Single-cell sequencing reveals transcriptional and immunogenetic diversity in stereotyped chronic lymphocytic leukemia

Key Points

• •Distinct metabolically active leukemic cells are enriched in aggressive CLL subsets and exhibit a BcR-mTOR-MYC-OXPHOS signature.
• •Most CLL cases exhibit intraclonal immunogenetic diversification suggestive of ongoing somatic hypermutation.

Visual Abstract

Abstract

Chronic lymphocytic leukemia (CLL) comprises immunogenetically defined stereotyped subsets of patients with distinct B-cell receptor (BcR) immunoglobulin features and clinical trajectories, yet the molecular pathways underlying subset-specific differences remain incompletely characterized. To resolve disease-relevant heterogeneity obscured in bulk analyses, we performed integrated single-cell transcriptomic and immunogenetic profiling of 48 557 malignant and bystander immune cells from 13 treatment-naïve primary patient samples representing poor-prognostic subsets #1 and #2 and the indolent subset #4. Despite interpatient variability, leukemic cells exhibited pronounced subset-specific transcriptional features, with enrichment of hypoxia-related genes in subset #1, oxidative phosphorylation (OXPHOS), MYC/E2F targets, and mechanistic target of rapamycin complex 1 (mTORC1) signaling in subset #2, and negative enrichment of hypoxia, apoptosis, and reactive oxygen species–related pathways in subset #4. Notably, compared with indolent subset #4, aggressive subsets #1 and #2 harbored increased proportions of metabolically active and recently emigrated/proliferative leukemic cells, characterized by a CXCR4^dim CD5^bright transcriptional phenotype and concerted enrichment of MYC target genes, mTORC1 signaling, OXPHOS, and BcR signaling. Immunogenetic analyses revealed dominant malignant clones with evidence of intraclonal immunogenetic diversification in 11 of 13 cases. T cells were dominated by effector-memory phenotypes exhibiting progressive differentiation toward terminal cytotoxic states, accompanied by exhaustion-associated programs, and expanded T-cell clones largely restricted to terminal states. Ligand-receptor analysis further indicated costimulatory and inhibitory signaling between leukemic and immune cells. Together, this study provides an integrated view of the transcriptional and immunogenetic landscape of 3 major stereotyped CLL subsets, linking BcR immunoglobulin configuration with proliferative capacity and microenvironmental cross talk, thereby shaping clonal behavior.

Introduction

Chronic lymphocytic leukemia (CLL) is a clinically heterogeneous B-cell malignancy with outcomes ranging from indolent to aggressive.^1,2 This variability reflects the underlying molecular diversity, and multiple biomarkers now inform refined patient stratification.³ One of the most established is immunoglobulin heavy variable (IGHV) gene somatic hypermutation (SHM) status, which divides patients into IGHV-mutated (M-CLL) and unmutated (U-CLL) subgroups with distinct prognoses.^4,5 Additionally, >40% of patients express stereotyped B-cell receptor immunoglobulins (BcR IG), defined by (quasi)identical variable heavy complementarity-determining region 3 (VH CDR3) amino acid motifs, that can be grouped into subsets with shared clinicobiological features, including outcome.⁶ Among the best characterized, subset #1 (IGHV1/5/7) comprises U-CLL with aggressive disease; subset #4 (IGHV4-34/IGKV2-30) represents IgG–switched M-CLL, and is a prototype of indolent disease; and subset #2 (IGHV3-21/IGLV3-21) displays intermediate levels of SHM, frequent SF3B1 mutations (up to 45%), and poor prognosis regardless of IGHV gene SHM status.6, 7, 8, 9, 10

Patients with comparable immunogenetic features also show corresponding transcriptional similarities. Previous comparisons of U-CLL and M-CLL using microarrays and bulk RNA-sequencing (RNA-seq) have revealed distinct transcriptional signatures. In particular, U-CLL typically shows elevated expression of activation- and proliferation-associated genes such as ZAP70, TCL1A, and CD38, whereas M-CLL is enriched for genes associated with a less activated transcriptional state, including TCF7, COBLL1, and ADAM29.11, 12, 13, 14, 15, 16 In contrast, the transcriptional features of individual stereotyped subsets remain less well defined.

Although bulk approaches may mask rare subpopulations and intratumoral heterogeneity, single-cell RNA-seq (scRNA-seq) overcomes these limitations by resolving gene expression at cellular resolution, enabling the identification of discrete functional states and subclonal populations.17, 18, 19, 20, 21 It further allows characterization of the tumor microenvironment, including bystander populations, such as T cells, natural killer (NK) cells, and monocytes, which modulate CLL evolution and therapy response.22, 23, 24 When combined with single-cell V(D)J sequencing (scV(D)J-seq), it also provides information about BcR IG and T-cell receptor (TR) diversity. Despite this potential, few studies have applied single-cell analyses to study CLL within immunogenetically defined subsets.^17,18

To explore these aspects, we performed integrated scRNA-seq and scV(D)J-seq on primary CLL samples from patients belonging to subsets #1, #2, and #4. This approach enabled us to resolve both the intersubset and intratumoral heterogeneity, whereas also profiling accessory immune populations. Importantly, we identified distinct malignant subpopulations enriched for metabolic and proliferative programs, which were more abundant in clinically aggressive subsets and resembled recently emigrated/proliferative cells. At the immunogenetic level, we detected evidence of ongoing SHM, indicating continuous intraclonal diversification. Together, these analyses establish a single-cell framework for studying stereotyped CLL subsets and provide a resource for future functional investigations.

