Plasmablast-like lymphoma cells as a distinct subpopulation confer multidrug resistance in PCNSL

Abstract

Background

Primary central nervous system lymphoma (PCNSL) is a highly aggressive subtype of non-Hodgkin lymphoma that is confined to the central nervous system. PCNSL is associated with a poor 5-year survival rate of 30%–40%, partly due to a high recurrence rate of 60%. A comprehensive understanding of the molecular signatures and mechanisms underlying drug resistance in PCNSL is crucial, as it has significant implications for therapeutic strategies.

Methods

A cohort of 56 newly diagnosed PCNSL patients was generated. The tumor specimens were obtained by biopsy. A combined approach, including single-cell transcriptomics and B-cell receptor sequencing, transcriptome-informed multiplex immunohistochemistry, and ex vivo drug response assays, was applied to reveal the transcriptional and immune microenvironment landscape of PCNSL.

Results

We identify four main B-cell subtypes, each characterized by distinct transcriptomic profiles, from malignant B cells in PCNSL. Among these subtypes, plasmablast-like lymphoma cells (PBLCs), characterized by upregulated genes typically associated with plasma cell differentiation, comprise 1.3%–8.1% of malignant B cells and correlate with poor prognosis in PCNSL. PBLCs exhibit reduced expression of CD20, Bruton tyrosine kinase, and FAS, resulting in drug resistance and immune evasion. In addition, PBLCs elevate the expression of critical transcription factors, including XBP1 and PRDM1, to maintain their phenotype. Clinical drugs, such as Lenalidomide, demonstrate potential cytotoxic effects on PBLCs.

Conclusions

These findings highlight PBLCs as a distinct subtype of malignant B cells that play a critical role in the multidrug resistance of PCNSL and reveal the molecular signature of PBLCs that can be targeted for therapy.

Graphical Abstract

Graphical Abstract.

Key Points.

• We identified a novel subset of lymphoma cells named plasmablast-like lymphoma cells (PBLCs) with plasma cell differentiation in primary central nervous system lymphoma.

• PBLCs are resistant to multiple drugs and associated with a poor prognosis.

Importance of the Study.

Primary central nervous system lymphoma (PCNSL) is an aggressive subtype of non-Hodgkin lymphoma confined to the CNS and eyes. As a devastating form of extranodal non-Hodgkin lymphoma, PCNSL poses significant treatment challenges due to high recurrence rates and refractoriness, with its therapeutic failures primarily rooted in its heterogeneity. In this study, we identified a novel subset of lymphoma cells with plasma cell differentiation in PCNSL, which are resistant to multiple drugs and associated with a poor prognosis. These findings highlight the difficulty of targeting this cell population and suggest that therapies aimed at plasma cell differentiation may provide a promising approach for treating PCNSL.

Primary central nervous system lymphoma (PCNSL) is an aggressive subtype of non-Hodgkin lymphoma localized to the CNS and eyes, representing approximately 4%–6% of extranodal lymphomas and 3%–4% of newly diagnosed malignant brain tumors.^1,2 Despite favorable responses to high-dose methotrexate (MTX)-based chemotherapy, PCNSL exhibits a worse prognosis than non-CNS systemic lymphoma, with a 5-year survival rate of 30%–40%.³ Additionally, 10%–15% of patients with PCNSL are refractory to MTX, and up to 66% experience relapse.^4,5 Following the failure of first-line treatment, outcomes remain bleak, with a median overall survival (OS) of 2 months for refractory PCNSL and 3.7 months for relapsed cases.^3,6 The lack of standard salvage treatments contributes to the poor prognosis,^2,3,5 highlighting the urgent need for new therapeutic strategies.

Pathologically, most PCNSL cases (90%) are classified as diffuse large B-cell lymphoma (DLBCL) with an activated B-cell-like (ABC) phenotype,⁷ which is associated with a poorer prognosis.⁸ However, the precise B-cell states in PCBSL remain uncertain.

Advances in genomic, transcriptomic, and immune microenvironment research have expanded the understanding of PCNSL pathogenesis, facilitating the development of novel therapeutic agents.⁹ Chronic activation of B-cell receptor (BCR) and Toll-like receptor signaling, along with mechanisms of immune evasion, is believed to support B-cell survival in PCNSL.¹⁰ Several novel agents targeting these dysregulated pathways have shown promise in clinical trials, offering new treatment options for refractory and relapsed PCNSL. However, recurrence remains common, even with combination therapies that target multiple mechanisms. This study aimed to characterize B-cell states in PCNSL to identify potential therapeutic targets.

Utilizing a comprehensive approach combining single-cell transcriptomics, B-cell receptor sequencing, transcriptome-informed multiplex immunohistochemistry, and ex vivo drug response assays, a novel lymphoma cell type was identified. These cells, which upregulate genes associated with plasma cell differentiation, exhibit resistance to standard first-line therapies and are associated with poor prognosis. These findings shed light on the heterogeneity of PCNSL and point toward a promising new therapeutic target.

Methods

Human Specimens

Human brain tissues and blood samples were collected at the Department of Neurosurgery, Second Affiliated Hospital, School of Medicine, Zhejiang University. Informed written consent was provided by all patients, and all research activities were approved by the Institutional Ethics Committee of the Second Affiliated Hospital, School of Medicine, Zhejiang University (IR2021001101). Details of the enrolled patients are provided in Supplementary Table 1.

