The protein secretion modulator TMED9 drives CNIH4/TGFα/GLI signaling opposing TMED3-WNT-TCF to promote colon cancer metastases

How cells in primary tumors initially become pro-metastatic is not understood. A previous genome-wide RNAi screen uncovered colon cancer metastatic suppressor and WNT promoting functions of TMED3, a member of the p24 ER-to-Golgi protein secretion family. Repression of canonical WNT signaling upon knockdown (kd) of TMED3 might thus be sufficient to drive metastases. However, searching for transcriptional influences on other family members here we find that TMED3 kd leads to enhanced TMED9, that TMED9 acts downstream of TMED3 and that TMED9 kd compromises metastasis. Importantly, TMED9 pro-metastatic function is linked to but distinct from the repression of TMED3-WNT-TCF signaling. Functional rescue of the migratory deficiency of TMED9 kd cells identifies TGFα as a mediator of TMED9 pro-metastatic activity. Moreover, TMED9 kd compromises the biogenesis, and thus function, of TGFα. Analyses in three colon cancer cell types highlight a TMED9-dependent gene set that includes CNIH4, a member of the CORNICHON family of TGFα exporters. Our data indicate that TGFA and CNIH4, which display predictive value for disease-free survival, promote colon cancer cell metastatic behavior, and suggest that TMED9 pro-metastatic function involves the modulation of the secretion of TGFα ligand. Finally, TMED9/TMED3 antagonism impacts WNT-TCF and GLI signaling, where TMED9 primacy over TMED3 leads to the establishment of a positive feedback loop together with CNIH4, TGFα, and GLI1 that enhances metastases. We propose that primary colon cancer cells can transition between two states characterized by secretion-transcription regulatory loops gated by TMED3 and TMED9 that modulate their metastatic proclivities.


Introduction
The mechanisms that drive the development of metastatic states within primary tumors remain ill defined [1]. These are likely at work in cells with stem cell properties, leading to the appearance of metastasis initiating cells, and at least in colon and breast cancers there is little correlation of the time of metastatic spread with primary tumor size [2,3]. Genomic analyses in colon or other cancers suggest that metastases reflect the distant expansion of specific cells already present in heterogenous primary tumors without common metastatic-specific driver mutations [4,5].
Previous work has shown that WNT-TCF signaling in colon cancer is anti-metastatic since its direct repression in grafted human cancer cells enhances metastatic behavior [6,7]. This anti-metastatic role of WNT-TCF signaling is consistent with the in vivo, unbiased identification of the positive WNT-TCF modulator TMED3 as an endogenous suppressor of distant colon cancer metastases [8].
TMED3 belongs to a family of p24 proteins involved in selecting cargo in COP vesicles in the secretory ER-Golgi network [9]. Given the large diversity of cargo and the existence of only 10 TMED p24 proteins, it is likely that each can affect multiple secretion events in direct and indirect context-dependent manners. Moreover, p24 proteins can exist as monomers or dynamic complexes where one can affect the stability of others [10][11][12][13]. They appear to be non-redundant [13,14] and affect multiple signaling pathways in mammalian cells [15][16][17][18][19]. In flies and mammals, specific TMED proteins control WNT secretion [8,[20][21][22], and both TMED3 and WNT-TCF signaling act as metastatic suppressors in human colon cancer cells [6][7][8]. How TMED3 may repress the establishment of prometastatic states, however, remains unknown.

TMED3 regulates the mRNA levels of other TMED family members
To elucidate how blockade of TMED3 promotes prometastatic states in primary colon cancer cells, we first investigated if it could affect the expression of other TMED family members. Knockdown (kd) of TMED3 was achieved in CC14 primary human colon cancer cells [23], which are E 1554 ->frameshift APC mutant [7], using a previously validated [8] specific short-hairpin RNA (shTMED3 with kd of 95%; Fig.  1a). TMED9 was the only one upregulated more than twofold, whereas several TMED genes were downregulated, out of which TMED7 showed the greatest decrease (Fig. 1a).

TMED9 is required for distant metastases
Little is known about the TMED family in cancer and, specifically, nothing is known about the possible participation of TMED9 and TMED7 in metastases. This was therefore tested by subcutaneously grafting CC14 cells transduced with lentivectors expressing either shTMED7 (with kd of 80%) or shTMED9 (with kd of 90%) and inspecting the lungs of the recipient mice 4 weeks later for distant metastases. We tracked CC14 cells expressing lentivirus-encoded β-galactosidase (CC14 lacZ ) in order to detect distant metastases after the X-Gal reaction at singlecell resolution [24]; Fig. 1b, c. Kd of TMED9, but not of TMED7, resulted in a significant change in the number of metastases with little change in grafted primary tumor volume (Supplemental Fig. 1a, b). TMED9 kd produced a similar reduction in micro and larger metastases (Fig. 1b,  Supplemental Fig. 1c-e). The requirement of TMED9 for distant metastases was recapitulated in primary human colon cancer CC36 lacZ cells [23] and in the human colon cancer cell line LS174T lacZ (Fig. 1b, c, Supplemental Fig.  1d). A second shRNA against TMED9 with kd of 96% used to validate the initial data yielded a similar result (Supplemental Fig. 1e). Rare liver metastases were also abrogated by kd of TMED9 (Supplemental Fig. 2).
The metastatic phenotypes were fully recapitulated by the Boyden chamber transfilter assay testing for cancer cell migration [25]; Fig. 1d. Using this assay, TMED9 was shown to be similarly required for the migration of human U251 glioblastoma cells (Fig. 1d), a tumor cell type that readily invades the brain parenchyma [26] used here to test whether TMED9 kd might also affect other tumor types.
As colon cancer metastases derive, at least in part, from CD133 + cancer stem cells [27] we quantified their abundance but did not find a difference between parental vs. shTMED9 pools (5% vs. 5.2% for CC14; 0.14% vs. 0.15% for CC36). This result suggests that the reduction of metastases is not simply due to the loss of CD133 + cancer stem cells upon kd of TMED9.

