Spatial and molecular profiling of the mononuclear phagocyte network in classic Hodgkin lymphoma

Classic Hodgkin lymphoma (cHL) accounts for between 12% and 15% of lymphoma diagnoses and occurs most frequently in adolescents and young adults, with another peak of incidence in older adults. Multiagent cytotoxic chemotherapy (including selective use of radiotherapy) is curative in between 80% and 95% of patients; in adults the potential for cure is largely determined by the disease stage.^1-3 However, for patients with relapsed or refractory disease, the prospect of cure or long-term disease control is significantly diminished.

A cHL tumor is defined by CD30⁺ Hodgkin Reed-Sternberg cells (HRSCs), comprising between 1% and 5% of cells, and derived from germinal center (GC) B cells.⁴ HRSC survive within a tumor microenvironment (TME) containing abundant T cells and a network of mononuclear phagocytes (MNPs).⁴ HRSCs establish an immunosuppressive niche through multiple mechanisms, including downregulation of major histocompatibility complex (MHC) proteins and enhanced expression of programmed cell death ligand-1 (PD-L1), resulting from 9p23-p24 copy number gain or amplification.^5,6 PD-L1 is also abundantly expressed by myeloid cells surrounding HRSCs, leading to the notion that MNPs in cHL are largely tumor-associated macrophages (TAMs).^7,8 

Recent cross-tissue surveys of MNP diversity, using single-cell approaches, have characterized multiple states of monocytes, macrophages, and dendritic cells, identifying significant differences in functional properties and roles in pathology.⁸ Although MNPs are frequently observed in cHL, details of their functional specialization are not known.

Here, we report the use of complementary high-resolution techniques to map MNP diversity in the microenvironment of cHL. We found that tumor-associated MNPs comprise multiple cell types and occupy distinct and prognostically relevant niches, and that MNPs adjacent to HRSCs exhibit immunosuppressive phenotypes. These results reveal the complexity of MNP networks in cHL, and critical intercellular communication pathways that may be therapeutically tractable.

Healthy lymph nodes, unaffected by neoplastic disease, were acquired from adult donors by donation after circulatory death at the time of organ retrieval for transplantation by the Cambridge Biorepository for Translational Medicine with ethical approval (15/EE/0152, East of England, Cambridge South research ethics committee) and consent from donor families (Table 1). Lymph node samples were collected from inguinal, mesenteric, and thoracic regions at the end of organ donation, within 1 to 2 hours of cessation of circulation under cold ischemic conditions. Lymph node specimens were transported in ice-cold 0.9% saline. Tissue was processed as previously described.⁹ Lymph node data (290b, 298c, and 302c) are shared with James et al and were processed with additional flow sorting.⁹ Patients who donated diseased tissue for single-cell RNA sequencing (scRNAseq) were enrolled in the “Investigating how childhood tumors and congenital disease develop” study (National Health Service [NHS] National Research Ethics Service [16/EE/0394]).

scRNA-seq data from cHL specimens were kindly shared by the laboratory of Christian Steidl (University of British Columbia) as a raw-counts matrix.¹⁰ Mapping and quantification of scRNA-seq data generated for this study (WSI data set) was performed using kallisto-bustools (version 0.24.4),¹¹ against the GRCh38 human reference. The unfiltered count matrix was profiled using defaultDrops in the DropletUtils package (version 1.10.3),¹² in R (version 4.0.4).

We performed data integration with batch correction (using donor as the batch key) and dimensionality reduction concurrently by training a single-cell variational inference model using the scvi-tools Python package (version 0.9.0a0).¹³ We completed quality control on the integrated data set (supplemental Methods; supplemental Figure 1A-B; available on the Blood website). A 15-dimensional latent representation was used as input for nearest-neighbor graph construction, uniform manifold and approximation (UMAP), and Leiden clustering in Scanpy. Marker genes were calculated using the tf-idf metric and genes were ranked by the tf-idf metric, and P values calculated by a hypergeometric test, corrected for multiple testing (Benjamini-Hochberg method).¹⁴ Cell type annotation was performed using marker genes and canonical marker expression (supplemental Figure 1C), supplemented with labeling using the celltypist Python package (0.1.15).¹⁵ Concordance between annotated cell-type labels and predicted reference annotations was assessed by calculating the Jaccard distance between cell label strings (supplemental Figure 1D-E). Cell-type predictions between the King et al B-cell atlas¹⁶ was calculated by training a multilayer perceptron classifier on the reference data and calculating predictions on unseen test data, using the scikit-learn package (0.24.2).¹⁷ 

Relative abundance analysis was performed using the MiloR package (0.99.1).¹⁸ The annotated data set was represented as a k = 30 nearest-neighbor graphs calculated in scVI latent space. This graph representation was partitioned into 22 573 overlapping neighborhoods, and cell counts per 10× channel was calculated per neighborhood. Differential abundance between disease and health was calculated using the testNhoods function, setting the design as ∼center + disease, and norm.method = “TMM.”