Methods

Study design and sample collection

Peripheral blood samples were collected from treatment-naïve patients with CLL diagnosed according to the 2018 International Workshop on Chronic Lymphocytic Leukemia criteria and belonging to stereotyped subsets #1 (n = 4), #2 (n = 5), and #4 (n = 4; supplemental Table 1) using previously established criteria: (1) IGHV gene usage from the same phylogenetic clan; (2) ≥50% amino acid identity and ≥70% similarity within the VH CDR3; (3) identical VH CDR3 length; and (4) identical offset of the shared amino acid motif.^2,6 Peripheral blood mononuclear cells were isolated using Ficoll density gradient centrifugation.

scRNA-seq and V(D)J-seq

Single-cell suspensions underwent droplet-based library preparation using Chromium Controller with Chromium Next GEM single-cell 5′ reagent kits version 2 (10X Genomics, Pleasanton, CA), targeting 5000 cells per sample. Libraries were sequenced on a NovaSeq 6000 (Illumina, San Diego, CA) using a 28–base pair (bp; R1) and 90–bp (R2) configuration to achieve a median read depth of 50 000 to 100 000 read pairs per cell for gene expression libraries and 5000 read pairs per cell for V(D)J (BcR IG and TR) libraries. For 1 subset #4 sample (P10), the BcR IG library was sequenced using a 2 × 150–bp configuration on a MiSeq (Illumina), with no corresponding TR library. FastQ files were evaluated with FastQC version 0.11.9, and reports were aggregated using MultiQC version 1.22.2.^25,26 Sequencing metrics are provided in supplemental Table 2A-C.

Data preprocessing

Raw scRNA-seq and scV(D)J-seq data were processed using the “multi” module of Cell Ranger version 9.0.0, with intronic reads excluded from the analysis. A reference genome and annotations from GENCODE Release 47 (GRCh38.p14) were used. For BcR IG and TR analyses, a reference was constructed using the International Immunogenetics Information System (IMGT) database. Cell Ranger metrics are provided in supplemental Table 3A-C.

scRNA-seq data analysis

scRNA-seq data were processed using Scanpy version 1.10.4.²⁷ Quality control and cell-level filtering were applied per sample based on quality control metric distributions (supplemental Figure 1A-C). Ambient RNA contamination was corrected with SoupX version 1.6.2 (supplemental Figure 2A), and doublets were identified using scDblFinder version 1.18.0 (supplemental Figure 1D).^28,29 Gene biotypes and the number of cells expressing each gene were assessed across the data set (supplemental Figure 2B-E). Library size normalization was carried out with scran version 1.32.0, and highly variable genes (HVGs) were selected using variance decomposition in 3 contexts: globally across the merged data set (no batch variable), for data integration (sample identity as the batch variable), and independently within each sample for per-sample analyses.³⁰ Data integration was performed using single-cell variational inference (scVI) and single-cell annotation using variational inference (scANVI) from scvi-tools version 1.3.1 and Harmony (via Scanpy).^31,32 Cell types were annotated with CellTypist version 1.6.3 and Azimuth version 0.5.0, and B- and T-cell identities were validated by scV(D)J-seq data.33, 34, 35 Dimensionality reduction was performed using uniform manifold approximation and projection (UMAP) and t-distributed stochastic neighbor embedding (t-SNE), and clustering was carried out using the Leiden algorithm. Pseudotime trajectories were inferred with Palantir version 1.4.1.³⁶ Differential gene expression (DGE) analyses were conducted using Scanpy for single-cell comparisons and DESeq2 version 1.42.0 for pseudobulk analyses.³⁷ Gene set enrichment analysis (GSEA) and gene set and transcription factor scores were computed with GSEApy version 1.1.9, decoupleR version 2.1.1, or Scanpy.^38,39 Further details are provided in the supplemental Methods.

scV(D)J-seq data analysis

Assembled contigs were reannotated using IMGT/HighV-QUEST version 1.9.5 and subsequently analyzed with Scirpy version 0.22.2.^40,41 Only cells with matched gene expression data were retained to ensure integrative interpretation. For BcR IG gene analysis, clonotypes were defined per sample based on CDR3 nucleotide sequence similarity and shared V and J gene usage across both heavy (IGH) and light (IGK/IGL) chains. A normalized Hamming distance threshold of 15% (corresponding to 85% sequence similarity) was applied, and a clonotype match was assigned if either chain met the similarity criteria. For TR, clonotyping was based on β/δ chains and complete CDR3 amino acid sequence identity combined with shared V and J gene usage. Public clonotypes (shared across samples) were allowed. Lineage trees were reconstructed with dowser version 2.3.⁴² Full details are provided in the supplemental Information.

Additional methods

Details on bulk IGHV-IGHD-IGHJ sequencing, external data sets, statistical analyses, and data visualization are provided in the supplemental Information.

The study was approved by the Swedish Ethical Review Authority (Etikprövningsmyndigheten), Stockholm, Sweden, and was conducted in accordance with the Declaration of Helsinki.

Results

Cohort overview and cell-type composition

We analyzed scRNA-seq data from 13 CLL samples representing 3 major stereotyped subsets: #1 (P1-P4), #2 (P5-P9), and #4 (P10-P13; Figure 1A). Integrated analysis of 48 557 cells showed that most were B cells (44 520 [91.7%]). The remaining cells corresponded to bystander immune populations, including T cells (2969 [6.1%]), NK cells (653 [1.3%]), monocytes (367 [0.8%]), and nearly negligible numbers (<0.1%) of dendritic cells and hematopoietic stem cell/multipotent progenitor cells (Figure 1B; supplemental Figure 3). On average, 3735 cells (range, 1127-9990) were captured per sample, with B cells consistently comprising most (mean, 3425; range, 940-9544; Figure 1C). Nonintegrated analysis revealed primarily patient-specific, and partially subset-associated, B-cell clustering, whereas bystander cells grouped according to their expected identities (supplemental Figure 4A-B). Per-patient embeddings confirmed stable clustering of B cells and bystander populations (supplemental Figure 5).

