Epithelial-to-Mesenchymal Transition Activates PERK–eIF2α and Sensitizes Cells to Endoplasmic Reticulum Stress

Carcinoma cells acquire key malignant traits by reprogramming their differentiation state via an epithelial-to-mesenchymal transition (EMT; refs. 1, 2). This transdifferentiation program, which was first described in developmental contexts, is phenotypically characterized by repression of epithelial markers, upregulation of mesenchymal markers, and changes in morphology associated with cell migration. EMT can be induced experimentally by overexpression of transcription factors, such as Snail or Twist, and, in some contexts, by treatment with TGFβ. Cancer cells that undergo an EMT become invasive and drug-resistant; such cells also efficiently seed primary and metastatic tumors, making them functionally indistinguishable from tumor-initiating cells or cancer stem cells (TIC or CSC; refs. 3–5).

To invade, EMT cells must remodel the extracellular matrix (ECM) by secreting matrix proteases and large scaffolding proteins that facilitate their migration. These scaffolding proteins, which include collagens, fibronectin (FN1), plasminogen activator inhibitor 1 (PAI1), and periostin (POSTN; ref. 6), interact to form networks that provide tensional forces and signals that are essential for migration. These quaternary interactions are often initiated within the cell before secretion. For example, collagens are partially assembled into triple-helical fibers within the endoplasmic reticulum (ER) before their secretion into the extracellular space.

Cells have evolved several quality control pathways that maintain ER homeostasis, collectively termed the unfolded protein response (UPR; ref. 7). The UPR is activated by misfolded proteins within the ER, which accumulate upon nutrient deprivation, hypoxia, oxidative stress, or viral infection (8–13). UPR signaling is initiated by three receptors localized to the ER membrane—ER-to-nucleus signaling 1 (ERN1/IRE1α), protein kinase RNA-like ER kinase (PERK), and ATF6 (14, 15). These receptors converge on shared downstream factors that increase ER protein-folding capacity, including BiP/GRP78 and GRP94; they also have unique signaling effects, namely, activated IRE1α induces splicing of XBP1 mRNA, resulting in the translation of a frame-shifted stable form of the protein that functions as a transcription factor [XBP1(S)]; and activated PERK phosphorylates eIF2α, inducing an integrated stress response associated with global translational repression and selective translation of repair proteins (e.g., ATF4).

Because they play a major role in both tumor progression and drug resistance, there is significant interest in finding vulnerabilities of cancer cells that have undergone an EMT. In this study, we addressed this question by using selectively toxic small molecules to probe EMT biology. This led to the discovery of a key vulnerability and the finding that EMT cells require UPR signaling for their malignancy.

To identify probes of EMT cell biology, we previously performed a large-scale chemical screen for small molecules with selective toxicity toward EMT cells (5, 16–18). Of 315,000 compounds tested, this screen identified a few structurally related small molecules (Cmp302, Cmp308, and Dev4) with EMT-selective toxicity (Fig. 1A). These compounds exhibited between 20-fold and >100-fold selective toxicity toward nontumorigenic (HMLE) and tumorigenic (HMLER) human mammary epithelial cells induced through an EMT by inhibition of E-cadherin (shEcad) or overexpression of Twist (Supplementary Fig. S1A and S1B). Treatment of GFP-EMT and DsRed–non-EMT cell cocultures with Cmp302, Cmp308, or Dev4 selectively depleted GFP-EMT cells from the cocultures, further confirming the selective toxicity of these compounds. In contrast, two common chemotherapy drugs, paclitaxel and doxorubicin, caused enrichment of GFP-EMT cells within cocultures (Supplementary Fig. S1C and S1D); this was consistent with previous reports indicating that EMT cells resist chemotherapies (5, 19). The substitution of a single atom in the pyrrolidine group of Cmp302 was sufficient to completely abolish toxicity (compound Dev2; Supplementary Fig. S1E and S1F; ref. 18).