Ligand-receptor analysis was performed using CellPhoneDB (2.1.7).¹⁹ The count matrix was normalized to 1e4 total counts before input to the Python implementation of CellPhoneDB, which was run with --iterations = 1000, --threshold = 0.1, and --result-precision = 3.

The GEO accessions GSE39133 and GSE17920 were downloaded using the GEOquery R package (version 2.60.0).²⁰ 

For GSE39133, we performed differential expression between GC and HRSC samples using limma (version 3.48.1)²¹ and derived gene sets by retaining genes with an absolute log fold change of >3 and an adjusted P value <.01.

For GSE17920, we performed WGCNA using CEMiTool (1.20.0),²² and annotated these modules using enrichment analysis against cell-type marker gene sets and Gene Ontology terms using ClusterProfiler (version 4.4.4).²³ 

Transcription factor regulons were acquired using the DoRothEA R package (version 1.3.3).²⁴ Regulons with confidence levels “A to C” were used to calculate transcription factor activities using run_viper.

We derived a network graph of transcription factors and targets, or ligands and receptors, using data from the OmnipathR R package (version 3.12),²⁵ plotting the interaction networks using the ggraph R package (version 2.0.5).

The study was approved by a UK NHS Health Research Authority Research Tissue Bank (Novopath Biobank, Newcastle upon Tyne Hospitals NHS Foundation Trust [17/NE0070]). Novopath released link-anonymized, formalin-fixed and paraffin embedded (FFPE) tissue, from patients with cHL. Tissue samples were surplus to diagnostic requirements and were released in accordance with the terms of the ethical approval.

Excised lymph node biopsies were selected from the pathology archives of Newcastle upon Tyne Hospitals NHS Foundation Trust, by members of Novopath, with appropriate ethical approval (17/NE0070). Limited, link-anonymized clinical data for patients in this study were collected by biobank staff. Hematoxylin and eosin (H&E)-stained and immunohistochemical (IHC) tissue sections were reviewed by a hematologist, together with original pathology reports. Fifty-four tumors were selected for the study, including biopsies from patients of all ages and histomorphological subtypes (Table 2). Patient biopsies were selected on the basis of high-quality, whole lymph node resection biopsies. Patients with coexisting malignant diagnoses were excluded from the study, even if these were not apparent in the tissue.

Ten of 54 FFPE cHL tissue biopsies used in the multiplexed immunofluorescence analysis were selected for in situ transcriptomics (Table 2), using the Nanostring GeoMx platform (Nanostring, Seattle, WA). Biopsies were selected based upon year of fixation (2014-2019), high-quality multiplex staining, and successful RNAscope quality control (QC) tests (Advanced Cell Diagnostics Inc, Newark, CA). Five test slides were prepared, each arrayed with 2 unique, whole tissue biopsy sections; 1 slide also included reactive lymph node tissue. For each test slide, 4 consecutive 4-μm-thick sections were taken from each tissue block, prepared per manufacturer’s instructions: slide 1, H&E; slide 2, CD45 and PAX5 fluorescent immunostains; slide 3, CD274 (PD-L1) and PDCD1 (PD-1), RNAscope probes, DNA counterstain; slide 4, oligonucleotide RNA probe hybridization.

