An integrated machine learning and mendelian randomization approach identifies SERPING1 as a prognostic biomarker associated with CD8 + T-cell infiltration in DLBCL

Abstract

Background

SERPING1, which encodes the C1 inhibitor (C1-INH) of the complement system, and plays a key regulator in regulating inflammatory responses and immune homeostasis. SERPING1 is downregulated in various disease, this downregulation occurs through the body’s negative feedback resulting from the overactivation of the complement system in diseases such as infections and acute inflammatory responses. Additionally, SERPING1 is vital for tumor immunomodulation. Diffuse large B-cell lymphoma (DLBCL) is a common and aggressive type of non-Hodgkin lymphoma. More treatment options are becoming available for this disease. However, some patients still experience recurrence or even disease progression during treatment. Consequently, elucidating the molecular underpinnings of DLBCL’s malignant behavior and identifying novel prognostic markers and therapeutic targets are paramount for improving patient outcomes.

Methods

This investigation utilized multi-omics integration combined with machine learning algorithms to identify CD8 + T cell-associated hub genes and clarify their roles in DLBCL pathogenesis. Data were obtained from public repositories including the Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA). Utilizing multi-omics and machine learning techniques such as differential analysis, WGCNA, feature selection, scRNA-seq analysis, molecular docking, and Mendelian randomization, we identified four key hub genes (MAFB, TMEM176A, SERPING1, and C1QB) linked to CD8 + T cells, highlighting SERPING1 as central. They have potential value in the diagnosis, prognosis, and immunotherapy of DLBCL.

Results

Analysis of six GEO datasets (GSE56315, GSE25638, GSE12453, GSE12195, GSE32018, GSE83632) and TCGA-DLBC RNA-sequencing data, following batch correction, revealed 209 differentially expressed genes (DEGs). CIBERSORT revealed significant immune infiltration disparities, with CD8 + T cells notably enriched in DLBCL. WGCNA identified the yellow module (151 genes) as strongly correlated with CD8 + T cell infiltration, yielding 40 core DEGs upon intersection with DEGs. Functional enrichment analysis highlighted their involvement in chemokine signaling and complement cascades. A machine learning framework utilizing 12 algorithms identified eight key genes: ANKRD22, C1QB, C1QC, CD163, MAFB, SERPING1, TMEM176A, and TNFSF13B. The combination of Least Absolute Shrinkage and Selection Operator (LASSO) with Linear Discriminant Analysis (LDA) demonstrated the highest diagnostic efficacy, achieving an area under the curve (AUC) of 0.799. Four hub genes (MAFB, TMEM176A, SERPING1, and C1QB) were identified from the eight key genes through the intersection of eight feature selection methods and validated using protein-protein interaction networks and receiver operating characteristic curve analysis, achieving an AUC greater than 0.9. SHapley Additive exPlanations analysis underscored the predictive dominance of MAFB, while a prognostic nomogram incorporating these genes demonstrated high accuracy (1-/3-/5-year calibration). scRNA-sequencing revealed hub gene enrichment in non-classical monocytes, and CellChat implicated their roles in intercellular communication via the APP/GALECTIN pathways. Molecular docking and dynamics confirmed the stable binding of SERPING1-doxorubicin (ΔG = − 7.6 kcal/mol), supported by the root mean square deviation/Rg stability.

Conclusion

Summary-data-based Mendelian randomization analysis identified a closed positive genetic association between SERPING1 overexpression and DLBCL incidence risk (P < 0.05), with no statistical evidence of horizontal pleiotropy, suggesting that SERPING1 overexpression may be a potential risk-related factor for DLBCL onset. SERPING1 is significantly upregulated in DLBCL tissues, suggesting that its expression is regulated in a disease-specific manner. SERPING1 exhibits a pattern characterized by “local selective activation and overall inhibition” through the unbalanced regulation of the complement system in the tumor microenvironment, acting as a pro-tumor hub gene in synergistic regulation with C1QB.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05340-2.

Introduction

DLBCL is a malignant B lymphocyte proliferation in the hematopoietic system, accounting for about 40% of non-Hodgkin lymphoma cases [1, 2]. Despite the increasing number of treatment drugs and methods, the treatment outcomes for DLBCL remain unsatisfactory, with over 30% of patients experiencing disease recurrence and chemotherapy drugs exhibiting poor efficacy [3]. Specifically, the prognosis is worse for patients with metastasis [4, 5]. The rising incidence of DLBCL, marked by its aggressive behavior and varied clinical outcomes, presents considerable challenges in both diagnosis and treatment.As the predominant non-Hodgkin lymphoma subtype, understanding its underlying molecular mechanisms is crucial for improving therapeutic strategies and patient management. Recent progress in genomic and transcriptomic technologies has enabled the the discovery of novel biomarkers and therapeutic targets, advancing personalized medicine in oncology [6–9].

The study highlights the crucial influence of the tumor microenvironment, especially immune cells, on the clinical outcomes of DLBCL. Immune evasion mechanisms, including immunosuppressive cell recruitment and activation, contribute to tumor progression and resistance to therapies [10]. Various studies have used high-throughput sequencing technologies to profile the immune landscape within DLBCL tumors, revealing infiltration patterns correlating with clinical outcomes [11]. However, our current knowledge still lacks comprehensive insights into how these immune cell populations interact with tumor cells and influence therapeutic responses.

To address these gaps, our study used a comprehensive approach that integrated multiple high-throughput datasets from public repositories(GEO, TCGA), consistent with the multi-omics and machine learning framework widely adopted in cancer biomarker research [7–9, 12, 13]. We conducted differential expression profiling and immune infiltration analysis using advanced computational techniques, such as WGCNA and machine learning algorithms, to identify CD8 + T cell-associated genes, which are key mediators of antitumor immunity.

Furthermore, conducting enrichment analyses including gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways will clarify immune regulatory pathways in DLBCL. Our methodology employs advanced techniques, such as single-cell RNA sequencing (scRNA-seq), to analyze the cellular heterogeneity of the TME. This approach allows for a more refined understanding of immune cell interactions and the identification of novel therapeutic targets [14].

This study aimed to advance DLBCL molecular understanding and develop predictive models that can guide clinical decision-making. We aimed to identify biomarkers that stratify patients by risk and immunotherapy response by examining the interaction between tumor cells and the immune microenvironment. This research seeks to augment the current corpus of scientific understanding and to provide insights that may facilitate the development of innovative therapeutic approaches for DLBCL [11, 15].

Materials and methods

Data acquisition

The study utilized data sourced from public databases. Transcriptome datasets were obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/), including six distinct DLBCL studies.

• GSE56315: 55 patients with DLBCL, 33 healthy controls, totaling 88 samples.

• GSE25638: 26 patients with DLBCL, 7 healthy controls, totaling 33 samples.

• GSE12453: 11 patients with DLBCL, 15 healthy controls, totaling 26 samples.

• GSE12195: 73 patients with DLBCL, 10 healthy controls, totaling 83 samples.

• GSE32018: 22 patients with DLBCL, 7 healthy controls, totaling 29 samples.

