TGFBR3 supports anoikis through suppressing ATF4 signaling

The balance between cell survival and death is critical for mammary epithelial morphogenesis. It depends on stimulation mediated by both cytokine receptors and integrins that sense the extracellular matrix (ECM). Loss of ECM ligands triggers a specific type of cell death, called anoikis, through the intrinsic pathway and a disruption in redox homeostasis (Schafer et al., 2009; Wang, 2021). Failure to eliminate disseminated cells appears to be the cause of malignant phenotypes.

Cancer cells develop a variety of strategies to resist anoikis, among which the integrated stress response (ISR) plays a critical role. ISR activation following ECM deprivation increases prosurvival signals by inducing activating transcription factor 4 (ATF4) pathway and cytoprotective ATF4 target genes (Dey et al., 2015). ATF4 signaling reduces reactive oxygen species and prevents anoikis. Another mechanism to protect cells from anoikis is growth factor receptor-mediated signaling. For example, loss of integrin engagement causes degradation of epidermal growth factor receptor (EGFR), which is an essential process for anoikis (Grassian et al., 2011; Haenssen et al., 2010; Schafer et al., 2009; Reginato et al., 2003; Muthuswamy et al., 2001). Overexpression of Erb-b2 receptor tyrosine kinase 2 (ERBB2) stabilizes EGFR and promotes cell survival in detached culture, underscoring the importance of cytokine receptors for regulating anoikis.

We previously reported that in the ECM-attached basal breast cell line (MCF10A-5E), the type III TGF-β receptor, TGFBR3, could be activated heterogeneously by various ligands, in turn leading to downregulation of the PAM50 basal marker and several stress-tolerance genes including JUND (Tai et al., 2022; Bajikar et al., 2017; Wang et al., 2014; Janes et al., 2010). This regulation is conserved within developing acini and also basal-like lesions during human premalignancy. TGFBR3 is a transmembrane protein with a large extracellular domain (766 amino acids) that contributes to the protein interacting with a variety of microenvironmental proteins (Diestel et al., 2013; Lin et al., 2011). It localizes exclusively on the basolateral side of polarized epithelia (Meyer et al., 2014), where cells form cell adhesions and interact with the basement membrane (Pampaloni et al., 2007). Ablation of TGFBR3 expression leads to disruption of normal morphogenesis (Wang et al., 2014). In addition, breast invasive ductal carcinoma and multiple types of human epithelial carcinomas have been found to be associated with TGFBR3 loss, supporting the tumor-suppressive role of TGFBR3 (Finger et al., 2008; Dong et al., 2007; Turley et al., 2007; Nishida et al., 2018).

Here, we report a tumor suppressor mechanism of TGFBR3, which supports apoptosis pathways and suppresses ATF4 signaling. We found that TGFBR3 accumulated in ECM-deprived cells, and that the accumulation is crucial for anoikis. Mechanistically, TGFBR3 functioned as a negative regulator of ISR by promoting destabilization of ATF4. These data show the important role of TGFBR3 in development, and connect cytokine receptor signaling to the cellular stress response.

To evaluate whether TGFBR3 was relevant to early breast cancer progression, we mined genomic and gene expression data from the 1904 primary breast tumors of The Cancer Genome Atlas (TCGA) and Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) datasets (Pereira et al., 2016; Gao et al., 2013; The Cancer Genome Atlas Network et al., 2012; Cerami et al., 2012). Although haploid insufficiency might correlate with TGFBR3 expression (Dong et al., 2007), across all breast cancer cases, TGFBR3 loci were unaltered – no patterns of mutations were observed (Fig. S1A). This could suggest that any potential role of TGFBR3-mediated tumor suppression is not associated with gene variations or functional mutations.

However, we reached a different conclusion when analyzing gene expression data, where invasive lobular carcinoma (ILC) displayed a significantly higher TGFBR3 expression, as compared with invasive ductal carcinoma (IDC) (IDC, n=1500; ILC, n=142) (Fig. 1A; Table S1). ILC reportedly exhibited very few Ki67-positive cells whereas IDC was highly proliferative (Huang et al., 2014). The main morphologic difference between the two histological subtypes is the single-file cell pattern in ILC (McCart Reed et al., 2015), primarily driven by poor ECM adhesion (Tasdemir et al., 2018) and lack of adhesion molecule E-cadherin (CDH1) (Fig. 1A; Table S2) (Desmedt et al., 2016; Ciriello et al., 2015). In human breast cancers, low TGFBR3 expression correlated significantly with high tumor grade and poor prognosis (Fig. 1B; Tables S3 and S4). These findings encouraged us to study the transcriptional changes in TGFBR3 associated with morphological phenotypes and growth properties of cells.

In previous experiments (Wang et al., 2014), TGFBR3 expression was heterogeneously induced during acinar morphogenesis and that resulted in normal acinar formation, whereas knockdown of TGFBR3 using short hairpin RNA resulted in large, budding structures in MCF10A-5E 3D culture, reflecting alternations of cell proliferation (Bajikar et al., 2017; Wang et al., 2014). In our original profiling dataset (Wang et al., 2014), ECM-attached cells on the outside of developing spheroids and inner, ECM-deprived cells were isolated by laser-capture microdissection (Fig. 1C), when the outer growth-arrested state and inner anoikis program had not yet fully engaged. Microarray transcriptional profiling of each subpopulation among 8259 genes revealed that TGFBR3 mRNA levels were proportionally higher in inner cells (Fig. 1D; Table S5). Surveying the entire transcript cluster that correlated with TGFBR3 within matrix-attached cells, we found that disruption of TGFBR3 cluster occurred due to ECM deprivation (Fig. 1E; Table S6). ECM-deprived cells showed decreased abundance of the TGF-β-family ligand GDF11, and unaltered expression of the TGF-β-family signaling ECM marker TGFBI. This indicated that intra-acinar TGFBR3 upregulation did not result from its regulatory circuit, which depended on ECM engagement (Wang et al., 2014). Previous studies have revealed that forced suspension of cells leads to alternation of cell-surface receptor levels (Grassian et al., 2011; Haenssen et al., 2010; Schafer et al., 2009; Reginato et al., 2003; Muthuswamy et al., 2001; Douma et al., 2004). Of the 52 detectable transmembrane receptors in the microarray data, we found that TGFBR3 was most highly upregulated in ECM-detached cells (Fig. S1B; Table S7), strongly suggesting that TGFBR3 levels are regulated transcriptionally after ECM deprivation.

