Diffuse large B cell lymphoma (DLBCL) is the most common hematological malignancy, and despite decades of research, roughly 40% of patients remain incurable with standard R-CHOP chemoimmunotherapy. The biological complexity underlying this treatment failure is partly understood in terms of cell-of-origin (COO) subtypes: germinal center B cell-like (GCB) DLBCL carries a better prognosis than activated B cell-like (ABC) DLBCL under standard regimens. Further genotypic classification systems such as LymphGen and the C1-C5 Chapuy subtypes have refined this picture, but none of these molecular classifications is sufficient to guide individualized treatment selection at diagnosis.
The missing layer - the tumor microenvironment: Beyond malignant B cell biology, the non-malignant cells surrounding the tumor, collectively the tumor microenvironment (TME), are known to profoundly influence clinical outcomes in DLBCL. Stromal signatures, inflammatory gene expression programs, and immune infiltration patterns have each been linked to survival in prior studies. Yet no study had previously performed a large-scale, systematic, unbiased characterization of the specific cell states within each TME lineage and how those states organize into coherent multicellular communities.
The technological gap: Single-cell RNA sequencing (scRNA-seq) can directly measure gene expression in individual cells, but prior scRNA-seq studies of lymphoma have been limited to small sample sizes, often fewer than a dozen tumors, making it difficult to identify cell states that are statistically robust and clinically relevant across diverse patient populations. Conversely, bulk transcriptomic datasets covering hundreds or thousands of DLBCL tumors exist, but bulk sequencing cannot directly resolve cell-type-specific expression without computational deconvolution.
This paper from Steen, Luca, and colleagues at Stanford (published in Cancer Cell, 2021, with PMC availability in 2022) addresses this gap by developing and applying EcoTyper, an integrative machine learning framework that bridges bulk and single-cell transcriptomics to map tumor cell states and multicellular ecosystems across 1,577 DLBCL tumors.
EcoTyper operates in a two-stage pipeline. First, it uses CIBERSORTx, a previously validated machine learning platform for digital cytometry, to deconvolve bulk RNA-seq or microarray expression data from complex tumor tissues. CIBERSORTx estimates the relative abundance of each of 13 cell types per tumor sample and then computationally purifies cell-type-specific gene expression profiles (GEPs) at single-sample resolution, without any physical cell isolation. Applied to the discovery cohort of 522 fresh/frozen DLBCL tumors (Schmitz et al.), this produced 6,786 digitally purified GEPs representing 13 cell types across each specimen.
Non-negative matrix factorization for state discovery: Within each cell type, EcoTyper applies a specialized non-negative matrix factorization (NMF) framework to decompose the collection of cell-type-specific GEPs into discrete transcriptional programs termed "states." The number of states per cell type is automatically determined by optimizing sensitivity and positive predictive value of cell state discovery while ensuring clustering stability. This unsupervised approach identified 44 distinct cell states ranging from 2 to 5 states per cell type across 13 lineages.
Dataset scale and validation design: The discovery cohort comprised 522 fresh/frozen DLBCL surgical biopsies from the Schmitz et al. RNA-seq dataset. Three additional independent DLBCL cohorts were held out as validation sets: Chapuy et al. (FFPE, microarray), Ennishi et al. (RNA-seq), and Reddy et al. (RNA-seq/FPKM), covering an additional 1,055 tumors. The paper also generated new scRNA-seq data from four de novo DLBCL tumors (three ABC, one GCB), three follicular lymphomas, and one tonsil, and integrated these with five publicly available lymphoid scRNA-seq datasets, assembling a pan-lymphoid single-cell atlas of 173,028 cells from 45 tissue specimens.
Validation of deconvolution accuracy: Before applying EcoTyper to clinical questions, the authors validated CIBERSORTx's deconvolution performance in lymphoid tissues by testing against simulated tissues reconstituted from scRNA-seq data with known cell compositions. Estimated cell fractions were significantly correlated with ground truth expectations (P < 0.05) regardless of whether real DLBCL transcriptomes with known tumor content or simulated mixtures were used.
EcoTyper identified five distinct transcriptional states of malignant B cells (S1 through S5) in DLBCL tumors. These states showed partial but imperfect alignment with COO subtypes. State S1 expressed canonical GCB marker genes including MME (CD10), LMO2, and MYBL1, consistent with a germinal center B cell origin. States S4 and S5 expressed ABC-associated markers (PTPN1 and PIM1, respectively). However, states S2 and S3 showed mixed COO representation, and many tumors classified by their most abundant B cell state were distributed across COO subtypes, underscoring that the five-state classification captures heterogeneity that COO subtyping alone does not resolve.