Circular regions of interest (ROIs; uniform diameter of 300 μm) were labeled upon the CD274/PDCD1 (PD-L1) RNAscope slides, with reference to H&E and IHC slides for histomorphological context. PD-L1 was used to identify HRSCs, rather than CD30, as fluorescent labeling with CD274 probes was more robust in optimization experiments. In all selected tumors, the HRSC-dense areas of tissue were known to be PD-L1⁺, and HRSCs were identified by 4′,6-diamidino-2-phenylindole (DAPI) signal and nuclear morphology. For each tumor, 3 ROIs were selected in PD-L1^high areas and 3 ROIs in PD-L1^low areas of tissue, unless otherwise stated. Assessments of PD-L1 intensity were made by visual inspection of normalized CD274 probe fluorescent signal within each tumor. Sclerotic bands, GCs, and areas of low cellularity were avoided in cHL tissues, with reference to DNA counterstaining and paired IHC. In the control reactive node tissue, 3 ROIs were labeled in germinal centers and 3 ROIs in interfollicular areas. RNA oligonucleotides from the Cancer Transcriptome Atlas (n = 1812 probes), coupled to photocleavable oligonucleotide tags (barcodes), were hybridized overnight. Hybridized barcodes were released from tissue by UV exposure from each ROI sequentially, before being counted and sequenced on the Illumina HiSeq 2500 platform.

QC excluded outlier probes, and data were normalized against the 75th percentile of signal from each tumor. After QC, 1372 (75.7%) probes were retained. The nanostring expression matrix (ROI by gene) was analyzed by principal component analysis using Scanpy, before constructing a shared–nearest neighbor graph representation. We identified k = 10 nearest neighbors in principal component analysis space using the FNN R package, before calculating the Jaccard distance between neighbor vectors per the approach of Levine et al.²⁶ The graph was clustered using the Leiden algorithm²⁷ with default settings.

Deconvolution was performed by first generating a matrix of 1-hot-encodings of marker genes per cell type, including HRSCs. We then used the rlm function in the MASS R package to perform robust linear regression. The resulting coefficients were treated as fractions to derive counts by multiplying by the number of nuclei in each ROI. These cell count estimates were then used to calculate differential abundance estimates for each cluster using DESeq2.²⁸ Differential expression analysis between clusters was performed using limma (version 3.48.1).²¹ The top 10 differentially expressed genes per cluster were selected for plotting (supplemental Figure 4C).

Multiplexed immunofluorescence (IF) was performed by sequentially immunostaining 4-μm-thick tissue sections from selected FFPE biopsies, with primary antibodies, secondary reagents, and unique fluorochromes. Each of the 2 multiplexes comprised 6 primary antibodies, followed by a nuclear counterstain, 4',6-diamidino-2-phenylindole, as per published protocols.^7,29,30 All immunostaining was performed on the automated Ventana Benchmark platform (Roche Diagnostics, Rotkreuz, Switzerland). Slides were air dried, mounted with Prolong Diamond antifade mounting medium (#P36965; Life Technologies), and stored at 4°C in a light-proof box until image acquisition. Target antigens, antibody clones, and dilutions for markers are listed in Table 3.

Test regions for multiplex IF analysis were initially identified for each tumor biopsy (n = 54) by reviewing H&E and immunofluorescence tissue sections (original magnification ×4). Three nonoverlapping ROIs were then acquired at original magnification ×20 for each multiplex panel. Each of these 3 test regions comprised 9 contiguous fields of view (FOVs), arranged as a 3 × 3 rectangular grid (measuring 2008 μm × 1508 μm in total). In total, there were 27 FOVs per tumor, for each multiplex.

Regions were selected to exclude tissue artifacts, best represent the overall tissue, and to include CD30⁺ HRSCs. Regions were imaged and deconvoluted using the Vectra multispectral imaging platform (Vectra 3, Inform 2.4.8, Akoya Biosciences, Marlborough, MA), using specific spectral libraries.

To overview the cellular ecosystem of cHL, we compiled a census of 243 753 single-cell transcriptomes from lymphoma-affected and unaffected lymph nodes. We sourced data from Aoki et al¹⁰ incorporating droplet encapsulation single-cell transcriptomes (10× Genomics) from reactive and cHL nodes (Figure 1A). In addition, we performed scRNA-seq (10× Genomics) with cell suspensions from lymph nodes from deceased organ donors, and 2 lymphoma lymph nodes, 1 with nodular sclerosis cHL (NSCHL), and 1 with nodular lymphocyte–predominant Hodgkin lymphoma/nodular lymphocyte–predominant B-cell lymphoma (NLPHL/NLPBL), a biologically distinct and recently reclassified subtype.³¹ The combined data set comprises tissue from 13 donors who did not have lymphoma (8 deceased donors, 5 donors with reactive lymph node hyperplasia), 1 NLPHL/NLPBL, and 23 cHL lymph nodes (Figure 1A; Table 4).