To simulate ECM deprivation, MCF10A-5E cells were placed in suspension. Together with TGFBR3, we monitored time-dependent changes in the levels of seven transcripts before the anoikis pathways became dominant: six reference genes were constitutively expressed (B2M, GAPDH, GUSB, HINT1, PPIA and PRDX6) (Kang et al., 2013) and one gene was not altered under ECM remodeling in 3D culture (TGFBI) (Fig. 1E,F). We observed that TGFBR3 was significantly upregulated, whereas other genes displayed weak, delayed kinetics (B2M and GAPDH), were modestly downregulated (HINT1), or were totally unaltered (GUSB, PPIA, PRDX6 and TGFBI) (Fig. 1F). This finding is consistent with the TGFBR3 increase observed in inner cells during 3D morphogenesis (Fig. 1C–E; Fig. S1B, Tables S5–S7), and proves that TGFBR3 is triggered by ECM deprivation.

To determine whether expression of TGFBR3 protein is induced by ECM deprivation, we examined the TGFBR3 abundance by immunoblotting. We found that the TGFBR3 core protein was present in cell membrane after loss of attachment and its abundance was continuously increased during suspension culture (Fig. 1G,H). TGFBR3 protein induction in cell aggregates was further verified with confocal immunofluorescence (Fig. 1H). We observed a significant increase in the percentage of TGFBR3-positive cells in suspension culture. TGFBR3 staining was localized on the surface of the cells, as expected for mature transmembrane receptors. Thus, we conclude that TGFBR3 upregulation in epithelial cells is triggered by ECM deprivation.

The best characterized TGF-β1 signaling pathways are the SMAD2 and SMAD3 (SMAD2/3) and SMAD1 and SMAD5 (SMAD1/5) axes (Shi and Massagué, 2003). TGF-β1 binds to the TGFBR2 (TGF-β type II receptor)–TGFBR1 (TGF-β type I receptor) complex, initiating TGFBR1-mediated phosphorylation of SMAD2 and SMAD3. TGFBR1 also activates the BMP type I receptor ACVR1, which phosphorylates SMAD1 and SMAD5 (Ramachandran et al., 2018). To investigate whether TGFBR3 affects canonical SMAD-dependent pathways, we perturbed TGFBR3 expression in MCF10A-5E cells by RNA interference (Fig. 2A,B). We compared the kinetics of SMAD2/3 and SMAD1/5 phosphorylation in response to TGF-β1 (Fig. 2C,D). TGF-β1 induced rapid SMAD2/3 phosphorylation and maintained sustained SMAD2/3 phosphorylation that was still detectable 3 h after stimulation. In contrast, TGF-β1 induced only transient SMAD1/5 phosphorylation that peaked at 1 h before rapidly returning to the baseline. SMAD2 phosphorylation was not affected by the loss of TGFBR3. Interestingly, the magnitudes of the TGF-β1-induced SMAD3 phosphorylation and SMAD1/5 phosphorylation were enhanced and reduced significantly by TGFBR3 depletion, respectively.

Chromatin immunoprecipitation followed by sequencing (ChIP-seq) for SMAD3 and SMAD1/5 has revealed that in human transformed breast epithelial cells the majority of SMAD1/5-binding sites are also bound by SMAD3 (Ramachandran et al., 2018), making it difficult to determine the distinct roles of the individual pathways. However, we found that knockdown of TGFBR3 significantly enhanced the expression of many genes which were bound and regulated by SMAD3, but not by SMAD1/5, including SERPINE1, FOXQ1, FOXF2, FOXC2, FN1, EDN1, TGFBI and ADAM19 in TGF-β1-treated cells (Fig. 2E; Table S8) (Ramachandran et al., 2018). These data are consistent with previous studies showing that TGFBR3 blocked the SMAD2/3 axis and favored the SMAD1 axis in mouse lung fibroblasts (Schwartze et al., 2014). This suggests that TGFBR3 is required differently for SMAD3 versus SMAD1/5 phosphorylation, which serve as the essential mediators of TGF-β signaling.

To investigate whether TGFBR3 could contribute to cell death, we examined the abundance of cleaved caspase 3 and cleaved poly(ADP-ribose) polymerase (PARP), a caspase 3 substrate that promotes the cellular response to stress signals (Boulares et al., 1999), in suspended cells. Cleaved caspase 3 was present in cells brought into suspension (Fig. 2F,G). This demonstrated that MCF10A-5E cells underwent caspase-dependent apoptosis upon attachment withdrawal. In contrast, production of cleaved caspase 3 was reduced when TGFBR3 signaling was blocked. The enhanced cleaved caspase 3 and caspase-3-mediated PARP cleavage was also not observed in the shTGFBR3-expressing cell population (Fig. S2A). In addition, PARP cleavage was observed mainly in TGFBR3-positive cells (Fig. 2H; Fig. S2B, Table S9). Assaying the metabolic activity as a viability indicator, we confirmed that shTGFBR3-expressing cells were less likely to undergo anoikis (Fig. 2I). These results indicate that breast epithelial cells express TGFBR3 as an apoptosis-supporting component in response to loss of cell adhesion.

To investigate how ligands influenced TGFBR3 signaling, we treated cells with TGF-β1 or GDF11. When comparing responses of detached cells to these ligands, we found that TGF-β1 alone appeared to be enough to cause cell death, whereas GDF11-mediated cell death required TGFBR3 (Fig. 2F,G). MFC10A-5E cells are known to form small multicellular aggregates to suppress anoikis in suspension (Wang et al., 2014). Ablation of TGFBR3 promoted aggregate assembly (Fig. 2J), suggesting that TGFBR3 is required to interfere with cell adhesion. Similarly, TGF-β1, but not GDF11, blocked the formation of large spheroid aggregates upon shTGFBR3 treatment. Together, these data suggest that TGFBR3 is required for the full activity of TGF-β family ligands and their effects on cell aggregation and anoikis.

As the BCL2-family of proteins regulates caspase-dependent cell death (Singh et al., 2019; Karbowski et al., 2006; Adams and Cory, 1998; Ouyang et al., 2010), we therefore examined the impact of TGFBR3 on the BCL2 family of prosurvival and proapoptotic regulators. Following prolonged detachment, BAX and BAK1 levels were not altered and the expressions of both transcripts were unaffected by TGFBR3 perturbations (Fig. S3A,B). Thus, TGFBR3 does not influence the production of pore-forming proteins. We next analyzed the expression of prosurvival BCL2 members in suspension culture. Endogenous BCL2 was eliminated by disengaging ECM components (Fig. S3C). BCL2A1 and MCL1 were weakly expressed upon detachment, and both were unaffected by TGFBR3 perturbations (Fig. S3D,E). TGFBR3 suppressed BCL2L1 expression; BCL2L1 expression was triggered by TGFBR3 knockdown and reduced upon TGFBR3 overexpression (Fig. 3A; Fig. S3F). However, in suspended cells, the shTGFBR3-dependent BCL2L1 induction was lost (Fig. S3F). These results suggest that BAX, BAK1 and some prosurvival BCL2 members are not downstream of TGFBR3 in suspension culture.

