Spatial transcriptomics identifies a suppressive, T-cell–excluded tumor microenvironment in extramedullary myeloma∗ Authors: Nicholas E. Bingham, Julie R. Boiko, Daniel C. Jones, Daniel Wong, Tiffany Khong, Sridurga Mithraprabhu, Kathleen S. Ensbey, Anna E. Elz, Evan W. Newell, Andrew Spencer, Geoffrey R. Hill Published: 2026 ABSTRACT Key Points•The tumor immune microenvironment in extramedullary myeloma is dominated by macrophages with a suppressive phenotype.•Bidirectional cellular interactions predict activation of canonical PC survival pathways and drive immunosuppressive macrophage phenotype. Introduction Multiple myeloma (MM) is a malignancy of terminally differentiated plasma cells (PC), with nonmalignant PC able to survive for many years because of a highly specialized and supportive niche within the bone marrow (BM) microenvironment.1 This niche is characterized by cellular interactions through direct contact and through secondary factors including the extracellular matrix (ECM), cytokines, and membrane-associated signaling proteins.2,3 In MM restricted to the BM, this specialized niche was thought to be highly advantageous to malignant PCs, providing signals that promote unrestrained growth and proliferation,4 although the existence of a universal MM niche has recently been questioned.5 Extramedullary disease (EMD), where MM cells survive outside the BM, reflects a significant change in MM pathophysiology. This disease phase portends a dismal prognosis,6 reflecting aggressive tumor kinetics and significant drug resistance. There are 2 types of EMD: extraosseous and true hematogenous EMD.7 Extraosseous EMD is more common, representing ∼83% to 88% of EMD tumors and arises because of direct extension from bony lesions into the surrounding soft tissues.8 EMD that arises by hematogenous spread, with distant sites of disease in soft tissues and organs, likely represents a different mechanism of dissemination and a different underlying biology.9,10 These differences are evidenced by patient outcomes because hematogenous EMD is associated with the worst outcomes, whereas the historically high-risk nature of extraosseous EMD is being overcome, at least in part, with the advent of novel therapies.8,11, 12, 13 These differences in clinical outcomes and mechanism of dissemination suggest a unique biology of hematogenous EMD tumors. Importantly, this also provides a rationale for the development of novel therapies that target aspects of this unique biology to improve patient outcomes. Although there is increasing literature around the pathobiology of EMD,14, 15, 16 as a relatively uncommon phenomenon, there is still a paucity of data surrounding the underlying biology. This is particularly true in regard to the unique ability of EMD MM cells to survive outside the BM microenvironment. Relative to BM-restricted MM, the limited improvement in outcomes with novel immunotherapies, including chimeric antigen receptor T cells and bispecific antibodies, suggests an important role of immune escape in these tumors.17, 18, 19, 20 The central role of the BM niche in the pathophysiology of BM-restricted MM suggests that the tumor microenvironment (TME) in EMD is likely a key feature in supporting tumor cells and promoting immune evasion. In this study, we aimed to characterize the TME of hematogenous EMD at high resolution through the use of subcellular spatial transcriptomics. We aimed to describe the immune microenvironment specifically and the nonimmune TME cells to define interactions that provide prosurvival signals to malignant PC and aid immune evasion in these tumors to facilitate the identification of novel therapeutic targets. Methods Study design and ethics Tumor samples were collected under protocol Project 343/24 (Alfred Health Ethics Committee). Written informed consent was obtained from all patients in accordance with the Declaration of Helsinki. Only hematogenous EMD was included. Sample collection and processing Formalin-fixed, paraffin-embedded biopsy specimens of EMD were selected. These samples have previously been investigated with whole-genome sequencing. Sections of 5-μm thickness were loaded onto a Xenium slide (4 samples per slide) before subsequent processing as per the manufacturer’s instructions. Xenium in situ The commercially available 5000 gene panel (Xenium Human 5K Prime) was used. Xenium slides were read on a Xenium Analyzer. Collected images underwent initial quality control on Xenium Explorer. Segmentation with Proseg followed, a probabilistic method of cell segmentation that improves the allocation of detected transcripts21 before reassignment of segmentation by the Xenium Ranger software and downstream bioinformatic processing. Post-Xenium histology Hematoxylin and eosin staining of the Xenium slides was performed upon completion. Bioinformatics summary Computational environments The computational environment included the following modules: R version 4.4.1; Seurat version 5.3.0, SeuratObject version 5.1.0; tidyr version 1.3.1; ggplot2 version 3.5.2; Matrix version 1.7.3; dplyr version 1.1.4; ktplots version 2.4.0; pheatmap version 1.0.12; and circlize version 0.4.16. CellPhone analysis used: Python version 3.8.20: cellphonedb version 5.0.1. Quality control and filtering Low-quality or artifactual segmented objects with <70 total transcripts were excluded based on a knee plot analysis. Cell