TMED9 is epistatic to TMED3
To establish an order of action of TMED3 vs.TMED9 we performed epistatic analyses using shRNAs to kd each gene alone and in combination in CC14 cells and measuring the number of distant metastases. Whereas shTMED9 decreased and shTMED3 increased distant lung metastasis compared with controls ( Fig. 2a, b), the simultaneous expression of these two shRNAs yielded an shTMED9-like phenotype with a drastic decrease of lung metastases (Fig. 2a, b). The increase in metastases by the repression of TMED3 is thus dependent on TMED9 activity. TMED3-WNT signaling could therefore act as an anti-metastatic brake in part by repressing TMED9. However, as repressing WNT signaling is sufficient to enhance metastases [6][7][8], these results raised the question of how TMED9 and WNT might interact.

TMED9 positively regulates genes with migratory/ invasive functions and represses WNT signaling
To gain insights into the function of TMED9 we compared the transcriptomes of CC14 cells expressing shTMED9 (CC14 shTMED9 ) and control infected sibling (CC14 control ) cells. Inspection of regulated genes (Fig. 2c, Supplemental  Fig. 3a) revealed potential mechanisms for the antimetastatic phenotype of kd of TMED9. For instance, genes with enhanced expression included metastatic suppressors (AKAP12), and genes with repressed expression included genes involved in EMT (MMP28, ADAM8, SNAI3) and cancer progression (DPEP1, LAMP3, GSPG4). Some of these have been correlated with poor prognosis in different cancer types [28,29]. Kd of TMED9 did not significantly alter the expression of other TMED genes.
Gene ontology enrichment analyses on the resulting RNAseq data revealed global alterations of components of the extracellular space and endoplasmic reticulum (ER) (Supplemental Fig. 3b). Mining public experimental data obtained from human colon cancer cells highlighted an upregulated enrichment of canonical WNT signaling in cells with kd of TMED9 (LEF1; Gene Set Enrichment Analysis (GSEA); Oncogenic Signatures normalized enrichment score (NES) = 2.05; Fig. 2d top). This was supported by the correlation of highly expressed genes in CC14 shTMED9 cells with those having TCF-binding sites within 4 kb of the start site (GSEA transcription factors; NES = 2.21; Fig. 2d bottom). Indeed, upregulated genes by kd of TMED9 included a considerable number of WNT signaling components, targets or modifiers in multiple contexts such as WNT11, WNT3, MUC16, VGLL1, SOSTDC1, and LGR5 (Fig. 2c, Supplemental Fig. 3 [30][31][32]. WNT11 has been widely reported to act in the non-canonical migratory/planar polarity/Ca ++ Fig. 1 Regulation of the TMED family by kd of TMED3 and metastatic phenotypes of cells with kd of TMED9. a Histogram of rt-qPCR results for TMED mRNA levels in CC14 cells expressing shTMED3. Numbers are ratios of normalized CT values of shTMED3 over control cells. Note that TMED8 has been suggested to lie outside of the p24 family [9]. TMED10 was not detected in CC14 cells. b Whole views of X-Gal-stained left lung lobes showing metastases in blue as noted. c Quantification of the number of lacZ + lung metastases per animal vs. primary xenograft volume under the different experimental conditions. Each animal carried one tumor per flank. The number of mice involved for CC14 grafts were seven for vector alone control cells and eight for shTMED9 cells; for CC36 grafts the numbers were seven for control cells and nine for shTMED9 cells; and for Ls174 grafts they were five for control cells and seven for shTMED9 cells. d Kd of TMED9 reduces cancer cell migration. Images and quantification of lacZ + cells (in blue) that have crossed the membrane in transfilter assays for different colon cancer (CC14, CC36, LS174T) and glioblastoma (U251) cells as noted, detected after X-Gal staining cells within the filter. Quantification derives from triplicate experiments with independent batches. In this and all figures, error bars are s.e.m. and p values from two-tailed Student's t -tests are in red if significant (p < 0.05). Scale bar = 0.25 cm for b and 80 µm for d pathways and repress canonical WNT-TCF signaling [33][34][35], although it may also act canonically [36]; see below. These results can be consistent with an anti-metastatic role of elevated canonical WNT signaling [6] but less so with a pro-metastatic function of non-canonical WNT signaling observed in other cancers [35,37,38].
Taken together, the data indicate that TMED9 normally represses different aspects of WNT signaling.

