan–Meier survival curves depicting overall survival (OS) were constructed based on the F-ICDRS within the training subset of TCGA and external validation sets (B), GSE39582 (C) and GSE17536 (D). E–G Receiver operating characteristic (ROC) curves, delineating the specificity and sensitivity of the F-ICDRS in predicting 1-year, 3-year, and 5-year OS, were generated utilizing data from the training subset of the TCGA training subset (E) and external validation sets (F, G)

First and foremost, we investigated the relationships between F-ICDs and clinicopathologic features in CRC patients. The results demonstrated that the number of F-ICDs could influence the T stage, N stage, M stage, and TNM stage (Fig. 4A–C). Moreover, we observed significant associations with overall survival (OS), suggesting that F-ICDs may serve as valuable biomarkers for prognosis and treatment stratification in patients with colorectal cancer. Moreover, the proportion of patients with T, N, and M stages was greater in the high-risk group (Fig. 4D–F). This further underscores the potential of F-ICDs as prognostic markers, highlighting their role in the stratification of patients on the basis of disease severity and progression. On the basis of the predictive performance of F-ICDs, we also assessed their ability to predict distant metastasis in CRC patients, revealing that F-ICDs had an AUC of 0.698 for predicting the M stage in these patients (Fig. 4G). Through Kaplan-Meier curve analysis, we found that F-ICDs not only possess strong predictive ability in pathological staging but also exhibit significant prognostic value across different age groups (Fig. 4H–K).

Fig. 4.

Comprehensive assessment of the F-ICDRS model. A The distribution of clinical characteristics alongside the expression patterns of model genes is analyzed in accordance with the risk scores derived from the F-ICDRS risk score. B Disparities in risk scores among staged T patients. C Examination of the correlation between the low- and high-risk groups identified by F-ICDRS and clinical characteristics. D–F Distribution analysis of T, N, and M stages within ICDRS risk subgroups. G ROC curves depicting the predictive performance of the F-ICDRS for metastasis in CRC. H–K Kaplan–Meier survival curves demonstrating consistent performance of the F-ICDRS across subgroups of CRC patients stratified by age and stage

Multivariate survival analysis revealed that F-ICDs were a significant independent risk factor for CRC patients (Fig. 5A). These findings highlight the potential of F-ICDs as critical biomarkers for identifying high-risk patients and tailoring individualized treatment plans to improve patient survival rates. On the basis of the above results, we constructed a nomogram using F-ICDs and clinicopathologic parameters (Fig. 5B). The measured values aligned well with the theoretical predictions (Fig. 5C). The DCA curve and timeROC analysis demonstrated that the nomogram and risk coefficients based on F-ICDs, exhibited excellent predictive performance (Fig. 5D, E). These results suggested that integrating F-ICDs with clinicopathologic parameters further consolidated their predictive utility, as evidenced by the alignment of measured values with theoretical predictions.

Fig. 5.

Construction of a predictive nomogram and pathway enrichment analysis of patients with various F-ICDRS. A Univariate analysis of clinical characteristics along with F-ICDRS regarding OS. B Development of a nomogram incorporating the F-ICDRS alongside pertinent clinical characteristics including age, grade, and stage. C Calibration curves depicting the performance of the nomogram for predicting 1-year, 3-year, and 5-year OS. D Decision curve analysis (DCA) illustrating the net benefit gained from the application of the nomogram. E Comparative analysis of the C index between the nomogram and alternative clinical characteristics. F Ridge plot visualizing the GO terms significantly enriched in the high-risk subgroup. G GSEA revealing enrichment of KEGG terms in the high-risk subgroup. H Examination of the correlation between risk score and the activity of hallmark pathways as determined by GSVA scores. I Evaluation of the correlation between risk scores and the activities of hallmark pathways assessed through GSVA scoring

We performed an in-depth exploration of the molecular mechanisms associated with F-ICDs and their impact on CRC prognosis through comprehensive functional enrichment analysis. The results revealed that the low-risk group was enriched in oxidative phosphorylation, MYC target, and DNA repair pathways (Fig. 5F). The high-risk group was enriched in NF-κB signalling, myogenesis, epithelial mesenchymal transition, and apical junctions (Fig. 5G). The results of GSVA revealed that the high-risk group was mostly related to the TGF-β signalling, epithelial mesenchymal transition, angiogenesis, and myogenesis pathways (Fig. 5H). We performed a rigorous correlation analysis between F-ICDs and pathway scores, further substantiating that F-ICDs are intricately aligned with the biological processes associated with cancer (Fig. 5I).