type annotation, niche, and DGE analysis Gene expression matrix was exported into a Seurat object. Normalization (SCTransform) was performed before dimensionality reduction with principal component analysis and shared nearest neighbor graphing. Clustering used FindClusters (Louvain/Leiden algorithm). Visualization used uniform manifold approximation and projection. Manual cell type annotation used canonical markers; cell type assignments were validated by spatial localization, marker coexpression, and program scoring using curated lineage- and state-specific gene sets.16,22, 23, 24, 25 Differential gene expression (DGE) used Seurat’s FindAllMarkers (Wilcoxon rank-sum; multiple testing correction with the Benjamini-Hochberg false discovery rate). Cell cycle analysis used CellCycleScoring. AddModuleScore assessed functional programs using defined gene sets (supplemental Methods). Results Xenium 5K Prime output quality control Eight patients were included. Patient demographic and disease data are shown in Table 1. The data acquired by Xenium were high-quality, with a median number of transcripts per sample of 24.4 × 106 (range, 15.3 × 106 to 76.2 × 106), and 83.5% of the transcripts were of high quality (Table 2). Cell segmentation using the Proseg algorithm improved the number of transcripts within cells and the median number of transcripts per cell (supplemental Table 1). The median number of cells detected was 61 000 per sample. The median number of transcripts per cell was 412, with a median of 307 genes detected per cell.Table 1.Patient demographic and disease dataPatientDemographicsMM diagnosis detailsEMD detailsAge, ySexMM subtypeFounder lesionSecondary lesionsSubtypeAnatomical siteEM0552FIgA λUnknown—SecondaryPancreas, thyroid, LN, and bladderEM0644FNSHypodiploiddel(17p)SecondaryPleura and muscleEM2058MIgG κNormal karyotype—SecondaryPelvic mass and LNEM2132Mκ LCHyperdiploidt(8;14)PrimaryAbdominal massEM2348MIgG κHypodiploiddel(1p)PrimaryMuscle, pleura, and LNEM2653MIgG κNormal karyotype—SecondaryLN, kidney, and muscleEM2754Mλ LCNon-t(4;14)/t(14;16)—SecondaryMuscle and CNSEM3062MIgG κHyperdiploidgain(1q) and del(1p)SecondarySubcutaneous and muscleBold text indicates the site of biopsy.CNS, central nervous system; F, female; IgA, immunoglobulin A; LC, light chain; LN, lymph node; M, male; NS, non-secretory.Table 2.Summary of Xenium data outputPatientTotal transcripts% transcripts within cellTotal cellsCells per 100 μm2Median transcripts per cellMedian genes per cellEM0517 902 89996.246 0400.372247196EM0676 163 28496.6187 1210.9207282EM2065 263 92699.6114 5320.351443328EM2119 199 87995.785 0640.42910∗8∗EM2315 350 21998.616 3970.168754497EM2627 277 77695.932 5040.319534360EM2721 653 23499.650 0590.378381287EM3070 764 62999.973 4380.398830555∗Areas of necrosis excluded from further analysis. Table 1.Patient demographic and disease dataPatientDemographicsMM diagnosis detailsEMD detailsAge, ySexMM subtypeFounder lesionSecondary lesionsSubtypeAnatomical siteEM0552FIgA λUnknown—SecondaryPancreas, thyroid, LN, and bladderEM0644FNSHypodiploiddel(17p)SecondaryPleura and muscleEM2058MIgG κNormal karyotype—SecondaryPelvic mass and LNEM2132Mκ LCHyperdiploidt(8;14)PrimaryAbdominal massEM2348MIgG κHypodiploiddel(1p)PrimaryMuscle, pleura, and LNEM2653MIgG κNormal karyotype—SecondaryLN, kidney, and muscleEM2754Mλ LCNon-t(4;14)/t(14;16)—SecondaryMuscle and CNSEM3062MIgG κHyperdiploidgain(1q) and del(1p)SecondarySubcutaneous and muscleBold text indicates the site of biopsy.CNS, central nervous system; F, female; IgA, immunoglobulin A; LC, light chain; LN, lymph node; M, male; NS, non-secretory. Patient demographic and disease data Bold text indicates the site of biopsy. CNS, central nervous system; F, female; IgA, immunoglobulin A; LC, light chain; LN, lymph node; M, male; NS, non-secretory. Table 2.Summary of Xenium data outputPatientTotal transcripts% transcripts within cellTotal cellsCells per 100 μm2Median transcripts per cellMedian genes per cellEM0517 902 89996.246 0400.372247196EM0676 163 28496.6187 1210.9207282EM2065 263 92699.6114 5320.351443328EM2119 199 87995.785 0640.42910∗8∗EM2315 350 21998.616 3970.168754497EM2627 277 77695.932 5040.319534360EM2721 653 23499.650 0590.378381287EM3070 764 62999.973 4380.398830555∗Areas of necrosis excluded from further analysis. Summary of Xenium data output Areas of necrosis excluded from further analysis. Interpatient heterogeneity is the main driver of EMD MM PC clustering All cells from all samples were combined (500 102 cells in total). Unsupervised clustering was performed (Figure 1A), identifying 21 cell clusters. Most clusters arose from individual patients (Figure 1B-C). Some clusters (eg, clusters 8, 9, 13, and 17) contained cells from all patients. Initial identification using known PC markers IRF4 and XBP1 (Figure 1D) identified that the single/low-patient clusters represent tumor cells (17 clusters). The mixed-sample clusters represent cells from the TME, with CD68 and CD8A expression demonstrating macrophages and T cells, respectively (Figure 1E).Figure 1.Analysis of all samples combined (total cells = 500 102). (A) Uniform manifold approximation and projection (UMAP) demonstrating unsupervised clustering of all cells; 21 clusters were identified. (B) The overlaying sample of origin demonstrates clusters consisting of cells from a single sample, whereas other clusters