Prognostic mutational subtyping in de novo diffuse large B-cell lymphoma

Molecularly defined DLBCL subgroups provide additional context for treatment decisions in addition to classical prognostic features (IPI, BCL2/MYC status).

Diffuse large B-cell lymphoma (DLBCL) is a genetically and clinically heterogeneous disease [1, 2]. Since 2000, gene expression profiling has been used to categorize DLBCL into distinct cell-of-origin (COO) subtypes, which include activated B-cell-like (ABC) and germinal center B-cell-like (GCB) subtypes [3‐5]. These subtypes arise from different stages of normal B-cell development and their prognostic impact has been confirmed in several studies. Notably, in retrospective analyses of patients with DLBCL receiving rituximab (R) plus cyclophosphamide, doxorubicin, vincristine, and prednisone (CHOP; R-CHOP), the ABC subtype has been associated with less favourable outcomes compared with the GCB subtype [6, 7]. However, the prognostic difference based on COO alone has not always translated to clinical observations in the R-CHOP era [8‐10]. Clinically relevant genetic subgroups that contribute to the molecular profile of the COO subtypes have been described and may yield greater discrimination and stratification than COO alone (e.g. CD79b or MYD88 mutations in ABC and BCL2 translocation in GCB) [11‐15].

With advances in next-generation sequencing, the ability to further interrogate genetic drivers to refine prognostic subgroups over and above COO subtypes has generated considerable interest. Using whole-exome and transcriptome sequencing of tumors from a cohort of 1001 newly diagnosed patients with DLBCL, Reddy et al developed a prognostic model comprising genetic alterations, COO DLBCL subtype, International Prognostic Index (IPI) score and dual MYC and BCL2 expression, that outperformed molecular or clinical factors alone (COO, MYC/BCL2 expression, IPI) in terms of prognostic ability for overall survival (OS) [16]. Further elucidation of some of the clinical and genetic heterogeneity in transcriptionally defined COO subsets from de novo DLBCL identified several common mutational profiles with distinct prognostic impact [2, 17]. Applying an unsupervised prognostic clustering approach, non-negative matrix factorization (NMF), Chapuy and colleagues were able to identify six distinct molecular clusters with potential clinical relevance leveraging whole exome sequencing; these included two distinct subsets of ABC (one enriched for mutations in MYD88 and CD79B, and another for BCL6 and NOTCH mutations) and a GCB subset enriched for BCL2 translocations and mutations in CREBBP and EZH2. These analyses demonstrated the value for incorporating multiple data types to identify prognostic subsets. As an additional step Wright et al. [18] developed a probabilistic algorithm, LymphoGen from multi-platform analyses to delineate clinically actionable subsets. These instrumental advances expand DLBCL subsets by including genetic data but rely on complex datasets/analyses from WES. FMI targeting panels have potential to make this type of analysis more accessible and therefore more actionable in the clinical setting.

In the current study, we applied the NMF-clustering approach on targeted exome-sequencing data from the Phase III GOYA study (NCT01287741)⁵ and the Phase Ib/II CAVALLI study (NCT02055820) [19, 20] to scrutinize known molecular clusters and novel prognostic groups in de novo DLBCL. The randomized GOYA study evaluated R-CHOP or obinutuzumab (G)-CHOP in patients with previously untreated advanced-stage DLBCL. Results from the final analysis of GOYA reported no significant difference in progression-free survival (PFS) in patients treated with G-CHOP vs R-CHOP [5], and hence the treatment arms were pooled for this secondary analysis. In the subsequent CAVALLI study, addition of the highly selective BCL2 inhibitor venetoclax to R-CHOP was associated with an improvement in PFS and OS in patients with previously untreated BCL2 positive (by immunohistochemistry [IHC]) DLBCL, when compared with a matched historical control group treated with R-CHOP alone in GOYA [19, 20]. Inclusion of data from the CAVALLI study in our genomic analysis enabled us to explore whether outcomes are improved for patients with BCL2-rich NMF-clustering characteristics (i.e. the Chapuy defined GCB subset that is enriched for BCL2 translocations) who are treated with venetoclax. We also evaluated a real-world DLBCL cohort from a de-identified clinico-genomic database to determine whether the clusters identified in the GOYA study could be replicated in a real-world cohort. Our aim was to demonstrate that molecular clusters of DLBCL can be identified using genomic information obtained from a clinically validated next-generation targeted sequencing platform and potentially inform treatment decisions.

Baseline disease characteristics were similar between patients in the biomarker-evaluable population and the ITT population for both the GOYA and CAVALLI study (Supplementary Table S1).

In recent years, a number of groups have made strides in capturing the heterogeneous genetic characteristics of DLBCL using complex multi-platform analyses [2, 16, 18], driving the need for simplified methods that enable clinical implementation and inform targeted therapeutic strategies. We applied NMF clustering to targeted exome-sequencing data utilizing the F1H panel on samples obtained from the Phase III GOYA (randomized R-CHOP versus G-CHOP) and Phase Ib/II CAVALLI (venetoclax + R-CHOP) clinical studies. Analysis of data from the GOYA study, in patients with de novo DLBCL, demonstrated that the mutation profiles, COO subset distribution, and NMF clusters previously identified using whole exome profiling [2‐5, 17] could largely be recapitulated using targeted mutational data from the F1H panel. The NMF clusters were also consistent with subsets identified using the probabilistic classification tool [18]. Importantly, these analyses validate the Chapuy NMF clusters [2] and their prognostic value for standard of care using large, independent, contemporary datasets.

