STAT3 Targets Suggest Mechanisms of Aggressive Tumorigenesis in Diffuse Large B-Cell Lymphoma

The signal transducer and activator of transcription 3 (STAT3) is a transcription factor that, when dysregulated, becomes a powerful oncogene found in many human cancers, including diffuse large B-cell lymphoma. Diffuse large B-cell lymphoma is the most common form of non-Hodgkin’s lymphoma and has two major subtypes: germinal center B-cell−like and activated B–cell—like. Compared with the germinal center B-cell−like form, activated B-cell−like lymphomas respond much more poorly to current therapies and often exhibit overexpression or overactivation of STAT3. To investigate how STAT3 might contribute to this aggressive phenotype, we have integrated genome-wide studies of STAT3 DNA binding using chromatin immunoprecipitation-sequencing with whole-transcriptome profiling using RNA-sequencing. STAT3 binding sites are present near almost a third of all genes that differ in expression between the two subtypes, and examination of the affected genes identified previously undetected and clinically significant pathways downstream of STAT3 that drive oncogenesis. Novel treatments aimed at these pathways may increase the survivability of activated B-cell−like diffuse large B-cell lymphoma.

In the present study, we sought to further understand the difference in STAT3 function between these two subtypes through mapping its binding regions (BRs) and analyzing gene expression in GCB and ABC patient tumor-derived cell lines. We performed ChIP-Seq (chromatin immunoprecipitation followed by DNA sequencing) experiments to map STAT3 binding sites and RNA-Seq to analyze the global gene expression patterns. We then synthesized these data to determine which genetic loci show both differential STAT3 binding and differential mRNA expression. We found that STAT3 likely upregulates a number of oncogenic pathways to promote aggressive tumor growth and migration.

MATERIALS AND METHODS
Cell lines were grown at 37°and 5% CO 2 . SU-DHL2, SU-DHL4, SU-DHL6, SU-DHL10, OCI-Ly7, and U-2932 were grown in RPMI 1640 media supplemented with 15% FBS and antibiotics. OCI-Ly3 and OCI-Ly10 were grown in IMDM media supplemented with 15% fetal bovine serum, antibiotics, and 55 mM beta-mercaptoethanol. Western blots were performed on whole-cell lysate, with equal protein loading in each lane, with the use of anti-STAT3 rabbit polyclonal antibody sc-482X (Santa Cruz Biotechnology, Inc.); anti-pSTAT3-Y705 mouse monoclonal antibody sc-8059X (Santa Cruz Biotechnology, Inc.); and anti-GAPDH mouse monoclonal antibody ab8245 (Abcam). ChIPsequencing was performed with the anti-STAT3 antibody sc-482X on formaldehyde-crosslinked pellets of 1 · 10 6 cells. DNA was mechanically sheared using a Branson sonicator, then immunoprecipitated for 16 hr. Bound DNA was recovered on protein A-agarose beads and purified via ethanol precipitation. mRNA for RNAsequencing was isolated directly from whole cell lysate using magnetic poly-dT beads (Dynabeads mRNA DIRECT Kit; Invitrogen), then chemically fragmented (RNA Fragmentation Reagents; Ambion). cDNA was synthesized using random hexamer primers. For library preparation, standard Illumina GA-IIx primers were ligated and gel purification was used to size-select DNA in the 150-to 300-bp range. Single-ended 36-bp reads were generated for both ChIP-and RNAsequencing runs.