contain cells from all patients. (C) Histogram demonstrating the composition of clusters by sample of origin. (D) Differences in gene expression are biologically relevant; patients with translocations involving MAF and MYC demonstrate higher average expression of MAF and MYC. (E) The clusters consisted of cells from largely 1 patient expressing PC markers, including IRF4 and XBP1. (F) Clusters containing cells from multiple patients comprise cells from the TME, here expressing CD68 (macrophages) and CD8A (T cells). For panels E-F, the color scale reflects normalized transcript counts. ID, identifier. Analysis of all samples combined (total cells = 500 102). (A) Uniform manifold approximation and projection (UMAP) demonstrating unsupervised clustering of all cells; 21 clusters were identified. (B) The overlaying sample of origin demonstrates clusters consisting of cells from a single sample, whereas other clusters contain cells from all patients. (C) Histogram demonstrating the composition of clusters by sample of origin. (D) Differences in gene expression are biologically relevant; patients with translocations involving MAF and MYC demonstrate higher average expression of MAF and MYC. (E) The clusters consisted of cells from largely 1 patient expressing PC markers, including IRF4 and XBP1. (F) Clusters containing cells from multiple patients comprise cells from the TME, here expressing CD68 (macrophages) and CD8A (T cells). For panels E-F, the color scale reflects normalized transcript counts. ID, identifier. Gene expression data appeared biologically relevant. Compared with a patient with a known t(14;16) translocation, there was higher MAF expression in PC than in other patients and higher MYC expression in PC from a patient with a known t(8;14) (Figure 1F). This confirms that differences in gene expression between samples reflect underlying biology and that the current technology was able to identify these differences. Suppressive macrophages dominate the TME in EMD We then sought to define the TME. The non-PC clusters were combined before dimensionality reduction and unsupervised clustering. Seven clusters were identified after excluding PC contamination (Figure 2A). The 5 TME clusters were comprised of a fibroblast-dominant cluster, 3 separate macrophage-dominant clusters, a T-cell–dominant cluster, and myocytes; the subsets of TME cells and associated lineage-defining genes are shown in Figure 2B. Each cluster included cells from each sample, aside from cluster 5 (myocytes) observed only in the 2 biopsies obtained from tumors within the skeletal muscle (Figure 2C).Figure 2.The TME in EMD. (A) UMAP and unsupervised clustering of TME cells demonstrating 6 major cell types. (B) Dot plot demonstrating expression of key lineage-defining genes in different TME cell types. (C) Cells of the TME are shared among patients and found in all samples, except myocytes found only in patients with biopsies from skeletal muscle. (D) Cellular composition of the TME in EMD samples, demonstrating that suppressive macrophages are the most numerous cells in the TME. (E) Stacked histogram showing the subtype of myeloid cells found within EMD tumors. Macrophages with a suppressive phenotype are the most numerous myeloid cells in all tumors. (F) Stacked histogram of the T-cell compartment showing heterogeneous T-cell infiltration and the composition of the T-cell compartment in each sample. (G) Radar plot showing the distribution of fibroblast functional phenotypes among fibroblast subtypes. Avg, average; DC, dendritic cell; gd, γ-δ; ID, identifier; IFN, interferon; NK, natural killer; TAM, tumor-associated macrophage; Treg, regulatory T-cells. The TME in EMD. (A) UMAP and unsupervised clustering of TME cells demonstrating 6 major cell types. (B) Dot plot demonstrating expression of key lineage-defining genes in different TME cell types. (C) Cells of the TME are shared among patients and found in all samples, except myocytes found only in patients with biopsies from skeletal muscle. (D) Cellular composition of the TME in EMD samples, demonstrating that suppressive macrophages are the most numerous cells in the TME. (E) Stacked histogram showing the subtype of myeloid cells found within EMD tumors. Macrophages with a suppressive phenotype are the most numerous myeloid cells in all tumors. (F) Stacked histogram of the T-cell compartment showing heterogeneous T-cell infiltration and the composition of the T-cell compartment in each sample. (G) Radar plot showing the distribution of fibroblast functional phenotypes among fibroblast subtypes. Avg, average; DC, dendritic cell; gd, γ-δ; ID, identifier; IFN, interferon; NK, natural killer; TAM, tumor-associated macrophage; Treg, regulatory T-cells. Macrophages were the most numerous non-PC cell type identified in the extramedullary TME, on average comprising 37% of nontumor cells (Figure 2D). The macrophage clusters also contain dendritic cells, as evidenced by low-level expression of markers such as CIITA and IRF4 (Figure 2B). The macrophage-dominant groups were combined, and cells were scored across 9 lineage- and state-specific macrophage gene expression programs, including proinflammatory and suppressive programs, previously conceptualized as canonical M1 and M2 polarization, respectively. Macrophage subsets are shown in Figure 2E. Interestingly, the canonical inflammatory phenotype was