Practical microenvironment classification in diffuse large B cell lymphoma using digital pathology

Abstract

Diffuse large B cell lymphoma (DLBCL) is a heterogeneous B cell neoplasm with variable clinical outcomes influenced by both tumor-derived and lymphoma microenvironment (LME) alterations. A recent transcriptomic study identifies four DLBCL subtypes based on LME characteristics: germinal center (GC)-like, mesenchymal (MS), inflammatory (IN), and depleted (DP). However, integrating this classification into clinical practice remains challenging. Here, we utilize deconvolution methods to assess microenvironment component abundance, establishing an LME classification of DLBCL using immunohistochemistry markers and digital pathology based on CD3, CD8, CD68, PD-L1, and collagen. This staining-based algorithm demonstrates over 80% concordance with transcriptome-based classification. Single-cell sequencing confirms that the immune microenvironments distinguished by this algorithm align with transcriptomic profiles. Significant disparities in overall and progression-free survival are observed among LME subtypes following rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone (R-CHOP) or R-CHOP with targeted agents (R-CHOP-X) immunochemotherapy. LME subtypes differed from distinct immune escape mechanisms, highlighting specific immunotherapeutic targets and supporting application of this classification in future precision medicine trials.

Keywords: algorithm; cell of origin; diffuse large B cell lymphoma; digital pathology; immune evasion; immunotherapy; lymphoma microenvironment.

Graphical abstract

Figure 1

Patient selection flowchart Flowchart of patient selection. 682 patients newly diagnosed with DLBCL-NOS or high-grade B-cell lymphoma (HGBCL) were included with RNA sequencing data. DNA sequencing data were available in 662 patients, and archived FFPE tissues were available in 315 patients.

Figure 2

Identifying transcriptome-based LME subtypes (A) Cellular and molecular components of LME quantified by single-sample gene set enrichment analysis (ssGSEA) scores. (B) Cellular components of LME quantified by deconvolution methods. (C and D) Kaplan-Meier curves of progression-free survival (C) and overall survival (D) according to transcriptome-based LME subtypes in the context of R-CHOP. (E and F) Kaplan-Meier curves of progression-free survival (E) and overall survival (F) according to transcriptome-based LME subtypes in the context of R-CHOP-X. LEC, lymphatic endothelium; VEC, vascular endothelium; CAF, cancer-associated fibroblasts; FRC, fibroblastic reticular cells; ECM, extracellular matrix; IS, immune suppressive; FDC, Follicular dendritic cells; TFH, follicular T helper cells; TIL, tumor-infiltrating lymphocytes.

Figure 3

Candidate staining marker selection (A–C) LME cellular components with top 10 AUC values discriminating one transcriptome-based LME subtype from all the other subtypes and classical staining markers of selected cellular components. (D–K) Correlation between selected gene expression levels and abundance of corresponding cellular components.

Figure 4

Staining and WSI of selected staining markers (A) Representative bright-field images of selected staining markers. The ruler represents 40 μm. (B) Representative whole-slide image (WSI) analysis markups of markers with moderate expression levels. CD3-, CD4-, CD8-, CD68-, and PD-L1-stained tissues were marked up with positive and negative cell detections; FAP- and collagen-stained tissues were marked up with positive and negative pixels; CD34-stained tissues were marked up with microvessel detections.

Figure 5

Developing and validation of the staining-based algorithm (A) RPART model for prediction, including CD3, collagen, CD8, CD68, and PD-L1. (B and C) Accuracy of LME subtypes in the training cohort and test cohort. (D and E) Kaplan-Meier curves of progression-free survival (D) and overall survival (E) according to transcriptome-based LME subtypes. (F and G) Kaplan-Meier curves of progression-free survival (F) and overall survival (G) according to RPART model-predicted LME subtypes. (H and I) Kaplan-Meier curves of progression-free survival (H) and overall survival (I) according to the external validation cohort.

Figure 6

Tumor microenvironment characteristics of staining-based LME subtypes by single-cell RNA sequencing (A) Distributions of six common cell clusters among different staining-based LME subtypes. (B) Gene expression profiles of six clusters. (C) Myeloid cells were divided into seven subclusters: cDC1, cDC2, pDCs, mature DCs, monocytes, macrophage-C1, and macrophage-C2. (D) Gene expression profiles of seven myeloid subclusters. (E) T/NK cells were further divided, and a total of 10 subclusters were identified, including 1 naive T, 2 CD4⁺ subclusters, 6 CD8⁺, and 1 NK subclusters. (F) Gene expression profiles of 10 T/NK subclusters. (G) Quantified percentages of the microenvironment components of each staining-based LME subtype.