Statistical analysis
Sequencing results were mapped to the human genome (hg19) using Bowtie (Langmead et al. 2009). STAT3 ChIP-sequencing peaks were compared with a non-IP'd genomic DNA control, and identified using the SPP peak caller (Kharchenko et al. 2008). Replicates were analyzed using irreproducible discovery rate analysis to identify strong, repeatable peaks for each cell line (Li et al. 2011). These lists were combined and any overlapping or abutting peaks were merged into broader BRs (Kasowski et al. 2010). BRs that occurred in only one cell line were eliminated from further analysis. The ChIP-sequencing data for each line were rescored to determine how many fragments mapped to each BR. RNA-sequencing mapped reads were intersected with the RefSeq database to generate read counts for each annotated gene. Both data sets were then normalized to library size with DESeq, and to contrast the two subtypes, the DESeq parameterized negative binomial model was used to obtained the P-values for each BR or gene (Anders and Huber 2010). To correct for multiple comparisons, P-values were adjusted with the Benjamini-Hochberg procedure (Benjamini and Hochberg 1995). ChIP-Seq BRs were associated with genes that they might regulate via the Genomic Regions Enrichment of Annotations Tool using its default association settings (GREAT; McLean et al. 2010) . This list of genes associated with significant changes in STAT3 binding was compared with the list of genes with significant expression changes using the online tool Galaxy (Giardine et al. 2005;Blankenberg et al. 2010;Goecks et al. 2010). Extended materials and methods are available as Supporting Information, File S1, complete data tables are available as File S2. All ChIP-Seq and RNA-Seq data are available in the NCBI GEO public database as series accession no. GSE50724. Within this accession, ChIP-Seq data are found in the subseries GSE50723, and RNA-Seq data are in the subseries GSE50721. Individual data files are assigned the consecutive accession numbers GSM1227193-GSM1227214.

RESULTS AND DISCUSSION
STAT3 is overexpressed and overactive in ABC DLBCL In normal B-cells, STAT3 is expressed at low endogenous levels in the cytosol. High levels of STAT3 expression and activation frequently are found in ABC DLBCL, suggesting that it plays a major role in the oncogenesis of this subtype (Ding et al. 2008;Wu et al. 2011). To investigate the levels of activation of STAT3 in the treatment-responsive GCB lymphoma relative to ABC lymphoma, we selected eight patient tumor-derived cell lines for study and examined levels of total and activated STAT3 by using immunoblot analysis. Four of these cell lines have been classified by Drexler (2010) as GCB (SU-DHL4, SU-DHL6, SU-DHL10, OCI-Ly7) and four as ABC (SU-DHL2, OCI-Ly3, OCI-Ly10, U-2932).
Whole-cell lysates from each line were prepared and probed with antibodies that recognize total STAT3 as well as those that recognize only the phosphorylated (Tyr705) activated form (Shuai et al. 1993;Wen et al. 1995;Zhang et al. 1995). STAT3 protein is present in all of the GCB and ABC cell lines; however, it is at much greater levels (7.2fold) in all four ABC cell lines ( Figure 1). Furthermore, the phosphorylated form of the protein exhibits an even larger difference (56-fold) between the GCB and ABC lines ( Figure 1).
STAT3 shows differential binding in ABC relative to GCB cells To better understand how STAT3 induces aggressive behavior in ABC cells, we next sought to determine the chromosomal targets of STAT3 in all eight cell lines. We performed ChIP-Seq by using an antibody prepared against the protein's C-terminal region, the area most divergent among STAT family proteins (Baker et al. 2007). Immunoblot analyses indicate that this antibody recognizes a single band of the expected molecular mass, which was confirmed as STAT3 by immunoprecipitation followed by mass spectrometry (see Figure S1). We Figure 1 Western blot analysis of STAT3 in ABC and GCB cell lines. Western immunoblots were performed on whole-cell lysate from DLBCL cell lines probed with antibodies against STAT3, tyrosinephosphorylated STAT3, and glyceraldehyde 3-phosphate dehydrogenase (GAPDH). Fold change was measured by densitometric quanitification using NIH ImageJ software, normalized to a GAPDH loading control.
performed ChIP-Seq for all eight cell lines along with genomic DNA input controls. Two to nine biological replicates were carried out for each line; replicates were sequenced to an average depth of 18 million mapped reads (87% mapping; see Table S1). All replicates were used in the analysis.
The ChIP-Seq binding peaks were scored using the SPP peak caller and irreproducible discovery rate analysis (Kharchenko et al. 2008;Li et al. 2011), which resulted in the identification of 100026500 strongly reproducible peaks per cell line (Table 1). Overlapping or abutting peaks were merged into BRs to facilitate comparison between cell lines, yielding a total of 10,337 distinct STAT3 BRs that are occupied in at least two of the eight cell lines. These BRs occur throughout the genome, with no obvious hot spots.
