Construction of a prognostic prediction model for diffuse large B-cell lymphoma patients based on ferroptosis-related LncRNAs Authors: Yirong Wang, Pengcheng Feng Published: 2026 ABSTRACT AbstractObjectiveTo investigate the prognostic value of a model based on ferroptosis-related long non-coding RNAs (lncRNAs) for patients with diffuse large B-cell lymphoma (DLBCL). Methods: Ferroptosis-related lncRNA expression matrices and corresponding clinical information for DLBCL patients were downloaded from the GEO public database (datasets GSE10846 and GSE11318). A prognostic risk model was constructed using the GSE10846 dataset through the following steps: identification of ferroptosis-related genes, Pearson correlation analysis with lncRNAs, Kaplan-Meier survival analysis, univariate Cox regression analysis, LASSO regression analysis, and multivariate Cox regression analysis. The accuracy of the model was evaluated using Receiver Operating Characteristic (ROC) curves. The model was subsequently validated using the GSE11318 dataset. Further systematic assessments of the model’s biological significance and clinical translational potential were conducted via immune infiltration analysis and drug sensitivity prediction. Results: A prognostic model comprising 9 lncRNAs was established. Time-dependent ROC curve analysis showed that the Area Under the Curve (AUC) values for 3-year and 5-year survival predictions were both greater than 0.75. Consistent results were obtained in the validation set, indicating good model robustness. Immune microenvironment analysis revealed that the high-risk group exhibited an immunosuppressive “cold tumor” phenotype, characterized by a significant reduction in the infiltration of effector immune cells such as cytotoxic cells and CD8⁺ T cells (all P < 0.001). Drug sensitivity analysis indicated that the high-risk group was more sensitive to PI3K/mTOR inhibitors (e.g., Dactolisib) but showed resistance to certain histone deacetylase inhibitors. Enrichment analysis suggested that these lncRNAs might influence DLBCL progression by regulating pathways such as protein degradation and RNA processing. Conclusion: The prognostic model constructed based on ferroptosis-related lncRNAs demonstrates considerable reliability. It not only effectively predicts patient survival but also reflects tumor immune status and drug response characteristics, showing potential as an auxiliary tool for prognosis assessment and personalized treatment in DLBCL.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-026-05145-3.Keywords: Diffuse large B-cell lymphoma, Ferroptosis-related lncRNAs, Prognostic model, Immune microenvironment, Drug sensitivity Introduction Diffuse large B-cell lymphoma (DLBCL) is the most common subtype of non-Hodgkin lymphoma (NHL). While most DLBCL patients achieve effective remission with standard rituximab combined with chemotherapy, approximately 40% still have a poor prognosis. Tumor heterogeneity is a crucial factor contributing to this significant disparity in treatment efficacy, and deciphering the molecular heterogeneity of tumors via multi-omics analysis combined with machine learning algorithms has become a research hotspot for constructing high-precision prognostic models in oncology. Recent studies have successfully revealed the prognostic significance of mitotic catastrophe and crotonylation heterogeneity in bladder cancer and clear cell renal cell carcinoma by these strategies [1, 2]. For DLBCL the molecular heterogeneity of ferroptosis-related regulatory networks remains largely unexplored, which limits the development of individualized prognostic tools. Therefore, establishing an individualized prognostic model based on ferroptosis-related lncRNAs is great important for guiding clinical therapy and filling the research gap of DLBCL molecular heterogeneity analysis. Ferroptosis is a novel, iron-dependent form of cell death distinct from autophagy and apoptosis [3]. Recent studies have shown that ferroptosis plays a significant role in the development and progression of various diseases, including cancer and brain injury [4–6], and its regulatory network is closely intertwined with the tumor immune microenvironment (TIME) remodeling—especially the formation of the immunosuppressive “cold tumor” phenotype characterized by insufficient effector immune cell infiltration [7–9]. Abnormal ferroptosis regulation can alter the release of damage-associated molecular patterns and the secretion of immunosuppressive cytokines, thereby shaping the immune infiltration landscape and further affecting tumor prognosis [10]. Consequently, investigating the expression patterns and prognostic value of ferroptosis-related lncRNAs is crucial for revealing the crosstalk between ferroptosis and TIME in diffuse large B-cell lymphoma (DLBCL), as well as optimizing disease treatment strategies based on immune and ferroptosis double targeting. In this work, we systematically explored the role of ferroptosis-related lncRNAs in DLBCL using bioinformatics approaches. A risk scoring formula was constructed to predict patient prognosis. These results will provide new perspectives for understanding the occurrence and development of DLBCL and offer theoretical guidance for clinical practice. More importantly, this ferroptosis-related lncRNA prognostic model is expected to be combined with the established International Prognostic Index (IPI) score to construct a more accurate and comprehensive prognostic assessment tool, which can make up for the limitations of traditional prognostic systems and provide incremental prognostic value for clinical stratification and individualized treatment of DLBCL