Audit of B-cell cancer genes

A central goal in cancer genomics is to delineate the genetic and molecular changes that contribute to the onset and progression of cancer. By definition, driver mutations afford a fitness advantage to cells that ultimately give rise to a neoplastic clone. The term “driver” can refer to the individual mutations that contribute to cancer and as shorthand for any gene whose mutation can promote oncogenesis. Driver genes are identified by comparing the mutation profiles of tumors from a cohort of patients with the same diagnosis or pathology.¹ Although the computational strategies to accomplish this vary, each seeks to identify the genes with a somatic mutation rate and/or pattern distinct from that expected by chance, often reported as significantly mutated genes. When applied to a single cancer diagnosis such as Burkitt lymphoma (BL), these genes can be considered candidate drivers within that entity, or “BL genes.” Although some driver genes are pervasive, with similar roles in many cancers, others appear mutated in a limited number of entities. For a surprising fraction of these genes, the mechanism by which their mutation contributes to lymphoma initiation or progression is merely speculative. In this review, we summarize the current landscape of knowledge for the established and suspected drivers across 3 major types of B-cell lymphomas: diffuse large B-cell lymphoma (DLBCL),^2-7 follicular lymphoma (FL),^2,8-11 and BL.^12-15 

An unwritten rule in cancer genomics is that each cohort-based study is accompanied by a significantly mutated gene list, typically comprising a mixture of known and novel genes. For any well-studied entity, this paradigm has 2 consequences: an inconsistency in the genes reported in each study, and a perpetually growing collection of drivers. Accordingly, there has been a surge of candidate lymphoma genes since the availability of massively parallel sequencing. The ongoing “discovery” of new cancer genes might not be surprising, considering the estimate that up to 5000 samples may be needed to saturate the drivers for some malignancies.¹ Based on this projection, one could reason that the genes unique to any study (orphan genes) merely reflect interpatient heterogeneity,¹⁶ and will eventually be confirmed in subsequent studies. We see this explanation as incompatible with the data. Instead, the literature represents a mix of true drivers alongside genes that have no relevance to lymphoma, and without addressing them, such errors continue to be propagated, and the collection of genes continue to grow. This leaves the scientific community with the challenge of how to prioritize these genes and on which to focus efforts for future experiments involving targeted sequencing or functional characterization.

Collectively, the 3 B-cell lymphomas covered by this review have been the subject of dozens of genomic screens. With this level of attention, one might naively assume that we can agree on which genes are the common target of mutations in each. One of our goals with this review was to address the following question: how many (and which) genes are relevant drivers in each? Here, we propose a framework to curate and categorize candidate drivers based on the supporting evidence. We provide annotated lists of candidate drivers for DLBCL, FL, and BL, which are in the supplemental Data and a public version–controlled repository, as a resource to the community that will continue to be actively maintained (https://github.com/morinlab/LLMPP/).

In a perfect world, every genomic survey of the same malignancy would arrive at essentially the same list of driver genes. In lieu of this, a single publication might be designated as the gold standard based on some objective criteria. Experimental design considerations such as the type and volume of data (coverage), source of nucleic acids (frozen/formalin-fixed paraffin-embedded) and the inclusion of matched constitutional DNA to remove germ line variants can affect the accuracy of a gene list.¹⁷ Small cohorts, insufficient coverage, and limited gene panels, can all hamper the detection of genes with lower mutation rates or miss genes completely. Inadequate removal of systematic artifacts or germ line variants can increase the rate of false positives. We compared gene lists from 12 DLBCL studies spanning a decade to estimate the accuracy of each result. By comparing the gene list from each study to the rest, the relative number of orphan genes can serve as a proxy for the false-positive rate. A caveat is that more recent gene lists naturally have higher orphan rates because there has been less opportunity for those genes to be subsequently confirmed. We partially mitigated this by comparing smaller sets of studies, grouping them by age (Figure 1A). These comparisons reveal a striking range of discordance in the published gene lists, with some studies having orphan rates of 82% (263 of 319 genes in Zhang et al¹⁶) and 55% (73 of 133 genes in Fan et al¹⁸).