Global opposite gene regulation by TMED9 and TMED3
To address if TMED9 and TMED3 might exert global antagonistic effects we analyzed the transcriptome of cells with kd of TMED3 and compared it with that of cells with kd of TMED9.
Analyses of transcripts altered in TMED3 kd cells over control levels revealed a group of 63 RNAs enhanced twofold or more that included several genes previously shown to promote pro-metastatic behavior in different cancers, including SOX8, ASCL2, FGF19, and CXCR4 (Supplemental Fig. 4). It also revealed a group of 146 transcripts, in addition to TMED3, repressed twofold or more that included multiple ZNF zinc finger proteins as well as transcription factor determinants such as DACH1 (Supplemental Fig. 4), a repressor of colon cancer tumor cell migration and invasion [42].
Comparison of shTMED3 vs. shTMED9 RNAseq data yielded 179 transcripts regulated by both (Supplemental Fig. 5). Plotting their expression as fold change over control in a graph with orthogonal axes showed that the great majority of genes (91%) was regulated in an opposite manner (Fig. 2e): 68% in the putative anti-metastatic shTMED9 high ;shTMED3 low quadrant and 23% in the putative pro-metastatic shTMED9 low ;shTMED3 high quadrant, versus only 8% in the high:high and 1% in the low:low quadrants.
The results show that TMED9 has pro-metastatic function and that it works below TMED3 in this context. Furthermore, they indicate that the antagonistic actions of these two TMED proteins control metastases via the global regulation of multiple genes that participate in the metastatic process. However, what is not known is if both TMED proteins regulate the same or diverse signaling events.
A common signature of shTMED9 in different primary colon cancer cells reveals genes encoding ER-Golgi network proteins To begin to address TMED9 actions in more detail we sought to delineate a conserved gene expression signature with which to track, albeit indirectly, TMED9 activity. Thus, to complement the findings on CC14 cells we determined the transcriptomes of CC36 control and CC36 shTMED9 cells since both CC36 and CC14 are primary colon cancer uncloned cell populations [23] and both respond to shTMED9 similarly (Fig. 1). Both have been previously determined to harbor active WNT signaling and produce enhanced metastases in response to WNT blockade by dnTCF [6,7]. Note that the vast majority are found either in the top left (shTMED9 high ;shTME-D3 low ) putative anti-metastatic or in the bottom right (shTMED9 low ; shTMED3 high ) putative pro-metastatic quadrants. Lists of potential anti-metastatic (blue background) and pro-metastatic (red background) genes are given on each side of the graph, ranked by shTMED9/ shTMED3 normalized fold change ratios. Scale bar = 0.25 cm for a Enrichment analyses of the CC36 shTMED9 vs. CC36 control transcriptomes did not highlight changes in WNT signaling (Supplemental Fig. 6), which may be owing to the different nature of the cells: CC36 cells do not form epithelial colonies as CC14 cells do. Instead, they display a mesenchymal phenotype with enhanced > 50-fold migration as compared with CC14 [23] and aspects of WNT signaling or its responses are already downregulated. For example, comparison of normalized baseline rt-qPCR ct gene expression values revealed generally higher expression of WNT signaling components/targets in CC14 vs. CC36 cells: Global comparison of gene expression changes shared between CC36 shTMED9 and CC14 shTMED9 cells over their respective controls thus allowed us to search for common effects of TMED9 kd beyond WNT signaling (Supplemental Fig. 6). Using FDR < 0.05 and fold change of two yielded 9 repressed and 11 induced candidates. These were then re-tested by rt-qPCR in three independent CC14 shTMED9 and CC36 shTMED9 batches as well as in LS174T shTMED9 cells vs. their respective controls. Four repressed genes were thus identified encoding CNIH4, Phosphatidylinositol glycan anchor biosynthesis class A (PIGA), the integral small membrane protein SMIM13, and the single-pass type I integral membrane protein C11orf24 (Fig. 3a). All four proteins are localized in the secretory network as are the p24 proteins [55,56]. The CORNICHON family displays phylogenetically conserved TGFα export function [57][58][59] and CNIH4 has also been involved in the secretion of G-protein coupled receptors [60] and defined as a cargo adaptor [61]. PIGA is the first enzyme required for the production of the GPI moiety of all GPI-linked membrane proteins [62], which require p24 function for export [63,64]. This conserved four-gene TMED9-dependent signature highlights secretion-transcription regulatory mechanisms and we use it henceforth to track TMED9 activity. Indeed, this signature also responded in opposite ways to kd of TMED3 vs. kd of TMED9 (Fig. 3b), in agreement with the global analyses shown above.

TMED9 pro-metastatic activity is linked to but separate from WNT signaling inhibition
Having determined a conserved TMED9-dependent signature, we addressed whether the actions of TMED9 may be solely effected via regulation of WNT signaling, perhaps in an opposite manner to TMED3. This was first addressed by directly repressing WNT signaling responses cell autonomously through the use of a dominant negative form of TCF4 (dnTCF), both in the absence and in the presence of kd of TMED9. Lentivector transduction of dnTCF alone resulted in the downregulation of AXIN2, a canonical WNT responsive gene, and of WNT11, encoding a WNT ligand, as expected. Moreover, it was also able to rescue their upregulation, and that of WNT3, by shTMED9 (Fig. 3c), indicating that TMED9 and WNT signaling may balance each other. Fig. 3 A conserved TMED9 kd signature and interaction with WNT-TCF signaling. a Heat map of shared repressed genes from deep sequencing (RNA seq) and rt-qPCR (PCR) values in three different colon cancer cells as shown. These four genes constitute a shTMED9 signature. b Heat map of rt-qPCR results testing for the expression of the four-gene TMED9-dependent signature in CC14 shTMED9 and CC14 shTMED3 cells. Note the opposite regulation by these two TMED factors. PCR was used here as the expression values of this signature in the shTMED3 RNA seq data did not pass the FDR cutoff. c Heat map of rt-qPCR normalized ratios for the three noted conditions (each over control) for the selected genes. The rescue of the expression of WNT pathway genes (yellow background), but not those of the TMED9 signature, is afforded by combined TMED9 kd and repression of WNT-TCF signaling by dnTCF4. Kd levels of TMED9 are in gray background. d Quantification of cells crossing the membrane in transfilter experiments with repressed WNT-TCF signaling (dnTCF4) and/or shTMED9 cells as noted. DnTCF4 does not rescue (enhance) the compromised migration of cells with kd of TMED9 Direct blockade of TCF function, however, was unable to rescue the repression of CNIH4, PIGA, SMIM13, or c11orf24 driven by shTMED9 (Fig. 3c), arguing for a separate effect of TMED9.
In functional assays, dnTCF induced a higher number of migrating CC14 cells in transfilter assays, consistent with the anti-metastatic function of WNT-TCF signaling [6]. However, it was unable to rescue the migration deficiency imposed by shTMED9 (Fig. 3d), paralleling the gene expression results above. We interpret these results to indicate that the pro-metastatic activity of TMED9 is not simply due to its blockade of canonical WNT function, arguing for the existence of additional TMED9-dependent pro-metastatic signals.
The TMED9 and TMED3 responsive gene CNIH4 is required for metastasis To validate the significance of the TMED9-regulated signature and to begin to investigate possible pro-metastatic events downstream of TMED9, we chose to test the function of CNIH4, which belongs to a family of evolutionarily conserved TGFα exporters [57,58].
Use of a lentivector-encoded shRNA reducing CNIH4 mRNA levels by 95% resulted in a 60% reduction of transfilter CC14 cell migration (Fig. 4a, b). Similar results were obtained with a second shRNA with kd of 90% (Supplemental Fig. 7a). Kd of CNIH4 did not affect the expression of TMED9 or the other TMED9-dependent signature genes (Fig. 4c). Conversely, enhanced CNIH4 levels using a cDNA in transiently transfected cells rescued the transwell migratory deficiency of shTMED9, although it did not enhance the basal level of control cells (Fig. 4d). Enhanced CNIH4 levels did not affect the expression of the other TMED9 signature genes (not shown). Importantly, and in line with the in vitro data, engrafted CC14 shCNIH4 cells produced tumors with reduced numbers of distant metastases without significantly altering primary tumor size, all as compared with CC14 control cells (Fig. 4e).
These results link CNIH4, a TMED9-and TMED3regulated gene, to the positive control of metastases.