After QC (supplemental Methods; supplemental Figure 1A-B), we performed data set integration and dimensionality reduction using single-cell variational inference¹³ and annotated cell types on the basis of marker genes and external data set validation (Figure 1B; supplemental Figure 1C-F).

The cellular ecosystem of cHL encompassed subsets of CD4⁺ and CD8⁺ T cells, including regulatory T cells (Treg), T-follicular helper cells (Tfh), and exhausted CD4⁺ (ThExh) and CD8⁺ T cells (CD8 TExh) (Figure 1B). ThExh matched the recently identified LAG3⁺ subset, and expressed checkpoint molecules and exhaustion markers including CD27, TNFRSF18, LAG3, and ICOS^10,32 (supplemental Figure 1G). The B-cell compartment split into 2 clusters of memory and naive subsets, in addition to plasmablasts, and GC B cells (Figure 1B; supplemental Figure 1C-F).

The inclusion of nonlymphoma lymph nodes, including healthy, nonreactive tissue, allowed us to robustly test the differential abundance of cells between lymphoma and nonlymphoma, using a neighborhood partitioning approach (MiloR, see “Methods”).¹⁸ Lymphoma was enriched for cytotoxic lymphocyte subsets including effector memory CD8⁺ T cells and natural killer (NK) cells, in addition to CD4⁺ T-cell subsets expressing checkpoint apparatus, including ThExh and Tfh cells (Figure 1C; supplemental Figure 2B). Within the lymphoma samples, cHL was enriched for exhausted T-cell subsets, in contrast with the NLPHL/NLPBL sample, which was dominated by cytotoxic subsets (NK cells, effector memory CD8⁺ T cells, and CD8 central memory T cells) (supplemental Figure 2C).

We did not detect HRSCs, likely because of the rarity, size, and fragility of these cells; this was confirmed by assessing for canonical copy number alterations (supplemental Fig 3A-C). To probe their transcriptional programs, which potentially orchestrate this immunoregulatory milieu, we leveraged a microarray data set profiling microdissected HRSCs and GCs (see “Methods”).³³ Differential expression between HRSCs and GCs identified an HRSC gene signature, including TNFRSF8 (CD30) (Figure 1D). Scoring of transcription factor regulons demonstrated nuclear factor κB (NF-κB) activation (NFKB1 and its activatory heterodimer partner RELA) in HRSCs (Figure 1E), consistent with previous reports.³⁴ Next, we performed in silico identification of interactions between active transcription factors and potential targets within the HRSC gene set (see “Methods”). This demonstrated an NF-κB–centric network coordinating the upregulation of the chemokines CCL5, CCL17, and CCL22 capable of the positioning and retention of ThExh via CCR5 and CCR4 ligation (Figure 1F; supplemental Figure 1G).

Within the myeloid compartment, we identified macrophages coexpressing CD14, CD68, and resident macrophage markers FOLR2 and MRC1 (Figure 1G-H). These cells were transcriptionally distinct from classical monocytes, which expressed a characteristic signature including S100A9, CD14, VCAN, and FCN1³⁵ (Figure 1G-H). Among DCs, we identified cDC1 (key transcripts: CLEC9A, CADM1, and IDO1) and cDC2 (key transcripts: CD1C, CLEC10A, and FCER1A) (Figure 1G-H). We also identified LAMP3⁺ DCs in both healthy and some lymphoma samples, expressing the chemokine receptor CCR7 and the chemokines CCL17 and CCL19, which we termed “activated DC” (aDC), consistent with the nomenclature of transcriptionally similar cells described in human thymus and spleen,^36,37 and in murine lung neoplasms³⁸ (Figure 1G-H; supplemental Figure 1E). We found a population of plasmacytoid DCs (pDC) with a dominant contribution from lymphoma samples (Figure 1I; supplemental Figure 2A), expressing IL3RA (CD123) and CLEC4C (Figure 1G-H).

Within the stromal cell compartment, we identified fibroblasts (key transcripts: LUM and DCN), and endothelial cells (key transcripts: CLDN5, PECAM1, and PLVAP [fenestrated endothelium]) (Figure 1B; supplemental Figure 1C).

We used this lymph node–wide account of the immune landscape of cHL as a reference to interrogate the cellular composition of cHL tissues, focusing on myeloid cells.

