AUTHORS
Steven Wang, MD, Amirsaman Ashtari, PhD2, Jordan Cheng, PhD3, Nick Phillips4, Jordan Scott Goldstein, MD, MSc5, Chih Long Liu, PhD6, Yavuz M. Bilgin, MD7, Yorick Sandberg8, Lara Bohmer, MD9, Otto Visser, MD, PhD10, Nicole de Graauw, MD11, Inge Ludwig, MD12, Henk van Zaanen, MD13, Yvonne Tromp, MD14, Rogier Mous, MD, PhD15, Ad Koster, MD16, Eva de Jongh, MD17, Marcel Nijland, MD, PhD18, Dirk Michiel Pegtel19, Hayley Warsinske20, Maximilian Diehn, MD, PhD21, Mark Roschewski, MD22, David Kurtz, MD, PhD23, Martine E.D. Chamuleau, MD PhD24, Ash A. Alizadeh, MD, PhD25
AFFILIATIONS
1 Amsterdam UMC - Location Vrije Universiteit, Amsterdam, Netherlands, Stanford University, Stanford, United States
2 Stanford University, Stanford, United States
3 Stanford University, Stanford, United States
4 Stanford University, Stanford, United States
5 Stanford University, Stanford, United States
6 Stanford University, Stanford, United States
7 Ardz, Goes, Netherlands
8 Maasstad Hospital, Rotterdam, Netherlands
9 Haga Teaching Hospital, The Hague, Netherlands
10 Isala Hospital, Zwolle, Netherlands
11 Bravis Hospital, Roosendaal, Netherlands
12 Bernhoven, Uden, Netherlands
13 Franciscus Hospital, Rotterdam, Netherlands
14 Saxenburgh Medical Center, Hardenberg, Netherlands
15 Utrecht UMC, Utrecht, Netherlands
16 VieCuri Medical Center, Venlo, Netherlands
17 Albert Schweitzer Hospital, Dordrecht, Netherlands
18 UMC Groningen, Groningen, Netherlands
19 Amsterdam UMC - Location Vrije Universiteit, Amsterdam, Netherlands
20 Foresight Diagnostics, Boulder, United States, Natera, Austin, United States
21 Stanford University, Stanford, United States
22 Memorial Sloan Kettering Cancer Center, New York, United States
23 Foresight Diagnostics, Boulder, United States
24 Amsterdam UMC - Location Vrije Universiteit, Amsterdam, Netherlands
25 Stanford University, Stanford, United States
ABSTRACT
Background: Genomic profiling of diffuse large B-cell lymphoma (DLBCL) has produced several molecular classifications based largely on coding mutations, copy-number alterations, and structural variants (SVs), defining subtypes with targetable pathways now under evaluation in biology-guided trials (e.g., GUIDANCE-01) (Zhang et al., Cancer Cell 2022; Shipp et al., Blood Cancer Discov 2026). A substantial limitation is that many cases remain unclassifiable or composite. Bal et al. (Nature 2022) showed that non-coding variants in super-enhancers and enhancers (SE/E) can deregulate DLBCL oncogene expression. We hypothesized that for patients with otherwise-unclassifiable tumors, a baseline assay could classify DLBCL from tissue or plasma, and that SE/E hypermutation profiles could adjudicate the molecular class.
Methods: We applied phased variant enrichment and detection sequencing (PhasED-Seq; CLARITY) (Kurtz et al., Nat Biotecnol 2021) to baseline tumor and plasma samples from 164 newly diagnosed DLBCL patients in the HOVON-902 study of curative-intent first-line immunochemotherapy. Tumors were classified by LymphGen v2.0, DLBclass, HMRN, and LymphPlex from single-nucleotide variants (SNVs) and SVs, and tissue–plasma concordance was assessed. Somatic SNVs were mapped to 228 recurrently hypermutated SE/E intervals (Bal et al.); each sample–region pair was scored as hypermutated using the Bal et al. (enrichment over the sample's background aberrant sometic hypermutation (aSHM) density [Benjamini-Hochberg FDR<0.05] plus ≥3 SNVs with ≥1 pair within 1 kb). Among high-confidence LymphGen-classified cases, per-region subtype enrichment was tested by Fisher exact test (BH-FDR<0.10), and a SE/E signature classifier was evaluated for resolving "other" and composite cases.
Results: Among 164 profiled patients, 85 (52%) were evaluable in having both paired baseline tissue and plasma. Plasma-based classification was highly concordant with tissue: LymphGen calls agreed across core subtypes in unambiguous cases (33/34; κ=0.96). Across all four classifiers, tissue–plasma concordance was 71–78% overall and 77–97% among unambiguous calls (κ = 0.62–0.96). A substantial fraction of detected SNVs fell in non-coding regions, and LymphGen left 46% (75/164) of cases unclassified ("other") or composite, while DLBclass had 46% unclassified at ≥70% confidence. Of 228 reference SE/E intervals, 68 (30%; 60 SE, 8 E) were recurrently hypermutated in this cohort. In total, 54% (89/164) of the cases were assigned a subtype by tissue and/or plasma. To identify cluster-defining SE/E profiles, we used a high-confidence LymphGen reference set (n=78; EZB=36, MCD=17, ST2=14, BN2=11), obtained after excluding 9 low-confidence cases (highest subtype probability <0.8) and 2 tissue-plasma discordant cases. Seven SE/E were subtype-enriched, consistent with prior associations of BCL2 (EZB, 89%), PIM1 (MCD, 35%), and SOCS1, ZFP36L1, and MIR142 (ST2); two IGH-linked intervals also enriched in MCD were excluded from classification given their potential physiological somatic hypermutation. A machine-learning (ML) classifier built from non-coding variants falling within these 5 non-IGH SE/E regions was able to correctly classify 45/48 (94%) of LymphGen classified cases. More importantly, this ML model adjudicated 20 of 75 (27%) otherwise unclassified or composite cases (EZB n=9; ST2 n=6; MCD n=5), with composite calls resolving coherently after reclassification (e.g., EZB/MCD to EZB or MCD), leaving 34% LymphGen unclassified.
Conclusion: Both coding and non-coding variants in baseline tissue or plasma support molecular classification with high tissue–plasma concordance, enabling subtyping when tissue is limited or of poor quality. Incorporating non-coding SE/E hypermutations resolved a meaningful fraction of otherwise-unclassifiable or composite cases, potentially extending biology-guided therapy to more patients. Validation in an independent cohort is ongoing. Because the phased-variant-based assay used at baseline also supports measurable residual disease assessment, these data support a unified, risk-adapted precision-medicine strategy in DLBCL.
DISCLOSURES
Steven Wang, MD: Amsterdam UMC - Patents & Royalties