• GSE83632: 76 patients with DLBCL, 87 healthy controls, totaling 163 samples.

RNA-seq data processed via the STAR pipeline and clinical information for DLBCL were obtained from TCGA (https://portal.gdc.cancer.gov). The expression profile data in TPM format were extracted from 48 patient samples. The detailed information is provided in Table 1.

Table 1.

DLBCL datasets information

Accession	Platform	Organism	Sample type	Experiment type	DLBCL	Control	Total
GSE56315	GPL570	Homo sapiens	Tissue	Expression profiling by array	55	33	88
GSE25638	GPL570	Homo sapiens	Tissue	Expression profiling by array	26	7	33
GSE12453	GPL570	Homo sapiens	Tissue	Expression profiling by array	11	15	26
GSE12195	GPL570	Homo sapiens	Tissue	Expression profiling by array	73	10	83
GSE32018	GPL6480	Homo sapiens	Tissue	Expression profiling by array	22	7	29
GSE83632	GPL5175	Homo sapiens	Whole blood	Expression profiling by array	76	87	163
TCGA-DLBCL	—	Homo sapiens	Tissue	Expression profiling by high-throughput sequencing	48	0	48

The training set comprised GSE56315, GSE25638, GSE12453, and GSE12195, while GSE32018 and GSE83632 were used as the validation set to evaluate the model’s generalization capability. The experimental design and procedural framework of this investigation are clearly illustrated in Fig. 1.

Fig. 1.

Schematic representation of the comprehensive workflow employed in this investigation

Data preprocessing and normalization

First, expression matrices and platform annotation files for each dataset were downloaded, and gene names were annotated using Perl (version 5.30.0). Subsequently, in the R (version 4.3.3) environment, The limma software package was employed to perform background adjustment and data normalization across all six microarray datasets. The sva package was used to remove batch effects, and the limma package was utilized for normalization across the combined training and validation datasets.

DEGs analysis

Differential expression analysis on the combined training set utilized R packages: tidyverse, DESeq2, and ggplot2. The selection criteria required an absolute log fold change greater than 2 and an adjusted p-value less than 0.05.Principal component analysis (PCA), volcano plots, and heatmaps were employed to visualize the results.

Immune infiltration analysis

CIBERSORT is a computational technique for determining the cellular makeup of complex tissues, such as blood or tumor samples, using gene expression data.Using the CIBERSORT algorithm, we assessed the infiltration levels of 22 unique immune cell populations in each training set sample based on gene expression data. To better understand these quantified infiltration levels, we visualized the differential immune infiltration patterns between disease and control cohorts, as well as the complex interactions among immune cell types. For this purpose, we employed multiple R packages, including reshape2, ggpubr, ggplot2, and dplyr, to generate stacked bar plots, comparative analyses between groups, and correlation heatmaps of immune cell distributions.

WGCNA

WGCNA evaluates gene co-expression by analyzing the correlation coefficients of normalized gene expression levels, clustering genes with similar expression patterns into modules. This method enables investigators to examine the associations among gene co-expression networks and immune cell populations, which can reveal their biological significance. The WGCNA package facilitated the construction of a gene co-expression network, enabling the identification of key gene modules linked to CD8 + T cell infiltration. The main parameters were set as follows: soft threshold power (power) = 13, minimum module size (minModuleSize) = 30, and module merge height (mergeCutHeight) = 0. The modules with the strongest association with CD8 + T cell infiltration were identified by analyzing the relationship between module eigengenes (ME) and CD8 + T cell infiltration levels. Subsequently, the genes in these selected modules were cross-referenced with the previously determined DEGs. This process identified a core set of DEGs linked to CD8 + T cell infiltration patterns.

Functional enrichment analyses

GO enrichment encompasses three categories: Biological Process (BP), Molecular Function (MF), and Cellular Component (CC). Additionally, KEGG serves as a bioinformatics database commonly used to identify enriched metabolic pathways from gene lists. The clusterProfiler package in R was used to perform GO and KEGG pathway enrichment analyses on core differentially expressed genes related to CD8 + T cells. Significance was determined by an adjusted p-value < 0.05 and a q-value < 0.25.

Gene set variation analysis (GSVA) evaluates the enrichment of predefined gene sets across samples without supervision. The GSVA algorithm was employed to compute pathway enrichment scores for individual samples, utilizing KEGG pathway gene sets as the reference database. Differential pathway analysis was conducted using the limma package to identify pathways with significant differences in enrichment scores (p < 0.05) between patient cohorts categorized by high and low CD8 + T cell infiltration.

Gene set enrichment analysis (GSEA) was performed to investigate the pathways differentially enriched between groups with high and low CD8 T cell infiltration. The clusterProfiler package conducted GSEA on groups with high and low CD8 + T cell infiltration, using the log2FC ranking of gene expression differences in disease samples from the training set. The c2.cp.all.v2022.1.Hs.symbols.gmt set was used as the reference gene set. The selection criteria mandated an adjusted p-value below 0.05 and a q-value under 0.25.

Feature gene selection based on machine learning

Core differential genes linked to CD8 + T cells were utilized to develop a machine learning framework incorporating 12 algorithms [7–9]. These include regularization techniques (LASSO, Ridge, Enet), generalized linear models (stepwise GLM, glmBoost, plsRglm), ensemble learning methods (RF, GBM, XGBoost), and pattern recognition models (SVM, LDA, NaiveBayes). Recursive feature elimination was used for the feature ranking. Subsequently, a predictive model was built using a stacking ensemble. The performance of all 113 predictive models was assessed using the area under the curve (AUC) metric under stratified 10-fold cross-validation. The optimal model was identified from which the most significant key feature genes were subsequently selected.

Analysis and validation of protein-protein interaction (PPI) networks

Feature genes were input into the STRING database (v12.0) to establish a PPI network with a confidence score threshold above 0.15. The differential expression profiles of these genes was validated in both the training and validation cohorts (|logFC| ≥ 2, adjusted p-value < 0.05), and their diagnostic performance was assessed using ROC curve analysis.

Hub gene selection based on multiple algorithm intersection

We integrated a machine learning framework using eight algorithms capable of feature selection, evaluated using 1000 iterations of 10-fold cross-validation: LASSO, Learning Vector Quantization (LVQ), Boruta, Bagged Trees, Random Forest (RF), Naive Bayes, SVM, and eXtreme Gradient Boosting (XGBoost). These algorithms are crucial in gene screening by limiting the focus to a more manageable gene set. Following this initial screening, key genes in DLBCL were further evaluated. Based on the classification performance, the genes selected by all eight algorithms were considered important hub genes for DLBCL.The R package “UpSet” was employed to visualize the interactions among the hub genes.

SHapley additive exPlanations (SHAP) interpretability analysis and prognostic model construction

SHAP analysis uses Shapley values from cooperative game theory to assign a SHAP value to each input feature, quantifying its contribution to the model’s prediction. This contribution can be either positive or negative. The magnitude and sign of the SHAP value directly indicate the strength and direction of the feature’s impact on the prediction. The SHAP algorithm was employed to interpret the predictions made by the optimally selected machine learning model. It systematically evaluated the relative importance of each hub gene in influencing the model’s decision-making process.

