Loss of Furin in β-Cells Induces an mTORC1-ATF4 Anabolic Pathway That Leads to β-Cell Dysfunction

Many protein hormones are synthesized as large inactive precursors that require proteolytic cleavage to be activated. Proprotein convertases (PCs) are subtilisin-like serine proteases that cleave a variety of precursor proteins in the secretory pathway, such as hormones, neuropeptides, extracellular matrix proteins, membrane receptors, and their ligands (1). Any alterations of this regulated process can result in pathological conditions. Typically, when such alterations include prohormones involved in the regulation of glucose and lipid homeostasis, they can result in metabolic disorders (2,3). For instance, PC1/3 and PC2 are both highly expressed in the secretory granules of neuronal and endocrine cells and process a number of hormones and neuropeptides involved in the regulation of glucose and energy homeostasis, such as insulin, glucagon, and proopiomelanocortin (2,4). In humans, mutations causing complete loss of PC1/3 enzymatic activity lead to hyperphagic obesity and impaired glucose homeostasis, among other endocrinopathies (4). In addition, common heterozygous PC1/3 mutations that cause partial loss of function contribute to variations in human BMI and plasma proinsulin and are associated with impaired regulation of plasma glucose levels (5). Genome-wide association studies have found associations between several common variants in the gene encoding human PC2 and diabetes risk and related traits in human populations (6,7).

In contrast, the role of PCs active in the constitutive secretory pathway, such as FURIN, PC5/6, PACE4, or PC7, in obesity and type 2 diabetes (T2D) has been much less explored. In the secretory pathway, FURIN is active in the trans-Golgi network and endosomes and at the cell membrane. In neuroendocrine cells, FURIN has also been found in immature secretory granules, where it has been proposed to participate in the processing of certain neuropeptides involved in energy metabolism, such as pro–brain-derived neurotrophic factor, pro–growth hormone–releasing hormone, and prokisspeptin (8–10). The FURIN gene is located near the PRC1 T2D association interval, and therefore, it has been suggested as a promising biological T2D candidate (11). In addition, polymorphisms in FURIN have been shown to be associated with the prevalence of metabolic syndrome (12). We previously showed that FURIN is crucial for acidification of secretory granules in mouse pancreatic β-cells via the cleavage of Ac45, an accessory subunit of the vacuolar-type ATPase (V-ATPase) proton pump (13). The V-ATPase complex acidifies intracellular organelles by pumping protons across membranes, and acidification is a critical step in β-cell granule maturation and proinsulin-to-insulin conversion by PC1/3 and PC2 (14). However, the significance of protein processing by FURIN in β-cell function is incompletely understood.

Recently, we discovered a major artifact in several transgenic mouse lines used in the diabetes field, negatively influencing normal β-cell function (15,16). More specifically, we found that human growth hormone (hGH), positioned downstream of the Cre transgene to ensure proper expression, was unexpectedly expressed into a bioactive protein. hGH activates prolactin receptors on β-cells, activating a pregnancy-like phenotypic switch. Moreover, changes unrelated to pregnancy, such as impaired glucose-stimulated insulin secretion (GSIS), were also observed. This complicates the interpretation of many studies in which these models were used, including our previous FURIN study (13). To unequivocally assess the in vivo role of FURIN in β-cells, we generated a conditional knockout mouse model lacking the hGH minigene. Metabolic studies were performed to characterize the function of FURIN in pancreatic β-cells in vivo. Genomic and proteomic studies and generation of a Furin knockout β-cell line allowed us to determine the pathways affected by the absence of FURIN and identify physiological FURIN substrates influencing β-cell function.

Human islets from healthy (HbA_1c <6.0%) and T2D donors (HbA_1c >6.5% at diagnosis) were obtained from the Alberta Diabetes Institute Islet Core (University of Alberta). Supplementary Table 8 contains general information about the donors, and additional information is available at https://www.epicore.ualberta.ca/IsletCore/. Islets were cultured overnight until they were handpicked and snap frozen. Protocols were approved by the human ethics committee of the Institut de recherches cliniques de Montréal.