To identify target regions differentially bound by STAT3 in the ABC and GCB subtypes, we used the parameterized negative binomial model in DESeq (Anders and Huber 2010) to identify those BRs that displayed significantly different normalized fragment counts between the ABC and GCB cell lines, then corrected for multiple comparisons using the Benjamini-Hochberg procedure (Benjamini and Hochberg 1995). Of the 10,337 high-confidence BRs, one third (n = 3524) are differentially bound by STAT3 between the two subtypes at a falsediscovery rate (FDR) , 0.05. (For a complete list of peaks by enriched subtype, see Table S4.) When these differentially bound peaks are clustered, the DLBCL cell lines cluster according to their GCB-ABC subtype ( Figure 2). Consistently, more BRs are strongly bound in ABC than in GCB, although BRs with increased STAT3 binding occur surprisingly frequently in the GCB cell lines given their low level of STAT3 expression: 44% of differentially bound BRs (n = 1550) show more STAT3 binding in GCB, whereas 56% (n = 1974) are more strongly bound in ABC ( Figure 3A). STAT3 protein levels are very low in GCB cell lines, so one possible explanation for its equivalent number of binding sites is simply that they represent lower occupancy regions. Indeed, we noticed the signal strength for the GCB lines was lower than that of the ABC cell lines and required additional ChIP-Seq replicate experiments to identify peaks.
STAT3 has a binding peak 14 bp upstream of the transcription start site (TSS) of JAK3, and this binding is much stronger (fold change = 12.47, FDR = 5.21 · 10 227 ) in the ABC subtype ( Figure  3B). JAK3 associates with the interleukin (IL)-2 family g-chain receptors, transducing the signal from cytokines such as IL-2, IL-7, IL-9, IL-15, and IL-21 (Suzuki et al. 2000). Although JAK2 is often regarded as the primary partner of STAT3, constitutively active JAK3:STAT3 signaling has been reported as a poor prognosis indicator in anaplastic large-cell lymphoma and in colon cancer cell lines (Lin et al. 2005;Han et al. 2006). Our results suggest that STAT3 activates key regulators of cytokine signaling.
Conversely, EIF3B shows increased STAT3 binding in the GCB subtype cell lines (fold change = 0.01, FDR = 2.21 · 10 23 ; Figure 3B). As with JAK3, the STAT3 binding peak is just upstream of the TSS, a common location for gene regulation by transcription factors. EIF3B is a major scaffolding subunit of the eukaryotic translation initiation factor 3 (eIF-3) complex, a complex required for the initiation of protein synthesis. Intriguingly, short hairpin RNA knockdown of eIF-3B induces apoptosis in glioblastoma (Liang et al. 2012), suggesting that eIF-3B may also play a role in GCB DLBCL.
Systematic analyses reveal that STAT3 binding associates with oncogenic pathways in the ABC subtype: After the BRs with significant binding differences were identified, we associated them with nearby genes and then analyzed enrichment across a variety of ontologies using GREAT (McLean et al. 2010). All but 19 BRs fall within the putative regulatory domain of at least one gene, and the majority are associated with one (n = 4203) or two (n = 6095) nearby genes.