A prognostic nomogram model, utilizing four hub genes to predict progression-free interval outcomes, was developed on the Xiantao platform (https://www.xiantaozi.com/), The analysis was performed using R packages survival and rms, versions 3.3.1 and 6.3-0, respectively. Calibration curves for 1-, 3-, and 5-year intervals were plotted to evaluate the model’s predictive accuracy, accompanied by risk factor plots demonstrating the impact of each factor.

Correlation analysis of hub genes with immune cell infiltration

Immune cell infiltration involves the gathering of various immune cells, including T and B lymphocytes and macrophages, within tumor microenvironments, inflammatory sites, or other pathological areas. This biological phenomenon represents the host immune system’s defensive reaction against malignant cells or microbial invaders.

We utilized CIBERSORT for computational analysis to examine the associations between the expression patterns of four key genes and the relative abundances of 22 distinct immune cell subtypes. Spearman’s rank correlation coefficient was used for statistical evaluation, and results were visualized using lollipop plots and correlation network graphs.Statistical significance was defined by a threshold of p < 0.05.

scRNA-seq analysis

Single-cell RNA sequencing data from three normal controls and three DLBCL patients were sourced from the heiDATA database (https://heidata.uni-heidelberg.de). Quality control (200 < nCount_RNA < 5000, nFeature_RNA > 200, log10 Genes Per UMI > 0.8, mitochondrial gene proportion < 0.2), normalization, and identification of highly variable genes (HVGs, top 3000) were performed using the Seurat package. Dimensionality reduction was achieved using PCA, and cell clustering was conducted utilizing the initial 30 principal components with a resolution of 0.8.

Cell Communication Analysis: The CellChat package was used to analyze the ligand-receptor interaction network between cells.

Pseudotime Analysis: The Monocle package was used to construct cell differentiation trajectories based on highly dispersed genes (dispersion ≥ 0.3, average expression ≥ 0.05) and to analyze the expression dynamics of hub genes at different differentiation stages.

Normalization Parameters and Operational Details: The core function involves using the NormalizeData() function from the Seurat package to perform normalization. The key parameters are set as normalization.method = “LogNormalize” (specifying the normalization method) and scale.factor = 10,000 (defining the scaling factor).

Dimensionality reduction analysis was performed using PCA, a core linear dimensionality reduction method, as follows: The core function is RunPCA() from the Seurat package; the input consisted of 3000 highly variable genes selected after quality control and normalization (identified by VariableFeaturePlot() and SelectVariableFeatures() functions, with nfeatures set to 3000). A key parameter, npcs, was set to 50 to extract the first 50 principal components, followed by ElbowPlot() analysis of explained variance to select the first 30 principal components as effective dimensions for cell clustering (dims = 30). Other parameters were set to their default values, with scale = TRUE to center and scale the gene expression data, eliminating bias caused by differences in gene expression levels.

UMAP dimensionality reduction, a core nonlinear dimensionality reduction technique used for visualization, was performed using the RunUMAP() function from the Seurat package. The input dimensions were the first 30 principal components determined by PCA analysis (dims = 1:30). Key parameters were set as n.neighbors = 30 (number of neighbors) and min.dist = 0.3 (minimum distance). This parameter setting ensures that cell clusters in the UMAP plot have clear boundaries while preserving biological associations between cells. This provides a reliable visual basis for subsequent cell annotation. The output consists of two dimensions, UMAP1 and UMAP2, which are used for subsequent visualization analyses of all cell clusters, marker genes, and hub gene expression.

Cell clustering validation: First, we construct a cell neighbor graph using the FindNeighbors() function by inputting the first 30 effective dimensions from PCA (dims = 1:30); then perform cell clustering using the FindClusters() function, setting the resolution = 0.8, which was determined through multiple optimization iterations to balance the granularity of cell grouping and biological relevance, which ultimately yielded 14 cell clusters with clear biological significance.

Parameters for differential expression testing of hub genes enriched in non-classical monocytes are as follows. Core analysis function: Using the FindMarkers() function from the Seurat package for differential analysis between a single cell cluster and all other cell clusters. Differential testing method: Setting test.use = “wilcox” (Wilcoxon rank-sum test), a widely used non-parametric test method for single-cell transcriptome differential expression analysis, suitable for the distribution characteristics of single-cell expression data and effectively avoiding bias from the assumption of data normality. Key screening parameters include setting logfc.threshold = 0.5 (minimum log fold change) to select differential genes exhibiting substantial expression changes, and setting min.pct = 0.1 (minimum expression percentage), which requires that a gene be expressed in at least 10% of cells within either the target cluster (non-classical monocytes) or control clusters, thereby excluding low-expressed rare genes and reducing false positive results.

Molecular docking and molecular dynamics (MD) simulation

Potential therapeutic agents for the identified hub genes were discovered using the Comparative Toxicogenomics Database (CTD, https://ctdbase.org/). Based on a detailed analysis of drug-gene interactions, four clinically significant compounds were prioritized for further investigation: Panobinostat, Paclitaxel, Doxorubicin, and Topotecan.

Molecular Docking: Drug 3D structures (SDF format) were sourced from PubChem, while target protein structures were retrieved from the PDB database. Proteins were preprocessed using PyMOL (version 3.1.0) by removing the water molecules and ligands. Binding energies were calculated using molecular docking on the CB-Dock platform (https://cadd.labshare.cn/cb-dock2/php/blinddock.php). The docking results were used for subsequent MD simulations.

MD Simulations are powerful computational techniques that model the dynamic behavior of molecular systems to predict their physical and chemical properties, such as diffusion coefficients, reaction rates, and structural conformations. GROMACS (version 2023) software was used to perform a 100ns MD simulation of the optimal binding conformation. The AMBER99SB-ILDN and GAFF force fields, along with a TIP3P water model, were used. The simulation process comprised energy minimization, NVT/NPT equilibration, and production dynamics. The structural integrity of the molecular complex was evaluated by analyzing the root mean square deviation (RMSD), root mean square fluctuation (RMSF), radius of gyration (Rg), and hydrogen bond count.

Summary-data-based Mendelian randomization (SMR)

SMR is a method used to infer causal relationships between genetic variants and traits through the application of aggregated data from genome-wide association studies (GWAS)。.

SMR analysis is an SNP-centric Mendelian randomization method that includes the HEIDI test to evaluate instrumental variable heterogeneity. This method uses SNPs as genetic instrumental variables. To investigate the potential causal association between gene expression profiles and the risk of developing DLBCL, we performed SMR analysis using expression quantitative trait loci as instrumental variables, together with GWAS summary data from the FinnGen consortium. The HEIDI test, with criteria of p < 0.05 for nominal significance and HEIDI p > 0.05 for non-significant pleiotropy, was used to evaluate horizontal pleiotropy. Finally, genes identified from the SMR analysis were intersected with hub genes exhibiting high expression levels identified in this study to determine the key genes potentially involved in DLBCL pathogenesis.