Figure 1.

Cohort overview, cell-type composition, and transcriptional variability across B cells. (A) Overview of the 13 patients with CLL included in the study, grouped by stereotyped subsets #1 (P1-P4), #2 (P5-P9), and #4 (P10-P13). (B) UMAP embedding of 48 557 cells after data integration, colored by major cell types. The pie chart shows the overall cell-type composition. HVGs (1000) were selected using VD and used for unsupervised scVI integration, followed by cell-type annotation refinement with scANVI. UMAP was run with default parameters using 30 latent dimensions and a minimal distance of 0.75. (C) Per-sample cell-type composition. The stacked bar plot shows the number of cells assigned to each cell type per sample. (D) Subsetting to 44 520 B cells, shown as t-SNE colored by sample identity, subset, and classification of malignant vs estimated normal B cells based on IGHV/IGKV/IGLV gene expression. (E) Mean-variance relationship of gene expression across B cells. The scatterplot shows total variance vs mean log₂ expression with HVGs (1000) highlighted. Highly variable IG genes are colored in red, and other HVGs are shown in dark gray. The ranked plot displays genes ranked by biological variance, with a dashed line indicating the HVGs used in subsequent analyses. (F) Hierarchical clustering of the HVGs (1000) across B cells. Rows represent genes and columns represent single cells, colored by sample identity and subset. Clustering was performed using Euclidean distance and Ward linkage. (G) Overlap between globally defined and per-sample HVGs (1000 each). The stacked bar plot shows the number of globally defined HVGs recovered across per-sample HVG selections, with highly variable IG genes colored in red and other HVGs shown in dark gray. The binary heat map displays the overlap across global and per-sample HVG sets (4149 unique HVGs in total). (H) Expression of subset-defining IGHV/IGKV/IGLV genes across subsets. The bar plot indicates the number and percentage of cells within each subset expressing the corresponding gene, and t-SNEs display their log₂ expression. For IGHV1-39/IGHV1D-39, mean log₂ expression across the 2 genes is shown, because they are paralogs with highly similar sequences and reads may align to either gene. HSC, hematopoietic stem cell; HVG, highly variable gene; IG, immunoglobulin; LD, latent dimension; MPP, multipotent progenitor; PC, principal component; PCA, principal component analysis; scANVI, single-cell annotation using variational inference; scVI, single-cell variational inference; t-SNE, t-distributed stochastic neighbor embedding; UMAP, uniform manifold approximation and projection; VD, variance decomposition.

Transcriptional variability across B cells

To assess transcriptional heterogeneity within the B-cell compartment, we analyzed all B cells in a nonintegrated manner while preserving patient-level structure (Figure 1D; supplemental Figure 6A). IG genes were observed as major contributors to variance, alongside HVGs such as CXCR4, DUSP1, and KLF6 (Figure 1E; supplemental Figure 6B). Hierarchical clustering of the top 1000 HVGs resolved cells by sample and subset identity (Figure 1F). Comparing global HVGs with those selected per sample (1000 HVGs each) showed a moderate overlap (382-540 HVGs per sample), with the remainder varying only within individual cases or shared among a few samples (Figure 1G; supplemental Figure 6C-D).

Assessment of IGHV/IGKV/IGLV gene expression demonstrated that >99.7% of B cells expressed subset-defining IG genes (Figure 1D), confirming the predominance of malignant B cells. A minor cluster, defined by the absence of CD5 and ROR1 expression (supplemental Figure 6E), likely corresponded to residual normal B cells. Although IG gene expression was generally consistent across subsets, subset #4 showed notably fewer cells expressing the subset-defining IGHV gene (72.5% for IGHV4-34) compared with subset #1 (99.1% for IGHV1-3 and IGHV5-10-1) and subset #2 (99.5% for IGHV3-21; Figure 1H).

Stereotyped subset-specific transcriptional programs

To identify global transcriptional differences associated with stereotyped subset membership, we performed pairwise DGE analyses using sample-level B-cell pseudobulk profiles. This revealed sets of upregulated and downregulated genes, amounting to 448 differentially expressed genes (DEGs) overall (Figure 2A; supplemental Figure 7A; supplemental Table 4A-C). Overlap analysis showed largely nonoverlapping, subset-specific genes (Figure 2B-C; supplemental Figure 7B). Among the most significant DEGs, LDOC1 and ANGPT2 were upregulated in subset #1, FGL2 in subset #2, and TCF7 in subset #4 (Figure 2D-E; supplemental Figure 7C). Consistent with their more aggressive phenotype, subsets #1 and #2 showed marked upregulation of ZAP70 and CD38 relative to subset #4. Other notable significant changes included COBLL1 downregulation in subset #1 and KLF3 and ITGAX in subset #2. At the single-cell level, most DEGs were expressed in ≤10% of cells within any subset, with only 13 genes detected in >50% of cells (supplemental Figure 8A-C). Hierarchical clustering of the 97 DEGs expressed in >10% of cells in any subset revealed partially subset-restricted profiles (supplemental Figure 8D). By comparing the scRNA-seq data with our published bulk RNA-seq data set, we observed a strong concordance in mean variance-stabilized expression and gene-wise variance, consistent fold-change directionality, and partially overlapping DEGs (supplemental Figure 9A-E).⁴³ Additional bulk RNA-seq comparisons stratified by IGHV gene SHM status, particularly within subset #2, showed that the predominant differences were more closely associated with stereotyped subset membership rather than IGHV status (supplemental Figure 9F).