Reproducibility across validation cohorts: State-specific marker genes and COO enrichment patterns were highly reproducible when EcoTyper was applied to the three independent validation datasets, confirming that the five B cell states are not artifacts of the discovery cohort or its specific RNA-seq platform. The results were also robust to dataset-specific variation in the COO subtype distribution between cohorts.
Single-cell validation and the "wishbone" model: Using scRNA-seq from paired DLBCL tumors with immunoglobulin VDJ sequencing (scVDJ-seq), the authors identified single dominant BCR clonotypes consistent with malignant cells defined by copy number analysis, confirming the identity of tumor B cells at the single-cell level. CytoTRACE, a tool for predicting developmental ordering from transcriptional diversity, placed S1 as the least differentiated state (most stem-like), while S2-S5 were progressively more differentiated. In GCB DLBCL tumors, malignant cells were distributed across all five states including S1. In ABC DLBCL tumors, malignant cells were concentrated in S2-S5 and were largely absent from S1, a pattern confirmed across 1,133 bulk tumors. This led the authors to propose a "wishbone" developmental model: GCB DLBCL arises in S1 (GC-like cells), while ABC DLBCL arises downstream of S1 in pre-memory B cell or pre-plasmablast precursors.
Prognostic significance: All five B cell states stratified overall survival (OS). State S1 was associated with favorable outcomes and state S5 (ABC-enriched) with adverse outcomes in univariable survival models. Critically, several states - including S5 and the COO-mixed S2 - remained independently prognostic after multivariable adjustment for COO, LymphGen, and C1-C5 subtypes, demonstrating that B cell state information adds prognostic value beyond existing molecular classifiers.
Beyond malignant B cells, EcoTyper characterized 39 distinct cell states from the 12 non-malignant lineages that comprise the DLBCL TME: CD8 T cells, CD4 T cells, CD4 regulatory T cells (Tregs), follicular helper T cells (Tfh), NK cells, monocytes/macrophages, dendritic cells, neutrophils, mast cells, plasma cells, endothelial cells, and fibroblasts. Together with the 5 B cell states, this produced a comprehensive 44-state atlas of DLBCL cellular heterogeneity.
Reproducibility of TME states: State recovery was validated against six lymphoid scRNA-seq datasets. Among cell types with more than 200 cells in at least one scRNA-seq dataset, 21 of 24 TME states (88%) were reliably recoverable by reference-guided annotation and permutation testing. When less-represented cell types were assessed using scRNA-seq data from other tumor types, 14 additional TME states were validated, bringing the total recovery rate to 91% (40 of 44 evaluable states including B cells). De novo state rediscovery experiments in two independent DLBCL datasets yielded 82% state reidentification in FFPE specimens and 66% in fresh/frozen follicular lymphoma microarray data.
Prognostic landscape of TME states: The majority of TME states showed significant survival associations in univariable models, and over 40% remained significant after multivariable adjustment for COO, LymphGen, or C1-C5 subtypes. Every TME cell type harbored states with reciprocal survival associations - favorable and adverse states within the same lineage - illustrating that the direction of immune influence on outcomes is state-specific rather than lineage-wide. Eight TME states were associated with favorable outcomes, with monocytes/macrophages S2 (M1-like phenotype) and CD4 T cells S3 (naive) among the most significant after COO adjustment.
COO-specific TME patterns: ABC and GCB DLBCL showed distinct TME compositions. For example, monocytes/macrophages state S3 (M2-like program) was most prevalent in ABC DLBCL, while Tregs state S2 (metabolically active) was most frequent in GCB DLBCL. In ABC DLBCL, tumor-infiltrating T cells across multiple lineages showed widespread overexpression of co-stimulatory and co-inhibitory molecules including LAG3 (a canonical T cell exhaustion marker) and OX40 (TNFRSF4, a costimulatory receptor under investigation as a therapeutic target). GCB DLBCL T cells showed the opposite pattern, being generally deficient in immunomodulatory molecule expression.
The 44 individual cell states do not exist in isolation within tumors; certain states tend to co-occur with each other at consistent relative abundances. EcoTyper identifies these co-association patterns through a clustering procedure that assembles cell states into "communities" that maximize co-occurrence. Applied to the discovery cohort without any prior knowledge of COO or other subtypes, EcoTyper revealed nine distinct multicellular ecosystems termed lymphoma ecotypes (LEs, LE1 through LE9).
