Cancer-Stimulated Mesenchymal Stem Cells Create a Carcinoma Stem Cell Niche via Prostaglandin E₂ Signaling

Carcinoma cells recruit mesenchymal cells into the tumor-associated stroma; these mesenchymal cells then proceed to modify the stroma, helping to establish a tissue microenvironment that favors tumor progression. Paracrine signals emanating from the resulting tumor-associated stroma can subsequently modulate the behavior of the carcinoma cells (1).

Among the recruited stromal cells are bone marrow–derived mesenchymal stem cells (MSC), which are known to exhibit multipotent differentiation potential (2). In the context of cancer pathogenesis, MSCs contribute to the formation of fibroblast and myofibroblast populations in the tumor-associated stroma (3, 4) and promote the growth, progression, and metastasis of tumors (3, 5, 6). Precisely how MSCs influence tumor progression is, however, poorly understood.

Elevated COX-2 mRNA and protein levels are found in many malignant tissues and are often associated with poor clinical outcome (7). The tumor-enhancing effects of COX-2 are generally ascribed to its role in producing prostaglandin E₂ (PGE₂), which has pleiotropic effects on cell proliferation, survival, angiogenesis, motility, and invasiveness (8). In addition to the neoplastic cells themselves, cells of the tumor-associated stroma contribute to elevated COX-2 expression (9, 10). However, it has been unclear whether the PGE₂ that promotes tumor progression derives from neoplastic cells, fibroblasts, macrophages, or some combination of these cell types.

Independent of these questions is the issue of heterogeneity of the neoplastic cells within carcinomas. Observations of a variety of human cancer types have revealed the existence of tumor-initiating cells (TIC), often called cancer stem cells (CSC), which coexist as minority populations within tumors, together with a majority population of cancer cells that lack tumor-initiating ability (11). Passage by neoplastic epithelial cells through an epithelial–mesenchymal transition (EMT) allows these cells to approach the stem cell state (12, 13). Moreover, EMT programs are known to be induced by heterotypic signals that epithelial cells receive from the microenvironment (14). However, the nature of these heterotypic signals and the identities of the stromal cells that release them remain poorly understood.

We show here that, in response to stimulation by carcinoma cells, MSCs express greatly elevated levels of PGE₂. The resulting PGE₂, together with cytokines also induced in the MSCs, contribute to entrance of nearby carcinoma cells into a stem cell–like state.

We initially studied the interactions in culture of LoVo and SW1116 human colorectal carcinoma cells with MSCs. Minimal PGE₂ accumulation was observed in pure LoVo, SW1116, or MSC cultures monitored over a 72-hour period (Fig. 1Aa and Supplementary Fig. S1A). However, PGE₂ levels increased by approximately 6.5-fold when the LoVo cells were cocultured with twice the number of MSCs for 48 hours and increased by approximately 60-fold after 72 hours of coculture. Correspondingly, levels of the COX-2 enzyme were also increased in the coculture (Fig. 1Ab). In contrast, there was no PGE₂ increase in SW1116/MSC cocultures. Of note, PGE₂ production was induced equally strongly in [LoVo][MSC] transwell cocultures, which only permitted their intercommunication via soluble factors (Supplementary Fig. S1B) and in MSCs treated with LoVo-conditioned medium (LoVoCM; Supplementary Fig. S1C and S1D). Hence, soluble factors secreted by LoVo cells were responsible for inducing PGE₂ production by MSCs.

Other work (in Supplementary Data, Section 1; Supplementary Fig. S1E) showed that the secretion of interleukin-1β (IL-1β) and IL-1α by carcinoma cell populations was correlated with their respective abilities to elicit PGE₂ production in cocultured MSCs. Thus, when MSCs were treated with recombinant IL-1β or IL-1α, the induced PGE₂ levels were similar to or higher than those induced by LoVoCM (Fig. 1Ba). Conversely, IL-1α– and IL-1β–neutralizing antibodies attenuated by 60% the production of LoVoCM-induced PGE₂ (Fig. 1Ba). In addition, short hairpin RNAs (shRNA) directed against IL-1α and IL-1β mRNAs expressed in LoVo cells decreased by 84% the ability of their conditioned media to induce PGE₂ in MSCs (Fig. 1Bb) and recombinant IL-1ra, a natural antagonist of the IL-1 receptor, attenuated by 73% the LoVoCM-induced PGE₂ production by MSCs (Fig. 1Bc). IL-1 α/β secreted by the carcinoma cells was largely responsible for the PGE₂ production by MSCs, whereas IL-1ra could antagonize this induction.

The levels of the IL-1α, IL-1β, and IL-1ra mRNAs and secreted proteins were quantified in 6 human breast carcinoma cell lines and the SW1116 colorectal carcinoma cell line, in addition to the LoVo cells examined above. Cells of the LoVo, HCC1806, BT549, SUM149, and SUM159 lines expressed elevated levels of IL-1α and/or IL-1β, but relatively low levels of IL-1ra; conversely, MDA-MB-453, MDA-MB-231, and SW1116 cells secreted little or no detectable IL-1α/β and/or relatively high levels of IL-1ra (Supplementary Fig. S1F and Fig. 1Ca).