Processing of Patient Samples

Freshly resected brain tissues were thoroughly washed with phosphate-buffered saline (PBS) to eliminate blood clots and minimize contamination. The tissues were cut into 2–4 mm³ pieces and dissociated using the gentle MACS Octo Dissociator with Heaters (Miltenyi Biotech). Tumors were dissociated with the Brain Tumor Dissociation Kit (130-095-942, Miltenyi Biotech), and healthy brain tissues were processed with the Adult Brain Dissociation Kit (130-107-677, Miltenyi Biotech) following the manufacturer’s instructions. Red blood cells were lysed using 1 mL of 1X Red Blood Cell Lysis Buffer (eBioscience). The single-cell suspensions were then washed and resuspended in PBS containing 0.04% bovine serum albumin (BSA, Sigma). Whole blood samples were diluted with PBS, and PBMCs were isolated using density gradient centrifugation with Ficoll-Paque for 20 min at 500 g without braking. The cloudy middle layer was collected and resuspended in PBS containing 0.04% BSA. All samples were processed within 15 min post-surgery.

Single-cell Library Construction and Sequencing

10X Genomics libraries were prepared according to the manufacturer’s protocols. Sixteen 3’ scRNA-seq libraries (8 tumors, 5 PBMCs, and 3 human brain specimens) and two 5’ scRNA-seq libraries were generated using 10X Genomics Chromium Single Cell 3’ v3 chemistry and 10X Genomics Chromium Single Cell 5’ v1.1 chemistry, respectively. V(D)J enrichment kits from Chromium Single Cell were used to capture full-length V(D)J segments of B cells. All libraries were sequenced on the Illumina NovaSeq platform with 150-bp paired-end reads at Novogene Bioinformatics Technology Co., Ltd. Paired-end scBCR-seq fastq files were mapped to human V(D)J reference (vdj_GRCh38_alts_ensembl-3.1.0) and processed with default parameters by CellRanger (v4.0.0) following the tutorials from 10X Genomics.

Basic Processing and Clustering Analysis of Single-Cell Transcriptome Data

Basic processing and visualization of scRNA-seq data were conducted using Seurat (v4.1.0).¹¹ Quality control filters were applied to select cells with less than 20% mitochondrial reads and 500–7000 unique genes. The filtered count matrix was normalized, and mitochondrial contamination was removed using the SCTransform function. Anchor genes were identified across samples, and all samples were integrated into a single dataset to minimize batch effects. Principal component analysis was performed on the scaled data, with the top 20 principal components retained. Clustering was performed using the FindClusters function, and the optimal resolution was determined based on cluster stability using the clusteree (v0.5.4) package. For visualization, nonlinear dimensionality reduction techniques, such as uniform manifold approximation (UMAP) and t-distributed stochastic neighbor embedding (t-SNE), were applied. Differentially expressed genes (DEGs) were identified using the Wilcoxon rank-sum test via the Seurat FindAllMarkers and FindMarkers functions. DEGs were defined based on 2 criteria: (1) Bonferroni-adjusted P-value <.05, and (2) |log2-fold change| >0.25.

For downstream analysis, single-cell copy number variation (CNV) was performed using inferCNV (v1.4.0), while functional enrichment analysis was conducted using Metascape. Pseudotime analysis was carried out with Monocle3 (v1.3.1) and Cytotrace (v0.3.3). Cell–cell interaction analysis was performed using InterCellDB (v0.9.2), and single-cell gene set activity analysis was conducted using AUCell (v1.10.0). Single-cell regulatory network inference was performed with pySCENIC (v0.11.2). Detailed descriptions of the bioinformatic methods are provided in the Supplementary Methods. The raw single-cell data files for each sample are available in the GSA-Human database under the access number HRA008594.

B-Cell State Identifier Development

Meta-module genes for the 4 B-cell states—dark zone (DZ), light zone (LZ), memory B cell precursors (PreM), and plasmablasts (PBL)—were obtained by calculating DEGs in tonsil GC B cell subclusters.¹² The Seurat AddModuleScore function was used to compute cell scores for these meta-modules. The 2-dimensional representation method¹³ was then applied to summarize cell distribution across these states. Cells were first classified into DZ/LZ versus PreM/PBL based on the sign of D = max(SC_DZ, SC_LZ) − max(SC_PreM, SC_PBL), which defined the y-axis. For DZ/LZ cells (D > 0), the x-axis was defined as log₂(|SC_LZ − SC_DZ| + 1), and for PreM/PBL cells (D < 0), the x-axis was defined as log₂(|SC_PBL − SC_PreM| + 1).

Quantification of the Proportion of PBLCs

Lymphoma cells were detected by multiplex immunohistochemistry (mIHC) using the Opal Polaris 780 Reagent Pack (NEL861001KT, Akoya Biosciences) following the manufacturer’s protocol. Details regarding the multiplex immunohistochemistry procedure are provided in the Supplementary Methods. For the quantification of PBLCs, QuPath (v0.3.2)¹⁴ was employed. Each sample was imaged and analyzed using at least 2 mIHC slices. In each slice, a minimum of 5 regions measuring 250 μm × 250 μm were initially labeled by 1 experimenter and subsequently reviewed by another experienced experimenter to ensure that the selected regions were located within the tumor tissue, avoiding nontumor areas. A background threshold was first established for each piece of PCNSL tissue, and the background intensity was subtracted from the entire image to reduce intertissue background variation. Subsequently, cells were identified based on DAPI staining using the QuPath cell detection algorithm. Finally, using cell intensity classification, CD79A-positive cells and JCHAIN/CD79A double-positive cells (PBLCs) were manually counted, and the positive proportion was calculated. To ensure accuracy, all slices were distributed randomly to 3 independent blinded experts for cell counting. The percentage of JCHAIN/CD79A double-positive cells was calculated for each labeled region, and the results were averaged by sample. For FAS expression quantification, CD79A-positive cells and PBLCs were identified as described above. The median fluorescent intensity of FAS was then measured for each identified cell using QuPath.