infrequent, and subclusters with inflammatory expression profiles (IL1B, NLRP3, CD14, TNF, TLR2, and CCR2) had the highest interferon-responsiveness scores (CXCL10, CXCL11, IFIT2, IFIT3, IRF7, and STAT1) and simultaneously expressed suppressive genes (MRC1, CD163, MSR1, MARCO, and IL10; supplemental Figure 2). Overall, macrophage cells with a suppressive phenotype dominate the TME in EMD. We then analyzed the T-cell–dominant clusters. Overall, 23.7% of the nontumor cells were T cells. These were further assessed using gene expression scores to identify T-cell subsets. The composition of T cells found within EMD tumors varied between patients (Figure 2F). The T-cell subtypes were scored for the expression of activation and exhaustion markers. γ-δ T cells and natural killer cells had higher activation scores than cytotoxic CD8 T cells (supplemental Figure 3). The coexpression of both activation and exhaustion markers was frequent. Twenty percent of nontumor cells were found in the fibroblast-dominant cluster, suggesting a potential role of fibroblasts in EMD pathobiology. Five fibroblast subclusters were scored using fibroblast-specific, functionally relevant gene expression signatures involving ECM production, ECM fibrosis/remodeling, immune interaction, and perivascular and angiogenesis signatures.26 Radar plots (Figure 2G) demonstrate the dominant phenotypes of each fibroblast subcluster, showing different predicted functions of the subclusters in the TME. Three recurrent immune microenvironments define extramedullary tumors Applying these annotations to individual samples, the spatial distribution of TME cell types was further assessed using unsupervised predictions of computational spatial niches, leveraging nearest neighbor analysis and grouping cells based on spatial proximity and phenotypic similarity (Figure 3A-B). The number of niches was the minimum number of niches that produced compositionally distinct and nonredundant niche profiles, defined by cellular composition. Three recurrent niches were identified (Figure 3C). The cellular compositions of these niches, corrected for PC percentage, invoke the spatial immune microenvironments proposed in BM-restricted MM,27 namely immune-suppressed (IS), immune-excluded (IE), and immune-permissive (IP) environments (Figure 3D). The IE niche was observed in all tumors, made up the bulk of all EMD tumors, and was characterized by a high proportion of tumor cells with minimal immune cell infiltrate (Figure 3C,E). The IS niche was characterized by the presence of suppressive macrophages and fibroblasts and was the second most frequent niche in EMD tumors. The IP niche, characterized by the highest nontumor cell composition, particularly interferon-responsive macrophages and T-cell infiltrates, was observed in only 2 samples.Figure 3.Spatial analysis of EMD tumors. (A) Distribution of cells of the TME in EMD tumors. (B) Spatial distribution of predicted niches in the same samples. (C) Composition of each sample by predicted niche, demonstrating that most of each tumor comprises the IE niche. (D) Composition of the TME of each niche, corrected for PC, highlighting differences in cellular composition. The IS niche contains the most endothelial cells, pericytes, and fibroblasts, whereas the IP niche contains the most inflammatory macrophages and CD8+ T cells. (E) Representative example H&E section of tumor with corresponding niches and TME cell distribution. (F) χ2 analysis of the distribution of cell types between niches. (G) Expression of T-cell exhaustion markers across samples. (H) Ratios of CLTA4, LAG3, and TIGIT to PDCD1 (PD-1) highlight relatively higher LAG3 and TIGIT expression, suggesting T-cell exhaustion because of TME suppression rather than PD-1–driven chronic antigen stimulation. DC, dendritic cell; gd, γ-δ; H&E, hematoxylin and eosin; IFN, interferon; IQR, inter-quartile range; NK, natural killer; TAM, tumor-associated macrophage; Treg, regulatory T-cells. Spatial analysis of EMD tumors. (A) Distribution of cells of the TME in EMD tumors. (B) Spatial distribution of predicted niches in the same samples. (C) Composition of each sample by predicted niche, demonstrating that most of each tumor comprises the IE niche. (D) Composition of the TME of each niche, corrected for PC, highlighting differences in cellular composition. The IS niche contains the most endothelial cells, pericytes, and fibroblasts, whereas the IP niche contains the most inflammatory macrophages and CD8+ T cells. (E) Representative example H&E section of tumor with corresponding niches and TME cell distribution. (F) χ2 analysis of the distribution of cell types between niches. (G) Expression of T-cell exhaustion markers across samples. (H) Ratios of CLTA4, LAG3, and TIGIT to PDCD1 (PD-1) highlight relatively higher LAG3 and TIGIT expression, suggesting T-cell exhaustion because of TME suppression rather than PD-1–driven chronic antigen stimulation. DC, dendritic cell; gd, γ-δ; H&E, hematoxylin and eosin; IFN, interferon; IQR, inter-quartile range; NK, natural killer; TAM, tumor-associated macrophage; Treg, regulatory T-cells. The immune cell composition of these niches varied. Using χ2 standardized residuals, cell types overrepresented in each niche were identified (Figure 3F). This identified suppressive macrophages and T cells, as well as fibroblasts and endothelial cells as overrepresented