When cultured individually, the carcinoma cell lines expressed low or undetectable PGE₂ levels (Fig. 1Cb). However, upon coculture with MSCs, the IL-1α/β–secreting carcinoma cells induced approximately 80- to 500-fold increases of COX-2 and PGE₂ production, whereas the carcinoma cells secreting low levels of IL-1α/β failed to stimulate COX-2 or PGE₂ formation (Supplementary Fig. S1G and Fig. 1Cb). In addition, IL-1α-/IL-1β–neutralizing antibodies and IL-1ra attenuated PGE₂ production induced in MSCs by the conditioned media from the various IL-1–secreting carcinoma cells (Supplementary Fig. S1H). Hence, the ability to stimulate PGE₂ in MSCs seemed to be a frequent but not universal property of breast and colon carcinoma cells and was directly correlated with their abilities to signal via secreted IL-1α/β. Because IL-1α and IL-1β were both capable of PGE₂ induction, we used the term “IL-1” to refer to both IL-1α and IL-1β in the text that follows.

In addition to PGE₂, GRO-α, IL-6, IL-8, and regulated upon activation, normal T-cell expressed, and secreted (RANTES) in the culture media of LoVo/MSC cocultures increased by 34-, 10-, 79-, and 21-fold, respectively, following 120 hours of coculture (Fig. 2A). In [LoVo][MSC] transwell cultures, in which direct contact between the LoVo cells and MSCs was prevented, PGE₂, GRO-α, IL-6, and IL-8 were induced to comparable levels; in contrast, RANTES expression was not elevated (Fig. 2A). Following direct coculture, RANTES production occurred far more rapidly than did the accumulation of the other cytokines or PGE₂ (Supplementary Fig. S2A).

We also found that GRO-α, IL-6, and IL-8, like PGE₂, were induced by LoVoCM in MSCs (Supplementary Fig. S1B). To determine whether IL-1 was able, by itself, to induce concomitant production in MSCs of the 3 cytokines and PGE₂, we assessed mRNAs levels in MSCs that had been treated for 48 hours with vehicle or recombinant IL-1 (Fig. 2B). The resulting 10-fold and 4-fold increases in MSCs of COX-2 and microsomal prostaglandin-E synthase-1 [mPGES-1 (a second PGE₂ biosynthetic enzyme)] mRNAs, the increase of COX-2 protein, and the 10- to 100-fold decrease of 15-hydroxyprostaglandin dehydrogenase (15-PGDH) mRNA (encoding the PGE₂-degrading enzyme) confirmed the key role of IL-1 in modulating the levels of enzymes governing PGE₂ production and accumulation in MSCs. Moreover, IL-1 treatment alone elicited substantial increases (36- to 440-fold) of IL-6, IL-8, and GRO-α mRNAs in MSCs. COX-2, IL-6, IL-8, and GRO-α mRNA induction was detectable within 30 minutes of IL-1 exposure and reached a maximum at 1 to 2 hours thereafter (Supplementary Fig. S2B). This key role of IL-1 was further confirmed by knocking down IL-1α/β mRNAs in LoVo cells, resulting in a 60% to 90% decrease in the induced levels of IL-6, IL-8, and GRO-α mRNAs (Supplementary Fig. S2C). Thus, IL-1 was both necessary and sufficient to induce PGE₂, IL-6, IL-8, and GRO-α production in MSCs.

We also confirmed that LoVoCM and IL-1 could induce comparable levels of PGE₂ (Supplementary Fig. S2D) and cytokines (Supplementary Fig. S2E) in other types of mesenchymal cells that may arise from the differentiation of MSCs (4), including human breast MSCs (bMSCs) isolated from a breast cancer patient, human colonic myofibroblasts (CCD-18co), and primary human mammary stromal fibroblasts.

IL-8 and IL-6 are known to be induced in certain cells by PGE₂ (15). To determine whether PGE₂ played a role in the induction of these cytokines in MSCs, PGE₂ production by MSCs was blocked with indomethacin, which inhibits the COX-1 and COX-2 enzymes. In LoVo/MSC cocultures, GRO-α, IL-6, and IL-8 protein induction was reduced by 85% to 98% by indomethacin; moreover, this induction could be partially rescued by providing PGE₂ to indomethacin-treated cocultures (Fig. 2C). Although PGE₂ treatment alone could not induce GRO-α, IL-6, or IL-8 expression in LoVo cells or MSC cells, additional PGE₂ potentiated the cytokine induction in LoVo/MSC cocultures and in IL-1–treated MSCs (Fig. 2C and Supplementary Fig. S2F).

