Regulatory network controlling tumor-promoting inflammation in human cancers

Using an inducible, inflammatory model of breast cellular transformation, we describe the transcriptional regulatory network mediated by STAT3, NF-κB, and AP-1 factors on a genomic scale. These regulators form transcriptional complexes that directly regulate the expression of hundreds of genes in oncogenic pathways via a positive feedback loop. This inflammatory feedback loop, which functions to various extents in many types of cancer cells and patient tumors, is the basis for an “inflammation” index that defines cancer types by functional criteria. We identify a network of non-inflammatory genes whose expression is well correlated with the cancer inflammatory index. Conversely, the inflammation index is negatively correlated with expression of genes involved in DNA metabolism, and transformation is associated with genome instability. Inflammatory tumors are preferentially associated with infiltrating immune cells that might be recruited to the site of the tumor via inflammatory molecules produced by the cancer cells.


INTRODUCTION
Tumor-promoting inflammation is a hallmark of cancer and plays important regulatory roles during cell transformation, invasion, metastasis and treatment resistance (Coussens and Werb, 2002;Mantovani et al., 2008;Grivennikov et al., 2010). An inflammatory reaction during tumor development occurs in near all solid malignancies (de Martel and Franceschi, 2009). Pre-existing inflammation promotes subsequent cancer development, accounting for 15%~20% of cancer deaths (Mantovani et al., 2008). Tumor-promoting inflammation also plays critical roles in immunosuppression (Grivennikov et al., 2010). For example, STAT3 promotes the expression of PD-L1 and PD-L2 in cancer cells, suppressing immune cell activity (Chen and Han, 2015).
Tumor-associated inflammation results from both an intrinsic pathway, in which mutations in cancer cells activate inflammatory gene expression, and an extrinsic pathway, in which cytokines and chemokines secreted by tumor-associated immune cells create inflammatory microenvironments (Mantovani et al., 2008). Oncogenes can trigger a gene expression cascade, resulting in activation or overexpression of pro-inflammatory transcription factors such as NF-κB, STAT3 and AP-1, with the resulting production of cytokines and chemokines (Coussens and Werb, 2002;Iliopoulos et al., 2009;Grivennikov et al., 2010). For example, NF-κB signaling pathway is activated upon activation of oncogenes RAS and MYC, through activation of IL-1β (Shchors et al., 2006;Guerra et al., 2007;Iliopoulos et al., 2009), and the proto-oncoprotein Src kinase can directly phosphorylate and activate STAT3 (Turkson et al., 1998).
In previous work, we have described an inducible model of cellular transformation in which transient activation of v-Src oncoprotein converts a non-transformed breast epithelial cell line into a stably transformed state within 24 hours (Iliopoulos et al., 2009;Hirsch et al., 2010). The stably transformed cells form foci and colonies in soft agar, show increased motility and invasion, form mammospheres, and confer tumor formation in mouse xenografts 4 (Iliopoulos et al., 2009;Hirsch et al., 2010). This epigenetic switch between stable nontransformed and transformed states is mediated by an inflammatory positive feedback loop involving NF-κB and STAT3 (Iliopoulos et al., 2009;Iliopoulos et al., 2010). By integrating motif analysis of DNase hypersensitive regions with transcriptional profiling, we found that > 40 transcription factors are important for transformation and identified putative target sites directly bound by these factors (Ji et al., 2018).
A few transcriptional regulatory circuits involved in this transformation model have been identified, and these are important in some other cancer cell types and human cancers (Iliopoulos et al., 2009;Iliopoulos et al., 2010;Iliopoulos et al., 2011b;Polytarchou et al., 2012). During transformation, STAT3 acts through pre-existing nucleosome-depleted regions bound by FOS, and expression of several AP-1 factors is altered in a STAT3dependent manner (Fleming et al., 2015). However, the connection between STAT3, NF-κB, and AP-1 factors as well as the underlying transcriptional regulatory circuits has not been described on the whole-genome level.
Here, we define the transcriptional network mediated by the combined action of NF-κb, STAT3 and AP-1 factors (JUN, JUNB and FOS) on a genomic scale in this breast transformation model. In contrast to previous studies (Fleming et al., 2015;Ji et al., 2018), this network is defined by genes that are induced during transformation by the binding of NF-κb, STAT3 and AP-1 factors to common target sites either through their cognate motifs or via protein-protein interactions. Based on this common NF-κb, STAT3 and AP-1 network, we develop a "cancer inflammation index" to define cancer types, both in cell lines and in patients, by functional criteria. As this inflammation index is based on the common regulatory network, it is distinct from and more specific than indices based simply on gene expression profiles that arise from multiple regulatory inputs. In addition, we identify many non-inflammatory genes whose expression is positively or negatively correlated with the cancer inflammation index, leading to the observation the inflammation is linked functionally to other aspects of cancer as well as genomic instability. Lastly, we show that inflammatory 5 tumor samples preferentially contain contaminating immune and stromal cells from the tumor microenvironment, consistent with the idea that immune cells might be recruited to the site of the tumor via inflammatory molecules produced by the cancer cells.
7 correlation values are lower than the correlation values between biological replicates ( Figure   S3B).
Additional results suggest that STAT3, NF-κB and AP-1 factors co-bind to target sites as a multiprotein complex. First, the median distance of peak summits for all pairwise combinations of AP-1 factors, STAT3, and NF-κB ranges between 15-30 bp, and these values are comparable to those obtained for biological replicates of the relevant individual factors ( Figure 1B). Second, co-immunoprecipitation experiments show that STAT3, NF-κB and AP-1 factors interact with each other in the nucleus ( Figure 1C).