For more details, including cell culture, animal procedures, electron microscopy, and stable isotope labeling of amino acids in cell culture (SILAC) methods, please see Supplementary Material. RIP-Cre^+/− mice were kindly donated by Dr. Pedro L. Herrera (University of Geneva Medical School). Furin^flox/flox mice have been previously described (17). Mice were backcrossed at least eight times to a C57Bl6J background (Janvier). Only male mice were used for experiments throughout the study. All experiments were approved by the Katholieke Universiteit (KU) Leuven Animal Welfare Committee, following the guidelines provided in the Declaration of Helsinki (KU Leuven project number 036/2015).

Microarray analysis was performed as described previously (15). Briefly, islets were isolated from male flox and β-cell–specific Furin knockout (βFurKO) mice (12 weeks of age) by collagenase injection in the pancreatic duct. Islet RNA was isolated using the Absolutely RNA Microprep Kit (Stratagene) according to the manufacturer’s protocol. RNA quantity and quality were determined using the Bioanalyser (2100; Agilent). Microarray analysis was performed on MoGene_1.0_ST arrays (Affymetrix). One hundred nanograms of total islet RNA was hybridized to the arrays according to the manufacturer’s manual 701880Rev4. Samples were analyzed pairwise, and P < 0.01 and fold change ≥1.5 were set as selection criteria. Functional enrichment, canonical pathway enrichment, and upstream regulator analysis were performed using ingenuity pathway analysis (IPA; Qiagen).

The mouse insulinoma cell line βTC3 was cultured in DMEM:F12 (1:1) supplemented with 10% heat-inactivated FBS, 100 units/mL penicillin, and 100 μg/mL streptomycin. A Furin knockout βTC3 cell line was generated using the CRISPR-Cas9 nuclease system, following the protocol as described before (18). Briefly, three 20-nucleotide guide sequences targeting fur exon 2 were designed using the online CRISPR Design Tool. The pSpCas9(BB)-2A-Puro construct (48139; Addgene) was used to clone the guide sequences. Optimal transfection and puromycin selection conditions for βTC3 cells were established in advance using a GFP-containing expression construct. To screen guide efficacy, transfected cells were harvested for DNA extraction using the QuickExtract solution, and indel mutations were detected by the SURVEYOR nuclease assay as described. Clonal cells were obtained by serial dilution in 96-well plates (Corning). Genomic microdeletions and insertions were verified by TOPO cloning and subsequent Sanger sequencing. Furin deficiency in βTC3 cells was corroborated by Western blot using an anti-FURIN antibody. Control cells received the same treatment and were grown in parallel with the knockout lines. Sequences for guide cloning in Addgene vector 48139 and SURVEYOR primer sequences are shown in Supplementary Table 9.

The microarray data were deposited in the Gene Expression Omnibus (GSE150312).

To determine the potential importance of FURIN in human islets, we first quantified mRNA expression levels of PCSK3 (encoding FURIN) and other PCs in islets from healthy and T2D donors (Fig. 1A). Genes encoding PC1/3 and PC2 (PCSK1 and PCSK2), which are responsible for the processing of proinsulin, proglucagon, and prosomatostatin, were, as expected, expressed highest. PCSK3 showed the highest expression of PCs mainly active in the constitutive secretory and endosomal pathways (PC5/6, PACE4, and PC7 encoded by PCSK5, PCSK6, and PCSK7, respectively), consistent with previously published RNA sequencing data sets from human islets and islet subsets (19,20) (Supplementary Fig. 1A and B). Expression levels of PCs that could potentially provide redundancy for FURIN activity were much lower, which could point to a nonredundant function of FURIN in human β-cells. Except for slight increases in PCSK5 and PCSK6 expression, we did not observe significant changes in PC expression in islets from T2D donors. INS expression (encoding INSULIN) was significantly reduced in patients with T2D in this cohort (Fig. 1B). At a protein level, FURIN was expressed at high levels in human β-cells and non–β-cells (Fig. 1C). We did not detect differences in control participants versus patients with T2D, although the study was too underpowered to make any substantial claims (n = 1 per group).