MSCs expressed 2 distinct PGE₂ cell-surface receptors, EP2 and EP4 (Supplementary Fig. S2G). To support the notion that MSC-derived PGE₂ acted in an autocrine fashion, the EP2 and EP4 receptor antagonists AH6809 and GW627368X were added to MSC cultures, along with either LoVoCM or IL-1. LoVoCM- and IL-1–induced IL-6, IL-8, and GRO-α mRNA was suppressed by 60% to 80% by these EP receptor antagonists (Fig. 2D). Figure 2E summarizes our model that (i) COX-2, mPGES1, and PGE₂ are initially induced in MSCs by IL-1 released by carcinoma cells and that (ii) the resulting PGE₂, acting in an autocrine manner on MSCs, then cooperates with ongoing IL-1 paracrine signaling to trigger IL-6, IL-8, and GRO-α production by MSCs (Fig. 2E).

Before further analyzing the interactions between the carcinoma cells and MSCs, we confirmed that, as reported by others (5, 16), MSCs are indeed recruited to IL-1–secreting tumors in vivo (described in Supplementary Data, Section 2; Supplementary Fig. S3A and B). Having done so, we further investigated the influence of MSCs on carcinoma cell behavior, more specifically by analyzing the effects of MSC coculture on the expression by the tumor cells of markers of EMT, a cell–biological program that imparts motility, invasiveness, and self-renewal to carcinoma cells (14). After culturing, either alone or together with tandem dimer (td) Tomato-MSCs (5 days), LoVo and HCC1806 cells were isolated by fluorescence-activated cell sorting (FACS) and analyzed for E-cadherin, vimentin, fibronectin, and β-actin protein expression (Fig. 3A). E-cadherin, a key epithelial marker, was decreased by 98% to 100% in both LoVo and HCC1806 cells cocultured with MSCs. Conversely, vimentin and fibronectin proteins—both mesenchymal markers—were robustly induced in both carcinoma cells (Fig. 3A). Moreover, expression of the SNAIL protein, an EMT-inducing transcription factor (EMT-TF), was increased 5- to 69-fold in the MSC/carcinoma cocultures (Fig. 3A).

We then determined whether PGE₂ and/or the cytokines produced by MSCs in LoVo/MSC cocultures could elicit an EMT-like response in LoVo cells. PGE₂ was able, on its own, to cause a decrease in E-cadherin protein in LoVo cells (Fig. 3B, ∼70% decrease) but failed to elicit concomitant robust increases of mesenchymal markers, that is, vimentin and the ZEB1, SNAIL, and TWIST1 EMT-TFs (Fig. 3B).

In contrast to PGE₂, IL-6 alone induced ZEB1 (7-fold), SNAIL (3-fold), and vimentin (3-fold) protein expression in LoVo cells, but was unable to decrease E-cadherin protein. However, treatment of LoVo cells with PGE₂ together with the 4 cytokines induced both a decrease of E-cadherin protein expression (∼80%) and increases of vimentin (9-fold), ZEB1 (14-fold), SNAIL (13-fold), and TWIST1 (10-fold) protein expression (Fig. 3B). Hence, activation of a more complete EMT program in the carcinoma cells required the concomitant activation of multiple signaling pathways, specifically those triggered in these cells by PGE₂ acting together with the indicated cytokines.

Significantly more carcinoma cells invaded in LoVo/MSC cultures than in cultures of LoVo cells alone (Fig. 3C). To examine whether PGE₂ played a critical role in the MSC-induced carcinoma cell invasiveness, LoVo/MSC cocultures were treated with NS398, a COX-2 inhibitor; this treatment resulted in an 80% reduction of MSC-induced LoVo cell invasiveness (Fig. 3C). Greater than 60% of this inhibition could be reversed by adding PGE₂ to these cocultures.

Also, LoVo cell invasion was significantly reduced by antibodies that neutralize either IL-6, GRO-α, or RANTES (Fig. 3D). Accordingly, the MSC-induced carcinoma cell invasiveness seemed to derive from a confluence of PGE₂, IL-6, GRO-α, and RANTES signals impinging on the LoVo cells.

We also examined possible effects of MSCs on carcinoma cell invasion in vivo. The carcinoma cells injected on their own formed reasonably well-encapsulated tumors (Fig. 3E, a, b, c). In contrast, the carcinoma cells coinjected with MSCs formed extensive invasive fronts that extended into adjacent muscle layers (Fig. 3E, d, e, f). In addition, we observed the intravasation of carcinoma cells into nearby microvessels (Fig. 3E, insert). To summarize, these data indicated that MSCs facilitated LoVo cell invasion both in vitro and in vivo.

Due to the fact that transformed epithelial cells that have undergone EMT contain larger subpopulations of TICs (12, 13, 17), we determined whether EMT induction of carcinoma cells by MSCs was similarly accompanied by an increase in tumor-initiating ability. When populations of 5 × 10⁴ IL-1–secreting carcinoma cells were coinjected with 2 × 10⁵ MSCs, their ability to give rise to palpable tumors was measurably increased (1/6 to 6/6 for LoVo, 1/6 to 6/6 for HCC1806, 0/6 to 5/6 for SUM159 and 1/6 to 5/6 for SUM149, Fig. 3F). In contrast, tumor initiation by 5 × 10⁴ MDA-MB-231 or MDA-MB-453 cells, neither of which secrete IL-1, was not increased by coinjection with 2 × 10⁵ MSCs.