Statistical methods

Statistical analyses were conducted using R software (v4.3.3) and the Xiantao online tool (https://www.xiantaozi.com). The Wilcoxon rank-sum test was utilized for statistical analysis to assess differences between the two study groups.Spearman’s correlation coefficients were used to analyze the relationships among immune cell populations. Statistical significance was determined at p-values of < 0.05 (*), < 0.01 (**), and < 0.001 (***).

Results

Data preprocessing and differential analysis results

PCA results revealed that before removing batch effects, the four training datasets formed four distinct clusters, indicating significant batch effects. After removing batch effects, the clustering of samples became intermixed, indicating effective correction of batch effects (Fig. 2A and B). A total of 209 DEGs were identified (186 upregulated, 23 downregulated), and the volcano plot displayed the distribution of these genes (Fig. 2C). The heatmap revealed that the top 15 upregulated genes showed high expression in the disease group, whereas the top 15 downregulated genes exhibited low expression levels in the same group (Fig. 2D).

Fig. 2.

GSE56315, GSE25638, GSE12453, and GSE12195 were merged as the train dataset for DEGs analysis. A PCA plot prior to batch correction. B PCA plot following batch correction. C Volcano plot illustrating the merged training dataset. D Heatmap of the merged train dataset

Immune infiltration analysis results

CIBERSORT analysis successfully quantified the infiltration levels of 22 unique immune cell populations in all samples (Fig. 3A).The disease group exhibited significant differences in the infiltration proportions of 20 immune cell types compared to the control group (Fig. 3B). Notably, CD8 + T lymphocytes, activated memory CD4 + T cells, γδ T lymphocytes, as well as M0, M1 and M2 macrophage subsets, were significantly enriched in the disease group.

Fig. 3.

Analysis of Immune Infiltration in the Training Dataset. Stacked bar charts depict the relative infiltration proportions of 22 immune cell types in both groups. Bar plots compare immune cell infiltration levels between the two groups. A correlation heatmap illustrates the relationships between various immune cell types

The correlation analysis of immune cells revealed the following findings:

• The top three strong positive correlations identified were between activated CD4 + memory T cells and resting mast cells (r = 0.36), neutrophils (r = 0.31), and M1 macrophages (r = 0.28).

• The top three strong negative correlations observed were between γδ T cells and resting natural killer (NK) cells (r = − 0.63), resting mast cells and activated mast cells (r = − 0.59), and CD8 + T lymphocytes and resting memory CD4 + T cells (r = − 0.44).

WGCNA analysis results

Setting the soft-threshold power (β) to 13 resulted in a scale-free topology fit index (R²) of 0.9 for the network (Fig. 4A). Nine gene modules were identified (Fig. 4B). The yellow module had the strongest correlation with CD8 + T cell infiltration (correlation coefficient r = 0.39, P = 2 × 10⁻7), containing 151 genes (Fig. 4C). The module eigengene of this module was also highly correlated with CD8 + T cell eigengenes (r = 0.55, P = 2.6 × 10⁻13) (Fig. 4D). The intersection of genes from the yellow module and DEGs yielded 40 core genes associated with CD8 + T cells The analysis identified 40 key genes shared by the yellow module and the DEGs, which were functionally linked to CD8 + T cell biology(Fig. 5A).

Fig. 4.

WGCNA of the combined training dataset. A Evaluation of scale independence and mean connectivity patterns in the integrated dataset. B Hierarchical clustering dendrogram illustrating gene expression patterns in the integrated genomic dataset. C Heatmap visualization illustrating correlations between modules and immune cells in the integrated dataset. D Scatter plot analysis showing the association between CD8 + T-cell-related genes and MEyellow-module membership

Fig. 5.

Identification and functional characterization of DEGs related to CD8 + T cells through comprehensive bioinformatics analyses. A Venn diagram depicting the overlap between DEGs and results from WGCNA. B The top five significantly enriched GO terms are presented in a lollipop plot, categorized by BP, CC, and MF. C Visualization of the 14 most significantly enriched KEGG pathways using a lollipop plot. D GSVA compares CD8 + T cell populations with high-expression versus low-expression profiles. E GSEA displays the five most upregulated signaling pathways in CD8 + T cell populations with differential expression patterns. F GSEA results show the five most downregulated pathways in CD8 + T cell populations stratified by high and low expression levels

Functional enrichment analysis results

The 40 core genes showed significant enrichment in molecular functions (MF) like chemokine activity and cytokine receptor binding, biological processes (BP) such as granulocyte chemotaxis, and KEGG pathways encompass complement and coagulation cascades, cytokine receptor interactions, and chemokine signaling pathways (Fig. 5B and C).

GSVA analysis: The cohort with elevated CD8 + T cell infiltration exhibited significant enrichment of oncogenic pathways, including pathways associated with colorectal carcinoma progression, ERBB family receptor signaling pathways, and WNT signaling cascades. Conversely, the cohort with low CD8 + T cell infiltration demonstrated predominant activation of immunological pathways, particularly those related to antigen processing and presentation, as well as NK cell-mediated cytotoxic responses (Fig. 5D).

GSEA analysis revealed enrichment of the B cell receptor (BCR) signaling pathway in the high-infiltration group, while the low-infiltration group exhibited significant enrichment of the IL-12 and interferon-α/β signaling pathways(Fig. 5E and F).

Machine learning feature gene selection results

Out of the 113 predictive models developed, the LASSO with LDA ensemble algorithm exhibited the best predictive performance, which is in line with multi-omics machine learning pipelines for prognostic modeling in malignancies [8, 9, 13], achieving an average AUC score of 0.799 in both training and validation cohorts (Fig. 6A). This model selected eight feature genes: ANKRD22, C1QB, C1QC, CD163, MAFB, SERPING1, TMEM176A, and TNFSF13B. The area under the ROC curve on the training set was 1.000, and the AUC values on validation datasets GSE32018 and GSE83632 were 0.701 and 0.696, respectively (Fig. 6B–D).

Fig. 6.

Integrated machine learning approaches for gene screening. A Schematic representation of the ensemble machine learning framework. B displays the ROC curve, demonstrating the diagnostic efficacy within the training cohort. C ROC curve illustrating diagnostic accuracy in the GSE32018 validation dataset. D ROC curve illustrating predictive capability in the GSE83632 independent dataset

The LASSO + LDA model constructed in this study demonstrated excellent discriminative ability in the training set (AUC = 1), but its performance decreased markedly in the independent validation cohort (AUC ranged from 0.696 to 0.701), which is below the commonly accepted threshold of AUC > 0.8 for clinical diagnostic tools, indicating that the model’s cross-dataset generalization ability is insufficient and currently lacks immediate clinical applicability.

PPI network construction and validation results

The PPI network constructed from the STRING database (version and date specified in the Methods section) contained eight nodes and 23 edges (Fig. 7A). The volcano plot indicated that all eight genes exhibited upregulation in the disease group relative to the control group (Fig. 7B). In the training set, these eight genes were significantly differentially expressed between the disease and control groups, with higher expression in the disease group (Fig. 7C). All exhibited excellent diagnostic efficacy (AUC > 0.9) (Fig. 7D).In the validation set, except for CD163, the other seven genes exhibited significant differential expression and good diagnostic efficacy (Fig. 7E and F).