Sequence motifs associated with binding of STAT3, NF-κb and AP-1 factors
To address which factors are primarily responsible for binding site specificity, we performed motif analysis on 50 bp sequences around peak summits. As expected, 60% of AP-1 factor binding peaks contain an AP-1 motif ( Figure S3C). Interestingly, 38% of STAT3 binding sites and 30% of NFKB1 sites also have an AP-1 motif, while 24% of STAT3 sites have a STAT motif and 16% NFKB1 sites have a NF-κB motif ( Figure S3C). Furthermore, the AP-1 motif is well located in STAT3 and NFKB1 peak summits ( Figure 1D). We did not find significant co-localization of AP-1, STAT and NF-κB motifs in the same CRRs. These results indicate that a significant fraction of STAT3 and NF-κB binding is mediated through the interaction with AP-1 factors. In contrast, only a small minority of AP-1 binding sites contain a STAT or NF-κb motif ( Figures 1D and S3C), and STAT and NF-κb motifs show modest enrichment around AP-1 peak summits ( Figure 1D). Thus, binding of AP-1 factors occurs predominantly via interactions with AP-1 motifs, presumably reflect a direct protein-DNA interaction. In contrast, in addition to directly binding via their motifs, STAT3 and NF-κb can also bind to AP-1 motifs, presumably via protein-protein interactions with AP-1 factors. However, we did not observed significant motif differences in AP-1 sites bound by AP-1 factors alone or those co-bound with STAT3 and/or NF-κb.

