Cross-tissue atlas of mucosa-associated lymphoid tissue lymphomas reveals intratumoral heterogeneity and microenvironmental subtypes Authors: Kai Kang, Yijun Wu, Xu Sun, Yuqi Yang, Wenyi Liang, Jinrong Yang, Yun Tang, Bing Xiang, Ting Niu, Ailin Zhao Published: 2026 ABSTRACT SummaryMucosa-associated lymphoid tissue (MALT) lymphoma, a common primary extranodal B cell lymphoma that arises in diverse mucosal tissues, has a pathogenesis closely linked to the supportive lymphoma microenvironment (LME). However, its microenvironmental heterogeneity and therapeutic implications remain unclear. Here, we report a cohort of 89 MALT lymphoma samples from multiple anatomical sites, profiled by bulk (n = 89), single-nucleus (n = 16), and spatial (n = 9) RNA sequencing. Integrative analysis of 120 bulk transcriptomes from our cohort and two independent cohorts, guided by cell-type-resolved signatures derived from single-nucleus data, identifies three LME subtypes: mesenchymal, inflammatory, and depleted. Single-cell and spatial mapping further reveal distinct cellular compositions and interaction networks across these subtypes, particularly in relation to B cell activation and angiogenesis, suggesting distinct biological states and subtype-specific therapeutic vulnerabilities. This study establishes an LME-based classification framework for MALT lymphoma with biological and clinical relevance.Keywords: MALT lymphoma, lymphoma microenvironment, microenvironmental subtype, intratumoral heterogeneity, single-nucleus RNA sequencing, spatial transcriptomics. Introduction Mucosa-associated lymphoid tissue (MALT) lymphoma constitutes the most common subtype of marginal zone lymphomas, characterized as a group of low-grade extranodal B cell malignancies that arise at diverse mucosal sites.1 Accumulating evidence indicates that MALT lymphomagenesis involves chronic antigen-driven immune dysregulation, leading to malignant transformation of clonally expanded B cells, with specific genetic alterations facilitating disease progression.2 Such persistent antigenic stimulation may originate from chronic infections (e.g., Helicobacter pylori [HP] at gastric mucosa3 and Chlamydophila psittaci in ocular adnexal tissues4) or autoimmune disorders (e.g., Sjögren syndrome5 and Hashimoto’s thyroiditis5,6). Owing to these diverse etiological drivers across anatomic sites, significant heterogeneity exists in molecular profiles and lymphoma microenvironment (LME), warranting further investigation to advance both mechanistic insight and precision medicine approaches. The molecular paradigm of MALT lymphomas is characterized by recurrent genetic alterations that dysregulate key signaling pathways, particularly the nuclear factor-kappa B (NF-κB) pathway, which plays a critical role in lymphocyte activation and survival.2 Notably, genetic translocations such as t(11;18)(q21;q21), which generates the BIRC3/MALT1 fusion, are frequently observed in gastric and pulmonary MALT lymphomas.7 This fusion leads to constitutive activation of NF-κB signaling, promoting B cell survival and lymphomagenesis, and also mediates resistance to conventional antibiotic therapy in gastric cases.1,7,8 In addition, t(14;18)(q32;q21) and t(1;14)(p22;q32) are recurrent genetic alterations in MALT lymphomas, giving rise to IGH/MALT1 and IGH/BCL10 fusions, respectively, which also contribute to activation of canonical and non-canonical NF-κB pathways.2,9 Furthermore, mutations in TNFAIP3, MYD88, and other NF-κB regulators dysregulate the same signaling pathway and exhibit site-specific prevalence⁠.2 Collectively, these genetic aberrations synergize with chronic immune stimulation to promote MALT lymphoma pathogenesis. LME plays a pivotal role in regulating tumor development, progression, and treatment responses. A key LME feature in MALT lymphoma is the synergistic interplay between chronic immune stimulation and genetic alterations, which collaboratively establish a niche that supports tumor cell survival, proliferation, and immune evasion. Although HP infection is implicated in gastric MALT lymphoma, it drives neoplastic B cell growth indirectly by activating specific tumor-infiltrating T cells through T-B cell interactions.10,11,12 Moreover, bystander T cell within LME can provide support via soluble ligands and cytokines such as CD40L and BAFF, thereby activating the non-canonical NF-κB pathway.13 However, the composition and functional interplays within the LME of MALT lymphomas remain poorly understood, particularly variations related to anatomical sites, immune landscapes, and genetic contexts. In this study, by leveraging a large cohort of formalin-fixed paraffin-embedded (FFPE) samples and integrating bulk, single-nucleus, and spatial RNA sequencing (RNA-seq), we systematically delineate the heterogeneous microenvironmental landscape and spatial architecture of multi-site MALT lymphomas. Additionally, an LME-based subtyping system was proposed to inform the development of synergistic therapeutic strategies and improve patient outcomes. Results Identification of LME subtypes of MALT lymphoma We retrospectively collected 89 FFPE samples pathologically diagnosed as MALT lymphoma from West China Hospital between January 2020 and June 2023 (WCH cohort), most of which originated from the stomach (n = 38) and ocular adnexa (n = 30) (Table S1). Considering the limitation that the retrospective cohort lacks normal controls, we further included two publicly available bulk RNA-seq cohorts containing both ocular adnexa MALT lymphoma and normal tissues (Figure 1A).14,15 In total, 145 samples were included in the integrated analysis, comprising 120 tumor samples and 25 normal controls. Among the tumor samples, 89 were derived from our in-house WCH cohort, 21 from the cohort reported by Shi et al. (GSE171059),14 and 10 from the cohort reported by Wolf et al. (GSE218618).15 The tumor samples were predominantly derived from the ocular adnexa (n = 61, 50.8%) and stomach (n = 38, 31.7%) (Figure 1B). Among the 120 tumor samples, 110 had clinical information available. The majority of patients were male (n = 61, 55.5%), aged ≥60 years (n = 58, 52.7%), and diagnosed at Ann Arbor stage I (n = 69, 62.7%) (Figure 1C; Table S1). After batch effect correction and data integration, principal-component analysis (PCA) revealed no obvious separation among different cohorts or tumor sites (Figure S1A). Compared with normal controls, MALT lymphoma samples displayed obviously increased expression of B cell-related genes, such as MS4A1, CD79A, and PAX5 (Figure S1B), together with enrichment of proliferation-related pathways such as E2F targets and G2/M checkpoint, as well as oncogenic pathways such as unfolded protein response and DNA repair (Figure S1C). Identification of LME subtypes of MALT lymphoma (A) Schematic illustration of the study design and workflow. (B) Bar plot showing the number of MALT lymphoma samples from different anatomical sites and normal controls in the integrated bulk RNA-seq dataset. Total n = 145. (C) Bar plots showing the distribution of samples according to clinical characteristics, including gender (left), age (middle), and stage (right). Total n = 110. (D) Schematic illustration of single-nucleus RNA-seq based on FFPE curls. (E) UMAP plot of 96,778 nuclei profiled by snRNA-seq, depicting six major cell lineages. (F) UMAP plot of 63,176 B cells, 10,001 T cells, 8,648 myeloid cells, and 7,290 stromal cells profiled by snRNA-seq, depicting finer-resolution cellular subpopulations. (G) Heatmap (left) of the activity scores of cell subpopulation signatures and clinical characteristics across 120 MALT lymphoma samples, denoting three LME subtypes. Line chart (right) showing the mean score for each signature across the three LME subtypes. The matrix of activity scores was standardized by row Z score transformation prior to visualization. Total n = 120. (H) Line chart showing the enrichment of cell subpopulations in different LME subtypes. Statistical significance was evaluated using Fisher’s exact test with Benjamini-Hochberg correction for multiple comparisons. (I) Relative proportion of anatomical sites (top) and BIRC3/MALT1 fusion (bottom) across LME subtypes. Statistical significance was assessed using Fisher’s exact test. Total n = 110. (J) Kaplan-Meier curves of PFS stratified by LME subtypes in the integrated cohort with available follow-up. Statistical significance was evaluated using the log rank test. To dissect LME subtypes across 120 MALT lymphoma samples, we performed single-nucleus RNA-seq (snRNA-seq) on FFPE samples using the 10× Genomics Flex platform. Given the technical requirements for sample quality and total cell nucleus count, 16 surgical resection samples were selected for subsequent snRNA-seq analysis (Figure 1D). After background noise removal and quality control, 96,778 nuclei were processed through normalization, highly variable gene selection, dimensionality reduction, batch correction, and clustering. Based on canonical markers, six major cell lineages were annotated: B cells, T cells, myeloid cells, stromal cells, hepatocytes, and alveolar cells, visualized by uniform manifold approximation and projection (UMAP) (Figures 1E and S2A). Further reclustering delineated multiple subpopulations within each lineage, characterized by distinct marker expression profiles (Figures 1F and S2B–S2D). To distinguish malignant B cells (B_Mal) from normal B cells (B_Nor), we applied inferCNV to score copy number variation (CNV) and gene set variation analysis (GSVA) to score with marginal zone B cell (MZB) signatures.16 Both CNV and MZB scores were markedly higher in B_Mal, consistent with the MZB origin of MALT lymphoma (Figures S3A–S3C). B_Mal represented the overwhelming majority of B cells, in line with previous single-cell studies of splenic marginal zone lymphoma.17 To minimize contamination from adjacent normal tissues, we evaluated Shannon entropy across anatomical origins. Five subpopulations with low entropy values, each showing strong enrichment (>80%) in lung or liver samples, were identified as likely derived from normal tissue within the FFPE specimen and were, therefore, excluded from downstream analyses (Figure S4A). We next constructed a MALT lymphoma-specific microenvironmental signature based on subclusters and applied GSVA for feature scoring, followed by non-negative matrix factorization (NMF) clustering (Figure S4B; Table S2). This approach identified three distinct LME subtypes: LME1 (mesenchymal), characterized by enrichment of stromal components; LME2 (inflammatory), enriched for lymphocytes; and LME3 (depleted), exhibiting globally attenuated microenvironmental features (Figure 1G). Importantly, the differences among these LME subtypes were highly concordant with results derived from GSVA using previously reported LME-related gene sets (Figure S4C).18 Based on the LME classification, the 16 snRNA-seq samples comprised four LME1 samples, eight LME2 samples, and four LME3 samples. PCA of the proportional composition of the remaining subpopulations in snRNA-seq data demonstrated partially consistent patterns across different LME subtypes (Figure S4D). Odds ratio analysis revealed that LME1 was enriched for endothelial cells, pericytes, neutrophils, and Mac_MARCO, whereas LME2 was enriched for T cells, dendritic cells (DCs), and Mac_SPARC, concordant with bulk RNA-seq-based LME classification (Figure 1H). Association with clinical features indicated that gastric MALT lymphoma and BIRC3/MALT1 fusions were more common in LME1 (Figure 1I). The frequency of TNFAIP3 inactivation mutations did not differ significantly among LME subtypes (Figure S4E). Survival analyses demonstrated that LME2, advanced stage, and BIRC3/MALT1 fusion were associated with significantly worse progression-free survival (PFS) (Figures 1J and S5A). Moreover, all three factors remained significant in both univariate and multivariate Cox regression analyses, indicating that they represent independent prognostic factors for PFS in patients with MALT lymphoma (Figures S5B and S5C). Gene set enrichment analysis demonstrated that, compared with LME3, LME1 was significantly enriched for TGF-β, extracellular matrix (ECM), and angiogenesis pathways, while LME2 showed prominent enrichment of interferon signaling and immune activation pathways (Figure S5D). In addition, we observed higher predicted activity of TGF-β1 and G-CSF in LME1, and of TNF-α and IFN-γ in LME2 (Figure S5E). Collectively, these findings support the biological distinctiveness and reliability of LME classification. Functional heterogeneity and therapeutic vulnerabilities of malignant B cells Compared with normal B cells, malignant B cells exhibited significant upregulation of cell proliferation and B cell receptor signaling pathways (Figure S6A). Building on previously established approaches for dissecting intratumoral heterogeneity,19 we analyzed the heterogeneity of malignant B cells using NMF to decompose their expression profiles into distinct gene expression programs. Programs that remained stable across variations in clustering parameters and were shared among multiple samples were defined as robust programs. These robust programs were further clustered based on the proportion of shared genes, ultimately yielding 7 meta-programs (MPs), each comprising 50 representative genes (Figure 2A; Table S3). Functional heterogeneity and therapeutic vulnerabilities of malignant B cells (A) Heatmap showing similarity among robust NMF programs based on their top 50 genes. Programs are ordered by clustering and grouped into MPs (marked by black dashed lines). (B) Bar plots showing enriched pathways identified by over-representation analysis of gene sets from each MP. (C) Pairwise correlations among MP signature scores in B_Mal. Correlation coefficients were calculated using Spearman’s correlation. Total n = 61,844. (D) Correlations between CytoTRACE2 scores and signature scores of each MP in B_Mal. Correlation coefficients were calculated using Spearman’s correlation. Total n = 61,844. (E) Kaplan-Meier curves of PFS stratified by B_Mal signature scores in the integrated cohort with available follow-up. Statistical significance was evaluated using the log rank test. (F) Forest plot depicting the PFS hazard ratio stratified by signature scores of each MP and B_Mal signature scores. Statistical significance was evaluated using the univariate Cox proportional hazards regression. Error bars indicate 95% confidence intervals. Total n = 88. (G) Correlations between MP7 signature scores and proportion of CD4_Treg within CD4+ T cells. Correlation coefficients and p values were calculated using Pearson’s correlation. Total n = 16. (H) Boxplots showing the signature scores of each MP across LME subtypes in snRNA-seq samples (top) and bulk RNA-seq samples (bottom). The centerline denotes the median, the lower and upper hinges represent the 25th and 75th percentiles, and whiskers extend to 1.5 times the interquartile range. Statistical significance was assessed using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons. All comparisons in snRNA-seq samples yielded adjusted p < 0.0001. Total n = 61,844 (top) and 120 (bottom). (I) Density curves of drug sensitivity distributions among LME subtypes. Dashed vertical lines indicate the median value for each LME subtype. Total n = 61,844. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. Functional annotation linked these MPs to distinct processes, including cell cycle G1/S (MP1), unfolded protein response (MP2), cell cycle G2/M (MP3), B cell activation (MP4), nucleic acid binding (MP5), immunoglobulin production (MP6), and immune modulation (MP7) (Figures 2A and 2B). Cell cycle and nucleic acid binding features have also been identified in single-cell studies of diffuse large B cell lymphoma, suggesting shared malignant B cell programs across B cell lymphoma entities.20 Notably, MP6 reflects an immunoglobulin program consistent with plasmacytic differentiation in MALT lymphoma,21 and its scores inversely correlate with both the proliferative MP3 and CytoTRACE2-inferred developmental potential, supporting a terminally differentiated malignant B cell state (Figures 2C and 2D). We derived a B_Mal signature from snRNA-seq-defined malignant B cells, such as canonical B cell markers CD79A and MS4A1, together with CD37, which is highly expressed in lymphoma (Table S2).22 The B_Mal signature was applied to bulk RNA-seq samples for scoring, and patients were stratified into two groups according to the median score (Figure 2E). Univariate Cox regression further identified MP1, MP3, and MP7 as adverse prognostic factors (Figure 2F). The adverse associations of MP1 and MP3 may reflect their link to heightened proliferative activity, while MP7 included chemokines CCL22 and CCL17, known to recruit regulatory T (Treg) cells.23,24 Consistently, MP7 scores positively correlated with the proportion of CD4T_Treg cells among CD4+ T cells (Figure 2G). Interestingly, MP scores from snRNA-seq and bulk RNA-seq were largely concordant across LME subtypes, despite the bulk transcriptomic data being confounded by non-malignant lineages. MP scores were not significantly stratified by clinical characteristics such as stage and age, suggesting that LME classification better captures functional states of malignant B cell (Figures 2H and S6B). For example, LME2 showed higher MP4 (B cell activation) and MP7 (immune modulation) scores, whereas LME3 was enriched for MP3 (cell cycle G2/M). Cell cycle scoring calculated using Seurat also supported that LME3 had a higher G2/M score and G2/M cell cycle proportion (Figure S6C). Drug sensitivity prediction using Beyondcell further revealed distinct therapeutic vulnerabilities: LME1 was more sensitive to vascular endothelial growth factor receptor (VEGFR) inhibitors (sorafenib, axitinib, and sunitinib), consistent with its endothelial-rich microenvironment; LME2 was more sensitive to the Bruton’s tyrosine kinase (BTK) inhibitor ibrutinib, consistent with its higher MP4 score; and LME3 showed sensitivity to microtubule-targeting agents such as vincristine and vinorelbine that cause G2/M arrest, consistent with MP3 enrichment (Figures 2I and S7A). Target-based drug sensitivity prediction using scRANK further corroborated these results, demonstrating higher sensitivity of LME1 samples to VEGFR2-targeting agents and of LME2 samples to BTK-targeting therapies (Figure S7B). Immunohistochemical analyses provided additional validation, showing higher expression of VEGFR2 and BTK in LME1 and LME2 samples, respectively, while differential Ki-67 expression further supported a higher G2/M cell cycle proportion in LME3 (Figure S7C). These subtype-specific response profiles underscore the potential clinical utility of LME classification for therapeutic stratification. Diversity of immune cell subpopulations across LME subtypes To delineate immune diversity across LME subtypes, we resolved major immune lineages into finer subtypes. T cells were divided into CD8+ subsets (CD8T_Teff, CD8T_Temra, and CD8T_Tex) and CD4+ subsets (CD4T_Treg, CD4T_Tfh, and CD4T_Tn). Subset proportions varied by LME: CD8T_Tex was significantly higher in LME2 than in LME1, whereas CD8T_Temra displayed the opposite pattern (Figures 3A and S8A). Guided by canonical marker expression, we classified CD8+ T cell subsets into three groups: CD8T_Teff, with high CCL5 and GZMK expression corresponding to an effector-memory phenotype25,26; CD8T_Temra, defined by GNLY, GZMB, and PRF1 corresponding to cytotoxic and terminal effector memory T cells (Temra); and CD8T_Tex, defined by HAVCR2, LAG3, and PDCD1 corresponding to an exhausted phenotype. For CD4+ T cells, CD4T_Treg expressed FOXP3 together with IL2RA and TNFRSF4 corresponding to Treg cells, CD4T_Tfh expressed CXCL13 together with PDCD1 and CXCR5 corresponding to follicular helper T (Tfh) cells, and CD4T_Tn expressed IL7R together with CCR7 and SELL corresponding to naive T cells (Figures 3B and S2C). These marker-based assignments were corroborated by enrichment scores from functional gene sets; for example, cytotoxicity scores were highest in the CD8T_Temra, whereas exhaustion scores were highest in the CD8T_Tex (Figure 3C). Immune and stromal cell heterogeneity across LME subtypes (A) Relative proportion of subpopulations of CD8+ T cells across LME subtypes (left, LME-level aggregation). Line graph showing differences in the proportions of subpopulations across samples in each LME subtype (right, sample-level aggregation), with error bars representing mean ± SEM. Statistical significance was assessed using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons. Total n = 4,207 (left) and 16 (right). (B) Bubble plot showing the expression of marker genes across T cell subsets. (C) Heatmap showing enrichment of representative functional signatures in T cell subsets. The matrix of enrichment scores was standardized by row Z score transformation prior to visualization. (D) Boxplots showing the signature scores of each CD8+ T cell subset across LME subtypes. The centerline denotes the median, the lower and upper hinges represent the 25th and 75th percentiles, and whiskers extend to 1.5 times the interquartile range. Statistical significance was assessed using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons. (E) Same visualization