patients. Materials and methods Sample screening and data acquisition The GSE11318 and GSE10846 datasets were downloaded from the GEO database (https://www.ncbi.nlm.nih.gov/geo/) for analysis. Samples with missing survival data or an overall survival (OS) of less than 50 days were excluded. LncRNA annotation information was obtained from GENCODE (https://www.gencodegenes.org/) and GeneCards (https://www.genecards.org/). Perl was used to convert probe IDs to gene names. A total of 113 ferroptosis-related genes (FRGs) validated in the literature were collected from the PubMed database (search strategy: (ferroptosis) AND (gene OR marker), up to December 2023). The inclusion criterion was clear association with ferroptosis regulation reported in at least two peer-reviewed papers. Univariate Cox regression analysis was performed on lncRNAs and FRGs to screen for molecules associated with survival. LncRNAs significantly correlated with FRGs (|R| ≥ 0.3, p < 0.05) were identified via correlation analysis for subsequent analysis. To explore the roles of lncRNAs in tumorigenesis and progression, functional annotation of candidate lncRNAs was performed using LncSEA (http://bio.liclab.net/LncSEA/) [11]. All data used in this study were sourced from public databases, eliminating the need for ethical approval. Construction and validation of the risk model Prognostic lncRNAs, corresponding survival data, and the “glmnet” package were used to perform Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression (10-fold cross-validation) to select the lncRNAs with the most significant prognostic value. Multivariate Cox regression analysis was then employed to construct the risk model. The risk score was estimated using the following formula: In this formula, ‘coef’ represents the coefficient obtained from multivariate Cox regression analysis. Risk scores were calculated for both the training and validation sets using this risk formula. Considering potential batch effects between different data sources, patients in each dataset were stratified into high-risk and low-risk groups based on their respective median risk scores. Kaplan-Meier curves and Receiver Operating Characteristic (ROC) curves were used to evaluate the model’s discriminative ability, including sensitivity and specificity. Calibration curves and decision curve analysis (DCA) were plotted to assess the model’s accuracy and clinical applicability. The independent dataset GSE11318 was used for validation. Construction of key LncRNA regulatory network and functional annotation To explore the potential functions of the aforementioned 9 key lncRNAs in DLBCL development, co-expression analysis and functional enrichment were performed based on the GSE10846 expression profile data. First, mRNAs significantly co-expressed with these 9 lncRNAs (|r| > 0.5, P < 0.05) were identified via Pearson correlation analysis. Subsequently, Gene Ontology (GO) enrichment analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis were conducted on this mRNA set using the R package clusterProfiler (FDR < 0.05) to systematically reveal potential biological processes and signaling pathways involved, thereby inferring the potential regulatory mechanisms of the related lncRNAs. Drug sensitivity analysis To assess the relationship between the prognostic model and treatment response, systematic drug sensitivity prediction analysis was performed. First, the oncoPredict R package was used to predict the in vitro half-maximal inhibitory concentration (IC50) of various drugs for the samples based on the GDSC2 database. Drugs with predicted IC50 values below 1 µM were selected for subsequent analysis. The predicted IC50 values were integrated with sample risk scores. Spearman correlation analysis was applied to test their association, and t-tests were used to compare differences between high- and low-risk groups (significance level P < 0.05, adjusted by FDR). Tumor immune microenvironment analysis To decipher the immune microenvironment features associated with the prognostic model, multi-level immune infiltration quantification analysis was performed. The ESTIMATE R package was used to calculate stromal score, immune score, and tumor purity for each sample. Furthermore, the ConsensusTME deconvolution algorithm was employed to quantify the abundance of specific immune cell subsets. After integrating all immune feature data with risk scores, the same statistical framework as for drug sensitivity analysis (Spearman correlation and t-tests) was applied to systematically evaluate their association with prognostic risk. Data analysis The primary software used in this study included R (version 4.2.2), Perl, and Python (version 3.9). The R package survival was mainly used for univariate Cox regression, LASSO regression, multivariate Cox regression, and ROC analyses. The R package glmnet was primarily used for LASSO regression analysis. The R package timeROC was used for ROC analysis. Python was utilized for data processing and analysis tasks such as merging survival data with expression matrices. Results Collection of survival-related LncRNAs We obtained 17,939 lncRNA gene symbols from GENCODE and 1,155 lncRNA genes from the annotation file GPL570. Univariate Cox regression analysis identified 414 lncRNAs significantly associated with survival. From the literature search, 113 FRGs were retrieved, of which 45 were survival-related. Through correlation analysis, 371 lncRNAs that were both survival-related and correlated with FRGs were obtained. LncSEA functional enrichment analysis indicated that these lncRNAs are involved in biological processes such as tumor proliferation, apoptosis, invasion, metastasis, and