STAT3, NF-κb and AP-1 factors co-regulate key genes in various oncogenic pathways
Co-binding of STAT3, NF-κB and AP-1 factors regulates gene expression and chromatin status. Genes up-regulated during transformation ( showing the most drastic effects. As there are multiple members of the NF-κb and AP-1 families, it is likely that the weaker effect on transcription are due to redundancy among family members. With respect to chromatin structure, regions bound by increasing numbers of these factors tend to have higher accessibility ( Figure S4A) and acetylation levels ( Figure   S4B). In addition, differential binding levels of STAT3, NFKB1 and AP-1 factors during transformation are positively correlated with dynamic chromatin accessibility (Pearson Correlation Coefficient >= 0.44) ( Figure S4C), and open chromatin regions bound by more factors tend to show increased accessibility ( Figure S4D). We identified 1,461 genes that are common targets of STAT3, NFKB1, JUN, JUNB and FOS and that show increased binding of at least four factors (>1.5 fold) during transformation and downregulation upon at least four factor knockouts (Table S1). These genes are enriched (Benjamini-Hochberg FDR < 0.005) in cancer-related processes ( Figure   S5), and they include genes involved in cell signaling cascades (e.g. RAB13, IF116, ZAK), inflammatory response (e.g. IL1B, IL1R1, SERPINA1), cell proliferation (e.g. CSF3, E2F7, E2F8) regulation of apoptosis (e.g. PIM1, CARD6, BCL2L1), angiogenesis (e.g. RHOB, CEGFC, EPAS1), and cell migration/metastases (e.g. LAMB3, CXCL3 and PLAU)( Figure   2D, E). Cancer stem cells are generated during the transformation process in our model (Iliopoulos et al., 2011a), and STAT3, NF-κb and AP-1 factors activate key genes (e.g. CD44, CXCR1 and ITGA7) mediating cancer stem cell formation. Cancer metabolism is a key component of oncogenesis (Pavlova and Thompson, 2016), and important metabolic enzyme genes (e.g. HK2 and NAMPT), are also regulated by these factors (Figure 2D, E).

9
Thus, STAT3, NF-κb, and AP-1 are the major factors involved in tumor-promoting inflammation, and they bind to and promote the expression of key genes in various oncogenic pathways ( Figure 2E).
Thus, the essence of the loop is that STAT3, NFKB1, and AP-1 directly activate upstream regulators that trigger the activation of intra-cellular signaling cascades, phosphorylate the transcription factors, and promote their nuclear localization and transcriptional activation. In accord with previous studies on IL6 (Iliopoulos et al., 2009), reducing the activation levels of JAK2, JNK, and IL1 receptor via inhibitors or antagonists results in decreased transformation efficiency ( Figure S6). We define the 27 genes in the IL6/STAT3, IL1/NF-κb and TNF/AP-1 signaling pathways ( Figure 3A) as the core positive feedback loop that maintains the transformed state.
An "inflammation index" to measure the inflammation level of a cancer cell line 10 The Cancer Cell Line Encyclopedia (CCLE) database contains gene expression data for 1,036 human cancer cell lines from over 20 developmental lineages (Barretina et al., 2012). To measure the inflammation level for each cancer cell line, we developed a scoring system called the "inflammation index", which is calculated as the expression values of 27 genes (normalized to the median expression levels in the 1,036 cell lines) in the positive feedback loop. Importantly, this index is not simply based on a subset of genes that are induced during transformation, but rather genes that are direct targets of the STAT3/NF-κB/AP-1 regulatory network. The inflammatory levels gradually increase during ER-Src transformation ( Figure 4D), and they are highly variable (>5 fold) among different cancer cell lines ( Figure 4A). Similarly, in a fibroblast cell transformation model involving stepwise addition of the SV40 and RAS oncogenes to immortalized fibroblast, the inflammatory levels increase upon transformation ( Figure 4E).
On average, head and neck as well as pancreatic cancer cell lines are most inflammatory, while the autonomic ganglia and blood cancer cell lines are least inflammatory ( Figure 4F). However, cancer cell lines from the same developmental lineage can show high variance in inflammatory levels, which is correlated with their genetic subtypes. For example, cell lines from non-small cell lung cancers are more inflammatory than those of small cell lung cancers ( Figure 4G), and triple negative breast cancer cell lines with p53 mutations are more inflammatory than other subtypes ( Figure 4H).