in IS environments, whereas IP environments contained interferon-responsive/inflammatory macrophages and T-cell subsets. We further examined T-cell exhaustion marker expression. Although there was a trend toward higher expression of exhaustion markers on T cells within the IP niche, this did not reach statistical significance when corrected for between-sample variability (per-sample Kruskal-Wallis P = .112). Across samples, TIGIT and LAG3 were more prevalent than PDCD1 (median TIGIT:PDCD1 and LAG3:PDCD1 ratios of 1.78 and 1.62, respectively), whereas CTLA4 was comparable to PDCD1 (Figure 3G-H). The inhibitory receptor burden per T cell was relatively low (median, 1.22 per cell). This pattern of gene expression is suggestive of a suppressive checkpoint landscape rather than a PD-1–dominant, chronic antigen-driven exhaustion. Given the association between macrophage phenotype and T-cell function, we then reviewed the spatial relationship between the two (supplemental Figure 3). Activation scores were moderately sample dependent (intraclass correlation = 0.26), whereas exhaustion scores showed minimal clustering by sample (intraclass correlation = 0.068). Proximity to inflammatory vs suppressive macrophages (using within-sample pairing) did not produce a detectable within-patient difference in T-cell exhaustion or activation scores. PC in EMD maintain a PC transcriptome We then examined the PC component of EMD tumors. Tumor cells make up the bulk of each tumor, averaging 82.5% of the cells within each sample (Figure 4A). As noted above, the PC clusters were largely separated into patient-specific groupings (Figure 1B-C). This interpatient variability was driven by canonical MM drivers, with genes such as IGHE, MYC, MYCN, CD44, IKZF2, and TENT5C among the top differentially expressed genes between samples (Figure 4B). There were 2 clusters that contained cells from more than 1 patient, with one demonstrating high expression of CCND2 and the other of TENT5C and CD44.Figure 4.PC gene expression profile in EMD. (A) Histogram demonstrating sample composition in terms of TME vs tumor cells. (B) Dot plot demonstrating average expression of key PC markers in PC from each patient, with notable differences in the expression of key clinically relevant genes including CD38 and TNFRSF17 (BCMA). (C) Lollipop plot of the top 20 highest-expressed genes in EMD PC. (D) PC in EMD maintain a PC transcriptome without evidence of dedifferentiation, with persistent expression of MM features XBP1, TNFRSF17, IRF4, and PRDM1. (E) Expression of HIF1A in 2 samples. (F) Histogram of the distribution of cell cycle phase in individual samples, showing marked variation in phase among patients. (G) Dot plot demonstrating that cell cycle phase is affected by niche, with cells in the IE niche overrepresented in the G2M/S phase. Avg, average. PC gene expression profile in EMD. (A) Histogram demonstrating sample composition in terms of TME vs tumor cells. (B) Dot plot demonstrating average expression of key PC markers in PC from each patient, with notable differences in the expression of key clinically relevant genes including CD38 and TNFRSF17 (BCMA). (C) Lollipop plot of the top 20 highest-expressed genes in EMD PC. (D) PC in EMD maintain a PC transcriptome without evidence of dedifferentiation, with persistent expression of MM features XBP1, TNFRSF17, IRF4, and PRDM1. (E) Expression of HIF1A in 2 samples. (F) Histogram of the distribution of cell cycle phase in individual samples, showing marked variation in phase among patients. (G) Dot plot demonstrating that cell cycle phase is affected by niche, with cells in the IE niche overrepresented in the G2M/S phase. Avg, average. The highest expressed genes in the PC compartment are shown in Figure 4C. They are, by and large, genes associated with classic MM pathways, namely unfolded protein response and endoplasmic reticulum stress, RNA processing, cellular metabolism, and cellular adhesion. Despite biological distinctions from BM-restricted MM and the associated ability to survive outside the BM microenvironment, EMD PC maintain a classic PC transcriptome defined by the persistent expression of key MM transcription factors XBP1, IRF4, and PRDM1 (Figure 5D). There was no suggestion of dedifferentiation, evidenced by the ongoing suppression of B-cell transcription factors PAX5, BACH2, and FOXP1.Figure 5.Cell-cell interaction modeling. (A) Heat map plot demonstrating the number of predicted interactions between cell types. Macrophages, endothelial cells, and fibroblasts have the most predicted cell-cell interactions, although they co-occur most frequently. (B) Circos plot of key predicted interactions from PC to cells of the TME, highlighting interactions that affect the ECM and reprogramming of myeloid cells. (C) Circos plot of key predicted interactions from the TME to PC, highlighting the numerous prosurvival signals observed. (D) Dot plot of the key interactions affecting T cells in EMD, demonstrating numerous suppressive signals from multiple cell types within the TME. (E) Dot plot of the key interactions from fibroblasts to other cells of the TME. Fibroblasts produce survival signals for malignant PC and suppressive signals for macrophages. (F) Diagrammatic summary of the TME of EMD. EMT, epithelial-mesenchymal transition; gd, γ-δ; IFN, interferon; min %, minimum percentage; NK, natural