HMGN2 deficiency drives DLBCL progression through impaired R-loop formation and PI3K-AKT hyperactivation

Abstract

Background

Diffuse large B-cell lymphoma (DLBCL) exhibits significant heterogeneity, with therapy resistance in Activated B-cell-like (ABC) and relapsed Germinal Center B-cell-like (GCB) subtypes being major challenges, with underlying drivers poorly understood.

Methods

We performed single-cell RNA sequencing (scRNA-seq) on de novo ABC-DLBCL, de novo GCB-DLBCL, and relapsed GCB-DLBCL specimens. Analyses included malignant cell subclustering, R-loop scoring, trajectory inference, and cell-cell communication. HMGN2 was functionally validated in vitro.

Results

We identified an aggressive, poor-prognosis B-cell subpopulation in de novo ABC and relapsed GCB-DLBCL, defined by significantly reduced R-loop formation.This phenotype was directly associated with the downregulation of High-Mobility Group Nucleosome Binding Domain 2 (HMGN2). Mechanistically, we validated that HMGN2 promotes R-loop formation; its loss derepresses IL4R expression, leading to hyperactivation of the oncogenic PI3K-AKT signaling pathway. Functional validation confirmed that HMGN2 knockdown in DLBCL cell lines enhanced proliferation and clonogenicity, whereas its overexpression was inhibitory. The tumor microenvironment in relapsed GCB tumors exhibited profoundly immunosuppressive features, including functionally impaired CD8⁺T cells and dominant inhibitory BTLA-TNFRSF14 interactions.

Conclusions

Our study identifies a potential HMGN2/R-loop/PI3K-AKT axis that may drive malignant cell-intrinsic fitness while shaping an immune-suppressive microenvironment. These findings position HMGN2 as a candidate regulator of DLBCL progression and a potential prognostic biomarker and therapeutic target to address treatment resistance.

Graphical Abstract

Supplementary information

The online version contains supplementary material available at 10.1186/s12967-026-08064-7.

Background

Diffuse large B-cell lymphoma (DLBCL) is the most prevalent subtype of non-Hodgkin lymphoma (NHL), comprising approximately 25–30% of all cases worldwide and exhibiting aggressive clinical behavior alongside substantial molecular heterogeneity [1, 2]. Currently, germinal center B-cell-like (GCB) and activated B-cell-like (ABC) are recognized as the two major subtypes of DLBCL based on their distinct genetic basis and responses to standard chemoimmunotherapy [3, 4]. Of note, ABC-DLBCL typically portends a worse prognosis attributable to constitutive NF-κB signaling and persistent antigen-driven B-cell receptor (BCR) activation [3, 4]. Although frontline immunochemotherapy regimens such as R-CHOP (rituximab combined with cyclophosphamide, doxorubicin, vincristine, and prednisone) have markedly improved outcomes, relapse manifests in up to 50% of patients, with particularly dismal 5-year overall survival rates below 50% in ABC-DLBCL and relapsed GCB-DLBCL cohorts. [5, 6] Such clinical disparities emphasize the imperative to elucidate the tumor microenvironment (TME), intratumoral cellular heterogeneity, evolutionary dynamics that underpin therapeutic resistance and unfavorable progression, and decode the underlying precise molecular programs particularly at single-cell resolution [7, 8].

It has been known that DLBCL oncogenesis is propelled by aberrant epigenetic and transcriptional regulatory networks, encompassing R-loop structures, a transient three-stranded nucleic acid configurations arising from transcription that, when dysregulated, engender genomic instability, DNA damage accumulation, and oncogenic signal transduction [9, 10]. Mounting evidence implicates R-loop dysregulation in the pathogenesis of B-cell malignancies, including DLBCL [9, 11]; nevertheless, its subtype-specific contributions to disease evolution, especially in relapse contexts where adaptive resistance may exacerbate R-loop-induced vulnerabilities, remain largely uncharted. Single-cell RNA sequencing (scRNA-seq) has transformed the interrogation of neoplastic ecosystems by unveiling subclonal architectures, cell-state transitions, and microenvironmental interactions that elude detection in bulk transcriptomic analyses [12]. Nevertheless, extant scRNA-seq investigations in DLBCL have largely focused on primary lesions or aggregate-level correlations, with scant incorporation of survival-stratified phenotypes, pseudotemporal trajectories, or TME crosstalk [13]. Although immune-infiltrative and tumor-microenvironment (TME) signatures have been described in DLBCL, most prior studies have not resolved subtype-specific (ABC vs GCB) cell–cell communication networks or the branch-dependent gene programs that govern shifts between proliferative and immune-suppressive states [14–16]. Recent single-cell and spatial datasets begin to map malignant B-cell states and TME ecosystems, yet an integrated trajectory-level view of interlineage signaling across molecular subtypes remains incomplete [17]. Moreover, how R-loop regulatory machinery (e.g., RNase H2, SETX, DDX helicases) intersects with PI3K–AKT signaling, an axis central to proliferation, invasion, and metabolic adaptation in B-cell lymphomas, has not been systematically profiled at single-cell resolution, limiting discovery of robust prognostic markers and actionable targets [18, 19].

In this study, we leveraged sc-RNA seq data to uncover the cellular ecosystems and potential heterogeneity in DLBCL. Specifically, through scRNA-seq profiling of de novo ABC-DLBCL, de novo GCB-DLBCL, and relapsed GCB-DLBCL specimens, we aim to map the single-cell atlas and survival-associated malignant B-cell subclusters, delineate pseudotemporal trajectories and branch-specific modules enriched in key oncogenic pathways, profile immune subsets with altered cytotoxicity in relapse, and then elucidate ligand-receptor axes fostering adverse phenotypes. By conducting in vitro experiments, we further validated that HMGN2 acts as a tumor suppressor by promoting R-loop formation and, consequently, restraining DLBCL proliferation. Our findings expose a novel regulatory network linking HMGN2, R-loop dynamics, and oncogenic signaling in DLBCL progression, identifying HMGN2 as a potent biomarker and a potential therapeutic target to overcome treatment resistance.