Robustness and prognostic significance of ecotypes: All nine LEs were detectable in more than 1,000 independent DLBCL tumors across validation cohorts, confirming their robustness. Notably, 89% of LEs (8 of 9) were significantly prognostic. Moreover, most LEs remained prognostic after multivariable adjustment for COO, LymphGen, and C1-C5 subtypes. In head-to-head comparisons, LEs outperformed direct NMF clustering of bulk tumor expression data for predicting OS, and were preferentially retained in stepwise survival models that also included existing molecular subtypes. LEs were also largely independent of four recently described TME subtypes in DLBCL (Kotlov et al., 2021).
Adverse-prognosis ecotypes (LE1-LE4): LEs 1 and 2 were characterized by cell communities linked to ABC DLBCL. LE3 was distinguished by communities associated with double-hit lymphoma (concurrent MYC and BCL2/BCL6 rearrangements, the highest-risk DLBCL genotype). LE4 was ABC-enriched and B cell-depleted, characterized by immunoreactive T cell states (CD8 S4, CD4 S2, Treg S4) showing widespread expression of co-inhibitory and co-stimulatory molecules, suggestive of a TME in which T cells are activated but ultimately dysfunctional.
Favorable-prognosis ecotypes (LE6-LE9): LE8 was uniquely enriched for GCB lymphoma and its related genotypic lesions (EZB, ST2, C3, and double-hit with favorable biology). LEs 6, 7, and 9 were generally independent of molecular subtypes but elevated in tumors with high stromal content. EcoTyper decomposed the previously described Stromal-1 favorable signature from Lenz et al. (2008) into distinct phenotypic programs: LAMA4 marks tumor-associated endothelial cells in LE7, while POSTN and THBS2 mark cancer-associated fibroblasts in LE9. Strikingly, LE9 was the single most prognostically significant ecotype (P < 10-6), outperforming even GC-like B cells and their associated ecotype LE8.
One of the most compelling applications of EcoTyper demonstrated in this paper is its use as a predictive biomarker discovery tool, illustrated through re-analysis of a negative phase III randomized clinical trial. The REMoDL-B trial enrolled 928 previously untreated DLBCL patients and tested whether adding bortezomib (a proteasome inhibitor) to standard R-CHOP (RB-CHOP arm) would improve progression-free survival versus R-CHOP alone. Based on preclinical rationale that bortezomib inhibits NF-kB signaling preferentially in ABC DLBCL, the trial was designed with COO as the primary stratification variable - but bortezomib failed to improve outcomes regardless of COO subtype.
EcoTyper identifies a responder subgroup invisible to COO: The authors enumerated all 44 DLBCL cell states in microarray gene expression profiles of 928 FFPE tumor specimens from the trial. For each state, they computed univariable OS associations within each treatment arm and developed an "adjusted OS z-score" metric that ranks states by their ability to predict differential therapeutic benefit (greater benefit in RB-CHOP versus R-CHOP). Across all 44 ranked states, CD8 T cell state S1 was the top-ranked predictive biomarker (P = 0.004, Q = 0.04 after multiple testing correction). Its parent ecotype, LE5, was also enriched among top-ranking states at the ecosystem level (P = 0.009, Q = 0.09 by gene set enrichment analysis).
Characterizing CD8 T cell S1 - the CXCR5+ stem-like population: Across the scRNA-seq atlas, CD8 T cell S1 expressed the highest levels of CXCR5 among all CD8 T cell states (top 1.2% of S1-enriched genes by log2 fold change). CXCR5 marks stem-like CD8 T cells with tissue-resident characteristics and robust effector potential in solid tumors, and CXCR5+ CD8 T cells have been reported to reside within B cell follicles with antitumor activity in follicular lymphoma. Known marker genes of CXCR5+ CD8 T cells (from Brummelman et al., 2018) showed significant concordance with the CD8 T cell S1 expression profile but not other CD8 T cell states. Spatial transcriptomics data from a normal human lymph node confirmed that CD8 T cell S1-rich zones localized significantly closer to B cell follicles than other CD8 T cell states.
In silico trial simulation: In a simulated trial where only patients with above-median CD8 T cell S1 abundance were selected for randomization, the RB-CHOP arm showed significantly longer OS and progression-free survival (PFS) than R-CHOP. Subgroup analysis revealed that the predictive effect was restricted to ABC DLBCL: among ABC patients with high CD8 T cell S1, bortezomib produced meaningful survival benefit, a finding invisible in the original trial that treated COO as the only relevant stratifier.