TGFα rescues the migratory deficiency of colon cancer cells with compromised TMED9 function
Given the nature of TMED9 as a secretory cargo selector and the requirement of TMED9 and CNIH4 for metastases shown above, TGFα as well as other ligands previously implicated in tumor progression or metastases-SONIC HEDGEHOG (SHH) [6], FGF1 [65], FGF19 [44], and TRAIL [66]-were tested for their ability to rescue the decreased migratory phenotype of shTMED9 cells.
This was performed using the transfilter system to ascertain their effects specifically on human tumor cells: cells were pre-treated with ligands for 48 h and plated with and without their continued presence. The only tested molecule able to rescue the transfilter migration deficiency of cells with kd of TMED9 under either experimental strategy was TGFα, which also induced the EMT-like disaggregation of CC14 epithelial colonies (Fig. 5a-c, Supplemental Fig. 8, 9). TGFα effects were reproduced in LS174T and CC36 cells, suggesting their widespread effects on human colon cancer cells (Fig. 5d, Supplemental  Fig. 10). TGFα did not rescue the transfilter deficiency of glioblastoma U251 cells with compromised TMED9 function, used here as outlier controls (Supplemental Fig. 10).
The colon cancer cells responding to TGFα ligand were determined by DNA sequencing to harbor oncogenic mutations (KRAS G13D and BRAF wt in CC36; KRAS wt and BRAF V600E in CC14, KRAS G12D , and BRAF wt in LS174T) downstream of the TGFα receptor EGFR. These cells might thus have been suspected a priori to be insensitive to signaling by EGFR ligands given the action of RAS-RAF signaling downstream of TGFα−EGFR function. Interestingly, EGF treatment was unable to mimic the effects of TGFα (Supplemental Fig. 8), suggesting the existence of EGFR ligand selectivity [67] for pro-metastatic function.

Blockade of the TGFα receptor EGFR and kd of TGFA decreases colon cancer cell migration and metastases
To investigate in more detail the mechanisms involved in controlling pro-metastatic states we have focused henceforth on patient-derived, primary CC14 cells as these display a clear epithelial morphology in vitro and in xenografts, mimicking local, and early primary colon cancers [6].
Events downstream of TGFα ligand action were probed by blocking the activation of its receptor, EGFR, with the monoclonal antibody Cetuximab [68]. Blocking EGFR resulted in a 50% reduction in the number of migrating cells in transfilter assays and abrogated the increase afforded by TGFα treatment when cells were incubated with both TGFα and Cetuximab (Fig. 6a). As expected, Cetuximab treatment decreased endogenous events downstream of EGFR in these cells as highlighted by the acute inhibition of p-ERK and AKT levels by 50% and 20%, respectively (Fig. 6b). Blocking EGFR thus yields the expected phenotype in cell migration predicted from the analyses of TGFα gain-offunction described above.
To directly test if endogenous TGFα participates in metastases, we independently used two shRNAs reducing its mRNA levels by 90%. CC14 shTGFA cells engrafted into NSG mice produced smaller tumors and had fewer metastases as compared with CC14 control cells (Fig. 6c, d).
In order to discern a direct effect on the metastatic process rather than as a possible secondary effect owing to reduced tumor size, shTGFA cells were tested in transfilter experiments: CC14 shTGFA produced an 80% reduction in transfilter migration as compared with controls (Fig. 6e, f). A second shRNA with kd of 80% reduced the number of cells crossing the filter by 60% (Supplemental Fig. 7b). These results show that TGFα and its receptor EGFR are required for experimental colon cancer metastases.