migration, and are significant for patient prognosis (Table 1; Fig. 1). Among them, GAS5, CRNDE, HOTTIP, BANCR, MALAT1, and AGAP2-AS1 were implicated in multiple biological processes. Table 1Role of lncRNAs in DLBCL progressionPhenotypeP ValueNumLncRNA IDPrognosis9.85E-2322LINC01016; THUMPD3-AS1; CASC15; MALAT1; AGAP2-AS1; LINC00668; CRNDE; FGF14-AS2; DANCR; GAS5; LINC00654; HOTTIP; TUSC7; UCA1; GHET1; ADAMTS9-AS2; LINC00628; LINC00115; HOTAIRM1; FOXD2-AS1; LINC00705; BANCRProliferation5.29E-1719CASC15; GAS5; CASC11; LINC00467; SOCS2-AS1; AGAP2-AS1; LINC00668; CRNDE; TUSC7; UCA1; HOTAIRM1; MALAT1; DLEU2; GHET1; LINC00628; BANCR; LINC00052 MIR100HG; HOTTIPInvasion1.45E-1112CASC15; GAS5; LINC00052; MALAT1; AGAP2-AS1; CRNDE; UCA1; DANCR; BANCR; HOTTIP; ADAMTS9-AS2; GHET1Migration4.25E-1112MALAT1; LINC00628; AGAP2-AS1; CRNDE; MIR100HG; ADAMTS9-AS2; GHET1; UCA1; BANCR; HOTTIP; LINC00052; GAS5Metastasis3.04E-108MALAT1; GAS5; AGAP2-AS1; GHET; BANCR; CASC11; HOTTIP; CRNDEApoptosis5.57E-1010LINC00467; MALAT1; AGAP2-AS1; CRNDE; TUSC7; UCA1; SOCS2-AS1; BANCR; GAS5; HOTTIP Role of lncRNAs in DLBCL progression Type and quantity of lncRNA corresponding to different DLBCL states LASSO regression analysis Based on the expression data of the 371 lncRNAs in the GSE10846 dataset, LASSO regression selected 9 lncRNAs for constructing the prognostic model(λ.min = 9, Supplementary Fig. 1,2). To verify the prognostic value and rationality of including each of the 9 ferroptosis-related lncRNAs in the final model, we first confirmed that all lncRNAs exhibited significantly different overall survival between high- and low-expression groups (log-rank test, all P < 0.05)(Supplementary Fig. 3). Subsequently, multivariate Cox regression analysis was performed, and the detailed results were presented as a forest plot (Supplementary Fig. 4). The forest plot showed that the combined prognostic signature of the 9 lncRNAs had extremely significant prognostic value for DLBCL patients (global Log-rank P = 3.7121e-16), with a good predictive accuracy (C-index = 0.72). Among the 9 lncRNAs, 5 lncRNAs (C12orf77, CHRM3-AS2, LINC00996, PDCD4-AS1, ZNF252P-AS1) had statistically significant independent prognostic value (all P < 0.05), while 3 lncRNAs (ELFN1-AS1, SERTAD4-AS1, SNHG19) showed a clear prognostic trend (P = 0.094, 0.101, 0.13, respectively). SNHG4 had a non-significant P value (P = 0.967) but an extremely small hazard ratio (HR = 1.03), which had no negative interference with the model. The marked differences in regression coefficients (0.0286 for SNHG4 to 2.9489 for ZNF252P-AS1) were consistent with the HR values in the forest plot, reflecting the differential magnitude of prognostic contribution of each lncRNA in the ferroptosis-related regulatory network: ZNF252P-AS1 (HR = 19.08, P < 0.05) was the core oncogenic lncRNA with the most prominent independent prognostic effect, while SNHG4 (HR = 1.03) had a weak individual effect but played a complementary role in the combined signature. The extremely significant global result of the multivariate model confirmed that the synergy between these lncRNAs effectively improved the prognostic discrimination ability of the signature, justifying their inclusion in the final model. Coefficients derived from multivariate Cox regression analysis were used to formulate the risk score. Risk scores were calculated for all samples. Using the median risk score as the cutoff, samples were divided into high-risk and low-risk groups. Kaplan-Meier survival curves indicated that high-risk patients had a significantly shorter overall survival (Fig. 2A). ROC curve results demonstrated that the model had good predictive performance (AUC = 0.74), indicating that the model is accurate, stable, and capable of predicting the overall survival of DLBCL patients (Fig. 2B). Reliability and accuracy analysis of the survival analysis model in the training set. A Survival analysis chart for high and low risk groups; B ROC curve analysis chart; C Calibration charts for predicted 1-, 3-, and 5-year survival rates versus observed survival in the training set; D Decision curve analysis chart Calibration curve analysis and decision curve analysis Calibration analysis for the training and validation sets showed that the predicted 1-, 3-, and 5-year survival rates were highly consistent with the actual survival rates in both sets, indicating the model’s reliability (Fig. 2C). Decision curve analysis for both sets revealed that the decision curve based on our model was distant from both the horizontal reference line (treat none) and the diagonal reference line (treat all), suggesting that our model has high clinical utility (Fig. 2D). Validation of the model’s prognostic value using GSE11318 The risk formula was applied to calculate risk scores for samples in the GSE11318 dataset to validate the model’s prognostic value. To minimize bias, samples in the validation set were stratified into high-risk and low-risk groups based on their own median risk score. The results from both ROC and Kaplan-Meier survival analyses were consistent with the conclusions from the training set (Fig. 3), further validating the model’s robustness. Reliability and accuracy analysis of the survival analysis model in the validation set. A Survival analysis chart for high and low risk groups; B ROC curve analysis chart; C Calibration charts for predicted 1-, 3-, and 5-year survival rates versus observed survival in the validation set; D Decision curve analysis chart Functional enrichment based on co-expression network reveals potential biological mechanisms of key LncRNAs Based on the mRNA set co-expressed with the 9 key lncRNAs (comprising 1076 genes, 4133 relationship pairs), GO and KEGG enrichment analyses revealed important biological functions and pathways potentially involved in DLBCL (Fig. 4). GO analysis indicated that these genes were significantly enriched in biological processes such as proteasome-mediated ubiquitin-dependent protein catabolic process (GO:0043161, P = 1.38 × 10⁻⁷), mRNA processing and spliceosomal complex assembly (e.g., GO:0050684, GO:0000380, P < 1.6 × 10⁻⁵), and positive regulation of DNA replication (GO:0045740, P = 3.21 × 10⁻⁵). Regarding cellular components, enrichment was primarily observed in nuclear speck (GO:0016607, P = 2.64 × 10⁻⁷), transport vesicle and its membrane (GO:0030133, GO:0030658, P < 5.4 × 10⁻⁵), and histone deacetylase complex (GO:0000118, P = 2.40 × 10⁻⁴), which are structures related to transcriptional regulation and intracellular transport. Functional enrichment analysis of mRNAs co-expressed with ferroptosis-related lncRNAs KEGG pathway analysis revealed that the co-expressed genes were significantly enriched in pathways such as vasopressin-regulated water reabsorption (hsa04962, P = 4.26 × 10⁻⁶), various virus infection-related pathways (e.g., Human cytomegalovirus infection, Hepatitis B, P < 5.6 × 10⁻⁵), and cancer-related pathways (e.g., Prostate cancer, Lipid and atherosclerosis, P < 2.1 × 10⁻⁴). Additionally, pathways related to cell communication and structural remodeling, such as cholinergic synapse, endocytosis, and regulation of actin cytoskeleton, were also significantly enriched. The results of enrichment analysis provide potential mechanistic directions for how the nine ferroptosis-related lncRNAs regulate DLBCL progression. These lncRNAs may modulate the expression and function of key ferroptosis regulators through processes including proteasome-mediated protein degradation, mRNA processing, PI3K/mTOR signaling pathway activation, and epigenetic modification (e.g., histone deacetylation). Furthermore, these regulatory processes are closely associated with the formation of the tumor immune microenvironment and drug response characteristics identified in our study, with the specific mechanistic links between lncRNA-mediated ferroptosis regulation and DLBCL pathobiology elaborated in the Discussion section. Tumor immune microenvironment analysis To systematically evaluate the relationship between the prognostic model and tumor immune status, we quantified the infiltration levels of 22 immune cell subsets in the high- and low-risk groups using the ConsensusTME algorithm. The analysis revealed that the high-risk group exhibited a broadly immunosuppressive state, characterized by a significant reduction in the infiltration of effector immune cells (Fig. 5 and Supplementary Table 1). Specifically, the infiltration level of cytotoxic cells (primarily including CD8⁺ T cells and NK cells) was lowest in the high-risk group (Pearson r = -0.434, P = 1.44 × 10⁻¹⁹), with an effect size (Cohen’s d) of -1.071, indicating a strong negative association. Concurrently, key immune subsets such as CD4⁺ T cells (r = -0.212, P = 2.12 × 10⁻⁵), dendritic cells (r = -0.238, P = 1.66 × 10⁻⁶), CD8⁺ T cells (r = -0.199, P = 6.76 × 10⁻⁵), and endothelial cells (r = -0.483, P = 1.72 × 10⁻²⁴) were all significantly decreased in the high-risk group (all P < 0.001). Furthermore, the cancer-associated fibroblast (CAF) score was also significantly lower (r = -0.409, P = 2.15 × 10⁻¹⁷). The ESTIMATE algorithm further confirmed that the high-risk group had higher tumor purity, and both immune and stromal scores were significantly lower than in the low-risk group (all P < 0.001). These results collectively reveal that high-risk DLBCL possesses a “cold tumor” immune phenotype, characterized by severely insufficient immune cell infiltration in the tumor microenvironment, suggesting potential poor response to immune checkpoint inhibitors. Forest plot for effect size assessment between high- and low-risk groups. Each dot represents the estimated Cohen’s d effect size, with horizontal lines indicating its 95% confidence interval. Dot colors are differentiated according to the effect size magnitude (Small, Medium, Large). The vertical dashed line (x = 0) denotes the reference line of no difference. The ranges for effect size magnitude (Small: |d| < 0.5, Medium: 0.5 ≤ |d| < 0.8, Large: |d| ≥ 0.8) are marked in the figure Drug sensitivity analysis To explore the predictive value of the prognostic model for treatment response, we predicted the IC50 values for 287 drugs in the DLBCL samples based on the GDSC2 database and selected 46 drugs with IC50 < 1 µM for in-depth analysis (Fig. 6 and Supplementary Table 2). The results showed that the IC50 of 17 drugs had a significant correlation with the risk score (FDR < 0.05). Among them, the high-risk group showed significantly higher sensitivity to several PI3K/mTOR pathway inhibitors, such as Dactolisib (dual PI3K/mTOR inhibitor, r = 0.267, P = 7.14 × 10⁻⁸, Cohen’s d = 0.557) and MK-1775 (Wee1 inhibitor, r = -0.337, P = 5.74 × 10⁻¹², Cohen’s d = -0.566). Comparison of key features between high- and low-risk groups. Box plots display the most significantly different features between the two groups. The box represents the interquartile range, the midline denotes the median, the black dot indicates the mean, and the scattered points show the sample distribution. The p-value, Cohen’s d effect size (interpreted as negligible, small, medium, large), and mean difference are annotated in the figure. * p < 0.05, p < 0.01, * p < 0.001 Additionally, the high-risk group exhibited significant sensitivity to Staurosporine (a multi-kinase inhibitor) (r = 0.393, P = 4.66 × 10⁻¹⁶, Cohen’s d = 0.918). Conversely, the high-risk group showed a trend of resistance to some epigenetic drugs such as LAQ824 (a histone deacetylase inhibitor) (r = -0.272, P = 4.15 × 10⁻⁸, Cohen’s d = -0.525). These results suggest that high-risk DLBCL may have a dependency on specific signaling pathways (e.g., activated PI3K/mTOR pathway), making it more sensitive to corresponding targeted therapies. The core reason for the enhanced sensitivity to PI3K/mTOR inhibitors such as Dactolisib is that the oncogenic ferroptosis-related lncRNAs (ELFN1-AS1, SNHG19, etc.) in the high-risk group are highly expressed and form a synergistic regulatory network, which abnormally activates the PI3K/mTOR signaling pathway through miRNA sponging and epigenetic modification. This pathway is a key node for the high-risk group to acquire ferroptosis resistance and maintain tumor cell proliferation; thus, targeted inhibition of PI3K/mTOR can effectively block the survival signal of tumor cells, reverse ferroptosis resistance, and induce ferroptotic death of DLBCL cells in the high-risk group. In addition, the PI3K/mTOR pathway activation in the high-risk group also leads to the abnormal proliferation of tumor cells with strong signal dependence, which further amplifies the killing effect of PI3K/mTOR inhibitors on tumor cells. Discussion With the deepening understanding of tumor biology, non-coding RNAs, especially long non-coding RNAs (lncRNAs), have been confirmed as key regulators in cancer development, progression, and therapeutic response. This study focuses on ferroptosis—a form of cell death closely associated with tumor suppression. By integrating bioinformatics analyses, we constructed and validated, for the first time, a prognostic prediction model for DLBCL based on ferroptosis-related lncRNAs. The model performed excellently in both the training set and an independent validation set, not only providing a new tool for prognostic stratification of DLBCL but also highlighting the importance of ferroptosis-related molecular networks in lymphoma biology. These findings not only enhance the biological interpretability of the model but also provide a theoretical basis for its future application in guiding personalized treatment (e.g., combination strategies of immunotherapy and targeted therapy), which is consistent with the core research goal of modern oncology to construct gene/lncRNA-driven prognostic models with both survival prediction and personalized treatment guidance value [12]. Most of the lncRNAs in our prognostic model have been implicated in tumor progression via epigenetic modification, miRNA sponging, and signal pathway regulation in various malignancies. Combined with the functional enrichment results of this study, the core regulatory mechanisms of ferroptosis, and our model’s risk coefficient characteristics, we propose a DLBCL-specific mechanistic hypothesis for how each lncRNA modulates ferroptosis to drive disease progression and shape the tumor immune microenvironment: C12orf77 (tumor-suppressive, negative coefficient): We speculate it inhibits the proteasome-mediated ubiquitin-dependent degradation of ferroptosis promoters (e.g., ACSL4, PTGS2), thereby upregulating their expression to enhance ferroptosis sensitivity in DLBCL cells, suppress tumor cell proliferation, and improve effector immune cell infiltration in the tumor microenvironment. CHRM3-AS2 (tumor-suppressive, negative coefficient): Based on the enriched histone deacetylation pathway and its known anti-apoptotic role, it may repress ferroptosis in DLBCL cells by regulating HDAC-mediated epigenetic activation of the ferroptosis inhibitor SLC7A11; its downregulation may further exacerbate the immunosuppressive “cold tumor” phenotype by promoting tumor cell migration and invasion [12]. ELFN1-AS1 (oncogenic, positive coefficient): It may sponge ferroptosis-related miRNAs (e.g., miR-183-3p, miR-497) to activate the PI3K/mTOR signaling pathway—an enrichment result of our study. PI3K/mTOR activation can induce Nrf2 nuclear translocation, upregulate downstream ferroptosis-inhibitory genes, and ultimately inhibit ferroptosis to drive DLBCL progression [13 – 15]. LINC00996 (tumor-suppressive, negative coefficient): Leveraging its known inhibitory effect on the JAK-STAT3 pathway and our immune microenvironment results, it may antagonize the JAK-STAT3-mediated ferroptosis-inhibitory pathway and regulate antigen presentation-related gene expression, thus enhancing ferroptosis sensitivity and reversing the “cold tumor” phenotype by increasing CD8⁺ T cell infiltration [16]. PDCD4-AS1 (oncogenic, positive coefficient): It may regulate the mRNA splicing and processing of the key ferroptosis regulator GPX4 (an enriched RNA processing pathway) to improve GPX4 mRNA stability and translation efficiency; meanwhile, its known role in stabilizing the tumor suppressor PDCD4 may coordinate the crosstalk between apoptosis and ferroptosis in DLBCL cells [17]. SERTAD4-AS1 (tumor-suppressive, negative coefficient): Combined with the enriched protein degradation pathway, it may inhibit the ubiquitin ligase-mediated degradation of GPX4. Its downregulation would lead to GPX4 inactivation and enhanced ferroptosis, and we speculate it fine-tunes ferroptosis levels to avoid excessive ferroptosis-induced immunosuppression in the tumor microenvironment [18]. SNHG19 (oncogenic, positive coefficient): It may activate the PI3K/mTOR/Nrf2 pathway via the miR-137/E2F7 axis (its known regulatory mode) to upregulate ferroptosis-inhibitory gene expression and suppress ferroptosis. Additionally, it may induce abnormal cancer-associated fibroblast (CAF) function to reduce