Transcriptomic classification of diffuse large B-cell lymphoma identifies a high-risk activated B-cell-like subpopulation with targetable MYC dysregulation

Abstract

Immunochemotherapy has been the mainstay of treatment for newly diagnosed diffuse large B-cell lymphoma (ndDLBCL) yet is inadequate for many patients. In this work, we perform unsupervised clustering on transcriptomic features from a large cohort of ndDLBCL patients and identify seven clusters, one called A7 with poor prognosis, and develop a classifier to identify these clusters in independent ndDLBCL cohorts. This high-risk cluster is enriched for activated B-cell cell-of-origin, low immune infiltration, high MYC expression, and copy number aberrations. We compare and contrast our methodology with recent DLBCL classifiers to contextualize our clusters and show improved prognostic utility. Finally, using pre-clinical models, we demonstrate a mechanistic rationale for IKZF1/3 degraders such as lenalidomide to overcome the low immune infiltration phenotype of A7 by inducing T-cell trafficking into tumors and upregulating MHC I and II on tumor cells, and demonstrate that TCF4 is an important regulator of MYC-related biology in A7.

Fig. 1. Unsupervised clustering identifies biological subpopulations within newly diagnosed DLBCL.

Unsupervised clustering was applied to a large cohort of patient-derived RNAseq data to identify biologically homogeneous subtypes of DLBCL. A Schematic of data transformation, unsupervised clustering, and classifier training methodology. Steps in black represent data objects, while steps in blue represent algorithmic processes. B Co-clustering frequency heatmap identifies sample clusters that consistently group together over repeated subsampling runs. C Cluster prevalence and breakdown by COO and TME26 classification. Bar heights represent the observed proportion in each cohort, and error bars represent the 95% confidence interval. D Top 50 up- and down-regulated genes per cluster from the Discovery dataset, replicated in each cohort. Source data are provided as a Source Data file.

Fig. 2. Cluster A7 demonstrates a poor clinical prognosis to induction R-CHOP.

Kaplan Meier curves showing event-free (MER, REMoDL-B, GOYA) or overall (Reddy) survival among RCHOP-treated patients in the replication cohorts, with two-sided, unadjusted log-rank p-values shown. Survival is shown stratified (A) by all seven of the SubLymE clusters; (B) by a binarized model comparing A7 to non-A7, and (C) by a binarized model comparing A7 to non-A7 in the ABC subpopulation only. Source data are provided as a Source Data file.

Fig. 3. Biological interpretation of SubLymE clusters.

A Heatmap of biologically relevant pathways across SubLymE classes in the ROBUST dataset. Heatmap colors represent signature-level normalized enrichment scores. B Association of class calls by SubLymE and other classification methods. The heatmap shows Cohen’s Kappa values measuring strength of association between specific pairs of classes, tested by binarizing each classifier with respect to the classes being compared. Positive values indicate classes that are commonly called together, while negative values indicate classes that are rarely called together. Kappa values not significantly different from 0 (p < 0.05) were set to 0 to highlight statistically significant associations. Associations are shown for MER, with the exception of the MHG comparisons, which were tested in REMoDL-B. Source data are provided as a Source Data file.

Fig. 4. Copy number features of A7.

Prevalence of copy number gains and losses in A7 and non-A7 subpopulations (ROBUST), with FDR-corrected significant differences highlighted with asterisks. Source data are provided as a Source Data file.

Fig. 5. A7 DLBCL tumors are characterized by high c-myc expression and activity.

A Results of GSEA Hallmark pathway analysis presented as positive or negative association with A7 cluster membership. B c-myc gene expression by A7 status across clinical cohorts. In each cohort, MYC is expressesd at significantly higher levels in A7 compared to non-A7 (two-sided, unadjusted Wilcoxon p-values: ROBUST, <2.2e16; MER, 2.9e-7; REMoDL-B, 2.2e-9; GOYA, 5.5e-8; Reddy, 1.8e−9). C Representative c-myc IHC staining in multiple A7 cases and non-A7 cases; samples were selected for similar tumor cellularity. D Gene expression of TCF4 in patients plotted with copy number alteration status of TCF4 (ROBUST, N = 299), value in parentheses, copy number of TCF4;. Box plots represent the median and upper/lower quartiles, with whiskers extending to the most extreme values. **** indicates p  <  0.0001 (TCF4 Diploid vs. Gain, p = 1.8e−5; Diploid vs. Amplification, p = 4.4e−14), N.S., not significant (two-sided, unadjusted Wilcoxon test). E TCF4 mRNA expression in DLBCL cell lines, value in parentheses is the copy number of TCF4 in each model. F Western blot analyzing expression of TCF4 in ABC-DLBCL cell lines. Representative blots from 3 independent experiments. G, H Western blot analysis and quantitation of expression of MYC and TCF4 in TCF4 knockdown ABC-DLBCL cell lines. GAPDH, loading control. Represetative blots from 2 independent experiments. I Cell proliferation assay of control (shNT) and TCF4 knockdown (shTCF4) in ABC-DLBCL cell lines. Each point represents the mean of technical triplicates with standard error of the mean too small to visualize. Source data are provided as a Source Data file.

Fig. 6. A7 DLBCL tumors have low levels of immune infiltration and reduced MHC expression.

A Deconvolution plot of inferred immune cell populations as enriched or depleted in A7 relative to non-A7 patients. B Scatter plot of the Discovery cohort in the space of Reddy COO score vs. TME26 score. C Single channel MIBI images for indicated markers in representative A7 (red) and non-A7 DLBCL cases (gray). D Multiplexed MIBI images for indicated markers in representative A7 (red) and non-A7 DLBCL cases (gray). Source data are provided as a Source Data file.

Fig. 7. Lenalidomide in combination with R-CHOP can overcome poor prognostic outcomes associated with A7 in ABC-DLBCL patients.

A ChIP-seq peaks demonstrating direct binding of Aiolos and Ikaros at representative promoters of MHC class I and II genes in U2932 DLBCL cells. B Surface mean fluorescent intensity of MHC class I (black bars) and MHC class II (gray bars) in DLBCL cells treated with vehicle or lenalidomide for 24 h as measured by flow cytometry. Error bars represent standard error of the mean. C Vehicle and lenalidomide treated target DLBCL cell viability following co-culture with EBV reactive CD8⁺ T cells overnight. % live target cells are marked by inverted triangles. ELISA measurements for granzyme B in resulting co-culture supernatants are marked by round dots. Error bars represent standard error of the mean. D Cellular composition for indicated population in hCRBN/eu-myc transplanted mice treated with vehicle or lenalidomide (30 mg/kg) for four days). Scale bars = 300 µm. N = 5. E Intratumoral cell counts in transplanted mice and transplanted mice treated with vehicle or lenalidomide by multiplex immunofluorescence for indicated cell population (**p < 0.01, ****p < 0.0001; B cells No transplant vs. Vehicle p = 5.7e−10, Vehicle vs. Len p = 1.4e−7; CD8+ T cells No transplant vs Vehicle, p = 1.8e−12, Vehicle vs. Len p = 4.4e−6). F Kaplan Meier EFS curves for A7 and non-A7 ABC-DLBCL populations treated with either R-CHOP or lenalidomide plus R-CHOP (R2-CHOP) in the ROBUST trial, with accompanying table of group-pairwise log-rank p-values. Source data are provided as a Source Data file.