If all gene lists are affected by false positives and/or false negatives, how can we determine how many driver genes exist for a given malignancy? The 5th edition of the World Health Organization Classification of Haematolymphoid Tumours (WHO-HAEM5) places the number of DLBCL genes at ∼150,⁴⁰ which is based on the list proposed by Reddy et al.¹⁹ Notably, this lacks important genes that were established as DLBCL genes in earlier studies. These include ZFP36L1 and HIST1H1C,² and DTX1, the latter having been reported as a driver by the same group.^16,41 Additional DLBCL drivers identified in separate studies, such as NOTCH1,^3,5,PRKDC,^5,20,ITPKB,^5,ID3,^5,7,HVCN1,^6,7 and SIN3A^6,7 are also missing from that list.^5-7,21 In all cases, the importance of mutations in these genes has been established by observation in a second study or characterization in model systems.^42-45 Along with these absences, the list of Reddy et al had the third highest orphan gene rate (31%), implying a significant number of false positives. Some of the more recent studies of large cohorts were more conservative, each reporting <100 genes.^5,6 Of the 99 genes nominated by Chapuy et al⁶ only 58 overlap with the list of Reddy et al. Based on this intersection, we can confidently state that there are at least 58 DLBCL genes, but this is far from complete.

A fair and thorough consolidation effort should somehow consider the gene lists from every study while avoiding propagating errors. The reason for this becomes clear if we instead accept every gene, even if reported by a single study. Exploring how this affects DLBCL as a function of time, the total number of genes grows from 81 at the end of 2011^2,3,28,30 to nearly 700 a decade later (Figure 1B), with most of these genes originating from a pair of papers from the same group.^16,18 Although this strategy yields a comprehensive list, it clutters our list with genes unlikely to be relevant to DLBCL, thus is of limited value without further contextualization. A conservative way to mitigate this issue is to limit our list to the genes reported by a reasonable number of studies. Requiring a consensus across many studies, however, is equally problematic. For example, only 7 genes lie at the intersection of the first 6 DLBCL lists. The result is similar if the more recent group of studies are compared (Figure 1A). A more balanced approach to reduce the perpetuation of false positives without discarding valid candidates is to prioritize the genes reported by at least 2 studies. Based on the 12 studies, this brings us to 153 DLBCL genes, a number close to the current WHO-HAEM5 estimate.

Based on these concepts, we implemented a flexible approach to consolidate, curate, and categorize candidate B-cell lymphoma drivers into a separate gene list for each pathology. We began by identifying all studies that promoted at least 1 gene as recurrently mutated (by protein-coding mutations) in DLBCL, BL, or FL. Starting with the superset of novel genes from all such surveys of a given pathology, we record the initial study mentioning the gene (ie, the “originating study”). Genes are assigned to tiers to reflect our confidence in their role in that entity, with Tier 1 and Tier 2, respectively, representing the high- and moderate-confidence genes. Genes of particularly low confidence can also be assigned to a third tier; however, not all lists contain Tier 3 genes. This process draws from gene lists in all available peer-reviewed studies that sequenced patient samples of the entity and, in some situations, manual review and/or reanalysis of the primary data (supplemental Methods).^41,46 Genes are assigned to Tier 2 by default, in which they can remain indefinitely. Promotion to Tier 1 typically results from additional evidence of recurrent mutation demonstrated by a subsequent study. Figure 1C shows the origin of each Tier 1 DLBCL gene and the confirmation of many of these across subsequent studies.