Survival Analyses

Survival was assessed using the Kaplan–Meier method and compared between groups using the log-rank test. A tertile cutoff was applied, with the upper tertile (patients with the highest third of PBLC proportion) classified as the “high-PBLC” group, and the lower tertile (patients with the lowest third of PBLC proportion) as the “low-PBLC” group.

Drug Sensitivity Score Calculation

Drug sensitivity scores were calculated using the Beyondcell R package (v1.3.3) and its integrated drug sensitivity signature collection (SSC) database, which includes CTRP, GDSC, and CCLE datasets with pretreatment transcriptional signatures predicting drug response. Briefly, expression signatures were obtained via limma-based differential expression analysis using the area under the curve as the drug response metric. Drug sensitivity scores were then calculated, rescaled between 0 and 1 based on their switch point, and aggregated to generate a combined drug sensitivity score.

Drug Prediction

The Beyondcell (v1.3.3) R package was used to identify drugs with potential efficacy against PBLCs. Drug perturbation (PSC) and drug sensitivity (SSC) collections in Beyondcell were employed to compute Beyondcell scores for each cell. The most differentially sensitive drugs targeting PBLCs were selected, and the top 500 genes downregulated by these drugs were listed. Drugs that significantly affected key PBLC TFs were considered potential candidates.

Drug Screening

PCNSL tissue was obtained as biopsy or surgical specimens from patients, with the post-surgery diagnosis of PCNSL confirmed by experienced pathologists at the institute. The tissue was sectioned into pieces of 2 mm³. To evaluate drug efficacy, each drug was mixed with a culture medium containing MTX (1000 nM), Rituximab (400 nM), Ibrutinib (400 nM), NVP-ADW742 (10 μM), NVP-TAE 684 (10 μM), RITA (10 μM), and Lenalidomide (10 μM). The culture medium consisted of RPMI solution, Penicillin (100 U/mL), and Streptomycin (0.1 mg/mL). For each drug, at least 2 PCNSL tissue pieces were placed in 12-well plates, each suspended in 1 mL of the prepared mixtures. All procedures were completed within 1 h of specimen acquisition. After 72 h of incubation, the tissue pieces were removed from the RPMI solution, fixed in 4% PFA overnight, and dehydrated in 30% sucrose solution for at least 2 days, as previously described.

Results

Single-Cell Transcriptional Profiling Charts of Cell Heterogeneity in PCNSL

To investigate the molecular characteristics of PCNSL, we first established a cohort of 56 patients with confirmed diagnoses of PCNSL, including those who underwent stereotactic biopsy or surgical resection of intracranial lesions. The median follow-up period was 20.5 months. Of these patients, 3 (5.4%) were refractory to initial therapy, while 34 (60.7%) experienced at least 1 relapse with a median recurrence time of 14 months (Supplementary Table 1).

We then developed an experimental pipeline to explore the tumor heterogeneity and identified the drug-resistant cell subsets in PCNSL (Figure 1A). Single-cell 3’ RNA sequencing was performed on 8 tumor specimens (Figure 1B, Supplementary Figure 1A). After filtering out low-quality cells and doublets, 39 231 cells were retained for further analysis. The number of cells (~4904 per sample) and genes detected (~3274 per cell) were consistent across the 8 samples (Supplementary Figure 1B). Unsupervised clustering revealed 3 distinct B-cell populations based on gene enrichment: B cells1 (FCRL5⁺), B cells2 (TOP2A⁺, MKI67⁺), and B cells3 (JCHAIN⁺). Additionally, tumor-associated macrophages (TAMs), CD4⁺ T cells, CD8⁺ T cells, oligodendrocytes (OLGs), and dendritic cells (DCs) were identified (Figure 1C). Clustering results were further validated by single-cell 5’ RNA sequencing in 2 cases, which demonstrated similar cell population distributions and proportions (Supplementary Figure 1C).

Figure 1.

Single-cell transcriptional profiling of primary central nervous system lymphoma (PCNSL). (A) Schematic of the experimental workflow detailing the single-cell landscape dissection and drug response exploration. (B) Uniform manifold approximation (UMAP) visualization of PCNSL cells from eight 3’ scRNA-seq samples, with cell types color-coded and annotated based on transcriptomic profiles. (C) Violin plots showcasing the expression levels of selected marker genes for each PCNSL cell cluster. (D) Bar graph representing the distribution of each subcluster in the scRNA-seq dataset, with color coding for distinct cell types. (E) Heatmap displaying the copy number variation (CNV) profiles of B cells (bottom panel), normalized against the “normal” cluster (top panel, immune cells). Chromosomal CNV changes are depicted in columns, and individual cells are represented in rows. (F) The IGKC fraction, defined as IGKC / (IGKC + IGLC2), was calculated for each B cell. Cells with an IGKC fraction of ≥0.5 were classified as κ+, while those with a ratio below 0.5 were classified as λ+. (G) Representative multiplex immunohistochemistry (mIHC) images of PCNSL tissues stained for CD8, CD68, CD4, CD3, and CD79A. Nuclei were labeled with DAPI. Scale bars: 1 mm (low magnification) and 200 μm (high magnification).

Notably, B cells in PCNSL exhibited markers relevant to cell-of-origin (COO) classification,⁷ such as the germinal center (GC)-associated gene BCL6 and CD10/MME in B cells1 and B cells2, while B cells3 expressed the post-GC marker IRF4 (Supplementary Figure 1D), aligning with previous findings that PCNSL shares characteristics with sDLBCL.¹⁵ Despite tumor heterogeneity, B cells comprised approximately two-thirds of the total cell population, while T cells made up 20% and tumor-associated macrophages (TAMs) accounted for 10% (Figure 1D).