Fig. 7.

PPI analysis and validation results for eight candidate genes in the combined test dataset (GSE32018 and GSE83632). A The constructed PPI network illustrates molecular interactions among the eight genes selected by machine learning. B Volcano plot analysis confirmed that all eight genes exhibited significantly elevated expression levels in disease samples. C Comparative analysis demonstrated statistically significant differential expression patterns for these genes in the training dataset compared to control samples. D ROC curve analysis demonstrated diagnostic potential for the disease, with all eight genes in the training dataset achieving area under the curve (AUC) values above 0.9. E Similar to the training dataset, differential expression patterns were observed for these genes in the test dataset. F ROC curve analysis confirmed the diagnostic efficacy of these biomarkers in the independent test cohort

Multi-algorithm intersection selection of hub gene results

Through intersection analysis of eight machine learning algorithms (LASSO, LVQ, Boruta, Bagged Trees, Random Forest, Bayesian, SVM, and XGBoost), four key hub genes for DLBCL were finally identified from the eight key genes: MAFB, TMEM176A, SERPING1, and C1QB (Fig. 8A–I).

Fig. 8.

Dentification of hub genes in DLBCL using eight distinct machine learning methods: A LASSO regression analysis; B Linear Quantile Vector (LQV) method; C Boruta feature selection algorithm; D Bagged Trees ensemble technique; E Random Forest (RF) classification; F Bayesian probabilistic modeling; G Support Vector Machine (SVM) analysis; H Extreme Gradient Boosting (XGBoost) implementation; and I Upset plot visualization showing genes overlapping among the eight machine learning methods

SHAP analysis and prognostic model construction results

SHAP analysis revealed that in the SVM model, the average SHAP values of the four hub genes were MAFB (0.126), TMEM176A (0.090), SERPING1 (0.088), and C1QB (0.085), with MAFB contributing the most to the prediction (Fig. 9B and C). The results of the prognostic models developed using data from the TCGA database were consistent with those obtained from the training dataset.The prognostic nomogram model, developed using these four genes (Fig. 9D), demonstrated high predictive accuracy in the calibration curves for 1, 3, and 5 years (Fig. 9E). The risk factor chart (Fig. 9F) showed that deceased patients were predominantly in the high-risk score group, while survivors were mainly in the low-risk score group.

Fig. 9.

SHAP analysis and prognostic model construction. A Diagnostic ROC curves for ten machine learning models. B Histogram of SHAP analysis for four hub genes. C Beeswarm plot of SHAP analysis for four hub genes. D The nomogram exhibited a prognostic risk prediction model for hub genes. E The calibration curves demonstrated the model’s accuracy and reliability by showing alignment between predicted and actual survival rates at 1, 3, and 5 years. F The risk factor graph illustrated the risk scores and transcriptional profiles of the four pivotal genes: MAFB, TMEM176A, SERPING1, and C1QB

Analysis of the relationship between central regulatory genes and immune cell populations

Correlation analysis revealed an association between each hub gene and specific immune cell subpopulations.

MAFB showed a positive correlation with resting dendritic cells, M2 macrophages, neutrophils, and CD8 + T cells, while exhibiting a negative correlation with memory B cells, naive B cells, and resting NK cells (Fig. 10A).

Fig. 10.

Immune infiltration profiling of immune cell subtypes and four hub genes. A Lollipop plot demonstrating the association between the expression level of MAFB and immune cell infiltration patterns. B Lollipop plot presenting the association between the expression level of TMEM176A and immune cell infiltration patterns. C Lollipop plot showing the association between the expression level of SERPING1 and immune cell infiltration patterns. D Lollipop plot revealing the association between the expression level of C1QB and immune cell infiltration patterns. E Interaction network illustrating associations between various immune cell populations and the expression levels of the four hub genes

TMEM176A showed a positive correlation with the abundance of neutrophils, naive CD4 + T cells, γδ T cells, and CD8 + T cells, while it exhibited a negative correlation with memory B cells, naive B cells, and activated mast cells (Fig. 10B).

SERPING1 shows positive correlations with naive CD4 + T cells, resting mast cells, M0 macrophages, and CD8 + T cells. It demonstrated negative correlations with memory B cells, naive B cells, and resting NK cells, as illustrated in Fig. 10C.

Statistical analysis identified significant positive correlations between C1QB expression levels and the infiltration of resting dendritic cells, activated natural killer cells, plasma cells, and CD8 + T lymphocytes.Conversely, significant negative associations were observed with memory B lymphocytes, naive B cells, and quiescent CD4 + memory T cell populations (Fig. 10D).

The network diagram illustrates relationships between immune cells and gene expression, with the thickness of the connecting lines indicating the strength of the correlation. Positive correlations are shown by orange lines, while negative correlations are depicted with green lines. Additionally, in the heatmap, darker colors corresponded to stronger correlations (Fig. 10E).

Single-cell analysis results

Cell annotation of the scRNA-seq data revealed 14 distinct clusters: basophils, CD8 + NKT-like cells, effector CD4 + T cells, megakaryocytes, memory B lymphocytes, CD4 + memory T cells, myeloid dendritic cells, naive B cells, naive CD4 + T lymphocytes, naive CD8 + T lymphocytes, NK lymphocytes, non-classical monocytes, plasmacytoid dendritic cells, and pro-B cells (Fig. 11A). All four hub genes exhibited significant overexpression in non-classical monocytes (Fig. 11E and F).

Fig. 11.

Cellular subpopulations delineated through scRNA-seq analysis of DLBCL specimens. A UMAP visualization with annotated cell populations identified in the DLBCL dataset. B Quantitative assessment of gene expression detection across distinct cell populations. C UMAP projection highlighting the most significantly expressed marker gene in each of the 14 identified cell populations. D Feature plots demonstrating expression patterns of the three most prominent marker genes in each cell populati(E)on. E Spatial distribution of expression profiles of four hub genes across all cell populations visualized through UMAP. F Feature plots illustrating the expression profiles of four hub genes across 14 cell populations

CellChat analysis identified robust signaling interactions among memory B cells, naïve CD4 + T cells, naïve CD8 + T cells, and non-classical monocytes (Fig. 12A–C). Non-classical monocytes acted as signal senders, mainly through pathways including APP and GALECTIN. As receivers, they communicated with other cells mainly through pathways, including MHC-II and CXCL (Fig. 12D–F).

Fig. 12.