We assessed whether these compounds were also selectively toxic toward breast cancer cells induced into EMT without any genetic modifications. Relative to cells in adherent culture, MDA-MB-157 cells cultured in suspension undergo an EMT as gauged by epithelial marker repression, upregulation of mesenchymal markers, acquisition of a mesenchymal morphology, and expression of stem-like surface markers (Supplementary Fig. S1G–S1I). In comparison with cells grown in adherent culture, MDA-MB-157 cells induced through an EMT by suspension culture exhibited increased sensitivity to Cmp302 and Cmp308 and reduced sensitivity to paclitaxel (Supplementary Fig. S1J and S1K).

To identify the intracellular effects of these EMT-selective compounds, we used microarrays to profile global gene expression in EMT and non-EMT cells after treatment with Cmp302. This revealed that Cmp302 strongly induced expression of UPR genes in EMT cells, but not in non-EMT cells (CHOP, ATF3, and GADD34; Supplementary Table S1A). This suggested that Cmp302 was selectively inducing ER stress in EMT cells. Gene set enrichment analysis (GSEA) demonstrated that Cmp302 significantly upregulated—selectively in cells that had undergone an EMT but not in those that had not—genes known from other work (20) to be induced by two well-established ER stressors, thapsigargin and tunicamycin (Fig. 1B, top and Supplementary Table S1B). In contrast, Cmp302 did not upregulate the expression of genes induced by either hypoxia or doxorubicin treatment (Fig. 1B, bottom and Supplementary Table S1B).

To more directly assess this hypothesis, we determined whether compound treatment affected UPR signaling pathways known to be activated by ER stress. In fact, Cmp302 and its more potent analog, Dev4, activated all three branches of UPR signaling in a dosage-dependent manner—causing increased XBP1 splicing, eIF2α phosphorylation (Fig. 1C), and ATF6 activity (Fig. 1D); expression of downstream UPR factors CHOP and BiP was also induced (Fig. 1C and E). Cmp302 and Dev4 activated the UPR at lower doses in EMT cells relative to non-EMT cells; this paralleled their selective toxicity toward EMT cells. In contrast, the nontoxic structural analog, Dev2, did not activate UPR signaling in EMT or non-EMT cells (Fig. 1C–E and Supplementary Fig. S2C).

Collectively, these findings strongly suggested that Cmp302/Dev4 was causing cell death by selectively inducing ER stress in EMT cells.

The ability to selectively induce ER stress in EMT cells could be a unique feature of Cmp302/Dev4, or might result from a generalized sensitivity of EMT cells to ER stressors. To distinguish between these possibilities, we assessed whether EMT cells were also selectively sensitive to four established chemical inducers of ER stress: thapsigargin, tunicamycin, dithiothreitol (DTT), and A23187. Notably, all four compounds caused activation of the PERK and IRE1 branches of the UPR at 8-fold to 100-fold lower doses in EMT versus non-EMT cells, as gauged by phosphorylation of eIF2α and splicing of XBP1, respectively (Fig. 2A). All four ER stressors also activated the downstream UPR factors CHOP, BiP, and GADD34 at lower doses in EMT versus non-EMT cells (Fig. 2A–C and Supplementary Fig. S2A).

Moreover, EMT cells were markedly more sensitive to cell death caused by all four ER stressors, and this was observed for both tumorigenic (HMLER) and nontumorigenic (HMLE) lines (∼10-fold for tunicamycin, ∼25-fold for thapsigargin, ∼4-fold for DTT, and ∼8-fold for A23187; Fig. 2D; Supplementary Fig. S2B). Cells induced to undergo EMT by TGFβ treatment also showed increased sensitivity to tunicamycin and thapsigargin (Supplementary Fig. S2C). Thapsigargin also selectively eliminated EMT breast cancer cells from cocultures of GFP-labeled EMT (HMLER_Twist_GFP) and DsRed-labeled non-EMT cells (HMLER_shGFP_DsRed), doing so in a dosage-dependent manner (Fig. 2E). This was accompanied by increased cleavage of Caspase-3, indicating that EMT cells activated apoptosis in response to ER stress (Supplementary Fig. S2D).