InferCNV¹³ was applied to examine CNV profiles in PCNSL. Our findings revealed that all 3 B-cell subsets exhibited chromosomal aberrations commonly associated with PCNSL, including frequent losses in chromosome 6 and variable gains in chromosomes 8, 12, 18, 21, and 22 (Figure 1E), consistent with previous reports.¹⁶ We also calculated the light-chain ratio (κ/κ+λ) for each B-cell cluster by assessing the expression of IGKC and IGLC2, revealing robust light-chain restriction—a hallmark of malignant B cells—across all clusters (Figure 1F). Additionally, mIHC validated the cellular composition of tumor specimens at the protein level, supporting our scRNA-seq data (Figure 1G). Collectively, these observations underscore the diverse cellular landscape of PCNSL.

Four Cell States Drive Malignant B-Cell Heterogeneity in PCNSL

Targeting malignant B cells through molecular therapy has become a pivotal treatment strategy for PCNSL. However, treatment failures persist due to the heterogeneous nature of B cell populations, posing a major clinical challenge.^10,17 To deepen insights into PCNSL tumor heterogeneity, B cells from PCNSL samples were re-clustered for comprehensive downstream analysis.

Unsupervised clustering identified 6 distinct clusters based on their transcriptomic profiles and functional characteristics (Figure 2A, Supplementary Table 2). The ccmB cluster displayed a proliferative phenotype with marked expression of cell-cycle-related genes (HMGB2, TUBB). In contrast, the mB1 cluster was characterized by GC markers (VPREB3, CXCR4, BTG1).¹⁸ The mB2 cluster was enriched with anti-inflammatory cytokine IL-10 and genes involved in the NF-kB pathway (BCL2A1, NFKBIA, NFKBID).^19,20 Elevated expression of MHC molecule-encoding genes (HLA-A, HLA-C) and memory B-cell markers (KLF2, BANK1) distinguished the mB3 cluster. Ribosomal protein genes (RPL12, RPS12, RPS18) were upregulated in the mB4 cluster (Supplementary Figure 2A). Furthermore, a unique cluster, characterized by high expression of plasma cell differentiation-related genes (JCHAIN, MZB1) and enhanced immunoglobulin production, was identified and termed plasmablast-like lymphoma cells (PBLCs) (Supplementary Figure 2B).

Figure 2.

Malignant B-cell heterogeneity in primary central nervous system lymphoma (PCNSL). (A) Uniform manifold approximation (UMAP) visualization showing the re-clustered B cells in PCNSL, with cell types color-coded and annotated based on transcriptomic profiles. (B) B cells from the single-cell B-cell receptor (BCR) sequencing dataset mapped to the results of Figure 2A. (C) Bar plot displaying the clonotype distribution across B-cell clusters from Figure 2B, with the top 5 clonotypes color-coded. (D) Violin plots depicting the logarithmic-scale expression levels of CXCR4, CD83, CCR6, and JCHAIN across malignant B cell types, with cells color-coded. (E) Representative multiplex immunohistochemistry (mIHC) images of PCNSL tissues stained for CD79A, JCHAIN, CCR6, CXCR4, and CD83. Nuclei labeled with DAPI. Scale bars: 150 μm. (F) Top: Monocle 3 pseudotime trajectory analysis of malignant B cells. Cells are color-coded by inferred pseudotime. Bottom: Loess smooth curves showing CytoTRACE scores (y-axis) along the pseudotime trajectory (x-axis), indicating differentiation states. (G) Box plots representing the distribution of CytoTRACE scores across B-cell clusters in PCNSL. ****P < .0001, Wilcoxon rank-sum test. (H) Heatmap comparing B-cell clusters in PCNSL (columns) with standard tonsillar B-cell clusters from scRNA-seq profiling (rows) (Holmes et al., 2020), with Pearson correlation coefficients calculated in R. (I) Two-dimensional representation of cellular states in PCNSL. Each dot corresponds to a malignant cell, with each quadrant representing a specific cellular state. The positioning of dots reflects the relative meta-module scores.

To validate the malignancy of all 6 subclusters, single-cell BCR sequencing was performed on 2 tumors and mapped to the identified clusters (Figure 2B). Our analysis showed that over 95% of the cells grouped into 5 BCR clonotypes (Figure 2C) with similar amino acid sequences in their light and heavy chains (Supplementary Figure 2C), suggesting a potential common ancestral origin for the B cells within each cluster, indicating their malignancy.

We then focused on the 4 major clusters, which displayed elevated expression of key markers associated with different stages of the GC reaction: CXCR4 (mB1, DZ), CD83 (mB2, LZ), CCR6 (mB3, PreM), and JCHAIN (PBLCs, plasmablast) (Figure 2D). The existence of these subclusters was further confirmed at the protein level via mIHC staining (Figure 2E).

A comprehensive analysis of the differentiation states of malignant B cells in PCNSL was conducted. We reconstructed the developmental trajectory using Monocle3²¹ and CytoTRACE,²² identifying 2 endpoints with significantly lower CytoTRACE scores in mB3 and PBLCs, indicating higher differentiation levels in these clusters (Figure 2F and G).

Malignant transformation of B cells typically preserves some characteristics of their original state.^23,24 PCNSL is believed to originate from late-stage GC B cells,²⁵ yet the intratumoral cellular heterogeneity remains underexplored. To reconstruct the development stages of the B-cell clusters in PCNSL, we used tonsil GC B cells (GSE 139891) as a standard comparison.¹² Five subclusters were identified: dark zone (DZ), light zone (LZ), intermediate (INT), memory B-cell precursors (PreM), and plasmablast (PBL) (Supplementary Figure 2D and E). Meta-module gene sets for DZ, LZ, PreM, and PBL were calculated based on their differential expression (Supplementary Table 3). Using these genes, a “B-cell state identifier” was developed to map the differentiation stages of lymphoma cells (Supplementary Figure 2F). When comparing tonsillar B-cell phenotypes with PCNSL subclusters, PBLCs showed specificity for PBL, mB3 for PreM, mB2 for LZ, and mB1 for DZ (Figure 2H). These results were consistent with the “B-cell state identifier,” demonstrating that PCNSL heterogeneity is driven by 4 primary cell clusters, each representing distinct stages of differentiation (Figure 2I). Additionally, observations were consistent with the single-cell sequencing data from PCNSL samples reported by Heming et al.²⁶ (Supplementary Figure 2G) and Liu et al.²⁷ (Supplementary Figure 2H), further reinforcing the reproducibility of these cellular compositions.