Cell-cell communication analysis in the DLBCL single-cell RNA sequencing dataset. A Network visualization depicting intercellular interactions among 14 distinct cell populations. B Network visualization highlighting communication patterns among memory B cells, naive CD4 + T cells, naive CD8 + T cells, and non-classical monocytes. C A heatmap visualizes interaction frequencies among immune cell populations, with sender cells on the horizontal axis and receiver cells on the vertical axis. D Eatmap visualization showing the outgoing signaling profiles of 14 immune cell populations across 27 signalingH molecules, with cell populations arranged along the x-axis and signaling molecules along the y-axis. E Heatmap visualization illustrating the incoming signaling patterns of 10 immune cell populations responding to 27 signaling molecules. F Bubble plot visualization depicting intercellular communication mediated by ligand-receptor pairs within specific signaling pathways

Pseudotime analysis: The constructed cell differentiation trajectory revealed that memory B cells were mainly distributed at the end of pseudotime, whereas non-classical monocytes were mainly distributed at the beginning (Fig. 13A–C). The heatmap revealed that MAFB was highly expressed in late pseudotime, whereas SERPING1, TMEM176A, and C1QB were highly expressed in early pseudotime (Fig. 13D).

Fig. 13.

Presents a pseudotemporal analysis of the DLBCL single-cell RNA-sequencing dataset. A Trajectory plot depicting pseudotime ordering. B Overall distribution of 14 distinct cell types along the pseudotemporal axis. C Visualization of the 14 cell types within the pseudotime framework. D Heatmap illustrating the expression dynamics of MAFB, TMEM176A, SERPING1, and C1QB across pseudotime

Molecular docking and MD simulation results

Molecular docking results revealed that all four drug-target combinations exhibited stable binding, with binding energies as follows:

• MAFB-panobinostat: −7.3 (Fig. 14A).

• TMEM176A-paclitaxel: −7.8 (Fig. 14B).

• SERPING1-toxorubicin: −7.6 (Fig. 14C).

• C1QB-topotecan: −8.0 (Fig. 14D).

Fig. 14.

The molecular docking results of MAFB, TMEM176A, SERPING1, and C1QB with candidate therapeutic agents are shown. A Docking of the MAFB protein with panobinostat, with hydrogen bonds indicated by yellow dashed lines. B Docking of the TMEM176A protein with paclitaxel. C Docking of the SERPING1 protein with doxorubicin. D Docking of the C1QB protein with topotecan

Although this is strictly an "in silico" prediction, it clearly reflects the well-established mechanism of action of Doxorubicin.

The 100 ns MD simulation indicated the following:

RMSD: All complexes reached equilibrium in the later stages of the simulation, with the SERPING1-doxorubicin complex having the lowest and most stable RMSD value (Fig. 15A–D).

Fig. 15.

MD simulation, RMSD, and RMSF of the four hub genes and predicted drugs. A RMSD of MAFB with panobinostat. B RMSD of TMEM176A with paclitaxel. C RMSD of SERPING1 with doxorubicin. D RMSD of C1QB with topotecan. E RMSF of MAFB with panobinostat. F RMSF of TMEM176A with paclitaxel. G RMSF of SERPING1 with doxorubicin. H RMSF of C1QB with topotecan

RMSF: The SERPING1-doxorubicin complex exhibited the lowest amino acid residue flexibility, whereas the TMEM176A-paclitaxel complex demonstrated slightly higher flexibility (Fig. 15E–H).

Rg: The TMEM176A-paclitaxel complex exhibited a more compact structure after binding, whereas other complexes maintained stable structures (Fig. 16A–D).

Fig. 16.

MD simulation of Rg and hydrogen bond numbers of four hub genes and predicted drugs. A Rg of MAFB with panobinostat. B Rg of TMEM176A with paclitaxel. C Rg of SERPING1 with doxorubicin. D Rg of C1QB with topotecan. E Hydrogen bond numbers of MAFB with panobinostat. Hydrogen bond interactions between TMEM176A and paclitaxel. The quantity of hydrogen bonds formed between SERPING1 and doxorubicin. H Hydrogen bond numbers of C1QB with topotecan

Hydrogen bonds: All the complexes formed stable hydrogen bonds during the simulation (Fig. 16E–H).

SMR analysis results

The SMR analysis identified 709 genes associated with DLBCL risk (Fig. 17A). The intersection of these genes with the hub genes identified in this study revealed that SERPING1 was the only gene simultaneously meeting the criteria of “pathogenic gene” and “high expression in the disease group” (Fig. 17B and C). Further analysis indicated that high SERPING1 expression was significantly positively correlated with DLBCL risk (p < 0.05) (Fig. 17D).

Fig. 17.

Presents an SMR analysis that combines eQTL data with extensive GWAS data from the FinnGen database. A The Manhattan plot illustrates the genomic loci significantly associated with DLBCL. B The Venn diagram demonstrates the overlap between SMR-identified pathogenic genes and highly expressed hub genes in the DLBCL patient cohort. C A second Venn diagram shows the intersection between SMR-identified protective genes and genes with low expression in the disease group. D Scatter plot exhibited that SERPING1 was positively associated with DLBCL risk

Discussion

In this experiment, the GSE83632 dataset includes whole blood samples, while the other GEO datasets are tissue samples. Due to this difference, we did not include it in the core analysis of the training set. Instead, we used it together with GSE32018 as an independent validation set to verify the universality of the feature genes and diagnostic models obtained from the training set (tissue sample integration), rather than involving it in the selection of differential genes, immune infiltration core analysis, WGCNA module construction, and other key steps. We performed batch effect correction for the differences in sample types and applied dual robustness treatment for biological differences. Additionally, we supplemented the analysis with a sensitivity analysis of the separate GSE83632 whole blood samples, which showed results completely consistent with the expression trends of the training set and the original combined validation set (Supplementary Fig. 1 ). Finally, the dataset for immune infiltration analysis was limited in sample size, so the circulating CD8 + T cells in whole blood did not affect the core conclusion of CD8 + T cell infiltration.

SERPING1 encodes C1 inhibitor (C1-INH), a crucial regulator of the complement system essential for managing inflammation and maintaining immune balance [16]. SERPING1 is downregulated in various disease states, this downregulation occurs through the body’s negative feedback resulting from the overactivation of the complement system in diseases such as infections and acute inflammatory responses [17–19]. This study has validated the upregulated expression characteristics of SERPING1 in six GEO datasets (training and validation sets) and the TCGA-DLBCL dataset, and confirmed its specific high expression in non-classical monocytes of the DLBCL microenvironment at the single-cell level. These results are consistent and reliable, providing sufficient experimental evidence for the aberrant upregulation of SERPING1 in DLBCL. The SMR analysis combined with the HEIDI test found a significant positive genetic association between SERPING1 overexpression and DLBCL incidence risk (P < 0.05). Additionally, there is no statistical evidence indicating the presence of horizontal pleiotropy, suggesting that SERPING1 overexpression may be a potential risk-related factor for DLBCL onset. However, its distinct expression patterns and regulatory mechanisms in inflammatory diseases and tumorigenic diseases indicate that SERPING1 expression is disease-specific.

C1QB encodes the β chain of the complement C1q complex, which is a core factor that initiates the activation of the classical complement pathway. After the C1q complex binds to immune complexes, it triggers a cascade reaction in the classical complement pathway, mediating complement lysis, cytotoxicity, and immune cell recruitment [20]. This study confirms its significant upregulation in DLBCL, serving as a core hub gene, which may be closely associated with the infiltration of immune cells, including CD8 + T cells and macrophages.

