Discovery and prioritization of somatic mutations in diffuse large B-cell lymphoma (DLBCL) by whole-exome sequencing

We performed solution-phase hybrid capture and whole-exome sequencing on paired tumor and germline (i.e., normal) DNA samples from 55 patients with primary DLBCL. We achieved 150-fold mean sequence coverage of targeted exonic regions, with an average of 97% of bases covered per patient (range, 91–98%). Such high coverage is important because tumor samples are often contaminated with normal cells (e.g., fibroblasts, immune cells), which can obscure the identification of somatically mutated alleles unless very deep coverage is achieved.

We excluded six samples from further analysis because of extremely low apparent mutation rates, most consistent with extensive stromal contamination. Of the remaining 49 patients, the mean nonsynonymous mutation rate was 3.2 mutations per megabase, with mutation rates varying widely (0.6–8.7 mutations per megabase), which is higher than the estimated mutation rate in other hematopoietic malignancies, such as chronic lymphocytic leukemia and other leukemias (<1 per megabase) (13, 14), and multiple myeloma (1.3 per megabase) (15) (Fig. 1A). There was no correlation between patient-specific mutation rate and frequency of particular types of mutation (e.g., *CpG→T, Cp(A/C/T)→T, A→G, transversions, or indels; Fig. 1C). Similarly, there was no correlation between observed mutation rate and average allelic fraction of those mutations observed in each patient, suggesting that the variation in mutation rate was not simply a function of variability in extent of stromal contamination (Fig. S1).

To define significantly mutated genes in DLBCL, we applied the MutSig algorithm to identify genes harboring mutations at a higher frequency than expected by chance (15, 16). To better estimate the significance of observed mutations, MutSig takes into account (i) the sample-specific mutation rate, (ii) the ratio of nonsynonymous to synonymous mutations in a given gene, and (iii) the median expression level of each gene in DLBCL [based on gene expression profiling datasets (7)]. This approach revealed 58 statistically significant genes with a false discovery rate cutoff of 0.1 (q₁ ≤ 0.1; Fig. 1B, Table 1, and Tables S1 and S2). We independently validated selected mutations by targeted resequencing in a subset of patients and obtained 97.9% validation rate (Table S5).

Among the significantly mutated genes were those for which a functional role in the pathogenesis of DLBCL has accumulated over many years. These include CD79B, TP53, CARD11, MYD88, and EZH2 (8–11). In addition, we discovered mutations in genes for which a pathogenic role in DLBCL has been suggested recently (17, 18). These include MLL2, TNFRSF14, BTG1, MEF2B, and GNA13, which are discussed in greater detail later.

We discovered mutations in genes not previously recognized as drivers of cancer. For example, we found β-actin (ACTB) mutations in five patients (Fig. 2). Actins are highly conserved proteins of the cytoskeleton and are involved in B lymphocyte activation (19). By using the xvar algorithm (based on evolutionary conservation and the nature of the observed amino acid change), the predicted functional consequence of the ACTB mutations observed in DLBCL is high. We also note a preponderance of ACTB mutations toward the amino terminus of the protein, but the functional significance of this remains unknown.

Another unexpectedly recurrently mutated gene is P2RY8, encoding a G protein-coupled purinergic receptor whose normal function has not been extensively characterized (20). P2RY8 is most notable for its involvement in a chromosomal translocation with CRLF2 in 7% of patients with B-progenitor acute lymphoblastic leukemia (ALL) and 53% of individuals with ALL and Down syndrome (21). The presumed mechanism of action of such P2RY8/CRLF2 fusions is activation of CRLF2 by its coming under the control of the P2RY8 promoter, which is highly active in B lymphoid cells (21). However, we note that P2RY8 has itself been reported to function as an oncogene in experimental models (22). In DLBCL, we identified six patients with coding mutations in P2RY8, two of whom harbored two mutations (Fig. 2). In three patients, the observed allelic fraction of these mutations is greater than 0.5, suggestive of deletion of the WT allele or amplification of the mutant allele. The functional consequence of P2RY8 mutation in DLBCL remains to be determined.