Methods

Single-cell transcriptomic data

Single-cell RNA sequencing (scRNA-seq) data was retrieved from a study of four patients with diffuse large B-cell lymphoma (DLBCL) whose data was deposited in the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo; GSE182436 [14]). Overall, this dataset includes three cases of activated B-cell-like (ABC) subtype de novo DLBCL and one case of germinal center B-cell-like (GCB) subtype de novo DLBCL. In these data, samples contain two sequencing libraries, including CD19+/CD20+ B cells and CD19-/CD20- non-B cells. In addition, single-cell data from two cases of GCB-subtype relapsed DLBCL were derived from a study that dissected intratumour heterogeneity of nodal B-cell lymphomas at the transcriptional, genetic and drug-response levels [20].

Bulk transcriptomic data

In addition to scRNA-seq of DLBCL, we also retrieved the bulk transcriptomic data of this disease from the Cancer Genome Atlas (TCGA) database, in which sequencing data of 47 DLBCL samples with survival information was included for the prognostic evaluation of the target gene. To validate the findings, public available sequencing data of DLBCL was obtained from GEO database (GSE87371 [20]), in which 121 samples with survival data were included.

Single-cell data quality control and cell annotation

Single-cell data processing was conducted using Seurat (version 4.1.1) [21]. Specifically, cells were removed if their mitochondrial content > 10%, hemoglobin content > 5%, or gene expression counts < 200 or >5000. Doublets were identified and removed using the DoubletFinder R package. Batch effects were corrected via the “RunHarmony” function in the harmony R package [22]. Prior to differential, enrichment and statistical analyses, data normalization (i.e., log-transformation), clustering, and dimensionality reduction analyses were performed to all scRNA-seq data using Seurat. For Doublet Detection, doublets were identified and removed using the DoubletFinder R package, the parameters were set to PCs = 1:20 and pN = 0.25, representing a simulated doublet rate of 25% based on the first 20 principal components. From the corrected expression matrix, 2000 highly variable genes were selected using the “FindVariableFeatures” function and principal component analysis (PCA) with the “RunPCA” function was conducted to select key components, with the top 20 principal components selected for downstream analysis. In addition, “FindClusters” function with resolution of 0.6 was applied for cell clustering and “RunUMAP” function was applied for the nonlinear dimensionality reduction. The identified cell clusters were further annotated based on the CellMarker 2.0 database and established cell-type marker genes.

Functional enrichment analysis

Differentially expressed genes (DEGs) for each cluster were identified using Seurat’s FindMarkers function. Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and HALLMARK pathways functional enrichment analysis performed for these identified DEGs using clusterProfiler R package (version 4.8.2) [23].

Identification of malignant and non-malignant B cells

Malignant and non-malignant B cells were distinguished using the infercnvpy package with default parameters. Cells from the tumor microenvironment (TME) served as normal controls (T cells were utilized as reference cells in the present study).

Identification of prognosis-associated malignant B-Cell subtypes

The Scissor algorithm (version 2.0.0) was employed to integrate survival outcomes from bulk TCGA-DLBC data with single-cell DLBCL profiles [24]. Survival status (i.e., deceased vs alive) in TCGA was determined from clinical follow-up data. The Scissor function was applied to epithelial cells with parameters: alpha = 0.003, family = “cox”. Cells with positive Scissor (+) status were associated with worse prognosis, whereas those with negative Scissor (-) status tend to exhibit favorable outcomes.

R-Loop scoring

The R-loop gene set was sourced from the R-loopBase database (https://rloopbase.nju.edu.cn/download.jsp)] [25]. R-loop scores for each malignant cell were computed using the AUCell function in the corresponding R package. AUCell ranks all genes in malignant cells by expression level, generating a gene ranking list where higher-expression genes rank foremost. Positions of R-loop gene set members within this list are identified for each malignant B cell, and the gene set activity score is calculated. Herein, the activity is quantified via the area under the curve (AUC) of the cumulative ranking of these genes, with a higher AUC value indicating elevated R-loop gene set expression in the respective malignant B cell.

Pseudotime analysis

Monocle (version 2.28.0) [26] was utilized to construct pseudotime trajectories based on gene expression profiles of malignant B cells. Following dimensionality reduction and cell ordering, all malignant B cells were projected and ordered onto trajectories with branching structures, where cells within the same branch are inferred to share similar cellular states. Branched Expression Analysis Modeling (BEAM) was further conducted to identify genes exhibiting branch-dependent expression patterns, thereby elucidating mechanisms underlying cell fate decisions.

Cytotoxicity gene set scoring in T-Cell subpopulations

The AUCell package was applied to score “Cytotoxicity” functional gene sets in T-cell subpopulations, with gene sets derived from the TCellSI (T Cell State Identifier) method [27]. Details of the T-cell “Cytotoxicity” gene set are provided in Table S1.

Cell-cell interaction analysis

CellChat (version 2.1.2) [28] was used to investigate potential interactions among different cell types. A merged Seurat object comprising malignant B cells and other TME cell types served as input. After creating the CellChat object, a reference database for secreted signaling pathways was created. The computeCommunProb and computeCommunProbPathway functions were then used to infer specific ligand-receptor interactions and communication probabilities between cell types, respectively.

Survival analysis

The optimal cutoffs for dichotomizing samples into high- and low-HMGN2 expression groups was determined using the “surv_cutpoint” function in the survminer R package. Results were visualized via Kaplan-Meier curves.

Cell culture

The human DLBCL cell lines Ly8 and U2932 cells were cultured in Dulbecco’s modified Eagle medium (DMEM-H) (Gibco) in a humidified atmosphere of 5% CO₂ at 37 °C.