The local selective activation of C1QB provides a ‘pro-inflammatory scaffold’ for the tumor microenvironment, promoting the infiltration of immunosuppressive cells [20, 21]. SERPING1 blocks the excessive activation of the classical complement pathway, preventing tumor cells from being cleared by the immune system. It can also reduce the increased vascular permeability of the tumor microenvironment by inhibiting the bradykinin system, promoting the stable formation of tumor vasculature, further supporting tumor proliferation [22]. The synergistic effect of the two genes creates a pro-tumor cycle of “recruiting immunosuppressive cells and blocking anti-tumor complement activation.” The “stage specificity” and “functional specificity” of complement activation by the two genes do not have a direct antagonistic relationship. Instead, they function complementarily to jointly shape a tumor microenvironment conducive to DLBCL progression, rather than acting through antagonistic effects of complement activation/inhibition [20–22].

Based on the experimental evidence of this study, including immune infiltration, PPI network, and single-cell expression analyses, the results suggest that both genes may be effective diagnostic biomarkers and may be correlated with immune infiltration patterns associated with tumor promotion.

MAFB is a transcription factor involved in macrophage differentiation and inflammatory responses [23–25]. However, TMEM176A is a transmembrane protein localized in lysosomes and is associated with inflammation and immune metabolism [26]. This study found that both genes are crucial for diagnostic and prognostic models and are significantly linked to immune cell infiltration, particularly CD8 + T cells. These results suggest that MAFB and TMEM176A may participate in shaping the DLBCL immune microenvironment by regulating the differentiation and function of specific immune cells.

Despite CD8 + T cells being the main antitumor effectors [27, 28], this study found their infiltration level to be notably higher in the DLBCL group compared to the control group. This paradoxical observation may be explained by the following reasons: (1) Functional exhaustion: These infiltrating CD8 + T cells may be in a state of functional exhaustion, characterized by impaired cytotoxic activity and inability to effectively kill tumor cells [29–31]. (2) Microenvironmental constraints, such as the elevated expression of genes like SERPING1, can impair the cytotoxic activity of CD8 + T cells, making them ineffective in eradicating tumor cells despite their readiness.

Emerging evidence indicates that the tumor microenvironment (TME) is crucial in tumor initiation, progression, and treatment response [32]. DLBCL, a malignancy arising from abnormal B lymphocyte development, can disrupt the TME by modifying cytokines that regulate cell proliferation, thus promoting tumorigenesis [33]. Research has shown that macrophages are the predominant immune-infiltrating cells in the DLBCL microenvironment, with CD8 + T cells also being significant [34]. In patients with DLBCL, elevated levels of LAG-3 and PD-1 on peripheral blood CD8 + T cells were associated with reduced CD8 + T cell functionality, leading to diminished tumor cell eradication and decreased patient survival [35]. Research on genes associated with CD8 + T cells in DLBCL is still scarce.One study reported activated CD8 + T cell infiltration in DLBCL, supporting research into CD8 + T cell-related gene expression in this disease [36].

Furthermore, studies have found that mutations in Tim-3 ligands and exhausted Tim-3 + CD8+ T cells correlate with a poor prognosis in DLBCL patients, highlighting the significance of CD8 + T cell-related gene expression. Zhang T et al. examined the genetic features of PD-1/PD-L1/L2 and CD73/A2aR pathways and their functions in the immunosuppressive microenvironment of DLBCL [37]. Single-cell RNA sequencing revealed that PD-1 and A2aR co-defined a dysfunctional CD8 + T cell subset, the abundance of which was significantly associated with poor prognosis [38]. These findings help explain the aforementioned paradoxical observations.

The results of GSVA/GSEA in this study (Fig. 5) exhibit that the high CD8 + T cell infiltration group is enriched in signaling pathways, including BCR signaling, whereas the low infiltration group is enriched in potentially beneficial pathways such as antigen presentation and NK cell cytotoxicity. This implies that the effectiveness of CD8 + T cell infiltration may hold greater significance than the sheer number of infiltrating cells.

Our study demonstrated that the four hub genes, specify gene names, were significantly overexpressed in non-classical monocytes.Analysis of cell communication indicated significant signaling interactions between non-classical monocytes, memory B cells, and T cells. Non-classical monocytes act as a central signaling hub, communicating with T cells via MHC-II and CXCL pathways and influencing other cells through the APP and GALECTIN pathways. Pseudotime analysis positioned non-classical monocytes at the trajectory’s start and memory B cells at its end (Fig. 13). The hub genes exhibited elevated expression during early pseudotime, indicating their potential critical role in the initial stages of disease onset by creating a foundational microenvironment for tumor progression.

The LASSO + LDA model, utilizing eight feature genes, exhibited strong predictive capabilities in both training and validation datasets (AUC > 0.7), suggesting its potential as a multi-gene diagnostic tool for DLBCL. The nomogram model selects the core feature genes of DLBCL, providing a preliminary idea for the molecular classification of DLBCL and laying a methodological foundation for subsequent model optimization.

Our SHAP analysis identified MAFB as the largest contributor to the the differentiation model between DLBCL patients and healthy controls (Fig. 9B), enhancing the interpretability and credibility of the model.

Molecular docking and MD simulations demonstrated a stable interaction between SERPING1 and doxorubicin, implying that doxorubicin may exert some of its therapeutic effects through this interaction.Although the main antitumor activity of Doxorubicin arises from its inhibition of DNA synthesis, we discovered a potential binding between SERPING1 and Doxorubicin, which suggests an interesting possibility. That is, the expression levels of SERPING1 may regulate the therapeutic effects of Doxorubicin through a non-classical mechanism. This provides a new perspective for studying chemotherapy resistance or personalized treatment strategies in DLBCL, but there is an urgent need to validate the existence of this interaction and its functional consequences through cellular and molecular-level experiments (such as surface plasmon resonance, co-immunoprecipitation, etc.). Other gene-drug pairs, including MAFB with panobinostat, TMEM176A with paclitaxel, and C1QB with topotecan, also demonstrated favorable binding affinities (Fig. 14).

Limitations of the study

The study’s conclusions rely on bioinformatics analyses of public databases and have not been independently experimentally validated. Although hypotheses were proposed, the specific regulatory mechanisms and molecular pathways involving these genes require further experimental validation. Furthermore, molecular docking and MD simulations are in silico predictions requiring validation through cellular and animal model experiments. The study has the following limitations:

1. One limitation is the lack of False Discovery Rate (FDR) correction in the analysis of correlations in immune infiltration. In subsequent studies, we will further validate the validity of all gene-immune cell associations through FDR correction, by increasing sample size, and other methods to strengthen the reliability of the research conclusions.