We also observed very common mutations in the gene PCLO (Piccolo), encoding a protein that functions as part of the presynaptic cytoskeletal matrix thought to be involved in regulating neurotransmitter release (23). A role for PCLO in calcium sensing has also been suggested (24), but a role in cancer has not been reported. We found a total of 23 nonsynonymous mutations in 17 patients (35%; Fig. S2), but the observed ratio of nonsynonymous to synonymous mutations (23:3) is consistent with that expected by chance, given that most of the observed mutations (12 of 23) are transversion mutations, which favor nonsynonymous outcomes by a ratio of nearly 5:1. It is thus possible that, for unknown reasons, the local rate of mutation at the PCLO locus is unusually high, giving rise to passenger mutations of no functional consequence in DLBCL. Additional work is clearly needed to resolve the role, if any, of PCLO mutations in DLBCL and other cancers.

Histone 1 (H1) family proteins are linker histone proteins that bind to the DNA entering and exiting the nucleosomal cores. Different forms are expressed at different stages of the cell cycle in different tissue types and are involved in chromatin compaction and possibly transcription (25). We observed a striking accumulation of mutations in H1 family proteins, with 59 nonsynonymous and 35 synonymous mutations among 31 histone H1 proteins in 34 patients (69%; Table S2). The functional significance of these mutations remains to be explored, but hotspot analysis as outlined later suggests that HIST1H3B and possibly other core histone proteins are subject to activation-induced cytidine deaminase (AID)–mediated SHM.

We also identified mutations in genes that were recently reported to be significantly mutated (17, 18). MLL2 is a histone methyltransferase of the SET1 family that is responsible for histone H3-lysine 4 trimethylation (H3K4me3) during oogenesis and early development (26). Inactivating mutations have been reported in medulloblastoma (27) and multiple myeloma (15), and chromosomal translocations involving the MLL family member MLL1 are well described in acute leukemias (28). The MLL2 mutations we observed in DLBCL are highly biased toward truncating events, the large majority being nonsense mutations and frameshift-inducing insertions and deletions (Fig. 2). As has been suggested previously, our data suggest that MLL2 may function as an important tumor suppressor in DLBCL (17, 29).

TNFRSF14, also known as LIGHT Receptor, belongs to the TNF-receptor superfamily most extensively studied in T cells. Interestingly, it can convey opposing signals based on its specificity for diverse ligands. In our data, five of nine mutations suggest loss of function (n = 4 nonsense mutations and n = 1 frameshift deletion), with an additional in-frame insertion and three missense mutations. No synonymous mutations were seen (Fig. 2). As has been suggested, these results strongly suggest a tumor-suppressive role of TNFRSF14 in DLBCL (17). It has been proposed that LIGHT-mediated triggering of TNFRSF14 renders B-cell lymphomas more immunogenic and sensitive to FAS-induced apoptosis (30)—a potential mechanism by which TNFRSF14 can act as a tumor suppressor.

Mutations in BTG1 were also observed to be relatively common (15 nonsynonymous and two synonymous mutations). BTG1 belongs to the BTG/Tob family of proteins that regulate cell cycle progression in a variety of cells. BTG1 is thought to confer DNA binding of sequence-specific transcription factors (31). Whether mutations in BTG1 result in aberrations in chromatin structure (as with MLL-family mutations) as opposed to nonhistone targets remains to be determined. Interestingly, we observed several patients with more than one mutation in BTG1, including two patients with three mutations and one patient with four mutations. In the latter patient, two sets of adjacent mutations (L37L plus L31F and P3R plus M1I) were never found in the same sequencing read, suggesting that they occurred on different alleles, i.e., in trans. We note that two patients had the identical silent mutation at codon 31, suggesting that this mutation, although it did not affect protein coding sequence, may alter codon use or mRNA stability. Overall, the nature of BTG1 mutations does not clearly point toward a gain-of-function or loss-of-function mechanism, although the biallelic involvement seen in some patients tends to favor loss of function (Fig. 2).