Cell transfection

Ly8 and U2932 cells were seeded overnight to grow to 50% confluence and were then transfected with siRNA-HMGN2. siRNAs that target the human HMGN2 gene (siRNA-HMGN2) and nontargeting negative control siRNA (siRNA-NC) were purchased from Sangon Biotech Co., Ltd. (Shanghai, China). Lipid-based transient transfections were performed using Lipofectamine 6000 Transfection Reagent (Beyotime Biotechnology, Shanghai, China) according to the manufacturer’s instructions, and the culture medium was replaced 6 h later. After 48 to 72 h of transfection, quantitative reverse transcription qRT-PCR was used to assess transfection efficiency.

In addition, we constructed an HMGN2 overexpression plasmid. HMGN2-overexpressing was packaged using a plasmid system, and Lipofectamine 3000 (Thermo Fisher, USA) was used as the transfection reagent. Primer sequences are listed in Table S2.

qRCR

To carry out our qRT-PCR analysis, cells were harvested and dissolved in Trizol reagent, and then 200 μl of chloroform solution was added. Extracts were then centrifuged at 12,000 rpm for 15 min. An equal volume of isopropanol precooled at 4 °C was added to the supernatant and then centrifuged again at 12,000 rpm for 10 min. Finally, the supernatant was purified using ethanol, and we detected the concentration of RNA. Using a HiScript II Q RT SuperMix kit (AU341; TRANS, China) for reverse transcription, the expression of genes was determined using SYBR Green PCR Mix (RR420A; Takara, Japan) as follows. Primer sequences used in this study are listed in Table S3.

All experiments were performed in triplicate and analyzed by the 2^−ΔΔCT method using β-ACTIN as the reference gene.

CCK8

We used the CCK8 assay to detect the viability of cells in accordance with the manufacturer’s protocol. Cells transfected with si-HMGN2 and si-NC were seeded at a density of 5,000 cells/well in 96-well plates. Then, 10 μl of CCK8 solution (CCK-8; Biosharp, Shanghai, China) was added to the cells after 0, 24, 48, 72, and 96 h, and the cells were incubated for another 1 h at 37 °C. After 1 h incubation, absorbance was measured at 450 nm with a microplate reader (BD Biosciences, USA). These data were analyzed and visualized using Prism software.

EdU

5-ethynyl-2-deoxyuridine (EdU) assay (Beyotime, Shanghai, China) was used to determine the cell proliferative capacity, following the manufacturer’s protocol. In brief, the proliferating Ly8 and U2932 cells were treated with 10 μM EdU working solution at 37 °C for 2 h. The cells were then fixed with 4% paraformaldehyde for 15 min, permeabilized with 0.1% Triton X-100 for 10 min and stained with DAPI for 10 min. The images were taken with a fluorescence microscope (Olympus, Tokyo, Japan). A minimum of five images were randomly selected from each of the three biological replicates per group. The number of EdU-positive cells within these images was then analyzed using Image J software.

Colony formation assay

For colony formation assay, cells were incubated in a complete medium on plates for 14 days. The medium was replaced at an interval of 3 days. Cell colonies were subjected to fixation with paraformaldehyde, and subsequently to 20 min of staining with crystal violet (0.1%). Visually available colonies were quantified. The number of colonies was analyzed using ImageJ software.

Immunofluorescence staining for cyclin A2

Cells were fixed with 4% paraformaldehyde for 30 min at room temperature, permeabilized with 0.1% Triton X-100 for 10 min, and blocked with 10% fetal bovine serum in PBS for 1 h. Cells were then incubated with primary antibody (Anti-Cyclin A2 [EPR17351], ab181591, Abcam) overnight at 4 °C, followed by secondary antibody (Goat anti-rabbit IgG H&L (Alexa Fluor® 555), ab150078, Abcam) for 1 h at room temperature in the dark. Nuclei were stained with DAPI for 10 min. Coverslips were mounted with ProLong® Gold Antifade Mountant. Images were acquired using an Axiovision-le_se64_491_SP2 fluorescence microscope. The percentage of Cyclin A-positive cells was quantified to characterize cells in the S and G2 phases of the cell cycle.

R-loop dot blot assay

DNA samples were loaded onto an Amersham Hybond-N+ membrane (Solarbio, China). The membrane was UV crosslinked for 5 minutes and washed with PBST. After blocking with 5% non-fat milk, the membrane was incubated with the S9.6 antibody (1:1000, Abcam) overnight at 4 °C. Dot blots were then incubated with HRP-conjugated anti-mouse IgG for 1 hour and visualized using an imaging system (Bio-Rad, USA).

Western blot

Cells and tissues were lysed with RIPA lysis buffer containing 1% PMSF. The lysates were then centrifuged, and the supernatant was collected. The quantified protein supernatant was supplemented with 4×protein loading buffer proportionally, boiled for 10 min to denature the protein, and stored at −80 °C. Proteins were then separated by 10% SDS-PAGE and electrophoretically transferred onto polyvinylidene fluoride membranes where they were blocked with 5% skim milk and incubated with anti-HMGN2 (1:1,000, #10953–1-AP; Ptoteintech, USA), anti-β-Actin (1:1,000, #AC026; Abclonal, China), anti-AKT (1:1,000, #JE75-09; HUABIO, China), anti-Phospho-AKT (1:1,000, #SY28-05; HUABIO, China), anti-PI3-kinase(1:1,000, # SC0580; HUABIO, China) and anti-Phospho-PI3K (1:1,000, #AF3242; Affinity, USA) overnight at 4 °C. Next, the membranes were incubated with horseradish peroxidase-conjugated anti-rabbit IgG. Antigen-antibody complexes were then detected with enhanced chemiluminescence reagent. The resulting images were processed and analyzed using ImageJ software.

DRIP-qPCR

DNA was extracted from Ly8 and U2932 cells transfected with si-HMGN2, oe-HMGN2, or corresponding negative controls (si-NC/oe-NC), followed by overnight incubation at 37 °C in RNase H buffer with or without treatment with RNase H (Beyotime, Shanghai, China). RNA-DNA hybrids from the digested nucleic acids, either treated or untreated with RNase H, were immunoprecipitated using the S9.6 antibody (Abcam, USA) and protein A/G beads (MCE, China) at 4 °C overnight in IP buffer. The beads were then washed four times with IP buffer for 10 minutes at room temperature, and nucleic acids were eluted with elution buffer at 55 °C for 1 hour. The immunoprecipitated DNA was analyzed by qPCR using primers targeting the IL4R promoter region(Table S3); MYOD1 served as a negative control locus.