We quantified these interactions more precisely by implanting limiting dilutions of LoVo cells at 4 dosages, together with 5 × 10⁵ admixed MSCs. The presence of admixed MSCs increased the tumor-initiating frequencies of 5 × 10⁵, 5 × 10⁴, and 5 × 10³ implanted LoVo cells from 4/6, 1/6, and 0/6 to 6/6, 6/6, and 6/6. (Fig. 3G). As judged by the extreme limiting dilution analysis [ELDA (18)], the frequency of TICs in cultured LoVo cells was 9 × 10⁻⁷ to 6 × 10⁻⁵; in the presence of admixed MSCs, this frequency increased to 1 × 10⁻³ to 1 × 10⁻² (Fig. 3H).

CSCs, which have some characteristics associated with normal stem cells, are cells defined operationally by their tumor-initiating ability (19). We validated the use of aldehyde dehydrogenase (ALDH) as a marker of tumor-initiating cells and thus CSCs (refs. 20, 21; described in Supplementary Data, Section 3; Supplementary Figs. S4A–S4C and S5A–S5B). We observed that a 5-day coculture of LoVo or HCC1806 cells with MSCs resulted in approximately 7-fold and 20-fold increases, respectively, in ALDH1 protein levels in the carcinoma cells (Fig. 4A and Supplementary Fig. S5C). To extend this observation, we cultured an unfractionated LoVo cell population or already sorted ALDH^high or ALDH^low LoVo cell subpopulations (Supplementary Fig. S4Ba), either alone or in the presence of a 2-fold excess of tdTomato-labeled MSCs, for 5 days. All 3 LoVo cell populations, when cocultured with MSCs, developed larger subpopulations of ALDH^high cells (59%, 38%, and 18%) than when cultured alone (38%, 20%, and 7%), that is, 1.6- to 2.6-fold increases in ALDH^high cells (Fig. 4Ba).

We also used FACS analysis to remove the tdTomato-MSCs from LoVo/MSC cocultures, then further propagated the LoVo cells for 5 days in the absence of MSCs. We found that the levels of the ALDH^high LoVo cell subpopulations that had previously been increased by coculture with MSCs reverted to levels comparable with those observed in LoVo cells that had never experienced MSC coculture (Fig. 4Bb). Our data indicated that maintenance of elevated numbers of ALDH^high LoVo cells depended on continuous interactions of the LoVo cells with MSCs.

Use of the ALDH marker as the sole stem cell marker is likely to have led to an underestimate in the increase of the number of CSCs, as ALDH^high cells are a CSC-enriched population rather than being a pure CSC population. To refine the markers used to identify CSCs, we determined that a LoVo subpopulation more enriched for CSCs could be identified by concomitant use of the ALDH^high and CD133⁺ markers, which have been used to define CSCs in various cancer cell populations (refs. 22–25; described in Section 3, Supplementary Data; Supplementary Fig. S6A–C). Indeed, in the LoVo cells cocultured for 5 days with MSCs, the ALDH^high/CD133⁺ LoVo cells were increased from 1.4% to 16.4% of the overall cell population, that is, a 11.7-fold increase (Supplementary Fig. S6C).

We also determined whether the MSC-induced increase of ALDH^high LoVo cells observed in culture was accompanied by an increase of TICs in LoVo cells. Limiting dilutions of sorted ALDH^high and ALDH^low LoVo cells were injected subcutaneously into severe combined immunodeficient (SCID) mice, either alone or together with MSCs (Fig. 4Ca). As calculated using ELDA, the TIC frequency of ALDH^high LoVo cells injected on their own was 4 × 10⁻⁵ to 3 × 10⁻⁴, whereas that of ALDH^high LoVo coinjected with MSCs was 2 × 10⁻³ to 1 (Fig. 4Cb). For ALDH^low LoVo cells injected alone, the TIC frequency was 7 × 10⁻⁷ to 5 × 10⁻⁶; coinjection with MSCs increased this frequency to 3 × 10⁻⁴ to 2 × 10⁻³ (Fig. 4Cb). TIC frequencies of both ALDH^high and ALDH^low LoVo cells were increased by several orders of magnitude when coinjected with MSCs.

The continued presence of MSCs in the implanted cell populations complicates interpretation of the direct effects of MSC on the tumor initiation of carcinoma cells, as the MSCs might affect tumor initiation by a number of different mechanisms. To address this issue, LoVo cells or HCC1806 cells were isolated by FACS analysis from 5-day carcinoma cell/tdTomato-MSC cocultures and were then immediately injected subcutaneously into SCID mice, in parallel with LoVo cells previously cultured alone, to determine the effect on the tumor initiation of the cell culture interactions (Fig. 4D and Supplementary Fig. S7A). Accordingly, TIC frequencies of LoVo cells were increased by one order of magnitude following a 5-day coculture with MSCs in vitro (Supplementary Fig. S7B). An increase of tumor initiation by prior coculture with MSCs was also observed on HCC1806 cells (Supplementary Fig. S7C).

