Practical Microenvironment Classification in Diffuse Large B-Cell Lymphoma Using Digital Pathology

Objective: Diffuse large B-cell lymphoma (DLBCL) is a highly heterogeneous disease, with patients experiencing varying clinical outcomes due to both tumor intrinsic factors and components of the lymphoma microenvironment (LME). A recent transcriptomics study identified four distinct subtypes of DLBCL based on their LME characteristics: Germinal center-like (GC), Mesenchymal (MS), Inflammatory (IN), and Depleted (DP). However, integrating this classification into routine clinical practice has proven challenging due to demanding tissue requirements, high costs, and lengthy processing times. To overcome these obstacles, this study aims to develop a staining-based algorithm for classifying LME in DLBCL patients.

Methods: This study enrolled 684 newly diagnosed DLBCL patients, including not otherwise specified or high-grade B-cell lymphoma. Transcriptome data from tumor tissues were utilized to establish the gold standard LME subtype. Among these patients, 315 had archived FFPE tissues, which were split into training (N=190) and test (N=125) cohorts at a ratio of 6 to 4. Potential markers were selected based on representative cells and highly expressed genes for each subtype. FFPE tissue sections from the cohort underwent immunohistochemistry and Mallory's trichrome staining, followed by quantification through whole-slide image analysis. The RPART decision tree algorithm was then employed for modeling. The resulting model was validated against the gold standard using the test cohort.

Results: The positive cell ratio of CD3, CD8, CD68, PD-L1, and the positive stained area ratio of collagen emerged as staining markers with discrimination exceeding 70% for the specified LME subtypes. Utilizing these five markers, the decision tree algorithm demonstrated 83.7% and 81.6% concordance rates with the gold standard in the training and test cohorts, respectively. Furthermore, significant disparities in overall and progression-free survival were observed among staining-based LME subtypes under RCHOP therapy. Multi-omics analyses uncovered distinct immune escape mechanisms among the four LME subtypes: GC and DP subtypes exhibited low immunogenicity, rendering them invisible to immune surveillance. Tumor growth in these subtypes depended on excessive stimulatory signals from follicular helper T cells and abnormal activation of MYC within tumor cells, respectively. Conversely, the MS and IN subtypes were highly immunogenic tumors, with their immune escape mediated through T cell exclusion caused by tissue damage repair and T cell exhaustion induced by immune checkpoints such as PD-L1, respectively.

Conclusion: This study provides a practical tool for classifying LME in DLBCL patients and sheds light on the underlying mechanisms of immune evasion employed by each LME subtype. Such information could serve as a guide for personalized immunotherapy in DLBCL patients.