To evaluate the generality of these findings, we assessed whether sensitivity to ER stressors correlated with the differentiation state of breast cancer lines. Breast cancers of the basal-B subtype are more stem-like and display increased activation of the EMT program relative to luminal subtype breast cancers (21–26). We therefore evaluated the sensitivity of a panel of 10 breast cancer lines comprising these two subtypes. Compared with the four luminal breast cancer lines, the six basal-B cell lines were significantly more sensitive to tunicamycin, thapsigargin, DTT, and A23187 (Fig. 2F and Supplementary Fig. S2E and S2F). Taken together, these data indicated that increased sensitivity to ER stress is a general characteristic of cells that have undergone an EMT.

To identify molecular factors underlying this increased ER stress sensitivity, we compared global transcriptional profiles of EMT and non-EMT breast epithelial cells (17). We analyzed 956 sets of functionally annotated genes for enrichment in cells that have undergone an EMT (27). ECM and secreted collagen gene sets were the most significantly enriched in EMT cells (P < 10⁻³), with many individual secreted genes being highly upregulated (Supplementary Fig. S3A and Fig. 3A).

Secretory cells often upregulate ER protein-folding and transport capacity to sustain their increased output (28, 29). Consistent with this, expression of 18 genes critically involved in secretory pathway components (SPCG; ref. 30) were upregulated in at least four of the five EMT lines relative to non-EMT controls (P < 1 × 10⁻¹⁰ with sign test; Supplementary Fig. S3B). Moreover, in highly secretory cells, increased ER capacity gene expression occurs together with increased vesicular transport from the ER to cis-Golgi. To assess vesicular flux, we transiently expressed GFP fused with Sec16, a core component of ER exit sites (ERES; ref. 31), and visualized ERES by confocal microscopy (32, 33). Quantification of Sec16-GFP foci revealed a significant 3-fold increase in ERES in EMT cells relative to controls, indicative of increased ER-to-cis-Golgi vesicular flux (Fig. 3B and Supplementary Fig. S3C).

To directly quantify secreted proteins, we used ³⁵S-methionine/cysteine to label secreted proteins, which were then harvested from the culture medium and visualized by gel electrophoresis and autoradiography. EMT cells (HMLE_shEcad, HMLE_Twist) exhibited an approximately 10-fold to 14-fold increase in total secreted proteins relative to isogenic control cells (Fig. 3C). As a control to confirm that the detected protein was secreted rather than being released from dying cells, we treated cells with the secretion inhibitor Brefeldin-A, which completely abrogated accumulation of labeled proteins in the culture medium (Supplementary Fig. S3D).

Along with the altered protein secretion capacity between EMT and control cells, significant differences in ER morphology were revealed using electron microscopy. In EMT cells, 75% of ER membranes had one or more branch points, with 30% having over 10 branch points; in contrast, only 10% of non-EMT cells had ER membranes with one or more branch points (Fig. 3D). Because professional secretory cells (e.g., pancreatic β cells) often display a highly developed ER network (34), this further suggested that as part of their function, EMT cells also have an increased demand for protein secretion.

To determine whether increased ECM secretion is a general feature of EMT, we examined the expression of ECM genes identified to be upregulated upon EMT (Fig. 3A) in basal-B and luminal breast cancer lines (35). Basal-B cancer lines (n = 9) expressed many EMT ECM genes—including FN1, COL1A1, COL1A2, COL4A1, COL5A1, POSTN, FBN1, and COL6A1—at significantly higher levels than luminal breast cancer lines (n = 13; Fig. 3E; ref. 24). In a subset of basal-B lines, EMT ECM genes were expressed at 10-fold to 100-fold higher levels than in luminal cancer lines (Supplementary Table S2; ref. 24). In contrast, ECM genes not upregulated upon EMT did not exhibit increased expression (COL4A3, COL10A1, COL13A1, and COL15A1; Fig. 3E). In support of these observations, ³⁵S-methionine/cysteine labeling showed that basal-B lines (Hs578T, BT549, MDA-MB-157, SUM159, MDA-MB-231, and 4T1) also exhibited, on average, a more than 40-fold increase in protein secretion relative to luminal lines (MCF7, T47D, BT474, and ZR-75-3; Fig. 3F).