immune cell infiltration and promote DLBCL progression [19]. SNHG4 (oncogenic, positive coefficient): Based on the enriched epigenetic modification pathway and its known miRNA-sponging ability, it may adsorb ferroptosis-related miRNAs (e.g., miR-204-5p, miR-377) to upregulate downstream ferroptosis inhibitors and repress ferroptosis; its regulation of cytoskeleton reorganization (a KEGG enrichment result) may further promote tumor cell migration and exacerbate immunosuppression [20–24]. ZNF252P-AS1 (oncogenic, highest positive coefficient): It may regulate the balance between ferroptosis and cell proliferation via the miR-324-3p/LY6K axis (its known regulatory mode). LY6K upregulation not only inhibits ferroptosis and significantly promotes DLBCL cell proliferation but also impairs cytotoxic T cell infiltration capacity, making it a key lncRNA in the formation of the “cold tumor” phenotype in DLBCL [24]. Notably, the nine lncRNAs in our risk score model form a synergistic regulatory network of ferroptosis in DLBCL, consisting of tumor-suppressive lncRNAs (C12orf77, CHRM3-AS2, LINC00996, SERTAD4-AS1) and oncogenic lncRNAs (ELFN1-AS1, PDCD4-AS1, SNHG19, SNHG4, ZNF252P-AS1). Oncogenic lncRNAs converge on inhibiting DLBCL cell ferroptosis through the PI3K/mTOR/Nrf2 pathway, epigenetic modification, and miRNA sponging, which collectively drives tumor progression and induces the immunosuppressive “cold tumor” phenotype. In contrast, tumor-suppressive lncRNAs antagonize this ferroptosis-inhibitory effect by regulating the protein degradation of ferroptosis-related factors and inhibiting pro-tumor signaling pathways (e.g., JAK-STAT3). An imbalance in this network—characterized by high expression of oncogenic lncRNAs and low expression of tumor-suppressive lncRNAs—leads to a marked enhancement of ferroptosis resistance in DLBCL cells, which is the molecular basis for the poor prognosis, prominent “cold tumor” phenotype, and increased sensitivity to PI3K/mTOR inhibitors observed in the high-risk group of our study. This synergistic regulatory network highlights the central role of ferroptosis-related lncRNAs in DLBCL pathobiology and forms a molecular bridge connecting ferroptosis resistance and the “cold tumor” immune phenotype in high-risk DLBCL, thus identifying potential combined targets for ferroptosis-based personalized therapy and immune checkpoint inhibitor sensitization. Our co-expression network and functional enrichment analyses further provide mechanistic hypotheses. The results indicate that these lncRNAs may be broadly involved in proteasome-mediated protein degradation, mRNA splicing regulation, DNA replication, and epigenetic regulation such as histone deacetylation. KEGG pathway analysis revealed enrichment in pathways related to viral infection, endocytosis, and cytoskeleton regulation, which are closely associated with the tumor microenvironment and cell communication. This implies that these ferroptosis-related lncRNAs may not be mere bystanders in cell death processes but rather key nodes in integrating intra- and extracellular signals, regulating global networks encompassing metabolism, stress response, and immune recognition. In DLBCL, they may affect tumor cell sensitivity to ferroptosis by interfering with these fundamental processes, thereby linking to chemotherapy resistance and poor prognosis. This study further revealed the clinical translational potential of the prognostic model through immune microenvironment and drug sensitivity analyses. Immune analysis indicated that high-risk DLBCL presents a typical “cold tumor” phenotype, characterized by significantly reduced infiltration of key immune effector cells such as cytotoxic cells, CD8⁺ T cells, and dendritic cells in the tumor microenvironment, along with lower immune and stromal scores —a phenotype that has been confirmed to be closely associated with poor prognosis in various hematological malignancies and solid tumors due to impaired anti-tumor adaptive immunity [7–8]. This feature suggests that high-risk patients may respond poorly to immune checkpoint inhibitors, aligning with clinical observations of immunotherapy tolerance in some DLBCL subtypes. The underlying molecular link between this typical “cold tumor” phenotype and the ferroptosis resistance mediated by the nine model lncRNAs is further elaborated below, and a mechanistic hypothesis for their causal association is proposed based on our study’s multi-omics analysis results. Linkage between ferroptosis resistance and the “cold tumor” immune phenotype in high-risk DLBCL Notably, recent studies have elucidated that the imbalance of ferroptosis regulatory networks is a key upstream driver of TIME polarization toward “cold tumor” status, as ferroptosis resistance can directly inhibit the maturation of dendritic cells and the recruitment of CD8⁺ T cells in the tumor microenvironment [9]. Combined with the ferroptosis regulatory mechanisms of the nine lncRNAs in our model, we propose a mechanistic hypothesis for the causal link between ferroptosis resistance and the “cold tumor” phenotype in DLBCL: the imbalance of the ferroptosis-related lncRNA regulatory network in high-risk DLBCL leads to enhanced ferroptosis resistance in tumor cells, which in turn shapes the immunosuppressive tumor microenvironment through three interconnected molecular pathways. First, ferroptosis-resistant DLBCL cells exhibit reduced release of damage-associated molecular patterns (DAMPs, e.g., ATP, HMGB1) caused by ferroptotic cell death; DAMPs are critical for activating dendritic cell maturation and cross-presentation of tumor antigens to CD8⁺ T cells, and their deficiency directly