PBLCs Are Associated With Poor Prognosis in PCNSL

In sDLBCL, tumors with increased plasma cell differentiation and reduced major histocompatibility complex class II (MHC II) expression are linked to poorer prognoses.^27,28 Gene set enrichment analysis (GSEA) on the DEGs in PBLCs from our scRNA-seq dataset indicated significant enrichment for the “Plasma cell vs B lymphocyte UP” gene set, suggesting a plasma cell differentiation profile (Supplementary Figure 3A). Additionally, PBLCs exhibited downregulation of MHC II molecules, including HLA-DQA1, HLA-DMB, HLA-DRA, HLA-DQB1, and HLA-DPB1 (Figure 5A). Notably, Steen et al. identified 5 distinct malignant B-cell subtypes in sDLBCL using Ecotyper,²⁹ and we mapped PCNSL B cells to these subtypes (Supplementary Figure 3B). Notably, State 5, associated with plasma cell differentiation and worse prognosis, showed a significantly lower plasmablast meta-module score compared to PBLCs, indicating a less favorable prognosis for PBLCs (Supplementary Figure 3C).

Figure 5.

Crosstalk between malignant B cells and immune cells in primary central nervous system lymphoma (PCNSL). (A, C) t-Distributed stochastic neighbor embedding (t-SNE) visualization of lymphoid cells (A) and myeloid cells (C) from tumor and blood samples of patients with PCNSL, with cell types color-coded and annotated based on their transcriptomic profiles. (B) Left: The same t-SNE plot as in (A), now color-coded by AUC scores for cytotoxic function. Right: Corresponding single-cell bar plot showing AUC scores for cytotoxic activity. (D) Left: The same t-SNE plot as in (C), color-coded by AUC scores for phagosome function. Right: Corresponding single-cell bar plot showing AUC scores for phagocytic activity. (E) Dot plot illustrating selected gene pairs involved in interactions between immune cells and plasmablast-like lymphoma cells (PBLCs) (left) or other tumor cells (right), as inferred using the InterCellDB R package. Gene pairs with P-values <.05 are displayed. Dot size indicates the interaction strength (calculated by multiplying the expression levels of both genes), and dot color represents the probability of interaction. (F) Representative multiplex immunohistochemistry (mIHC) images of PCNSL tissues stained for FAS, JCHAIN, and CD79A, with nuclei labeled using DAPI. Scale bars: 50 μm. *P < .05, paired t-test.

To assess the clinical relevance of PBLCs, their proportions were quantified via mIHC staining in 45 tumor specimens from the 56 patients. For accurate identification in paraffin sections, we compared PBLCs with mB1–mB4 cells and focused on genes associated with plasmablast differentiation. PBLCs showed elevated expression of 2 key plasmablast membrane receptors, CD27 and CD38, along with significant upregulation of JCHAIN, which encodes the joining chain of multimeric IgA and IgM (Figure 3B). mIHC staining confirmed the co-expression of CD27, CD38, JCHAIN, and CD79A in PBLCs (Figure 3C), and revealed that PBLCs constituted approximately 1.3%–8.1% of malignant B cells (Supplementary Figure 3D). Based on PBLC proportions, patients were stratified into low, medium, and high subgroups. The high PBLC group had a significantly higher proportion of PBLCs than the low PBLC group (Supplementary Figure 3E). Survival analysis demonstrated that patients with lower PBLC proportions had significantly better outcomes, with a median OS of 33 months (IQR 21.5–76.5 months) and a median progression-free survival (PFS) of 30 months (IQR 15.5–61.5 months), compared to those with higher PBLC proportions (OS: median 15 months; IQR 11.5–48 months; PFS: median 14 months; IQR 9–21 months) (Figure 3D and E).

Figure 3.

Prognostic significance of plasmablast-like lymphoma cells (PBLCs). (A) heatmap illustrating the expression of genes encoding major histocompatibility complex (MHC) class II molecules in each subcluster of primary central nervous system lymphoma (PCNSL). (B) Scatter plot highlighting the upregulated plasma cell differentiation-related differentially expressed genes (DEGs) (log₂FC > 0.25) in PBLCs compared to other malignant B cells (mBs). The x-axis shows the percentage difference, while the y-axis reflects log₂FC. DEGs are categorized as Receptor, Immunoglobulin-related, Transcription Factors (TFs), or others, with color coding based on a combined z-score of percentage difference and log₂FC. (C) Representative multiplex immunohistochemistry (mIHC) images of PCNSL tissues stained for CD38, CD27, JCHAIN, and CD79A, with nuclei labeled using DAPI. Scale bars: 150 μm. (D, E) Kaplan–Meier survival curves illustrating overall survival (OS) and progression-free survival (PFS) for patients with PCNSL. The x-axis represents time (months), while the y-axis indicates the probability of OS or PFS. Patients are divided into PBLC-high and PBLC-low groups, with survival curves compared using the log-rank test. (F) Two-dimensional plot depicting the cellular states of relapse-phase malignant B cells based on scRNA-seq data from Heming et al. (G) Representative mIHC images comparing first-diagnosed (left) and relapse-phase (right) PCNSL tissues, stained for CD38, CD27, JCHAIN, and CD79A, with nuclei labeled using DAPI. Scale bars: 150 μm. *P < .05, paired t-test.

