Transcriptional signature with differential expression of BCL6 target genes accurately identifies BCL6-dependent diffuse large B cell lymphomas

Abstract

Diffuse large B cell lymphomas (DLBCLs) often express BCL6, a transcriptional repressor required for the formation of normal germinal centers. In a subset of DLBCLs, BCL6 is deregulated by chromosomal translocations or aberrant somatic hypermutation; in other tumors, BCL6 expression may simply reflect germinal center lineage. DLBCLs dependent on BCL6-regulated pathways should exhibit differential regulation of BCL6 target genes. Genomic array ChIP-on-chip was used to identify the cohort of direct BCL6 target genes. This set of genes was enriched in modulators of transcription, chromatin structure, protein ubiquitylation, cell cycle, and DNA damage responses. In primary DLBCLs classified on the basis of gene expression profiles, these BCL6 target genes were clearly differentially regulated in “BCR” tumors, a subset of DLBCLs with increased BCL6 expression and more frequent BCL6 translocations. In a panel of DLBCL cell lines analyzed by expression arrays and classified according to their gene expression profiles, only BCR tumors were highly sensitive to the BCL6 peptide inhibitor, BPI. These studies identify a discrete subset of DLBCLs that are reliant on BCL6 signaling and uniquely sensitive to BCL6 inhibitors. More broadly, these data show how genome-wide identification of direct target genes can identify tumors dependent on oncogenic transcription factors and amenable to targeted therapeutics.

BCL6 is a BTB/POZ domain transcription repressor that is required for normal germinal center (GC) development and expressed by the majority of normal GC B cells and a subset of diffuse large B cell lymphomas (DLBCLs) (1, 2). BCL6 favors the survival and proliferation of GC B cells, which undergo somatic hypermutation of Ig variable regions and Ig class switch recombination; down-regulation of BCL6 is necessary for post-GC B cell maturation (3–5). Deregulation of BCL6, by chromosomal translocation or aberrant somatic hypermutation, is the most common genetic abnormality in DLBCL (6). Conclusive evidence for the oncogenic role of BCL6 comes from murine models in which constitutive BCL6 expression results in the development of a lymphoid malignancy resembling DLBCL (7, 8). Although deregulated BCL6 clearly plays a pathogenetic role in a subset of human DLBCLs, other DLBCLs may simply express this transcriptional repressor because the lymphomas are derived from normal BCL6⁺ GC B cells. Identification of BCL6-dependent tumors has important therapeutic implications because a recently described specific BCL6 peptide inhibitor (BPI) inhibits the growth of some but not all DLBCLs (9).

To delineate functionally relevant DLBCL subsets, we and others have used gene expression signatures. In an earlier approach known as the cell of origin (COO) classification, subsets of DLBCLs were associated with specific types of normal B cells [GC B cells (GCB) or in vitro activated peripheral blood B cells (ABC)] or left unassigned if the tumors did not closely resemble either B cell category (Other) (10). Although GCB DLBCLs had more abundant BCL6 transcripts, there was no association between BCL6 genetic abnormalities and this tumor subset.

More recently, we applied consensus clustering methods to the transcriptional profiles of two large independent series of primary DLBCLs to identify the dominant substructure a priori (i.e., to classify DLBCLs in an unbiased manner) (11). The obtained consensus clusters were highly reproducible and included three groups of DLBCLs, termed B cell receptor/proliferation (BCR), oxidative phosphorylation (OxPhos), and host response (HR) tumors; these DLBCL subsets were unrelated to the developmentally defined COO tumor groups (11). HR tumors are defined, in part, by their brisk host inflammatory/immune response and histologic and clinical similarities to the WHO pathologic subtype, T cell/histiocyte-rich LBCL. HR tumors rarely exhibit the genetic lesions seen in other DLBCLs (11, 12). In contrast, OxPhos DLBCLs have increased expression of genes involved in oxidative phosphorylation and mitochondrial function and more common structural abnormalities of intrinsic and extrinsic apoptotic pathway components (11, 12). BCR tumors have increased expression of cell cycle regulatory genes, components of the BCR signaling cascade, and certain B cell-specific transcription factors such as BCL6; these DLBCLs also exhibit more frequent translocations of the BCL6 locus (11, 12).