As BH3-only proteins interact with prosurvival BCL2 regulators and neutralized their effect (Mailleux et al., 2007; Reginato et al., 2003), we next investigated whether TGFBR3 affected the expression of BH3-only proteins. We observed time-dependent changes in the levels of four BH3-only factors during suspension: one sensitizer gene (BAD) and three activator genes (BBC3, BCL2L11 and BID). BID was unaltered, whereas the other genes were upregulated continuously (BCL2L11) or dramatically at later time points (BBC3 and BAD) (Fig. 3B–E). Expression of BCL2L11 and BID were unaffected by TGFBR3 perturbations. In contrast, detachment-induced BBC3 and BAD expression was completely eliminated by TGFBR3 knockdown. Together with the fact that TGFBR3 enhanced the expression of BBC3 and BAD (Fig. 3F), we conclude that TGFBR3 is sufficient to induce BBC3 and BAD in suspension cultures.

During basal acinar morphogenesis, inner cells are metabolically stressed (Wang et al., 2011; Darini et al., 2019) and displayed high levels of BTG1 and SOD2 expression (Figs 1C,D, and 4A). We examined BTG1 and SOD2 expression in cells that were forced to detach. As expected, levels of both transcripts increased 4-to-5-fold between 24 and 48 h of suspension (Fig. 4B). TGFBR3 ablation did not affect BTG1 expression; however, it reduced the levels of SOD2 (Fig. 4C; Fig. S4A). Similar effects were observed on the expression of the antioxidant response gene SESN1 (Fig. S4B). We found that knockdown of SOD2 resulted in reduced cell survival in suspension culture (Fig. 4D–F). This observation is consistent with reports showing that SOD2 suppresses anoikis in breast cancer cell lines and human mammary epithelial cells (He et al., 2019; Kamarajugadda et al., 2013). The effects of TGFBR3 were distinct from the SOD2 antioxidant response, indicating that cells with low TGFBR3 might exploit different strategies to overcome anoikis.

Cancer cells encounter stress during cancer development (Wang, 2021). Thus, we asked whether TGFBR3 loss is one of the reasons for increased cancer cell survival in stress. For this purpose, we firstly examined the expression pattern of TGFBR3 in breast cancer subtypes. Breast cancer can be classified for clinical purposes based on the expression of estrogen receptor (ER), progesterone receptor (PR), and HER2 (ERBB2) (Sørlie et al., 2001). TGFBR3 abundance in hormone receptor-negative specimens was typically low (Fig. S4C–E). We used gene expression data from TCGA and METABRIC datasets to perform integrative clustering (IntClust) classification (Pereira et al., 2016; Gao et al., 2013; The Cancer Genome Atlas Network et al., 2012; Cerami et al., 2012; Curtis et al., 2012). TGFBR3 loss was significantly associated with ER-positive subtypes that had poor prognosis, and hormone receptor-negative subgroups, including IntClust10 (basal-like enriched) and IntClust4ER­­– [triple negative breast cancer (TNBC)] (Fig. S4F). We classified these breast cancer patient tumors by integrating PAM50 and claudin-low subtyping: 679 samples (36%) were luminal A, 461 (24%) were luminal B, 220 (12%) were HER2-enriched, 140 (7%) were normal-like, 199 (10%) were basal-like, and 199 (10%) were claudin-low. Expression of TGFBR3 was high in normal-like, luminal A and claudin-low subtypes, and low in basal-like, luminal B and HER2-enriched subtypes (Fig. 4G; Table S10). Basal-like and claudin-low tumors form the majority of aggressive TNBCs (Herschkowitz et al., 2007). The significant difference in TGFBR3 expression between basal-like and claudin-low molecular subtypes demonstrated that there is a unique subtype-specific TGFBR3 expression pattern in TNBCs.

The aggressiveness of TNBC cells is driven by the integrated stress response (ISR) (Gonzalez-Gonzalez et al., 2018; Wang, 2021). ISR activation suppresses global protein synthesis through the phosphorylation of eukaryotic translation initiation factor 2 subunit α (eIF2α, encoded by EIF2S1), facilitating translation of ATF4 (Wek, 2018). ATF4 transcripts were expressed at significantly higher levels in basal-like TNBCs when compared to claudin-low TNBCs (Fig. 4H; Table S10). Interestingly, TGFBR3 levels were was significantly anti-correlated with ATF4 levels and every detectable transcript that encoded ATF4 signaling components, including three well-described eIF2α kinases EIF2AK1 (HRI), EIF2AK2 (PKR) and EIF2AK3 (PERK) in TNBCs (basal-like and claudin-low subtypes), but not in other subtypes of breast cancer (Fig. 4I,J; Tables S11–S15). In TNBCs, some ATF4 target genes displayed no correlation [CTH, DDIT3 (CHOP), FGF21, and HSPA5 (BiP, GRP78)] or positive correlation (ATF3 and PPP1R15A) with TGFBR3 levels (Fig. 4I; Table S11). However, we identified a small subset of three well-characterized ATF4 target genes (ASNS, TRIB3 and ATF5) whose levels were significantly anti-correlated with TGFBR3 in TNBC-specific RNA-seq data (Fig. 4I,J; Table S11).

To assess the protein abundance of ATF4 and ATF4-related genes in TNBCs, we examined the data from three quantitative proteomic databases of breast cancers. The study by Lawrence et al. (2015) compared protein abundance across breast cancer cell lines and primary TNBCs, to a depth of more than 13,000 distinct proteins, which is higher than 95% of the estimated number of expressed genes (Lawrence et al., 2015; Nagaraj et al., 2011). In TNBCs, EIF2S1 and four eIF2α kinases [EIF2AK1, EIF2AK2, EIF2AK3, and EIF2AK4 (GCN2)] were detected (Fig. 4K), with EIF2S1, EIF2AK2 and EIF2AK4 being the most commonly expressed. However, ATF4 appeared to be absent in population-averaged measurements. This is likely to be because of the inherent limitations of proteomic technology – both highly variable and low-abundance proteins cannot be reliably detected (Shi et al., 2016; Kim et al., 2007). Similarly, proteins encoded by some ATF4 target genes were under the limit of detection. However, three ATF4 targets (ASNS, CTH and HSPA5) were ubiquitously expressed, suggesting ATF4 signaling remained activated in TNBCs. Differences in the abundance of ATF4 pathway proteins are not unique to this dataset. A similar pattern was observed in the data from two recent studies, which quantified more than 10,000 proteins across 40 and 105 patient samples each, consisting of multiple molecular subtypes of breast cancer (Fig. S4G,H) (Tyanova et al., 2016; Mertins et al., 2016).

