New insights into the biology and origin of mature aggressive B-cell lymphomas by combined epigenomic, genomic, and transcriptional profiling

Aberrant DNA methylation is a hallmark of cancer. Virtually all cancer types are associated with alterations of the methylome. These include global DNA hypomethylation, mostly targeting DNA repeats, and hypermethylation of CpG islands located in the promoter regions of tumor suppressor genes.^1–4  It is widely accepted that tumor suppressor gene inactivation by DNA hypermethylation allows the tumor clone to obtain a selective (eg, proliferative) advantage. However, recent reports have provided evidence for an instructive mechanism behind aberrant DNA methylation in cancer, which might indicate that specific sequences are predisposed to acquire epigenetic alterations.^5–9  Remarkably, 3 independent reports have recently shown that a highly significant proportion of genes becoming hypermethylated in cancer were already repressed at the embryonic stem cell (ESC) stage by polycomb group (PcG) marks.^7–9  These findings are considered to support the “cancer stem cell theory” in which aberrant epigenetic changes of PcG target genes occurring in a cell with stem cell features might represent the initial event in tumorigenesis.^2,6,10 

The current concept of tumorigenesis in the lymphatic system seems to contradict the stem cell origin of cancer. The model of B-cell lymphomagenesis proposes that transformation is driven during B-lymphocyte ontogenesis (eg, during the transit of B cells through the germinal center) through the acquisition of recurrent, so-called primary, genetic aberrations, such as chromosomal translocations affecting the immunoglobulin heavy chain (IGH) locus.^11,12  For instance, the translocation t(8;14)(q24;q32) places the oncogene MYC in 8q24 under the control of regulatory regions of the IGH locus in 14q32. As a result, MYC expression becomes deregulated and the cell loses its normal differentiation and proliferation program.¹²  However, the sole presence of this translocation is not sufficient to induce lymphomagenesis, and additional genetic hits, so-called secondary genetic changes, are necessary for lymphoma development and progression.¹² 

Different lymphoma subgroups are supposed to reflect various stages of normal lymphocyte differentiation. In line with this, different lymphomas, for example, mature aggressive B-cell non-Hodgkin lymphomas (maB-NHLs) such as Burkitt lymphoma (BL) and diffuse large B-cell lymphoma (DLBCL), display typical morphologic, immunophenotypic, and gene expression features associated with the lymphocyte maturation step they resemble. According to the current genetic model of lymphomagenesis, the frozen differentiation stage might reflect the cell type in which the primary translocation took place leading to maturation arrest. An alternative hypothesis is that cells acquiring primary chromosomal translocations are able to further differentiate, and that maturation is arrested later in the differentiation process.

Classical methods to diagnose and classify lymphomas are based on cell and tissue morphology as well as the expression of immunophenotypic markers. Using gene expression profiling (GEP), morphologic subtypes of lymphoma have been further characterized allowing thus a more objective diagnosis.^13,14  In addition, GEP studies have identified new lymphoma subgroups based on their molecular signature.^14–19  For instance, DLBCL comprises at least 3 gene expression subgroups, for example, germinal center B cell–like (GCB), activated B cell–like (ABC), and primary mediastinal B-cell lymphoma. Additional studies have used GEP to define the boundaries between maB-NHLs such as BL and DLBCL.^16,17  In a recent study, we combined genetic and transcriptional profiling to propose a biologic definition of BL, which was based on a GEP index and the presence of chromosomal translocations fusing the MYC gene to an IG locus.¹⁷ 

The aim of the present study was to analyze patterns of DNA methylation of cancer-associated genes in maB-NHL. To that end, we performed a microarray-based DNA methylation study of different maB-NHL subtypes characterized by genome-wide genetic and transcriptional profiling, as well as in lymphoma cell lines and control samples.

Our results provide evidence that DNA methylation profiles of maB-NHL subtypes are not strictly associated with phenotypic or molecular features. Furthermore, the groups of genes commonly and differentially hypermethylated among various subtypes of maB-NHL show distinct biologic features, including enrichment for PcG targets in ESCs. Our findings indicate that various morphologic, genetic, and transcriptional subtypes of maB-NHLs share a stem cell–related epigenetic pattern. Thus, different maB-NHLs may all derive from precursor cells with stem cell–like features that have acquired aberrant DNA methylation of PcG target genes. Alternatively, such stem cell–like features could have been acquired in the process of lymphomagenesis by epigenetic remodeling through, for example, chromosomal changes.