We first identified spatially distinct microenvironments using targeted spatial transcriptomic profiling (Nanostring GeoMx Cancer Transcriptome Atlas), defining 300-μm-diameter ROIs in both PD-L1^high and PD-L1^low regions of 9 NSCHL and 1 mixed-cell cHL (MCCHL) lymph nodes and follicular and interfollicular regions of 1 control reactive lymph node (supplemental Figure 4A-B). We represented these transcriptional profiles in a shared–nearest neighbor graph and identified 5 clusters (Figure 2A-B). We then deconvoluted the cell composition of each cluster using our scRNA-seq reference, followed by cell-type count estimation and differential cell-type abundance analysis (Figure 2C, see “Methods”). The clusters encompassed 2 “neoplastic” PD-L1^high HRSC-enriched clusters (clusters 3 and 4), 2 “nonneoplastic” PD-L1^low HRSC-depleted clusters (clusters 1 and 5), and 1 intermediate neighborhood (cluster 2).

Clusters 1 and 5 were PD-L1^low and represented fibrotic regions and healthy follicular microenvironments, respectively. Cluster 1 was enriched for stromal (fibroblast and endothelial) cells, classical monocytes, and pDCs but was depleted of cDCs (Figure 2C). Differentially expressed genes in this cluster included genes encoding signaling mediators of fibrosis FGFR4, TGFB2, and PTCH1 (supplemental Figure 4C). In contrast, cluster 5 was exclusively derived from control samples and enriched for GC B cells, Tfh, and plasmablasts (Figure 2A-C).

The PD-L1^high neoplastic clusters exhibited divergent leukocyte enrichment. Cluster 3 was enriched for ThExh, Th, and NK cells. In contrast, cluster 4 represented a myeloid-enriched niche, characterized classical monocyte, macrophage, and cDC2 infiltration (Figure 2C). Marker genes for cluster 4 highlighted inflammatory signaling, with upregulation of chemokines associated with Th2 responses and CCR3-dependent eosinophil recruitment (CCL18, CCL13, CCL24, CCL26, and CCL23) and granulocyte-attracting chemokines CXCL1 and CXCL6 (supplemental Figure 4C).

Next, we sought to identify transcriptional correlates of these microenvironments in publicly available gene expression data. We performed weighted correlation network analysis (WGCNA) using data from 130 diagnostic cHL lymph nodes.^22,39 This analysis yielded 6 modules of coexpressed genes (modules A-F) (Figure 2D). We performed enrichment analysis using both scRNA-seq references and Gene Ontology terms to annotate these modules and found that whereas module C corresponds to the cell cycle program (supplemental Figure 4D), the other modules strikingly mirror the tissue niche patterns identified in our spatially resolved (Nanostring) transcriptomics data (Figure 2D), suggesting a highly conserved organization of pathological niches in cHL.

Module A enriched for stromal cell signatures, similarly to PD-L1^low nanostring cluster 1. In contrast, module B represented myeloid-skewed inflammation, enriching for monocyte, macrophage, and cDC2 signatures, aligning to nanostring cluster 4 (Figure 2D).

We then interrogated the relationship between module gene expression and treatment outcomes. High expression of the fibrosis/stroma enriched module was associated with treatment success, whereas high expression of the myeloid enriched module was associated with early treatment failure. Expression of module F (B-cell follicle) was not associated with differences in outcome (Figure 2D).

Given that variation in microenvironmental myeloid infiltration appears to be a conserved and prognostically important feature of cHL pathobiology, we sought to define MNPs within the HRSC niche at single-cell resolution. Using marker gene coexpression patterns in the scRNA-seq data (supplemental Figure 4E), we designed multiplexed immunofluorescence panels to identify MNPs in cHL tissue sections (n = 54 tumors, representative images: Figure 2E-G). We identified CADM1⁺/CD11c⁺ cDC1, CD1c⁺/CD11c⁺ cDC2, LAMP3⁺ aDCs with distinctive dendritic morphology, CD123⁺ pDC, CD11c⁺ monocytes and macrophages (CD11c⁺ ONLY), and CD30⁺ HRSCs (Figure 2E). We then phenotyped segmented cells and performed neighborhood analyses, taking a 25-μm-radius neighborhood around each CD30⁺ HRSC and measuring the relative enrichment of MNP subsets in aggregated neighborhoods, compared with “nonneighborhood” regions (Figure 2F). This revealed enrichment of cDC2 and CD11c⁺ monocytes in the immediate vicinity of HRSCs. In contrast, both pDC and aDC were excluded from the HRSC niche and occupy regions with a low density of CD11c⁺ cells (Figure 2E-H; Figure 3A-H).