as (A), but for macrophages. Total n = 5,213 (left) and 16 (right). (F) UMAP plots of 5,213 macrophages profiled by snRNA-seq, along with the expression patterns of MRC1 and FOLR2. (G) Representative gene expression profiles associated with MRC1 and FOLR2 expressions. (H) Heatmap showing enrichment of representative functional signatures in macrophage subsets. The matrix of enrichment scores was standardized by row Z score transformation prior to visualization. (I) Radar plots showing the variation in representative pathways among different myeloid cell subpopulations across LME subtypes. (J) Same visualization as (A), but for fibroblasts. Total n = 4,423 (left) and 16 (right). (K) Heatmap showing enrichment of representative pathways in fibroblast subsets. The matrix of enrichment scores was standardized by row Z score transformation prior to visualization. (L) Heatmap showing the regulatory potential of top-ranked ligands on Fibro_CCL19 signature genes, as determined by NicheNet analysis. (M) Same visualization as (L), but for Fibro_LRRC15. (N) Heatmap showing representative PROGENy-derived pathway activity in fibroblast subsets. The matrix of activity scores was standardized by column Z score transformation prior to visualization. (O) Heatmap showing representative pySCENIC-derived transcription factor activity in fibroblast subsets. The matrix of activity scores was standardized by column Z score transformation prior to visualization. (P) Boxplots showing the signature scores of Fibro_CCL19 and Fibro_LRRC15 across LME subtypes. The centerline denotes the median, the lower and upper hinges represent the 25th and 75th percentiles, and whiskers extend to 1.5 times the interquartile range. Statistical significance was assessed using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. Functional comparison revealed a graded decline in cytotoxicity across CD8+ subsets from LME1 to LME2 to LME3, while exhaustion peaked in LME2 (Figure 3D). Activities of cytokine and chemokine programs, including IFN-α and IFN-γ, mirrored these patterns (Figure S8B). Taken together, these results show that T cell heterogeneity across LME subtypes is both compositional and functional, reflecting differences in subset proportions and in cytotoxicity and exhaustion states. The prominent exhaustion signature in LME2 further suggests a higher likelihood of benefit from immune checkpoint blockade directed at reversing T cell exhaustion. Macrophages were subdivided into Mac_SPARC, Mac_STAB1, Mac_MARCO, and Mac_CHIT1. Mac_SPARC was more abundant in LME2 than in LME1, whereas Mac_MARCO was significantly enriched in LME1 (Figure 3E). Notably, these four subtypes were segregated by expression of two classical markers, MRC1 (encoding CD206) and FOLR2, yielding Mac_SPARC (FOLR2+ MRC1−), Mac_STAB1 (FOLR2+ MRC1+), Mac_MARCO (FOLR2− MRC1+), and Mac_CHIT1 (FOLR2− MRC1−) (Figures 3F and 3G). We also observed mutually exclusive expression of FOLR2 and TREM2 within macrophages, consistent with prior reports in solid tumors.27,28 Macrophage subtypes displayed pronounced functional heterogeneity. In the context of canonical M1/M2 polarization, CD206 serves as a classical marker of the M2 phenotype. Accordingly, CD206− Mac_SPARC and Mac_CHIT1 exhibited higher M1 scores, while CD206+ Mac_MARCO was associated with higher M2 scores (Figure 3H). Notably, Mac_MARCO scored highest in angiogenesis, phagocytosis, and antigen presentation, indicating a central hub role within the LME (Figure 3H). While both Mac_SPARC and Mac_MARCO strongly influenced T helper (Th) differentiation, Mac_SPARC was biased toward Th1 and Mac_MARCO toward Th2, in line with their respective M1/M2 profiles (Figure 3H). Immune response enrichment analysis (IREA) identified distinct upstream cytokines for these macrophage states, with TNF-α associated with Mac_SPARC and IL-4/IL-13 associated with Mac_MARCO (Figure S8C). As canonical Th2 cytokines, IL-4 and IL-13 promote CD206 expression and drive alternative (M2) macrophage activation.29 Gene regulatory network analysis likewise revealed heterogeneity in putative upstream regulatory programs across subpopulations (Figure S8D). DCs were categorized as cDC1_XCR1, cDC2_CLEC10A, and pDC_CLEC4C. LME1 was enriched for cDC2_CLEC10A, whereas LME2 was enriched for cDC1_XCR1 and pDC_CLEC4C (Figure S8E). IREA suggested TNF-α as a principal driver of cDC1_XCR1, IL-1α/IL-1β for cDC2_CLEC10A, and IFN-α1/IFN-β for pDC_CLEC4C (Figure S8F). Functionally, conventional DCs (cDCs) scored higher for antigen presentation overall; cDC1_XCR1 preferentially presented via MHC class I and cDC2_CLEC10A via MHC class II, aligning with cDC1 and cDC2 identities, respectively (Figure S8G). When comparing all myeloid subtypes across LME subtypes, LME1 showed broadly higher angiogenesis and neutrophil chemotaxis, in keeping with its unique cellular composition (Figure 3J). By contrast, T cell and lymphocyte chemotaxis varied with the predominant macrophage subtype, exemplified by stronger signals from Mac_MARCO in LME1 and from Mac_SPARC in LME2, indicating interdependence between myeloid subtype function and LME composition (Figure 3I). Diversity of stromal cell subpopulations across LME subtypes To identify stromal diversity, we performed refined subclustering of fibroblasts and endothelial cells. Fibroblasts comprised Fibro_PI16, Fibro_CCL19, and Fibro_LRRC15, with Fibro_CCL19 significantly enriched in LME2 relative to LME1 (Figure 3J). Together with prior reports and CytoTRACE2-inferred developmental potential, Fibro_PI16 likely represents a precursor fibroblast state capable of differentiating into the other two subsets (Figure S9A).30 Functionally, Fibro_CCL19 showed a higher level of antigen processing and presentation, lymphocyte chemotaxis, and TNF-α signaling via NF-κB, whereas Fibro_LRRC15 was enriched for ECM-related programs (Figure 3K). Leveraging subpopulation-specific gene signatures, NicheNet implicated interactions, including TNF-NFκB2, as upstream drivers of Fibro_CCL19 and identified TGFB1/TGFB2 as key ligands for Fibro_LRRC15 (Figures 3L and 3M). The subpopulation specificity of ligand-receptor programs was independently supported by PROGENy-inferred pathway activities (Figure 3N). Gene regulatory network analysis indicated that Fibro_CCL19 is characterized by inflammation-associated transcription factors, including IRF1, STAT1, NFKB1, and NFKB2 (Figure 3O). Integrating these results with prior reports, we classified Fibro_CCL19 as an inflammatory fibroblast state and Fibro_LRRC15 as a myofibroblast state.31,32 Even within the same fibroblast subtype, function varied by LME subtypes. Fibro_CCL19 and Fibro_LRRC15 exhibited higher TNF-α signaling via NF-κB activity in LME2 and higher ECM organization in LME1, with similar patterns across additional pathways (Figures 3P and S9B). Thus, fibroblast heterogeneity in MALT lymphoma spans both subpopulation identity and LME context. Endothelial cells comprised three subtypes: Endo_venous, Endo_arterial, and Endo_lymphatic (Figure S9C). Endo_venous and Endo_arterial were assigned to the vascular endothelium based on high PECAM1 and VWF expression. Within this compartment, Endo_venous displayed high ACKR1 and SELP and was annotated as venous endothelium, whereas Endo_arterial displayed high GJA5 and SEMA3G and was annotated as arterial endothelium. Endo_lymphatic showed high expression of the lymphatic markers PROX1, FLT4, and LYVE1, annotated as lymphatic endothelium. Endo_arterial-specific gene signatures implicated VEGFA-related interactions as major drivers of its transcriptomic profile (Figure S9D). Consistently, angiogenesis-related pathway scores were higher in Endo_arterial than in Endo_venous (Figure S9E). Together with CytoTRACE2-inferred cellular potency, these data suggest that Endo_arterial may progress toward a more differentiated endothelial state under VEGF signaling (Figure S9F). Across LME subtypes, vascular endothelium in LME1 exhibited higher angiogenesis, whereas lymphatic endothelium in LME2 showed higher lymphocyte chemotaxis and interferon signaling (Figures S9E and S9G). These functional contrasts align with the defining features of the LME subtypes, with LME1 favoring stromal and vascularization and LME2 favoring lymphocyte recruitment and activation. Thus, endothelial states not only mirror the prevailing microenvironment but also actively contribute to shaping it. Distinct cellular modules across different LME subtypes After defining subpopulation heterogeneity across LME subtypes, we integrated all subpopulations and, using correlations among their proportional abundances, identified cellular modules (CMs) in MALT lymphoma. CM1 comprised CD8T_Tex, CD4T_Treg, CD4T_Tfh, DCs, Mac_SPARC, Fibro_CCL19, Endo_lymphatic, and FDC_CR2, whereas CM2 comprised vascular endothelial subsets, Neu_G0S2, Mac_MARCO, and Peri_PDGFRB (Figure 4A). CM1 was more prominent in LME2, whereas CM2 was more prominent in LME1, consistent with the defining features of these LME subtypes (Figure 4B). Cellular modules and intercellular communication networks (A) Heatmap showing pairwise correlations of subpopulation proportions among MALT lymphoma samples, with CMs marked by black dashed lines. Correlation coefficients were calculated using Spearman’s correlation. (B) Boxplots showing the proportion of each CM across LME subtypes. The centerline denotes the median, the lower and upper hinges represent the 25th and 75th percentiles, and whiskers extend to 1.5 times the interquartile range. Statistical significance was assessed using the two-sided Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons. ∗p < 0.05. Total n = 16. (C) Bar plots comparing intercellular interaction strength across LME subtypes, shown for all cell subpopulations, CM1 constituents, and CM2 constituents. (D) Relative information flow of representative pathways highlights the varying importance of each pathway across different LME subtypes. (E) Scatterplots showing outgoing and incoming interaction strengths, facilitating the visualization of subpopulation-level differences across LME subtypes. (F) Heatmap (bottom left) showing the relative overall interaction strength of CM1 constituents across CXCL, CD40, ICAM, and CCL pathways in different LME subtypes. Bar plot (top) displaying the summed relative strengths of each cell type across these pathways. Bar plot (right) displaying the summed relative strengths of each pathway across these cell types. (G) Bubble plot showing the communication probabilities of ligand-receptor pairs in the CXCL and ICAM pathways across different cell-cell interactions within CM1. (H) Same visualization as (F), but for CM2 constituents across