DNA samples from 83 maB-NHLs were analyzed. These samples have previously been morphologically classified by a panel of expert hematopathologists and characterized by genome-wide genetic and transcriptional profiling.¹⁷  Lymphomas were heterogeneous with regard to current diagnostic parameters, including cell morphology (eg, BL vs DLBCL), chromosomal aberrations (presence or absence of translocations such as IG-MYC and IGH-BCL2 fusions, and different chromosomal imbalances), and gene expression signatures (molecular BL [mBL] and non-mBL/DLBCL of the activated B-cell [ABC] and germinal center B-cell [GCB] subtypes). A summary of the patient features is shown in Table 1 and Table S1 (available on the Blood website; see the Supplemental Materials link at the top of the online article). In addition, we studied 7 maB-NHL lymphoma cell lines (Table S2A), 10 hematopoietic control samples (Table S2B), and 6 adult stem and progenitor cells (ASCs), which included 2 cell lines each from mesenchymal stem cells (MSCs) and 2 multipotent adult progenitor cells (MAPCs) as well as 2 isolated CD34⁺ cell samples (Table S2C). The study was performed as part of the network project “Molecular Mechanisms in Malignant Lymphomas (MMML),” for which central and local ethics approval was obtained from University Hospital Schleswig-Holstein.

We used the GoldenGate Methylation Cancer Panel I (Illumina, San Diego, CA) for DNA methylation analyses as previously described.²⁰  The panel is developed to assay 1505 CpG sites selected from 807 genes, which include oncogenes and tumor suppressor genes, previously reported differentially methylated or differentially expressed genes, imprinted genes, genes involved in various signaling pathways, and genes responsible for DNA repair, cell- cycle control, metastasis, differentiation, and apoptosis.

To validate the DNA methylation data generated by BeadArray technology, conventional methylation analyses such as methylation-specific polymerase chain reaction (MSP) and bisulfite sequencing (BS) were performed as previously described.^21–24  Bisulfite pyrosequencing (BPS) was performed according to standard protocols with brief modifications²⁵  and evaluated with the Pyro Q-CpG 1.0.9 software (Biotage AB, Uppsala, Sweden). The primer sequences as well as product length and annealing temperature used in MSP, BS, and BPS polymerase chain reactions (PCRs) are shown in Table S3.

CpGs located on imprinted genes and X-chromosomal genes with sex-specific methylation were excluded from the analysis. To select sex-specific CpGs, methylation values on chromosome X were compared between sexes for a subset of arrays with no copy number aberrations on chromosome X (n = 32). CpGs that showed a mean difference between sexes of at least 0.1 and P value less than .001 (Wilcoxon test) were marked as sex specific and therefore excluded from further analyses. Imprinted genes were identified from publicly available databases (http://igc.otago.ac.nz/home.html²⁶  and http://www.geneimprint.com/site/genes-by-species²⁷ ) and a published review.²⁸ 

Agglomerative hierarchic clustering was performed on maB-NHL and healthy hematopoietic controls excluding CpGs located in imprinted genes and X-chromosomal genes with sex-specific methylation (1343 CpGs and 741 genes were included in the analysis). For clustering arrays and CpGs, Euclidean distance and complete linkage were used.

A detailed explanation of the statistical analyses used to define the different DNA methylation subgroups is shown in Document S2.

Proportions of PcG marks and promoter classes in different methylation groups were compared using the closed test procedure based on the generalization of Fisher exact test to multiple groups. A genome-wide mapping of Polycomb target genes in ESCs was available as supplemental material²⁹  and was kindly provided by Dr A. Bracken and Dr K. Helin (University of Copenhagen, Copenhagen, Denmark) for embryonic fibroblasts.³⁰  To analyze whether promoter regions of the methylation groups showed different CpG compositions, we used a recently described classification into promoters with high (HCP), intermediate (ICP), and low (LCP) CpG content.³¹  A genome-wide list of classified gene promoters is available from one of the authors (D.S.). Annotation lists of PcG marks and promoter classes were compared with the genes analyzed for methylation via gene symbol or locuslink ID.