ChIP-qPCR

For DNA immunoprecipitation, a DNA ChIP Assay Kit (CST, USA) was utilized. Ly8 and U2932 cells transfected with Flag-tagged HMGN2 overexpression plasmid (oe-HMGN2) or empty vector control (oe-NC) were fixed with 1% formaldehyde and quenched with glycine (0.2 mol L^{− 1}). After sonication, the chromatin was directly incubated with anti-Flag beads overnight at 4 °C. Finally, the DNA was released from the immunoprecipitates for qPCR analysis. Based on genomic annotation of the IL4R promoter region, three pairs of ChIP-qPCR primers (IL4R-1, IL4R-2, and IL4R-3) were systematically designed around the transcription start site (TSS) to detect HMGN2 chromatin occupancy at different subregions of the IL4R promoter (Table S3).

Statistical analysis

Differences in continuous variables between examined groups were compared using the independent-samples Mann-Whitney U test (for two groups) or Kruskal-Wallis test (for three groups). Comparisons for categorical variables were conducted with chi-square tests. All statistical analyses were performed in R (version 4.0.5). Two-tailed p-values < 0.05 were considered statistically significant.

Results

Single-cell landscape

The workflow of this study is shown in Fig. S1. As shown in Fig. 1, a total of six single-cell RNA-seq samples (Fig. 1A) were included in the present study. Among them, three were de novo ABC-type DLBCLs, one was de novo GCB-type DLBCL, and the other two were relapsed GCB-type DLBCLs (Fig. 1B). After quality control, a total of 23,334 cells were retained and included for dimensionality reduction and unsupervised clustering analyses, which resulted in 19 clusters (Fig. 1C). Based on canonical marker expression, these clusters were annotated as four major cell types, including T cells, B cells, myeloid cells, and NK cells (Fig. 1D–E). Among them, B cells were the most abundant (n = 12,104), whereas NK cells were the least abundant (n = 222) (Fig. 1F). Cell-type proportions were comparable across samples, suggesting successful data integration (Fig. 1G). Using inferCNV, we calculated CNV score for each B cell (Fig. 1H) and defined 7,449 putative malignant B cells, with the highest fraction of malignant B cells found in the de novo ABC-type group (Fig. 1I).

Fig. 1.

Umap-based single-cell atlas. A Cells colored by sample. B Cells colored by clinical group. C Cells colored by cluster. D Cells colored by annotated lineage. E Selected marker genes per lineage. F Cell counts per lineage. G Lineage proportions per sample. H CNV score for each B cell. I Proportions of malignant vs. normal B cells in each group

Malignant B-cell analysis

With the identified malignant B cells, we then conducted clustering analyses and identified 13 subclusters (Fig. 2A). By integrating with bulk transcriptomics phenotypes associated with survival outcome (i.e., Scissor), we classified malignant subclusters into “worse-survival,” “good-survival,” and “background” groups (Fig. 2B) with their spatial distribution shown in Fig. 2C. Comparisons of malignant B-cell proportions across these three groups indicated that “good-survival” malignant subclusters were chiefly derived from de novo ABC-type and de novo GCB-type samples, whereas subcluster indicating “worse-survival” originated from both de novo ABC-type and relapsed GCB-type samples (Fig. 2D).

Fig. 2.

Malignant B-cell subclusters. A UMAP of malignant B cells colored by cluster. B UMAP showing Scissor-defined outcome-associated cells (red, worse; blue, good). C UMAP colored by clinical group. D Group-wise composition of “worse,” “good,” and “background” malignant subclusters. E UMAP of R-loop scores. F Boxplot of R-loop scores across clinical groups. G Boxplot of R-loop scores across malignant subclusters

Based on the defined 1,185 R-loop regulators, we calculated an AUCell-based R-loop score for each malignant B cell (Fig. 2E). We found that de novo GCB-type cases had the highest R-loop scores, whereas the score was lowest among relapsed GCB-type cases (Fig. 2F). Meanwhile, within “worse-survival” malignant subclusters, R-loop scores were significantly lower in relapsed GCB-type than in de novo ABC-type samples, a trend that appeared to persist in the “good-survival” and “background” groups (Fig. 2G). In addition, “worse-survival” malignant subclusters exhibited significantly lower R-loop scores than “background” malignant subclusters (Fig. 2G).

Pseudotime analysis of B cells

To explore the dynamic evolution and cell fate decision, we performed pseudotime trajectory analysis and identified five malignant B-cell states (Fig. 3A–D). As shown in Fig. 3, cells in “state 1” constitute a putative root and “state 2” progresses along the right trajectory after the branchpoint2, whereas “state 3” progresses along the left trajectory. Analysis of subcluster compositions across these states suggested a potential drift toward worse-survival phenotypes with pseudotime, as “state 3” contained a higher proportion of worse-survival cells (19%) compared to state 1 (8%) (Fig. 3C).Furthermore, BEAM identified top 50 branch-dependent genes that regulate differentiation from the pre-branch (State 1) to post-branch fates (Cell fate 1: State 2; Cell fate 2: State 3). Based on expression similarity, these genes were partitioned into four modules. As shown in the heatmap (Fig. 3E), gene expressions in Cluster 2 presented an overall increasing trend along the left trajectory after branchpoint 1 (Cell fate 2: State 3). The HALLMARK enrichment analysis indicated that genes in Cluster 2 were mainly enriched in Hypoxia, TNFα signaling via NF-κB, and IL6/JAK/STAT3 signaling pathways (Fig. 3F), suggesting that poorer survival outcomes might be associated with progressively reinforced proliferative capacity, metabolic reprogramming, and immune-evasion programs.

Fig. 3.