The Gene Ontology Biological Processes (GO BP) ontology shows that STAT3 BRs occupied in ABC DLBCL are located near genes involved in B-cell activation and mature B-cell differentiation ( Figure 4A). This result is consistent with STAT3's role in B-cell maturation and with the theory that more mature B-cells give rise n  Figure 4C). Although there are nearly as many STAT3 BRs with strong binding in the GCB cell lines as in the ABC lines, these BRs and their associated genes do not point to a clear oncogenic program like the ABC BRs do. Far fewer categories are enriched, and at lower significance levels; furthermore, the 1550 GCB-specific BRs are not enriched for any disease categories. The enriched PANTHER Pathway category "apoptosis signaling pathway" primarily contains genes that induce apoptosis for the GCB-bound BRs, but the ABC subtype list includes genes that both up-and down-regulate apoptosis ( Figure 4B). Taken together, these data show that in the ABC subtype, overactive STAT3 substantially changes its binding behavior and promotes a genetic program that blocks apoptotic signaling. On the basis of a 1-bp overlap, 3932 (38%) of the 10,337 STAT3 BRs observed in our DLBCL cell lines are also found in GM12878 cells, whereas 1143 (11%) are found in LoVo and 1063 (10%) in HeLa S3 (see Figure S2). It is not surprising that GM12878 has the greatest concordance with DLBCL, as it is the most similar cell type. DLBCL and GM12878 share 2958 BRs not found in either HeLa S3 or LoVo. As determined using GREAT (McLean et al. 2010), these BRs are found near genes enriched for type I interferon and cytokine signaling, many B-cell properties including activation and differentiation, DNA damage response, and the JAK/STAT pathway itself (including JAK1 and JAK3; STAT1, 3, 4, 5A, and 5B; and the down-regulators SOCS1, PIAS2, and PTPRC). More than one-third (36%, n = 1406) of these regions are subtype-specific in DLBCL, and three quarters of those (74%, n = 1049) are more strongly bound in the ABC subtype. This bias toward ABC regions may be due to the presence of the Epstein-Barr virus used to transform GM12878; EBV is a well-known oncovirus that promotes the development of hematopoetic cancers (Pallesen et al. 1991;Yu et al. 1996;Siu et al. 2002;Brady et al. 2007;Castillo et al. 2008) and whose presence contributes to poorer survival in DLBCL specifically (Park et al. 2007).
The 1063 STAT3 binding sites shared between DLBCL and HeLa S3 are enriched for genes involved in a number of cancer-related functions: chromatin structure, single-stranded DNA binding, DNA damage checkpoints, abnormal interferon levels, and the apoptotic signaling pathway. One third (n = 353) of these regions are subtypespecific in DLBCL, and the vast majority (90%, n = 316) are more strongly bound in the ABC subtype. The ABC-specific BRs in HeLa S3 and GM12878 fall near genes that enrich many of the same ontological categories, including immune response, the JAK/STAT pathway, interferon signaling and response, and cytokine signaling and response. They also share Disease Ontology categories that indicate immune system activation, such as leukemias of various types, allergies, hepatitis, and response to infection. In contrast, the 1143 shared LoVo BRs are primarily enriched for genes involved in RNA processing, splicing, and translation, rather than cancer-related categories. Bias toward ABC-specific BRs is not seen in the LoVo colon carcinoma line, where BRs specific to each DLBCL subtype are equally represented. Only 143 STAT3 BRs are found in all four cell types. These are generally not near genes enriched for any pathways or ontological categories. However, there is overlap with a curated gene sets for genes that are targets of c-Myc.

Subtypes show differential gene expression
To better understand the biology of the GCB and ABC subtypes and to relate their gene expression profiles to identified STAT3 targets, we examined global mRNA expression in both subtypes using RNA-Seq analysis (Morin et al. 2008). RNA-Seq was performed on mRNA isolated from all eight cell lines. Two biological replicates per cell line were analyzed with an average of 22 million mapped reads (86% mapping; see Table S2).
To identify genes that are differentially expressed in each subtype, we used the parameterized negative binomial model in DESeq (Anders and Huber 2010) to identify annotated RefSeq genes that displayed significantly different normalized read counts between the ABC and GCB cell lines, then corrected for multiple comparisons by using the Benjamini-Hochberg procedure (Benjamini and Hochberg 1995). We found many genes that exhibit differential gene expression in the GCB and ABC cell types, including several genes that play a role in apoptosis and cancer. This analysis also successfully re-identified many genes that have previously been shown to be biomarkers for each DLBCL subtype (Rosenwald et al. 2002;Tirado et al. 2012), including BCL2, IRF4, MME (CD10), and BCL6 (Table 2). Surprisingly, the established ABC subtype biomarker MUM1 (Hans et al. 2004;Sweetenham 2005;Anderson et al. 2009) showed no significant expression change (FDR = 0.92) but did have a STAT3 BR 344 bp upstream of its TSS with significantly more binding in GCB (fold change = 0.46, FDR = 4.89 · 10 23 ).