Tumor mutation analysis in ferroptosis-immunogenic cell death

To investigate the associations between F-ICDs and the genomic characteristics of tumors, we performed mutant allele tumor heterogeneity (MATH) analysis. The higher the risk of CRC patients was, the higher the MATH score was (P = 0.017, Fig. 6A). Patients with higher MATH scores had a worse prognosis (P = 0.037, Fig. 6B). We further found that patients with high-F-ICDs and high-MATH scores had a significantly worse prognosis than patients with low F-ICDs and MATH scores did (Fig. 6C).

Fig. 6.

Genetic changes associated with the stratification of individuals into low- and high-risk groups based on F-ICDRS. A The results illustrating the disparity in mutant allele tumor heterogeneity (MATH) scores. B Survival analysis revealed differences in OS between the low- and high-risk groups. C Kaplan-Meier curve analysis assessing OS based on the combined MATH scores and F-ICDRS risk scores. D, E Waterfall diagram depicting the somatic mutation landscape in low- and high-risk patients. F Correlation analysis of the initial 20 mutated genes within the high-risk subgroup

To investigate gene mutation patterns in different F-ICD groups, we constructed waterfall plots, and found that the mutational spectrum was significantly distinct between the two groups (Fig. 6D, E). We subsequently analysed the correlations among the top 20 mutated genes in the high-risk group. The results indicated that with the exception of APC and TP53, a strong correlation was observed among the majority of the mutated genes (Fig. 6F).

Characterization of ferroptosis-induced Immunogenic cell death in single cells

To explore the role of F-ICDs in the TME at the single-cell level, we analysed the expression levels of CD24, CXXC5, TMEM59, FUS, HADHB, INPP5D, and PLAC8. The results revealed that the seven genes are expressed predominantly in immune cells, particularly CD8⁺ T cells and B cells. (Fig. 7A). Only CD24 was highly expressed in tumor cells. We performed GSEA in the high-risk group defined by seven genes. The results suggested that the allograft rejection and MYC target pathways were the top 2 enriched terms in the high-risk group (Fig. 7B).

Fig. 7.

Characterization of F-ICDRS in single cells. A CD24, CXXC5, TMEM59, FUS, HADHB, INPP5D, PLAC8 were analyzed by scRNA-seq. B GSEA revealed the enrichment of GO terms in the seven genes. C The ligand-receptor interactions originating from high-risk tumor cells. D, E The plots illustrating the CXCL and ITGB2 signaling pathway networks. F–H The Immune score, ESTIMATE score, and stromal score were utilized to ascertain the distinct immune statuses between the low- and high-risk groups

Next, we analysed the associations between intracellular communication and functional pathways in the high-risk group of F-ICDs. The functional manifestation of these genes is primarily observed in B cells and tumor cells (Fig. 7C). Moreover, the interaction between these genes and the tumor microenvironment was critical. The CXCL and ITGB2 signalling pathways were identified as the most active mediators of cell-to-cell communication within the TME (Fig. 7D). These pathways play pivotal roles in orchestrating the complex interactions between tumor cells and the surrounding stromal and immune cells.

Finally, we calculated the immune score, stromal score, and ESTIMATE score via the ESTIMATE algorithm. The results revealed that the stromal score and ESTIMATE score were not significant different between the high-risk and low-risk groups (Fig. 7E). The high-risk group presented a higher immune score than did the low-risk group.

The expression level and role of a single gene in ferroptosis-induced Immunogenic cell death

To understand the expression patterns and potential roles of each gene, we conducted a detailed analysis of their expression levels in tumors and their associations with patient prognosis. CXXC5, FUS, and INPP5D exhibited high expression levels in tumors (Fig. 8A), whereas CD24, TMEM59, HADHB, and PLAC8 presented low expression levels in tumors. CD24, TMEM59, FUS, HADHB, and PLAC8 were good prognostic genes. CXXC5 was a poor prognosis gene (Fig. 8B–H). To determine the expression of seven genes in tumor cells, we detected their expression levels in three CRC cell lines. The results revealed that the expression of seven genes was essentially consistent across the three cell lines (Fig. 8I). TMEM59, FUS, and HADHB had the highest expression levels in tumor cells.

