Application of the LymphGen classification tool to 928 clinically and genetically‐characterised cases of diffuse large B cell lymphoma (DLBCL)

We recently published results of targeted sequencing applied to 928 unselected cases of diffuse large B-cell lymphoma (DLBCL) listed in the Haematological Malignancy Research Network (HMRN) registry.1 Clustering allowed us to resolve five genomic subtypes. These subtypes shared considerable overlap with those proposed in two independent genomic studies,2, 3 suggesting there is potential to use genetics to stratify patients by both risk and biology. In the original studies, clustering techniques were applied to sample cohorts to reveal the molecular substructure but left open the challenge of how to classify an individual patient. This was addressed by the LymphGen classification tool.4 LymphGen assigns an individual case to one of six molecular subtypes. The tool accommodates data from exome or targeted sequencing, either with or without copy number variant (CNV) data. Separate gene expression data allows classification of a seventh, MYC-driven subtype, defined by a double hit or molecular high-grade (MHG) gene expression signature.5-7 Our large cohort of unselected registry patients, with comprehensive clinical and molecular annotation, provides an opportunity to examine the prognostic implications of the LymphGen classifier and to compare the robustness of cluster assignment across studies. Our sequencing panel provided only limited CNV data; therefore, we chose to enter exclusively mutation data. The LymphGen tool is able to accommodate mutation-only data, recognising that without CNV data the A53 subtype cannot be identified. We previously saw a strong negative prognostic effect of a truncating NOTCH1 mutation1; we therefore modified our original classification to annotate all patients with a truncating exon-34 NOTCH1 mutation as a distinct subgroup. We compared the classification assigned by our own clustering to that assigned by the LymphGen classifier. Our original clustering assigned a molecular subtype to 73% of cases. LymphGen assigned a unique classification in 53% (489) cases (Fig 1A and Table SI). Forty-six percent of the cases remained unclassified and 1% was assigned to overlapping categories of uncertain significance. We restricted further analysis to the 477 cases confidently classified in both our study and by the LymphGen classifier, to establish the extent of agreement at the level of individual samples (Fig 1B). We saw strong consensus among these cases, with 86% classified to the analogous LymphGen subtype (Fig 1B). In particular, we saw a 95% overlap between MYD88 and MCD subgroups and 96% overlap between BCL2 and EZB subgroups. Our SOCS1/SGK1 and TET2/SGK1 clusters represented subdivisions of the ST2 cluster with 89% of ST2 cases corresponding to one of these subgroups. This considerable overlap between separate classification strategies, identified using independent statistical approaches, demonstrates the robust reproducibility among the 'core' members of these molecular subtypes. However, 47% of our patients did not receive a unique LymphGen classification. In part, this may relate to the lack of CNV data precluding A53 identification. However, the A53 group represented only 7% of cases in the LymphGen study. Accordingly, even with full CNV data the original LymphGen publication classified only 57% of cases. In contrast, the original Chapuy publication assigned a classification to 96% of patients. Taken together, we conclude that analogous subgroups identified across studies represent the same robust, biological entities but that different classifications tolerate differing thresholds of uncertainty when assigning a subtype (Fig 1C). That is to say, the main variation between classifications is whether a case is classified at all, rather than the movement of confidently classified cases from one subgroup to another. We then looked at the prognostic implication of the LymphGen classifier in our cohort of patients. Strengths of our registry cohort include the large patient number, meticulous clinical annotation and comprehensive enrolment of every DLBCL diagnosis, without confounding referral bias. The LymphGen classifier suggests use of gene expression to identify a MYC-driven subgroup of the EZB cluster. Since gene expression was not available for every patient, and in an attempt to probe the utility of a mutation-only strategy, we took advantage of a recent observation that MYC mutations at codons 57–60 associate strongly with MYC-rearranged or MHG DLBCL.6, 8 We used the presence of these mutations to define a MYC-driven subgroup of the EZB cluster. Our previous analysis emphasised the importance of considering prognostic impact in homogeneously treated patients.1 Therefore, we restricted our analysis to patients receiving full dose R-CHOP (rituximab, cyclophosphamide, doxorubicin hydrochloride, vincristine, prednisolone). We excluded patients treated with regimens considered R-CHOP-like, who frequently received considerably attenuated chemotherapy and were not equally distributed across genomic subtypes1 (Table SII). We saw poor survival amongst patients assigned to the N1 group, a finding consistent across studies (Fig 2A,B). The MYC-EZB subgroup was also associated with poor survival, consistent with MYC/BCL2 rearranged DLBCL.9 In contrast, the ST2 subgroup was associated with favourable outcome. However, the prognostic impact of the remaining subtypes (MCD, EZB, BN2) did not achieve significance in R-CHOP treated cases. Unclassified cases had an intermediate survival (Fig 2A,B). Comparison with the international prognostic index (IPI) suggests that clinical factors remain a dominant determinant of survival in DLBCL, but that genetic classification provides independent prognostic information over and above the IPI (Fig 2C,D; Table SII). In summary, we conclude that mutation-only data from targeted sequencing allows a confident LymphGen classification in just over 50% of patients. These cases show strong consensus across different classification strategies, reinforcing the robust reproducibility of the core disease subgroups. Identification of the A53 subgroup will require either exome data or a panel specifically designed to provide the required CNV data. Both N1 and MYC-EZB, were associated with a markedly inferior prognosis, whilst ST2 showed consistently favourable outcome. We did not observe a significant prognostic impact from MCD, EZB and BN2 subgroups in R-CHOP treated patients. Nevertheless, the greatest potential of this classification will be to allow biological stratification of a disease where genetic heterogeneity will otherwise stymie our ability to assess the benefit of biologically targeted therapy, where efficacy may be restricted to specific biological subtypes. Whilst knowledge of the molecular subtype may not yet define the optimal therapy for an individual patient it will allow us to design and interpret clinical trials of these agents in the future. HR was funded by a studentship from the Medical Research Council (MRC). DH was supported by a Clinician Scientist Fellowship from the Medical Research Council (MR/M008584/1). The Hodson laboratory receives core funding from Wellcome and MRC to the Wellcome-MRC Cambridge Stem Cell Institute and core funding from the Cancer Research UK (CRUK) Cambridge Cancer Centre. HMRN is supported by BCUK 15037 and CRUK 18362. PB reports consultancy for Karus Therapeutics (Oxford, UK), OncoDNA (Gosselies, Belgium) and Everything Genetic (London, UK). HR performed data analysis. SC performed statistical analysis. SL, SB, PB, DP, AS, ER, CB and RT discussed and interpreted data. DH wrote the manuscript with input from HR. All authors read and approved the final manuscript. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.