RNA-seq data revealed that, in MCF10A and TNBC cell lines with either basal-like or claudin-low molecular profiles, ATF4 levels correlated positively with JUND (R=0.54, P<0.05) and negatively with TGFBR3 levels (R=−0.52, P<0.05) (Fig. 4L) (Daemen et al., 2013). We analyzed the protein expression of TGFBR3 and ATF4 of MCF10A-5E (Basal-B subtype) and 12 TNBC cell lines (seven of these Basal-A cell lines resemble basal-like tumors and five of these Basal-B cell lines resemble claudin-low tumors) (Heiser et al., 2012). MCF10A-5E expressed a high level of TGFBR3 protein and an undetectable level of ATF4 protein (Fig. 4M,N). The majority of the basal-like cell lines expressed very low or undetectable levels of TGFBR3 protein (5 of 7; 71%) (Fig. 4M,N). In all basal-like cell lines, ATF4 proteins were ubiquitously expressed. In contrast, moderate (1 of 5; 20%) and strong (4 of 5; 80%) TGFBR3 expression was observed in all claudin-low cell lines. Four of the five claudin-low cell lines (80%) expressed undetectable levels of ATF4, and only one claudin-low cell line (MDA-MB-231) expressed very low levels of ATF4 protein. We observed a very strong negative correlation between TGFBR3 protein levels and ATF4 protein levels in MCF10A-5E and TNBC cell lines (Fig. 4O), except for HCC1937 cells in which the TGFBR3 program is constitutively active (Bajikar et al., 2017; Wang et al., 2014). These results are consistent with mRNA expression pattern of TGFBR3 and ATF4 in human TNBCs found in TCGA datasets. Collectively, these results suggest that in TNBC tumors and TNBC cell lines, reduced TGFBR3 levels correlate with increased levels of ATF4 and ATF4 signaling.

We next asked whether TGFBR3 could affect the bioavailability of ATF4. TGFBR3 knockdown had no effect on basal ATF4 mRNA levels (Fig. 5A), and only slightly induced ATF4 expression in TGF-β1-treated cells (Fig. S5A). However, shTGFBR3 affected ATF4 at the protein level (Fig. 5B), suggesting a post-transcriptional mechanism of action for TGFBR3 on ATF4. Endoplasmic reticulum stress is known to be activated in cells with accumulated proteins (Oakes and Papa, 2015). As expected, we observed that ATF4 protein was induced upon treatment with the proteasome inhibitor MG-132 (Fig. 5C). shTGFBR3 and MG-132 combination treatment caused synergistically increased ATF4 levels. These results support the hypothesis that TGFBR3 plays a role in ATF4 protein destabilization.

To evaluate whether TGFBR3 alone was sufficient to regulate ATF4 activity, we surveyed the signaling capabilities of ATF4. By quantifying expression of a panel of ATF4 target genes, we discovered that the measured expressions were divided into two groups with distinct patterns. Group 1 was not affected by TGFBR3 signaling (Fig. S5B). In contrast, the genes of group 2, ATF5 and TRIB3, received a key, negative input specifically from TGFBR3, as knockdown of TGFBR3 significantly enhanced the levels of ATF5 and TRIB3 transcripts (Fig. 5D). The specificity of the effects was confirmed by reversal of the ATF5 expression upon TGFBR3 overexpression (Fig. 5E,F).

ATF4 was undetectable at the protein level in MCF10A-5E cells (Fig. 4M,N). To evaluate whether ATF4 promotes resistance to anoikis, we overexpressed ATF4 in this cell line (Fig. 5G). Overexpression of ATF4 selectively enhanced the levels of ATF5 and TRIB3 transcripts (Fig. 5H; Fig. S5C), consistent with the TNBC specificity in comparison with other ATF4 target genes (Fig. 4I,J). ATF4-expressing cells were placed in suspension culture for 4 days. We confirmed that ATF4 overexpression enhanced cell survival following detachment (Fig. 5I). To determine whether TGFBR3 loss-dependent ATF4 induction is important for acquisition of anoikis resistance, we depleted ATF4 by RNAi in TGFBR3-knockdown cells (Fig. 5J). Inhibiting ATF4 upregulation resulted in profound apoptosis in detached TGFBR3-knockdown cells (Fig. 5K). Thus, ATF4 upregulation specifically suppressed detachment-induced cell death. We examined the function of shTGFBR3-induced ATF4 protein among cell population by confocal immunofluorescence. In TGFBR3-knockdown cell aggregates, the expression of ATF4 protein and cleavage of PARP were mutually exclusive (i.e. did not occur in the same cells) (Fig. 5L; Fig. S5D, Table S16). These results together associate the TGFBR3 upregulation with a cellular deficiency in bioactive ATF4 in human epithelial cells. These data further support the fact that TGFBR3-mediated anoikis is likely to be mediated through inhibition of ATF4.

To assess the clinical significance of TGFBR3 downregulation in breast cancer, we performed patient survival analysis using the Kaplan–Meier plotter (Győrffy, 2021; Györffy et al., 2010). In human breast cancers, low TGFBR3 expression was associated with poor relapse-free survival (RFS), overall survival (OS), distant metastasis-free survival (DMFS), and post-progression survival (PPS) (Fig. 6A,B, Fig. S6A,B). In TGFBR3-low breast cancers, high levels of ATF4 (Fig. 6C,D; Fig. S6C,D), ATF5 (Fig. 6E,F; Fig. S6E,F) and TRIB3 (Fig. 6G,H; Fig. S6G,H) were associated with poor RFS, OS, DMFS and PPS. Taken together, these data support our conclusion that TGFBR3 loss indicates poor patient survival, and the upregulation of ATF4 and ATF4 target genes (ATF5 and TRIB3) are prognostic factors for poor patient survival in TGFBR3-low breast cancer.

During the lumen formation of mammary epithelial cells and dissemination of breast cancer cells, detachment induces a specialized cell death, termed anoikis. Cancer cells acquire anoikis resistance through several mechanisms, including activation of cell surface receptor signaling. Our transcriptional profiling revealed that TGFBR3 was the most upregulated receptor in ECM-detached cells. However, clinical data from TCGA and METABRIC datasets showed that breast cancer patients with high expression of TGFBR3 had much longer RFS, OS, DMFS and PPS than those with low TGFBR3 expression (Fig. 6A,B; Fig. S6A,B; Györffy et al., 2010). A low TGFBR3 level correlated with high tumor grade and poor prognosis (Fig. 1B) (Pereira et al., 2016; Gao et al., 2013; The Cancer Genome Atlas Network et al., 2012; Cerami et al., 2012). These findings raised the possibility that detachment-induced TGFBR3 had the potential to serve as a tumor suppressor and that its inactivation could promote breast cancer progression.