A total of 1545 genes are differentially expressed between subtypes at an FDR , 0.05 ( Figure 5), which is approximately 7% of the transcriptome. The vast majority of these genes (81%, n = 1251) are more highly expressed in the ABC cell lines. In contrast, only 294 genes (19%) are more highly expressed in the GCB cell lines. Interestingly, half (n = 765) of the genes with greater ABC subtype expres-sion demonstrate very low read counts (,5) in the GCB cell types. Conversely, only 21 genes that are more highly expressed in GCB are unique to that subtype ( Figure 6A). The prevalence of such "on/off" genes indicates that the major differences between ABC and GCB DLBCL are due almost exclusively to additional gene expression in ABC, rather than the two subtypes having divergent but equally active genetic programs.
Genes up-regulated in ABC subtype: Genes with GO BP categories related to cell adhesion, motility, and chemotaxis are enriched in ABC DLBCL cells relative to GCB cells (Figure 7). The expression of these pathways would prime tumors to spread rapidly from their lymph nodes of origin, and may partially explain why ABC lymphomas are so much more aggressive than their GCB counterparts.
The underlying gene list includes a large number of chemokine ligands and receptors, including several with known roles in cancer (Table 3). CCR7 shows expression limited to the ABC subtype ( Figure  6B). Increased CCR7 expression is clearly associated with metastasis to the lymph nodes in a broad variety of cancers (Schimanski et al. 2006;Kodama et al. 2007;Nakata et al. 2008;Pan et al. 2008;Arigami et al. 2009;Ishida et al. 2009). Furthermore, CCR7 and CXCR5 contribute to apoptosis resistance in B-cell2derived acute and chronic lymphocytic leukemia (Hu et al. 2004;Burkle et al. 2007).
Also up-regulated in the ABC subtype is transforming growth factor alpha (TGFA), which is a ligand for the epidermal growth factor receptor. Epidermal growth factor receptor in turn activates a signaling pathway for cell proliferation, differentiation, and development that significantly affects the progression of many human cancers, including brain, lung, breast, ovarian, prostate, and pancreatic tumors (Lu et al. 2001;Yang et al. 2011).
Genes in GO BP categories related to angiogenesis and hypoxia are also substantially enriched in the ABC cell lines (Figure 7). Angiogenesis is known to be important and prognostic in malignant lymphomas (Koster and Raemaekers 2005). The factors up-regulated in ABC DLBCL include FGF2, a major lymphangiogenic growth factor that promotes tumor growth and metastasis (Cao et al. 2008;Da et al. 2008). A number of growth stimulation factors and receptors are also up-regulated, such as platelet-derived growth factor alpha (PDGFA) and the beta subunit of the PDGF receptor (PDGFRB), both of which play essential roles in blood vessel development. Hypoxia induced genes include ARNT2, which has been reported to promote VEGF expression in teratomas and to protect neuronal cells from hypoxia (Maltepe et al. 2000). ARNT2 expression has not been reported previously as a mechanism in non-Hodgkin lymphoma. Thus, these analyses suggest new molecules that may be involved in ABC tumorigenesis.
Genes up-regulated in GCB subtype: GO analysis of the differentially expressed genes reveals that the GCB-enriched transcripts are involved in activation of lymphocytes (Figure 7). This list includes BCL6, a major biomarker for the GCB DLBCL subtype. Another notable GCB up-regulated gene is protein kinase C alpha (PKCa), which is encoded by PRKCA ( Figure 6B). PKCa is a protein kinase that can act as an oncogene (Besson and Yong 2000;Huttemann et al. 2012). However, in response to inflammatory signaling, PKCa also can negatively regulate NF-kB2induced genes by promoting the synthesis of the inhibitor IkBa (Han et al. 1999). Constitutive NF-kB signaling is a prominent feature of ABC DLBCL and includes crosstalk with STAT3 via up-regulation of IL-6 and IL-10 ( Motoyama et al. 2005). The high PKCa and low NF-kB levels in our GCB cell lines suggests that its predominant role in DLBCL may not be as an oncogene but rather as an inhibitor of the NF-kB pathway.