Raw datasets from U133A and U95A Affymetrix gene chips (Santa Clara, CA) were processed using R statistical software (http://www.R-project.org, Vienna University of Economics and Business Administration) in conjunction with the Bioconductor open source software (Fred Hutchinson Cancer Research Center, Seattle, WA).³²  We computed “absent,” “marginal,” or “present” detection calls of each probe set and microarray sample according to the Affymetrix Microarray Analysis Suite version 5.0 (MAS).³²  For that purpose, we used the implementation of the detection call algorithm from Bioconductor's Affy-package (version 1.12.0)^32,33  on raw data with standard parameters. Table S4 contains a list of the lymphoma and control samples (including GEO identifiers and references³⁴ ) used to generate gene expression calls, which are summarized in Table S5.

A hypergeometric test procedure³⁵  was used to determine enriched Gene Ontology terms³⁶  within the selected methylation groups. The genes of each group were considered as a subset of all genes from the array, except that all imprinted and sex-specific genes on chromosome X were excluded as described in “Methods.”

We have quantified the methylation status of 1505 individual CpGs from 807 promoters of genes involved in cancer in 106 samples using the BeadArray technology.²⁰  The complete dataset is provided in Document S3.

A total of 98 array-based DNA methylation analyses were performed in duplicate. The high reproducibility of the technique was demonstrated by calculating coefficients of determination (R²), which showed a median value of 0.991 (IQR = 0.985-0.993) (Figure S1A). Representative scatterplots are shown in Figure S1B. In addition, MVA (variability [M] as a function of mean [A]) plots were calculated. The median standard deviation (SD) of MVA plots was 0.034 (IQR = 0.029-0.043), whereas the median bias was 0.0002 (− 0.003 to 0.005; Figure S1A). As the technical duplicates agreed well with each other, the mean of pair-wise beta values was calculated and used for further statistical analyses. As global DNA methylation data per case shows a bimodal distribution in which less than 0.25 defines the unmethylated CpGs and more than 0.75, the methylated CpGs, beta values below 0.25 and above 0.75 were selected as cutoffs to define unmethylated and methylated CpGs (Figure S1D).

Microarray-based DNA methylation data were validated in 2 different ways. First, the accuracy of the DNA methylation values generated by the array (beta values, ranging from 0 for unmethylated to 1 for completely methylated) was studied by BS and BPS. Second, MSP, BS, and BPS were applied to study whether the CpGs analyzed with the methylation-specific array were representative for the methylation status of a given promoter region.

Three CpGs showing variable DNA methylation (beta) values in the array analyses were selected for validation by BS: INSR_P1063_R, IGSF4_P454_F, and CCND2_P898_R. A total of 10 clones from 4 cases with different methylation values were sequenced per CpG region. INSR was additionally sequenced in 2 negative controls (peripheral blood and unmethylated DNA), one methylated control, and the cell line Raji. A significant positive correlation was observed (Spearman correlation coefficient = 0.873, P < .001) between beta values and data obtained by BS in 4 primary cases (3 CpGs each) and the cell line Raji (1 CpG; n = 13; Figure S2A). Likewise, a significant correlation between beta values and BPS-based DNA methylation analyses was observed for the CpG CCND2_P898_R (n = 44, Spearman correlation coefficient = 0.913, P < .001; Figure S2B). These data confirm the accuracy of the BeadArray to quantify DNA methylation levels.

Due to the impracticability of validating whether all the CpGs on the array were representative for the associated promoter regions, a selection of 13 genes commonly methylated in cancer was investigated by widely used MSP assays²¹  in the 7 B-cell lymphoma cell lines. As shown in Figure S3, quantitative BeadArray and qualitative MSP methylation data are highly concordant (median of beta values for genes unmethylated by MSP was 0.05, whereas for genes methylated by MSP it was 0.91). Few discrepancies were observed (Figure S3C,D), which is not surprising considering the different location of PCR primers and the CpG investigated in the BeadArray. BS in 3 specific CpGs from the BeadArray also shows that their methylation status was representative for the surrounding CpGs (Figure S4). The same was observed by BPS of the region around CpG CCND2_P898_R (Figure S5). These data and the frequent bimodal distribution of DNA methylation observed in promoter regions of normal tissues (eg, either methylated or unmethylated)³⁷  show that individual CpGs from the array can be taken as surrogate markers for the methylation status of the respective promoter regions.