To further explore the relationship between PBLCs and tumor recurrence, scRNA-seq data from cerebrospinal fluid tumor cells of a patient with recurrent PCNSL, as recently reported by Heming et al.,²⁶ were re-analyzed (Supplementary Figure 3F–G). This revealed a significant increase in PBLCs in recurrent PCNSL, comprising approximately 60% of the malignant B-cell population (Figure 3F). mIHC staining of specimens from patients with recurrent PCNSL also confirmed a marked increase in PBLCs proportions following relapse (Figure 3G). These results suggest a strong association between PBLC presence and PCNSL recurrence.

PBLCs Exhibit Resistance to First-Line Drugs Used in PCNSL

The current treatment strategy for PCNSL primarily involves a combination chemotherapy regimen, such as high-dose methotrexate (HD-MTX), thiotepa, and cytarabine, along with adjunct therapies such as molecular targeted agents (rituximab and ibrutinib), immunotherapy (nivolumab and CAR-T cell therapy), and autologous stem cell transplantation³⁰ (Supplementary Figure 4A). Although the integration of novel therapeutics has improved PCNSL prognosis, outcomes remain poor due to the high recurrence rate.^31,32 This underscores the importance of identifying drug-resistant cell subpopulations and potential therapeutic targets.

To explore drug resistance within malignant B cells, the AUCell toolkit was employed, revealing an upregulation of a known drug resistance gene set³³ in PBLCs (Supplementary Figure 4B), alongside increased expression of the multidrug resistance protein ABCB1³⁴ (Supplementary Figure 4C). Additionally, PBLCs exhibited significant downregulation of pathways associated with cell replication, intrinsic apoptotic mechanisms, and B cell activation signaling (Figure 4A). These findings suggest that PBLCs may demonstrate enhanced drug resistance.

Figure 4.

Drug response exploration of plasmablast-like lymphoma cells (PBLCs). (A) Functional enrichment analysis of differentially expressed genes (DEGs) in PBLCs compared to other malignant B cells (mBs). (B) Negatively enriched gene sets identified by gene set enrichment analysis (GSEA). (C, E, G) Representative immunofluorescence images of primary central nervous system lymphoma (PCNSL) tissues stained for CD79A and JCHAIN in control and drugs-treated groups. Scale bars: 100 μm. n = 3 per group. *P < .05, Student’s t-test. (D, F) Box plot displaying the expression level of MS4A1 (D) and BTK (F) across PCNSL subclusters. ****Adjusted P-value <.0001, Wilcoxon rank-sum test (PBLCs vs others). (H) Ridge plot illustrating combined drug sensitivity scores, derived by summing individual scores for methotrexate (MTX) and ibrutinib. (I) Correlation analysis between drug sensitivity scores from Figure 4H and patient recurrence times from Figure 1C. Pearson correlation coefficient = 0.79, P-value = .019.

Subsequently, a thorough screening of key drugs used in PCNSL treatment followed. MTX-based chemotherapy is a critical therapeutic strategy widely employed in treating PCNSL. GSEA revealed notable downregulation of cell-cycle-related pathways in PBLCs (Figure 4B), indicating resistance to cell-cycle-dependent chemotherapeutic agents like MTX. To validate this, an in vitro PCNSL slice culture model was developed (Supplementary Figure 4D). After 3 days of MTX treatment, a marked increase in the ratio of JCHAIN-positive cells to CD79A-positive cells was observed (Figure 4C), further confirming the resistance of PBLCs to MTX-induced apoptosis.

Rituximab, a monoclonal antibody widely used in PCNSL treatment, targets CD20 and induces cell death in malignant B cells. However, our analysis demonstrated a significant downregulation of MS4A1 (encoding CD20) in PBLCs (Figure 4D). Additionally, key regulators of CD20 expression (ETS1, CHD4) and its interacting partners (CD53, CD82) were also markedly downregulated (Supplementary Figure 4E). To assess drug response in vitro, PCNSL slice cultures were treated with rituximab for 3 consecutive days. RNA sequencing of tumor tissues from both treated and control groups revealed 191 genes significantly downregulated post-treatment, forming a distinct “rituximab treatment signature” (Supplementary Figure 4F). Notably, this treatment signature was significantly downregulated in PBLCs (Supplementary Figure 4G), indicating their resistance to rituximab. Immunofluorescence assays further showed a substantial increase in JCHAIN-positive cells within the CD79A-positive B-cell population post-treatment (Figure 4E).

Ibrutinib, a selective Bruton’s tyrosine kinase (BTK) inhibitor, disrupts the survival, proliferation, and migration of malignant B cells by irreversibly inhibiting BTK in PCNSL. In our study, BTK expression was significantly downregulated in PBLCs (Figure 4F), along with its regulators (NFKB1) and interacting partners (CD19, LYN, SYK, BLNK) (Supplementary Figure 4H). After 3 days of BTK inhibitor treatment, we observed an increased proportion of PBLCs within the CD79A-positive B-cell population in PCNSL slice cultures (Figure 4G).

Subsequently, we conducted drug sensitivity analysis using the Beyondcell toolkit,³⁵ which confirmed the reduced sensitivity of PBLCs to both MTX (Supplementary Figure 4I) and ibrutinib (Supplementary Figure 4J). To investigate the prognostic impact of drug sensitivity, individual drug sensitivity scores for MTX and ibrutinib were calculated across the 8 scRNA-seq datasets, and a combined drug sensitivity score was generated by summing these individual scores (Figure 4H). Our results revealed that patients with lower combined drug sensitivity scores were more likely to experience early relapse after chemotherapy (Figure 4I). In summary, these results highlight PBLCs’ resistance to current therapies, contributing significantly to early relapse in patients with PCNSL.