Compromised TGFα biogenesis in cells with kd of TMED9
To investigate how TMED9 might control TGFα function, we investigated the localization of the latter in cells with normal or compromised TMED9 function. Epitope tagged TGFα [69] was immunolocalized in permeabilized cells as a smooth layer with fine puncta over the secretory network as expected ( [69]; Fig. 7a). In cells with compromised TMED9 this pattern was commonly replaced by large aggregates (Fig. 7b). Double immunolabeling against co-transfected HA-tagged-TGFα and MYC-tagged-TMED9 revealed partial colocalization in structures close to the nucleus (Fig. 7c upper and lower panels): TGFα localized largely to the Calreticulin + ER and less to the TGN46 + Golgi (Fig. 7d, f). Conversely, in shTMED9 cells TGFα strongly localized within the Golgi and weakly in the ER (Fig. 7e, g). As TGFα was not localized in EEA1 + endosomes or LAMP1 + lysosomes (Supplemental Fig. 11) it appeared retained in the Golgi in cells with kd of TMED9. Coimmunoprecipitation analyses failed to yield positive results, likely due to the small amounts of ligand present. The possibility that TGFα was retained intracelullarly in cells with compromised TMED9 was tested by immunolocalizing membrane-bound tagged TGFα precursors in nonpermeabilized cells using maximal projection of z-stacked confocal images to visualize the entire cell surface. TGFα precursors were normally localized in the membrane of control GFP + cells but their abundance was greatly diminished in cells with kd of TMED9, with the strongest effect yielding little TGFα signal (Fig. 7h-k). Quantification of immunopositive dots reveled a 10-fold decrease in shTMED9 vs. control cells (Fig. 7h, i). As membrane localization of TGFα is required for membrane signaling as well as for the cleavage to produce secreted ligands, its biogenesis and overall function is compromised in cells lacking normal TMED9 function.
To complement the immunofluorescence results, we analyzed overall cell surface TGFα levels using cell surface biotinylation and pull-down of biotinylated proteins, followed by Western blotting in CC14 vectoralone vs. CC14 shTMED9 cells expressing tagged TGFα. Using this method we observed that the 36 kD membrane form of TGFα was only detected in the membrane fraction, as expected, but also that it was reduced twofold in shTMED9 cell surface fraction as compared with control cells. As controls, EGFR, was enriched 2.5-fold in the membrane and the cytosolic protein HSP70 inversely enriched twofold in the non-membrane fraction (Supplemental Fig. S12a).
In addition, enzyme-linked immunosorbent assay (ELISA) immunodetection of secreted TGFα present in the condition media of CC14 vectoralone and CC14 shTMED9 cells showed twofold reduction, on average, in the latter (Supplemental Fig. 12b).
The levels of TGFA and CNIH4 predict disease outcome Taken together with previous data [8], the results presented above indicate that TMED3 and TMED9 act as gatekeepers for protein secretion events that lead to the regulation of TGFα and WNT signaling, establishing antagonistic regulatory modes that control metastatic states within primary tumors. To analyze the possibility that their expression predict disease outcome in patients we have correlated their expression levels with disease-free survival using public cohorts.
TMED9 quintiles showed a trend where top expressors showed poorer outcomes, but these did not reach significance of p = 0.05 (Supplemental Fig. 13), much as reported for TMED3 [8]. These results may reflect the fact that each TMED protein controls the secretion of different cargoes and the activity of several signaling pathways [70], and that TMED proteins are detected at all colon cancer stages (https://www.proteinatlas.org/search/TMED). We therefore analyzed the correlation of the TMED regulated genes TGFA and CNIH4 with disease-free progression. High TGFA [71] or CNIH4 quintiles showed a strong correlation with disease outcome and the combination of both TGFA high , CNIH4 high vs. TGFA low , CNIH4l ow was also positively correlated with poor outcome (Fig. 8a).
TGFα shows a mutually antagonizing relationship with WNT-TCF signaling and rescues the gene signature repressed by kd of TMED9 Having determined the requirement of TGFα, we asked how TGFα and the WNT pathway may affect each other since we show that TMED9 represses WNT11 and other signaling components and responses (Fig. 2) and TCF represses (dnTCF enhances) TMED9.
We found that dnTCF enhanced TGFA levels, suggesting the coordinate repression of TMED9 and TGFA by WNT-TCF. Inversely, TGFα ligand treatment repressed WNT11 (Fig. 8b) and TGFA kd-induced WNT11 2.3-fold. Moreover, TGFα compensated the boost of WNT11 induced by shTMED9 (Fig. 8b). This shows the ability of TGFα to revert a number of TMED9 kd phenotypes (see also Fig. 5) and its ability to repress WNT11. TGFA mRNA levels were also enhanced by TGFα treatment (Fig. 8b), pointing to its auto-induction as in keratinocytes [72].
We then asked if TGFα could antagonize WNT signaling using a TCF luciferase reporter: WNT11 enhanced and TGFα repressed endogenous TCF activity, whereas together they yielded an intermediate phenotype, arguing that they antagonize each other's effects (Fig. 8c). Indeed, such antagonism was also revealed through the exogenous activation of WNT signaling using active N'-mutated βCA-TENIN (βCATENIN*): increasing doses of TGFα cotransfected with βCATENIN* led to the repression of the activating effect of the latter on a TCF Luciferase reporter to levels similar to those detected after the repression by dnTCF (Fig. 8d).
At last, whereas kd of TGFA did not alter the TMED9dependent gene signature (not shown), treatment with TGFα ligand enhanced the expression of CNIH4, PIGA, and TMED9 itself and, notably, it rescued the downregulation of the TMED9-dependent signature by shTMED9 (Fig. 8b). This result, together with the requirement of TGFA shown above (Fig. 6), indicates that TGFα mediates many, but not all, aspects of TMED9 function, that it shows a mutually antagonizing relationship with WNT-TCF signaling, and that TGFα signaling can be self-reinforcing (Fig. 8j).
GLI1 rescues the kd phenotypes of TMED9, CNIH4 and TGFα, establishing a pro-mestastatic regulatory loop Previous work has identified GLI activity in colon cancer as pro-metastatic acting in part to antagonize WNT-TCF signaling [6,23]. As this parallels the roles of TMED9 and TGFα described above, we addressed the possibility that GLI1 could mediate aspects of TMED9-TGFα activity, as we found that both shTMED9 and shCNIH4 decreased GLI1 expression levels (Fig. 8e), and GLI1 was able to enhance the expression of TMED9 by 40% as well as the TMED9-dependent signature, thus mimicking shTMED3 (Fig. 8f).
Functionally, cells expressing lentivector-encoded shRNAs against GLI1 were not viable enough to perform assays (not shown). Nevertheless, we were able to determine that GLI1 rescued the transfilter migratory deficiency imposed by kd of TMED9, TGFA, or CNIH4: Exogenous GLI1 (yielding a 3-5-fold increase in mRNA levels over controls) induced a twofold enhancement in the number of CC14 cells crossing the filter and, importantly, rescued the phenotype of shTMED9, shCNIH4 and shTGFA to control levels (Fig. 8g). Inversely, the positive effect of GLI1 was unopposed by shTMED9 but decreased in cells with kd of CNIH4 or TGFA (Fig. 8g). This indicates that whereas GLI1 rescues all three phenotypes, full GLI1 function requires endogenous CNIH4 and TGFA activities.
The results then prompted us to investigate the possibility of a mutual dependency between GLI1 and TGFA. We found that GLI1 was able to enhance the mRNA levels of TGFA although TGFα did not affect GLI1 expression ( Fig.  8f and not shown). However, TGFα did enhance GLI1 activity threefold, measured with a GLI-binding-site-> luciferase reporter (Fig. 8h).
Taken together the results are consistent with the idea that pro-metastatic actions of TMED9, CNIH4, and TGFα establish a positive regulatory loop with GLI1. Further support for this possibility derived from the finding that 5/6 shTMED9 low ;shTMED3 high pro-metastatic genes (Fig. 2e) tested were GLI1-responsive (Fig. 8i). Globally, the idea that protein secretion-transcription loops operate to determine metastatic proclivities in primary tumor cells is also suggested by the finding that TMED9, CNIH4, as well as TGFA [73], harbor high-confidence consensus GLI-binding sites within 5 Kb upstream of the transcriptional start site, whereas TMED3 harbors TCF-binding sites instead (Supplemental Fig. 14).  Fig. 10 for CC36 data. Concentrations of TGFα (gray triangles) were 10 and 25 ng/ml Discussion A critical step in metastasis is the decision of cells within primary tumors to acquire a pro-metastatic state, which when realized will allow cells to migrate, disperse, and eventually colonize distant sites. How such states are initially established is not understood, in part owing to the difficulty in deciphering context-dependent actions and the complexity of cross-talk among signaling pathways. Here, we provide evidence that a critical mechanism in the positive control of pro-metastatic states, and the resulting distant metastases, in human colon cancer cells is regulated protein secretion, as highlighted by the function of the protein secretion cargo selector TMED9.