CpGs from X-chromosomal genes with sex-specific methylation and known imprinted genes were removed from the analysis. These genes are partially methylated under physiologic conditions and, thus, represent a confounding variable to classify cases according to their DNA methylation profile. In fact, the sex of maB-NHLs studied herein as well as genetic imbalances of the X chromosome could be identified by the DNA methylation profile of X-chromosomal genes (Figure S6A,B).

A hierarchic cluster analysis of the remaining 1343 CpGs (741 genes) is shown in Figure 1. Morphologic features, gene expression, and genetic aberrations defined by presence or absence of t(14;18) (IGH/BCL2 fusion) and MYC translocations, as well as genomic complexity measured as number of imbalances by array-based comparative genomic hybridization (Table 1; Table S1), are shown in the upper part of the heat map. This analysis shows that DNA methylation profiling is not able to accurately differentiate maB-NHL subtypes with regard to morphology, immunophenotype, genetic aberrations, or gene expression signatures. However, a certain degree of clustering is observed in mBL. Eleven mBLs cluster together, whereas the remaining 7 are scattered throughout dataset (Figure 1).

The 741 genes under analysis have been selected for their general involvement in cancer. Therefore, one could argue that the expression of this selected set of genes might not be suitable to classify genetic and transcriptional subgroups of lymphoma. However, clustering of the very same cases as well as an extended set of maB-NHLs based on Affymetrix U133A-derived expression values (data available from 714 of the 741 genes studied with the BeadArray) allowed for discrimination of mBL from non-mBL (Figure S7A,B). In contrast, gene expression values of the genes studied in the BeadArray could not differentiate non-mBL of the ABC and GCB subtypes (Figure S7C,D). This suggests that the array is likely not sufficiently representative to evaluate whether the different subtypes of DLBCL display distinct methylation profiles. Overall, these findings indicate that DNA methylation patterns are not strictly associated with morphologic, genetic, or transcriptional features of maB-NHLs.

Using a supervised approach, we detected a total of 26 genes differentially methylated between mBLs and non-mBLs. Only 5 genes were differentially methylated between ABC- and GCB-like non-mBL.

Remarkably, we identified a larger subset of genes (n = 56) unmethylated in 10 healthy adult hematopoietic controls but methylated in all subtypes of maB-NHL regardless of the presence or absence of specific chromosomal changes. To obtain further insights into the observed methylation patterns, we classified all 741 evaluated genes into the following categories: genes (1) unmethylated in lymphomas and controls (umL/umC, n = 360), (2) methylated in lymphomas and controls (meL/meC, n = 85), (3) methylated in all lymphomas but not in controls (meL/umC, n = 56), and (4) differentially methylated in transcriptionally defined lymphoma subtypes and unmethylated in controls (dmeL/umC, n = 22) (list of genes in Table S6). Although 26 and 5 genes were detected as differentially methylated between mBL versus non-BL and ABC versus GCB, only 22 genes met the criteria to be included in the dmeL/umC group, which was caused by the more rigorous statistical approach used for the latter (Document S2). Genes commonly unmethylated (n = 8) or differentially methylated in maB-NHL and methylated in controls (n = 10; Table S6) were excluded from further statistical analyses because of the low statistical power associated with the small number of genes in each group. As highly stringent criteria were applied to classify genes into the methylation groups described in “Methods,” 474 CpGs from 358 genes could not be unambiguously assigned to any of the methylation groups and were also excluded from further analyses (Document S2).

