A unified spatial & molecular atlas of anatomically restricted large B-cell lymphomas defines stereotyped cell ecosytems

Abstract

Abstract Background Large B-cell lymphomas (LBCLs) are clinically and biologically heterogeneous, yet specific extranodal variants can display strikingly stereotyped features reflecting their anatomical predilections. For instance, LBCLs arising in immune-privileged sites—such as primary central nervous system (PCNSL) or testis (PTL)—commonly harbor NF-κB pathway lesions and an activated B-cell phenotype, correlating with poor outcomes. In contrast, primary bone LBCLs (PBL), usually linked to favorable outcomes, often harbor immune modulatory and epigenetic mutations and a germinal center B-cell phenotype. These patterns suggest that anatomical origin may serve as a meaningful basis for LBCL subtype classification, potentially reflecting local immune responses to each tumor. To clarify these relationships, here we describe a Unified Spatial and Molecular Atlas of Anatomical Restricted LBCLs. By integrating genomic and transcriptomic data with spatial proteomic measurements and clinical outcomes, we provide a comprehensive, multimodal resource. This atlas enables robust evaluation of established classifiers across anatomical subtypes and lays the groundwork for identifying novel, site-specific therapeutic vulnerabilities in LBCL. Method We studied archival diagnostic formalin-fixed paraffin-embedded (FFPE) biopsy specimens from well-annotated LBCL cases at Leiden University Medical Center (n=68): 22 PCNSL, 9 PTL, 21 PBL, and 18 localized nodal LBCL. The same FFPE blocks were sourced for DNA, RNA and spatial proteomics. Genomic profiling was done with custom amplicon sequencing (111 genes) and transcriptome profiling with a custom probeset (800 genes, nCounter). Spatial proteomics employed imaging mass cytometry (41-marker panel, Hyperion). We developed the computational pipeline “Unhuddle” for refined single-cell profiling. Absolute cell abundances were quantified as cells/mm² and stable case level clustering was achieved with hierarchical clustering. Case classifications were performed by existing algorithms that rely on tumor somatic genotype [LymphGen (Wright et al. 2020) and DLBclass (Chapuy et al. 2025)], and microenvironment decomposition from transcriptomic profiles [LymphoMAP (Li et al. 2025), Cluster (Ciavarella et al. 2018), LME (Kotlov et al. 2021), EcoTyper (Steen et al. 2021)]. Results: We identified 3 stereotyped ‘spatial protein’ cell ecosystems across anatomically defined LBCL tumors, including Cytotoxic Predominant, Complex Immune, and Immune Depleted patterns. Nodal DLBCL and PBL were enriched for the Complex Immune ecosystem, which were conspicuously absent from PCNSL and PTL tumors (OR=52, p<0.001). In contrast, PCNSL tumors were enriched for Cytotoxic Predominant ecosystem (OR=6, p=0.01). The Immune Depleted showed no specific enrichment by anatomical site. Each spatial protein ecosystem strongly aligned with distinct corresponding gene expression and genomic signatures. Tumors with the Complex Immune ecosystem had better overall survival outcomes, and were enriched for the LN LymphoMAP (OR=3, p=0.02), ‘Hot’ Ciavarella cluster (OR=7, p<0.001), Mesenchymal LME (OR=10, p<0.001), and C3 DLBclass (OR=4, p<0.01). In contrast, tumors with Cytotoxic Predominant ecosystem were enriched for TEX LymphoMAP (OR=8, p<0.01), Inflammatory LME (OR=8, p<0.01), and DLBclass C4 (OR=4, p<0.05). Tumors with Immune Depleted ecosystem were enriched for FMAC LymphoMAP (OR=5, p=0.01), ‘Cold’ Ciavarella cluster (OR=4, p=0.01), Depleted LME (OR=9, p<0.001), and C5 DLBclass (OR=4, p=0.02). DLBclass C5 is more closely associated with anatomical site than with spatial protein ecosystem (Cramer's V: 0.5 vs 0.3, p=0.04). Conclusions: This atlas provides a direct link between genomic and transcriptomic profiles and their corresponding spatial immune ecosystems in LBCLs. PTL and PCNSL are defined by immune exclusion or cytotoxic exhaustion, explaining their poorer survival. In contrast, complex immune ecosystem confers a survival benefit and is almost exclusively found in bone and nodal LBCL. Tumor genomic profiles align more strongly with anatomical location than with the immune ecosystem, indicating that immune contexture can sometimes override underlying risk conferred by the tumor genome. The spatial proteomic profiling framework we describe can provide a ground truth for immune ecosystems previously inferred from genomic or transcriptomic data and adds crucial biological and clinical nuance for defining site-specific vulnerabilities in LBCL.