Little is known on the function of individual TMED proteins in normal human cells or in cancer. TMED9 (aka p24alpha2 or p25) has been shown to form complexes with other TMED proteins ( [10,12]; https://string-db.org) but its role in metastasis remained largely unexplored [74,75]. In the present study, we show that TMED9 is transcriptionally negatively regulated by TMED3, and that TMED9 has a pro-metastatic function and works in an epistatic manner in relation to TMED3. We find that TMED9 and TMED3 activities balance each other to determine metastatic outcomes and control in opposite manners a global gene cohort that includes multiple factors implicated in the regulation of metastases.
A previous unbiased in vivo RNAi screen revealed TMED3 as a positive modulator of WNT signaling [8] and blocking WNT-TCF signaling with dnTCF has been shown to promote metastases from multiple primary colon cancer cells [6,7]. Consistently, here we show that WNT-TCF signaling represses TMED9 and in turn TMED9 represses WNT-TCF pathway components and responses. However, we show that although the metastatic outcome of repression of WNT signaling requires TMED9, the pro-metastatic function of TMED9 is separate from repression of WNT signaling. Therefore, TMED9 not only exhibits a mutually repressive interaction with TMED3-WNT but must also positively regulate pro-metastatic signals.
Defining TMED9-regulated genes and searching for factors able to rescue TMED9 kd-induced functional deficiencies allow us to uncover TGFα, a high-affinity ligand for EGFR [69,76], as a mediator of pro-metastatic TMED9 function: TGFα ligand, rescues the phenotypes of kd of TMED9 and provokes the dispersion of colon cancer epithelial colonies, as in other cancer cells [77]. Although endogenous cargoes remain to be identified for TMED9, four lines of evidence suggest that TMED9 promotes TGFα activity: (i) We find that cancer cells with compromised TMED9 function have reduced membrane and secreted TGFα levels, indicating that TGFα requires TMED9 for normal biogenesis and thus function; (ii) TMED9 is present in TGFα-containing and in COPII secretory vesicles [78,79]; (iii) TMED proteins associate with GRASP55/65 (GORASP1/2 [80]), which binds precursor and membrane forms of TGFα [81]; (iv) We show that CNIH4, which encodes a member of the CORNICHON family of evolutionarily conserved TGFα exporters, is required for metastases and is regulated by TMED9 activity, and like TMED9, CNIH4 is also found in COPII vesicles [79]. TMED9, however, is likely to regulate signals other than TGFα. Moreover, it remains unclear if CNIH4 directly regulates TGFα secretion given that cells with overexpressed CNIH4 via lentivector transduction were not viable (not shown) to perform ligand localization analyses, and CNIH4 has been also shown to export G-protein coupled receptors [60].
The results presented above, the presence of secretiontranscription regulatory loops suggested by the control of the expression of CNIH4 and TGFA by TMED9 and TMED3, the positive regulation of GLI1 by TMED9 and CNIH4, the requirement of normal TGFα activity, as well as by the ability of GLI1 to rescue the kd phenotypes of TMED9, CNIH4, and TGFA, together with previous findings on the roles of WNT and GLI signaling in colon cancer metastases [6][7][8], allow us to propose the existence of two antagonistic, self-sustaining loops that regulate the prometastatic states of colon cancer cells (Fig. 8j).
1-A basal pro-tumorigenic but anti-metastatic state is afforded by fully activated WNT signaling, usually via APC mutation, as well as by TMED3-mediated secretion of WNT ligands, as these are required for full WNT-TCF pathway activation even in the presence of pathwayactivating mutations [82]. We posit that full WNT-TCF activity is self-reinforcing and expands the tumor. Moreover, it keeps tumor growth local through repression of prometastatic TMED9-TGFα signaling via its downregulation of TMED9, CNIH4, TGFA, and GLI1.
2-The self-sustaining local, anti-metastatic WNT-TCF loop in colon cancer cells is proposed to rebalance in favor of pro-metastatic TMED9, CNIH4, TGFα, GLI signaling during the metastatic transition, which involves WNT signaling repression (e.g., WNT ligand and TMED3 downregulation) and the enhancement of the expression levels and function of TMED9, CNIH4, TGFA, and GLI1 (Fig. 8j). This may happen in single or small numbers of cells and is consistent with both the wide distribution of βCATENIN, with high levels commonly due to APC mutation, (https://www.proteinatlas.org/ENSG00000168036-CTNNB1/pathology/tissue/colorectal+cancer#img) and the heterogenous expression of high levels of TGFα Analysis of the phosphorylation of AKT1 and ERK1/2 by Western blotting in sibling cells to those used in a 5' after treatments. Total AKT1, total ERK1/2 and GAPDH were used as controls. c, d Representative images c and quantification d of lacZ + lung metastases in mice carrying CC14 lacZ/control or CC14 lacZ/shTGFA xenografts. The number of metastases is compared with the size of xenografts in d.  How the stable TMED3-WNT-TCF autocrine loop is derailed is unclear. This might involve the upregulation of WNT inhibitors (e.g. DKK1 and SFRP1) as observed in metastatic vs. non-metastatic colon cancer patients [6], but also oncogene-mediated increases in GLI function [24,83] and cell-intrinsic metastatic reprogramming [84]. Enhanced and self-sustaining TMED9-gated ligand signaling in turn is predicted to further downregulate the TMED3-gated WNT-TCF loop and promote pro-metastatic states.
In this model, GLI1 can act both upstream and downstream of TMED9, CNIH4, and TGFα. Whereas direct GLI activation by oncogenic signals may help break the WNT-TMED3 loop (see above), its downstream regulation by TMED9-TGFα signaling is consistent with the ability of peptide growth factors, as well as oncogenic downstream mediators such as RAS, MEK, and AKT (reviewed in [83,85], to boost positive GLI function and enhance metastases [6,23]. Moreover, TMED3/TMED9 coregulated genes include genes shown to be GLI-responsive such as CLDN2, CXCR4, and FGF19 [73], although whether the combined action of TGFα and GLI1 can drive a gene set distinct from that driven by GLI1 alone, as shown for EGF [73], remains to be determined. Enhanced GLI1 activity can thus establish a regulatory loop with TGFα (this work) and decrease WNT-TCF activity [6].
An additional mechanism to reinforce the balance in favor of TMED9-TGFα-GLI signaling may rely on NAKED2 (NKD2), as post-Golgi TGFα (after TMED9 and CNIH4 function) requires NKD2 for vesicular export [86,87]. Mammalian NKD2 also targets the essential WNT pathway component Disheveled for degradation thus further inhibiting WNT signaling [88].
We note that in normal development and regeneration, these signaling pathways and components are very sensitive to dosage, working as morphogens. Small changes in the activities of WNT and TGFα in cancer cells might therefore have important metastatic outcomes. In this sense, the antagonism of the TMED9-TGFα and TMED3-WNT loops may not act solely cell autonomously, with the final balance and outcome being influenced by both signaling within tumor cells and from the surrounding stroma: the cancer cell-intrinsic and the stroma-to-cancer cell pro-metastatic functions of TGFα may therefore coexist and cooperate ( [89,90]; this work).
The balancing act of opposing TMED3-WNT-TCF and TMED9-TGFα signaling loops in the determination of prometastatic fates may be akin to those found in development and regeneration. WNT and TGFα participate in the architecture of normal and regenerating crypt-villus axes of the intestine and are expressed at opposite ends: canonical WNT ligands are made in the bottom part of the crypt and promote crypt stem cell self-renewal [91]. In contrast, TGFα activity is enriched in the villus and promotes cellular differentiation and migration towards the tip, also affecting cells in the crypt [92][93][94][95]. Epithelial cells in villi loose responsiveness to TGFα upon differentiation and are eventually shed into the lumen [96]. Metastatic cells may re-differentiate and repress TGFα signaling, regaining an epithelial morphology and reentering a local mode of growth to establish metastatic colonies. In metastatic colon cancer, such a reversal might be accomplished by re-establishing the primacy of WNT signaling for local expansion through non-cell autonomous niche-mediated mechanisms [97].
More generally, we suggest that TMED protein secretion cargo selectors may have context-dependent roles in the regulation of migratory/invasive/metastatic behaviors in different cancer types. For example, TMED3 can promote antimetastatic WNT signaling in colon cancer cells [8], whereas it has been reported to promote IL11 signaling and tumor progression in liver cancer cells [70]. Thus, whereas TMED9 levels per se are not significantly correlated with disease-free survival, the levels of mediators and markers we identify -CNIH4 and TGFα-predict colon cancer outcome. We propose that antagonistic secretion-transcription loops gated by TMED9 and TMED3 represent a key modulatory network for the control of the number of colon cancer metastatic cells.