We predicted that differential regulation of BCL6 target genes would identify tumors specifically driven by BCL6. We postulated that among DLBCLs, BCR tumors would be more likely to rely on deregulated BCL6 expression and be uniquely sensitive to BPI treatment. For these reasons, we used a chromatin immunoprecipitation (ChIP)-on-chip approach to identify BCL6 target genes in a B cell lymphoma cell line and asked whether these BCL6 target genes contributed to the signature of a specific DLBCL subset. Here, we demonstrate that BCR DLBCLs exhibit coordinate regulation of the identified BCL6 target genes. In addition, the BCL6 signature and BCR subtype designation have important functional consequences because only BCR DLBCL growth is inhibited by BPI treatment. The BCL6 target gene signature provides important insights into the biology of BCR DLBCLs and identifies these tumors as candidates for rational targeted BCL6 inhibition.

Results and Discussion

Identification of BCL6 Target Genes.

We predicted that BCL6-dependent DLBCLs would have a transcriptional signature that was defined, at least in part, by the differential expression of BCL6 target genes. To identify such genes, we performed high-throughput ChIP on chip in the Ramos B cell lymphoma cell line, which is frequently used to evaluate BCL6 function (13–15). Chromatin fragments were immunoprecipitated with an antibody directed against BCL6 or an irrelevant control (actin), and the resulting products were amplified by ligation-mediated PCR (LMPCR). Specific enrichment of BCL6 target genes was validated by single-locus quantitative-PCR ChIP (Q-ChIP) before and after LMPCR.

Thereafter, the resulting amplicons were labeled and cohybridized with input chromatin to high-density oligonucleotide arrays containing a 1.5-kb sequence of 24,275 gene promoters, each of which was represented by 15 consecutive 50-mer oligonucleotides. “Hits” were captured through a highly stringent approach employing random permutation analysis on a sliding window of oligonucleotide probes (i.e., on groups of three consecutive probes) (see Materials and Methods). The threshold of positivity was set at the enrichment level of the known BCL6-binding site in the CCL3 promoter (16), which corresponded to the 95th percentile confidence interval for this method [Fig. 1A and supporting information (SI) Fig. 4]. Only genes that were captured by all three replicates and that displayed overlapping peak enrichment were considered positive. Examples of BCL6 peak oligonucleotide enrichment at target gene regulatory regions are shown in Fig. 1B.

Fig. 1.

Identification of BCL6 target genes. ChIP on chip was performed in triplicate on a 24,000 promoter array. (A) Selection of BCL6 target genes. (Left) Density plot of the normalized log ratio of fluorescence intensity (BCL6 vs. input) of the oligonucleotide probes. The probes showing relative enrichment by BCL6 antibodies are clustered in the right tail of the distribution curve. (Right) Density plot of the maximal enrichment peak of each promoter (black line). The gray line represents a similar plot generated by using a random distribution of probes. The indicated cutoff point for selection of positive hits (shaded) was set at the 95th percentile of the random probe curve (see also SI Fig. 4). The y axis for both panels represents the local relative frequency of events within each level of fluorescence intensity represented on the x axis, corresponding to probe frequency (Right) and peak frequency (Left). (B) Representative BCL6 target genes. Shown are the peak BCL6 vs. input enrichment at negative and positive control promoters (CD20 and FCER2, respectively) and four selected gene promoters (SUB1, CR1, CBX3, and MBD1) that met the following criteria: (i) inclusion in an enriched GO category; (ii) validation by single-locus Q-ChIP; (iii) up-regulation after BPI treatment; and (iv) inclusion in the leading edge of the BCL6 target gene set in GSEA (for details, see Materials and Methods). In each graph, the y axis shows fold enrichment by BCL6 antibodies (gray field) vs. a control IgG (black field). The x axis indicates the relative position of the different oligonucleotides relative to the transcriptional start site as annotated in the National Center for Biotechnology Information (NCBI) human genome assembly version 35 (May 2004).

