Gene Expression-Based Classifier Reclassifies Burkitt Lymphoma, HGBL NOS, DLBCL NOS, and Double/Triple Hit Lymphomas into Subtypes with More Uniform Mutational and Microenvironment Landscapes and Treatment Response

Introduction: Aggressive lymphomas possess unique molecular features; however, high-grade B-cell lymphoma (HGBL), NOS is an entity without reproducible diagnostic criteria with limited genomic data. Thus, we performed a comprehensive genomic analysis across aggressive lymphomas, aiming to identify subgroups by more uniform biology and clinical course.

Methods: We identified cases of HGBL with MYC/BCL2 (DH) with/without BCL6 (TH) rearrangement, Burkitt lymphoma (BL), diffuse large B cell lymphoma, NOS (DLBCL), and HGBL, NOS from the University of Miami (UM). Transformed and EBV-positive (except for BL) cases were excluded. Expert hematopathologists independently reviewed all cases with diagnoses established by consensus. We performed whole exome (WES) and bulk RNA sequencing (RNAseq) and selected somatic mutations using a supervised machine-learning approach. LymphGen was used for sample subclassification.

Results: Ninety-two cases passed DNA and RNA quality control: HGBL NOS: 34, DLBCL: 27, DH/TH: 17, and BL: 14. We developed a classifier to distinguish DLBCL, DH/TH, and BL while excluding HGBL, NOS. Gene expression (GE) in each subtype was compared against the two others and top genes for each subtype were selected (n= 265 unique genes), and 122 genes were subsequently added. DLBCL and BL samples from public datasets divided into training (n= 4005) and test (n= 211) sets were used for additional training, feature selection and testing. Lowly expressed genes in the UM dataset were filtered from this initial set of features (n= 183 genes), leaving behind only genes present in all public datasets used for training/testing (n= 135). We used the BorutaShap algorithm on the training set for feature selection, identifying 27 genes important for classification. Then, we trained the light gradient-boosting machine (LGBM) classifier with hyperparameters selected by GridSearchCV. The LGBM classifier demonstrated weighted precision = 0.97 and weighted recall = 0.95. The classifier was additionally validated in our and the test cohort (AUC >0.95 in both).

Expression patterns of the 27 selected genes separated HGBL, NOS into two subtypes: one resembling DLBCL (n= 28, 82.3%; DLBCL-like) and another resembling BL (n= 6, 17.7%; BL-like). The most impactful genes were CYB5R2, BATF, and BMP7. Applying the classifier to all cases reassigned 41% of DH/TH and 7.4% of DLBCL cases as BL-like. Moreover, 21.4% of BL cases were considered DLBCL-like. Transcriptomic features of the DLBCL-like and BL-like cases were similar to those of DLBCL and BL, with significant differences between the groups (e.g., Th2, follicular dendritic cells, protumor cytokines, lymphatic endothelium, fibroblastic reticular cells, M1 signature, B cell traffic, CAF, antitumor cytokines, hypoxia, NF-kB, and TNFa). Tumor microenvironment analysis showed that both BL and BL-like cases had a predominantly depleted microenvironment. The BL-like cases showed inferior overall survival (OS, P= 0.02) since they were treated with R-CHOP based on the original diagnosis. Compared to DH/TH and DLBCL, DLBCL-like cases had a more depleted lymphoma microenvironment. Mutation in KMT2D, CD79a, and MYD88 occurred more often in DLBCL and DLBCL-like cases, while MYC, TP53, and GNA13 mutations were more common in BL and BL-like cases. DH/TH demonstrated features of both groups and high mutation incidence in BCL2 and CREBBP.

To validate the classifier, we analyzed publicly available datasets with DLBCL (n= 4021 treated with R-CHOP), BL (n= 195), and DH (n= 78) samples. According to the LGBM classifier, 3.2% of DLBCL cases were reclassified as BL-like, 1.5% of BL as DLBCL-like, and 28 (36%) and 50 (64%) of DH as BL-like and as DLBCL-like, respectively. Again, BL-like cases demonstrated lower OS (P= 0.002). We showed that the probability of a tumor classified by the LGBM Classifier as BL-like has the most significant impact on treatment selection and survival.

Conclusions: Genomic analysis and classification of well-annotated aggressive lymphoma cases divided these histologies into DLBCL-like and BL-like subsets, supported by mutation frequency and transcriptomics data. The Cox regression model showed BL as a more favorable diagnosis; however, DLBCL reclassified as BL-like showed shorter survival when treated with R-CHOP. This study underscores the need for genomic analysis in treatment selection for patients with aggressive lymphomas.