In this study, we investigated the mechanisms and mediators through which TGFBR3 influenced cell fate decisions during ECM deprivation. We showed that TGFBR3 synthesis is induced in detached cells, and demonstrated the ability of TGFBR3 to inhibit cell survival in ECM-poor microenvironments. ECM deprivation was associated with increased oxidative stress, and an induced ISR was critical for anoikis resistance. Blocking TGFBR3-mediated signaling enhanced ATF4, thus reducing anoikis of cells in suspension.

The initiation of anoikis involves changes in the interactions between integrin, ECM and other cell surface receptors. Although integrin–receptor interactions are well reported, the studies emphasized their role in the suppression of the intrinsic apoptotic pathway. TGFBR3 was the first TGF-β receptor cloned (López-Casillas et al., 1991; Wang et al., 1991). Having a short cytoplasmic domain with no intrinsic kinase activity, TGFBR3 had classically been thought to act as a coreceptor that selectively recognized ligands and promoted receptor complex assembly (López-Casillas et al., 1993; Henen et al., 2019). However, the genetic evidence of early embryonic lethality in Tgfbr3^−/− mice (Stenvers et al., 2003), together with the ability of TGFBR3 to drive epithelial–mesenchymal plasticity (Kirkbride et al., 2008), suggests that there are unique and non-redundant roles for TGFBR3 during development and tumorigenesis. During 3D organoid culture, the ECM-deprived subpopulation of cells expressed TGFBR3. This finding was recapitulated with cells forced to grow in suspension (Fig. 1F,G). We investigated whether TGFBR3 signaling was critical in modulating the transcriptional or translational, and post-translational response to cell dissemination. By considering ECM deprivation as a key variable, we uncovered TGFBR3-dependent activation of life–death decision proteins within the intrinsic pathway of apoptosis, including caspase 3 and BCL-2 family members (BCL2L1, BAD, and BBC3). We propose that this is one of the mechanisms leading to anoikis, indicating a role of TGFBR3 in preventing tumorigenesis.

TGF-β signaling is involved in the progression of human breast cancers (Wang et al., 2014, 2012; Wang and Janes, 2014). TGF-β1 initially binds to TGFBR2, which activates TGFBR1. Activated TGFBR1 phosphorylates SMAD2 and SMAD3, and activates the SMAD2/3 axis. In mouse lung fibroblasts and human breast cancer cell lines, TGF-β1 can induce phosphorylation of SMAD1 and SMAD5 (Ramachandran et al., 2018; Schwartze et al., 2014). In this study, we showed that the coreceptor TGFBR3 played a role in supporting TGF-β1-induced SMAD1/5 phosphorylation and in attenuating TGF-β1-induced SMAD3 phosphorylation in human epithelial cells (Fig. 2C,D). TGFBR3 loss might endow cells with survival ability through the SMAD2/3-dependent axis.

Under stress conditions, the ATF4 pathway is activated to preserve homeostasis. ATF4 is overexpressed and responsible for the aggressiveness in TNBCs (Gonzalez-Gonzalez et al., 2018; van Geldermalsen et al., 2016). We demonstrated that TGFBR3 blocked ATF4 protein bioavailability in epithelial cells, and selectively prevented the expression of ATF4 target genes ATF5 and TRIB3. The pseudokinase Tribbles homolog 3 (TRIB3) is a stress sensor. TRIB3 simultaneously activates the ERK1/2 pathway and modulates TGF-β1-mediated transcriptional activity in breast cancer. Its expression is associated with poor patient survival (Wennemers et al., 2011; Izrailit et al., 2013). Similarly, activating transcription factor 5 (ATF5) promotes survival of malignant cells by stimulating the expression of BCL2-family antiapoptotic regulators BCL2 and MCL1 (Dluzen et al., 2011; Sheng et al., 2010). ATF4 and ATF5 are also the key regulators of the mitochondrial unfolded protein response (UPR^mt). In breast cancer, UPR^mt is common in metastatic lesions (Kenny et al., 2019, 2017). Using SOD2 as a surrogate marker, we observed that ECM deprivation caused UPR^mt (Figs 1D, 4A,B), which could be rescued by shTGFBR3-mediated TGFBR3 knockdown (Fig. 4C). Future work will be required to evaluate whether UPR^mt-responsive ATF4 (Quirós et al., 2017) and ATF5 (Fiorese et al., 2016), and their induction upon TGFBR3 loss, protect epithelial cells against mitochondrial stress during tumorigenesis.

In conclusion, we highlight a mechanism of anoikis regulation whereby TGFBR3 is synthesized during ECM deprivation. TGFBR3 increased the selectivity for ligands that induced tumor-suppressive properties. We further demonstrated that TGFBR3 destabilized ATF4 protein. In light of our current data, it will be important to investigate in what tumor contexts TGFBR3 loss is required for upregulation of ISR, and more broadly, how TGF-β1 pathways are rewired in matrix-detached cancer cells and contribute to anoikis resistance.

The MCF10A-5E clone was previously reported and maintained as described for the parental MCF10A cell line (American Type Culture Collection, ATCC; Debnath et al., 2003; Janes et al., 2010). MCF10A-5E cells were obtained from the laboratory of Kevin Janes at University of Virginia, USA, and cultured in Dulbecco's modified Eagle's medium with nutrient mixture F-12 medium (Gibco) plus 5% horse serum (Gibco), 20 ng ml⁻¹ animal-free recombinant human epidermal growth factor (EGF) (Peprotech), 10 µg ml⁻¹ insulin from bovine pancreas (Sigma), 0.5 µg ml⁻¹ hydrocortisone (Sigma), 100 ng ml⁻¹ cholera toxin from Vibrio cholera (Sigma), and 1× penicillin and streptomycin (Gibco). This cell line from a female, is grown in a 5% CO₂ humidified incubator at 37°C, authenticated by short tandem repeat profiling by ATCC, and was confirmed negative for mycoplasma contamination. Lysates of the triple-negative breast cancer cell lines (HCC1599, HCC1937, HCC1143, MDA-MB-468, HCC70, HCC1806, HCC1187, BT-549, Hs578T, SUM159PT, MDA-MB-231, and MDA-MB-436) were also a kind gift from Professor Kevin Janes.