Trajectory analysis of malignant B cells. A Pseudotime coloring. B Coloring by normal B, worse-survival malignant B, and background malignant B. C Coloring by state. D Coloring by clinical group. E Heatmap of dynamic gene expression along pseudotime. F HALLMARK enrichment for cluster 2 genes

Re-clustering of immune cells

Given the intricated relationship between tumor progression and the tumor microenvironment (TME), we further re-clustered T cells. Through dimensionality reduction and clustering analyses, we identified four T-cell subsets based on marker genes annotation, including naïve T, CD4⁺ T, CD8⁺ T, and regulatory T (Treg), which present across all the three clinical groups but with heterogeneous proportions (Fig. 4A–C). Specifically, CD8⁺ T cells were the most abundant subset across these groups (Fig. 4D). In terms of composition of these T-cell subsets, the proportion of CD8⁺ T cells was highest in relapsed GCB-type DLBCL, whereas the proportion of CD4⁺ T cells was lowest in this group (Fig. 4E). Furthermore, we examined the distribution of expression of cytotoxicity-related genes across these three clinical groups and found that the expression in CD8⁺ T cells was reduced in relapsed GCB-type cases as compared to the levels in the other two groups (Fig. 4F). AUCell-based “Cytotoxicity” scores indicated that CD8⁺ T cells in de novo GCB-type cases were significantly more cytotoxic than those in other groups, with the lowest scores observed in relapsed GCB-type cases (Fig. 4G).

Fig. 4.

Re-clustering of T cells. A UMAP of T cells across groups. B UMAP colored by T-cell subset. C Marker genes for each subset. D Bar chart of T-cell subset counts. E Subset proportions per clinical group. F Bubble plot of subtype-signature gene expression (bubble size, % expressing cells; color, scaled expression). G AUCell “Cytotoxicity” scores across T-cell subsets

Similarly, we conducted cell subset clustering analysis for profiled myeloid cells and identified four subclusters, including macrophages, myeloid dendritic cells (mDCs), plasmacytoid dendritic cells (pDCs), and monocytes (Fig. S2A–C), among them mDCs was the most abundant subset (Fig. S2D). In line with the cell subclustering in T-cell, all identified subsets of myeloid cell were presented in these three groups but with different proportions (Fig. S2E). Specifically, in relapsed GCB-type DLBCL, pDCs has the highest proportion, whereas the proportions of mDCs and monocytes were comparatively lower in this group as compared the proportions in other groups (Fig. S2F).

Cell–cell communication

To interrogate intercellular signaling in the TME, we constructed receptor–ligand interaction networks in DLBCL, including total interaction counts and outgoing signaling strengths between lineages (Fig. 5A–B). As shown in Fig. 5, communication patterns differed markedly across cell types. Notably, as compared to good-survival B cells subset, B cell subset indicating a worse-survival had increased regulatory outputs toward CD4⁺ T, NK, and naïve T cells, but decreased outputs toward monocytes and macrophages (Fig. 5C–D). Furthermore, the ligand–receptor analyses between T-cell subsets and outcome-defined malignant B cells revealed that IL-2 and BTLA pathways were predominantly enriched in worse-survival malignant B cells that act as signaling senders, whereas the LT (LTA) pathway exclusively presented in good-survival malignant B cells that act as signaling senders (Fig. 5E). Specifically, worse-survival malignant B cells specifically regulated the CD4 T, CD8 T, naïve T, and Treg cells via IL7–(IL7R + IL2RG) (IL-2 pathway) and BTLA–TNFRSF14 (BTLA pathway). In contrast, good-survival malignant B cells regulated the same T-cell subsets via LTA–TNFRSF1B, LTA–TNFRSF14 pathways.

Fig. 5.

Cell–cell communication in the TME. A Heatmap of total interaction counts between cell types. B Heatmap of outgoing signaling strength. C–D Chord diagrams of interactions from worse-survival (C) and good-survival (D) malignant B cells to other cell types. E Bubble plot of selected ligand–receptor pairs between T-cell subsets and outcome-defined malignant B cells

Identification of key genes

We then performed differential expression and functional enrichment analyses for genes in identified groups with different survival outcomes in de novo ABC-type and relapsed GCB-type groups. As compared to others (good-survival + background) groups, worse-survival malignant B cells had significantly activated PI3K–AKT signaling pathway in both de novo ABC-type and relapsed GCB-type groups (Fig. 6A–D), a pattern echoed with the known roles of this pathway in proliferation, invasion, migration, and metabolic reprogramming [29, 30]. Moreover, IL4R was significantly up-regulated in worse-survival malignant B cells and enriched within the PI3K–AKT pathway.

Fig. 6.

Key-gene discovery. A Volcano plot: worse-survival vs. other malignant B cells (de novo ABC-type). B KEGG enrichment of up-regulated genes (de novo ABC-type). C Volcano plot (relapsed GCB-type). D KEGG enrichment of up-regulated genes (relapsed GCB-type). E Venn diagram of down-regulated genes (worse-survival vs. others) intersecting across groups. F Violin plots of HMGN2 expression in outcome-defined malignant B cells (left: de novo ABC-type; right: relapsed GCB-type). G Kaplan–Meier curves for HMGN2 (left: GSE87371; right: TCGA-DLBC). H Scatter plots showing HMGN2–IL4R expression correlations in malignant B cells (left: de novo ABC-type; right: relapsed GCB-type)

Intersecting down-regulated genes in worse-survival malignant B cells (vs. others) between de novo ABC-type and relapsed GCB-type groups yielded 93 shared genes (Fig. 6E). Among these, HMGN2 was significantly down-regulated in worse-survival malignant B cells in both groups (Fig. 6F). Prognostic validation in bulk datasets (GSE87371 and TCGA-DLBC) suggested better survival in cases with higher HMGN2 expression (Fig. 6G). In malignant B cells, HMGN2 and IL4R expression were significantly negatively correlated in both de novo ABC-type and relapsed GCB-type cohorts (p < 0.05, r < 0) (Fig. 6H). These data suggest that low HMGN2 may facilitate IL4R up-regulation, thereby activating PI3K–AKT to promote malignant-cell proliferation and survival, implicating HMGN2 as a potential prognostic biomarker and therapeutic target, particularly in ABC-DLBCL.