STAT3 induces an aggressive genetic program in DLBCL
To better understand the downstream consequences of STAT3 binding, we next integrated the binding and gene expression results. Genes associated with STAT3 BRs were correlated with their n  expression levels from the RNA-Seq data. Both the STAT3 binding and mRNA expression were evaluated for statistical significance, resulting in a list of genes that shows both nearby STAT3 binding changes and an accompanying relative mRNA expression level for the two cell types. ChIP-Seq identified a total of 10,337 STAT3 BRs, whereas the RNA-Seq analysis compared the expression level of 20,992 genes. Using GREAT analysis, we found that more than half of these genes (60%, n = 12,631) have a STAT3 BR associated with them. Many of these genes have multiple nearby BRs; likewise, many BRs are associated with two genes because they are located in a region in which putative regulatory domains overlap. Intersecting the ChIP-Seq and RNA-Seq datasets resulted in a total of 16,341 gene/BR matches (see Table S4). Forty percent (n = 6580) showed significant differences (FDR , 0.05) in at least one category: either gene expression or STAT3 binding differed markedly between the two DLBCL subtypes. A total of 823 of these gene/ BR matches, corresponding to 445 individual genes, were significantly different in both categories (Figure 8).
If STAT3 binding did not influence DLBCL gene expression, we would expect the nature of STAT3 binding near a gene to have no correlation with its expression level (Zhu et al. 2012). Instead, we observe that an increase in STAT3 binding is correlated with an increase in expression in most of these pairs (83%, n = 686), an association that is highly significant at P = 2.3 · 10 228 (Table 4). The strong association between binding and increased expression indicates that STAT3 binding is a significant contributor to the gene expression pattern in GCB and ABC DLBCL. The analysis indicates that STAT3 primarily functions as a transcriptional activator whereas a subset of its binding might contribute to transcriptional repression, which is consistent with its roles as previously described (Ouyang et al. 2009).
High STAT3 binding and gene expression in ABC: The most enriched category of BR/gene pairs consists of genes with greater STAT3 binding in the ABC subtype as well as increased gene expression in ABC: 531 pairs correspond to 287 different genes. STAT3 itself is among these, as are several of its known targets, including PRDM1 and SOCS3 (Table 5). We also confirm the findings from array CGH studies that NFKBIZ and the oncogene PIM2 are elevated in ABC lymphoma (Tirado et al. 2012), but with the added information that STAT3 shows elevated levels of binding in the regulatory domains of these genes. STAT3 also binds near many genes that are up-regulated in ABC cells relative to GCB cells; these genes have a wide range of functions that are pertinent to oncogenesis, especially apoptosis and cellular signaling. The CXCR5 chemokine receptor, which contributes to apoptosis resistance in B-cell-derived leukemias, is 6.83-fold more expressed in the ABC subtype (FDR = 8.83 · 10 27 ) and also has a distinct STAT3 peak 43 bp upstream of its TSS that is occupied only in ABC cell lines (Figure 9). STAT3 is also characterized as regulating the BCL2 protein family, which influences apoptosis. We specifically identify BCL2A1 (formerly BFL1), which is associated with three strongly bound STAT3 BRs and demonstrates a 4.43-fold increase in expression. High levels of BCL2A1 have been proposed as a mechanism of chemoresistance in chronic lymphocytic leukemia (Olsson et al. 2007).  STAT3 also has a binding peak just 7 bp upstream of the TSS of BATF, which is highly expressed in ABC cells. BATF belongs to the AP-1/ATF superfamily of transcription factors and dimerizes with the Jun family, acting as a negative regulator (Echlin et al. 2000). Although BATF typically promotes growth arrest and terminal differentiation (Liao et al. 2011a), chronically elevated BATF levels often are seen in lymphoid tumors and promote an impaired response to apoptotic signals, apparently mediated by the BCL2 family (Logan et al. 2012). Because normal levels of BATF stop cell division, this pathway presents a compelling target for drug modification.
We also identify APOL2 as a potential target of STAT3, suggesting a novel mechanism of apoptosis resistance in ABC DLBCL cells. A STAT3 BR is located 322 bp downstream of the APOL2 TSS (fold change = 2.64, FDR = 2.44 · 10 26 ), and its mRNA expression level rises 36.76-fold in the ABC subtype (FDR = 6.80 · 10 214 ). APOL2 has recently been found to protect endothelial cells from IFNg cytotoxic signaling (Liao et al. 2011b), and could play the same role in DLBCL in response to inflammatory signals.