impairs anti-tumor adaptive immunity and leads to insufficient effector T cell infiltration. Second, our functional enrichment and drug sensitivity analyses confirmed abnormal activation of the PI3K/mTOR pathway in high-risk DLBCL—one of the core pathways mediating ferroptosis resistance. PI3K/mTOR activation not only upregulates ferroptosis-inhibitory genes (e.g., SLC7A11, GPX4) but also induces the secretion of immunosuppressive cytokines (e.g., IL-10, TGF-β) by tumor cells and cancer-associated fibroblasts (CAFs), further inhibiting the proliferation and cytotoxic function of infiltrating CD8⁺ T cells and promoting the formation of the “cold tumor” phenotype. Third, the nine lncRNAs in our model regulate epigenetic modification (e.g., histone deacetylation) and RNA processing, which not only modulates ferroptosis sensitivity in tumor cells but also represses the expression of immune checkpoint molecules (e.g., MHC-I) and chemokines (e.g., CXCL9, CXCL10) in tumor cells; the downregulation of these molecules impairs the recruitment of CD8⁺ T cells to the tumor microenvironment, ultimately solidifying the immunosuppressive “cold tumor” state. Collectively, ferroptosis resistance is not only a molecular feature of high-risk DLBCL but also a key upstream driver of the “cold tumor” immune phenotype, and the nine ferroptosis-related lncRNAs form a regulatory bridge between these two core characteristics of high-risk DLBCL. Guiding precise targeted therapy for high-risk patients Drug sensitivity analysis showed that the high-risk group exhibited significantly higher sensitivity to PI3K/mTOR pathway inhibitors (e.g., Dactolisib) and some multi-target kinase inhibitors (e.g., Staurosporine, Trametinib), while showing resistance to certain epigenetic drugs (e.g., LAQ824). The high sensitivity of the high-risk DLBCL group to PI3K/mTOR inhibitors has important and direct clinical implications, which can provide a clear direction for the personalized treatment of DLBCL patients stratified by our ferroptosis-related lncRNA prognostic model. Clinically, DLBCL patients with poor prognosis (high-risk group in this model) often have limited efficacy of conventional rituximab combined with chemotherapy, and the lack of specific targeted therapeutic targets is a major clinical dilemma. Our results confirm that the high-risk group has an intrinsic dependency on the PI3K/mTOR pathway, and PI3K/mTOR inhibitors (e.g., Dactolisib) can be used as a first-line or second-line targeted therapeutic option for this subgroup of patients, which is expected to improve the treatment response rate and overall survival of high-risk DLBCL patients. Providing a theoretical basis for the development of ferroptosis induction combined with PI3K/mTOR inhibition therapy. The PI3K/mTOR pathway is not only a key signaling pathway for tumor cell proliferation but also a core mediator of ferroptosis resistance in the high-risk group. Targeted inhibition of this pathway can not only directly kill tumor cells but also reverse ferroptosis resistance of high-risk DLBCL cells, making them sensitive to ferroptosis inducers. This suggests that the combination of PI3K/mTOR inhibitors and ferroptosis inducers (e.g., erastin, sorafenib) may be a potential effective treatment strategy for high-risk DLBCL, which can exert a synergistic anti-tumor effect and avoid the drug resistance caused by single-target therapy. Assisting in clinical treatment decision-making and prognostic re-evaluation: Our prognostic model can not only stratify the prognosis of DLBCL patients but also predict the therapeutic response of patients to PI3K/mTOR inhibitors based on risk stratification. In clinical practice, the model can be used as a companion diagnostic tool: for patients classified as high-risk by the model, PI3K/mTOR inhibitor-based targeted therapy can be prioritized; for low-risk patients with low PI3K/mTOR pathway activation, conventional chemotherapy can still be used as the main treatment, avoiding the unnecessary side effects and economic burden caused by blind targeted therapy. In addition, the therapeutic response to PI3K/mTOR inhibitors can also be used as a secondary prognostic evaluation index for high-risk patients, and patients with a good response to targeted therapy may have a better prognostic outcome, which enriches the clinical prognostic evaluation system of DLBCL. The high-risk group’s resistance to histone deacetylase inhibitors (e.g., LAQ824) is also closely related to the abnormal activation of the PI3K/mTOR pathway. The PI3K/mTOR pathway can regulate the expression and activity of histone deacetylases through epigenetic modification feedback, leading to the insensitivity of high-risk tumor cells to epigenetic drugs. This finding reminds clinicians to avoid using histone deacetylase inhibitors alone for high-risk DLBCL patients stratified by our model, which can reduce the occurrence of clinical ineffective treatment and improve the rationality of drug selection. Clinical translational potential of the model: combination with IPI score and incremental value As the most classic and widely used prognostic assessment tool for DLBCL in clinical practice, the International Prognostic Index (IPI) score is limited by its inability to reflect the intrinsic molecular characteristics of tumors (e.g., ferroptosis resistance, immune microenvironment status, signaling pathway activation) and has insufficient discriminatory power for patients in the intermediate-risk group (IPI 2–3), who often exhibit significant heterogeneity in prognosis and treatment response. In contrast, our ferroptosis-related lncRNA prognostic model is constructed based on the molecular