Figure 2.

Stereotyped subset-specific transcriptional programs. (A) Pairwise DGE comparisons of B cells between subsets (#1 vs #2, #1 vs #4, and #2 vs #4) using sample-level pseudobulk profiles. Volcano plots show significantly upregulated genes (log₂FC >1; FDR <0.05) in red, and downregulated genes (log₂FC <-1; FDR <0.05) in blue. (B) Overlap of DEGs across comparisons. (C) Hierarchical clustering of the 448 unique DEGs identified across comparisons. Clustering was performed using Euclidean distance and Ward linkage method. (D) Bar plots showing selected top subset-specific DEGs based on log₂FC (nonannotated ENSG genes not shown), identified as shared across both comparisons involving the respective subset, and ranked by log₂FC and colored by -log₁₀FDR. (E) t-SNE showing log₂ expression of representative DEGs. (F) Summary of GSEA results across comparisons. The dot plot displays NES for pathways that were significant in both comparisons involving the respective subset. Dot size corresponds to the number of leading-edge genes, and dot color indicates -log₁₀FDR. DEG, differentially expressed gene; FC, fold change; FDR, false discovery rate; IG, immunoglobulin; NES, normalized enrichment score; t-SNE, t-distributed stochastic neighbor embedding.

We next performed GSEA using the DGE results and MSigDB Hallmark gene sets (supplemental Figure 10A).⁴⁴ Subset #1 showed enrichment for hypoxia-related genes compared with subsets #2 and #4 (Figure 2F), and, among others, for tumor necrosis factor α (TNF-α) signaling via NF-κB and p53 pathway relative to subset #2 (supplemental Figure 10B). Subset #2 was enriched for oxidative phosphorylation (OXPHOS), MYC and E2F target genes, fatty-acid metabolism, and mechanistic target of rapamycin complex 1 (mTORC1) signaling compared with subsets #1 and #4, and, among others, for glycolysis-related genes relative to subset #4. In contrast, subset #4 displayed negative enrichment scores for hypoxia, adipogenesis, apoptosis, and reactive oxygen species–related pathways when evaluated against subsets #1 and #2. Only a limited number of DEGs (0-5 per pathway) overlapped with the enriched sets (supplemental Figure 10C).

Single-cell inference of pathway activity

To examine pathway activity at cellular resolution, we computed per-cell gene set scores (supplemental Figure 11A). As a proof of concept, we applied the poor- and good-survival signatures from Hüttmann et al (Figure 3A), which, as expected, showed higher poor-survival and lower good-survival scores in subsets #1 and #2, whereas subset #4 displayed the opposite pattern.¹¹ Interestingly, across Hallmark gene sets, several pathways showed restricted activity, with OXPHOS, MYC targets, and mTORC1 signaling, among others, marking distinct subpopulations of cells within the main sample clusters (Figure 3A; supplemental Figure 11B). Gene sets related to BcR signaling from Reactome and KEGG also showed elevated scores in subsets #1 and #2 and overlapped with these subpopulations.^45,46 In addition, we scored recently emigrated/proliferative (CXCR4^dim CD5^bright) and quiescent/resting (CXCR4^bright CD5^dim) states using signatures from Calissano et al and Pozzo et al (Figure 3A; supplemental Figure 11C).^14,47 As expected, these scores were inversely correlated (supplemental Figure 11D). Cells with high proliferative scores formed localized clusters that typically overlapped with those described earlier and occurred more frequently in subsets #1 and #2, whereas cells with high resting scores were more uniformly distributed across subsets. Consistent with this, CXCR4 and CD5 displayed opposing expression trends (supplemental Figure 11E). CXCR4 expression showed the strongest correlation with BTG1, an antiproliferative factor linked to cellular quiescence (supplemental Figure 11F).⁴⁸

Figure 3.

Identification of metabolically and proliferatively active malignant B-cell subpopulations. (A) Single-cell–based gene set score distributions across subsets for good- and poor-survival signatures from Hüttmann et al,¹¹ selected Hallmark and Reactome gene sets, and recently emigrated/proliferative (CXCR4^dim CD5^bright) and quiescent/resting (CXCR4^bright CD5^dim) signatures from Calissano et al.¹⁴ For each gene set, raincloud plots show per-cell scores and summarize subset-specific differences. Scores were derived using a univariate linear model–based gene set scoring approach. Statistical significance between subsets is indicated as ∗P <.05; ∗∗P <.01; and ∗∗∗P <.001, based on 2-tailed Welch t tests. The red shading indicates the percentage of cells with scores exceeding 1 standard deviation above the global mean (z score >1), calculated across cells. (B) Pairwise Pearson correlation matrix of gene set scores across all assessed gene sets. A module of highly correlated gene sets is highlighted. (C) Radar plot showing z score–normalized gene set scores across the gene sets in the correlated module identified in panel B, with points colored by the mean z score across these gene sets (referred to as the composite activity score). The histogram shows the distribution of composite scores, with cells exceeding a mean z score of >1 classified as active-state cells. (D) t-SNE highlighting active-state cells in red. The bar plot indicates the percentage of active-state cells per sample. (E) DGE analysis comparing active-state cells with the remaining cells. Volcano plot shows significantly upregulated genes (log₂FC >1; FDR <0.05) in red, and downregulated genes (log₂FC <-1; FDR <0.05) in blue, defining the transcriptional signature of the active-state population. FC, fold change; FDR, false discovery rate; t-SNE, t-distributed stochastic neighbor embedding.