We next considered the possibility that increased ECM secretion by EMT cells was directly responsible for their increased sensitivity to ER stressors. If this were indeed the case, then reducing ECM levels would attenuate UPR activation in response to ER stressors. To examine this, we analyzed the proteins secreted by two nontumorigenic EMT lines (HMLE_shEcad and HMLE_Twist) that, by ³⁵S-methionine/cysteine labeling, strongly upregulated secretion of a limited number of proteins (Fig. 3C). Mass spectrometry of conditioned medium from these two EMT-associated lines revealed that, relative to the corresponding controls, they secreted two major proteins, PAI1 and FN1.

We next inhibited PAI1 and FN1, both singly and in combination, with multiple shRNAs (Fig. 4A). Consistent with their abundance by mass spectrometry, dual inhibition of FN1 and PAI1 greatly reduced the total protein secreted by EMT cells into conditioned medium (Supplementary Fig. S4A). To examine whether the reduction in PAI1 and FN1 levels was biologically significant, we assessed the migratory properties of double-knockdown cells. Dual inhibition of PAI1 and FN1 also significantly reduced the migration of both the HMLE_shEcad and HMLE_Twist EMT cells (Fig. 4B and Supplementary Fig. S4B), consistent with prior reports (36). Dual inhibition of PAI1 and FN1 also strongly abrogated UPR induction in response to either Dev4 or thapsigargin treatment (Fig. 4C and D, HMLE_shEcad and HMLE_Twist cells, respectively; Supplementary Fig. S4C). These findings indicated that ECM secretion was required for EMT cells to migrate, while also increasing their sensitivity to ER perturbations.

Nascent polypeptides en route to secretion are folded by critical chaperone proteins that reside within the ER. Because EMT cells are more secretory and therefore have a higher ER load, we hypothesized that they might also be more sensitive to reductions in chaperone proteins. To test this, we used shRNAs to inhibit the key ER chaperone BiP (37) in a cell line model in which EMT could be induced within 3 days by addition of 4-hydroxytamoxifen (4-OHT; HMLE_ER_Twist; ref. 3). Using two different shRNAs, a 65% to 75% reduction in BiP had negligible effects on the viability of this line in the uninduced (non-EMT) condition (Fig. 5A and B). However, induction of EMT in these shBiP lines caused significant reduction of cell growth (8-fold less in mesenchymal vs. epithelial cells), and the surviving cells were clustered in epithelial islands (Fig. 5B). This indicated that the reduced BiP levels, although sufficient for the needs of epithelial cells, were not sufficient for cells to survive EMT. Inhibition of BiP also differentially affected the viability of basal-B (EMT-like) and luminal (non-EMT-like) breast cancer lines. Although BiP inhibition only modestly affected the viability of two luminal lines (MCF7, T47D), it caused significant death in two basal lines (MDA-231 and BT549) together with CHOP upregulation (Fig. 5C and D), suggesting that ER stress was more readily induced in BiP-deficient EMT cells. This was confirmed by examining UPR signaling, which revealed that the UPR was activated upon BiP inhibition in the basal-B cancer cells, but not the luminal cancer cells (Supplementary Fig. S5).

Before their differentiation, progenitors of secretory cells activate UPR pathways in anticipation of an increased ER load (38, 39); this UPR activation is not a response to ER stress but rather a means of preventing it. Because EMT cells are also highly secretory, we examined whether, in the absence of ER stressors, they also activate one or more UPR pathways.

Compared with non-EMT cells, EMT cells had reduced PERK protein mobility suggestive of its phosphorylation, increased eIF2α phosphorylation (Fig. 6A), and increased expression of the downstream gene GADD34 (Fig. 6B). In contrast, IRE1 signaling was not increased in EMT or non-EMT cells in the absence of exogenous ER stressors (Supplementary Fig. S6A). To confirm that PERK was in fact phosphorylated in EMT cells, we also performed phosphatase treatments and immunofluorescence with a phospho-specific PERK antibody. Treatment of lysates with lambda phosphatase before Western blotting abolished the reduced PERK mobility present in EMT cells under basal conditions; as a control, phosphatase treatment also abolished the reduced PERK mobility caused by thapsigargin in both EMT and non-EMT cells (Fig. 6C). Immunofluorescence with a phosphorylation-specific antibody also showed that PERK was constitutively activated upon EMT, but not in non-EMT cells (Fig. 6D). Consistent with these findings, cells induced through an EMT by TGFβ treatment also activated PERK but not IRE1 signaling (Supplementary Fig. S6B).