To understand more precisely how the signals exchanged between carcinoma cells and MSCs led to increases in ALDH^high CSCs, we first ascertained whether PGE₂ and/or cytokines produced by MSCs in LoVo/MSC cocultures could elicit ALDH1 expression. Only PGE₂ was able, on its own, to elicit an increase in ALDH1 expression in LoVo cells (Fig. 5A). Moreover, when PGE₂ was combined with the 4 cytokines, there was no further increase in ALDH1 expression.

EP4 is the only PGE₂ receptor highly expressed by LoVo cells (Supplementary Fig. S2G). To elucidate in more depth the effects of PGE₂ on LoVo cells, we treated these cells with vehicle or PGE₂ for 5 days. The vehicle-treated cultures contained 10.3% ALDH^high LoVo cells, whereas the LoVo cells treated with PGE₂ contained approximately 25% ALDH^high LoVo cells (Fig. 5B). We added GW627368X (the EP4 receptor antagonist) to LoVo cells to determine whether the basal, unperturbed levels of ALDH^high cells depended on ongoing PGE₂ autocrine signaling; this treatment reduced by approximately 60% the basal level of ALDH^high LoVo cells (Fig. 5B). These data suggested that within LoVo cells, subpopulations of cells are maintained in an ALDH^high state, in part, through ongoing, low-level autocrine PGE₂ signaling. Moreover, this signaling and associated entrance into the ALDH^high state could be enhanced by exogenously supplied PGE₂.

The observed increase in ALDH1 expression induced by PGE₂ was not confined to LoVo cells. PGE₂ treatment induced elevated ALDH1 expression (2.9- to 12.9-fold) in LoVo, SUM149, SUM159, and BT549 cells (Fig. 5C), all of which were previously found to elicit increased PGE₂ production from cocultured MSCs.

The LoVo cells that had been treated ex vivo with vehicle or PGE₂ for 5 days were implanted subcutaneously in SCID mice. Tumors derived from control LoVo cells occurred in 3 of 16 injected hosts, whereas the corresponding PGE₂-treated cells formed tumors in 18 of 24 hosts (Fig. 5D). Hence, a substantial increase in ALDH^high cells and TICs could be achieved by PGE₂ treatment of LoVo cells ex vivo. Moreover, treating LoVo cells with the cocktail of cytokines (IL-6, IL-8, GRO-α, and RANTES) in addition to PGE₂ did not significantly increase the tumor initiation beyond that observed for PGE₂ treatment alone (Supplementary Fig. S8). In addition to its effects on TICs, PGE₂ is likely to contribute to the maintenance of CSCs in vivo by increasing tumor angiogenesis (Supplementary Fig. S9, discussed in Supplementary Data, section 4).

To confirm that the above-described role of PGE₂ could explain the ability of MSCs to induce CSC formation, we inhibited PGE₂ synthesis with NS398, or PGE₂ signaling with GW627368X, in cocultures of LoVo cells and tdTomato-MSCs. Blocking PGE₂ signaling by either route prevented most of the increases in ALDH^high LoVo CSCs by MSCs (described in Section 5, Supplementary Data; Supplementary Fig. S10). LoVo cells that had been cocultured under the various conditions with MSCs for 5 days were sorted by FACS analysis to eliminate tdTomato-MSCs and were then injected subcutaneously into SCID mice. The increase in TIC frequencies resulting from a 5-day coculture with MSCs (from 4/20 to 14/19) was prevented by introducing either NS398 or GW627368X into the cocultures (from 14/19 to 3/17 or 1/12, respectively, Fig. 5E). Adding PGE₂ to the cocultures along with NS398 restored the tumor-initiating frequency from 3/17 to 7/8.

To validate the role of COX-2 in these properties, we knocked down COX-2 in MSCs. The ability of the resulting MSCshcox-2 cells to produce PGE₂ in response to LoVoCM was reduced by 90% (Fig. 5Fa). The tumor-initiating frequency of 5 × 10⁴ LoVo cells was increased from 1/12 to 10/12 by coinjection with MSCs expressing a control, scrambled shRNA (MSCshsc cells; Fig. 5Fb). In contrast, in mice coinjected with MSCshcox-2 cells, the tumor-initiating frequency of LoVo cells was increased from 1/12 to only 3/12, supporting the notion that COX-2-dependent PGE₂ induced in MSCs was required for the observed robust increases of LoVo TICs.