The availability of functional experimental data demonstrating the role of a mutation in lymphoma biology also allows a gene to be promoted to Tier 1. For example, B2M was a Tier 2 DLBCL gene as of its first description in 2011.² It became eligible for Tier 1 when another group demonstrated their role in modulating recognition by immune cells.⁴⁷ In rare situations, the originating study may include functional data demonstrating the effect of the mutations in an appropriate model system, in which case the gene is immediately eligible for Tier 1. MYD88, MS4A1, CREBBP, and EP300 in DLBCL are examples of such genes.^28,30,34 We do not consider hits from generic pooled CRISPR screens as sufficient evidence to validate a candidate driver because they do not model the effect of the specific mutations observed. In general, determining the adequacy of functional data involves a more nuanced assessment of each study that is discussed hereafter.

Another complication affecting the genetic analysis of these lymphomas is the influence of aberrant somatic hypermutation on the genome, which causes an elevated mutation rate in certain genes.²⁷ Many genes affected by aberrant somatic hypermutation have been assigned to Tier 1 because of the mere recurrence of mutations.⁴⁸ Additional caution should be used when considering these mutations as drivers without additional evidence of their relevance. Importantly, although these may not all represent drivers, these mutations contribute to the classification of DLBCL,^6,48 BL,¹³ and FL^49,50 into genetic subgroups. The Tier 1 list for DLBCL currently contains 126 protein-coding genes, which represents the lower bound for the full set of drivers (supplemental Table 1). The experimental demonstration of a gene or mutation represents a laborious and costly barrier to the growth of Tier 1 genes. We anticipate the promotion of additional Tier 2 genes as more genetic and functional data emerge and, in a subsequent section, discuss how this might be accelerated.

Although Tier 1 genes can be considered bona fide drivers, the genes in Tier 2 should represent only those with a reasonable chance of eventually joining Tier 1. Across the Tier 2 genes, we observed a substantial range in the quantity and quality of supporting data. To reduce the size of Tier 2, genes can be moved to Tier 3 if they meet either of 2 strict criteria. First, a gene is assigned to Tier 3 if a minority of the variants in that gene fail to meet minimal quality thresholds upon manual review of the original data.⁴¹ This is based on the rationale that, if stricter criteria were used, the gene would not likely have been nominated in the first place. Secondly, we have reduced confidence in any study whose results cannot be reproduced upon reanalysis of the same data. Accordingly, if a subsequent analysis of the same samples fails to reproduce at least half of the driver mutations reported in the initial study, all Tier 2 genes from that study are assigned to Tier 3. Thus far, this criterion has only been applied to 2 studies. More information is provided in supplemental Table 7 and the supplemental Results.^16,18 Collectively, this results in 387 Tier 3 DLBCL genes and 70 Tier 3 BL genes.

Although this practice intends to enrich Tier 2 for genes with potential merit, the contents of the Tier 3 lists should not be entirely dismissed. Although there is minimal evidence that their mutation contributes to lymphomagenesis, some may nonetheless be relevant to lymphoma biology. For example, mutations in SYK and MYB have not been confirmed but these genes respectively have established roles in B-cell receptor signaling and hematopoiesis.¹⁹ It is conceivable that the reliance of a few studies on RNA sequencing for detecting mutations could bias their gene list toward genes that are highly expressed in malignant cells.^2,14,51 We acknowledge that low mutation incidence alone is not evidence against a gene acting as a driver. TET2 and NOTCH1, respectively mutated in 5% and 7% of DLBCLs, were recognized as drivers in this entity relatively recently, likely escaping detection earlier because of this low incidence. Tier 3 is significantly enriched for genes that have an elevated rate of rare germ line variants in the human population (supplemental Results).⁵² This suggest that many of the genes in Tier 3 result from inadequate removal of germ line variants.