BCL6 was recruited to 436 promoters, potentially regulating 485 target genes (SI Table 2), including known target genes such as FCER2 and CCL3 (16, 17). To determine the accuracy of BCL6 target gene discovery, single locus quantitative ChIP was performed on 54 of the candidate BCL6 target genes, using the known targets CCL3 and FCER2 as positive controls. Eighty-one percent of the examined candidate BCL6 target genes were confirmed with this stringent approach (SI Fig. 5).

Functional Classification of BCL6 Target Genes.

To gain insights into the functions of identified BCL6 target genes, we evaluated their associated Gene Ontology (GO) Biological Process terms (www.geneontology.org/GO.doc.shtml). GO terms annotate genes and their products based on described biological functions. The GO term frequency in a given gene set (i.e., BCL6 target genes) can be compared with the global GO term frequency to identify functional categories that are represented more frequently in the examined gene set. We specifically compared the representation of GO terms in the BCL6 target gene set with that in the total analyzed gene pool (i.e., all of the genes in the GO database) (18) (Table 1). The BCL6 target list was enriched in genes regulating transcription, DNA damage responses, chromatin modification of the cell cycle, and protein ubiquitylation (Table 1). Because the mechanism(s) through which BCL6 mediates the GC reaction and lymphomagenesis are largely unknown, these data provide insights regarding BCL6 function in these processes.

Table 1.

GO term analysis of BCL6 target genes

GO term	GO term frequency in BCL6 target gene set	Global GO term frequency	P value	FDR
Transcription	37/418 (0.0885)	0.0386	0.0000	0.0003
Protein ubiquitylation	15/418 (0.0359)	0.0096	0.0000	0.0006
Cell cycle	14/418 (0.0335)	0.0103	0.0001	0.0035
Ubiquitin cycle	12/418 (0.0287)	0.0083	0.0002	0.0043
Chromatin modification	6/418 (0.0144)	0.0023	0.0003	0.0053
Response to DNA damage stimulus	3/418 (0.0072)	0.0004	0.0004	0.0053
Regulation of transcription, DNA-dependent	41/418 (0.0981)	0.0567	0.0005	0.006
Ubiquitin-dependent protein catabolism	7/418 (0.0167)	0.0039	0.0011	0.013

Enrichment was evaluated using the GeneMerge program (18), which compares the frequency of GO categories represented in the nonredundant list of SwissProt/TrEMBL accession nos. of BCL6 target genes (n = 418) versus the global frequency of GO categories in the population gene file [19,168 nonredundant SwissProt/TrEMBL accession nos. corresponding to the known genes in NCBI human genome assembly version 36 (March 2006)]. GO enrichment analysis was carried out with GO biological process terms, and obtained P values were corrected for multiple-hypothesis testing by false-discovery rate (FDR) (24, 25). P values (E scores) for GO term overrepresentation in the BCL6 target genes were obtained using the hypergeometric distribution (18, 29). The hypergeometric distribution quantifies the overrepresentation of a given GO term in a sample of a specific size that is drawn from a larger population (18).

BCL6 Target Gene Expression in DLBCL Subtypes.

We predicted that differential expression of BCL6 target genes would identify DLBCLs in which BCL6 plays a dominant oncogenic role, so we assessed the relative abundance of BCL6 targets in the respective DLBCL consensus clusters (11). In this analysis, we used gene set enrichment analysis (GSEA) to determine whether the set of BCL6 target genes was expressed differentially in a specific DLBCL subtype (19). Because the signature of HR tumors is largely defined by genes expressed by the tumor-infiltrating normal inflammatory and immune cells, we focused the GSEA on BCR vs. OxPhos DLBCLs. Although BCL6 likely functions as a direct transcriptional repressor, the absolute levels of specific target genes may depend on BCL6 cooperation with other transcription factors, binding to different corepressors or additional epigenetic modifications of chromatin. For these reasons, we ranked the genes discriminating between BCR and OxPhos phenotypes according to absolute (rather than positive or negative) signal-to-noise ratios (SNRs) and assessed the enrichment of BCL6 target genes in the ranked dataset. In our series of 176 primary DLBCLs, the BCR vs. OxPhos ranked gene list was significantly enriched for BCL6 target genes (P < 0.0001), indicating that the BCL6 signature contributes to the difference between BCR and OxPhos tumors.