VEGF pathways. (I) Bubble plot showing the communication probabilities of ligand-receptor pairs in the VEGF pathway across different cell-cell interactions within CM2 across LME subtypes. CellChat analysis revealed stronger overall intercellular interaction strength in LME1 and LME2 than in LME3, likely reflecting the sparser cellular composition of LME3 (Figure 4C). Within modules, interactions among CM1 constituents were strongest in LME2, whereas interactions among CM2 constituents were strongest in LME1, mirroring module proportions (Figure 4C). Relative information flow indicated that stromal-associated pathways, including VEGF, TGF-β, PDGF, and collagen, were more dominant in LME1, whereas pathways related to lymphocyte and B cell activation, including ICAM, CXCL, CD40, and BAFF, were more dominant in LME2 (Figure 4D). At the subpopulation level, Mac_SPARC, FDC_CR2, and Fibro_CCL19 exhibited stronger interactions in LME2 than in other LME subtypes, while Neu_G0S2 and Mac_MARCO were stronger interactors in LME1 (Figure 4E). For malignant B cell activation, the key pathways CD40 and BAFF signaling both peaked toward B_Mal in LME2 (Figure S10A).33 However, the two pathways differed in source. CD40 signaling to B_Mal was predominantly paracrine from other cell types, whereas BAFF signaling was largely autocrine from B_Mal, indicating distinct activation dependencies (Figures S10B–S10D). We further examined ligand-receptor differences across LME subtypes within modules. In CM1, the aggregate signaling strength, calculated as the sum of relative contributions from all constituent cell types, was higher in LME2 for the CXCL, ICAM, and CD40 pathways (Figure 4F). At the level of individual ligand-receptor pairs, interactions linking Mac_SPARC, FDC_CR2, and Fibro_CCL19 to the T cell subsets contained in CM1 were broadly stronger in LME2 (Figures 4G and S10E). In CM2, aggregate signaling strength of VEGF and TGF-β was higher in LME1, with Neu_G0S2 being most discriminative (Figure 4H). Neu_G0S2 and Mac_MARCO, as senders to vascular endothelial cells, especially Neu_G0S2, showed broadly increased signaling in LME1, particularly via VEGFA-related ligand-receptor pairs, thereby accounting for the heightened angiogenic capacity observed in LME1 (Figures 4I and S10F). Given that LME1 predominates in gastric MALT lymphoma and that HP infection plays a central role in gastric MALT lymphomagenesis, we further investigated the relationship between neutrophils and HP infection-related gene signatures, as well as angiogenesis signatures. Analyses of gastric MALT lymphoma samples from this study, together with two independent transcriptomic microarray datasets,34,35 consistently revealed a positive correlation between neutrophil-associated signatures, HP infection-related gene signatures, and angiogenesis-related pathways (Figures S11A and S11B).36 Altogether, cell-cell interaction analysis indicated that subtype-specific interactions driving lymphocyte recruitment and functional activation, together with angiogenesis, shape the distinctive cellular landscapes among LME subtypes. Spatial mapping of cell-cell interactions To resolve spatial organization, we performed 10× Visium spatial transcriptomics on nine FFPE samples (three from each LME subtype), selected to ensure balanced representation across LMEs and inclusion of samples from multiple extranodal sites (Figures 5A and S12; Table S1). Integration with snRNA-seq using cell2location enabled spatially resolved mapping of cell-type distributions. Colocalization analysis identified five spatial factors (SFs) that captured recurrent patterns of subpopulation co-occurrence. CM1-associated subpopulations were primarily enriched in SF2 and SF3, whereas CM2-associated subpopulations mapped to SF5, supporting spatial colocalization within each module (Figure 5B). MISTy-derived interaction communities showed similar organizations (Figures S13A and S13B). Spatial patterns varied across SFs. SF4 was enriched in stromal regions, while SF2 and SF5, linked to lymphoid and endothelial niches, respectively, showed more dispersed distributions within the microenvironment (Figure 5C). In line with module proportions, SF2 and SF3 were more prominent in LME2, whereas SF5 was more prominent in LME1 (Figure 5D). Representative sections illustrated that CM1 cell types preferentially localized near B_Mal, while CM2 cell types, although internally colocalized, were overall more distant from B_Mal (Figures 5E and 5F). Spatial organization of cell-cell interactions (A) Schematic illustration of spatial RNA-seq and mIF based on FFPE slides. (B) Bubble plot showing the importance of cell subpopulations in each SF. (C) Spatial visualization of H&E staining and the distribution of SFs in a representative slide. (D) Line chart showing the mean importance for each SF across the three LME subtypes. (E) Spatial mapping of cell types onto spatial RNA-seq slides utilizing the cell2location method in CM1 constituents and B_Mal. Estimated cell abundance is depicted by color intensity for all cell types (left) as well as individually for each cell type (right). (F) Same visualization as (E), but for CM2 constituents and B_Mal. (G) Schematic illustration (left) showing the calculation of nearest distance between B_Mal+ spots (red) and other cell type-positive spots (gray). If two spots overlap, the nearest distance is defined as 0. Ridge plots (right) depicting the distribution of nearest distances between B_Mal+ spots and indicated cell types across LME subtypes. Dashed vertical lines indicate the median distance within each LME subtype. (H) Schematic illustration (left) showing the calculation of the nearest distance between focal spots (red) and cell type-positive spots (gray). Bubble plot (right) depicting the correlation between ligand expression and the distance to the indicated cell type across LME subtypes. Correlation coefficients were calculated using Spearman’s correlation. (I) Representative images of mIF staining indicating T cell subpopulations in the LME. Arrows highlight example cells with colocalized fluorescence signals on the slides. (J) Boxplots showing the average distance from PAX5+ cells to CD3+ CD8+ CD4− PD-1+ cells (top) and CD3+ CD4+ CD8− FOXP3+ cells (bottom) across samples in each LME subtype. n = 5. (K) Representative images of mIF staining indicating vascular endothelial cells and related myeloid subpopulations in the LME. Arrows highlight example cells with colocalized fluorescence signals on the slides. (L) Boxplots showing the proportion of CD31+ cells (top) and average distance from CD31+ cells to CD11b+ CD15+ cells (bottom left) and CD206+ MARCO+ cells (bottom right) across samples in each LME subtype. n = 5. In (J and L), the centerline denotes the median, the lower and upper hinges represent the 25th and 75th percentiles, and whiskers extend to 1.5 times the interquartile range. Statistical significance was assessed using the two-sided Student’s t test. ∗p < 0.05, ∗∗p < 0.01. To quantify these spatial relationships, we calculated the nearest distance from each B_Mal+ spot to surrounding CD8T_Tex+ spots as an example. Distances were shorter in LME2, and similar proximity was observed for B_Mal relative to CD4T_Tfh, FDC_CR2, and Mac_SPARC (Figure 5G). Compared with other myeloid subsets, Neu_G0S2 and Mac_MARCO were positioned closer to vascular endothelial cells, consistent with their roles in angiogenesis (Figure S13C). We further evaluated the relationship between ligand expression and intercellular distance. Using CD4_Tfh as an example, we calculated the nearest distance from each spot to CD4T_Tfh+ spots and correlated this with ligand expression within the spot. In LME2, CD40LG showed weaker distance correlations with CD4_Tfh, Mac_SPARC, and FDC_CR2, indicating higher ligand expression in regions proximal to these cell subpopulations. Similarly, in LME1, VEGFA exhibited weaker distance correlations with Neu_G0S2, Mac_MARCO, Endo_venous, and Endo_arterial (Figure 5H). These spatial associations were further supported by MISTy, which identified strong intra-spot importance between CD4T_Tfh and CD40LG in LME2, and highlighted angiogenic signaling in both intra-spot and immediate neighborhood contexts in LME1 (Figure S13D). To validate spatial relationships among cell populations across LME subtypes, we performed multiplex immunofluorescence (mIF) staining on 15 samples, with 5 representative cases from each subtype. Given the numerical predominance of malignant B cells in the LME of MALT lymphoma, PAX5+ cells were considered malignant B cells (Figure 5I). In LME2, malignant B cells tended to localize closer to FOXP3+ CD4+ T cells and PD-1+ CD8+ T cells (Figure 5J). Based on snRNA-seq-derived marker definitions and prior reports, CD11b+ CD15+ cells were annotated as Neu_G0S2 and CD206+ MARCO+ cells as Mac_MARCO (Figure 5K).37 Quantitative analysis revealed that CD31+ vascular endothelial cells were most abundant in LME1, where both Neu_G0S2 and Mac_MARCO were positioned in closer proximity to endothelial cells, consistent with spatial transcriptomic findings (Figure 5L). Moreover, the frequencies of VEGFA+ Neu_G0S2 and Mac_MARCO were higher in LME1 (Figure S13E), and VEGFA+ subsets of these populations localized closer to endothelial cells than their VEGFA− counterparts, in line with the pro-angiogenic role of VEGFA (Figure S13F). Collectively, our results support the classification of LME subtypes and underscore the role of distinct cellular interactions in shaping microenvironmental heterogeneity in MALT lymphoma (Figure S14). Discussion Chronic antigenic stimulation from persistent infection or autoimmunity elicits sustained immune and inflammatory responses in MALT lymphoma, creating conditions that enable the malignant transformation of reactive B cells and support their growth and survival, pointing to the microenvironment as a critical driver of this process.1,2 In this study, we analyzed a large, multi-organ MALT lymphoma cohort using bulk, single-nucleus, and spatial transcriptomics and validated key findings with mIF. This integrative approach established an LME-based classification framework with three subtypes (mesenchymal, inflammatory, and depleted) that captures biological heterogeneity and provides insights for clinical stratification. Our study identified distinct LME subtypes varying from cellular composition and functional states to spatial organization and clinical characteristics. The mesenchymal subtype reflects a microenvironmental state shaped by chronic inflammation-driven angiogenesis. Neutrophils and MARCO+ macrophages act as proximate drivers of angiogenesis, colocalizing with