characteristics of DLBCL, which can accurately reflect the biological behavior of tumors (e.g., ferroptosis resistance, “cold tumor” immune phenotype) and predict drug sensitivity (e.g., PI3K/mTOR inhibitor response). The complementary advantages of the clinical factor-based IPI score and the molecular characteristic-based lncRNA model lay a solid foundation for the construction of a combined prognostic tool with higher accuracy. For DLBCL patients stratified by the traditional IPI score (low, intermediate, high risk), further stratify the prognosis of each subgroup using our lncRNA model’s risk score (high/low), and reclassify patients into more refined prognostic subgroups (e.g., IPI intermediate-low/lncRNA low risk, IPI intermediate-low/lncRNA high risk, IPI intermediate-high/lncRNA low risk, IPI intermediate-high/lncRNA high risk). This reclassification can effectively resolve the prognostic heterogeneity of the IPI intermediate-risk group, which is a major clinical pain point. In addition, exploring the genetic crosstalk between different diseases and applying the research results of one disease to another has become an important way to advance precision diagnosis and treatment [25]. From “prognosis stratification” to “prognosis + treatment” dual guidance: The traditional IPI score only provides prognostic information and cannot guide the selection of specific therapeutic regimens for patients. Our lncRNA model not only predicts prognosis but also reflects the tumor immune microenvironment status (e.g., “cold tumor” phenotype) and drug sensitivity characteristics (e.g., high sensitivity to PI3K/mTOR inhibitors, resistance to histone deacetylase inhibitors) of patients. When combined with the IPI score, the new tool can realize one-step assessment of “patient prognosis + tumor molecular characteristics + targeted therapy response”, which can directly guide clinicians to formulate individualized treatment strategies (e.g., PI3K/mTOR inhibitor targeted therapy for IPI high-risk/lncRNA high-risk patients, immune checkpoint inhibitor combined with ferroptosis inducer for IPI intermediate/lncRNA high-risk patients with “cold tumor” phenotype). Providing a dynamic prognostic evaluation index for clinical treatment response: The IPI score is a static assessment tool based on baseline clinical factors and cannot be dynamically updated according to the patient’s treatment response and disease progression. The expression level of ferroptosis-related lncRNAs in our model can be detected in peripheral blood or tumor tissue samples during the treatment process, and the dynamic change of the lncRNA risk score can reflect the dynamic change of tumor biological behavior (e.g., reversal of ferroptosis resistance, transformation of immune microenvironment) after treatment. Combined with the IPI score, the tool can realize dynamic prognostic re-evaluation of patients during treatment, and timely adjust the treatment plan for patients with an increased lncRNA risk score, which is conducive to improving the overall therapeutic effect of DLBCL. This study also has certain limitations. Firstly, the analysis is entirely based on retrospective data from public databases and requires validation in future prospective clinical cohorts. Secondly, the precise biological mechanisms of the model, particularly how each lncRNA exactly regulates ferroptosis or interacts with known DLBCL driver genes, need further exploration through in vitro and in vivo experiments. Additionally, integrating the model with existing clinical staging systems (e.g., IPI) to evaluate its incremental prognostic value is a necessary step towards clinical translation. Conclusion In conclusion, we successfully developed and validated a robust prognostic signature for DLBCL based on ferroptosis-related lncRNAs. This work not only provides a promising molecular prognostic prediction tool but, more importantly, reveals the potential central role of ferroptosis-related lncRNAs in the pathobiology of DLBCL and exhibits significant clinical translational potential. This lncRNA model can form a highly complementary prognostic assessment system with the traditional IPI score: the combined tool can significantly improve the prognostic discrimination ability of DLBCL patients (especially for the intermediate-risk IPI subgroup), realize the integration of “prognosis prediction + tumor molecular characteristic analysis + targeted therapy guidance”, and provide important incremental value for clinical individualized treatment and dynamic prognosis monitoring of DLBCL. Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Acknowledgements We thank Changzhi Medical College for its encouragement and support, and the faculty of the School of Basic Medicine for their assistance. Author contributions Yirong Wang was contributions to Provision of study materials or patients, Collection and assembly of data, Data analysis and interpretation, Manuscript writing and Final approval of manuscript. Pengcheng Feng was responded to Conception and design, Administrative support, Manuscript writing, Final approval of manuscript. Funding There was no Funding. Data availability The data utilized in this study were obtained from publicly available datasets. All datasets employed have been referenced within the manuscript. The results analyzed in this study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Data Availability Statement The data utilized in this study were obtained from publicly available datasets. All datasets employed have been referenced within the manuscript. The results analyzed in this study are available from the corresponding author upon reasonable request.