The inflammatory loop is active in tumors from cancer patients
Although cancer cell lines are derived from tumors, long-term propagation of cell lines under artificial conditions raises the possibility that cell line data might be misleading with respect to cancer. We addressed the relevance of the inflammatory loop in human tumors using RNA-seq data in the Cancer Genome Atlas database (Cancer Genome Atlas Research et al., 2013). In accord with the results in cancer cell lines, genes in the IL6/STAT3, IL1/NF-κB and JNK/AP-1 inflammatory loop are co-expressed in human breast tumors ( Figure 5A).
Similarly, triple negative breast tumors are more inflammatory than other types of breast tumors ( Figure 5B). Moreover, the median inflammatory index value of all tumor samples from a given developmental lineage is highly correlated with that of cells lines from the same developmental lineage (Pearson correlation coefficient = 0.82; P-value < 10 -5 ) ( Figure 5C).
As we combined all tumors for a given lineage, it seems unlikely that this analysis is significantly compounded by differences in the tumor microenvironment. These observations indicate that, with respect to inflammation and gene regulation profiles, cancer cell lines are good models for tumors, and the inflammatory loop is very relevant for many types of human cancer.

Identification of other genes whose expression is correlated to the inflammation index
Previous analyses of the ER-Src model designed to identify oncogenically relevant genes have relied on differential gene expression upon transformation and/or direct regulation by NF-κB, STAT3, and AP-1 factors. However, it is highly likely that this approach will miss critical genes that are part of the tumor-promoting inflammation process.
As an alternative approach, for every gene, we calculated the Spearman correlation coefficient between its expression level and the inflammation index across cancer cell lines for various developmental lineages. The correlation values are significantly conserved among developmentally distinct cancers, again indicating a widespread role of the inflammatory loop and it target genes ( Figure 6A).
Genes showing higher positive correlation with the inflammation index are more likely to be direct transcriptional targets of STAT3/NF-κB/AP1 and tend to be upregulated during transformation ( Figure 6B, 6C and 6D). We identified 1,303 genes showing significant positive correlation of expression with the inflammation index (median Spearman's Rank Correlation Coefficient across cancer cell types > 0.18, false discovery rate < 0.005)( Figure   6B). Aside from genes involved in the inflammatory response, these genes are enriched in biological pathways such as angiogenesis, cell proliferation, apoptosis, intracellular signaling cascade and cell migration ( Figure 6E). These pathways, which are defined by the inflammatory index and not transformation per se, are in excellent accord with the pathways accord with pathways activated during ER-Src transformation ( Figure 6E), thereby providing independent evidence that the inflammatory loop and oncogenic pathways uncovered in the ER-Src cell transformation model are highly relevant for different types of human cancer.

A correlation between inflammation and genome instability in cancer cell lines
1,369 genes show negative correlation between expression levels and inflammation index, and these are enriched in pathways including DNA metabolic process, DNA replication, DNA repair and cell cycle (Benjamin FDR < 10 -13 ). This observation suggests that oncogenic-associated inflammation is inversely related to genome stability, and indeed the genes showing most negative correlation (e.g. MSH2, FANCF, BRCA1) are regulators of genome instability. Interestingly, the genes related to genome instability are not transcriptionally regulated during ER-Src cell transformation.
To examine the link between inflammation and genome instability, we performed micronucleus staining of cells during ER-Src transformation. Indeed, more cells contain micronuclei as transformation processes ( Figure 6F, G), with 2.2% cells containing micronuclei after 72 hr after induction of transformation as compared to virtually no cells before transformation. Thus, the inflammation-mediated process of transformation is associated with decreased genome stability.