killer; TAM, tumor-associated macrophage; Treg, regulatory T-cells. Figure created with biorender.com. Bingham, N (2026). https://BioRender.com/lno7zum. Cell-cell interaction modeling. (A) Heat map plot demonstrating the number of predicted interactions between cell types. Macrophages, endothelial cells, and fibroblasts have the most predicted cell-cell interactions, although they co-occur most frequently. (B) Circos plot of key predicted interactions from PC to cells of the TME, highlighting interactions that affect the ECM and reprogramming of myeloid cells. (C) Circos plot of key predicted interactions from the TME to PC, highlighting the numerous prosurvival signals observed. (D) Dot plot of the key interactions affecting T cells in EMD, demonstrating numerous suppressive signals from multiple cell types within the TME. (E) Dot plot of the key interactions from fibroblasts to other cells of the TME. Fibroblasts produce survival signals for malignant PC and suppressive signals for macrophages. (F) Diagrammatic summary of the TME of EMD. EMT, epithelial-mesenchymal transition; gd, γ-δ; IFN, interferon; min %, minimum percentage; NK, natural killer; TAM, tumor-associated macrophage; Treg, regulatory T-cells. Figure created with biorender.com. Bingham, N (2026). https://BioRender.com/lno7zum. Accordingly, canonical MM cell markers remain expressed, including therapeutic targets such as SLAMF7, TNFRSF17 (BCMA) and, to a lesser extent, CD38. CD44, an adhesion molecule previously associated with EMD development, is expressed in our cohort. Additional cellular adhesion genes were also expressed. This includes ITGA4, associated with VCAM1 and fibronectin adhesion, as well as aiding CX3CR1 signaling. Most genes in the highest-expressed list are related to the unfolded protein response, the evasion of which is central to MM pathogenesis. Combined with the high expression of RNA processing genes such as EEF1G, SRSF1, and SND1, the PC of EMD maintain a protein production phenotype. Cellular metabolism genes were also present in the top 50 expressed genes in EMD PC. LDHA was highly expressed across all tumors, suggesting a degree of glycolysis. Taken with the widespread expression of HIF1A (Figure 4E), this is consistent with a hypoxic TME. Cell cycle analysis was also performed (Figure 4F). There was substantial variation in the fraction of proliferating tumor cells between samples (16%-80%); some tumors are highly proliferative. After adjusting for sample, cells in the IE niche were more likely to be in the S/G2M phase compared with those in the IP or IS niches (odds ratio, 0.80 and 0.87, respectively; both P ≤ .01), although the absolute difference (48% vs 42% or 44%) is small (Figure 4G). Interestingly, PC subsets were not evenly distributed between niches, suggesting PC within each niche differ and therefore either drive and/or respond to the niche microenvironment. We performed subclustering of PC separately within each sample. Aside from proliferation, noted above, no recurrent transcriptional programs were identified across samples. In some samples, the PC subclusters were spatially distinct, but in others, there was no spatial discrimination between PC subclusters. This implies that interpatient variability is the greatest driver in PC subsets and that there is no recurrent EMD-related transcriptional program. Bidirectional cross talk between tumor cells and the microenvironment suppresses T-cell function The differing cellular compositions of the spatial niches and the nonrandom distribution of cell types are consistent with cell-cell interactions within the EMD TME affecting tumor biology. Using the previously identified cell clusters and cellular composition of the niches, in conjunction with DGE data, CellPhoneDB28 was used to identify predicted cell-cell interactions. Only cell types that co-occurred in a sample and niche were assessed for cell-cell interactions, with each cell type requiring a minimum of 20 cells and to be >0.5% of cells within the niche for inclusion. A heat map of the relative number of interactions demonstrates that suppressive macrophages, fibroblasts, and endothelial cells had the most interactions with other cell types, although they occur in more combinations (Figure 5A). We began by analyzing the signals from malignant PC to the cells of the TME (Figure 5B). These were broadly characterized into interactions affecting myeloid cell programming, fibroblast/ECM remodeling, angiogenesis and vasculature, and immune suppression/exclusion. The myeloid reprogramming signaling centered on GAS6 expression by PC, predicted to interact with AXL and MERTK from suppressive macrophages; this interaction is known to promote the “M2” polarization of macrophages and the suppressive phenotype identified above. Other suppressive signals included SEMA4D interacting with PLXNB2 and CD72. In addition, CD47 interacts with SIRPA/SIRPG to inhibit phagocytosis. PC were also predicted to affect the ECM, with PC signaling through THBS1, THBS2, and THY1 predicted to lead to fibroblast activation and promote remodeling through CD36 and integrins expressed by fibroblasts. Furthermore, BMP6 and/or BMP8A signaling was observed in all samples, interacting with BMPR/ACVR in fibroblasts and macrophages to promote fibrosis and fibroblast activation. PC were also predicted to affect angiogenesis through the expression of