While BCL2/EZH2 (C3, EZB), MYD88/CD79b (C5, MCD), and NOTCH/TNFAIP3 (C1, BN2) subset identification and effects on prognosis are generally consistent in GOYA compared with the consolidated datasets used in Chapuy et al [2] and Wright et al [18], the panel is limited in terms of copy number amplification measurements, making detection of NMF clusters based on these types of variants (C2, A53) less likely. Consistent with this limitation, TP53 could not be assigned to the C2-like group due to its absence. Given the effects on prognosis of TP53 and the potential to target these variants, this is an identified limitation of this simplified platform. Similarly, C4 (ST2) was not observed despite reasonable coverage of genes encoding histone core and linker genes, as well as signaling genes on the F1H panel. A potential reason for the lack of C4 identified in GOYA is that the C4 heterogeneous cluster is defined by at least three mechanistic molecular profiles; histone regulators, signaling pathways and immune evasion. Unlike the other clusters containing key driver gene profiles, this subset is more heterogeneous and therefore may be more difficult to consistently capture. More generic differences could be attributed to variations in the patient populations and treatments with GOYA being more contemporary and homogeneous.

Furthermore, there were differences between GOYA, CAVALLI and FH-FMI CGDB across key genes likely due to reduced sample sizes. While BCL2 translocations and EZH2 variants were present, BCL2 SNVs were not assigned within the BCL2/EZH2 cluster in CAVALLI. Additionally, within the FH-FMI CGDB clusters, several key genes were not significantly assigned to clusters of patients. For example, MYD88, CD79B, NOTCH2 and TNFAIP3 were missing due to low prevalence and small sample size.

Given the heterogeneous mutational profile of de novo DLBCL [1, 16] differences between this analysis and other datasets (i.e. Chapuy) are not altogether unexpected. In contrast to Chapuy et al [2] who reported that genomic clusters were prognostic when adjusted with several clinical variables (e.g. IPI), our study indicated that NMF was not, and was of similar prognostic value to COO. Potentially, the more contemporary GOYA dataset (data cut-off January 31, 2018 vs CAVALLI data cut-off June 28, 2019) and more homogeneous treatment regimen (single trial compared with those for Chapuy, where 85% of patients were uniformly treated with R-CHOP) may account for these differences.

One of the key findings of the NMF clustering analysis using the GOYA and CAVALLI datasets is that while we were able to identify and confirm the prognostic potential of the high-risk NMF subsets (BCL2/EZH2 and MYD88/CD79B) to standard of care, validation of these NMF-defined subsets using the CAVALLI dataset showed distinct clinical response profiles; MYD88/CD79B exhibits high-risk/poor clinical response to the venetoclax R-CHOP regimen while in contrast, the BCL2/EZH2 subset appears to benefit and thus may be predictive for targeted therapies such as venetoclax. This difference in PFS outcome between GOYA and CAVALLI (Supplementary Fig. S5) is potentially explained by the mechanism of action of venetoclax, a highly selective BCL2 inhibitor [19], and its potential to improve treatment outcomes for patients with DLBCL with elevated levels of BCL2 [19, 20, 25, 26]. However, there are multiple other reasons this could occur and further confirmation is necessary. The addition of venetoclax to standard of care immunochemotherapy resulted in categorization of the BCL2/EZH2 cluster as low-risk; while the MYD88/CD79B cluster is maintained as high-risk across therapeutic settings. However, as the MYD88/CD79B cluster also includes BCL2 amplifications, we cannot explain why venetoclax might specifically benefit patients with either BCL2 mutations or translocations over amplifications. These data also highlight the limitations of BCL2 IHC alone as insufficient to define prognosis. Interestingly, elevated BCL2 levels by IHC may further stratify low-risk NMF clusters into high- and low-risk categories. In these clusters, different mechanistic drivers of BCL2 expression and activation are again present (mutations, amplifications and translocations) and likely contribute to prognostic or predictive potential warranting further investigation.

The real-world data cohort had the greatest differences compared with GOYA, where MYD88 and CD79B mutations had low prevalence (9/53 patients and 6/53 patients, respectively), NOTCH2 mutations were detected in only two patients and TNFAIP3 were detected in 8/53 patients. We attempted to corroborate our findings in a real-world setting, but differences in the prevalence of key variants and small sample size limited interpretation.

In summary, the aim of our study, was to identify a simplified platform for the identification of molecularly-defined subsets of DLBCL to inform treatment in clinical practice. Using the F1H assay, we were able to recapitulate previously defined prognostic subsets in contemporary datasets and demonstrate the potential to identify molecular subsets sensitive to targeted rational therapies, such as venetoclax, that require further validation.