Next, we investigated which signals might coordinate the positioning of MNP subsets and T cells found in close association with HRSCs. To interrogate ligand-receptor interactions (LRIs), we calculated the statistical enrichment of candidate LRIs between MNPs and T- cells in our scRNA-seq data, using the CellPhoneDB tool.¹⁹ This analysis predicted paracrine CCL3 and CCL4 signaling by macrophages and classical monocytes to CCR5- and CCR1-expressing cDC2 and ThExh (Figure 4A). Furthermore, classical monocyte–derived CXCL10 was predicted to signal to cDC1, ThExh, and Tfh via CXCR3. Nominated inhibitory interactions from ThExh included TIGIT signaling via NECTIN2 expressed by cDC2, classical monocytes, and macrophages (Figure 4A). Analysis of putative LRIs inferred from Nanostring cluster 4 and WGCNA module B also highlighted CCR1- and CCL5-mediated signaling (supplemental Figure 5A). A proposed schematic illustrates potential interactions between HRSCs, cDC2s, classical monocytes, and TExh (Figure 4B); NF-κB–driven expression of CCL22, CCL17, and CCL5 by HRSCs coordinates recruitment of TExh and cDCs; the niche is further coordinated by CCL3, CCL4, and CXCL10 signaling from monocytes and macrophages.

We were surprised that the HRSC-associated MNP network includes cDC2 and occasional cDC1, which play key roles in immunosurveillance. Although macrophage expression of PD-L1 is established, upregulation of inhibitory ligands and receptors by DCs is less well described in cHL. To test this, we designed a second IF panel to examine expression of PD-L1, IDO1, and TIM-3 by CD11c⁺/CD68⁻ cDC1 and cDC2, CD11c⁺/CD68⁺ monocytes, and CD11c⁻/CD68⁺ macrophages, and to assess their spatial relationships to CD30⁺ HRSC (Table 3).

HRSCs exhibited extensive PD-L1 expression, as expected. We also found enrichment of PD-L1⁺ CD11c⁺/CD68⁺ and CD11c⁺/CD68⁻ MNPs in CD30⁺ HRSC neighborhoods (Figure 5A-D). Similarly, we found variable coexpression of the coinhibitory receptor TIM-3, and IDO1, an immunomodulating tryptophan catabolizing enzyme, on CD11c⁺/CD68⁺ and CD11c⁺/CD68⁻ MNPs in close proximity to HRSCs (Figure 5A-E; supplemental Figure 6). MNPs expressing ≥1 inhibitory molecules are enriched around HRSCs and associated with other such MNPs; CD68⁺ CD11c⁻ MNPs are distributed evenly across the tissue, with no enrichment around HRSCs (Figure 5D-E).

The proportion of CD11c⁺/CD68⁻ and CD11c⁺/CD68⁺ MNPs expressing inhibitory molecules increased with age at diagnosis (P < .05), but there were no significant differences in the proportion of MNPs expressing inhibitory molecules based on sex, histological subtype, or Epstein-Barr virus status (supplemental Figure 7A-D). This finding suggests that immunosuppressive signaling in pediatric and young-adult cHL may differ from cHL in older adults.

HRSCs have the remarkable ability to cohere an immunosuppressive microenvironment, including the recruitment of MNPs expressing immunoregulatory molecules. Through high-resolution analysis, we found that MNPs comprise diverse subsets of monocytes, macrophages, and DCs. We highlight several new features of this MNP network: the exclusion of pDCs and aDCs from the vicinity of HRSC; that immunoregulatory molecules are present on DCs close to HRSCs, at equivalent levels to macrophages; that multiple soluble and cell-to-cell interactions underpin this structure, communicating with HRSCs and lymphocytes. This work complements recent detailed analysis of lymphocytes in cHL, showing enrichment of exhausted LAG3⁺ CD4⁺ T cells around MHC-II⁻ HRSCs,¹⁰ and elaboration of CXCL13 by Tfh-like cells in lymphocyte-rich cHL.³² 

We have shown that conventional DCs are closely associated with PD-L1⁺ HRSCs. These cells are usually involved in immunosurveillance and presentation of neoantigens to T cells, however, this mechanism appears to be stalled in cHL, likely as a result of immunosuppressive niche signaling. Colocalization and LRI analyses suggest that MNPs have proximal roles in directing and amplifying this signaling. These MNPs cooperate with HRSC-derived and NF-κB–directed chemokine expression, providing chemoattractant and inhibitory signals that recruit and instruct immunosuppressive T cells. Our analysis identifies classical monocytes, in particular, as important signaling hubs, controlling retention of cDC2 and ThExh via CCR1-, CCR4-, CCR5-, and CXCR3-dependent signaling (Figure 4A-B).