We next investigated whether the genes in the 4 methylation categories were among those repressed by the polycomb repressive complex 2 (PRC2) in ESCs.²⁹  As shown in Figure 2A, meL/umC genes were highly enriched for loci repressed by PcG marks (OR = 8.2, P < .001), whereas dmeL/umC genes showed a weak, but still significant, enrichment (OR = 2.9, P = .03). Interestingly, the content of PcG marks was significantly different (P = .02) between both groups showing de novo methylation in lymphomas, with the meL/umC group being highly enriched for the simultaneous presence of all 3 PRC2 marks: EED, SUZ12, and 3mK27-H3 (Figure 2B). Particularly, EED target genes were notably enriched in the meL/umC group (P = .003) (Figure 2C). A still significant, but less marked, enrichment in both meL/umC and dmeL/umC groups was observed for PcG target genes in embryonic fibroblasts (Figure S8). We performed MSP analyses in maB-NHL and controls of 4 PcG target genes shown to be de novo methylated in lymphomas (ie, HOXB13, CALCA, NEFL, and PROK2), which confirmed the data obtained with the GoldenGate platform (Figure S9). These findings indicate that various subtypes of maB-NHLs exhibit a common set of de novo methylated genes, which is highly enriched for targets of PcG proteins in ESCs (Table 2).

To analyze the methylation status of genes de novo methylated in lymphomas (ie, meL/umC and dmeL/umC) in ASCs, we used the BeadArray platform to generate DNA methylation profiles of 6 ASC samples (Document S3) derived from the bone marrow of healthy donors. These included 2 cell lines each from MSCs and MAPCs, as well as 2 samples of sorted CD34⁺ cells (> 95% purity; Table S2C). Fifty-three (95%) of the 56 genes from the meL/umC group and 21 (95%) of the 22 genes from the dme/umC group were unmethylated in ASCs (Figure S10). The 4 genes (ie, EYA4, HOXA11, MT1A, and HIC1) methylated in ASCs and lymphomas but not in controls were unmethylated in CD34⁺ bone marrow cells, whose methylation profile was similar to the controls (Figure S10). Overall, these findings indicate that genes de novo methylated in lymphomas are mostly unmethylated in ASCs.

A recently published study has classified gene promoters according to their CpG density into high (HCP), intermediate (ICP), and low (LCP) CpG content.³¹  To analyze whether the promoter regions of the genes in the different methylation groups showed a specific CpG composition, we classified them into HCP, ICP, and LCP. Our analysis demonstrated that the umL/umC and meL/meC groups were enriched for HCP- and LCP-containing genes, respectively (Table 2). In contrast to the meL/meC group, genes de novo methylated in lymphomas (ie, meL/umC and dmeL/umC) were predominantly associated with the HCP subtype (Table 2; Figure 2D). This analysis demonstrates that the vast majority of the genes de novo methylated in maB-NHLs contains promoters with a high CpG content, which is in line with the concept that genes de novo methylated in cancer are frequently associated with dense CpG islands.

To study expression features of genes from the 4 DNA methylation groups in lymphoma samples, we analyzed microarray-based expression profiles of the 83 lymphoma cases studied here and additional cases studied by Hummel et al,¹⁷  as well as of published gene expression datasets from lymphomas and various normal tissues^38,39  (Figure 3; Figure S11). Using Affymetrix gene expression data from the mBLs and non-mBLs studied with the BeadArray, we observed that the percentage of absent detection calls of genes in the meL/umC group was significantly higher (P < .001) than that of genes in the umL/umC group (Figure 3). Analyses from other maB-NHL studied in Hummel et al¹⁷  (ie, additional mBLs and non-mBLs and maB-NHLs classified as intermediate; Figure S11) and from an independent series of maB-NHLs (Figure 3) show the very same effect. The different series of maB-NHLs^17,38,39  also showed that genes in the dmeL/umC group displayed a significantly (P < .001), but less marked, higher proportion of absent detection calls compared with the umL/umC group (Figure 3; Figure S11).

Intriguingly, genes in the meL/umC and dmeL/umC groups showed a significantly higher number of absent detection calls also in 2 independent series of normal hematopoietic tissues (Figure 3). In contrast, we did not observe such effect in normal tissues of nonhematopoietic origin (Figure 3). These findings suggest that genes undergoing de novo methylation across various subtypes of maB-NHLs are expressed at comparably low levels also in healthy hematopoietic tissues but not in other healthy tissues (Table 2).