TGF-β1 (Peprotech) was reconstituted at 50 µg ml⁻¹ in 10 mM citric acid (pH 3.0) as recommended by the manufacturer. GDF11 (Peprotech) was reconstituted in sterile water at 250 µg ml⁻¹. Cytokines were prepared at 20× final concentration in serum-free medium. A 1/20th volume of cytokines was added to cells for stimulation.

pLKO.1 shTGFBR3 puro (Addgene #58696, shTGFBR3#1) and pBabe neo TGFBR3-HA (Addgene #83095) were previously described (Wang et al., 2014; Bajikar et al., 2017). pLKO.1 puro (Addgene #8453), Tet-pLKO-neo (Addgene #21916), pLKO.1 GFP shRNA (Addgene #30323) (Sancak et al., 2008), pBABE-neo (Addgene #1767) (Morgenstern and Land, 1990), pDONR223 LacZ (Addgene #25893), pDONR223_ATF4_WT (Addgene #82190), and pLX304 (Addgene #25890) (Yang et al., 2011) were obtained commercially.

The shRNA target sequences of human TGFBR3, SOD2, ATF4 and luciferase from Promega, and Escherichia coli lacZ from the RNAi Consortium shRNA Library were as follows: shTGFBR3#1, 5′-CCAAGCATGAAGGAACCAAAT-3′ (TRCN0000033430); shTGFBR3#2, 5′-GGAGTTGGTAAAGGGTTAATA-3′ (TRCN0000359081); shSOD2#1, 5′-GGATGCCTTTCTAGTCCTATT-3′ (TRCN0000320665); shSOD2#2, 5′-TTGGTTCCTTTGACAAGTTTA-3′ (TRCN0000350349); shLuc, 5′-AGAATCGTCGTATGCAGTGAA-3′ (TRCN0000072250); shATF4#1, 5′-TGGATGCCCTGTTGGGTATAG-3′ (TRCN0000329696); shATF4#2, 5′-GCCAAGCACTTCAAACCTCAT-3′ (TRCN0000013575); and shLacZ, 5′-GTTCCGTCATAGCGATAACGA-3′ (TRCN0000072238). shTGFBR3#2 was cloned into pLKO.1 puro as previously described (Pereira et al., 2020). shSOD2#1, shSOD2#2, shLuc, shATF4#1, shATF4#2 and shLacZ were cloned into Tet-pLKO-neo. Targeting sequences were modified to replace the XhoI restriction site in the shRNA loop (CTCGAG) with a PstI site (CTGCAG). Oligonucleotides were heated to 95°C in annealing buffer (10 mM Tris-HCl pH 7.5, 100 mM NaCl, and 1 mM EDTA) for 5 min and annealed by cooling slowly to room temperature. Annealed primers were phosphorylated with T4 polynucleotide kinase (NEB #M0201) and then cloned into pLKO.1 puro or Tet-pLKO-neo that had been digested with EcoRI and AgeI. For ATF4 expression, pDONR223 LacZ (control) and pDONR223_ATF4_WT were recombined with pLX304 by using Gateway LR recombination (Invitrogen) to obtain pLX304 LacZ-V5 blast and pLX304 ATF4-V5 blast. All plasmids were verified by sequencing.

Lentiviruses and retroviruses were prepared in 293T/17 cells (ATCC) by triple transfection of the lentiviral vector (pLOK.1, pLX304) together with lentiviral gag-pol packaging vector psPAX2 (Addgene #12260) and VSV-G envelope expressing plasmid pMD2.G (Addgene #12259) or double transfection of the retroviral vector (pBABE) with retroviral packaging vector pCL ampho (Novus) as described previously (Liu et al., 2022, 2020; Kumari et al., 2021; Sahoo et al., 2019; Wang et al., 2014). Briefly, 1.25 µg of lentiviral plasmid was mixed with 0.75 µg psPAX2 and 0.5 µg pMD2.G, or 1.25 µg of retroviral plasmid was mixed with 1.25 µg pCL ampho. 10 µl 2.5 mM CaCl₂ was added to the mixture. 0.1× Tris-EDTA (TE) buffer (pH 7.6) was added to the mixture to make a final volume of 100 µl. 100 µl of 2× HEPES-buffered saline was added to the mixture and mixed by pipetting. The mixture was allowed to stand at 1 min, and the entire mixture was added dropwise to a well of 293T/17 cells. The precipitates were incubated with the cells for 4–6 h at 37°C, then the medium was aspirated and carefully replaced with 1 ml of growth medium. Viral supernatant was collected at 48 h (for retrovirus) or 24 and 48 h (for lentivirus) and sterile filtered. Cells were transduced with 500 µl virus in culture medium containing 8 µg ml⁻¹ polybrene. Cells were selected in puromycin (2 µg ml⁻¹; MP Biomedicals #0210055225), G418 (300 µg ml⁻¹; Sigma #A1720) or blasticidin (10 µg ml⁻¹; Invitrogen #R21001) until control plates had cleared.

Brightfield images were acquired with a 4× objective with a Q-Color3™ digital imaging system (Qimaging) as described previously for 3D spheroids (Bajikar et al., 2017).

RNA purification was performed exactly as described previously (Wang et al., 2014; Bajikar et al., 2017; Pereira et al., 2020). Briefly, RNA was extracted from cells by lysing cells plated on an individual 3.5-cm-diameter dish in 1 ml Trizol. The aqueous phase was separated by adding 100 µl chloroform replacement separation reagent, shaking the samples vigorously for 15 s. The samples were incubated at room temperature for 3 min, and then centrifuged for 15 min at 12,000 g on a benchtop centrifuge at 4°C. The aqueous phase of the samples that contained RNA was transferred to a fresh microcentrifuge tube. 0.5 ml isopropanol was added to the aqueous phase and then incubated at 4°C for 30 min. Samples were centrifuged at 12,000 g for 10 min on a benchtop centrifuge at 4°C and washed with 1 ml 75% ethanol. Pelleted RNA was dried and resuspended in 20 µl nuclease-free water. The RNA concentration was quantified on a Nanodrop machine.

cDNA synthesis and quantitative real-time PCR (qRT-PCR) were performed as previously described (Wang et al., 2014; Bajikar et al., 2017; Pereira et al., 2020) with the primers listed in Table S17. Briefly, first-strand cDNA synthesis was performed on 100–500 ng DNA-free RNA with reverse transcriptase and oligo(dT)₂₀. qRT-PCR was performed using 0.1 µl cDNA template and 5–10 pmol forward and reverse primers together with a master mix used at a final concentration of 1× PCR buffer with 20 mM Tris-HCl pH 8.8, 10 mM KCl, 10 mM (NH₄)₂SO₄, 2 mM MgSO₄, 0.1% Triton X-100, 400 µM dNTPs, 0.2× SYBR Green I Nucleic Acid Gel Stain (Invitrogen #S7563) and 0.02 U µl⁻¹Taq DNA polymerase in a final reaction volume of 15 µl. Real-time reactions were tracked on a Bio-Rad CFX Connect instrument, using the following thermal protocol: 95°C denaturation (1.5 min), 40 cycles of 95°C denaturation (10 s), 60°C annealing (10 s), 72°C elongation (12 s) with a fluorescence read at the end of elongation. Melt curves were measured by a 65°C to 95°C touch up in 0.5°C increments with a fluorescence read after each increment. Fluorescence measurements were analyzed using Bio-Rad CFX Manager 3.1 ISO software. Samples were normalized to the geometric mean of GAPDH, PRDX6, HINT1, B2M, GUSB and PPIA levels.