The β-catenin signaling pathway has been implicated in maintaining stem cell and CSC homeostasis in most epithelial tissues (26, 27). Relevant here is the finding that PGE₂ treatment leads to Akt activation; activated Akt subsequently stimulates β-catenin signaling in several ways (28, 29). Accordingly, we examined activation of Akt/glycogen synthase kinase-3 (GSK-3)/β-catenin signaling axis in LoVo cells treated with vehicle, PGE₂, GW627368X, or GW627368X + PGE₂. We found that PGE₂ treatment led to a 12-fold increase in Akt phosphorylation at Thr473 (Fig. 6A), which indicates functional Akt activation (30). Conversely, inhibiting PGE₂ signaling with GW627368X in PGE₂-treated LoVo cells blocked 50% of the PGE₂-induced Akt phosphorylation (Fig. 6A). Moreover, β-catenin activity is positively affected by Akt-mediated phosphorylation of its Ser552 residue (31, 32). PGE₂ treatment of LoVo cells, which caused a 12-fold increase of Akt phosphorylation, also increased β-catenin phosphorylation at Ser552 (11-fold), an increase that was blocked entirely by GW627368X treatment (Fig. 6A).

Akt also enhances β-catenin by phosphorylating and thereby inactivating GSK-3; this prevents inactivation of β-catenin by unphosphorylated GSK-3, as Ser21-unphosphorylated GSK-3α and Ser9-unphosphorylated GSK-3β phosphorylate β-catenin at its Ser33/Ser37/Thr41 residues, leading to its degradation (33, 34). PGE₂ strongly increased the phosphorylation of GSK-3α at Ser21 (56-fold above basal level, Fig. 6A). Although Ser9 of GSK-3 was already phosphorylated before PGE₂ treatment, this basal level of GSK-3β Ser9 phosphorylation was reduced (70%) by adding GW627368X (Fig. 6A), presumably by blocking basal autocrine PGE₂ signaling. Inhibition of PGE₂ signaling by GW627368X prevented the inactivating phosphorylation of GSK-3α and GSK-3β and, conversely, increased phosphorylation of β-catenin at Ser33/Ser37/Thr41 residues that lead to its degradation by 15-fold (Fig. 6A). These data indicated that inhibition of PGE₂ signaling leads to β-catenin phosphorylation, a prelude to its proteasomal degradation.

A strong inhibition of β-catenin signaling is also achieved by E-cadherin, which recruits β-catenin to adherens junctions associated with the plasma membrane, thereby preventing its nuclear localization and its actions in promoting transcription (35). PGE₂ decreased the E-cadherin and ZO-1 protein located at the cell junctions (Fig. 6B and C). Correspondingly, in PGE₂-treated LoVo cells, β-catenin was found in cell nuclei, rather than being sequestered by E-cadherin in adherens junctions (Fig. 6B and C). PGE₂ caused a 5.3-fold increase of β-catenin in the nuclear fraction; this increase was blocked by adding the EP4 antagonist during PGE₂ treatment (Fig. 6D). We also found, by analyzing expression in LoVo carcinoma cells of a number of β-catenin/TCF-regulated genes, that PGE₂-induced nuclear β-catenin was functionally active (Fig. 6E).

PGE₂ induces Akt phosphorylation in carcinoma cells in a phosphatidylinositol 3-kinase (PI3K)-dependent manner (28). To determine whether the activation of the Akt/GSK-3/β-catenin signaling axis was required for PGE₂-induced ALDH^high LoVo CSCs, we treated ALDH^high LoVo cells with vehicle, PGE₂, GW627368X, LY294002 (a PI3K inhibitor), FH353 (a β-catenin/TCF inhibitor), or cardamonin (a β-catenin inhibitor) for 5 days at concentrations that did not cause significant cell death.

LY294002 functioned as efficiently as the EP4 antagonist by completely blocking the exogenous PGE₂-induced increase of ALDH^high LoVo cells and by decreasing the endogenous PGE₂-maintained basal ALDH^high LoVo cells (Fig. 6F, a and d). Although the 2 β-catenin inhibitors FH353 and cardamonin blocked the PGE₂-induced increase of ALDH^high LoVo cells (Fig. 6Fb and Fc), these inhibitors did not function as effectively as the EP4 antagonist or the PI3K inhibitor. PGE₂/EP4 signaling, acting through the Akt/GSK-3/β-catenin signaling axis (28), contributes to induction of the ALDH^high LoVo cell phenotype.

PGE₂ is metabolically unstable and is thought to act within tissues over short distances, doing so in both an autocrine and a paracrine manner. Wishing to pursue this notion further, we examined whether ALDH^high CSCs were located near MSCs (or their mesenchymal derivatives) in tumors that arose following the coinjection of these 2 cell types. ALDH^high LoVo cells (Fig. 7Aa, red signal) were often surrounded by tdTomato-labeled MSCs (Fig. 7Aa, green signal) or their derivatives in tissue sections of LoVo/MSC xenografts; in addition, many of these mesenchymal cells expressed COX-2 (Fig. 7Aa, cyan signal). Moreover, the mesenchymal cells associated with the ALDH^high tumor cells expressed FSP, the fibroblast marker (Fig 7Ab, red signal; discussed in Supplementary Data, section 6; Supplementary Fig. S11), indicating that the MSC-derived fibroblasts rather than their MSC precursors were largely responsible for forming the stromal microenvironment of these ALDH^high cells. The conclusion that PGE₂ induces the formation of ALDH^high CSCs was further supported by human colorectal adenocarcinoma studies in which we found the juxtaposition within tumors of such CSCs with stroma expressing COX-2, the likely source of PGE₂ production (discussed in Supplementary Data, section 7; Supplementary Fig. S12A and B). These observations lead us to propose that MSCs (or their differentiated derivatives) create a niche within tumors, leading to induction and/or maintenance of CSC subpopulations.