Importantly, some genes in Tier 3 for 1 entity represent high-confidence genes for another. For example, 11 of Tier 3 BL genes are high-confidence DLBCL genes, including NOTCH1, BRAF, CD79B, ETS1, and BTG2. Each of these were attributed to BL by Love et al or Panea et al.^14,53 The existence of DLBCL-associated mutations in BL cohorts could be explained by insufficient pathology review. Consistent with this, a classifier trained on DLBCL and BL genetics assigned a third of the patients from that study to the DLBCL group.⁵⁴ When cost is a consideration when designing custom sequencing panels, we recommend that Tier 3 genes should be considered low priority and excluded if necessary.

The gene lists for BL, FL, and DLBCL naturally contain significant overlap because many of the same genes are mutated in multiple B-cell neoplasms.⁵⁵ In particular, it has long been recognized that FL and germinal center B-cell–like DLBCL share many driver mutations.^2,20,49 For example, EZH2 hot spot mutations are a feature of FL and DLBCL but are rarely found in any of the other common B-cell lymphomas.⁵⁴ In contrast, BRAF is commonly mutated across a variety of solid tumors including epithelial cancers,⁵⁶ whereas these mutations are relatively rare across the mature B-cell neoplasms, with the exception of hairy cell leukemia. Interestingly, there are more documented drivers in DLBCL than any other mature B-cell neoplasm, perhaps reflecting that DLBCL represents an umbrella term comprising at least 5 to 7 subtypes with distinct molecular underpinnings, most sharing some biological features with another B-cell malignancy.⁵⁷ Accordingly, the number of genes that are unique to DLBCL as a whole appears low.

Despite sharing many drivers with DLBCL, the FL list is substantially shorter, comprising 54 Tier 1 genes and 58 Tier 2 genes. This might partially be a consequence of fewer large-scale studies that have focused on this pathology. In just the past few years, multiple studies focusing on FL have emerged (Figure 2) and we anticipate that additional such studies will allow the more recent Tier 2 genes to be promoted to Tier 1. We also expect additional Tier 1 DLBCL genes to eventually enter the FL list. The separate gene list for FL is justified, in part, for genes that are scantily mutated in DLBCL and unlikely to be attributed to both entities (Figure 3). Currently, this is limited to MAP2K1, ATP6AP1, ATP6V1B2, CTSS, and VMA21. MAP2K1 mutations are limited to pediatric-type FL,⁵⁸ whereas the others are seen in adult FL. For completeness, we anticipate that researchers studying either FL, BL, or DLBCL may prefer to pool the Tier 1 genes from all lists (supplemental Table 1), which is only slightly longer than the DLBCL list (supplemental Table 2) because of overlap with the others (supplemental Tables 3 and 4).

Although having been the subject of multiple whole-genome sequencing,^12,54 whole-exome sequncing,^14,15,53 and targeted sequencing studies,^62,65 BL currently has only 34 Tier 1 genes (Figure 2) with another 77 genes residing in Tier 2. Some of the earliest exome-based studies nominated a substantial number of genes but most have since been assigned to Tier 3 (Figure 2; supplemental Table 4).⁵³ As the most recent genome-wide studies have individually contributed a small number of new genes, the Tier 1 BL list may be approaching saturation (supplemental Table 5). This would suggest that BL is more genetically homogeneous than FL or DLBCL. Interestingly, more than a third of the BL genes overlap with DLBCL and FL, including KMT2D, FOXO1, GNA13, TP53, CCND3, ARID1A, and SMARCA4. There is further sharing of driver genes between Tier 1 BL genes and DLBCL that does not extend to FL, such as DDX3X, HNRNPU, SIN3A, and RHOA. Some of these likely represent mutations that are only relevant in conjunction with other common BL drivers. For example, DDX3X mutations protect against the collateral proteotoxic stress experienced by cells harboring IGH::MYC.⁶⁶ Only 6 of the Tier 1 BL genes are not shared with either the FL or BL Tier 1 lists whereas many more Tier 2 BL genes overlap the Tier 1 DLBCL list (Figure 3).