Spatial topology remains incompletely characterized: EcoTyper infers cell states and ecosystems from bulk gene expression data using computational deconvolution and from single-cell suspensions. While the paper uses spatial transcriptomics data from one normal human lymph node to validate that certain co-associated states are spatially proximate, the spatial organization within individual DLBCL tumors - which cell states physically neighbor which others, and how spatial geography influences signaling interactions - is not fully characterized. The authors acknowledge that future studies will be needed to map the spatial topology and interaction networks within LEs.
Mechanistic questions around bortezomib response: While the identification of CXCR5+ CD8 T cells as a predictive biomarker for bortezomib benefit is intriguing, the precise mechanism remains unclear. Bortezomib has known immunomodulatory anti-tumor effects on T cells beyond NF-kB inhibition in malignant B cells, but how a resident CXCR5+ CD8 T cell population would specifically benefit from proteasome inhibition - whether through enhanced antigen presentation, altered cytokine milieu, or direct effects on T cell function - requires experimental validation.
Dissociation bias and scRNA-seq limitations: The scRNA-seq datasets used to validate EcoTyper states involve tissue dissociation, which is known to distort the representation of certain cell types and alter expression profiles of stress-sensitive populations. The authors acknowledge this limitation and note that EcoTyper's primary discovery platform (bulk transcriptomics with digital deconvolution) avoids dissociation artifacts. Nevertheless, the single-cell validation component inherits these limitations, particularly for rare cell populations such as mast cells and dendritic cells with inherently sparse scRNA-seq representation.
Prospective validation requirement: All survival associations between cell states/ecotypes and outcomes are derived from retrospective analyses of previously collected clinical datasets. The predictive biomarker finding for bortezomib was demonstrated by re-analyzing existing trial data rather than in a prospective trial designed to select patients by CD8 T cell S1 status. Prospective validation - including a clinical trial that pre-specifies CD8 T cell S1 or LE5 abundance as the patient selection criterion for bortezomib-based therapy - is required before this biomarker strategy can influence clinical practice.
EcoTyper's compatibility with clinical biospecimens: A critical practical advantage of EcoTyper is its applicability to formalin-fixed paraffin-embedded (FFPE) tissue, the standard format for clinical tumor archiving. The paper directly demonstrates this by applying EcoTyper to deconvolve more than 1,500 FFPE gene expression profiles from the REMoDL-B trial and the Chapuy et al. validation cohort. This compatibility means EcoTyper does not require fresh frozen tissue or specialized collection protocols, and could in principle be applied to routinely collected diagnostic biopsies using standard bulk RNA profiling assays - including cost-effective amplicon-based RNA panels or existing clinical-grade sequencing platforms.
Therapeutic targeting opportunities: The mapping of 44 cell states and 9 ecosystems with prognostic and predictive significance opens multiple avenues for therapeutic intervention. For example, the identification of LAG3 and OX40 co-expression on T cells enriched in ABC DLBCL provides a cellular and spatial context for testing combination checkpoint strategies in this subtype. The characterization of M1-like macrophage state (monocytes/macrophages S2) as a favorable-prognosis TME constituent suggests potential benefit from macrophage-activating strategies. The CD47-SIRPalpha "don't eat me" signal, expressed on tumor cells, is a target for macrophage-mediated phagocytosis and is under active clinical investigation in DLBCL.
Generalizability to other lymphomas and cancers: The REMoDL-B re-analysis used FFPE specimens from a multi-center international trial, and the EcoTyper DLBCL atlas was validated in datasets profiled across different platforms (RNA-seq, microarray), tissue preservation methods (fresh/frozen, FFPE), and clinical settings. The de novo state rediscovery in follicular lymphoma microarray data (66% of states reidentified) suggests partial conservation of DLBCL cell state architecture in related B cell malignancies. The broader EcoTyper framework has been applied to solid tumors including lung, breast, and colorectal cancer, indicating that the "top-down" ecosystem discovery paradigm is generalizable beyond lymphoma.
The EcoTyper web resource: To accelerate adoption by the research community, the authors made EcoTyper freely available for non-profit academic use at https://ecotyper.stanford.edu/lymphoma, along with the code, trained models, and the complete 44-state DLBCL atlas. Researchers can apply these pre-learned cell state definitions to new DLBCL datasets without repeating the discovery process, enabling rapid cell state quantification in new clinical cohorts and trial datasets. This open science infrastructure positions EcoTyper as a reusable community resource for dissecting TME heterogeneity in DLBCL and related malignancies.