To determine whether elevated IL-1 production correlates with COX-2/PGE₂ expression in human primary carcinomas, we compared the normalized IL-1α, IL-1β, and COX-2 mRNA levels across 19 human normal colon and 123 human colon carcinoma samples and across 178 human invasive breast carcinoma samples (Cancer Genome Atlas analyzed by Oncomine). IL-1α and IL-1β were expressed at significantly higher levels (9.6-fold and 4.7-fold in colon carcinoma and 2.1-fold and 2.1-fold in breast carcinoma) in aggressive subtypes of samples—colon mucinous carcinoma and triple-negative breast carcinoma (TNBC), respectively (Fig. 7B). Moreover, the COX-2 mRNA levels were correlated with the IL-1α and IL-1β mRNA levels; COX-2 mRNA was elevated 5.4-fold and 5.3-fold in colon mucinous carcinoma and TNBC when compared with normal colon samples and other breast cancers (Fig. 7B). These correlations suggested that the signaling mechanisms described above (Fig. 7C), involving IL-1–activated expression of COX-2/PGE₂, may be relevant to understanding the pathogenesis of colon mucinous carcinomas and TNBCs.

Elevated IL-1 expression has been correlated with increased malignant progression and more aggressive phenotypes in many types of cancer (36). However, the mechanism(s) underlying this correlation have been unclear. We describe here a bidirectional, reciprocal interaction between carcinoma cells and the MSCs (Fig. 7C). The signaling is initiated through the release of IL-1 by carcinoma cells. We present evidence that seems to explain this connection between IL-1 production and increased tumor aggressiveness in many, and perhaps all, IL-1–secreting carcinomas.

IL-1 secreted by carcinoma cells induces COX-2 and mPGES1 expression in MSCs. The 2 enzymes collaborate in MSCs to generate PGE₂ levels, which can increase by 80- to 500-fold (Fig. 1Cb). Of note, we show these responses operate equally well in MSCs and in their more differentiated descendants (Supplementary Fig. S2D and S2E). We note that others recently reported that IL-1 secreted by head and neck squamous cell carcinoma cells induce PGE₂ from fibroblasts (37). Consequently, fibroblasts and myofibroblasts, both of which accumulate in large numbers in the stroma, may also represent sources of the PGE₂ and the subsequently produced cytokines described here.

Importantly, although COX-2 is highly expressed in both neoplastic and stromal cells in tumors, not all COX-2-expressing cells can produce PGE₂. Thus, despite large variations in COX-2 expression, the colorectal carcinoma cell lines that we examined produced only about 10 to 40 pg/mL of PGE₂ (Supplementary Fig. S1A). These levels are dwarfed by the 15,000 to 40,000 pg/mL of PGE₂ produced by the IL-1–stimulated MSCs studied here (Fig. 1Cb). This failure by COX-2-expressing carcinoma cells to produce significant levels of PGE₂ may be due to the absence in many carcinoma cells of significant levels of mPGES1 expression (Supplementary Fig. S1A and S1D). Hence, COX-2 expression, on its own, is unlikely to provide an accurate indication of PGE₂ production by carcinomas.

MSC-produced PGE₂ acts in 2 ways within such tumors—in an autocrine fashion on the MSCs that produced it and in a paracrine fashion on the nearby IL-1–releasing carcinoma cells. The MSC autocrine signaling elicits a second wave of signaling responses: In collaboration with ongoing paracrine IL-1 signaling from carcinoma cells, the autocrine PGE₂ induces IL-6, IL-8, GRO-α, and RANTES cytokines in the MSCs. Together, these MSC-derived molecules induce a third wave of responses that profoundly alter the carcinoma cells that initiated this signaling cascade (Fig. 2E and 7C).

The changes induced in carcinoma cells by the confluence of PGE₂ and cytokine signals are all components of the complex cell–biological program termed EMT. Although this program has been increasingly implicated in the acquisition of phenotypes associated with high-grade malignancy (14), major questions concerning EMT have remained unanswered. Among them are the paracrine signals, ostensibly of stromal origin, that trigger EMT in carcinoma cells. Here we present a scenario that explains how EMT can be induced in carcinoma cells by a reactive stroma.

This work also addresses another longstanding puzzle about EMT: Is it usually activated as a single, coherent program or, alternatively, are distinct components of this program activated separately, each by a distinct set of heterotypic signals? Our observations indicate that the latter scenario is more likely. For example, PGE₂ caused a decrease of E-cadherin expression in carcinoma cells, whereas cytokines were required to induce the vimentin and ZEB1 expression that is usuallydepicted as intrinsic components of the EMT program (Fig. 3B). Such responses suggest the possibility that, during the course of spontaneous tumor progression, some carcinoma cells may receive only a subset of these signals and accordingly only activate portions of the EMT program, whereas the others receiving the complete suite of heterotypic signaling molecules may pass through an entire EMT program.