As genes studied with the methylation-specific BeadArray were selected for their involvement in cancer, they will be by definition enriched for functions deregulated in cancer. Even considering this bias, we performed a hypergeometric test to determine whether the 4 DNA methylation groups were especially enriched for certain Gene Ontology terms. The most significantly enriched biologic terms are shown in Table 3(a full list is provided as Table S7). Genes from the umL/umC group were mostly enriched for terms associated with apoptosis and cell death, whereas genes from the meL/meC group were involved in metabolic processes as well as chromosome segregation and DNA repair. Genes from the dmeL/umC group showed enrichment for processes involving cell differentiation. Finally, genes from the meL/umC group were highly enriched for signal transduction pathways, which included phospholipase C activation, G-protein signaling, and second messenger–mediated signaling.

In the present report, we have studied the relationship between DNA methylation patterns and morphologic, genetic, and transcriptional features in a series of molecular Burkitt lymphomas and nonmolecular Burkitt lymphomas,¹⁷  the latter being mostly DLBCL lymphomas of the GCB or ABC subtype.¹⁵ 

Initially, we observed that DNA methylation did not strictly correlate with any phenotypic, genetic, or transcriptional feature. Using a stringent supervised statistical approach (Document S2), we identified 56 (72%) genes de novo methylated in all maB-NHLs studied and 22 (28%) genes differentially methylated in distinct maB-NHL subtypes. Genes from the latter group can be potentially used as diagnostic markers to differentiate maB-NHL subtypes (Table S6).

To obtain further insights into the nature of the groups of genes showing specific DNA methylation patterns in lymphomas and controls, we studied several biologic characteristics, which are summarized in Table 2. In these analyses, we observed that umL/umC and meL/meC genes were enriched for HCP and LCP promoters, respectively. This is in line with data on normal fibroblasts, in which HCPs are predominately unmethylated, whereas LCPs are frequently methylated.³¹  Genes suffering de novo methylation in lymphomas (meL/umC or dmeL/umC) belonged mostly to the HCP group. This finding further supports the widely accepted model in which de novo methylation in cancer predominantly affects genes with dense CpG islands.^1–4,40 

We also studied gene expression profiles from the very same cases analyzed with the methylation-specific BeadArray.¹⁷  In addition, gene expression profiles from an independent dataset of maB-NHLs were studied.³⁸  Genes from the meL/umC and dmeL/umC groups showed a higher proportion of absent detection calls than genes of the umL/umC in both lymphoma series (Figure 3). Interestingly, genes from the meL/umC and dmeL/umC groups also showed a significant increase of absent detection calls in 2 independent series of normal hematopoietic tissues but not in nonhematopoietic normal tissues (Figure 3). This finding is in line with a previous study showing that genes de novo methylated in colon cancer are expressed at low levels not only in colon cancers but also in normal colon samples.⁵  The finding that genes of the meL/umC group show low or reduced expression in maB-NHL is in complete agreement with recent studies on, for example, DCC, EPHA7, DBC1, or HTR1B applying quantitative reverse-transcription (qRT)–PCR and Northern blot techniques.^41–44 

The meL/umC group was significantly enriched for genes encoding proteins involved in signal transduction pathways such as phospholipase C activation, G-protein signaling, and second messenger–mediated signaling. Interestingly, genes previously identified as hypermethylated in solid tumors are also enriched for the very same signal transduction pathways.⁵  This suggests that genes de novo methylated in distinct cancer types are involved in similar biologic functions, and points to the importance of certain signaling pathways conserved across various cancer types.

Recent reports indicate that DNA methylation in cancer mostly targets genes repressed by PcG proteins at the ESC stage.^7–9  Therefore, aberrant DNA methylation might occur in a cancer precursor cell, and the detection of epigenetic changes in tumor biopsies might represent a kind of memory of stemness or early stages of tumorigenesis. Here, we observed that genes commonly methylated across maB-NHLs with different genetic and phenotypic features were highly enriched for such PcG targets (OR = 8.2). In contrast, genes differentially methylated in maB-NHL subtypes showed a lower enrichment (OR = 2.9; Figure 2A). This indicates that genes acquiring de novo methylation in lymphomas, especially from the meL/umC group, were repressed by PcG-related histone marks in ESCs.