endothelial niches and exhibiting stronger pro-angiogenic signaling in the mesenchymal subtype. The mesenchymal subtype predominates in gastric MALT lymphoma, suggesting that its emergence may be driven by chronic HP infection in the stomach, which promotes neutrophil accumulation and, in turn, enhances vascular endothelial angiogenesis.38 In gastric MALT lymphoma, eradication of HP with specific antibiotics remains the standard therapy.39 By removing the infectious trigger, this approach reduces neutrophil and other inflammation-associated cell populations, thereby disrupting the microenvironmental support on which malignant B cells depend. At more advanced stages, tumors may lose their dependency on chronic infection through the acquisition of additional genetic alterations.38 In such cases, our LME-based classification suggests that anti-angiogenic strategies could be particularly relevant for the mesenchymal subtype. In contrast, the inflammatory subtype represents an immune-enriched state in which malignant B cells are sustained by dense immune crosstalk rather than vascular support. Coordinated signaling through CXCL, CD40, and ICAM pathways establishes reciprocal recruitment and activation loops. These interactions provide potent paracrine stimulation to malignant B cells, particularly through CD40 signaling. Previous studies have shown that blockade of the CD40-CD40L axis with monoclonal antibodies reduces the abundance of CD8+ T cells, Treg cells, and DCs in the LME and simultaneously suppresses MALT lymphoma lesions.40,41 In addition, BAFF signaling is elevated in the inflammatory subtype, with a greater contribution from autocrine secretion by malignant B cells. This observation is consistent with prior reports linking autocrine BAFF signaling to infection-independent growth of MALT lymphoma.42 Together, these findings suggest that the inflammatory subtype may be particularly susceptible to interventions targeting CD40 and BAFF pathways, both of which converge on B-cell activation. The depleted subtype was characterized by attenuated microenvironmental signatures coupled with high proliferative signatures in malignant B cells. Owing to the paucity of microenvironmental features, this subtype appears less responsive to therapies targeting the LME but more susceptible to agents directed against malignant B cell proliferation, such as microtubule inhibitors. A comparable LME subtype with reduced microenvironmental activity has also been reported in diffuse large B cell lymphoma and peripheral T cell lymphoma, and its emergence may be associated with distinct DNA methylation patterns.18,43 In conclusion, this study demonstrates that MALT lymphoma is governed by distinct microenvironmental programs that shape tumor behavior and therapeutic sensitivity. The LME-based classification framework, therefore, provides a biologically grounded rationale for treatment stratification, with potential implications for precision management. Limitations of the study First, owing to the long-standing lack of suitable cellular and animal models for MALT lymphoma, we were unable to experimentally validate the therapeutic vulnerabilities predicted for each LME subtype. Consequently, these findings should be interpreted with caution, particularly with respect to their translational relevance. Second, the retrospective nature of our cohort limited sample availability and precluded the incorporation of genomic and epigenomic analyses, which could provide deeper mechanistic insight into the processes underlying the development of distinct LME subtypes. In addition, certain anatomical sites, such as the thyroid, were under-represented or absent in our cohort, which may constrain the interpretation of microenvironmental heterogeneity across different tissues. Finally, although we propose a potential link between HP infection, neutrophil accumulation, and angiogenesis in gastric MALT lymphoma, this hypothesis remains speculative. Validation using dedicated experimental systems, such as HP co-culture models, will be important to substantiate this proposed mechanism. Further research is necessary to guide precise medicine strategies based on different LME subtypes. Resource availability Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Ailin Zhao (zhaoailin@wchscu.cn). Materials availability This study did not generate new unique reagents. Data and code availability •Raw sequencing data have been deposited in the Genome Sequence Archive for Human at the National Genomics Data Center under accession number HRA013147 (https://ngdc.cncb.ac.cn/gsa-human/browse/HRA013147). Processed sequencing data are available at Zenodo (https://zenodo.org/records/19388891). The publicly available bulk RNA-seq data re-analyzed in this study were obtained through the Gene Expression Omnibus (GEO) database under accession numbers GSE171059, GSE199517, GSE218618, GSE39577, and GSE48047. Raw sequencing data have been deposited in the Genome Sequence Archive for Human at the National Genomics Data Center under accession number HRA013147 (https://ngdc.cncb.ac.cn/gsa-human/browse/HRA013147). Processed sequencing data are available at Zenodo (https://zenodo.org/records/19388891). The publicly available bulk RNA-seq data re-analyzed in this study were obtained through the Gene Expression Omnibus (GEO) database under accession numbers GSE171059, GSE199517, GSE218618, GSE39577, and GSE48047. •Scripts used for the analyses presented in this study are available at GitHub (https://github.com/k2med/MALToma_atlas). Scripts used for the analyses presented in this study are available at GitHub (https://github.com/k2med/MALToma_atlas). •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This work was supported by National Natural Science Foundation of China (82370192 and U24A20680); Natural Science Foundation of Sichuan Province (2026NSFSC0677, 2026NSFSC1696, and 2026NSFSC0651); “Qimingxing” Research Fund for Young Talents (HXQMX0101, HXQMX0141, and HXQMX0189), West China Hospital, Sichuan University; and 1.3.5 Project of High Altitude Medicine (GYYX24003), West China Hospital, Sichuan University. The authors thank Li Li, Fei Chen, and Chunjun Bao from the Institute of Clinical Pathology, West China Hospital, Sichuan University, for processing histological staining. Author contributions Conceptualization, Y.W., T.N., and A.Z.; data curation, K.K., X.S., Y.Y., W.L., J.Y., Y.T., and B.X.; formal analysis, K.K., Y.W., and X.S.; funding acquisition, Y.W., T.N., K.K., and A.Z.; methodology, K.K., Y.W., and X.S.; project administration, K.K., Y.W., and X.S.; supervision, Y.W., T.N., and A.Z.; visualization, K.K., Y.W., X.S., Y.Y., W.L., J.Y., Y.T., and B.X.; writing – original draft, K.K., Y.W., and X.S.; writing – review and editing, all authors. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesAnti-human CD8α antibodyCell Signaling TechnologyCat# 70306; RRID:AB_2799781Anti-human CD4 antibodyCell Signaling TechnologyCat# 48274; RRID:AB_3076699Anti-human PD-1 antibodyCell Signaling TechnologyCat# 86163; RRID:AB_2728833Anti-human CD3ε antibodyCell Signaling TechnologyCat# 85061; RRID:AB_2721019Anti-human CD206 antibodyCell Signaling TechnologyCat# 24595; RRID:AB_2892682Anti-human VEGFR2 antibodyCell Signaling TechnologyCat# 2479; RRID:AB_2212507Anti-human BTK antibodyCell Signaling TechnologyCat# 8547; RRID:AB_10950506Anti-human Ki-67 antibodyCell Signaling TechnologyCat# 9027; RRID:AB_2636984Anti-human VEGFA antibodyAbcamCat# ab1316; RRID:AB_299738Anti-helicobacter pylori antibodyAbcamCat# ab231433Anti-human CD11b antibodyHUABIOCat# HA722075; RRID:AB_3096756Anti-human MARCO antibodyHUABIOCat# HA722114Anti-human FOXP3 antibodyBiossCat# bsm-60760rAnti-human PAX5 antibodyBiossCat# bsm-60221rAnti-human CD31 antibodyMaixinCat# MAB-0720Anti-human CD15 antibodyMaixinCat# MAB-0779OmniMAP anti-mouse HRPVentanaCat# 760–4310; RRID:AB_2885182OmniMAP anti-rabbit HRPVentanaCat# 760–4311; RRID:AB_2811043Biological samplesHuman MALT lymphoma samplesThis paperN/ACritical commercial assaysSMARTer Stranded Total RNA-Seq Kit v2 - Pico Input MammalianTakaraCat# 634411NanoHema Panel v2.0NanodigmbioCat# 1001722Bioyou® Nuclei lsolation Kit for FFPE TissueShanghai Biotechnology CorporationCat# 52009-10Chromium Next GEM Single Cell Fixed RNA Sample Preparation Kit10x GenomicsCat# 1000414Visium Human Transcriptome Probe Set v2.010x GenomicsCat# 1000466Deposited dataRaw sequencing dataThis paperGSA-Human (https://ngdc.cncb.ac.cn/gsa-human/browse/HRA013147)Processed sequencing dataThis paperZenodo (https://zenodo.org/records/19388891)Software and algorithmsHISAT2 (v2.0.4)Githubhttps://github.com/DaehwanKimLab/hisat2Stringtie (v1.3.0)Githubhttps://github.com/gpertea/stringtiesva (v3.42.0)Bioconductorhttps://bioconductor.org/packages/release/bioc/html/sva.htmllimma (v3.50.0)Bioconductorhttps://bioconductor.org/packages/release/bioc/html/limma.htmlclusterProfiler (v4.2.2)Bioconductorhttps://bioconductor.org/packages/release/bioc/html/clusterProfiler.htmlGSVA (v1.42.0)Bioconductorhttps://bioconductor.org/packages/release/bioc/html/GSVA.htmlNMF (v0.23.0)CRANhttps://cran.r-project.org/web/packages/NMF/index.htmlComplexHeatmap (v2.10.0)Bioconductorhttps://www.bioconductor.org/packages/devel/bioc/html/ComplexHeatmap.htmlCytoSig (v0.1)Githubhttps://github.com/data2intelligence/CytoSigArriba (v2.4.0)Githubhttps://github.com/suhrig/arribaFusionCatcher (v1.33)Githubhttps://github.com/ndaniel/fusioncatcherGATK-Mutect2 (v4.6.2)Githubhttps://github.com/broadinstitute/gatksurvival (v3.5-7)CRANhttps://cran.r-project.org/web/packages/survival/index.htmlsurvminer (v0.4.9)CRANhttps://cran.r-project.org/web/packages/survminer/index.htmlrms (v6.2)CRANhttps://cran.r-project.org/web/packages/rms/index.htmlCellRanger (v7.0.1)10x Genomicshttps://10xgenomics.comCellBender (v0.3.1)Githubhttps://github.com/broadinstitute/CellBenderSeurat (v4.3.0)Githubhttps://github.com/satijalab/seuratharmony (v0.1.0)Githubhttps://github.com/immunogenomics/harmonyinferCNV (v1.10.1)Bioconductorhttps://www.bioconductor.org/packages/release/bioc/html/infercnv.htmlNebulosa (v1.0.1)Bioconductorhttps://www.bioconductor.org/packages/release/bioc/html/Nebulosa.htmlAUCell (v1.16.0)Bioconductorhttps://bioconductor.org/packages/release/bioc/html/AUCell.htmlMAST (v1.20.0)Bioconductorhttps://bioconductor.org/packages/release/bioc/html/MAST.htmlCytoTRACE2 (v1.0.0)Githubhttps://github.com/digitalcytometry/cytotrace2beyondcell (v2.2.2)Githubhttps://github.com/cnio-bu/beyondcellscRANK (v1.0.0)Githubhttps://github.com/ZJUFanLab/scRankpySCENIC (v0.12.1)Githubhttps://github.com/aertslab/pySCENICnichenetr (v2.0.5)Githubhttps://github.com/saeyslab/nichenetrPROGENy (v1.16.0)Bioconductorhttps://www.bioconductor.org/packages/release/bioc/html/progeny.htmlCellChat (v2.1.2)Githubhttps://github.com/jinworks/CellChatspaceranger (v2.0.1)10x Genomicshttps://10xgenomics.comcell2location (v0.1.4)Githubhttps://github.com/BayraktarLab/cell2locationmistyR (v1.2.1)Bioconductorhttps://www.bioconductor.org/packages/release/bioc/html/mistyR.html Experimental models and study participant details Study design and participants The retrospective collection of FFPE samples in the WCH cohort was approved by the Ethics Committee on Biomedical Research of West China Hospital of Sichuan University. In total, 89 samples were obtained from 82 patients who were pathologically diagnosed with MALT lymphoma and had undergone surgical resection or biopsy between January 2020 and June 2023. Patients were staged with Ann Arbor staging classification for lymphoma. All clinical samples were collected with the approval of the ethics committee (No. 2023–1001) and in accordance with the Declaration of Helsinki, after obtaining written informed consent from the patients. Details of the characteristics are summarized in Table S1. We incorporated two publicly available MALT lymphoma cohorts with normal tissues. The cohort reported by Shi et al. consisted of two datasets, with lymphoma and normal samples deposited separately in the GEO database under accession numbers GSE171059 and GSE199517.14 The cohort reported by Wolf et al. included both lymphoma and normal samples within a single dataset GSE218618.15 In addition, two independent transcriptomic microarray cohorts of gastric MALT lymphoma (GSE39577 and GSE48047) were included.34,35 Owing to differences in platforms, these datasets were not integrated with the high-throughput sequencing cohorts but were instead analyzed separately to independently evaluate the association between HP infection–related signatures and neutrophil-associated gene signatures. Method details Bulk transcriptomics Library preparation and sequencing For each sample, 10 ng of total RNA was used to construct library using SMARTer Stranded Total RNA-Seq Kit v2 - Pico Input Mammalian kit which incorporates a technology that enables removal of ribosomal cDNA following cDNA synthesis (Takara Bio, Europe). The libraries were sequenced on NovaSeq 6000 sequencing system (Illumina), and 150 base pair (bp) paired-end reads were generated. The reads were aligned to the human reference genome (GRCh38) using HISAT2 (v2.0.4). The read counts of each gene were obtained using Stringtie (v1.3.0). Data integration and dimensionality reduction Bulk RNA-seq data from the WCH cohort and the two publicly available cohorts were first converted from raw read counts to transcripts per million (TPM). Batch effects across datasets were corrected using the ComBat function implemented in the R package sva (v3.42.0),44 with batches defined as each dataset (i.e., the WCH cohort dataset and the three datasets from the two public cohorts). Negative values in the integrated matrix were set to zero, followed by a binary log transformation using log1p. PCA was then performed with the prcomp function in the R package stats (v4.1.2) to evaluate global expression patterns and potential group separation. Differential gene expression and enrichment analysis After stratifying samples by lymphoma versus normal tissues or by LME subtype, differential expression analysis was performed using R package limma (v3.50.0).45 Genes were ranked by log2 fold change (log2FC) values from highest to lowest, and GSEA was conducted using R package clusterProfiler (v4.2.2).46 Genes with log2FC > 0 were defined as upregulated, whereas genes with log2FC > 2 and an adjusted p value <0.05 were considered significantly upregulated. Evaluation and clustering of microenvironmental features Microenvironmental features were assessed based on signature gene lists from snRNA-seq in this study. To avoid confounding effects from malignant B cells, genes associated with B-cell subpopulations were excluded from signature construction, as their expression levels are heavily influenced by tumor cell abundance and therefore do not reliably reflect the microenvironment. The R package GSVA (v1.42.0) was applied to calculate signature scores for each MALT lymphoma sample.47 Since negative values were present in the GSVA-derived scores, min–max normalization was performed prior to clustering. The R package NMF (v0.23.0) was then used for unsupervised clustering,48 with the optimal rank determined at the inflection point of the residual sum of squares (RSS) curve.49 For microenvironmental features derived from previously reported gene sets,18 GSVA was applied in the same manner. Heatmaps displaying LME subtypes alongside microenvironmental and clinical features were generated using the R package ComplexHeatmap (v2.10.0).50 Cytokine response analysis Based on bulk RNA-seq data from MALT lymphoma samples, cytokine signaling activities were inferred using CytoSig (v0.1).51 CytoSig provides a curated database of cytokine-regulated target genes and a predictive framework for estimating cytokine signaling activity from transcriptomic profiles. The results with prior Z score transformation were used for subsequent visualization and statistical analyses. Target sequencing and genomic alterations Targeted RNA and DNA sequencing was performed using the NanoHema panel (v2.0), which employs hybrid-capture technology to enrich genes recurrently involved in hematologic malignancies. Targeted RNA sequencing was conducted in all 89 samples, whereas targeted DNA sequencing was performed in a subset of 31 samples for which sufficient material was available. For RNA sequencing data, BIRC3/MALT1 fusion events were identified using Arriba (v2.4.0) and FusionCatcher (v1.33).52,53 Given the limited sensitivity of individual fusion-calling algorithms, results from both tools were integrated and further supported by available pathological validation using fluorescence in situ hybridization (FISH) to determine the fusion status of each sample. For DNA sequencing data, somatic mutations were identified using GATK-Mutect2 (v4.6.2). In this study, particular attention was given to loss-of-function mutations in TNFAIP3. Survival analysis In the WCH cohort, 67 treatment-naïve individuals were included, and the corresponding FFPE specimens were obtained at initial diagnosis. PFS data were collected during follow-up. Patients received heterogeneous treatment modalities depending on disease site and clinical context, including surgery, radiotherapy, chemotherapy, targeted therapy, and anti-infective treatment. PFS was defined as the time from the date of treatment initiation to disease recurrence, progression, or death from any cause. Patients without an event were censored at the date of last follow-up. Additional 21 patients with available PFS data from the cohort reported by Shi et al. were included, resulting in a total of 88 individuals for survival analysis. Kaplan–Meier survival curves were generated using the survfit function in the R package survival (v3.5-7) and visualized with the R package survminer (v0.4.9), with statistical significance assessed by the log rank test. Univariate and multivariate Cox regression were performed using the coxph function implemented in the survival package. The nomogram was developed using the R package rms (v6.2). Single-nucleic transcriptomics Library preparation and sequencing For the Chromium Single Cell Gene Expression Flex workflow, 50 μm FFPE curls were collected from FFPE blocks, and then were dissociated with the Bioyou Nuclei Isolation Kit for FFPE Tissue (Shanghai Biotechnology Corporation, China). FFPE specimens with a tissue cross-sectional area exceeding 0.3 × 0.3 cm were selected. RNA was extracted for quality assessment, and only samples with DV200 values, which are the percentage of RNA fragments >200 nucleotides, greater than 50% were subjected to nuclei isolation. Samples yielding more than 5 × 105 nuclei with nuclear membrane integrity exceeding 80% were retained for subsequent library preparation and sequencing. Sequencing libraries were generated following the user guide for fixed RNA profiling (CG000477). Sequencing libraries were quantified using a High Sensitivity DNA Chip (Agilent) on a Bioanalyzer 2100 and the Qubit High Sensitivity DNA Assay (Thermo Fisher Scientific). Libraries were sequenced on NovaSeq 6000 sequencing system (Illumina), and 150 bp paired-end reads were generated. Reads from snRNA-seq were aligned and quantified against the human reference genome (GRCh38) using CellRanger (v7.0.1). Quality control and data processing Unique molecular identifier (UMI)–based snRNA-seq count matrices were first processed with the Python package CellBender (v0.3.1) to remove ambient RNA molecules and artifacts arising from random barcode swapping.54 The filtered count matrices were then imported into R and converted into Seurat (v4.3.0) objects. Nuclei with 100–2,500 detected genes and <10% mitochondrial transcript content were retained for downstream analysis. Data preprocessing followed the standard Seurat workflow: expression matrices were normalized, highly variable genes were selected, and the data were scaled with regression on mitochondrial ratio, UMI counts, and gene counts. PCA was then performed on the scaled expression matrix of highly variable genes, followed by batch correction using R package harmony (v0.1.0).55 According to the elbow plot, the top 20 principal components were retained for downstream analysis. Clustering was subsequently carried out using the FindNeighbors and FindClusters functions in Seurat, which apply a graph-based approach to construct a shared nearest neighbor graph followed by community detection with the Louvain algorithm. After the initial clustering, subclusters were preliminarily assigned to major cell lineages based on the expression of canonical marker genes. For each lineage, highly variable features were re-identified, and subclustering was performed following the same procedure as described above. Subclusters were annotated according to their marker gene expression profiles. Subclusters that exhibited simultaneous expression of distinct lineage-specific markers were classified as doublets and excluded from subsequent analyses. For B cells, a tailored approach was applied given that most were malignant and exhibited substantial heterogeneity. Highly variable features were re-identified for B cells from each individual sample, and the SelectIntegrationFeatures function was used to rank features by the number of datasets in which they were variable, with ties broken by the median variable feature rank across datasets. The top 1,500 features were