We propose that there are distinct and recurring tissue niches within cHL lymph nodes. Enrichment for the inflammatory cDC2-monocyte-macrophage niche is associated with early relapse following treatment, expanding upon previous reports that CD68⁺ TAM infiltration and macrophage gene expression is associated with poor outcomes in advanced cHL.^39,40 Conversely, a fibrosis/stroma-rich tissue niche is associated with favorable outcomes. Computational pathology approaches and high-throughput multiplexed technologies may permit the characterization of these diverging cell networks in diagnostic tissue sections, before treatment.^41,42 Future strategies might aim to shift the dominant pathology toward a fibrotic or MNP-depleted phenotype (eg, CCL2 inhibition or CSF-CSF1R blockade), and away from an inflammatory and immunosuppressive phenotype.^43-45 DCs may be augmented by Flt3-ligand,⁴⁶ CD40 agonists, or lentivirally-transduced chimeric antigen receptors.⁴⁷ 

Therapeutic efforts to reverse tumor-associated immune suppression have targeted T cells, with recent studies suggesting alternative immune checkpoints, in addition to PD-1.^10,48 The coexpression of inhibitory molecules and chemokines across MNPs suggests extensive redundancy in immunosuppressive signaling.^49,50 This may partially explain resistance to PD-1 blockade, and offer opportunities for tailoring therapies on the basis of molecular profiling. In a murine model, it has been shown that TAMs rapidly transfer anti–PD-1 complexes away from T cells, through a FcɣR-mediated process.⁵¹ Intriguingly, a recent report demonstrated that HRSC eradication, after induction therapy with Nivolumab, was associated with a reduction in PD-L1⁺ TAM, rather than detectable cytotoxic responses.⁵² This suggests initial vulnerability of MNP networks to PD-1 blockade, such as dependence on tonic signaling by T cells or HRSCs, and may support the hypothesis of reverse signaling on HRSCs.⁵³ 

Overall, these atlas-scale molecular and spatially resolved data provide evidence for distinctly organized tissue niches defining the pathology of cHL. Our results highlight a dominant, interacting MNP network associated with HRSCs with immunosuppressive phenotypes, the rational targeting of which may present valuable therapeutic opportunities.

The authors thank Christian Steidl (The University of British Columbia) for generous sharing of scRNA-seq data. The authors thank the Newcastle University/Newcastle upon Tyne NHS Foundation Trust Molecular diagnostics laboratory (NovoPath; https://www.novopath.co.uk/) for tissue handling and technical assistance.

B.J.S. is supported by a Clinical Doctoral Fellowship from the Wellcome Trust (216366/Z/19/Z). C.D.C. is supported by grant funding from the JGW Patterson Foundation (https://jgwpattersonfoundation.co.uk), Bright Red (https://brightred.org.uk), the Lymphoma Research Trust (https://www.lymphoma-research-trust.org.uk), a Medical Research Council Confidence in Concepts award, and a Wellcome Trust Small project grant award. S.B. is supported by the Wellcome Trust (206194).

Contribution: B.J.S. performed data analysis and wrote the manuscript; M.F. performed data analysis and contributed to the manuscript; M.D.Y. contributed to data curation; C.J., A.S., A.B., C.M.B., V.R., J.R.F., K.R.J., and N.C. performed scRNA-seq experiments; K.T.M., L.H., N.J., and K.S.-P. acquired deceased-donor lymph node tissue; M.C., M.R.C., and S.B. supervised the project; and C.D.C. performed microscopy experiments, analyzed data, wrote the manuscript, and supervised the project.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Sam Behjati, Wellcome Sanger Institute, Wellcome Genome Campus, Hinxton, Saffron Walden CB10 1RQ, United Kingdom; e-mail: sb31@sanger.ac.uk; and Christopher D. Carey, Translational and Clinical Research Institute, Newcastle University, Newcastle upon Tyne, NE1 7RU, United Kingdom; e-mail: christopher.carey@newcastle.ac.uk.