Whole-mount immunofluorescence of 0 h, day 2 or day 3 MCF10A-5E cell aggregates was performed as previously described (Wang et al., 2014; Bajikar et al., 2017; Pereira et al., 2020). Briefly, cell aggregates were fixed with 3.7% paraformaldehyde for 15 min at room temperature and permeabilized with 0.3% Triton X-100 in PBS. Washed cell aggregates with 0.1% Tween-20 in PBS. Cell aggregates were blocked for 1 h at room temperature with 1× western blocking reagent (Roche #11921673001) diluted in 0.1% Tween 20 in PBS. The following primary antibodies were added into blocking buffer at the indicated dilutions and incubated overnight at 4°C: TGF-β Receptor III (Cell Signaling Technology #2519, 1:200, RRID:AB_390707), E-cadherin (Clone 36) (BD Biosciences #610182, 1:200, RRID:AB_397581), cleaved PARP (Asp214) (BD Biosciences #552596, 1:200, RRID:AB_394437), and ATF-4 (D4B8) (Cell Signaling Technology #11815, 1:200, RRID:AB_2616025). Cell aggregates were incubated in secondary antibody diluted in blocking buffer for 1 h at room temperature, with DAPI (0.5 µg ml⁻¹) used as a nuclear counterstain. Autofluorescence was quenched with 10 mM CuSO₄ in 50 mM NH₄Ac (pH 5.0) for 10 min. Slides were washed once in 0.1% Tween 20 in PBS and mounted with 0.5% N-propyl gallate in 90% glycerol in PBS (pH 8.0). The edges of the coverslip were sealed in clear nail polish and allowed to air dry. Slides were stored at 4°C until imaging. Full blots for the images presented in the figures are shown in Fig. S7.

Quantitative immunoblotting was performed as previously described (Wang et al., 2014, 2009; Bajikar et al., 2017; Pereira et al., 2020; Cirit et al., 2010; Weiger et al., 2009). Briefly, samples were prepared using an Mem-PER Plus Membrane Protein Extraction Kit (Thermo Fisher Scientific # 89842) or in dithiothreitol (DTT)-containing Laemmli sample buffer to a total volume of 20–40 µl. Polyacrylamide gels (8, 10, 12 or 15%) were cast, and samples were electrophoresed in Tris-glycine running buffer (25 mM Tris base, 250 mM glycine, and 0.1% SDS) at 130 V. Proteins were transferred to a PVDF membrane (Millipore; Immobilon-FL, 0.45 µm pore size) in a Mini Trans-Blot Electrophoretic Transfer Cell (Bio-Rad) in transfer buffer (24 mM Tris base, 190 mM glycine and 10% methanol) without SDS at 100 V for 1 h on ice. Membranes were blocked with 0.5× Odyssey blocking buffer in PBS (for detection of total proteins) or in TBS (for detection of phosphorylated targets) for near-infrared fluorescent detection. Primary antibodies were diluted with 0.5× Odyssey blocking buffer plus 0.1% Tween-20. Primary antibodies recognizing the following proteins or epitopes were used: TGF-β Receptor III (Cell Signaling Technology #2519, 1:1000, RRID:AB_390707), sodium/potassium ATPase [EP1845Y] (Abcam #ab76020, 1:1000, RRID:AB_1310695), vinculin (Clone V284) (Millipore #50-386, 1:10,000, RRID:AB_309711), phospho-Smad2 (Ser465/467) (138D4) (Cell Signaling Technology #3108, 1:1000, RRID:AB_490941), Smad3 (phospho S423+S425) [EP823Y] (Abcam #ab52903, 1:1000, RRID:AB_882596), phospho-SMAD1/5/9 (Ser463/465/467, D5B10) (Cell Signaling Technology #13820, 1:1000, RRID:AB_2493181), SMAD2 (Clone L16D3) (Cell Signaling Technology #3103, 1:1000, RRID:AB_490816), Smad3 (C67H9) (Cell Signaling Technology #9523, 1:1000, RRID:AB_2193182), Smad1 (D59D7) (Cell Signaling Technology #6944, 1:2000, RRID:AB_10858882), GAPDH (Clone 6C5) (Ambion #AM4300, 1:10,000, RRID:AB_437392), tubulin (Abcam #ab89984, 1:20,000, RRID:AB_10672056), caspase-3 (Cell Signaling Technology #9662, 1:1000, RRID:AB_331439), ATF4 (D4B8) (Cell Signaling Technology #11815, 1:1000, RRID:AB_2616025) and HA (3F10) (Roche #11867423001, 1:2000, RRID:AB_390918). Membranes were washed and probed with secondary antibody diluted with 0.5× Odyssey blocking buffer. Membranes for fluorescence imaging were scanned on an Odyssey Classic infrared scanner (LI-COR) at 169-μm resolution and 0-mm focus offset. Raw 16-bit images of exposures were analyzed in ImageJ with the gel analysis tool.

The suspension assay was performed as previously described (Wang et al., 2014). Briefly, MCF10A-5E cells expressing the indicated constructs were trypsinized and plated at 400,000 cells ml⁻¹ in assay medium containing 5 ng ml⁻¹ EGF on poly-(2-hydroxyethyl methacrylate) (poly-HEMA)-coated tissue culture plates. At the indicated time points, medium was removed by centrifugation at 800 g for 3 min.

For RNA isolation and purification, cells were lysed directly in Trizol (1 ml per well for a six-well plate) in a chemical fume hood. The denatured genomic DNA in each sample was sheared by pipetting the lysate up and down several times with a P200 micropipet. After the lysate had become less viscous, the samples were incubated at room temperature for 5 min to complete the dissociation of mRNAs from ribonuclear proteins, and then were stored at −80°C. For immunoblotting analysis, cells were resuspended in total 1 ml ice-cold PBS and were spun at 800 g for 3 min. Supernatant was aspirated and cells were lysed with 100 µl RIPA buffer. The lysate was vortexed, incubated on ice for 15 min, and centrifuged at 21,000 g for 15 min at 4°C. The supernatant (avoiding the pellets) was collected into a new microcentrifuge tube, and then was stored at −80°C.