Functional impact of HMGN2 in tumor cells

To better understand the function of HMGN2 signaling in tumor cells, we generated stable HMGN2 knockdown and overexpression in DLBCL cell lines (Ly8, U2932). qRT-PCR and Western blot confirmed the effective HMGN2 down-/up-regulation (Fig. 7A–B). Meanwhile, CCK-8 assays showed that HMGN2 knockdown increased cell viability, whereas overexpression decreased viability (Fig. 7C). Also, EdU assays indicated enhanced proliferation upon knockdown and suppressed proliferation upon overexpression (Fig. 7D). Likewise, colony formation assays demonstrated increased clonogenicity with HMGN2 knockdown and reduced clonogenicity with overexpression (Fig. 7E). Given that the Cyclin A–CDK2 complexes are key drivers of S-/G2-phase progression and genome duplication fidelity, we quantified Cyclin A⁺ cells by immunofluorescence and found that the proportion of Cyclin A⁺ cells increased with HMGN2 knockdown but decreased with HMGN2 overexpression (Fig. 7F).

Fig. 7.

HMGN2 regulates DLBCL cell phenotypes. A qRT-PCR for HMGN2 mRNA (Ly8, U2932). B Western blot for HMGN2 protein (Ly8, U2932). C CCK-8 viability after HMGN2 knockdown/overexpression. D EdU proliferation assay (scale bar: 200 μm). E Colony formation (crystal violet). F Immunofluorescence for Cyclin A⁺ cells (scale bar: 20 μm). *data are mean ± SD. *p < 0.05; **p < 0.01; ***p < 0.001; ***p < 0.0001

HMGN2 promotes R-loop formation and suppresses PI3K–AKT signaling

Initial immunodot blot assays using the S9.6 antibody suggested that HMGN2 knockdown significantly reduced global R-loop levels, whereas its overexpression led to a marked increase in DLBCL cells (Fig. 8A). To address potential cross-reactivity with double-stranded RNA (dsRNA) and ensure the specificity of our detection, we incorporated RNase H pretreatment. The near-complete elimination of the S9.6 signal upon RNase H digestion across all conditions validated that the observed signals were specifically derived from DNA:RNA hybrids (Fig. 8A). Consistent with a model in which HMGN2 deficiency drives oncogenic signaling, Western blot analysis revealed that HMGN2 knockdown robustly upregulated AKT, p-AKT, PI3K, and p-PI3K, whereas its overexpression exerted the opposite effect in Ly8 and U2932 cells (Fig. 8B).To establish the functional necessity of HMGN2-mediated R-loop homeostasis in the repression of IL4R and the subsequent activation of the PI3K/AKT cascade, we performed rescue experiments via RNase H1 overexpression in HMGN2-depleted cells (Fig. 9A–D). While HMGN2 knockdown reduced global R-loop levels, subsequent RNase H1 overexpression led to further clearance of these structures. Crucially, the significant induction of IL4R mRNA and protein, as well as the concomitant hyperactivation of the PI3K/AKT pathway triggered by HMGN2 loss, were markedly reversed upon RNase H1-mediated R-loop depletion (Fig. 9E–F, 9I–T). Furthermore, functional CCK-8 assays confirmed that the enhanced proliferative capacity of HMGN2-deficient cells was significantly suppressed by RNase H1, demonstrating that R-loop dynamics are functionally essential for the phenotypic shifts observed upon HMGN2 loss (Fig. 9G–H).

Fig. 8.

HMGN2 controls PI3K–AKT via R-loop modulation. A Dot-blot analysis of R-loop formation after HMGN2 knockdown/overexpression in Ly8 and U2932. B Western blots of AKT, p-AKT, PI3K, p-PI3K following HMGN2 perturbation. *data are mean ± SD. *p < 0.01; **p < 0.001; ***p < 0.0001

Fig. 9.

HMGN2 controls IL4R/PI3K–AKT signaling and proliferation via R-loops. A-D R-loop detection by S9.6 dot-blot assay. (A, C) Representative images showing R-loop formation in Ly8 and U2932 cells treated with siRNA or overexpressing RNase H (oe-RNase H) to deplete R-loops. (B, D) R-loop intensity quantification in Ly8 and U2932 cells. HMGN2 knockdown reduces R-loop formation, and RNase H further decreases R-loops in HMGN2 knockdown cells. E-F IL4R mRNA expression in Ly8 (E) and U2932 (F) cells. G-H Cell proliferation in both Ly8 (G) and U2932 (H) cells. I-T Western blot analysis of PI3K/AKT signaling proteins. Western blots for IL4R, AKT, p-AKT, PI3K, and p-PI3K in Ly8 (I) and U2932 (Q) cells. Data are presented as means ± SDs. *p < 0.01; **p < 0.001; ***p < 0.0001

We further substantiated the requirement of IL4R in this process by co-silencing IL4R in the si-HMGN2 background. Evaluation of cellular fitness via CCK-8 assays showed that the pro-proliferative effects induced by HMGN2 deficiency were abrogated by the simultaneous silencing of IL4R (Fig. 10A–B). Western blot analysis corroborated these phenotypic findings, showing that IL4R blockade effectively suppressed the upregulation of IL4R protein and the phosphorylation of PI3K and AKT, thereby identifying IL4R as the primary mediator of signaling activation in this context (Fig. 10C–N). To rigorously confirm pathway dependence, we employed the PI3K inhibitor copanlisib. Pharmacological intervention significantly attenuated the proliferative advantage of HMGN2-depleted cells and neutralized the phosphorylation of PI3K and AKT without altering total protein levels, indicating that the malignant phenotype is contingent upon sustained signaling activation rather than altered protein expression (Fig. 11A–L).

Fig. 10.