To validate these observations in an independent dataset, GSEA was performed in an additional large series of transcriptionally profiled primary DLBCLs with available COO and consensus cluster designations (11, 20). In this independent series, BCL6 targets were similarly enriched in ranked genes discriminating between BCR and OxPhos signatures (SI Table 3). In contrast, BCL6 target genes were not significantly enriched in either dataset when the DLBCLs were sorted with respect to the GC B vs. ABC classification (SI Table 3).

To determine which BCL6 targets were more (or less) abundant in BCR vs. OxPhos DLBCLs, the BCL6 target genes were clustered with respect to these tumor types (SI Table 4). The top-scoring BCL6 target genes [the “leading edge” (see Materials and Methods) (19)] are visually displayed in Fig. 2A, which also included normal tonsillar GC B cells for comparison. Consistent with the known role of BCL6 as a transcriptional repressor, a number of BCL6 target transcripts were less abundant in BCR DLBCLs than in OxPhos tumors (Fig. 2 and SI Table 4); the majority of these BCL6 targets were also less abundant in normal GC B cells (Fig. 2A). However, additional bona fide BCL6 targets were more abundant in BCR tumors and normal GC B cells than OxPhos DLBCLs (Fig. 2A and SI Table 4). This unexpected observation prompted us to analyze directly the BCL6 dependence of candidate target genes in a panel of informative DLBCL cell lines.

Fig. 2.

BCL6 target genes in primary BCR and OXP DLBCLs and DLBCL cell lines. The top-scoring BCL6 target genes from the GSEA leading edge were clustered with respect to the DLBCL BCR and OxPhos phenotypes and represented visually. Each individual column represents a tumor, and each individual row corresponds to a gene. For comparison, the relative expression of these BCL6 target genes in normal GC B cells is also shown. The color scale at the bottom indicates relative expression. (A) Primary DLBCLs (11). (B) DLBCL cell lines (BCR lines: Ly1, Ly7, SU-DHL4, SU-DHL6, and Farage; OxPhos lines: Ly4, Toledo, Kaspas 422, and Pfeiffer). (C) BCL6 target gene abundance in BCR and OxPhos cell lines after BPI treatment. BCR (SU-DHL6, SU-DHL4) and OxPhos (Toledo, Ly4) cell lines were treated with 20 μM BPI or control peptide for 8 h, and the transcript abundance of the indicated BCL6 targets was evaluated with real-time PCR thereafter. The y axis indicates fold activation of genes after treatment with BPI vs. control peptide based on the ΔΔ^CT normalized to the expression of hypoxanthine-guanine phosphoribosyl transferse (HPRT). BPI treatment increased the expression of each BCL6 target gene in the BCR cell lines but did not alter the expression of these genes in OxPhos lines. BPI treatment increased the abundance of BCL6 targets that were less abundant in BCR than OxPhos cells at baseline (SUB1, ZNF443, CR1, and CBX3; shaded in blue) and others that were more abundant in BCR tumors at baseline (CD74, CCN1, and MBD1; shaded in red). A known BCL6 target gene FCER2 was used as a positive control (shaded in pink).

BCL6 Actively Represses Its Target Genes in BCR but Not OxPhos Tumors.