Gene expression files (mRNA expression z-scores) and genomic profiles (mutations and putative copy-number alterations from DNAcopy) of The Cancer Genome Atlas (TCGA) and METABRIC datasets were downloaded via the TCGA Data Portal (Gao et al., 2013; Cerami et al., 2012). We used the data set of 1904 primary breast tumors that were deposited from the following studies: Invasive Breast Carcinoma/Breast Cancer (METABRIC datasets EGAS00000000083 and EGAS00001001753 from references Curtis et al., 2012 and Pereira et al., 2016, respectively) (The Cancer Genome Atlas Network et al., 2012). We downloaded the mRNA expression z-scores of TGFBR3, CDH1, ATF4, HSPA5, EIF2S1 (missing information), EIF2AK3, EIF2AK4 (missing information), EIF2AK2, EIF2AK1, ASNS, TRIB3, DDIT3, PPP1R15A, ATF3, ATF5, FGF21 and CTH. Cases of breast tumor were classified by their clinical attributes (Cancer Type Detailed, Neoplasm Histologic Grade, Nottingham prognostic index, ER Status, ER status measured by IHC, PR Status, Integrative Cluster, and Pam50+Claudin-low subtype) using cBioPortal (https://www.cbioportal.org/). Statistical significance for the difference in gene expression levels between subgroups was determined by using Welch's t-test. Co-expressions were studied by Pearson and Spearman correlation tests.

Transcriptomic expression data on ECM-attached and ECM-detached MCF10A-5E cells in 3D culture were reported previously (Wang et al., 2014). The relative expression in inner/outer cells (Log2 fold change) and s.e.m. of 8259 genes were downloaded from the NCBI Gene Expression Omnibus data with accession number GSE41527, and were subsequently visualized by using Igor Pro (WaveMetrics).

The RNA sequencing data of MCF10A cell line and TNBC cell lines in JWGray Breast Cancer Cell Line Panel (Daemen et al., 2013) were downloaded from Synapse (ID: syn2346643). We used the data set that contained the transcriptome of a normal epithelial cell line (MCF10A), and 11 TNBC cell lines (HCC1599, HCC1937, HCC1143, HCC38, HCC70, HCC1806, BT549, Hs578T, MDA-MB-231, MDA-MB-453, and HCC1395). Relative abundances in transcripts per million (TPM) of TGFBR3, ATF4 and JUND were calculated by converting fragments per kilobase of exons per million mapped reads to TPM. To avoid infinite values in log calculations, a value of 1 was added to all TPM values before log₁₀ transformation. Co-expressions were studied by Pearson correlation test, and were subsequently visualized by using Igor Pro (WaveMetrics).

Kaplan–Meier survival analysis was performed as described previously (Győrffy, 2021). The prognostic values of TGFBR3 for patients with breast cancer were determined using the Kaplan–Meier plotter database (http://kmplot.com/analysis/) that contained a total of 7.830 patients with breast cancer with survival data from Gene Expression Omnibus, European Genome-Phenome Archive and TCGA. For each gene symbol, the Affymetrix probe IDs were individually entered to obtain Kaplan–Meier plots. Patients were divided into high and low expression groups according to the ‘auto select best cutoff’. Information on relapse-free survival (RFS), overall survival (OS), distant metastasis-free survival (DMFS), and post-progression survival (PPS) was extracted. In subgroup analyses, the samples were further divided into high and low TGFBR3 expression groups based on the median TGFBR3 mRNA level. The prognostic values of ATF4, ATF5 and TRIB3 for patients with TGFBR3-low breast cancer were determined using the same method as mentioned above. The number of cases, hazard ratios (HRs), and log rank P-values were obtained from Kaplan–Meier plotter.

Clinical data for patient survival analysis were obtained from Kaplan–Meier plotter under accession numbers: E-MTAB-365, E-TABM-43, GSE11121, GSE12093, GSE12276, GSE1456, GSE16391, GSE16446, GSE16716, GSE17705, GSE17907, GSE18728, GSE19615, GSE20194, GSE20271, GSE2034, GSE20685, GSE20711, GSE21653, GSE2603, GSE26971, GSE2990, GSE31448, GSE31519, GSE32646, GSE3494, GSE37946, GSE41998, GSE42568, GSE45255, GSE4611, GSE5327, GSE6532, GSE7390 and GSE9195.

In-depth proteomics data for 20 breast cancer cell lines and four primary breast tumors were downloaded from Lawrence et al. (2015). We used the reported intensity-based absolute quantitation (iBAQ) profile of each sample as protein absolute abundances. iBAQ values represent the summed peptide intensity normalized by total proteome intensity for that sample and by the number of observable peptides for the protein. A breast tumor proteomic data set with 40 different breast cancer patient samples was downloaded from Tyanova et al. (2016). The most comprehensive database of 105 genomically annotated breast cancers was downloaded from Mertins et al. (2016). In brief, we downloaded the protein expression data of EIF2AK1, EIF2AK2, EIF2AK3, EIF2AK4, EIF2S1, ATF4, ASNS, ATF3, ATF5, CTH, DDIT3, FGF21, HSPA5, PPP1R15A and TRIB3. Except ATF4, proteins that were undetectable (ATF3, ATF5, DDIT3, FGF21, PPP1R15A, and TRIB3) were removed from the dataset. Data are presented as mean±s.e.m.

For all experiments in this study, the value and meaning of n, the definition of the central estimate, and the measure of dispersion are provided in the figure legend. The log-rank test was used to compare Kaplan–Meier survival curves. Correlation of mRNA–mRNA pairs of the gene set in cell lines was analyzed calculating the Pearson's and Spearman's correlation coefficients. All other two-sample comparisons were performed by Welch's two-sided t-test. No method was used to test the assumptions of parametric methods. All hypothesis testing was performed at a type I error rate of α=0.05.

We thank Kevin Janes for plasmid reagents, MCF10A-5E clone, and TNBC cell lysates. We thank Eleanore Sturgill and Yi-Ching Yeh for critically reading this manuscript and copyediting. We would also like to thank Cheng-Yu Wu, Anson Liu, and Chien-Wei Lee for assistance with protein BCA quantification. We acknowledge research support from the confocal imaging core in NTHU (sponsored by MOST 108-2731-M-007-001-).

The peer review history is available online at https://journals.biologists.com/jcs/lookup/doi/10.1242/jcs.258396.reviewer-comments.pdf.