To study the methylation status of genes de novo methylated in maB-NHL in stem cells, we performed methylation profiling of ASCs. We found that the great majority of the genes de novo methylated in maB-NHLs was unmethylated in ASCs (Figure S10). Furthermore, 73 from the 78 genes de novo methylated in maB-NHLs were also unmethylated in ESCs, as demonstrated by analyzing data from a recent study applying the same BeadArray platform to study ESCs.²²  This finding indicates that although a high proportion of genes de novo methylated in maB-NHL is repressed by PcG proteins in ESCs, the genes are mostly unmethylated in this pluripotent cell type.

As the GoldenGate methylation platform targets only 807 cancer-related genes, we performed a pilot study of maB-NHLs with the Infinium HumanMethylation27 BeadChip (Illumina), which allows measuring the methylation status of a total of approximately 27 000 CpGs from the approximately 14 000 best annotated genes of the human genome (Document S3). We studied 20 of the same maB-NHLs (7 mBLs, 6 non-mBLs [ABC], and 7 non-mBLs [GCB]) analyzed with the GoldenGate platform and 5 control samples (3 peripheral blood samples, 1 bone marrow sample, and 1 B-lymphoblastoid cell line). Using this new array, we identified 969 genes to be de novo methylated in maB-NHL (meL/umC). From those, 44.1% (398 of 906) were PcG targets in stem cells (OR = 10.1, P = .001, comparing genes of the meL/umC group vs genes of the umL/umC and meL/meC groups). Therefore, an independent platform analyzing an unbiased set of genes clearly confirms the high PcG enrichment among hypermethylated genes in maB-NHL identified with the GoldenGate platform. A summary of the data derived from the 27K platform is shown in Figure S12.

Overall, our findings suggest that all the different maB-NHLs studied herein might originate from lymphoma precursor cells with similar stem cell–like features that have acquired aberrant methylation of PcG target genes. This new hypothesis seems to contradict current models of lymphomagenesis, in which it is accepted that lymphomas are initiated through chromosomal translocations taking place at different stages of lymphocyte differentiation (Figure 4A). In this context, we interpret our results in 2 ways. On the one hand, it might be that aberrant DNA methylation of PcG target genes in cells, probably with stem cell–like features, represents the initial event of lymphomagenesis (Figure 4B). On the other hand, it might happen that lymphomagenesis is initiated through chromosomal changes occurring in cells with stem cell features (Figure 4C1), which is supported by the fact that multipotent hematopoietic cells are more susceptible to suffer chromosomal rearrangements than mature cells.⁴⁵  Alternatively, initial chromosomal changes might take place in a differentiating cell instead of in a stem cell. This cell would then acquire stem cell–like features (eg, PcG repression of ESC-related genes) through gene deregulation caused by a chromosomal translocation (Figure 4C2). This hypothesis might be supported by several recent studies indicating that overexpression of key transcription factors such as OCT4, SOX2, KLF4, and MYC is able to reprogram somatic cells into stem cells.^46,47  Interestingly, MYC is overexpressed in lymphomas with chromosomal translocations affecting the MYC locus.¹²  Subsequently, these cells with stem cell–like features bearing a chromosomal translocation would suffer aberrant DNA methylation of PcG target genes. Finally, this would be followed by additional genetic and epigenetic changes to develop a maB-NHL (Figure 4C).

A recent study published by Opavsky et al has shed light on the relationship between genetic and epigenetic changes in lymphomagenesis taking advantage of transgenic mice overexpressing MYC.⁴⁸  The authors found that T-cell lymphomas arising in those mice were associated with a specific signature of aberrant DNA methylation. Interestingly, the DNA methylation pattern of cells showing and lacking MYC overexpression was identical, suggesting that MYC deregulation does not directly induce aberrant methylation.⁴⁸  In addition, they found that epigenetic changes were detectable only in late-stage tumors, which led the authors to suggest that specific profiles of CpG island methylation arise from the expansion of rare neoplastic cells.⁴⁸  Remarkably, using a published dataset of PcG targets in mouse ESCs,⁴⁹  we found that 74% of the genes hypermethylated in the MYC-induced T-cell lymphomas studied by Opavsky et al⁴⁸  were targets of at least one PcG protein in ESCs (data not shown). Thus, it is tempting to hypothesize that these T-cell lymphomas were originated from rare cells with stem cell–like features, which in addition to MYC over-expression, would acquire aberrant DNA methylation of PcG target genes.