VEGFA/VEGFB, interacting with NRP1/NRP2 to promote angiogenesis. Importantly, PC also generated direct-acting suppressive cues through prostaglandin E2 and TGFB3 production to directly suppress T-cell function and promote fibrosis to further exclude T-cell infiltration. We then reviewed the signaling from the TME to malignant PC (Figure 5C). These interactions were broadly categorized into prosurvival signals and adhesion/retention signals. There were fewer interactions involving the epithelial-mesenchymal transition, apoptosis, and metabolic/hormonal pathways. The prosurvival signals were predicted to activate canonical PC survival pathways, including TNFSF13 and TNFSF13B via TNFRSF17 (APRIL/BAFF to BCMA), IGF2 to IGF1R, and PDGFA/PDGFB/PDGFC to PDGFRA/PDGRFB, as well as NCAM1 with FGFR1. The adhesion and retention signals of note include PECAM1 and CXCL12 to CD38 and CXCR4, respectively. The net impact is that the TME provides prosurvival and proadhesion/prohoming signals to malignant PC, predicted to promote PC growth and survival. We then reviewed the signals affecting T cells (Figure 5D). These interactions are broadly categorized as inhibitory, costimulatory, chemokine, and immunomodulatory effects. The inhibitory checkpoints include canonical T-cell checkpoints TIGIT, CD96, PDCD1 (PD-1), and CTLA4, interacting with NECTIN2, PVR, PDCD1LG2 (PDL2), and CD86, respectively. Costimulatory signals of note were CD28, CD6, and CD2 interacting with CD86, ALCAM, and CD58 from the TME. Chemokine signals included TME production of CXCL9, CXCL10, and CXCL11 to provide proinflammatory T-cell retention signals. There was significant CXCL12 production, predicted to interact with CXCR4 on T cells to putatively promote T-cell exclusion by maintaining T cells in the IS niche. Further immunomodulation of T-cell function is predicted through PTGER4/PTGER2 receptors in T cells and the expression of PGE2 from tumor cells and the TME more broadly. Given the high proportion of fibroblasts in the TME, we reviewed fibroblast interactions with the TME (Figure 5E). Fibroblast interactions were categorized into interactions promoting adhesion, angiogenesis, guidance cues, immune suppression, and tumor prosurvival signals. Fibroblasts further promoted the suppressive phenotype of macrophages through the expression of CSF1, GAS6, TGFB3, and PGE2, interacting with macrophage CSF1R, MERTK/AXL, TGFBR1, and PTGER4. These interactions are predicted to drive M2 polarization. In summary, the predicted cell-cell interactions suggest that the TME provides significant supportive signals to malignant PC, particularly through macrophages and, to a lesser extent, fibroblasts. The PC directly and indirectly modulate the ECM, which promotes a fibrotic and suppressive microenvironment, promoting a suppressive phenotype in tumor macrophages. This results in a profoundly suppressive immune microenvironment with resultant predicted T-cell dysfunction and exclusion, despite the relatively minimal expression of inhibitory molecules on PC themselves. These key interactions are summarized in Figure 5F. Discussion EMD is a poor prognostic feature in patients with MM, associated with aggressive disease kinetics and therapy resistance. In this study, we describe the TME of EMD in MM, identifying a highly immunosuppressive environment dominated by suppressive macrophages, in association with T-cell dysfunction and exclusion from the bulk of tumors. There is bidirectional signaling between the TME and MM cells, with canonical MM survival pathways activated by cells of the TME. We identify recurrent microenvironment immune niches, shared between patients, defined by TME cell composition. We show that, similar to BM-restricted MM,29 interpatient variability in gene expression remains a striking feature in EMD. Despite this, the PC of patients with EMD maintain a PC transcriptome, without evidence of significant dedifferentiation or lineage plasticity. Significant proportions of EMD tumor cells are proliferating, consistent with the clinically aggressive behavior of EMD tumors. Recently, spatial transcriptomics was applied to BM-restricted MM.5 This study concentrated on the PC compartment; the authors did not identify a recurrent single MM niche in the BM microenvironment in terms of cellular composition, but the phenotype of TME cells was not exhaustively described. We identified 3 recurrent TME niches in EMD, which invoke the IS, IP, and IE environments proposed in BM-restricted MM.27 The IE niche makes up the bulk of the tumor, characterized by extensive malignant PC with extremely few immune cells. The IS niche is the next most abundant niche, characterized by the presence of suppressive macrophages, with endothelial cells and cancer-associated fibroblasts, and cytotoxic T cells. The smallest niche, the IP niche, contains the most inflammatory macrophages and has the highest proportion of infiltrating T cells, although the macrophages maintain expression of suppressive markers. The pattern of T-cell coinhibitory markers suggests dysfunction driven by the immunosuppressive TME, rather than chronic antigen stimulation. The expression of coinhibitory markers in MM is not necessarily predictive of T-cell function30,31; therefore, the activity of T cells in EMD cannot be confirmed without further experimental validation. A recent publication used single-cell (ie, nonspatial) sequencing