Research by ourselves and others has connected EMT with entrance into a stem cell–like state, both in normal and neoplastic epithelial cells (12, 13, 17). These findings are also echoed by this work, in which we observed a concomitant entrance into the mesenchymal and stem cell states in response to MSC-derived heterotypic signals. PGE₂, which activated portions of the EMT program, was able to increase both the number of CSCs and the frequency of tumor initiation (Fig. 5A–D). Our findings here further clarify the connection of EMT with entrance into a stem cell–like state by showing that the partial EMT induced by PGE₂, which represses cell–cell junctions without inducing mesenchymal traits, suffices to increase CSCs. The observation of MSC-induced increases in CSCs is consistent with the recent finding of prostaglandin-induced increases in the number of CD44⁺ tumor cells (38, 39). The unique contribution of PGE₂ is underscored by the observation that other MSC-derived cytokines, when combined with PGE₂, had only marginal effects on further increasing the TIC frequency (Supplementary Fig. S8).

In earlier work (27), we documented an alternative means of activating the EMT program that involves canonical and noncanonical WNTs, together with TGFβ. Those findings echoed the present observations, as both studies showed that multiple distinct heterotypic signals, acting in concert, are required to activate an EMT in carcinoma cells. These earlier findings left open the possibility that other EMT-inducing signals beyond those documented at the time may converge on the WNT and TGF-β signaling pathways to activate EMT programs. We note here that PGE₂-activated signals do, indeed, converge on one of these signaling cascades by activating β-catenin signaling, the same pathway that represents the main signaling channel lying downstream of canonical WNT signaling.

In tumors that arise from IL-1–producing carcinoma cells, we find that this interleukin plays a critical role in the tumor cell–induced COX-2/mPGES1/PGDH/PGE₂ response in MSCs that is required for tumor progression. IL-1 blockage has been used in thousands of patients to control infection and inflammatory disease and has a remarkable safety record (40). On the basis of our findings and the existing clinical use of IL-1 inhibitors, IL-1 inhibition may present a promising alternative to COX-2 inhibitors for cancer therapy. In addition, limited therapeutic options are currently available for TNBCs, which produce higher levels of IL-1 (Fig. 7B), because they are unresponsive to standard receptor–mediated treatments. Accordingly, our findings suggest a possible option for treating these aggressive subtypes of breast and colon cancer.

Human carcinoma cell lines HCC1806, BT549, MDA-MB-231, MDA-MB-453, SW1116, and LoVo were obtained from American Type Culture Collection and human bone marrow–derived MSCs (Sciencell) were obtained from Sciencell. SUM149 and SUM159 cells were provided by S.P. Ethier (Wayne State University, Detroit, MI). The human carcinoma cell lines SUM149, SUM159, BT549, MDA-MB-231, MDA-MB-453, SW1116, and LoVo were authenticated by microarray analysis. HCC1806 and MSC were not passaged more than 6 months after receipt.

All research involving animals complied with protocols approved by the MIT Committee on Animal Care. In experiments evaluating tumor initiation and growth, the tumors were isolated and weighed at the end of each experiment. To measure TIC frequency, serial dilutions of cancer cell suspensions were injected subcutaneously into nude mice. TIC frequencies of the samples were determined using the ELDA webtool (18).

The concentrations of PGE₂ and cytokines were determined by ELISA as described in the manufacturers' protocols. PGE₂ levels were measured using a PGE₂ direct Biotrack assay kit (GE Healthcare). Human cytokine levels were measured using Quantikine kits (R&D Systems).

Cell invasion was evaluated by using BD Matrigel Invasion Chambers, 8.0 μm (BD Biosciences). The cells that migrated through the membrane during the incubation period were counted in 5 randomly selected regions.

See Supplementary Materials and Methods for more information.

No potential conflicts of interest were disclosed.

Conception and design: H.-J. Li, H.R. Herschman, R.A. Weinberg

Development of methodology: H.-J. Li, H.R. Herschman

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): H.-J. Li, F. Reinhardt

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): H.-J. Li, H.R. Herschman

Writing, review, and/or revision of the manuscript: H.-J. Li, H.R. Herschman, R.A. Weinberg

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): H.-J. Li, H.R. Herschman

Study supervision: H.R. Herschman, R.A. Weinberg

The authors thank members of the Weinberg laboratory (Wai Leong Tam, Michael Hwang) and the Herschman laboratory (Tomo-o Ishikawa, Art Catapang) for discussion and technical support; the Whitehead Flow Cytometry Core for technical support; and Sarah Dry and the UCLA Translational Pathology Core Laboratory for providing colon adenocarcinoma samples.

The work was supported by Breast Cancer Research Foundation (R.A. Weinberg), Susan G. Komen for the Cure (H.-J. Li), NIH (R.A. Weinberg: U54CA163109 and H.R. Herschman: R01CA123055 and P50CA086306), and Ludwig Center for Molecular Oncology (R.A. Weinberg).