Fig. 8.

Validation of the expression level and prognostic significance of 7 F-ICDRS genes. A The expression level of CD24, CXXC5, TMEM59, FUS, HADHB, INPP5D, PLAC8 in CRC. B–H The prognostic significance of CD24, CXXC5, TMEM59, FUS, HADHB, INPP5D, PLAC8 in CRC. I The expression level of 7 F-ICDRS genes in 3 human colon cancer cell lines

Discussion

In our study, we developed a novel ferroptosis- and immunogenic cell death related model using 101 machine learning algorithms. Our findings highlight the interplay between ferroptosis, a regulated form of cell death characterized by iron-dependent lipid peroxidation, and immunogenic cell death, a process in which dying cells activate an immune response. Overall, we identified seven promising genes (CD24, TMEM59, HADHB, CXXC5, FUS, INPP5D, and PLAC8) that play crucial roles in the regulation of ferroptosis and immunogenic cell death. These genes emerged as key regulators through our rigorous analysis via 101 machine learning algorithms, each contributing uniquely to the interplay between these two cellular processes.

In recent years, there has been extensive research into the role of immunogenic cell death (ICD) in colorectal cancer, revealing significant insights. ICD is a unique form of cell death that not only eliminates tumor cells but also triggers an immune response, thereby promoting antitumour immunity [22]. Studies have shown that certain chemotherapeutic agents and radiation therapies can induce ICD, leading to the release of damage-associated molecular patterns (DAMPs), which are crucial for activating the immune system [23]. Aspirin induces ICD in colon cancer cells, increasing the efficacy of anti-PD-1 and anti-CTLA-4 antibodies and suggesting its potential use as a CRC therapy in combination with tumor antigens [24]. Wang et al. reported that camptothetes could induce ICD in colorectal cancer cells via clathrin-mediated endocytosis, increase the efficacy of immune checkpoint inhibitors, and reduce toxicity and increase the anti-CRC efficacy [25]. Multiple studies have utilized marker genes of the ICD to identify tumor subtypes with specific characteristics of the ICD, enabling more accurate predictions of patient responses to different treatment options in CRC [11, 26, 27]. Although existing classification models have revealed the expression patterns of ICD-related genes to some extent, their limitation lies in the inability to fully integrate the complex biological mechanisms of ICD and its multiple roles in the tumor microenvironment.

Ferroptosis is an iron-dependent cell death mechanism characterized by the accumulation of lipid peroxides in the cell membrane [28]. In recent years, the importance of ferroptosis in CRC research has increasingly emerged. Ferroptosis has been reported to release damage-associated molecular patterns (DAMPs) and lipid metabolites that modulate the immune system, confirming its role as analogous to that of ICD [29]. Studies have revealed that in tumor groups enriched with ICD characteristics, ferroptosis-related marker genes exhibit elevated expression levels [30]. This correlation suggests that tumors with high ICD activity may also be more susceptible to ferroptosis, highlighting a potential overlap in the mechanisms driving these forms of cell death. Zhao et al. found that ferroptosis can enhance the ICD of tumor cells by inducing antigen exposure, thus reversing the immunosuppressive microenvironment [31]. The elevated expression of ferroptosis-related marker genes in tumors with high ICD characteristics underscores the potential for integrating these pathways in cancer therapy. In this study, we integrated the marker genes of the two modes of death, allowing for more accurate identification of CRC prognosis-related characteristics, which has not been reported in previous studies.

In this study, we analysed the expression patterns, mutation statuses, and associations with the immune microenvironment of F-ICD characteristics in colorectal cancer cohorts, via multidimensional algorithm models and single-cell data analysis. The expression patterns of F-ICD-related genes revealed distinct clusters within the CRC cohorts, indicating heterogeneity in how tumors may respond to ferroptosis and ICD-inducing therapies. This clustering could serve as a basis for classifying patients into subgroups with specific therapeutic susceptibilities. Our investigation of the immune microenvironment revealed significant associations between F-ICD characteristics and various immune cell populations. Tumors with high F-ICD activity showed increased infiltration of immune cells, such as NK cells, T cells, and a subset of B cells, indicating a more robust immune response. This enhanced immune presence suggests that F-ICD may not only directly kill tumor cells but also create a more immunogenic tumor microenvironment conducive to effective antitumour immunity. The integration of multidimensional algorithm models and single-cell data analysis provided deeper insights into the spatial and temporal dynamics of F-ICDs within the TME. Single-cell analysis, in particular, allowed us to map the interactions between F-ICD markers and specific immune cell subsets at a high resolution, revealing potential targets for combination therapies.