Tumor engraftment and distant metastases
Colon cancer cells expressing the lacZ tracer were transduced with additional lentivectors at a MOI of 2-3. Cells infected with lentivectors carrying GFP tracers, were sorted by fluorescence-activated cell sorting (BD biosciences ARIA III) 72 h afterwards and replated. 48 h after sorting, cells were trypsinized, counted. For pLKO.1 lentivectors, 48 h after infection puromycin (5 µg/ml) selection was performed and cells were plated for experiment after further 48 h. In total, 5 × 10 5 cells were re-suspended in 100 µl of Hanks' Balanced Salt Solution (HBSS) solution for subcutaneous injections per site into the flanks of 8-12 week immunocompromised NSG (NOD-scid IL2rg null ) mice purchased from Charles River. Xenografts of shCNIH4 cells and controls were performed in NUDE mice purchased from Janvier Laboratories. Mice were allocated to control or experimental samples randomly from the cage but not blindly. All mice were female and were cared for and kept at the SPF University of Geneva's Animal Facility. Sample sizes and power were determined with G*Power3. Xenografted subcutaneous tumor growth was monitored by measuring its volume with a caliper. Mice were sacrificed before tumors reached legal limits (according to Swiss (Office Cantonal de Vétérinaire de Genève) animal care standards) and xenografts and lungs harvested following approved protocols (GE/77/17). Rare (< 1 in 30) outliers with tumor sizes above or below two standard deviations from the average were excluded from the study. Lungs were fixed in fresh 4% paraformaldehyde (PFA) for 4 h and stained with X-Gal for 6 h to identify β-Galactosidase + cells. Stained lungs were washed with phosphate-buffered saline (PBS) 4-5 times and positive cells/colonies were counted under a dissecting microscope as described in [24]. LacZ + cells were counted on both sides of all lung lobes per mouse.
Transfilter assays CC14 lacZ , CC36 lacZ , or LS174T lacZ cells transduced with different lentivectors were plated as adherent cultures to reach 70-75% confluence the next day. Cells were washed twice with HBSS and replenished with 0.5% FBS containing medium (serum-deprived medium) for 16 hours. They were then trypsinized to obtain single cells. 2 × 10 5 cells were re-suspended in 100 µl of serum-deprived medium and placed on the upper chamber of prehydrated transwells (6.5 mm diameter, 8.0 µm pore size, Corning #3422). In total, 650 µl of medium supplemented with 10%FBS was placed in the lower chamber of the transwell insert to serve as attractant. CC14 lacZ and LS174T lacZ cells were incubated for 72 h, whereas CC36 lacZ needed only 24 h. Subsequently, the contents of the upper chambers were carefully aspirated and the membranes washed with PBS, fixed for 5 min in 4% PFA, re-washed with PBS thrice and stained with X-Gal. The upper part of the membrane was subsequently cleaned gently with a cotton bud to remove any remaining cells on the surface of the filter. X-Gal stained cells were counted and photographed with an inverted optical microscope. All assays were performed at least twice in triplicates. In the case of GLI1 transfilter assay, DNA transfection with 2 µg of GLI1 plasmid was performed with 10 6 cells in P60 dishes using Lipofectamine LTX or with Amaxa Nucleofection reagent. After 48 h, the cells were plated for transfilter assays as described above. All assays were performed at least twice in triplicates.
CC14 cells were treated with 2 µg/ml Cetuximab (Biovision, #A1047), TGFα (10 ng/ml) or with both at the same time for 72 h and plated for transfilter assays. For western blot analyses, cells were plated at the density of 2 × 10 5 cells in 10 cm plates and treated the day after with Cetuximab for 72 h in 0.1% serum condition medium followed by 5 min treatment with TGFα at 10 ng/ml.