Because there are several kinases upstream of eIF2α, we next examined whether its phosphorylation in EMT cells was dependent on PERK. Suppression of PERK activity with a specific chemical inhibitor strongly decreased both PERK and eIF2α phosphorylation in EMT cells (Supplementary Fig. S6C). Similarly, PERK inhibition by shRNA also decreased eIF2α phosphorylation in two basal-B breast cancer lines (Fig. 6E). Collectively, these observations established that the PERK–eIF2α–ATF4 branch of the UPR is selectively and constitutively induced by cells that have undergone an EMT.

Depending on the context, UPR signaling can either promote survival or induce apoptosis in cells challenged with ER stress (7). Inhibition of PERK in EMT cells with a chemical inhibitor dramatically increased their sensitivity to thapsigargin (Fig. 6F), indicating that activation of the PERK pathway is adaptive and beneficial for the survival of cancer cells that have undergone an EMT. We next examined whether PERK signaling also contributed to the malignant properties of EMT cells. PERK inhibition strongly reduced the ability of EMT cells to form tumorspheres (Fig. 6G) and migrate in Transwell assays (Fig. 6H); at the same dose, the PERK inhibitor minimally affected cell proliferation (Supplementary Fig. S7A and S7B). Pretreatment of metastatic 4T1 cells with either the PERK inhibitor or thapsigargin resulted in significantly diminished metastatic capacity, as assessed by lung tumor burden 15 days after tail-vein injection (Fig. 6I). Collectively, these findings indicated that disruption of the PERK pathway significantly compromises the malignant phenotype of EMT cancer cells and further increases their sensitivity to ER stressors.

We next examined the clinical relevance of the above findings by assessing primary human tumors. Primary cancer cells (<3 passages) from breast tumors expressing EMT markers had elevated PERK and BiP expression, and increased eIF2α phosphorylation, when compared with primary breast cancer cells that did not express EMT markers (Fig. 7A and B). Primary cancer cells expressing EMT markers were also more sensitive to the ER stressor thapsigargin as indicated by UPR pathway activation (Fig. 7B). Consistent with this, these cells also exhibited significantly reduced viability upon treatment with ER stressors (Fig. 7C and Supplementary Fig. S8).

To assess whether these findings extended to other tumor types, we analyzed gene-expression microarray data from patient tumors to test for associations between the expression of EMT, ECM, and UPR pathway genes (see Methods for details). This analysis revealed that the expression of EMT and ECM genes is strongly correlated across patient tumors and could be observed in five datasets spanning 792 breast, colon, gastric, and lung tumors, as well as metastatic tumors (Fig. 7D). EMT and ATF4 genes were also strongly correlated in their expression (mean corr. = 0.80), whereas a significant correlation was not observed between the expression of EMT and XBP1 genes (mean corr. = −0.14; Fig. 7D). These findings established that EMT is strongly associated with PERK but not IRE1 signaling across a spectrum of tumor types.

Given the central role of EMT in tumor metastasis and therapy resistance, there is a vital need to identify pathways and processes that modulate either the survival or malignancy of cancer cells that have undergone EMT. In this study, we have assessed the effects of EMT-selective small-molecule probes in the context of global transcriptional profiling. This revealed that EMT cells, by virtue of their increased secretion of ECM, are highly sensitive to ER stress. This finding is noteworthy because EMT cells are resistant to a wide range of chemotherapies, and because the secretory output of a cell has not previously been shown to influence its sensitivity to chemicals that cause ER stress.

Our findings are consistent with prior studies linking EMT induction with ECM secretion. However, although the importance of ECM for tumor progression is well established (36), our study is the first to suggest that ECM secretion, while promoting malignancy, also creates a key cellular vulnerability. Thus, the acquisition of invasive and metastatic ability—by virtue of increased ECM production—might invariably lead to increased vulnerability to ER stress.