βFurKO was achieved by intercrossing RIP-Cre^+/− mice (Tg [Ins2-cre]23Herr) (21) lacking the hGH expression enhancer (15,16) with Furin-floxed mice (17). As expected, islet Furin mRNA (Fig. 2B) and FURIN protein (Fig. 2C) expression were strongly reduced. Residual levels were likely caused by expression in islet α- and δ-cells and blood vessel endothelial cells, where the Cre transgene is not expressed (21). There was no compensatory increase in expression of PCs that might provide redundancy (Supplementary Fig. 2A). Blood glucose levels were mildly but significantly elevated in male βFurKO mice at 10 weeks of age in fed and 6-h fasted states (Supplementary Fig. 2B). At this age, male βFurKO mice were significantly glucose intolerant compared with control mice, which worsened as the mice aged (Fig. 2D–F). Insulin tolerance tests showed normal peripheral insulin sensitivity (Fig. 2G–I), and instead, the mice showed severely affected GSIS (Fig. 2J). Interestingly, no significant differences in glucose tolerance were observed between βFurKO and control females at 10 weeks of age, indicating that FURIN is not critical for glucose homeostasis in females at a young age (Supplementary Fig. 2C). Because glucose homeostasis was unaltered in male flox, Cre, and heterozygous mice, nonspecific transgene-related effects could be ruled out. Therefore, flox controls were used throughout the rest of the study. It should be noted that Cre expression in this mouse model was not restricted to β-cells but was also found in orexigenic RIP-expressing neurons in the hypothalamus (22). However, body weight was not significantly different between βFurKO and control mice on a normal chow diet (Supplementary Fig. 2D), suggesting that lack of Furin does not have marked effects on body weight regulation by these neurons.

Islets from βFurKO mice appeared smaller than control islets during islet isolations. Indeed, β-cell mass and proliferation were significantly reduced in βFurKO mice at 24 weeks of age (Fig. 3A and B), although islet apoptosis was not different between groups (Fig. 3C). Insulin immunoreactivity and pancreatic and islet insulin content were reduced (Fig. 3D–F). Ultrastructural analyses showed that Furin-deficient β-cells were smaller than controls (Fig. 3G and H). Despite similar numbers of secretory granules per cell area (Fig. 3I), granules in βFurKO islets were smaller, showing reduced halo and core areas (Fig. 3J). Furthermore, β-cells displayed an increase in the number of immature secretory granules (electron-light core and distinctive halo) (Fig. 3K). These results suggest that FURIN is essential for β-cell homeostasis and secretory granule biogenesis. Ex vivo islet GSIS was unchanged when secreted insulin was adjusted for total islet insulin content (Fig. 3L), and the ratio of secreted proinsulin to secreted insulin did not differ between genotypes (Fig. 3M), suggesting that there were no defects in insulin secretion or proinsulin processing in βFurKO islets. Instead, reduced functional β-cell mass was likely to be the main driver of the phenotype.

To explain the perturbations observed in this model, microarray analysis was performed on βFurKO islets. Although this method did not directly identify FURIN substrates, the notion of differential gene expression might have elucidated affected cellular processes, potentially leading to upstream FURIN substrates. We found 126 differentially expressed genes in βFurKO islets (cutoff P < 0.01 and |fold change| ≥1.5) (Fig. 4A and Supplementary Table 1). Several mRNAs encoding amino acid (AA) transporters, enzymes involved in AA metabolism, and aminoacyl-tRNA synthetases were significantly upregulated (Fig. 4B and Supplementary Table 1). IPA pointed to significant activation of AA metabolism (Supplementary Table 2). The upstream regulator tool predicted activation of activating transcription factor 4 (ATF4) (P = 2.86E–30) (Fig. 4B and Supplementary Table 3), a key transcription factor involved in the integrated stress response (23). Consistently, many stress-induced genes were also upregulated (Fig. 4A and B and Supplementary Table 1).