Inflammatory tumor samples contain increased levels of non-cancer cells
As shown above, the inflammation indices of cancer cell lines for a particular cancer type are highly correlated with the corresponding tumor samples ( Figure 5C). However, unlike samples from cancer cell lines, tumor samples not only contain the cancer cells, but also immune and stromal cells from the tumor microenvironment. Such tumor impurity complicates the analysis of gene expression profiles of tumor cells, particularly as immune cells are also inflammatory. We therefore took several approaches to disentangle the inflammatory nature of cancer cells from other cells in a tumor.
First, for numerous tumor samples we analyzed the relationship between the inflammatory index and tumor purity (fraction of cancer cells in a sample), which has been estimated from cancer-specific genetic mutations or cell type-specific gene signatures (Aran et al., 2015). In general, the inflammation level increases as the tumor purity decreases As an alternative approach, we selected breast patient samples with similar levels of tumor purity but different levels of inflammation (divided into 4 bins) and analyzed expression of non-inflammatory genes whose expression is strongly correlated with the inflammatory index ( Figure 7B). Expression of these transformation-related, but noninflammatory genes, which are direct targets of STAT3/NF-κb/AP1 and are involved in angiogenesis, apoptosis and cell migration, increases as the inflammation index increases.
Similar results are observed in all other types of cancer we examined such as melanoma, lung, head and neck, and glioblastomas ( Figures 7C and S8). These results indicate that the 14 inflammatory loop and associated network are active in various types of cancers, although we can't exclude the formal possibility that non-malignant cells in the various samples might be in different states.
Lastly, we created a "progression index" that is strongly correlated with the inflammatory index ( Figure 7D), but is based on 98 genes that regulate "migration/metastasis", "apoptosis" and "angiogenesis", but are not currently annotated as inflammatory or part of the immune response. This progression index, should reflect the regulatory circuits involved in cellular transformation and tumor formation, but not normal immune cells. Indeed, across the large set of tumor samples, the score of non-inflammatory index inversely correlated with tumor purity ( Figure 7E). All of these observations indicate that tumors containing inflammatory cancer cells preferentially contain non-cancer cells.
Thus, tumor-associated inflammation level is positively correlated with the complexity of microenvironment and the presence of immune cells.

DISCUSSION
NF-κB, STAT3, and AP-1 factors are the core of the positive feedback loop controlling

tumor-promoting inflammation
In the ER-Src cellular transformation model, a transient inflammatory stimulus mediates an epigenetic switch from a stable non-transformed cell to a stable transformed cell (Iliopoulos et al., 2009). Epigenetic switches are the basis of multicellular development, and they occur by activating a positive feedback loop that maintains the altered state. In the ER-Src model, Src activates the inflammatory transcription factors STAT3 and NF-κB that form the basis of the inflammatory feedback loop that is required for maintenance of the transformed state (Iliopoulos et al., 2009;Iliopoulos et al., 2010;Fleming et al., 2015;Ji et al., 2018).
Here, we show that AP-1 factors play a critical role in the inflammatory feedback loop.
AP-1 factors are not only important for transformation, but they form complexes with STAT3 and/or NF-κB that bind target sites. Specifically, these factors coimmunoprecipitate, and their binding profiles are coincident at many target sites. Most, and perhaps all, of the sites where co-binding occurs contain AP-1 motifs, suggesting that the AP-1 factors directly interact with DNA, whereas STAT3 and NF-κB are often recruited via interactions with the AP-1 factors. At some sites, STAT3 and NF-κB bind via their own motifs in the absence of AP-1 factors. At present, it is unclear whether AP-1, NF-κB, and STAT3 can form a ternary complex at individual sites or if NF-κB and STAT3 form independent complexes with AP-1 factors at these sites.
In addition to their roles in inflammation per se, STAT3, NF-κB and AP-1 work together to regulate key genes in oncogenic pathways such as angiogenesis, apoptosis, cell migration and epithelial to mesenchymal transition. Expression of many genes in these pathways is induced upon transformation in a manner that is linked to increased binding of all these factors. Moreover, expression of many common genes is reduced upon knockouts of individual factors, indicating that all of these factors contribute to expression of these genes. Our identification of a common STAT3/NF-κB/AP-1 network is distinct from, but not inconsistent with, previous observations that STAT3 and NF-κB have different (i.e. nonoverlapping) binding sites and gene expression effects during the transformation process (Fleming et al., 2015).
The positive feedback loop elucidated here is considerably more complex than