RNA sequencing and transciptome analyses
Oligo dT-selected mRNAs isolated from three independent experiments for CC14 vectoralone vs. CC14 shTMED9 and for CC36 vectoralone vs. CC36 shTMED9 cells were subjected to deep sequencing on the Illumina platform of the University of Geneva's Genomic Facility. The sequencing was performed with 50 or 100 bp reads and~30 million of reads per sample. Differentially expressed genes were identified using edgeR [101] software and the P values adjusted with a 1% false detection rate (FDR) using the Benjamini-Hochberg correction unless otherwise noted. Analyses of regulated genes were performed with Enrichr (http://amp.pharm. mssm.edu/Enrichr/), GSEA (http://software.broadinstitute. org/gsea/index.jsp) and GOrilla (http://cbl-gorilla.cs. technion.ac.il/) for enrichment analyses. RNA sequencing data are deposited to GEO (accession number GSE125282).

Cell surface protein isolation
In total, 2 × 10 6 of each CC14 control and shTMED9 cells were plated in p100 dishes and transfected 2 days after with 24 µg of HA-tagged TGFA plasmid. After 24 hours of transfection, cell surface protein isolation was performed using a cell surface protein isolation kit (ab206998 Abcam) following the manufacturer's instructions. In brief, cells were first labeled with Sulfo-NHS-SS-Biotin, cells were then lysed and the labeled cell surface proteins isolated using streptavidin beads. Cell surface proteins were finally eluted with 0.1 M DTT and analyzed by western blotting with the antibodies of interest.
TGFα ELISA assay of CC14-conditioned medium For conditioned media preparation: 1 × 10 6 of each CC14 control and shTMED9 cells were seeded in p60 dishes with complete medium. The day after cells were washed twice with PBS and once with serum-free medium, and covered with 5 ml of serum-free medium. After 48 hours of incubation media were collected, centrifuged and filtered. This conditioned media was immediately used to measure the concentration of TGFα using the Human TGFα ELISA kit (MBS700091 MyBioSource). The optical densities of standards and of our duplicate samples were measured at 450 nm using a Victor3 1420 plate reader.

Immunofluorescence
Cells were seeded over glass coverslip in 24-wells plates and plasmid transfected the day after with Lipofectamine-2000. The following day cells were fixed with fresh 4% PFA pH8.0, permeabilized and blocked with PBS containing 0.1% Triton X-100 and 10% heat-inactivated normal goat serum (PBSTH). Primary antibodies were incubated overnight at 4°C at dilutions recommended by the providers in PBSTH. After washing in PBS, secondary fluorescently tagged antibodies were incubated for 1 h at room temperature in PBSTH. For non-permeabilized immunofluorescence with anti-HA antibodies the protocol of Briley et al. [100] was followed with the exception of the replacement of BSA for 10% heat-inactivated goat serum. The efficiencies of transfection and the total numbers of HApositive cells in each experiment were determined per condition. Images were taken with an LSM700 confocal microscope (0.4 µm slices) at the University of Geneva's Imaging Facility and processed using FIJI software. Primary antibodies used were anti-: HA- Tag

Data mining for Kaplan-Meier disease outcome plots and binding site predictions
Mining of data were performed using the GSE17538 set containing 238 tumor samples. The clinical follow-up including disease-free survival and event of death for all 238 patients are available in www.ncbi.nlm.nih.gov/geo/ query/acc.cgi?acc=GSE17538 [103,104]. The probes were as mentioned in [84]. or as follows; TMED9-208757_at, TMED2-214658_at, TMED7-214658_at, TMED1-203679_at, TGFA-205016_at, CNIH4-228437_at. p values were calculated and survival Kaplan-Meier plots were made with PRISM GRAPHPAD program. For multiple gene analyses, 50% cohorts were independently used for each probe and common samples (e.g., high or low for the genes tested) kept in the multiple cohort.
For each gene, fasta sequence of 5 kb upstream from the transcription start site and 5′-UTR region was obtained from UCSC table browser on the hg19 human genome. The fasta sequences thus obtained were searched for the presence of binding motifs using the program FIMO [105]. FIMO [v4. 10.0] was run with default parameters. The position weight matrices (PWMs) for the transcription factors were obtained from the JASPAR database [106]. The following PWMs were searched for binding sites GLI: M01042_GLI1, MA0734.1_GLI2; TCF: MA0523.1_TCF7L2, MA0830.1_TCF4 Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons. org/licenses/by/4.0/.