Pairwise correlations across gene sets revealed a strong association between metabolic and proliferative programs, indicating a coordinated transcriptional state within these subpopulations (Figure 3B). To derive a composite activity score across the correlated gene sets, scores were z score–normalized across cells and averaged per cell. Cells with the composite activity score >1 standard deviation above the mean (z score >1) were classified as “active-state” cells (8.6% of all cells; Figure 3C). These discrete groups of cells were most frequent in aggressive subsets #1 (4.2%-18.3%) and #2 (9.5%-28.0%) and were rarer in subset #4 (0.5%-6.3%; Figure 3D). DGE analysis comparing active-state cells with the remaining cells revealed 295 upregulated genes, defining an active-state transcriptional signature that included metabolic genes (ENO1, LDHA, COX5A, PYCR1, and FABP5), genes associated with recent BcR activation and signaling (CCL3, CCL4, DUSP4, and RGS10), as well as genes involved in immunoregulation and microenvironmental interactions (LILRA4, LILRB4, LGALS1, and TNFRSF9; Figure 3E; supplemental Figure 11G; supplemental Table 4D). Notably, 94 of these genes overlapped with the proliferation-associated gene sets reported by Calissano et al and Pozzo et al (supplemental Figure 11H).^14,47 Transcription factor activity inference using the Collection of Transcriptional Regulatory Interactions (CollecTRI) showed high MYC and NF-κB transcription factor scores within these subpopulations, along with increased activity of major histocompatibility complex II–associated regulators (CIITA, RFX5, RFXANK, and RFXAP), consistent with a metabolic and proliferative state accompanied by an immune interaction program (supplemental Figure 11I-K).⁴⁹

Intratumoral transcriptional heterogeneity

We next examined whether broader transcriptional heterogeneity across B cells was detectable within individual samples. Clustering revealed discernible clusters in 9 of 13 cases; a single dominant malignant B-cell cluster was present in 2 of 4 subset #4 cases, but in only 1 case each in subsets #1 and #2 (Figure 4A; supplemental Figure 12A). Healthy B cells, when present, typically appeared as small, separate clusters (eg, P3 and P13). To illustrate intratumoral heterogeneity, we present 1 informative sample per stereotyped subset, prioritizing samples with multiple transcriptional clusters. In the subset #1 case (P1; Figure 4B; supplemental Figure 13A), clustering resolved 4 malignant B-cell clusters. Clusters 2 and 3 displayed transcriptional profiles consistent with relatively quiescent states. Cluster 1, the largest, showed partial enrichment for signatures associated with recently emigrated/proliferative cells, whereas cluster 4, the smallest, demonstrated the strongest enrichment for MYC targets, BcR signaling, mTORC1 signaling, and OXPHOS, as well as for our active-state signature (supplemental Figure 12B-C), indicating a population of metabolically and proliferatively primed cells. The subset #2 case (P5; Figure 4C; supplemental Figure 13B) exhibited the most pronounced heterogeneity, with cluster 2 positioned distinctly in the UMAP space. Among the 3 clusters, only cluster 3 showed clear enrichment for the active-state signature, as well as BcR signaling, MYC targets, and mTORC1 signaling gene sets. Because this was an SF3B1-mutated case, we asked whether the observed heterogeneity reflected underlying genetic subclonality. Targeted genotyping for the SF3B1 p.K700E variant yielded data for only 35 cells, with mutated cells broadly distributed across the embedding rather than confined to any cluster (supplemental Figure 13C). The subset #4 case (P13; Figure 3D; supplemental Figure 13D), despite the lower number of total cells analyzed, also displayed intrasample heterogeneity, with 3 malignant B-cell clusters and 1 healthy B-cell cluster. Among the malignant clusters, cluster 2 showed features of active-state cells, whereas cluster 3, a small subpopulation, exhibited selective enrichment for TNF-α signaling with upregulation of many NF-κB–related genes.

Figure 4.

Cluster-level transcriptional heterogeneity within selected stereotyped subset samples. (A) Overview of cluster composition across B cells of individual samples. The stacked bar plot shows the percentage of cells assigned to each cluster per sample. HVGs (1000) were selected by VD. UMAP was run with default parameters using 10 PCs and a minimal distance of 0.25. Clustering was performed using Leiden clustering with sample-specific resolution parameters. (B) Detailed cluster characterization for sample P1. The UMAP embedding is colored by clusters, with the number of cells per cluster indicated. The heat map with hierarchical clustering summarizes cluster-specific DEGs, identified by comparing cells within each cluster with the remaining cells in the sample. Clustering was performed using Euclidean distance and the Ward linkage method. The dot plot summarizes overrepresentation analyses based on upregulated DEGs for each cluster. Dot size corresponds to the number of genes contributing to each enrichment, and dot color indicates -log₁₀FDR. Only the top 10 enriched gene sets per cluster are shown. (C) Same as panel B but for sample P5. (D) Same as panel B but for sample P13. DEG, differentially expressed gene; FDR, false discovery rate; PC, principal component; PCA, principal component analysis; Res, resolution; UMAP, uniform manifold approximation and projection; VD, variance decomposition.