In conclusion, our study further provides evidence for an instructive mechanism leading to DNA methylation not only in solid cancer but also in maB-NHLs. Moreover, our data support the concept that aberrant DNA methylation of genes already repressed by PcG in ESCs is an epigenetic hallmark of a wide variety of cancer types.⁵⁰  Furthermore, we show that maB-NHLs with different morphologic, genetic, and transcriptional backgrounds share a PcG-associated signature of DNA methylation. This suggests that maB-NHL might not originate from differentiating cells but rather from precursor cells with stem cell–like features that have acquired aberrant methylation of PcG target genes. Alternatively, lymphocytes in the process of malignant transformation might have acquired a stem cell–like epigenetic pattern, for example, through chromatin remodeling induced by chromosomal changes.⁴⁸  In light of these new findings, current models of lymphomagenesis might have to be revisited.

The authors thank Björn Hihn for his assistance in performing the Infinium HumanMethylation27 BeadChip experiments.

The Network Project “Molecular Mechanisms in Malignant Lymphoma” is supported by the Deutsche Krebshilfe (Bonn, Germany; grant 70-3173-Tr3). R. Siebert and W.K. are funded by the KinderKrebsInitiative Buchholz/Holm-Seppensen (KKI; Buchholz, Germany). M.E. and E.B. were funded by BFU2004-02073/BMC and CSD2006-49 grants from the Spanish Ministry of Education and Science (MEC; Madrid, Spain). M.E. was supported by the Health (FIS01-04) and Education and Science (I+D+I MCYT08-03 and Consolider MEC09-05) departments of the Spanish government (Madrid, Spain), the European Grant Transfog LSHC-CT-2004-503438 (Brussels, Belgium), and the Spanish Association Against Cancer (AECC; Madrid, Spain). S.B.C. is supported by the Swiss Group for Clinical Cancer Research (SAKK; Bern, Switzerland). The work performed in the laboratory of D.S. was supported by the Novartis Research Foundation (Basel, Switzerland). M.K. is supported by a predoctoral grant (GRK 1034) of the Georg August University of Göttingen (Göttingen, Germany).

Contribution: J.I.M.-S. and R. Siebert designed experiments; H.-W.B., S.B.C., M.L.H., M.H., W.K., P.M., G.O., A.R., and H.S. acquired tumor samples and performed a central pathological review; X.A., H.B., H.G.D., R.K., R.A.F.M., C.P., F.P., and M.S. provided tumor, stem cell, and control DNA samples; J.I.M.-S., M.B., E.W.-G., J.R., L.L.-S., O.A., E.B., J.-B.F., C.S., S.W., M.E., and D.B. generated experimental data; V.C., M.F.F., D.S., and M.W. provided experimental data; J.I.M.-S., M.K., S.B., M.R., H.B., M.L., R. Spang, D.H., and R. Siebert analyzed the data; B.K., B.S., and L.T. provided logistical and intellectual support; and J.I.M.-S. and R. Siebert wrote the paper.

Conflict-of-interest disclosure: M.B., E.W.-G., J.-B.F., and D.B. are (or were at the time of the research) employees and stockholders of Illumina. J.I.M.-S. has received an honorarium for speaking in a meeting organized by Illumina. The remaining authors declare no competing financial interests.

A complete list of members of the Molecular Mechanisms in Malignant Lymphomas (MMML) Network Project appears in Document S1.

Correspondence: José I. Martín-Subero, Institute of Human Genetics, University Hospital Schleswig-Holstein, Campus Kiel, Schwanenweg 24, D-24105 Kiel, Germany; e-mail: jimartin@idibell.org.