retained for downstream analyses. UMAP visualization was generated using the RunUMAP function from Harmony-corrected PCA embeddings to project the cells into a two-dimensional space. Copy number variation inference For B cells, CNVs were inferred using the R package inferCNV (v1.10.1),56 which estimates CNV profiles based on gene expression intensities across the genome. For each sample, B cells were analyzed using T cells from the same sample as the reference population. This approach was applied to help distinguish malignant from normal B cells. For each cell, a CNV score was computed as the mean squared deviation of the estimated CNV value from the diploid state across all genes. Visualization was performed using the R package Nebulosa (v1.0.1),57 which applies a weighted kernel density estimation method to overlay two-dimensional density plots onto UMAP embeddings. Gene signature scoring Gene sets were obtained from the MSigDB database, including Hallmark, GO, KEGG, and Reactome collections, as well as from previously published literature and curated functional gene sets in this study (Table S4).58,59,60 Signature scores for each cell were calculated using the R package AUCell (v1.16.0).61 For visualization, density plots were generated with the R package Nebulosa (v1.0.1),57 and heatmaps were created with the R package ComplexHeatmap (v2.10.0) after Z score transformation of the signature scores.50 Assessment of cell-type enrichment To characterize the distribution of cell subpopulations across LME subtypes, ORs were calculated to quantify enrichment. For each combination of cell subpopulation i and LME subtype j, a 2 by 2 contingency table was constructed, containing: (1) the number of cells from subpopulation i in subtype j; (2) the number of cells from subpopulation i in other subtypes; (3) the number of non-i cells in subtype j; and (4) the number of non-i cells in other subtypes. Fisher’s exact test was then applied to each contingency table to estimate the OR and the corresponding p value. Multiple testing correction was performed using the Benjamini–Hochberg method implemented in the p.adjust function of the R package stats (v4.1.2). For visualization, ORs were log-transformed (lnOR), such that lnOR >0 indicated enrichment and lnOR <0 indicated depletion. To evaluate the distribution of cell subpopulations across tissues, Shannon entropy was calculated from their proportional abundances across different tissue sites. For each subpopulation, entropy was computed as: where pi represents the proportion of that subpopulation in a given tissue. Higher entropy values indicated a more even distribution across tissues, whereas lower values reflected stronger tissue specificity. In this study, an entropy threshold of 0.5 was selected based on the observed distribution of entropy values across all subpopulations. Subpopulations with entropy values below this threshold (n = 5) exhibited strong enrichment (>80%) in a single anatomical site, and their transcriptional profiles were biologically consistent with normal tissue–derived cell types. Based on the concordance among entropy distribution, tissue specificity, and biological annotation, this threshold was applied to exclude subpopulations likely representing normal tissue contamination from downstream analyses. Differential gene expression Based on annotated subpopulations, marker genes were identified using the R package MAST (v1.20.0).62 For each subpopulation, candidate marker genes were filtered according to the following criteria: (1) a minimum 25% difference in the fraction of cells expressing the gene inside versus outside the subpopulation; (2) a log2FC greater than 0.25 compared with other clusters; and (3) an adjusted p value <0.05. For each subpopulation, differentially expressed genes (DEGs) were identified, and when more than 100 DEGs were detected, the top 100 DEGs ranked by log2FC were selected; these genes were then used as microenvironmental features to evaluate bulk RNA-seq samples. Meta-programs of transcriptional heterogeneity To capture transcriptional heterogeneity, we applied NMF to malignant B cells from each sample, following previously described approaches.19 For each sample, the scaled expression matrix of highly variable genes was used as input after setting negative values to zero. NMF was performed with the R package NMF (v0.23.0)48 using ranks from 4 to 9, yielding a total of 39 programs per sample. Each program was summarized by the top 50 genes ranked by NMF coefficients. To identify robust transcriptional programs, NMF results were compared across different ranks and across samples. Programs were retained if they recurred within a sample (≥70% gene overlap between programs derived from different ranks) and showed similarity across samples (≥20% overlap of the top 50 genes). Within each sample, redundant programs (≥20% overlap with a higher-ranking program) were removed. Robust NMF programs were then clustered by similarity, and MPs were defined by the most recurrent genes, limited to 50 genes per MP. An MP suspected to reflect doublets, based on high similarity to T cell expression profiles, was removed. Enrichment analysis For each MP, the top 50 genes were subjected to over-representation analysis using the R package clusterProfiler (v4.2.2).46 To assess differences between malignant and normal B cells, following differential expression analysis, genes were ranked by log2FC values from highest to lowest, and GSEA was performed with the R package clusterProfiler (v4.2.2).46 Cellular potency analysis Cellular potency scores were predicted using the R package CytoTRACE2 (v1.0.0),63 which estimates single-cell developmental potential. The method provides both discrete potency categories and continuous scores (range 0–1), smooths predicted values using diffusion, and rescales them through a combination of binning and k-nearest neighbor smoothing. In this framework, CytoTRACE2 scores closer to 1 indicate a less differentiated state with higher developmental potential, whereas scores closer to 0 reflect a more differentiated state. Identification of therapeutic heterogeneity Drug sensitivity was inferred at single-cell resolution for malignant B cells using the R packages beyondcell (v2.2.2) and scRANK (v1.0.0).64,65 Beyondcell was applied to predict the relative sensitivity of individual cells to a broad range of therapeutic agents based on previously curated drug response gene signatures. In parallel, drug target–based sensitivity prediction was performed using scRank, which infers drug-responsive cell types from untreated single-cell RNA-seq data by modeling in silico drug perturbations through target-perturbed gene regulatory networks. Cytokine response analysis To investigate the role of cytokines in driving transcriptional differences among subpopulations, cytokine responses were assessed using IREA (https://www.immune-dictionary.org) with a hypergeometric test.66 For macrophage and DC subpopulations, differential expression analysis was performed using the following criteria: a minimum 25% difference in the fraction of cells expressing the gene, log2FC > 0.25, and an adjusted p value <0.05. DEGs were then converted between mouse and human homologs using the National Center for Biotechnology Information (NCBI) HomoloGene database (release 68) and subsequently used for IREA to evaluate cytokine activation. Gene regulatory network analysis Transcription factors activated in each subpopulation were identified using the Python package pySCENIC (v0.12.1).67 The references used for cistarget was downloaded from the cisTarget database (https://resources.aertslab.org/cistarget). The results with prior Z score transformation were used for subsequent visualization. Ligand-receptor interaction analysis by NicheNet To predict ligands potentially driving gene expression changes in specific cell subpopulations, ligand–receptor interactions were assessed using the R package nichenetr (v2.0.5).68 For subpopulations within fibroblasts or vascular endothelial cells, differential expression analysis was performed using a criterion that required a log2FC > 0.25 and an adjusted p value <0.05. DEGs were used for downstream ligand prediction analysis. The top five ligands and corresponding receptors were subsequently displayed. Pathway activity prediction by PROGENy To infer the activities of signaling pathways, such as TNF-α, NF-κB, and TGF-β, in fibroblasts, the R package PROGENy (v1.16.0) was applied.69 For each pathway, the top 500 genes were selected according to significance, while other parameters were kept at default settings. The results with prior Z score transformation were used for subsequent visualization. Identification of cellular modules To investigate potential multi-cellular co-enrichment patterns in MALT lymphoma, the proportions of individual subpopulations within major lineages (T cells, myeloid cells, and stromal cells) were calculated for each sample. Pairwise correlations between subpopulations across samples were assessed using Spearman’s correlation with the cor function in the R package stats (v4.1.2). The resulting correlation matrix was clustered with the R package ComplexHeatmap (v2.10.0)50 using hierarchical clustering with complete linkage. Distinct CMs were identified based on the clustering results. Considering intercellular modules require sufficient representation of multiple cell types to capture meaningful interactions, we retained CM1 containing 10 subpopulations and CM2 containing 8 subpopulations for downstream analyses. For each sample, the proportions of subpopulations within a given module were summed and compared across different LME subtypes. Cell–cell interaction analysis Cell–cell communication was inferred, analyzed, and visualized using the R package CellChat (v2.1.2) applied to single-nucleus RNA-seq data.70 The computeCommunProb function was used to calculate communication probabilities between cell types within each LME subtype. To reduce dropout effects commonly observed in single-cell data with shallow sequencing depth, particularly the zero expression of ligand or receptor subunits, smoothed projected data were used for the analysis by adjusting the parameter. In addition, the proportion of each cell group across all sequenced cells was considered to minimize artifacts related to population size. Prior to generating heatmaps and line charts of relative interaction strength, values were transformed using the reciprocal of the natural logarithm. Spatial transcriptomics Library preparation and sequencing FFPE samples that passed RNA quality control (DV200 > 30%) were used for spatial transcriptomic library construction and