BcR immunoglobulin gene profiling

Of the 44 520 B cells, 44 411 (99.8%) contained V(D)J information, corresponding to cells with at least 1 captured heavy (IGH) or light (IGK/IGL) chain gene rearrangement. We assessed IG gene usage (supplemental Figure 14A-C) and performed clonotyping using a chain-flexible approach that relied on either chain, crossvalidated in cells with paired chains. This was necessary to recover clonotypes in cells with incomplete heavy-chain capture (supplemental Figure 14D-E), most notably in subset #4. We identified 13 dominant malignant clonotype clusters, each corresponding to 1 of 13 samples (Figure 5A; supplemental Figures 15A-B and 16A), alongside 108 minor normal B-cell clonotypes, which typically comprised single cells, pairs, or groups of 3 cells (Figure 5B; supplemental Figure 15C). Consistent with the transcriptomic classification, nearly all B cells with clonotype calls belonged to malignant clonotypes (44 291 [99.7%]). Malignant clonotypes showed the expected stereotyped VH–VK/L CDR3 pairings (supplemental Figure 15D-E), with minor VH CDR3 sequence variation in a small fraction of cells. IGHV gene SHM status followed the subset assignments: unmutated in subset #1, mixed in #2, and mutated in #4, with corresponding trends in IGKV/IGLV genes (supplemental Figure 15F-G).

Figure 5.

Immunogenetic profiling and clonal architecture of B cells. (A) Clonotype clustering of B cells based on CDR3 nucleotide sequence similarity. Each circle represents a group of cells sharing an identical CDR3 sequence, with circle size reflecting the number of cells, and colors indicating chain-pairing configurations. Lines connect groups whose CDR3 sequences are highly similar, thereby defining a clonotype cluster. (B) Distribution of clonotype cluster sizes (1 cell, 2 cells, 3 cells, or ≥904 cells), alongside t-SNE in which cells carrying malignant or healthy B-cell clonotypes are colored by sample identity. (C) Subclonality analysis of malignant B cells based on full-length IGH nucleotide sequence identity. Circles mark major clones and their subclones, with circle size reflecting the number of cells, and colors indicating IGHV gene identity. Outlined red circles mark missense mutations, and black dots denote mutations within CDRs. t-SNE shows IGHV gene identity and the distribution of subclones, colored by sample identity. CDR, complementarity-determining region; t-SNE, t-distributed stochastic neighbor embedding.

Of all leukemic cells with clonotype information, 38 528 displayed full-length IGH sequences (considering only sequences shared by at least 2 cells), enabling subclonality analysis. We detected intraclonal diversification in 11 cases, with prominent subclones (≥10 cells) in 2 subset #1 (P1, P3), 1 subset #2 (P8), and 1 subset #4 cases (P10; Figure 5C; supplemental Figure 16B; supplemental Figure 17A-C). These subclones showed no transcriptomic divergence from the major clones, intermixing across the t-SNE space. Similar but less pronounced subclonal patterns were observed for IGK/IGL (supplemental Figures 15H and 16C). Bulk IGHV-IGHD-IGHJ sequencing independently validated the scV(D)J-seq–derived sequences, showing complete concordance for the major clones and for several subclones arising from intraclonal diversification (supplemental Figure 18A-E).

Characterization of circulating immune cells and leukemic-immune cross talk

Bystander immune populations were dominated by T cells (2969 cells; 6.1% of all cells), largely derived from 5 cases spanning all stereotyped subsets. The T-cell compartment was enriched for CD8⁺ effector memory (T_EM; 51.4%) and CD4⁺ central memory (T_CM; 30.7%) T cells, with all other subtypes individually accounting for 0.3%-4% (Figure 6A-B; supplemental Figure 19A). Pseudotime trajectory analysis of CD8⁺ T_EM cells revealed a continuum from early to terminal effector-memory/cytotoxic (CTL) states (Figure 6C), defined by reciprocal early (GZMK, TCF7) and late (GNLY, GZMB) gene expression modules (supplemental Figure 19B). Along this trajectory, CD8⁺ T_EM cells upregulated exhaustion-associated genes (LAG3, TIGIT, EOMES, TOX), resulting in elevated exhaustion scores also observed in double-negative and regulatory T cells, whereas CD4⁺ subtypes displayed a weaker exhaustion signature (Figure 6D-E; supplemental Figure 19C). CD4⁺ T-cell polarization was heterogeneous and overall limited, indicating predominantly uncommitted states (supplemental Figure 20A-C). To contextualize these T-cell states, we assessed genes related to immune activation, inhibition, and leukemic-immune interactions (Figure 6F). CD27 and CD28 were broadly expressed across T-cell subtypes but reduced in terminal CD8⁺ T_EM cells, whereas CD40LG and ICOS were enriched in CD4⁺ T_CM/T_EM/CTL and regulatory T cells populations. TNFRSF4 (OX40) localized mainly to CD4⁺ T_CM/T_EM cells, whereas TNFRSF9 (4-1BB) was restricted to early CD8⁺ T_EM cells.

Figure 6.