Biased Cellular Communication Contributes to the Immune Evasion of PBLCs

Strategies targeting the immune microenvironment have emerged as pivotal and transformative approaches for treating B-cell malignancies.^36–38 However, PCNSL frequently exhibits loss of HLA class I and II expression, allowing it to evade immune detection and destruction.³⁹ Given the limited understanding of immune evasion mechanisms and the complex interactions between immune and tumor cells in PCNSL, there is an urgent need for in-depth investigations of its immune microenvironment to uncover new therapeutic opportunities.

To explore transcriptomic changes in immune cells infiltrating tumors, a comparative analysis between tumor-infiltrating immune cells and their counterparts from peripheral blood and brain tissue was conducted (Supplementary Figure 5A). Nine distinct lymphocyte populations were identified (Supplementary Table 2): CD4⁺ central memory T cells (CD4 Tcm), CD4⁺ effector memory T cells (CD4 Tem), mucosa-associated invariant T cells (MAIT), CD4⁺ Tregs, CD8⁺ central and effector memory T cells (CD8 Tcm/Tem), CD8⁺ effector T cells (CD8 Teff), CD8⁺ exhausted T cells (CD8 Tex), proliferating T cells (Prolif T cells), and γδT (Figure 5A, Supplementary Figure 5B). A notable increase in CD8⁺ T cells was observed in PCNSL compared PBMCs (Supplementary Figure 5C). Using the AUCell toolkit, we assessed the cytotoxic functions of various lymphocyte populations. Our analysis revealed elevated cytotoxic activity in the CD8 Tcm/Tem, γδT, CD8 Teff, and CD8 Tex clusters (Figure 5B). Notably, CD8 Tex exhibited the highest AUCell score for cytotoxic activity, suggesting that tumor-reactive CD8 T cells are pervasively exhausted in PCNSL. Furthermore, the CD8 Tcm/Tem, γδT, CD8 Teff, and CD8 Tex clusters also displayed significant upregulation of genes associated with T-cell activation, cytotoxicity, and exhaustion compared to their PBMC counterparts (Supplementary Figure 5D). These findings highlight the central role of these clusters in T-cell-mediated cytotoxicity.

The investigation into the heterogeneity of myeloid cells within PCNSL identified 5 distinct subpopulations (Supplementary Table 2): 1 monocyte subset (Mono), 3 macrophage subsets (Macro1-3), and 1 DC subset (DC) (Figure 5C). To better characterize these myeloid cells, “Mono-sets” and “Micro-sets” scores were calculated using AUCell, based on gene sets derived from DEGs between monocytes and microglia (Supplementary Table 3). The results consistently indicated higher “Mono-sets” scores for myeloid cells within PCNSL (Supplementary Figure 5E), suggesting a monocyte lineage for the macrophages. Notably, the Macro1 subcluster displayed significant enrichment for genes involved in phagocytosis (Figure 5D), complement activation, chemotaxis, antigen presentation, and angiogenesis (Supplementary Figure 5F), underscoring its role in modulating the immune microenvironment of PCNSL.

We then explored intercellular communication using InterCellDB⁴⁰ toolkit. The analysis revealed stronger molecular interactions between T cells (CD4 Tem, CD8 Teff, and CD8 Tcm/Tem) and malignant B cells than with myeloid cells (Supplementary Figure 5G). These interactions were particularly robust within the mB1, mB2, and mB3 subclusters, but significantly weaker in the ccmB, mB4, and PBLCs subclusters (Supplementary Figure 5G).

Consistent with previous findings,^26,41 several checkpoint molecules, such as PDCD1 and CTLA4, were implicated in interactions between malignant B cells and T cells (Figure 5E). Signaling between these cells also involved chemotaxis-related gene pairs, while these gene pairs were absent in PBLCs, suggesting a diminished capacity for immune cell recruitment in PBLCs. Additionally, significant interactions between JCHAIN (in PBLCs) and immune cell receptors like Vcam1, CD7, and CTLA4 were observed, indicating a possible immune regulatory role for immunoglobulins secreted by PBLCs. Furthermore, a key interaction was identified between FASLG (CD8 T cells) and FAS (tumor cells), a mechanism through which antigen-specific T cells induce programmed cell death in tumor cells. In contrast, PBLCs showed significant downregulation of FAS at both the transcriptional (Supplementary Figure 5H) and protein levels (Figure 5F), likely contributing to their ability to evade immune responses. In summary, these altered cellular interactions could hinder the immune system’s ability to effectively target and eliminate PBLCs.

Novel Drug Predictions Targeting the Unique Regulatory Landscape Within PBLCs

Transcription factors (TFs) frequently act as key drivers of oncogenic transformation, proliferation, and survival.⁴² Recent advancements in chemical biology have challenged the notion that TFs are “undruggable,”⁴³ leading to significant progress in TF-mediated gene regulation, which has shown success in cancer treatments.⁴⁴

Subsequently, we explored the potential regulatory activity within PBLCs that drives cell survival and drug resistance. Using the pySCENIC toolkit,⁴⁵ several candidate TFs were identified with increased regulon activity in PBLCs (Supplementary Figure 6A, Figure 6A). We then computed co-expression scores between each TF and its target genes. The results revealed notably elevated co-expression scores for XBP1 and PRDM1, particularly with DLBCL driver genes,⁴⁶ plasma cell differentiation, and B-cell survival (Figure 6B, Supplementary Figure 6B).

Figure 6.