Patient tumor samples contain cancer cells and immune (and other non-cancer) cells, all
of which contribute to the transcriptional profile. Interestingly, over a large number of tumor samples, there is an inverse relationship between the inflammatory index and the estimated degree of sample purity. In principle, this relationship could merely reflect the fact that immune cells also express many inflammatory genes, which could significantly contribute to the observed inflammation index of the sample.
We attempted to distinguish the contributions of cancer and immune cells to the transcriptional profile by analyzing many non-inflammatory genes (either in specific pathways or as an overall index) whose expression is very strongly correlated with the inflammatory index and hence to the STAT3, NF-κB, AP-1 regulatory network. These observations suggest a dynamic interplay between cancer cells and immune cells that is linked to the inflammation index of the cancer cells in the sample. We propose that cytokines

Cell culture
The inducible model of cellular transformation involves MCF-10A, a non-transformed mammary epithelial cell line (Soule et al., 1990) containing ER-Src, a derivative of the Src kinase oncoprotein (v-Src) that is fused to the ligand-binding domain of the estrogen receptor (Aziz et al., 1999). Cells were cultured in DMEM/F12 medium with the supplements as previously described (Iliopoulos et al., 2009;Hirsch et al., 2010). Tamoxifen

CRISPR Knockouts
CRISPR-blasticidin lentiviral plasmid was constructed by replacing puromycin resistance gene with blasticidin resistance gene in LentiCRISPR V2 plasmid (Addgene, #52961). Table S2 is the list of the oligo sequences used to clone into CRISPR-blasticidin plasmid. CRISPR-blasticidin plasmid and three lentiviral plasmids, VSV-G, GP and REV were co-transfected into 293T cells to produce lentiviruses as the previous publication (He et al., 2011). After CRISPR lentiviruses infection, ER-Src cells were selected with blasticidin 10 μg/ml for 3 days to generate CRISPR knockout stable cell lines. The knockout efficiencies of the transcription factors were assessed by Western blotting.

Analyses of transcription factor binding
Transcription factors tend to localize into cis-regulatory regions (CRRs) that regulate gene expression. We merged overlapping peaks of all factors to define CRR regions. For each CRR, we measured factor binding levels as Reads per Million (RPM) using ChIP-seq data, and chromatin accessibility based on DNase-seq data. The peak summit of each factor binding site was defined based on MACS (Zhang et al., 2008). In Figure 1B, we plotted the distance between peak summits of paired transcription factors, located in the same CRR. For each factor binding site, we took 50 nt around the peak summits to perform the motif analyses, using the HOMER (Heinz et al., 2010). We used the position weight matrices (PWM) of STAT, NF-κb and AP1 in HOMER (Heinz et al., 2010). In Figure 1D, we plotted the distribution of STAT, NF-κb and AP1 motifs around peak summits of transcription 22 factors. As the control, we shuffled the nucleotide positions of PWM, and kept the A/T/G/C occurrence frequency of the motifs as the same. We created the shuffled motifs for 50 times for each motif and plotted their occurrence around the factor peak summits. To examine the contribution of factor binding to gene expression, we assigned factor binding peaks to the closest expressed genes within a distance of 200 megabases, summed up the ChIP signal, and calculated the binding level fold-change during transformation.

Gene Ontology analyses
The Database for Annotation, Visualization and Integrated Discovery (DAVID) (Huang da et al., 2009) was used for gene ontology analyses.

Data availability
All sequencing data that support the findings of this study have been deposited in the National Cancer for Biotechnology Information Gene Expression Omnibus (GEO) and are accessible through the GEO series accession numbers GSE115597, GSE115598 and GSE115599. All computational codes are available from the authors upon request.
As the control, we randomly shuffled AP-1, STAT and NF-κb consensus motifs and plotted the motif distribution around peak summits.

Figure S8. Relationship between tumor purity and inflammation
Randomly picked tumor samples with different inflammation levels and similar purity were examined for expression levels of genes in the indicated oncogenic pathways.

Figure 6
A.

Figure S3
AP B.

Lung cancers
Head and neck cancers