Characterization and immunogenetic profiling of T cells. (A) UMAP embedding of 2969 T cells after data integration, colored by T-cell subtype annotation. The pie chart shows the overall cell subtype composition. HVGs (1000) were selected using VD and used for unsupervised scVI integration, followed by cell-type annotation refinement with scANVI. Harmony-based supervised integration was used as an intermediate step to assist cell-type annotation. UMAP was run with default parameters using 30 latent dimensions and a minimal distance of 0.25. (B) Per-sample T-cell subtype composition. The stacked bar plot shows the number of cells assigned to each cell subtype per sample. For P5, no T cells were detected, and for P10, scV(D)J-seq of the TR was not performed. (C) Pseudotime analysis of CD8⁺ T_EM cells illustrating a continuum from early to terminal effector-memory states. Line plots show 2 gene modules with distinct expression dynamics along pseudotime, corresponding to early- and late-associated transcriptional programs. (D) Dot plot showing mean expression of transcriptional regulators and checkpoint receptors associated with T-cell exhaustion across T-cell subtypes. Dot size indicates the percentage of cells expressing each gene per cell subtype, and dot color represents mean log₂ expression. (E) Composite exhaustion score derived from exhaustion-associated genes, shown as raincloud plots and the UMAP embedding. Score was computed as the average expression of these genes relative to a randomly sampled reference set of genes. (F) UMAP embedding showing expression of genes mediating direct contact and costimulatory signaling to malignant B cells. Dot size denotes the percentage of cells expressing a gene per subset, and dot color encodes mean log₂ expression. (G) Clonotype analysis of T cells based on β/δ-chain CDR3 amino acid sequence identity. The pie chart shows the percentage of cells with available clonotype information and those lacking V(D)J information or β/δ chain assignments, with corresponding UMAP embedding. The bar plot shows the distribution of clonotype sizes (1 cell, 2 cells, 3 cells, 4 cells, or ≥5 cells), with the zoomed-in bar plot illustrating the composition of expanded clones. Expanded clones (≥5 cells) are highlighted in the UMAP embedding and colored by sample identity. dnT, double-negative T cells; gdT, gamma delta T cells; HVG, highly variable gene; LD, latent dimension; MAIT, mucosal-associated invariant T cells; scANVI, single-cell annotation using variational inference; scVI, single-cell variational inference; UMAP, uniform manifold approximation and projection; T_CM, central memory T cells; T_EM, effector memory T cells; T_REG, regulatory T cells; VD, variance decomposition.

We next assessed genes involved in costimulatory, survival, and inhibitory signaling to delineate interactions between malignant and bystander immune cells, including NK cells, monocytes, and dendritic cells. Within this broader cellular context, the BAFF/APRIL axis was evident, with pronounced expression of TACI in subset #2, but also coinhibition mediated by CTLA4 (supplemental Figure 21A-B). Mapping of inferred ligand-receptor interactions further highlighted putative communication networks between leukemic and bystander populations (supplemental Figure 21C-E). Although interaction patterns were largely comparable across subsets, subset #1 showed increased LAG3–major histocompatibility complex II involvement, whereas subset #2 was enriched for LILRB2-associated interactions.

Immunogenetic profiling of T cells

Of 2969 T cells, 2251 (75.5%) contained V(D)J information (supplemental Figure 22A), reflecting capture of at least 1 TR chain. Clonotyping based on β and δ chains identified 1554 β-chain clonotypes across 2114 cells and 2 δ-chain clonotypes across 2 cells (supplemental Figure 22B). A subset of expanded clones (≥5 cells; 28 clonotypes across 391 cells) was detected primarily in 2 samples (P3 and P6) among terminal CD8⁺ T_EM and CD4⁺ CTL cells (Figure 6G; supplemental Figure 22C). TRBV20-1 and TRBV28 were the most frequently used TRBV genes across all clones, whereas the expanded clones predominantly used TRBV19 and TRBV28 (supplemental Figure 23).

Discussion

The biological basis underlying the distinct clinical trajectories and treatment responses of stereotyped CLL subsets is only partly understood. Although previous bulk transcriptomic studies have provided important insights into CLL biology, they extensively mask cellular heterogeneity, underscoring the need for single-cell approaches.6, 7, 8^,10 In this study, we integrated single-cell transcriptomics with immunogenetic profiling to systematically characterize transcriptional programs, intratumoral heterogeneity, and the bystander immune landscape across representative subsets.

As expected, patients in subsets #1 and #2 generally had shorter time to first treatment than those in subset #4. Consistent with this clinical behavior, both pseudobulk and single-cell analyses recapitulated the corresponding transcriptional signatures. Subset #1 was characterized by activation of prosurvival and stress-related pathways, including TNF-α signaling via NF-κB and p53, in line with its aggressive phenotype and previous reports of heightened NF-κB activity in U-CLL and increased incidence of NFKBIE deletions in this subset.^50,51 Subset #2 displayed a pronounced BcR-mTOR-MYC-OXPHOS axis signature, previously linked to increased metabolic demand and proliferative capacity in poor-prognosis CLL.⁵² Subset #1 also exhibited these features to a lesser extent, whereas, in contrast, subset #4 showed a more quiescent profile consistent with its more indolent course.

Building on these established observations, our data further demonstrate that, at single-cell resolution, these programs were not uniformly expressed across malignant cells but instead confined to distinct cellular fractions. Specifically, the composite activity score revealed that metabolic and proliferative programs were largely restricted to small fractions of active-state cells, most prominent in subsets #1 and #2. These cells were enriched for MYC target genes, mTORC1 signaling, OXPHOS, BcR signaling, and the CXCR4^dim CD5^bright transcriptional phenotype, with proportions consistent with their reported frequencies in the peripheral blood.^14,47 Although their identification relied on transcriptional signatures rather than direct functional assays, the observed profiles align with previously described recently emigrated/proliferative cells defined by phenotypic studies.^14,47 Notably, various metabolic genes, including LDHA, ENO1, and COX5A, overlapped with the signature reported by Calissano et al, supporting the contribution of metabolic programs to this population. Beyond this overlap, our active-state signature also included genes such as CCL4, DUSP4, EGR3, PYCR1, and TNFRSF9, suggesting a wider activation state encompassing recent BcR stimulation and metabolic rewiring, together with potential engagement with the microenvironment. In line with this, LGALS1, encoding galectin-1, has been implicated in CLL microenvironmental support and immune modulation, whereas recent work on galectin-9/TIM-3 interactions supports a broader role for galectin family members in CLL immune cross talk.⁵³