sequencing. Tissue sections of 5 μm thickness were mounted onto Visium Spatial Gene Expression Slides (10x Genomics). H&E staining was performed using Mayer’s hematoxylin (Millipore Sigma), bluing reagent (Dako, Agilent), and alcoholic eosin (Millipore Sigma). Stained slides were scanned, decrosslinked, and incubated with the human whole-transcriptome probe panel, followed by processing with CytAssist (10x Genomics) for probe hybridization and ligation. Libraries were prepared according to the manufacturer’s protocol and sequenced on an Illumina NovaSeq 6000 platform, generating 150-bp paired-end reads. The reads were aligned to the Visium Human Transcriptome Probe Set (v2.0), and the expression matrix was extracted using the spaceranger (v2.0.1) pipeline. Mapping cell types in spatial transcriptomics To comprehensively map cellular architecture by integrating single-nucleus and spatial transcriptomics, the Python package cell2location (v0.1.4)71 was used to resolve fine-grained cell types within coarse-grained spatial transcriptomic data. Annotated snRNA-seq data were used as a reference to derive cell-type–specific expression signatures, which were subsequently projected onto spatial transcriptomic profiles to estimate the abundance of each cell type in individual spots. Quality control and gene filtering were performed prior to mapping. Genes were retained if they were detected (mRNA count >0) in more than 5% of cells, or if detected in a smaller fraction of cells (0.03% < cell proportion ≤5%) but with a mean expression greater than 1.12 across non-zero cells. Mitochondrial-encoded genes were excluded before mapping, as their expression is more strongly associated with technical variation rather than true mitochondrial abundance. For spatial transcriptomic data, only genes retained after snRNA-seq filtering were used to ensure consistency between datasets. Among the nine spatial transcriptomic samples, seven had matched single-nucleus RNA-seq data generated from the same tissue, ensuring the reliability of this mapping. Spatial colocalization of cell types Based on the mapping results obtained from the integration of single-nucleus and spatial transcriptomics, spatial colocalization of cell types was assessed using the run_colocation function in the Python package cell2location (v0.1.4),71 which extracts spatial factors to capture patterns of colocalization across cell types. In addition, we applied the R package mistyR (v1.2.1).72 For MISTy analysis, intra-spot was defined as the presence within the same spot, whereas the immediate neighborhood was defined as the surrounding ring of adjacent spots, representing the closest spatial context outside direct colocalization. After obtaining the predictor–target importance matrix using the run_misty function, the plot_interaction_communities function was applied to generate network community plots. Spatial proximity and ligand dynamics To quantitatively assess the spatial distribution of cell types and the variation of key ligand expression as a function of distance, we performed distance-based analyses on spatial transcriptomic data. For each tissue section, cell-type proportions per spot were obtained from cell2location. Using a two-class k-means clustering approach, spots were divided into two clusters for each cell type, and the midpoint between the cluster means was used as a threshold to classify spots as positive or negative for that cell type. Pixel coordinates from the spatial transcriptomic data were converted into micrometer coordinates to construct a two-dimensional matrix, from which pairwise Euclidean distances between spots were calculated. If a spot was classified as positive for two cell types simultaneously, the nearest distance between these two cell types at that spot was defined as zero. Based on these results, we derived the distribution of nearest distances between pairs of cell types and evaluated how ligand expression levels varied as a function of distance from specific cell types. In addition, the R package mistyR (v1.2.1)72 was applied to evaluate the dynamic associations between key ligands or pathway activities and cell-type proportions. VEGF signaling at each spot was quantified using the R package PROGENy (v1.16.0).69 Using an approach similar to the colocalization analysis described above, we further assessed how ligand expression and pathway activity varied with cell-type proportions within intra-spot and immediate-neighborhood contexts. Multiplex immunofluorescence Slides from FFPE samples were used and stained with Opal 6-Plex Manual Detection Kit (Akoya, NEL861001KT) and the following primary antibodies: rabbit anti-PAX5 (Bioss, bsm-60221r, 1:200), mouse anti-CD8α (CST, 70306, 1:200), rabbit anti-CD4 (CST, 48274, 1:100), rabbit anti-PD-1 (CST, 86163, 1:200), rabbit anti-FOXP3 (Bioss, bsm-60760r, 1:200), rabbit anti-CD3ε (CST, 85061, 1:200), mouse anti-VEGFA (Abcam, ab1316, 1:100), rabbit anti-MARCO (HUABIO, HA722114, 1:2000), rabbit anti-CD206 (CST, 24595, 1:200), mouse anti-CD31 (Maixin, MAB-0720, prediluted), rabbit anti-CD11b (HUABIO, HA722075, 1:2000), mouse anti-CD15 (Maixin, MAB-0779, prediluted), mouse anti-helicobacter pylori (Abcam, ab231433, 1:100). The following panels were performed: Panel 1: PAX5-Opal 520, CD8α-Opal 620, CD4-Opal 570, PD-1-Opal 690, FOXP3-Opal 480, CD3ε-Opal 780, DAPI; Panel 2: VEGFA-Opal 520, MARCO-Opal 620, CD206-Opal 570, CD31-Opal 690, CD11b-Opal 480, CD15-Opal 780, DAPI; Panel 3: helicobacter pylori-Opal 690, CD11b-Opal 480, CD15-Opal 780, DAPI. Slides were processed on the staining module for de-paraffinization, epitope retrieval, and endogenous peroxidase quenching. Multiplex staining was performed in sequential rounds, each consisting of nonspecific site blocking (Discovery Inhibitor and Discovery Goat Ig Block, Ventana), incubation with unlabeled primary antibodies, detection with horseradish peroxidase (HRP)-conjugated secondary antibodies (Discovery OmniMap anti-rabbit, anti-goat, and anti-mouse; Ventana), and labeling with Opal reactive fluorophores (Opal 480, 520, 570, 620, 690, 780, Akoya), which covalently tag the primary epitope. Each staining round was followed by a heat-induced denaturation step to remove both primary and secondary antibodies before the next cycle. Nuclear counterstaining was performed with spectral DAPI (Akoya). Images were acquired using the Vectra Polaris Automated Quantitative Pathology Imaging System (Akoya). Image analysis was performed using the HALO platform (v3.0.311.314; Indica Labs). Scanned immunofluorescence images were analyzed for positive cell counts, colocalization, and spatial distance metrics. Regions of interest were delineated along the tissue sections, with hemorrhagic or necrotic areas manually excluded as appropriate. Positive signals were selected for each channel, and consistent thresholding criteria were applied across all slides for the same marker. Nuclei were identified and cytoplasmic boundaries expanded, followed by stepwise quantification of parameters such as positive cell counts and spatial distances under high-power magnification. Immunohistochemistry staining Immunohistochemistry was performed on FFPE tissue sections using the Dako REAL EnVision Detection Kit (Dako, K5007), according to the manufacturer’s instructions. The primary antibodies used included anti-VEGFR2 monoclonal antibody (CST, 2479), anti-BTK monoclonal antibody (CST, 8547), and anti-Ki-67 monoclonal antibody (CST, 9027). Whole-slide images were acquired and quantitatively analyzed using the Aipathwell digital pathology platform (v2.3.1). Immunoreactivity was evaluated using the histochemistry score (H-score), which integrates both the proportion of positive cells and staining intensity to provide a semi-quantitative assessment of protein expression. The H-score was calculated as the sum of the percentage of cells at each staining intensity multiplied by the corresponding intensity score. Quantification and statistical analysis Statistical methods and thresholds are described within Figure legends or STAR Methods. Statistical analyses were performed in R (v4.1.2). Where presented, error bars represent mean ± s.e.m. All statistical tests were two-sided. For analyses involving multiple comparisons, p values were adjusted using the Benjamini–Hochberg method. The Kaplan–Meier method was used to estimate survival functions, and the log rank test was used to assess differences in PFS. Statistical significance was defined as p or adjusted p < 0.05. Published: April 30, 2026 Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.xcrm.2026.102785. Contributor Information Yijun Wu, Email: wuyj01029@wchscu.cn. Ting Niu, Email: niuting@wchscu.cn. Ailin Zhao, Email: zhaoailin@wchscu.cn. Supplemental information Document S1. Figures S1–S14 and Tables S3 and S4 Table S1. Clinical characteristics of samples in this study, related to Figure 1 Table S2. LME signatures based on the snRNA-seq data in this study, related to Figure 1 Document S2. Article plus supplemental information References Associated Data Supplementary Materials Document S1. Figures S1–S14 and Tables S3 and S4 Table S1. Clinical characteristics of samples in this study, related to Figure 1 Table S2. LME signatures based on the snRNA-seq data in this study, related to Figure 1 Document S2. Article plus supplemental information Data Availability Statement •Raw sequencing data have been deposited in the Genome Sequence Archive for Human at the National Genomics Data Center under accession number HRA013147 (https://ngdc.cncb.ac.cn/gsa-human/browse/HRA013147). Processed sequencing data are available at Zenodo (https://zenodo.org/records/19388891). The publicly available bulk RNA-seq data re-analyzed in this study were obtained through the Gene Expression Omnibus (GEO) database under accession numbers GSE171059, GSE199517, GSE218618, GSE39577, and GSE48047. Raw sequencing data have been deposited in the Genome Sequence Archive for Human at the National Genomics Data Center under accession number HRA013147 (https://ngdc.cncb.ac.cn/gsa-human/browse/HRA013147). Processed sequencing data are available at Zenodo (https://zenodo.org/records/19388891). The publicly available bulk RNA-seq data re-analyzed in this study were obtained through the Gene Expression Omnibus (GEO) database under accession numbers GSE171059, GSE199517, GSE218618, GSE39577, and GSE48047. •Scripts used for the analyses presented in this study are available at GitHub (https://github.com/k2med/MALToma_atlas). Scripts used for the analyses presented in this study are available at GitHub (https://github.com/k2med/MALToma_atlas). •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.