We noted large disparities in the sampling of Tier 1 genes across panel-based studies. For example, the panel used by Burkhardt et al included only 22 of the Tier 1 BL genes, with notable exclusions such as KMT2D, HNRNPU, and USP7.⁶² Similarly, Crouch et al sequenced 44 of the 54 Tier 1 FL genes, excluding important genes such as ATP6AP1, ATP6V1B2, and CTSS.⁵⁰ In designing a general panel for application to any of these 3 entities, we suggest the inclusion of the 144 genes that exist in at least 1 of the Tier 1 lists (supplemental Table 1). Importantly, none of the panel-based studies sequenced all 144 Tier 1 genes.⁶⁴ The panels that included the largest fraction of these genes were used in Krysiak et al and Russler-Germain et al (supplemental Table 6).^31,32 

The recurrent mutation of a gene within a disease entity is suggestive that the altered function of the gene product contributed to initiation or progression of malignancy. However, the strongest evidence of a driver comes from experimental demonstration of a lymphoma-driving impact on cell behavior or fitness. Although the list of drivers may seem manageable, with only 144 Tier 1 genes across BL, DLBCL, and FL, we published studies detailing the function for a mere 58 of these (Figure 4). Functional characterization of drivers is complicated by the fact that most genes have many different recurrent mutations, not all of which can be considered functionally equivalent. Moreover, the consequence of any mutation may involve genetic interactions, which can only be revealed in combination with specific alterations in other genes. Finally, to exert their relevant effect, the effect of some mutations may require modeling within a specific developmental stage of B-cell biology and/or microenvironmental setting. The importance of modeling each mutation in the appropriate biological context is supported by contrasting the mutation patterns in the same gene between entities. For example, the dominant class of mutations in CREBBP is distinct between DLBCL and FL, with the latter predominantly harboring missense rather than truncating mutations.⁴⁹ Although this observation implies different selective pressures, it remains unclear whether this is because of different microenvironmental pressures or cellular context of the premalignant cells (cell of origin). Biological sex is another variable that must be considered, as illustrated by the distinct mutation patterns seen in DDX3X among male and female patients with BL.⁶⁷ 

Understanding the mechanistic contribution of each driver gene to lymphoma biology and deciphering this complex network of cooperating mutations is an essential challenge we must overcome if we are to exploit the full potential of genomic medicine in lymphoma. The evidence required for this functional annotation may come in many forms, typically starting by drawing inferences from the mutation pattern. For example, genes with a preponderance of truncating mutations are presumed to be tumor suppressors whereas those with mutation hot spots are expected to exhibit gain of function. Among the lymphoma genes, counter-examples of each of these have been documented.⁹⁷ The strongest support, sufficient to validate a driver, is established using an experimental system that allows quantification of a biological phenotype after the introduction of a defined genetic perturbation in an experimentally controlled manner. The technology available for such experiments is constantly evolving, allowing many genetic perturbations to be interrogated in high throughput in multiple different cellular contexts. Scalability is paramount when we consider the goal of modeling hundreds or thousands of mutations in combinations of 2, 3, or 4, the number of permutations to model rapidly becomes very large.

Perhaps the most direct evidence of a lymphoma-promoting function of an individual mutation is provided by using genetically modified mouse models. Intercrossing allows multiple mutant alleles to be combined. These models have the advantage of recapitulating the immune microenvironment, and conditional gene targeting allows introduction as specific stages of B-cell development. Numerous such models have demonstrated the ability of specific genetic combinations to drive transformation of B cells to diseases that mimic human lymphomas.^101-105 However, transgenic mouse models are costly, slow to generate and unsuited to high-throughput analysis. Moreover, the requirements for lymphocyte transformation in the mouse do not always faithfully recapitulate the situation in human cells.