MEF2B mutations were observed in 18% of patients with DLBCL, similar to data reported recently (17, 18), predominantly in the MADS box or MEF2 domains. MEF2B belongs to a family of calcium-regulated transcription factors that recruit histone-modifying enzymes. Further supporting an oncogenic role for MEF2 proteins, MEF2C has been identified as a T-cell acute lymphoblastic leukemia oncogene that is activated by chromosomal rearrangement (32).

A hallmark of cancer genes (i.e., genes that harbor “driver” mutations, which contribute to the formation and progression of cancer) is the preponderance of nonsynonymous mutations compared with synonymous mutations (typically with an expected ratio of ∼2.8:1). Curiously, we observed a striking opposite effect in the BCL2 gene—a known driver in some DLBCLs. We observed a very high mutation rate in BCL2, but with a depletion of nonsynonymous mutations, with 18 nonsynonymous mutations and 28 synonymous mutations, for a ratio of 0.64, far below that expected by chance (P = 8.8 × 10⁻⁶; Fig. 3). We hypothesized that this phenomenon might be caused by two processes: first, a locally high mutation rate at the BCL2 locus occurring via SHM, and second, negative selection against functionally deleterious mutations.

We first explored the possibility that the high rate of BCL2 mutations observed in our patients might be related to SHM. This physiologic process in B lymphocytes has been best studied at Ig gene loci, where it facilitates the development of antibody diversity (33), but other genes have been shown to be subject to aberrant SHM (including the PIM1 gene, in which we also observed hypermutation with 30 mutations in 16 patients; Fig. S2) (34). Chromosomal translocations of the BCL2 gene into the Ig heavy chain (IgH) locus occur in approximately 20% of DLBCL tumors, and the resulting BCL2 overexpression provides an antiapoptotic signal to the lymphoma cells (35). We hypothesized that BCL2 hypermutation in our patients might be explained by the BCL2 gene adopting the IgH locus's normal process of SHM as a result of the translocation (36). If this hypothesis were correct, those tumors with elevated BCL2 mutation rates would be expected to also harbor BCL2/IgH translocations. We tested for BCL2/IgH translocation in 26 patients with DLBCL (13 with BCL2 mutation and 13 others selected randomly). As predicted, the vast majority of patients with BCL2 hypermutation (10 of 13, 77%) had BCL2/IgH translocation, whereas only one of 13 patients (8%) lacking BCL2 mutation had the translocation (P = 0.0005, one-sided Fisher exact test; Fig. S3). To determine whether BCL2 or other genes may be targets of SHM, which is mediated by AID, we asked whether mutations occur preferentially in the context of WRCY motifs, which are known target sequences of AID. This analysis indeed revealed that BCL2, as well as other genes, including PIM1, have significant enrichment of mutations in WRCY motifs (Table S3).

Although these results likely explain the hypermutation at the BCL2 locus, they do not explain the preponderance of synonymous mutations observed in these patients. We hypothesized that they represent a vestige of negative selection against deleterious mutations in BCL2, on which the tumor cells are dependent for survival. If this hypothesis were correct, we would expect that any nonsynonymous mutations would be confined to functionally nonessential domains of the protein, whereas the synonymous mutations would not. To address this, we focused on the BCL2 BH domains that mediate interactions with proapoptotic proteins (37). We observed 18 nonsynonymous mutations, but only three of these fell within BH domains, whereas 15 of 28 synonymous mutations fell within BH domains (Fig. 3). We used permutations to determine the probability of nonsynonymous mutations being located within versus outside BH domains, corrected for domain size. To increase the power of this analysis, we included the BCL2 mutations observed in the present study and those recently reported (17). This analysis indicated a significant enrichment of nonsynonymous mutations falling outside of BH domains (P = 0.041). These results argue that the nonsynonymous mutations preserve BCL2 function by avoiding critical BH domains. On the contrary, we found the synonymous mutations to be preferentially located within BH domains (P = 0.02). This can be explained by nonuniform distribution of these mutations along the gene (P = 0.045) with preferential clustering at the 5′ end. This is consistent with previous demonstrations of aberrant SHM preferentially occurring closer to the 5′ end (34).