We first identified representative BCR or OxPhos DLBCL cell lines (BCR: Ly1, Ly7, SU-DHL4, SU-DHL6, and Farage; and OxPhos: Ly4, Toledo, Karpas 422, and Pfeiffer) based on their transcriptional profiles (Materials and Methods and http://www.pnas.org/cgi/content/full/0611399104/DC1SI Methods). Thereafter, we performed GSEA for BCL6 targets by using the cell line gene list, ranked according to absolute SNR values for the BCR vs. OxPhos distinction. As was the case in primary DLBCLs, BCL6 target genes were highly enriched in the ranked cell line gene list (P < 0.001). In addition, certain BCL6 target transcripts were less abundant in BCR than in OxPhos cell lines, whereas other BCL6 targets were more abundant in BCR DLBCLs (Fig. 2B).

We then treated four of the BCR and OxPhos cell lines with BPI and evaluated the transcript abundance of representative BCL6 targets (Fig. 2C). The BCL6 targets chosen for this analysis were: (i) validated by Q-ChIP; (ii) included in a significantly enriched GO category; and (iii) most differentially expressed in BCR and OxPhos tumors (i.e., included in the leading-edge gene set). We specifically selected certain candidate BCL6 targets that were less abundant in BCR than in OxPhos cells (SUB1, ZNF443, CR1, CBX3) at baseline (shaded in blue in Fig. 2C) and others that were more abundant in BCR tumors (CD74, CCN1, MBD1, FCER2) (shaded in red in Fig. 2C). BPI treatment increased the expression of each of these BCL6 targets in the BCR DLBCL cell lines, but it did not alter the expression of these genes in OxPhos cell lines (Fig. 2C). These data suggest that BCL6 is biologically active in BCR but not OxPhos tumors and show that BCL6 represses its target genes in BCR DLBCLs regardless of baseline target transcript levels.

The BCR Signature Predicts for BCL6-Dependent DLBCL Survival.

Because BPI selectively increased BCL6 target expression in BCR DLBCLs, we predicted that these tumors would be more dependent on BCL6-regulated gene pathways than OxPhos DLBCLs. We had shown (9) that BPI specifically blocked BCL6 activities in vitro and in vivo and inhibited the growth of certain BCL6-positive lymphomas. For this reason, we treated the five BCR and four OxPhos DLBCL cell lines with BPI and subsequently evaluated tumor cell proliferation. In these experiments, cell line identity was blinded until after the functional data were analyzed independently.

BCR cell lines had significantly lower BPI IC₅₀ values than OxPhos lines, which were uniformly resistant to the peptide inhibitor (BCR vs. OxPhos DLBCL IC₅₀, 12.7 ± 3.49 μM vs. 50.15 ± 4.43 μM, P < 0.0001; Fig. 3 A and B). To characterize further the differential sensitivity of BCR vs. OxPhos cell lines, we exposed the panel to 20 μM BPI for 48 h. BPI inhibited cellular proliferation of BCR DLBCL cell lines by 65–90% but had little effect on OxPhos tumors (Fig. 3C). Therefore, the designation of BCR (vs. OxPhos) DLBCL accurately predicted the response to BCL6 inhibition. Importantly, the consensus cluster designation was more effective in predicting response to BPI therapy than simple baseline BCL6 expression (SI Fig. 6).

Fig. 3.

BCR and OxPhos DLBCL cell lines exhibit differential sensitivity to BPI. (A) BPI IC₅₀ for BCR and OxPhos DLBCL cell lines. BCR and OxPhos cell lines were exposed to increasing doses of BPI, and cellular proliferation was assessed at 48 h. The IC₅₀ ± SEM for triplicate samples of each cell line in a representative experiment are shown. (B) Mean BPI IC₅₀ (±SD) for BCR and OxPhos DLBCL cell lines. (C) Proliferation of BPI-treated BCR and OxPhos DLBCL cell lines after BPI treatment. Cell lines were exposed to 20 μM BPI for 48 h, and cellular proliferation was evaluated thereafter.