We have shown that EMT cells constitutively activate the PERK branch of the UPR, which is required for them to invade, metastasize, and form tumorspheres. The selective activation by EMT cells of PERK–eIF2α–ATF4 signaling, but not the IRE1 branch of the UPR, raises the possibility that this branch may be specifically required for ECM production. In support of this notion, mouse models have shown that PERK deficiency specifically compromised ECM production by osteoblasts (38); in contrast, XBP1 loss prevented the maturation of antibody-secreting plasma cells (40). Because PERK is activated in both cancerous and noncancerous cells following EMT, it may contribute to normal (non-neoplastic) functions of the EMT state. For example, during wound healing, epithelial cells must undergo EMT to secrete new ECM and migrate to close the wound, and interfering with EMT induction significantly impairs this process (41). If PERK signaling is required for ECM secretion during wound healing, its activation in EMT cancer cells may be a consequence of the normal functional properties of the EMT state.

Cancer cells that undergo an EMT are, in many cases, functionally indistinguishable from CSCs (3). This raises the possibility that CSCs may also exhibit increased sensitivity to ER stressors. In support of this, expression profiling of CSCs has revealed significant upregulation (relative to non-CSCs) of secreted ECM components also upregulated upon EMT, including COL1A1 and COL1A2 (42); we have also observed that CSC-enriched subpopulations from breast cancer cells display increased eIF2α phosphorylation (data not shown).

The finding that EMT cancer cells are vulnerable to ER stress has implications for the treatment of malignant tumors. Investigational agents that cause ER stress (43) may be most effective against tumors containing a high proportion of EMT cancer cells, such as breast tumors of the basal-B subtype. In such tumors, ER stressors could directly cause the death of EMT cancer cells, or interfere with their ability to secrete ECM and thereby mitigate tumor malignancy. In addition, ER stressors may effectively target disseminated cancer cells that have undergone EMT; they may also prove useful for eradicating EMT cells when they only constitute a small fraction of a tumor, provided that another therapy is used to eradicate the bulk population. Because PERK pathway inhibitors strongly abrogated the malignant traits of EMT cells, they also warrant further exploration as potential cancer therapies.

HMLE and HMLER cells expressing shRNAs targeting GFP (shGFP), E-cadherin (shEcad), or the coding sequence of Twist (Twist) were generated from Dr. Robert A. Weinberg's laboratory, and maintained in a 1:1 mixture of DMEM + 10% FBS, insulin (10 μg/mL), hydrocortisone (0.5 μg/mL), EGF (10 ng/mL), and Mammary Epithelial Cell Growth Medium (MEGM). The HMLE/HMLER_shEcad cells were validated by loss of E-cadherin expression, and the HMLE/HMLER_Twist cells were validated by overexpression of Twist. MCF7, T47D, BT474, ZR-75-3, Hs578T, MDA-MB-157, and MDA-MB-231 cells were obtained from ATCC and were cultured in DMEM + 10% FBS. BT549 and 4T1 cells (ATCC) were cultured in RPMI + 10% FBS. SUM159 cells were obtained from Asterand, and were cultured in F12 + 5% FBS, insulin (10 μg/mL), and hydrocortisone (0.5 μg/mL). The cell lines from ATCC have not been independently validated in our laboratory. PERK inhibitor was purchased from Toronto Research Chemicals Inc. (Cat G797800). Cmp302 (acyl hydrazone 1), Cmp308 (acyl hydrazone 2), and Dev4 (ML239) have been previously reported (18) and were identified in a Molecular Libraries Probe Production Centers Network screen conducted at the Broad Institute (Cambridge, MA).

Mammosphere formation assays were performed as described previously (5), but with 0.6% methylcellulose (R&D Systems) added to the medium. Five thousand cells were plated per well in low-adherence, 24-well plates and cultured for 5 to 8 days before being counted and photographed.

For analysis of Cmp302-treated expression data, Tm, Tg, and Dox gene sets were defined respectively as the top 100 genes induced by tunicamycin (GSE24500), thapsigargin (GSE24500), or doxorubicin (GSE39042). The hypoxia gene set consisted of the top 80 genes induced by both low oxygen tension (1%) and dimethyloxalylglycine (DMOG) treatment (GSE3188).