Our results are most consistent with purifying selection (i.e., negative selection) against mutations that inhibit or decrease BCL2 function. Such mutations likely result in loss of or impaired tumor cell viability, leaving behind only those mutations that are synonymous or that affect nonrelevant domains. Interestingly, we note the presence of recurrent synonymous mutations at codons N11 and K22 (Fig. 3), suggesting a preference for certain sites to acquire mutations. Whether there is a functional consequence of these synonymous mutations, for example by affecting mRNA stability, remains to be determined.

The preceding sections focused on the identification of cancer genes based on their mutation frequency across the whole genome. However, additional genes may harbor functionally important mutations that fall short of statistical significance (in the standard mutation frequency test) given the modest size of our dataset. Some of these genes are recognizable based on their known importance in the pathogenesis of other malignancies. For example, we observed two mutations in KRAS (G13D), four mutations in NOTCH1, and two mutations in BRAF, often involving specific amino acids documented to be sites of recurrent mutation in the Catalogue of Somatic Mutations in Cancer (COSMIC) database. It therefore seems likely that these mutations similarly play a causal role in DLBCL, albeit at a low frequency. To perform this search in a rigorous manner, we also tested each gene whether it has an increased mutation rate only in sites previously reported in the COSMIC database. Table S4 shows a list of the most significant genes identified by this test.

We also sought to prioritize rare mutations based on the clustering of mutations within particular regions of the coding sequence. Mutations occurring by chance (i.e., passenger mutations) would be expected to be scattered randomly across the protein, whereas functionally consequential mutations may cluster within critical domains. First, we assumed that mutations have a higher likelihood of functional impact if they are found at the same site or in close proximity to each other. For example, highly clustered mutations are observed in MYD88, CARD11, CD79B, and EZH2 (Fig. S4). Second, we assumed that mutated genes are more likely to be functionally relevant if the affected amino acid residue is more conserved across species. Our algorithm thus returns a list of mutated genes that is rank-ordered by a composite score reflecting sequence conservation and clustering of mutation sites (Table 1). This analysis prioritized genes known to be mutated in DLBCL and to play a role in its pathogenesis and also novel genes that would not otherwise be considered as likely drivers. For example, the tyrosine kinase SYK was identified as a likely driver based on these criteria, and it has recently been demonstrated that SYK inhibitors have significant clinical activity in non-Hodgkin lymphoma, including DLBCL (38, 39). Similarly, the serine/threonine kinase SGK1 was identified by this analysis despite its not reaching statistical significance based on frequency alone (q₁ > 0.1), but a recent study confirms the recurrent nature of SGK1 mutation in DLBCL (17).

A total of 55 patients with DLBCL provided DNA for this study. This study was reviewed and approved by the human subjects review board of the Mayo Clinic, the University of Iowa, and the Broad Institute, and written informed consent was obtained from all participants. DNA was extracted from lymph node samples (tumor) and blood (normal) as previously described and processed as detailed in SI Materials and Methods.

Massively parallel sequencing data were processed by using two consecutive pipelines developed at the Broad Institute: “Picard” generates a single BAM file representing the sample; “Firehose” starts with the BAM files for each DLBCL sample and matched normal sample from peripheral blood (hg19) and performs various analyses (15, 16). We evaluated the fraction of all bases suitable for mutation calling whereby a base is defined as covered if at least 14 and eight reads overlapped the base in the tumor and in the germline sequencing, respectively. Subsequent analysis is described in more detail in SI Materials and Methods.

A PCR assay was used for detection of the t(14;18) translocation, which targets the joining region of the IgH gene and distinct regions of the BCL2 locus (InVivoScribe Technologies). Details are provided in SI Materials and Methods.