Our approach of combining stringent genomic localization by ChIP on chip with large-scale functional genomics and the use of a specific transcription factor inhibitor highlight the important contribution of an oncogenic transcription factor to the transcriptional programming of a human tumor. Specifically, our studies identify a subset of DLBCLs, the BCR tumors, in which BCL6 plays a critical biological role. The BCR consensus cluster designation was more accurate in predicting BPI sensitivity than either BCL6 protein expression alone or the absolute levels of BCL6 target genes. This result is not surprising because the consequences of BCL6 binding to its target genes depend on which corepressors, additional transcription factors, and epigenetic modifications are present at those genes at a given time. Rather than absolutely repressing all of its targets, BCL6 likely modulates target gene expression in specific cellular contexts. In this regard, there may be differences in the BCL6 target gene signature in BCL6-dependent tumors and normal GC B cells. Such differences could explain why the consensus clusters more accurately identify BCL6-dependent tumors than the COO categories, which relate DLBCLs to subsets of normal B cells.

From a clinical standpoint, our data indicate that patients with BCR DLBCLs may represent the best candidates for therapeutic trials of BCL6 inhibitors. Standard diagnostic methods will not delineate these patients. Development of methods to identify tumors most likely to respond to targeted therapy is an important advance because it allows for molecular stratification of patients to therapeutic arms most likely to be of benefit. More broadly, these data show how integration of genome-wide transcription factor binding and gene expression profiling can provide important insights into tumor biology, identify the presence of gene regulatory programming by oncogenic transcription factors, and direct selection of tumors for targeted therapeutic agents.

Materials and Methods

Q-ChIP.

ChIPs were performed as described in ref. 9 with 5 × 10⁶ Ramos cells (13–15) and rabbit antisera directed against BCL6 (N3 antibody; Santa Cruz Biotechnology, Santa Cruz, CA) or actin (Sigma–Aldrich, St. Louis, MO). DNA fragments enriched by ChIP were quantified by real-time PCR using a SYBR green kit (Applied Biosystems, Foster City, CA) and an Opticon Engine 2 (MJ Research, Waltham, MA). The known BCL6-binding sites in the CCL3 and FCER2 genes (16, 17) were used as positive controls for BCL6 target gene enrichment, whereas the CD20 gene, which is not a BCL6 target, was used as a negative control. ΔC_T values (antibody − input) were expressed relative to control antibody by using the ΔΔC_T method (21). The same approach was used to validate an additional 44 (of 54 promoters tested) target genes identified by ChIP on chip. Primers used in these experiments are listed in SI Table 5.

ChIP on Chip and Data Processing.

After validation of enrichment by real-time PCR, BCL6, or actin, ChIP products and their respective input genomic fragments were amplified by ligation-mediated PCR (22). Q-ChIP was repeated after amplification to verify that the enrichment ratios were retained. The genomic products of three biological ChIP replicates were labeled with Cy5 (for ChIP products) and Cy3 (for input) and cohybridized on a NimbleGen human promoter array representing 1.5 kb of promoter sequence from 24.275 genes (human genome version 35, May 2004) according to manufacturer's protocol (NimbleGen Systems, Madison, WI). The enrichment for each promoter was calculated by computing the log ratio between the probe intensities of the ChIP product and input chromatin, which are cohybridized on the same array. Thereafter, for each of the 24,175 promoter regions, the maximum average log ratio of three neighboring probes in a sliding window was calculated and compared with random permutation of the log ratios of all probes across the entire array. The positive threshold was defined by using the CCL3 signal that corresponds to the 95th percentile in random permutation of the log ratios.

The putative BCL6-binding regions were calculated from triplicate experiments, represented as enrichment peaks of BCL6 over control antibody signal and aligned with chromosome positions (NCBI human genome assembly version 35, May 2004). Thereafter, by using the NimbleGen 24K promoter array annotation file, the peak signals of BCL6 binding were assigned to the respective regulatory regions of candidate BCL6 target genes. In addition, all peaks were inspected by using BLAT (The BLAST-like Alignment Tool) to identify genes on opposite strands that could be regulated from the same bidirectional promoter. Two genes were considered to be bidirectional partners when they were located on the opposite strands in a “head-to-head” orientation and their transcription start sites were separated by <1 kb (23). In previous studies, 90% of promoters meeting these criteria were bidirectionally active in functional assays (23).