HMLE_shGFP and HMLE_Twist cells were treated with 5 or 10 μmol/L of Cmp302, or DMSO solvent for 6 hours. Total RNA were extracted using the Qiagen RNeasy Kit, and the integrity and quality of the RNA met the quality requirements for Human Genome U133 Plus 2.0 arrays (Affymetrix, Inc.) recommended by the company. All of the subsequent experimental procedures, including labeling, hybridization, and scanning, were processed according to the standard Affymetrix protocols. Raw CEL files were generated by Affymetrix GCOS 1.2 software, and the present/absent calls were defined with global scaling to target value of 500. By R software, the CEL files were normalized to a median-intensity array, and model-based expression values were calculated using PM/MM difference model. Alteration of gene expression by Cmp302 was calculated by comparing the expression of each gene in the DMSO- and Cmp302-treated groups, in both HMLE_shGFP and HMLE_Twist cells. The gene-expression data have been deposited in the public database Gene Expression Omnibus (GEO; GSE55604).

Cells were fixed in 2.5% gluteraldehyde, 3% paraformaldehyde with 5% sucrose in 0.1 mol/L sodium cacodylate buffer (pH 7.4), pelletted, and postfixed in 1% OsO₄ in veronal-acetate buffer. The cell pellet was dehydrated and embedded in Embed-812 resin. Sections were cut on a Reichert Ultracut E microtome with a Diatome diamond knife at a thickness setting of 50 nm, stained with uranyl acetate, and lead citrate. Sections were examined using a FEI Tecnai spirit at 80 kV and photographed with an AMT CCD camera.

HMLE and HMLE_shEcad cells (2 × 10⁵/well of a 6-well plate) were transfected with 1-μg pmGFP-Sec16S (Addgene 15775) using 2.5 μL of FuGENE (Roche). Twenty-four hours after transfection, cells were replated and allowed to adhere onto cover slides before fixation with 4% paraformaldehyde, and confocal imaging. Images were captured in accordance with the manufacturer's protocols (PerkinElmer).

NOD/SCID mice were purchased from The Jackson Laboratory. All mouse procedures were approved by the Animal Care and Use Committees of the Massachusetts Institute of Technology (Cambridge, MA). For lung metastasis analysis, 2 × 10⁶ cancer cells were suspended in 100 μL PBS and injected into the tail vein of each mouse. Lung tissues from experimental animals were harvested at the various time points indicated in the text. All animals were randomized by weight.

Cells were plated in 100 μL of medium per well in 96-well plates, at a density of 3,000 cells per well. Twenty-four hours after seeding, compounds were added at eight different doses with three replicates per dose per cell line. The same volume of DMSO was added in three replicates per line as a control. Cell viability was measured after 72 hours with the CellTiter-Glo AQueous Non-radioactive Assay (Promega). Paclitaxel, doxorubicin, tunicamycin, thapsigargin, A23187, and DTT were purchased from Sigma-Aldrich. Cmp302 and Cmp308 were purchased from Interbioscreen.

A total of 7.5 × 10⁵ cells were seeded on 3.5-cm plates 18 hours before wounding. Cells were washed two times with PBS, re-fed with culture medium, and allowed to migrate for 7 hours before visualization.

Western blotting was performed as previously described (5). Antibodies used for immunoblotting were as follows: E-cadherin (BD Transduction; 610182), fibronectin (Abcam; Ab6328), β-actin (Cell Signaling Technology; 12620), β-tubulin (Cell Signaling Technology; 5346), CK8/18 (Cell Signaling Technology; 4546), PERK (Cell Signaling Technology; 9956), BiP (Cell Signaling Technology; 9956), eIF2α (Cell Signaling Technology; 9722), p-eIF2α (Cell Signaling Technology; 3597), caspase-3 (Cell Signaling Technology; 9665), CHOP (Cell Signaling Technology; 5554), and p-PERK (Santa Cruz Biotechnology; Sc-32577).