Many interleukins and interleukin receptors show both elevated expression and STAT3 binding, including IL-6, IL-10, IL-10RA, and IL-12A. Both IL-6 and IL-10 are known to induce signaling via STAT3 upon binding their extracellular receptors; in turn, STAT3 up-regulates expression of the IL-10 and IL-6 genes to form an activation loop. NFKBIZ, which is also upregulated by STAT3, is a nuclear IkB protein that activates IL-6 transcription and may also play a role in this loop (Motoyama et al. 2005). Polymorphisms in IL-10 and its receptor subunit IL-10RA, also up-regulated in this study, have been implicated in the development of non-Hodgkin lymphoma generally and B-cell lymphoma specifically Rothman et al. 2006), suggesting that IL-10 pathway dysregulation promotes hematopoetic tumors.
IL-2RB produces the beta subunit of the IL-2 cytokine receptor. Its expression level increases 935.76-fold from GCB to ABC (FDR = 6.45 · 10 216 ) and is associated with increased STAT3 binding at a distal site about 27 kb downstream (FDR = 1.33 · 10 219 ). IL-2RA, which encodes the other half of the IL-2 receptor complex, is also upregulated in the ABC subtype, although it lacks a proximal STAT3 peak (fold change = 8.00, FDR = 1.95 · 10 23 ). These findings point to a potential therapy target for refractory cases of ABC DLBCL. The existing anticancer treatment denileukin diftitox (trade name Ontak) is an engineered protein combining IL-2 and diphtheria toxin, which binds to IL-2 receptors and introduces the diphtheria toxin into cells, killing them (Lansigan et al. 2010). Ontak has been specifically approved by the Food and Drug Administration for the treatment of cutaneous T-cell lymphoma but should be effective in treating any lymphoma that displays an elevated number of IL-2 receptors. Furthermore, increased expression of the IL-2RB gene is known to increase the efficacy of Ontak treatment (Shao et al. 2002).
High STAT3 binding and gene expression in GCB: A total of 95 genes have greater STAT3 binding and gene expression in GCB cells. Among these is MYBL1, a known biomarker for the GCB subtype (Blenk et al. 2007), which is overexpressed as expected in GCB cell lines and also has a statistically significant STAT3 peak 199 bp upstream of its TSS. Also included in this list are two tumor suppressor genes: ABI3 and TRIB2. In thyroid tumors, ABI3 expression blocks tumor progression on many fronts, including reducing anchorageindependent growth and in vivo tumor formation while increasing cellular senescence (Latini et al. 2011). In the GCB subtype, it demonstrates greater expression that corresponds to the presence of STAT3 424 bp downstream of its TSS (Figure 9). TRIB2 inhibits Red dots indicate STAT3 BRs that have significant differences in binding (FDR , 0.05) between subtypes and are also associated with genes that display significant differences in expression (FDR , 0.05); dark gray dots indicate STAT3 BRs that either show significant differences in binding or whose associated genes display a significant difference in expression; and light gray dots indicate other BRs.
n JNK, which ultimately reduces the antiapoptotic activity of BCL2 (Gilby et al. 2010). The loss of TRIB2 expression in ABC DLBCL is particularly problematic, as STAT3 increases the expression of BCL2 and its family members.
Anticorrelated STAT3 binding and gene expression: Seventeen percent (n = 137) of BR/gene pairs have strong STAT3 binding in one subtype but strong gene expression in the other. These 107 genes may be regulated by other factors, or STAT3 may serve as a repressor at these sites, potentially by blocking the binding of other factors. Fully 40% (n = 43) of these potential anticorrelated genes occur due to a mixture of STAT3 binding behavior: the gene is associated with many BRs, some of which are more occupied by STAT3 in ABC while others show stronger binding GCB. We observe greater CD22 expression in GCB DLBCL, associated with a STAT3 peak 138 bp upstream of its TSS. However, CD22 also has two STAT3 BRs near its 39 end that are more strongly bound in ABC (Figure 9). This finding suggests that STAT3 binding at the CD22 promoter drives transcription, whereas binding at the two downstream sites suppresses it. CD22 is a cell surface antigen expressed in a wide variety of non-Hodgkin lymphomas, and anti-CD22 monoclonal antibody treatments currently are being developed (Lu et al. 2006;Li et al. 2013). However, they may be less successful in treating ABC DLBCL due to lowered expression of this antigen.