GO Term Enrichment Analysis.

GO term enrichment analysis was performed with the online version of GeneMerge program (18). Enrichment was assessed by comparing the frequency of GO Biological Process categories represented in the nonredundant list of SwissProt/TrEMBL accession numbers of BCL6 target genes (n = 418) versus the global frequency of GO categories in the population gene file containing 19,168 nonredundant SwissProt/TrEMBL accession numbers that corresponded to known genes in NCBI human genome assembly version 36 (March 2006). SwissProt/TrEMBL ID codes of remaining BCL6 target loci were not available. All SwissProt/TrEMBL ID codes were obtained from the Affymetrix genome annotation file supporting U133 Plus 2 GeneChip (version July 2006). Obtained P values were corrected for multiple hypothesis testing by false-discovery rate analysis (24, 25).

GSEA.

GSEA was performed by using the GSEA version 1.0 program (19), the BCL6 target gene set, and two independent series of primary DLBCLs with gene expression profiles and consensus cluster and COO designations (11, 20). Because the signature of HR tumors is largely defined by normal tumor-infiltrating host inflammatory and immune cells, the analysis was focused on BCR and OxPhos DLBCLs. GSEA was performed as described previously, with minor modifications. The top 15,000 genes selected with a median absolute deviation-based variation filter were first ranked with respect to the phenotype, BCR vs. OxPhos, by using an absolute value (rather than positive or negative) SNR. With this approach, the final position in the ranked gene list depended only on the strength of the gene in discriminating between phenotypes rather than specific up- or down-regulation in a given phenotype. Represented members of the BCL6 target gene set were then located within the ranked gene list, and the proximity of the BCL6 target gene set to the most differentially expressed BCR vs. OxPhos genes (i.e., those with the highest absolute SNR value) was measured with a weighted Kolmogorov–Smirnov statistic [ES, enrichment score (ref. 19)], with a higher score corresponding to a higher proximity. The observed ES score was then compared with the distribution of 1,000 permuted ES scores (gene tag permutations) to assess statistical significance. Similar results were observed with the permutation of the class template (data not shown). The query gene set included the 309 (of a total 485) BCL6 target genes present in the 15,000 ranked genes; these 309 BCL6 targets were represented by 477 Affymetrix probe sets. BCL6 target gene enrichment was also assessed in the gene list ranked for the positively defined COO phenotypes GC B vs. ABC (10, 11) sorted by absolute SNRs.

GSEA was also performed in an independent dataset of 218 primary DLBCL patients with available COO designations (llmpp.nih.gov/DLBCL/DLBCL patient data NEW.txt; ref. 20) and consensus cluster assignments (11). Affymetrix IDs of BCL6 target genes were translated to Lymphochip IDs with current and archival UniGene cluster IDs and used as the query gene set. Enrichment was assessed as described above by ranking the genes with respect to the absolute SNR values for the comprehensive cluster phenotypes BCR vs. OxPhos or COO phenotypes GC vs. ABC.

The top-scoring BCL6 target genes, described as the leading edge genes, appear in the ranked list at or before the point where the ES running sum reaches its maximum deviation from zero (26). The leading-edge genes can be interpreted as the core of a gene set that accounts for the enrichment signal (19). These top-scoring BCL6 target genes were clustered with respect to the BCR vs. OxPhos tumor phenotypes and represented on heat maps by using the dChip 2006 program. For comparison, the heat maps also included normal CD19⁺ sIgD⁻ CD38⁺ GC B cells that were isolated as described (27) and transcriptionally profiled at the same time as the primary DLBCLs (11).

Cell Line Culture.

DLBCL cell lines Ly1 and Ly7 were grown in medium containing 90% Iscove's medium, 10% FBS (Gemini Bio-Products, Woodland, CA), and penicillin G/streptomycin. DLBCL cell lines Farage, Toledo, SU-DHL4, SU-DHL6, Karpas 422, and Pfeiffer were cultured in medium containing 90% RPMI medium, 10% FCS, 2 mM glutamine, 10 mM Hepes (ptt 7.2–7.5), and penicillin G/streptomycin.