An alternative approach has been to use human lymphoma cell lines.¹⁰⁶ These are ideally suited to high-throughput screening technologies such as genome-wide CRISPR screening and have proven a workhorse tool for mechanistic investigation of lymphoma functional genomics. However, cell lines do not exist for some lymphoma types. For instance, all cell lines associated with FL come from tumors that have transformed to DLBCL. The absence of lines representing untransformed FL may reflect the essential dependence of this lymphoma type upon the microenvironment. Moreover, cell lines may have been growing for decades during which time they have evolved a complex and skewed repertoire of genetic mutations, that is already perfectly optimized for in vitro growth. This limits the capacity to test the introduction of new mutations on a specific genetic background, but also makes it hard to test the ability of any genetic perturbation to increase the fitness of the cell line. To overcome these issues, we developed a novel coculture system.^107,108 By mimicking the survival signals provided by the lymph node microenvironment we were able to support the growth and proliferation of normal, nonmalignant germinal center B cells purified from discarded pediatric tonsillectomy tissue. Although these cells cannot be transduced with conventional vesicular stomatitis virus G pseudotyped virus, the use of a gibbon ape leukemia virus and murine leukemia virus fusion envelope allowed high-efficiency transduction with both lentivirus and retrovirus. This allows for efficient genetic modification of nonmalignant human GC B cells and the potential to overexpress or knock out by CRISPR any desired putative cancer driver gene, in its wild type or mutant form and into a genetically normal background or in any desired genetic combination. Moreover, this system is scalable and suited to high-throughput screening using targeted or whole genome CRISPR libraries, or over expression of wild type or mutant open reading frame libraries. Established lymphoma cell lines are most amenable to dropout screening, which identifies genes essential for survival of malignant B cells in the context of an existing set of somatic mutations.⁹⁹ In contrast, the use of primary cells allows for the controlled introduction and identification of genetic perturbations that afford a gain of fitness to the cell, providing an ideal system for the assessment of putative lymphoma driver genes. However, this powerful system is not without limitation, most prominently the inability to model interaction with the host immune system. In an attempt to model this microenvironmental interaction several groups have established autologous coculture systems from normal¹⁰⁹ or malignant lymphoid tissue¹¹⁰ to establish in vitro organoid models of the germinal center or of B-cell lymphomas. These have not yet been combined with genetic modification but may provide a future strategy to assess the potential of lymphoma driver genes to modify interaction with the host immune system.

The technology for genetic editing and phenotypic interrogation continues to see rapid improvements. CRISPR provides a strategy to mimic loss of function mutations by knockout of the gene, which is routinely done at genome scale. Single-nucleotide variants can be introduced by targeted CRISPR editing to recapitulate specific mutations, but this is relatively inefficient and laborious. In contrast, newer technologies such as base editing now allow near nucleotide-level interrogation of individual nucleotide changes in high throughput in cell lines or primary B cells.¹¹¹ Improvements in single-cell RNA sequencing now allow hundreds of thousands of cell transcriptomes to be profiled in a single experiment. These can be combined with barcoded perturbations such as CRISPR knockout, base editing, or open-reading-frame libraries to profile the transcriptional impact of many hundreds of genetic perturbations in a single experiment, providing mechanistic insight into the functional impact of each genetic change beyond its impact on competitive fitness.

Each experimental strategy balances advantages and limitations, but together these model systems offer a complementary set of tools to interrogate the function of putative lymphoma genes, at the resolution of individual nucleotide changes. They provide the tools to do this at scale and to screen for synergy between cooperating mutations. As techniques for genetic editing, cell barcoding and phenotypic interrogation continues to advance, we hope to witness an acceleration in our understanding of the role and function of lymphoma drivers.

The authors are grateful to the courageous patients who selflessly donated their valuable samples to further cancer research.

Contribution: R.D.M. and D.J.H. conceived of the study; K.M.C., R.D.M., and D.J.H. wrote the manuscript; and K.D. and R.D.M. created the figures and tables.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Ryan D. Morin, Simon Fraser University, 8888 University Dr, Burnaby, BC V5A 1S6, Canada; email: rdmorin@sfu.ca.