Flow cytometry analysis was performed according to the manufacturer's protocol (BD Biosciences) with at least 10,000 live events captured per analysis. Reagents used were as follows: allophycocyanin (APC)-conjugated anti-CD44 antibody (clone G44-26), phycoerythrin (PE)-conjugated anti-CD24 antibody (clone ML5), and propidium iodide (5 μg/mL; BD Biosciences).

p5xATF6-GL3 and hRluc constructs were obtained from Addgene (Plasmid 11976 and 24348). Twenty-four hours after cotransfection of 0.3-μg p5xATF6-GL3 and 0.05-μg hRluc plasmids, cells were treated with indicated compounds for an additional 6 hours, after which ATF6 activity was measured by a dual luciferase assay (Promega).

Equal numbers of cells were cultured in the presence of ³⁵S-methionine/cysteine in medium with reduced methionine/cysteine content and 0.5% serum. At the indicated time points, aliquots of medium were extracted for analysis. Medium was centrifuged at 800 × g for 2 minutes to pellet any whole-cell contaminants. An equal volume of medium was reduced in loading buffer, separated by SDS-PAGE, and analyzed by autoradiography.

Ten million HMLE_shGFP and HMLE_shEcad cells were seeded in serum-free medium, and culture medium was collected at 48 hours. Protein from the culture medium was precipitated using 10% trichloroacetic acid (TCA) on ice. After centrifuging at 15,000 × g for 15 minutes, the pellet was washed in acetone and dissolved in reducing loading buffer. After separation by SDS-PAGE, the gel was silver stained, and bands were cut out for analysis by LC/MS-MS (44).

Gene-expression sets for correlation analyses were defined as follows: the EMT core gene set consisted of the top 100 genes upregulated upon EMT induction; the ECM gene set was obtained from MolSigDB; the XBP1 (GSE40515; ref. 45) and ATF4 (GSE35681; ref. 46) gene sets consisted of the most downregulated genes in XBP1 and ATF4 knockout cell lines, relative to the corresponding controls. For every gene set, a composite expression score was calculated for each sample by summing the log-normalized expression of the genes in the set. Genes in the EMT gene set that were also present in the ECM, XBP1, or ATF4 gene sets were excluded to eliminate any overlaps before calculating correlations. Tumors from five human cancer datasets were analyzed (GSE41998, GSE37892, GSE26942, GSE4573, and GSE11360). Spearman ρ was used as the measure of correlation, and for each comparison a P value was empirically determined by Monte Carlo sampling to generate a null distribution of 1,000 correlations from random gene sets of the same size as those being compared.

All data unless otherwise specified are presented as mean +/- SEM from at least three independent experiments. A Student t test was performed for comparisons between two groups of data. Two-way ANOVA tests were performed when comparing the responses of different groups of cells to various drug treatment doses.

No potential conflicts of interest were disclosed.

Conception and design: Y.-x. Feng, E.S. Sokol, Q. Wang, P.B. Gupta

Development of methodology: Y.-x. Feng, E.S. Sokol, H.L. Ploegh, P.B. Gupta

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Y.-x. Feng, E.S. Sokol, C.A. Del Vecchio, S. Sanduja, J.H.L. Claessen, T.A. Proia, D.X. Jin, H.L. Ploegh, Q. Wang

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Y.-x. Feng, E.S. Sokol, C.A. Del Vecchio, S. Sanduja, D.X. Jin, H.L. Ploegh, P.B. Gupta

Writing, review, and/or revision of the manuscript: Y.-x. Feng, E.S. Sokol, S. Sanduja, D.X. Jin, H.L. Ploegh, P.B. Gupta

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Y.-x. Feng, F. Reinhardt, Q. Wang, P.B. Gupta

Study supervision: P.B. Gupta

The authors thank Dr. George Bell and Dr. Inmaculada Barrasa for assistance with dose–response data analysis, Eric Spooner for mass-spectrometry analysis, Nicki Watson for electron microscope analysis, Dr. Jan Reiling for helpful discussions, and Tom DiCesare for assistance with graphical design. The p-mGFP-Sec16S plasmids were kindly provided by Dr. Benjamin Glick (University of Chicago).

This research was supported by grants from the Richard and Susan Smith Family Foundation and the Breast Cancer Alliance (to P.B. Gupta), and the National Science Foundation Graduate Research Fellowship (Grant No. 1122374; to E.S. Sokol).