IL4R blockade reverses PI3K/AKT activation and proliferation induced by HMGN2 knockdown. A–B CCK-8 proliferation assays in Ly8 (A) and U2932 (B) cells treated with si-NC, si-HMGN2, or si-HMGN2 + si-IL4R. C–H Western blot analysis (C) and quantification (D–H) of IL4R, AKT, p-AKT, PI3K, and p-PI3K expression in Ly8 cells. I–N Western blot analysis (I) and quantification (L–N) in U2932 cells. Data means SD. *p < 0.01; **p < 0.001; ***p < 0.0001

Fig. 11.

PI3K/AKT pathway dependence in HMGN2-deficient cells upon copanlisib treatment. A-B CCK-8 proliferation assays in Ly8 (A) and U2932 (B) cells treated with si-NC, si-HMGN2, or si-HMGN2 + copanlisib. C-H Western blot analysis (C) and quantification (D-H) of IL4R, AKT, p-AKT, PI3K, and p-PI3K expression in Ly8 cells. HMGN2 knockdown increases IL4R expression and activates PI3K/AKT signaling, whereas co-silencing IL4R reverses these effects. I-N. Western blot analysis (I) and quantification (L-N) in U2932 cells show similar results, confirming that IL4R blockade suppresses PI3K/AKT activation induced by HMGN2 deficiency.Data are presented as means ± SDs. *p < 0.01; **p < 0.001; ***p < 0.0001

Finally, to delineate the locus-specific impact of HMGN2, we conducted DNA:RNA immunoprecipitation followed by quantitative PCR (DRIP-qPCR) at the IL4R locus. Using the transcriptionally silent MyoD1 locus as a negative control, we found that HMGN2 depletion reduced, and overexpression increased, R-loop enrichment specifically at the IL4R gene (Fig. 12A–G). These enrichment signals were completely sensitive to RNase H treatment, while no significant changes were observed at the MyoD1 locus across any experimental conditions. These results provide high-resolution evidence that HMGN2 serves as a critical, site-specific regulator of R-loop stability at the IL4R locus, thereby maintaining transcriptional repression of the IL4R–PI3K/AKT signaling cascade.

Fig. 12.

HMGN2 regulates R-loop formation at the IL4R locus and occupies the IL4R promoter. A–D DRIP-qPCR analysis of R-loops at the IL4R locus and control locus (MyoD1) in Ly8 and U2932 cells. E–G ChIP-qPCR analysis of HMGN2 occupancy across the IL4R promoter region (IL4R-1 to IL4R-3). Significant HMGN2 enrichment is detected at IL4R-1, IL4R-2, and IL4R-3 upon HMGN2 overexpression in both cell lines. Data are mean ± SD. *p < 0.01; **p < 0.001; ***p < 0.0001

Discussion

By integrating scRNA-seq of DLBCL specimens, our study revealed malignant and immune ecosystems with patient group–specific features and linked them to outcome-associated cell states. Our findings extend and refine prior work that profiled DLBCL tumor cell states and microenvironmental archetypes using single-cell and bulk approaches, including distinct malignant programs, variable T/NK/myeloid composition, and prognostic TME classes. Specifically, our findings reveal subtype- and outcome–stratified intercellular communication maps from malignant B cells to immune subsets, identify branch-dependent gene modules that accompany divergence toward proliferative/immune-suppressive fates. Notably, our results point toward a potential HMGN2 regulated R-loop–PI3K–AKT axis that may mechanistically constrains malignant growth and potentially shape immune interactions. Together findings collectively nominate HMGN2 and IL4R–PI3K–AKT signaling as putative coupled prognostic/therapeutic nodes in DLBCL and suggest BTLA versus LTA pathway biases as possible microenvironmental correlates of survival [14].

One of our major findings is the identification of malignant subclusters with a “worse-survival” phenotype that are present in both de novo ABC-DLBCL and relapsed GCB-DLBCL. This observation challenges the rigid dichotomy of the cell-of-origin model, which has historically classified ABC and GCB subtypes as distinct diseases with separate oncogenic drivers [3, 31]. This finding suggests a possible evolutionary convergence, where cells from different subtypes may acquire a common transcriptional program conferring a selective advantage, potentially through treatment resistance and immune evasion. This aligns with emerging concepts of “molecular high-grade” lymphoma, which are defined by mutational and transcriptional signatures rather than subtype alone [32]. The pseudotime trajectory analysis further supports this, demonstrating a clear branching point where malignant cells evolve towards a state enriched for “worse-survival” cells, characterized by the activation of NF-κB, hypoxia, and JAK-STAT signaling—pathways universally associated with proliferation, metabolic adaptation, and poor outcomes [33].

Our discovery that aggressive malignant subclusters exhibit significantly lower R-loop scores suggests a potential role of R-loop biology in DLBCL. Traditionally, R-loops are viewed as dangerous DNA:RNA hybrids that threaten genomic integrity by inducing replication stress and DNA damage [33]. However, our data suggests that their regulated presence is essential for maintaining a non-malignant cellular state. The lowest R-loop scores in relapsed GCB cases imply that a loss of certain R-loops, rather than their gain, may be a key step in malignant evolution. This is not entirely inconsistent with the established literature; instead, it suggests that both excessive and insufficient R-loop formation can be oncogenic, depending on the genomic context [19]. The aggressive, R-loop-low state may represent a cell that has bypassed the need for the strict transcriptional control that R-loops provide, allowing for the unchecked expression of pro-survival genes. This novel hypothesis positions the targeted restoration of specific R-loops as a potential therapeutic strategy.

Our deep profiling of the TME reveals a coordinated landscape of immune dysfunction in relapsed GCB-DLBCL. The combination of functionally exhausted CD8⁺ T cells—evidenced by their significantly reduced cytotoxicity scores—and an expanded population of tolerogenic plasmacytoid dendritic cells (pDCs) indicate a profoundly suppressed anti-tumor immune response [34, 35]. Crucially, our analysis moves beyond correlation to suggest causation. The ligand-receptor interaction analysis reveals that “worse-survival” malignant B cells are not passive beneficiaries of this immunosuppression but active architects. Their specific engagement of the BTLA-and IL-2 inhibitory pathways with T cells is a masterstroke of immune subversion. The BTLA-HVEM axis is a potent, yet less explored, immune checkpoint that directly inhibits T cell activation [36]. That aggressive tumor cells selectively use this pathway, while “good-survival” cells use immunostimulatory pathways like LTA, reveals a sophisticated level of communication and nominates BTLA-HVEM as a compelling therapeutic target for relapsed/refractory DLBCL.

