AUTHORS
Reyhan Keser, PhD, Jansen N Seheult, MBBCh, BAO, MSc, MS, MD2, Yanglan Ou, D.Sc., M.E.3, Janek S. Walker, PhD4, Joseph Peter Novak, MS5, Elizaveta Flerova6, Ayo Samuel Falade, MD, MBA7, Steven Huang8, Yurim Lee, PhD9, Prasad Bandodkar, Ph.D.10, Raphael Mwangi, MS11, Thomas Tavolara, PhD12, Aishwarya Ravindran, MBBS13, Patrizia Mondello, MD, PhD14, Pedro Horna, M.D.15, Sarah Gibson, MD16, Cheryl Willman, MD17, Philippe Armand, MD18, Thomas Witzig, MD19, Stephen Maxted Ansell, MD, PhD20, James R. Cerhan, MD, PhD21, Rebecca L King, MD22, Matthew J. Maurer, Sc.D., M.S.23, Anne J. Novak, PhD24, Wenchao Han, PhD25, Eric Hsi, M.D.26
AFFILIATIONS
1 Mayo Clinic, Rochester, Minnesota, United States
2 Mayo Clinic, Rochester, Minnesota, United States
3 Mayo Clinic, Rochester, Minnesota, United States
4 Mayo Clinic, Rochester, Minnesota, United States
5 Mayo Clinic, Rochester, Minnesota, United States
6 University of Pennsylvania, Philadelphia, Pennsylvania, United States
7 Mayo Clinic, Rochester, Minnesota, United States
8 Mayo Clinic, Rochester, Minnesota, United States
9 Mayo Clinic, Rochester, Minnesota, United States
10 Mayo Clinic, Rochester, Minnesota, United States
11 Mayo Clinic, Mayo Clinic, Minnesota, United States
12 Mayo Clinic, Rochester, Minnesota, United States
13 Mayo Clinic, Rochester, Minnesota, United States
14 Mayo Clinic, Rochester, Minnesota, United States
15 Mayo Clinic, Rochester, Minnesota, United States
16 Mayo Clinic, Phoenix, Arizona, United States
17 Mayo Clinic, Rochester, Minnesota, United States
18 Mayo Clinic, Rochester, Minnesota, United States, Mayo Clinic, Rochester, Minnesota, United States
19 Mayo Clinic, Rochester, Minnesota, United States
20 Mayo Clinic, Rochester, Minnesota, United States, Mayo Clinic, Rochester, Minnesota, United States
21 Mayo Clinic, Mayo Clinic, Minnesota, United States, Mayo Clinic, Rochester, Minnesota, United States
22 Mayo Clinic, Rochester, Minnesota, United States
23 Mayo Clinic, Mayo Clinic, Minnesota, United States
24 Mayo Clinic, Rochester, Minnesota, United States
25 Mayo Clinic, Rochester, Minnesota, United States
26 Mayo Clinic, Rochester, Minnesota, United States
ABSTRACT
Background: While gene expression profiling (GEP) stratifies diffuse large B-cell lymphoma, not otherwise specified (DLBCL, NOS) into germinal center B-cell-like (GCB) and activated B-cell-like (ABC) cell-of-origin (COO) subtypes with prognostic and therapeutic implications, it is not routinely available. We previously developed LymphoVision, a lymphoma-specialized pathology foundation model trained on H&E from 9,155 archival lymphoma and reactive lymphoproliferative cases, and showed that it predicts COO at diagnosis from Hans-immunohistochemistry (IHC) training labels (n=358: mean test AUC 0.93, accuracy 84.5%, Blood 2025;146 Suppl 1:117), outperforming general-purpose foundation models. We now aim to demonstrate LymphoVision’s accuracy against GEP in newly diagnosed DLBCL specimens and in the relapsed/refractory (r/r) setting, and whether the prediction is stable between paired newly diagnosed and r/r specimens from the same patient.
Methods: Four cohorts were assembled, none overlapping the 358-case training set. (1) Diagnostic GEP validation: 142 DLBCL specimens with NanoString GEP or gneSeqCOO v1.1 RNA sequencing COO calls. (2) R/r GEP validation: 79 r/r DLBCL specimens with gneSeqCOO v1.1 RNA sequencing COO calls, with predictions generated on the relapse specimen H&E. (3) Exploratory outcome assessment: 68 diagnostic cases with mature follow-up for event-free survival at 24 months (EFS24) and cause of death; all were excluded from model development, and selection was independent of outcome. (4) Paired concordance: 42 patients with a diagnostic and at least one r/r specimen, yielding 45 pairs. GEP-unclassified cases were excluded rather than counted as discordant. Metrics included accuracy with Wilson confidence intervals (CI), sensitivity, specificity, Cohen kappa, area under the receiver operating characteristic curve (AUC), and Spearman correlation of model probability with the GEP linear predictor score (LPS). Outcomes used cause-specific models with non-lymphoma death as a competing event. Deaths of undetermined cause (n=4) were treated as non-lymphoma deaths.
Results: In the newly diagnosed GEP-labeled cohort (n=142, 57% GCB), LymphoVision achieved AUC 0.94 (95% CI 0.90-0.98), accuracy of 85.2% (78.4-90.1) and F1 of 0.84; manually segmenting lymphoma in the 6 cases where tumor occupied less than 10% of the tissue area raised cohort accuracy to 89.4%. In r/r specimens (n=72, 54% GCB), AUC was 0.89 (0.81-0.97) with accuracy of 83.3% (73.1-90.2) and F1 of 0.88. Discrimination was not significantly lower than at diagnosis and was preserved in needle core biopsies (82.5%, n=40). Model probability correlated with the GEP LPS (Spearman rho 0.670).
In an exploratory analysis of 68 held-out newly diagnosed cases (63% GCB; median follow-up 109.5 months by reverse Kaplan-Meier), AI-predicted ABC had higher EFS24 failure (36.0% vs 25.6%; OR 1.64, 0.56-4.75) and roughly double the lymphoma-specific mortality (cumulative incidence 25.3% vs 12.0% at 60 months; cause-specific HR 1.54, 0.52-4.59). Most deaths among GCB cases were not lymphoma-related (13 vs 2 non-lymphoma deaths; Gray’s P=0.017). With 20 EFS24 failures and 13 lymphoma deaths the cohort is not powered for significance, but effect sizes match those reported for GEP-defined COO in the literature.
Across 45 diagnostic-relapse pairs, label concordance was 71.1% (56.6-82.3), with 9 GCB to ABC and 4 ABC to GCB switches (McNemar p=0.27). Expected agreement under independence at these accuracies is 70.1% (66.8-71.7). Concordance declined with time to relapse: 82.4% within 9 months (n=17), 75.0% at 9-24 months (n=16) and 50.0% beyond 24 months (n=12).
Conclusions: The LymphoVision H&E COO classifier is accurate against GEP at initial diagnosis and generalizes to r/r DLBCL without loss of discrimination. Predicted ABC carried the expected excess risk in an exploratory cohort. Discordance between paired specimens matched the level expected from assay reproducibility and increased with time to relapse. Serial GEP shows the same gradient at lower absolute discordance (0%, 4.2% and 10.6%, respectively; Hilton, JCO 2023); serial Hans-IHC, the training label source for LymphoVision-COO, changes in 20% of relapse samples (Maeshima, Hum Pathol 2024). These results support potential use of LymphoVision-COO as a rapid prescreen, including when tissue may be insufficient for GEP.
DISCLOSURES
Reyhan Keser, PhD: No relevant conflicts of interest to declare.