Detectable only at single-cell resolution, this heterogeneity reflects functional diversity among malignant B cells, arising from varying proportions of recently emigrated lymph node–derived cells and more quiescent circulating counterparts, and highlights the dynamic interplay between proliferative and resting states. Our observations also warrant caution when interpreting bulk transcriptomic data, because signals from these active-state fractions may disproportionately influence aggregated expression profiles.¹⁴

Across individual samples, clustering revealed discernible but modest transcriptional differences among malignant B cells, with more pronounced heterogeneity observed in subsets #1 and #2 than in subset #4. Although multiple clusters were identified in several cases, these showed limited differences beyond the small active-state fractions described earlier, suggesting that most variability reflects functional states rather than stable subclonal lineages. The SF3B1-mutated case is noteworthy, because SF3B1 mutations are known to induce transcriptional and splicing changes in CLL; however, cells carrying the mutation did not segregate into a separate cluster.

Parallel V(D)J profiling revealed evidence of ongoing intraclonal diversification through SHM in most cases. Bulk IGHV-IGHD-IGHJ sequencing confirmed complete concordance for major clonotypes and provided orthogonal support for the intraclonal variants detected at single-cell level, reinforcing the robustness of the single-cell V(D)J assemblies. These subclones were transcriptionally indistinguishable from the dominant clone, suggesting that despite immunogenetic diversification, CLL subclones maintain a stable cellular transcriptional state.

Because the CLL transcriptome is strongly shaped by extrinsic cues, we examined the circulating components of the tumor microenvironment. Although the limited number of bystander cells restricted detailed subset-specific analyses, we observed a T-cell compartment dominated by CD8⁺ T_EM cells across subsets, followed by CD4⁺ T_CM cells and smaller specialized subtypes. This composition is consistent with the characteristic low CD4-to-CD8 ratio and expansion of cytotoxic populations in CLL.^54,55 The relatively high representation of CD8⁺ T_EM cells and expanded clonotypes implies sustained antigenic stimulation, aligning with earlier reports of chronic activation and functional impairment in CLL-associated T cells.^53,56,57 Exhaustion-associated transcripts spanned a continuum from GZMK⁺ early to GZMB⁺ terminal effectors, suggesting early engagement of transcriptional programs such as the NFAT–TOX/NR4A axis.^58,59 Although progressive exhaustion coexisted with retained helper potential across subsets, this underscores T-cell functional plasticity and its contributions to immune modulation and disease persistence in CLL. In this context, the concurrent expression of activating and inhibitory molecules suggests that leukemic cells receive survival cues while limiting pressure from immune cells. Although we did not observe major differences in T-cell composition or functional states between subsets, suggesting that these immune features are largely shared, higher expression of ANGPT2 in subset #1 may reflect increased microenvironmental interactions. Previous studies have shown that leukemic cells can secrete ANGPT2 and that its expression may be induced by hypoxic and inflammatory stimuli, which is consistent with the enrichment of hypoxia-related genes and TNF-α signaling via NF-κB observed in subset #1.^60,61

This study reports subclonal features of aggressive and indolent stereotyped subsets that cannot be resolved by bulk RNA-seq; however, some limitations should be considered. The number of patients included within each stereotyped subset was limited, and larger cohorts would be required to assess the generalizability of the observed heterogeneity. Increasing the number of profiled cells or using single-cell technologies with higher sensitivity may further improve the detection of subtle populations, facilitate the characterization of intraclonal immunogenetic diversity, and enable a more comprehensive assessment of bystander immune cells, including T cells. Analyses of other lymphoid tissues, such as lymph nodes or bone marrow, would provide additional insights into spatially restricted subpopulations and their microenvironmental interactions. Finally, although the scV(D)J-seq coverage was comprehensive, incomplete recovery of heavy-chain sequences was observed in some cases, particularly in subset #4, potentially reflecting the higher SHM burden characteristic of this subset.

In summary, our integrative single-cell approach provides a comprehensive view of transcriptional programs across the 3 major stereotyped CLL subsets, while simultaneously resolving intratumoral heterogeneity, ongoing BcR IG diversification, and phenotypic characteristics of bystander cells. A key observation is the presence of metabolically and proliferatively active-state subpopulations that contribute to functionally relevant heterogeneity and reflect the clinicobiological differences between subsets. Although bulk sequencing remains indispensable for large-scale analyses, single-cell profiling offers complementary resolution of cellular states, clonal architecture, and microenvironmental context, thereby refining our understanding of CLL biology across immunogenetically defined subsets.

Conflict-of-interest disclosure: B.O. reports employment with Basic Genomics. S.A.P. reports honoraria from AbbVie, AstraZeneca, BeiGene, Genentech, Janssen, Kite Pharma, Merck, MingSight Pharmaceuticals, NovalGen, Pharmacyclics, and Sobi; and research funding from AstraZeneca, Genentech, Janssen, and Merck. N.E.K. reports an advisory board role with AbbVie, AstraZeneca, BeOne Medicines, Janssen, and Pharmacyclics; a data safety monitoring committee role with AstraZeneca, Bristol Myers Squibb, Celgene, and Dren Bio; and research funding from AbbVie, Acerta Pharma, AstraZeneca, Genentech, Janssen, Merck, and Pharmacyclics. K.S. reports honoraria from AbbVie, AstraZeneca, and Janssen; and research funding from AbbVie, Gilead Sciences, and Janssen. R.R. reports honoraria from AbbVie, AstraZeneca, Eli Lilly, Illumina, Janssen, and Roche. The remaining authors declare no competing financial interests.