The most significant mechanistic insight from this study is the elucidation of HMGN2‘s tumor-suppressive role in DLBCL. We identify HMGN2 not merely as a differentially expressed gene, but could act as a central regulatory node that restrains oncogenic signaling. It has been known that HMGN2 loss enhances proliferation, clonogenicity, and cell cycle progression by promoting entry into S-phase, as confirmed by increased Cyclin A⁺ cells—a direct marker of S/G2 phase progression [37]. Our functional validation experiments revealed that RNase H1 overexpression successfully reversed the IL4R induction and PI3K/AKT hyperactivation triggered by HMGN2 deficiency, hinting that R-loop homeostasis may be functionally involved in HMGN2 mediated gene repression. Furthermore, our ChIP-qPCR data analysis indicated potential HMGN2 occupancy at the IL4R promoter, suggesting it may act as a localized chromatin architectural regulator. The potential functional dependence of this axis was further confirmed through IL4R blockade and pharmacological inhibition with copanlisib [38], which appeared to neutralize the proliferative advantage of HMGN2-low cells. These observations collectively indicate a previously unknown function for HMGN2 in promoting R-loop formation that bridges the gap between our observational scRNA-seq data and the core oncogenic PI3K-AKT pathway, a hallmark pathway dysregulated in DLBCL subtypes and drives proliferation and survival [39–41]. We thus propose a putative mechanistic model wherein the loss of HMGN2 may lead to reduced R-loop formation, potentially resulting in the derepression of IL4R expression, subsequent hyperactivation of the PI3K-AKT signaling pathway, and ultimately uncontrolled malignant B-cell proliferation and survival. This framework is corroborated by our observations in patient-derived malignant B cells and independent bulk RNA-seq cohorts (GSE87371 and TCGA-DLBC), which suggests an HMGN2‘s association with favorable outcomes, positioning it as a potential biomarker for identifying high-risk DLBCL cases—a role that aligns with its documented tumor-suppressive activities in other cancer types, albeit remaining largely uncharacterized in the context of lymphoma [41–44]. Collectively, this HMGN2-R-loop-IL4R-PI3K-AKT axis unveils a novel therapeutic vulnerability, whereby HMGN2-low tumors may display enhanced susceptibility to inhibitors targeting the PI3K-AKT or IL4R pathways. Such strategies build upon the proven clinical efficacy of these agents in PI3K-AKT-dependent DLBCL subtypes, even in the absence of HMGN2-stratified clinical trial data to date [45–48].

While our study provides a rich resource, certain limitations must be acknowledged. The cohorts leveraged in the present study, especially for validation, are generally small, which may limit the broad generalizability of the findings. Although we have mapped HMGN2 occupancy and R-loop regulation to the IL4R promoter, the global landscape of HMGN2-regulated R-loops remains to be fully explored. Future work integrating ChIP-seq and DRIP-seq will be essential to define the complete regulatory repertoire of HMGN2. Finally, the efficacy of targeting the identified pathways in HMGN2-low models requires further verification in pre-clinical in vivo studies.

Conclusions

In summary, our study characterizes a cellular atlas that provides mechanistic insights into DLBCL pathogenesis. We propose that aggressive disease may be defined by a convergent transcriptional state that likely remodels its microenvironment and may be influenced by the loss of the epigenetic regulator HMGN2. By linking R-loop biology to a core oncogenic pathway, we suggest HMGN2 as a candidate regulator of lymphomagenesis. These findings provide a new biological framework for potentially stratifying DLBCL patients and suggest a suite of candidate therapeutic targets, including the BTLA-HVEM axis and the IL4R-PI3K/AKT pathway.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

DLBCL

Diffuse large B-cell lymphoma

ABC

Activated B-cell-like

GCB

Germinal Center B-cell-like

scRNA-seq

Single-cell RNA sequencing

HMGN2

High-Mobility Group Nucleosome Binding Domain 2

NHL

Non-Hodgkin lymphoma

BCR

B-cell receptor

TME

The tumor microenvironment

GEO

Gene Expression Omnibus

TCGA

The Cancer Genome Atlas

PCA

Principal component analysis

DEGs

Differentially expressed genes

GO

Gene Ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

Author contributions

J.Z., J.T. and D.X.: Conceptualization; Y.H. and Y.X.: methodology; C.C., J.L. and C.Z.: Investigation; J.Z. and J.T.: writing—original draft; D.X.: writing—review & editing; J.Z., J.T. and D.X.: funding acquisition; C.C., J.L. and C.Z.: resources; J.Z., J.T. and D.X.: supervision. All authors read and approved the final manuscript.

Funding

This work was supported by grants from Guangdong Basic and Applied Basic Research Foundation (grant no. 2022A1515140022, 2024A1515140108), the Medical Scientific Research Foundation of Guangdong Province (grant no. A2024577, A2025173), the Scientific Research Start Plan of the Eighth Affiliated Hospital, Southern Medical University (grant no. SRSP2021038, SRSP2023013, SRSP2023019, SRSP2024034, SRSP2024038, SRSP2025032 and SRSP2025035).The funders provided financial support only and had no role in the study design, in the collection, analysis, or interpretation of the data, in the writing of the report, or in the decision to submit the manuscript for publication. Despite author affiliations with The Eighth Affiliated Hospital, Southern Medical University, the funder did not influence the results or outcomes of the study.

Data availability

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/Supplementary Material. All data are available from the GEO (https://www.ncbi.nlm.nih.gov/geo/) database within the article. GEO database under accession number GSE182436 and GSE87371. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Material availability

This study did not produce any new unique reagents.

Declarations

Ethical approval

Not applicable.

Competing interests

The authors declare no competing interests.