BCL6 is the most extreme example of this mixed binding phenomenon. It is the primary biomarker for the GCB subtype of DLBCL and is known to be regulated by STAT3. We found that expression of BCL6 is extremely high in GCB lines, three times the level found in our ABC cell lines. However, its regulation is complex: there are 24 STAT3 BRs within the putative regulatory domain of BCL6, which includes a large upstream gene desert that contains considerable STAT3 binding activity. Seven of these BRs show a significant binding differential between the subtypes. Two of them are more strongly bound in ABC, but the other five are more strongly bound by STAT3 n n STAT3, signal transducer and activator of transcription 3; TSS, transcription start site; ChIP, chromatin immunoprecipitation; FDR, false-discovery rate; ABC, activated B-cell-like; FDR, false-discovery rate; obs, observed counts; exp, expected counts.
in GCB (see Figure S3 and Table S3). This complexity suggests that the overall sum of STAT3 binding contributes to the final regulation of BCL6.
The combination of whole-genome profiling approaches shows differential STAT3 binding sites near almost a third of all genes that differ in expression between GCB and ABC DLBCL, leaving little doubt that STAT3 is a regulator of the aggressive clinical phenotype and drug resistance displayed by the ABC subtype. The STAT3 binding and expression results suggest possible models by which STAT3 regulates biological pathways to promote aggressive oncogenic behavior in ABC DLBCL. Ontological analysis of the genes upregulated in ABC DLBCL with increased nearby STAT3 binding indicates that STAT3 promotes cell migration, blood vessel development, and the inflammatory response ( Figure 10). Furthermore, it ensures its own continued high expression by up-regulating its pathway partners and upstream signaling molecules, such as IL-6 and IL-10, and promotes cross-talk with the nuclear factor-kB pathway. Most critically, STAT3 strongly up-regulates many mechanisms of apoptosis resistance. These effects synergize and likely lead to increased cell proliferation and distributions of cells, which are hallmarks of aggressive lymphomas. These results provide novel insights into the genetics of oncogenesis in the ABC form of diffuse large B-cell lymphoma and identify a substantial number of plausible downstream mechanisms for the treatment-resistant phenotype of this subtype. Figure 9 STAT3 peaks associated with gene expression changes. Examples of differentially occupied STAT3 BRs are shown. Blue tracks represent gene expression; orange tracks represent STAT3 binding; pink bars show noteworthy STAT3 BRs. CXCR5 shows increased gene expression correlated with STAT3 binding (chr11: 118,754,2592118,754,737) in the ABC cell lines: SU-DHL2, OCI-Ly3, OCI-Ly10, and U-2932. ABI3 shows strong increased gene expression correlated with STAT3 binding (chr17: 47,287,858247,288,168) in the GCB cell lines: SU-DHL4, SU-DHL6, SU-DHL10, and OCI-Ly7. CD22 shows increased gene expression in the GCB cell lines, but a mixture of STAT3 binding behavior: STAT3 has more GCB binding near the TSS (chr19: 35,819,821235,820,060) but more ABC binding at two sites near the 39 end (chr19: 35,837,564-35,837,836; chr19: 35,838,409235,838,860).

ACKNOWLEDGMENTS
We thank the ENCODE Consortium for generating some of the data and techniques used in this work. We would also thank the laboratories of Dr. Louis Staudt at the National Cancer Institute and Dr. Ash Alizadeh at the Stanford School of Medicine for the generous donation of some of the DLBCL cell lines used in these experiments. This research was supported in part by the National Institutes of Health "Analysis of Human Genome Using Integrated Technologies" Centers of Excellence in Genomic Science Award Grant no. P50 HG002357 and by the ENCODE Consortium under National Institutes of Health grant NIH 5U54 HG 004558.