2. This study used SMR analysis to explore the association between SERPING1 and the risk of DLBCL. Although it validated the strength of the instrumental variable and provided suggestive evidence of no horizontal pleiotropy, it did not comprehensively validate all core assumptions of Mendelian randomization, thus failing to establish a causal relationship between SERPING1 and DLBCL risk. Additionally, due to the limitations of the public dataset information, this study did not include clinical confounding factors for adjustment. Future research should integrate clinical data with functional experiments to further validate the regulatory effect and causal relationship of SERPING1 in the occurrence and development of DLBCL. There is also a lack of direct experimental evidence to elucidate the specific molecular mechanisms by which SERPING1 regulates CD8 + T cell function. However, we can propose a hypothesis: SERPING1, as a key inhibitor of the complement system (C1 esterase inhibitor), may indirectly affect the recruitment, activation, or functional status of T cells by inhibiting the activation of the classical pathway of the complement system through its elevated expression. This hypothesis requires further experimental validation.

3. The scRNA-seq analysis in this study only included samples from 3 DLBCL patients and 3 healthy controls, resulting in a considerably small sample size. This small sample size limited the statistical robustness for identifying specific expression differences of hub genes in non-classical monocytes between disease and control groups, and cannot fully exclude the confounding effects of individual heterogeneity. Moreover, this study lacked validation in an independent and larger-scale single-cell RNA-sequencing cohort, which compromised the reproducibility and generalizability of the single-cell findings. In addition, spatial transcriptomics was not applied to confirm the precise cellular localization of hub genes in the DLBCL tumor microenvironment, so the spatial distribution and in-situ cell–cell interaction patterns of these key genes remain unclarified. Taken together, these limitations restrict the reliability and translational potential of the conclusion that hub genes are enriched in non-classical monocytes, and their exact regulatory roles in DLBCL initiation and progression need to be verified by expanded sample size and multi-modal spatial transcriptomic evidence in future work.

4. This research model is only aimed at distinguishing DLBCL from healthy controls and does not include other lymphoma subtypes or reactive lymphoproliferative samples. Therefore, it cannot address the differential diagnosis issues of DLBCL and similar diseases in clinical practice, revealing a significant gap with the clinical diagnostic needs in the real world. Our future plans include expanding sample types to align with clinical differential diagnosis needs, optimizing model features and structure to enhance generalization ability, incorporating multi-dimensional clinical features to improve model practicality, and conducting validation of clinical samples to evaluate the model’s performance in real-world clinical settings, thereby further enhancing clinical applicability.

Future research directions and conclusion

Future studies should verify the expression patterns of the four hub genes in independent clinical samples using quantitative real-time PCR, Western blotting, and immunohistochemistry. Furthermore, cell models, including DLBCL cell lines and CRISPR/Cas9 gene editing, should be used to explore the specific molecular mechanisms of genes, including SERPING1, in DLBCL initiation and progression. The impact of combining SERPING1 or complement system inhibitors with current chemotherapy or emerging immunotherapies, particularly chimeric antigen receptor T-cell therapy, merits thorough clinical investigation.

This study identified 8 core feature genes of DLBCL through machine learning algorithms. The constructed LASSO + LDA model can achieve a preliminary distinction between DLBCL and healthy controls. However, the model’s effectiveness has not reached the clinical application threshold and cannot solve the practical diagnostic problems in clinical settings. At this stage, it only provides preliminary insights into the molecular mechanisms of DLBCL. Future work needs to expand the clinical sample size, optimize model structure, and integrate multimodal features to further enhance the clinical applicability of the model.

In conclusion, this study successfully identified and validated SERPING1 as a highly promising biomarker closely related to the immune microenvironment of DLBCL, CD8 + T cell infiltration, and patient prognosis from multiple perspectives through a comprehensive and robust computational framework. SERPING1 exhibits a pattern characterized by “local selective activation and overall inhibition” through the unbalanced regulation of the complement system in the tumor microenvironment, acting as a pro-tumor hub gene in synergistic regulation with C1QB. It should be noted that SERPING1, as a potential therapeutic target, requires future systematic in vitro and in vivo experimental validation for its efficacy and safety.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Abbreviations

C1-INH

Inhibitor of the complement system

DLBCL

Diffuse large B-cell lymphoma

GEO

Gene Expression Omnibus

TCGA

The Cancer Genome Atlas

WGCNA

Weighted Gene Co-expression Network Analysis

DEGs

Differentially expressed genes

TME

Tumor microenvironment

GO

Gene ontology

KEGG

Kyoto encyclopedia of genes and genomes

scRNA-seq

RNA sequencing

CIBERSORT

Estimating Relative RNA Transcript Subsets for Cell-type Identification

BP

Biological process

MF

Molecular function

CC

Cellular component

GSVA

Gene set variation analysis

GSEA

Gene Set Enrichment Analysis

LASSO

Least Absolute Shrinkage and Selection Operator

SVM

Support vector machine

LDA

Linear discriminant analysis

AUC

Area Under the Curve

PPI

Protein-Protein Interaction

ROC

Receiver operating characteristic

LVQ

Learning Vector Quantization

RF

Random Forest

XGBoost

eXtreme Gradient Boosting

SHAP

SHapley Additive exPlanations

MD

Molecular Docking and Molecular Dynamics

SMR

Summary-data-based Mendelian Randomization

HEIDI

The heterogeneity in dependent instruments

TPM

Transcripts Per Million

FDR

False Discovery Rate

Author contributions

Jinhui Wang and Zhihui Li contributed to data curation, formal analysis, funding acquisition, methodology, validation, visualization, original draft writing, and investigation. Xinrong Zhan, Yanping Zhang, Zhongliang Wang, Pengtao Xing, and Mengmeng Liu contributed to data curation, formal analysis, funding acquisition. Mengyi Guo, Kaili Xu, and Haoyan Wang contributed to data curation, formal analysis, methodology, validation, and visualization.

Funding

The authors acknowledge receipt of financial support for both the conduct of this research and the preparation of the manuscript. This study received funding from the Henan Provincial Medical Science and Technology Joint Construction Program (LHGH20250856).

Data availability

The original data in this study are from these public databases.Researchers can access data pertinent to our study from the GEO database (https://www.ncbi.nlm.nih.gov/geo/) using the accession numbers GSE56315(https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE56315), GSE25638(https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE25638), GSE12453(https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE12453), GSE12195(https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE12195), GSE32018(https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE32018), and GSE83632(https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE83632).Downloaded and organized the RNA-Seq data from the TCGA database (https://portal.gdc.cancer.gov). Click the link in the Supplementary Materials section to view the TCGA cohort code for the DLBCL samples, CIBERSORT and CTD.The analysis also includes the following datasets and resources: heiDATA (https://doi.org/10.11588/DATA/VRJUNV), eQTL datasets for SMR analysis(https://download.gcc.rug.nl/downloads/eqtlgen/cis-eqtl/2019-12-11-cis-eQTLsFDR-ProbeLevel-CohortInfoRemoved-BonferroniAdded.txt.gz), FinnGen GWAS summary data(https://storage.googleapis.com/finngen-public-data-r12/summary_stats/release/finngen_R12_C3_DLBCL_EXALLC.gz), and STRING v12.0 (https://cn.string-db.org/). KEGG and GO enrichment analyses are both performed and visualized using the the Xiantao online tool (https://www.xiantaozi.com/).

Declarations

Ethics approval and consent to participate

This study complies with the data access policies of the databases used and mainly relies on public datasets; Thus, no further ethical review is required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.