Unique regulatory landscape and novel drug predictions for plasmablast-like lymphoma cells (PBLCs). (A) Ridge plot showing the regulon activities of selected transcription factors (TFs) across malignant B-cell subclusters. (B) Bar plot illustrating the co-expression scores of transcription factors with gene sets associated with plasma cell differentiation and diffuse large B-cell lymphoma (DLBCL) driver genes. (C) Scatter plot of the mean residuals (x-axis) versus switch points (y-axis) for PBLCs. The central quadrants highlight drugs to which PBLCs show differential sensitivity, compared to other cell populations. (D) Eight predicted drugs with potential efficacy against PBLCs, identified through the Beyondcell R packages. (F) Representative immunofluorescence images of PCNSL tissues stained for CD79A and JCHAIN in control and drugs-treated groups. Scale bars: 5 μm. n = 5 per group. *P < .05, One-way ANOVA.

Following the identification of key TFs regulating survival pathways in PBLCs, potential drugs targeting these TFs were predicted using the Beyondcell (Figure 6C). Eleven transcriptionally effective compounds against XBP1, PRDM1, and their downstream targets (Figure 6D, Supplementary Table 4) were identified. Notably, lenalidomide, an immunomodulatory agent with documented single-agent efficacy in refractory/recurrent PCNSL, demonstrated enhanced sensitivity in PBLCs (Supplementary Figure 6C) and significantly reduced the proportion of JCHAIN⁺ CD79A⁺ cells in vitro (Figure 6E). In conclusion, these results suggest that targeting XBP1 and PRDM1, along with their downstream gene networks, may offer promising therapeutic strategies for treating PCNSL.

Discussion

PCNSL is a highly aggressive B-cell malignancy, with therapeutic challenges largely stemming from its significant heterogeneity.⁴⁷ In this study, we revealed that B cells within PCNSL share substantial transcriptional similarities with those found in the GC. A key discovery was the identification of a distinct subset of PBLCs, which exhibited multidrug resistance. The abundance of PBLCs was found to be associated with poor patient prognosis. Furthermore, our analysis pinpointed TFs XBP1, PRDM1 as critical regulators driving the multidrug-resistant phenotype in these cells.

Single-cell transcriptional profiling confirmed the distinct phenotype of PBLCs. Plasma cell differentiation is a hallmark feature of malignant B cells across various lymphomas, such as follicular, marginal zone, and diffuse large B-cell lymphomas. This plasmablast-like population was independently confirmed by Kobayashi et al.,⁴⁸ who further demonstrated that patients with a PBL signature and low CD3⁺ cell infiltration exhibited a worse prognosis. While their work focused on spatial heterogeneity and microenvironmental interactions, our study provides mechanistic insights into PBLC-mediated resistance through CD20, BTK, and FAS downregulation, while also identifying lenalidomide as a potential therapeutic strategy. Together, these independent observations underscore the critical role of plasmablast-like populations in PCNSL pathogenesis and treatment resistance.

Recent studies underscore the prognostic significance of CD8⁺ T-cell infiltration in PCNSL.^48,49 While our data reveal substantial CD8⁺ T-cell activation, we also observed pervasive upregulation of immune checkpoint molecules (eg, PDCD1, LAG3, HAVCR2) in these cells, consistent with an exhausted phenotype. Similarly, Liu et al. observed that both CD8⁺ Tex and proliferating CD8⁺ T cells in PCNSL extensively express both exhausted genes and cytotoxic genes. By constructing a developmental trajectory, they identified pre-exhausted CD8⁺ T cells with self-renewal capabilities, as well as terminally exhausted CD8⁺ T cells that highly express exhaustion markers.²⁷ Another independent study also identified pervasive immune checkpoint expression in PCNSL-infiltrating T cells, thereby recognizing immune evasion signals from PCNSL cells.²⁶ Together, these observations suggest that despite their activated state, CD8⁺ T cells in PCNSL exhibit functional impairment due to exhaustion, highlighting immune checkpoint inhibition as a promising strategy to restore antitumor cytotoxicity.

The present study has several limitations. First, the analysis of tumor cells from all 8 patients as a single group may obscure potential intertumoral heterogeneity in cellular composition and molecular features. Although PBLCs were consistently present across all samples, future studies should assess the variation of PBLCs between patients and explore their relationship with molecular subtypes. Additionally, the relatively low incidence of PCNSL limits the sample size, which may affect the generalizability of the findings. Furthermore, our study was limited by the lack of refractory/recurrent samples. Comparing treatment-refractory versus responsive cases would better reveal PBLC-mediated resistance mechanisms. Larger, more diverse cohorts are necessary to validate these results and evaluate their clinical relevance for PCNSL treatment. Finally, this study did not investigate the interactions between PBLCs and other tumor cell types. Future research should focus on understanding the plasticity of PBLCs and the transitions between different B-cell states to further explore their role in PCNSL.

In conclusion, this study identified a novel subset of lymphoma cells with plasma cell differentiation in PCNSL, which are resistant to multiple drugs and associated with a poor prognosis. These findings underscore the challenge of targeting this cell population and suggest that therapies aimed at plasma cell differentiation may offer a promising approach for treating B-cell malignancies.

Supplementary material

Supplementary material is available online at Neuro-Oncology (https://academic.oup.com/neuro-oncology).

Author Contributions

F.L., X.Z., and A.P. contributed equally to this work. L.S. and J.Z. designed the research. F.L., X.Z., A.P., Y.G., H.L., L.L., F.Y., J.W., and J.L. performed the experiments. X.Z., A.P., and Z.J. contributed to bioinformatics and statistical analysis. F.Z., J.X., J.C., J.Z., and F.L. provide the human specimens. X.Z. wrote the original manuscript. F.L. and L.S reviewed and polished the manuscript.

Conflict of interest statement

The authors declare no conflict of interests.

Funding

None declared.

Data Availability

The raw single-cell data files for each sample are available in the GSA-Human database under the access number HRA008594. All other data are available upon reasonable request from the corresponding author, Ligen Shi (slg0904@zju.edu.cn).