GSEA in Cell Lines.

Total RNA was extracted from a panel of DLBCL cell lines, processed, hybridized to U133A and B Affymetrix oligonucleotide microarrays, scanned, and analyzed as described (11). Cell lines were then assigned to consensus clusters by using an ensemble classifier incorporating multiple independent predictors (SI Methods, SI Table 6, and SI Fig. 7). Cell lines that were assigned to BCR or OxPhos categories with the highest probability were selected for GSEA and additional functional analyses. GSEA was performed as described above, using the top 12,666 genes that met threshold and variation index criteria (28); genes were ranked according to absolute SNR values for the phenotype BCR vs. OxPhos. The proximity of the BCL6 target gene set to the top of the ranked list was measured with an ES, and the significance of the ES was determined by using 1,000 gene tag permutations, as described above.

Treatment with BPI.

Peptides (BPI and control) were obtained from Bio-Synthesis, Inc. (Lewisville, TX) and stored at −20°C until reconstituted with sterile pure water immediately before use. BPI purity was determined by HPLC-MS to be 98% or higher. We exposed 25 × 10⁴ DLBCL cells to BPI or control peptide (0, 1, 2.5, 5, 10, 20, 40, and 80 μM) for 48 h. Cellular proliferation was assessed by MTS assay (Cell Titer 96 Aqueous One; Promega, Madison WI) according to the manufacturer's instructions, by using eight replicates per treatment condition. The proliferation of BPI-treated cells (T) was normalized to their respective peptide concentration controls (C) as follows: (T/C)_corr (%) = (T/C) / UT × 100. The growth inhibition (IC₅₀) values were estimated by a linear least-squares regression of the (T/C)_corr values versus the concentration of BPI (or control) peptide; T/C_corr values of 50% were extrapolated. The difference in BPI IC₅₀ values of BCR and OxPhos cell lines was assessed with a two-sided Student's t test. Previous studies confirmed that BPI efficiently enters lymphoma cell lines that were subsequently identified as OxPhos DLBCLs (9).

BCL6 Target Gene Expression.

After treatment with 20 μM BPI or control peptide for 8 h, RNA was extracted from 10⁴ DLBCL cells by using the RNeasy kit (Qiagen, Valencia, CA). cDNA was synthesized by using a SuperScript III first-strand cDNA synthesis kit (Invitrogen, Carlsbad, CA). The mRNA levels of SUB1, CBX3, CR1, ZNF433, CCN1, MBD1, CD74, FCER2, and HPRT (housekeeping control) were detected by using a SYBR green kit and an Opticon Engine 2 thermal cycler (MJ Research, Waltham, MA). Primer sequences for real-time PCR are listed in SI Table 7. The C_T values of the genes of interest were normalized to HPRT (ΔC_T). ΔC_T values of the BPI-treated cells were expressed relative to control peptide-treated cells by using the ΔΔC_T method. The fold change in expression of each gene in BPI-treated vs. control peptide-treated cells was determined by the expression: 2−ΔΔ^CT with ΔΔC_T + s and ΔΔC_T − s where s is the standard deviation of the ΔΔC_T value for triplicates. Results were represented as fold expression ± SD.

Abbreviations

ABC

activated peripheral blood B cells

BCR

B cell receptor/proliferation

BPI

BCL6 peptide inhibitor

COO

cell of origin

DLBCL

diffuse large B cell lymphoma

ES

enrichment score

GC

germinal center

GC B cells

germinal center B cells

GO

Gene Ontology

GSEA

gene set enrichment analysis

HPRT

hypoxanthine-guanine phosphoribosyl transferase

HR

host response

OxPhos

oxidative phosphorylation

Q-ChIP

quantitative-PCR ChIP

SNR

signal-to-noise ratio.