Using SILAC, we found 203 differentially expressed proteins in βFurKO islets (Fig. 4C and D and Supplementary Table 4) (negative log P value cutoff = 2.13682; P = 0.0072). Consistent with the microarray data, IPA showed significant changes in AA metabolism (Supplementary Table 5), and many AA synthesis enzymes showed an increase in expression (Fig. 4C and D). In addition, canonical pathway analysis predicted activation of tRNA charging (Supplementary Table 6), with nine aminoacyl-tRNA synthetase proteins being significantly upregulated (Fig. 4C and D and Supplementary Table 4). Upstream analysis predicted activation of ATF4 (P = 3.59E–11) and PKR-like endoplasmic reticulum (ER) kinase (PERK; EIF2AK3; P = 3.21E–15) (Fig. 4C and D and Supplementary Table 7). PERK is an ER-proximal sensor of unfolded proteins that induces ATF4 translation via phosphorylation of Ser51 on eukaryotic initiation factor 2α (eIF2α) (23). These data suggest that β-cells lacking Furin induce an anabolic program involving upregulation of AA transporters and aminoacyl-tRNA synthetases, likely through activation of the ATF4 transcription factor. Consistent with our data shown in Fig. 3F, we observed a significant decrease in insulin I by means of SILAC (∼30% decrease; P = 0.01) (Supplementary Table 4). Supplementary Fig. 3 shows a good correlation (r = 0.4186; P < 0.0001) of microarray and SILAC data for all proteins with at least a 25% increase or decrease in expression.

Several kinds of stress conditions increase eIF2α-dependent ATF4 translation, with subsequent upregulation of ATF4 target genes (23). Accordingly, we observed significantly increased islet expression of the ATF4 target gene Trib3 (Fig. 5A), whereas Atf4 mRNA levels were unaltered. At the protein level, both ATF4 and CHOP (another ATF4 target gene) were increased in βFurKO islets, confirming activation of ATF4 (Fig. 5B). Unexpectedly, βFurKO islets did not exhibit increased phosphorylation of PERK or eIF2α (Fig. 5B), nor were other ER stress markers increased (Supplementary Fig. 4A and B). Emerging evidence suggests that ATF4 translation can also be induced by mammalian target of rapamycin complex 1 (mTORC1) activation (24,25) (Fig. 5C). In agreement with this idea, phosphorylation levels of the mTORC1 substrates p70 ribosomal protein S6 kinase (S6K) and eukaryotic translation initiation factor 4E-binding protein 1 were augmented in βFurKO islets (Fig. 5D). This suggests an activated mTORC1-ATF4 axis in islets from βFurKO mice.

To substantiate our in vivo findings in a cellular model, we generated a Furin knockout β-cell line using the CRISPR-Cas9 system (Supplementary Fig. 5A and B). WST-1 assays demonstrated that cell proliferation and/or viability was significantly reduced in Furin knockout βTC3 cells (Fig. 5E). In addition, Furin knockout βTC3 cells showed increased Trib3 mRNA (Fig. 5F) and CHOP and ATF4 protein levels (Fig. 5G), suggesting ATF4 activation. Like βFurKO islets, we did not observe increased phosphorylation of eIF2α (Fig. 5G), implying that ATF4 upregulation is not mediated by eIF2α. Additionally, Furin knockout cells displayed higher mTORC1 activity, as demonstrated by the increased phosphorylation levels of S6K (Fig. 5G). Treatment with the mTORC1 inhibitor rapamycin reduced ATF4 levels in this cell line, substantiating the idea that mTORC1 activation causes ATF4 upregulation (Fig. 5H). We conclude that the Furin knockout βTC3 cell line can be used as a representative model to investigate the molecular mechanisms caused by Furin deficiency in β-cells in vivo.