to assess EMD lesions.14 The cellular composition they described identified natural killer and T cells as the predominant cell type in EMD. This likely reflects differences between the technologies employed, with single-cell sequencing techniques potentially biased against specific cell types such as macrophages and fibroblasts.32 In addition, the macrophages we identified are in higher numbers in the IS niche; sampling of just IE regions, the bulk of tumors, would generate a different cellular profile. Cell-cell interactions were predicted using CellPhoneDB,28 with these predictions supporting the spatial features described above. Macrophage/PC and fibroblast/PC interactions are predicted to provide significant support to PC growth and survival through quintessential MM survival pathways, including CD38 and B-cell maturation antigen (BCMA), with both APRIL and BAFF. Interestingly, the PC/macrophage interactions are predicted to drive the “M2” polarization of macrophages, contributing to tumor immune suppression. Fibroblasts further drive this suppressive macrophage phenotype, generating an ECM that additionally excludes T cells from the bulk of the tumor cells and directly contributes to T-cell dysfunction. Our findings have implications for therapeutic approaches. Consistent with published response rates in the literature, the high proportion of proliferating cells suggests an ongoing role for traditional chemotherapeutic agents in debulking EMD tumors.11,33 Because of the known short duration of response with these agents, additional novel approaches are required. The centrality of CD38 and BCMA signaling in our predicted cell-cell interactions suggests that targeting these pathways may be useful. However, the spatial analysis suggests that because the bulk of the tumor is found within the IE niche, there may be poor infiltration of immune cells that may contribute to the limited efficacy. This is consistent with published literature revealing poor outcomes from CD38-targeting therapies in EMD.6 Furthermore, there is variable expression of these markers among patients. The marked immunosuppressive TME and T-cell exclusion we describe will likely limit the impact of single-agent immunotherapies targeting coinhibitory molecules. It may be that an approach of tumor debulking with cytotoxic chemotherapy to reduce disease burden, followed by therapies that target the TME, in conjunction with T-cell approaches, either bispecific antibody or chimeric antigen receptor T-cell based, may improve outcomes relative to using these in isolation. Therapies to delete suppressive macrophages could be studied as an adjunct to these therapies; CSF-1/CSF-1R–targeting approaches show promise in this area.34 In addition, the impact of fibrosis,35 and thus fibroblasts, on MM outcomes, taken with our observations in EMD, provides rationale for further investigation of fibroblast-directed therapies such as transforming growth factor β pathway inhibition with agents like pirfenidone.36 Our study has limitations that should be acknowledged. The sample size is relatively small, particularly given the variations among samples in immune cell infiltrates. In addition, because of the breadth of the genes assessed, the depth of coverage was lower than that observed with targeted approaches. This affects the ability to precisely define some cell types without overclustering. Spatial transcriptomics using Xenium is also not truly quantitative in the same way as sequencing-based techniques because of potential differential performance of primer sets relative to one another. Reassuringly, we demonstrated that biologically relevant features were detectable using our technique, for example, higher MYC expression in a patient with known t(8;14) than in the rest of the cohort. Finally, despite the >5000 genes analyzed, some relevant genes were unavailable for assessment, such as SDF1 (CD138), which underestimates key cell-cell interactions. In sum, our study describes the TME of EMD, noting significant interpatient heterogeneity in the PC transcriptome with the presence of recurrent immune microenvironments. Our approach allows reliable identification of macrophages and fibroblasts, which have been, to date, underestimated in EMD tumors analyzed using alternative techniques. The TME is immunosuppressive, driven by suppressive macrophages and fibroblasts, with cell-cell interactions from TME cells that augment PC survival and exclude T cells from the bulk of tumors. The reciprocal interactions between PC and TME cells drive this immunosuppressive environment with targetable interactions. We provide a rationale for treatment regimens involving initial tumor debulking and targeting the specific interactions between PC and the TME to improve treatment responses. Conflict-of-interest disclosure: G.R.H. has consulted for Generon Corporation, NapaJen Pharma, iTeos Therapeutics, Commonwealth Serum Laboratories, Cynata Therapeutics, Neoleukin Therapeutics, and Incyte Pharma; and reports research funding from Compass Therapeutics, Syndax Pharmaceuticals, Applied Molecular Transport, Serplus Technology, Heat Biologics, Laevoroc Oncology, iTeos Therapeutics, Genentech, Incyte Pharma, and Commonwealth Serum Laboratories. E.W.N. is a cofounder, advisor, and shareholder of ImmunoScape Pte Ltd. The remaining authors declare no competing financial interests.