To further stratify the risk of CRC patients, we constructed a prognostic model using seven F-ICD genes (CD24, TMEM59, HADHB, CXXC5, FUS, INPP5D, and PLAC8) and investigated the expression profiles and prognostic implications of each gene. By incorporating these F-ICD genes into our model, we aimed to increase the accuracy of prognostic predictions and identify patients who may benefit from targeted therapies on the basis of their unique genetic profiles. Recent studies have demonstrated that inhibition of CD24 N-glycosylation leads to paclitaxel-induced ferroptosis. According to Zhang et al., HSBP1 interacts with FUS, playing a crucial role in suppressing hypoxia-induced ferroptosis in pancreatic cancer, suggesting a potential therapeutic target. The potential involvement of other genes in ferroptosis and ICD warrants further investigation. High expression levels of CXXC5, for example, are associated with poor prognosis, indicating its potential role as a biomarker for aggressive disease. Conversely, elevated expression of HADHB and CD24 was correlated with improved survival rates, suggesting their protective roles in CRC. Previous studies have shown that FUS promotes tumor progression and metastasis in various cancers and is an indicator of poor prognosis [32–34]. However, our findings in CRC suggest a more nuanced role for FUS. Despite its high expression in tumors, FUS serve as an indicator of favour survival in CRC, suggesting a seemingly counterintuitive scenario. Chen et al. indicated that the recruitment of FUS promotes the occurrence and metastasis of CRC [35]. FUS may promote metastasis by facilitating processes such as epithelial-to-mesenchymal transition (EMT) or regulating cell migration and invasion pathways. On the other hand, its role as a favorable prognostic marker may be related to its involvement in the regulation of tumor suppression pathways or immune modulation, which could lead to enhanced immune response and improved survival outcomes in certain contexts. Future research should aim to delineate the conditions under which FUS acts as an oncogene or tumor suppressor in CRC, thereby informing the development of more precise and effective therapeutic interventions.

There were several limitations in this study. One major limitation was the reliance on retrospective data, which may not fully capture the dynamic and heterogeneous nature of CRC progression and the role of F-ICDs in real-time. Additionally, our study utilized primarily in vitro models and bioinformatic analyses to explore the role of F-ICDs in CRC. While these approaches are valuable for identifying potential mechanisms and correlations, they may not fully recapitulate the complexity of tumor biology in vivo. Although our analysis suggests that certain genes may influence immunogenicity and TME composition, we did not experimentally assess key immune features such as dendritic cell activation, T cell priming, or immunological memory formation. Future research should incorporate functional immunological assays, such as ELISPOT, cytokine profiling, under conditions of ferroptosis or ICD induction.

Conclusion

Overall, the development of this F-ICD gene-based signature represents a significant advancement in the field of CRC prognostics. By providing a more nuanced understanding of patient risk and informing treatment decisions, this signature holds promise for improving patient outcomes and advancing the move towards more personalized and precise cancer care.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: X.L.; methodology: F.Z.; formal analysis: H.L.; software: F.Z.; writing—original draft preparation: J.L.; writing—review and editing: H.L.; supervision: F.Z.; funding acquisition: Y.W. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by grants from the Fundamental Research Program of Shanxi Province (202203021222418), the Fundamental Research Program of Jincheng People’s Hospital (JSY-2024D008), and the Research Program of Shanxi Provincial Health Commission (2023061).

Data availability

The raw data are available from the corresponding author for various reasons. And the datasets generated and analyzed during the current study are available in the TCGA (https://portal.gdc.cancer.gov/) and GEO (https://www.ncbi.nlm.nih.gov/geo/) database.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Clinical trial number

Not applicable.

Competing interests

The authors declare no competing interests.