Background: Molecular classifications of diffuse large B-cell lymphoma (DLBCL) can identify patient groups that benefit from distinct, innovative therapies. However, their implementation in routine practice has yet to be demonstrated. DLBclass (Chapuy, 2025) is a probabilistic, neural network–based classifier that assigns DLBCL cases to one of five genetic clusters (C1–C5). MSK-IMPACT Heme is a clinical NGS panel for detecting somatic mutations and allele-specific copy number alterations (CNAs) (Ptashkin, 2023). Here, we demonstrate how MSK-IMPACT Heme data collected during routine clinical care across various disease states enable the classification of cases into DLBclass clusters.

Methods: DLBCL samples underwent MSK-IMPACT Heme sequencing along with clinical evaluation. This enabled the generation of DLBclass inputs: CNAs were called with FACETS, mutations annotated per Ptashkin (2023), structural variants identified by FISH, and cell-of-origin classified using the Hans algorithm.

Technical validation: A total of 531 unique samples were initially assessed: 139 were excluded for low tumor purity (<20%) or failed FACETS quality control, resulting in a final cohort of 392 samples. This included 279 de novo DLBCL cases (184 pre- & 95 post-treatment) and 113 transformed indolent NHL (tiNHL) cases. Gene coverage of DLBclass cluster assignments ranged from 52% to 98% across the MSK-IMPACT Heme panels. Among the pre-treatment de novo DLBCL samples, cases were distributed across clusters as follows: 11% in C1, 28% in C2, 31% in C3, 12% in C4, and 17% in C5. Cluster C3 was significantly enriched compared to the published reference DLBclass datasets (p < 0.001), while the distribution of the other clusters mirrored that of the reference cohort. Using a 0.7 confidence threshold, we observed a lower proportion of high-confidence predictions in clusters C1, C4, and C5 relative to the DLBclass data. Across 45 matched samples with both WES and MSK-IMPACT Heme sequencing, cluster assignments were consistent in 71% of cases. Among the 13 mismatches, 12 involved cluster C2, highlighting variability in CNA-driven clusters between platforms. Given that C2 is primarily defined by CNAs, we investigated the impact of CNA removal on cluster assignment. Excluding CNAs led to a reduction in C2-classified samples, from 33% to 3.6%. Notably, in post-treatment samples, C2 had comprised 42%, underscoring its relevance in that clinical context. Upon CNA removal, many C2 samples shifted toward C3, suggesting that in the absence of CNAs, mutation profiles and panel breadth play a dominant role in cluster classification. Applying the DLBclass framework to routine MSK-IMPACT Heme data reproduced cluster patterns seen in reference datasets, validating its use in clinical samples and supporting further study of cluster biology and phenotypes.

Biological and clinical correlations: In the de novo DLBCL cohort, the C3 cluster was significantly enriched for germinal center B-cell (GCB)–type cases, while C5 was more frequent among non-GCB cases (p < 0.001), both consistent with Chapuy (2025). Additionally, C3 and C5 showed enrichment for double/triple-hit lymphomas and primary CNS lymphomas (PCNSL), respectively. Cluster C2 was associated with low tumor mutational burden (TMB), whereas C4 correlated with high TMB. Patterns of progression-free survival mirrored those reported in the original DLBclass cohort, suggesting consistent clinical behavior across datasets. Among 15 patients with sequential samples, 80% retained stable cluster assignments over time. Changes in cluster confidence were mostly observed within C2, indicating overall low genomic heterogeneity across serial samples. Applying the DLBclass classifier to samples obtained after histologic transformation, we found that C3 was enriched in patients with prior follicular lymphoma, C2 in those with transformed chronic lymphocytic leukemia and marginal zone lymphoma (p < 0.001). MZL samples also showed enrichment in C5 (p < 0.003).

Conclusion: Our findings are the first to demonstrate the feasibility and utility of applying the DLBclass genomic classifier to routine clinical samples using the MSK-IMPACT Heme assay. In our cohort, the classifier reproduced cluster distributions similar to reference datasets, establishing a basis for incorporating genomic subtyping into clinical workflows and improving molecular classification to guide personalized therapeutic strategies in DLBCL.

This content is only available as a PDF.
2025
Sign in via your Institution