We next sought to determine the mechanism driving mTORC1 hyperactivation in βFurKO cells. We focused on the V-ATPase proton pump, because it serves as part of a docking and activation site for mTORC1 on lysosomes (26). Two critical V-ATPase subunits, ATP6AP1/Ac45 and ATP6AP2/prorenin receptor (PRR), are known FURIN substrates (13,27) and are highly expressed in β-cells in mice and humans (13,20,28). The FURIN cleavage sites of the corresponding precursors were shown to be at the RVAR²³¹ and RKSR²⁷⁸ sites, respectively (13,27) (Fig. 6A), whereas PRR can also be cleaved by SKI-1 at an RTIL²⁸¹ site (29). As such, we hypothesized that impaired Ac45 and/or PRR cleavage would lead to disturbed lysosomal acidification, affecting mTORC1 activity in β-cells. We first investigated the processing of Ac45 and PRR in βFurKO cells. Mature Ac45 is heavily glycosylated and appears as a broad smear at ∼60 kDa, but the cleaved form can be detected as a 24-kDa peptide when treated with N-glycosidase F (13) (Fig. 6B). We observed a complete lack of the 24-kDa form, indicating that Ac45 cleavage does not occur in βFurKO cells (Fig. 6B). In addition, whereas the intracellular amount of cleaved PRR did not seem to be affected, the soluble secreted form was significantly reduced in the medium of βFurKO cells (Fig. 6C). To measure V-ATPase function indirectly, we performed ImmunoGold electron microscopy on islets that were incubated with the acidotrophic agent 3-(2,4-dinitroanilino)-3′amino-N-methyldipropylamine (13) and quantified the acidification of lysosomes. We observed a decrease of ∼57% in the number of gold particles per lysosome, demonstrating impaired acidification of lysosomes in β-cells from βFurKO animals (Fig. 6D and E). Lysosomal size was not different between genotypes (Fig. 6F). Disrupted lysosomal acidification has been shown to induce mTORC1 signaling in osteoclasts (30). To test this in β-cells, wild-type βTC3 cells were treated with the V-ATPase inhibitor bafilomycin A1. This induced the mTORC1-ATF4 axis, confirming that V-ATPase dysfunction can cause the phenotype (Fig. 6G). Knockdown of Atp6ap1/Ac45 showed a significant upregulation of ATF4 but no activation of mTORC1 signaling (Fig. 6H). No indications for activation of the mTORC1 pathway or ATF4 were observed after knockdown of Atp6ap2/Prr (Fig. 6I). These results likely reflect the compound nature of the phenotype, whereby loss of FURIN results in impaired cleavage for several substrates, leading to the phenotype. Alternatively, the uncleaved proprotein might lead to additional effects that a genetic knockdown cannot replicate.

In this study, we demonstrated that FURIN is essential for β-cell function and that the dysregulation of its activity can affect β-cells through the induction of the stress factor ATF4 in an mTORC1-dependent manner (Fig. 7). The most striking phenotype in βFurKO mice is a strong reduction in functional β-cell mass, which might be below the threshold required to maintain adequate glucose homeostasis and could directly lead to impaired pulsatile insulin secretion (31). Consistent with our previous study (13), we did not observe impairment of total proinsulin processing in this mouse model. In that study, we showed that although maturation of the less abundant proinsulin II was impaired, labeling of total insulin did not reveal differences. Because the major proinsulin I processing enzyme PCSK1 has a relatively broad pH optimum, the impairment of acidification in the βFurKO model is not sufficient to block insulin processing. Moreover, this was also apparent when the ultrastructure of CPE^fat/fat (32) and Pcsk1^−/− β-cells (33) were compared with our model (Fig. 3G). In the former two, virtually all granules contain electron-lucent (i.e., grayish) cores, which was not the case in β-cells lacking Furin. These data suggest that although total insulin content is strongly decreased in βFurKO cells, processing is largely unaffected.

Whole-genome expression profiling, proteomic analyses, and in vitro studies pointed toward activation of the transcription factor ATF4 by mTORC1. Consistent with our data, recent studies in other cell types have shown that ATF4 can be activated by mTORC1 independently of its canonical induction via phosphorylated eIF2α. For instance, ATF4 stimulates the de novo purine synthesis pathway and transcriptionally regulates AA transporters, metabolic enzymes, and aminoacyl-tRNA synthetases in an mTORC1-dependent manner (24,25). Despite increasing evidence in various other cell lines, the importance of this mTORC1-ATF4 axis in β-cells has not been explored.

mTORC1 is critical in the regulation of β-cell survival and proliferation (34). However, chronic hyperactivation of mTORC1 promotes progressive hyperglycemia and hypoinsulinemia accompanied by a reduction in β-cell mass (35), which is likely caused by ER stress and impaired autophagic response, leading to β-cell failure (36). Islets isolated from patients with T2D exhibit increased mTORC1 activation, and genetic or chemical inhibition of mTORC1-S6K signaling in these islets restores insulin secretion (37), suggesting that hyperactive mTORC1 impairs human β-cell function. Therefore, it is highly likely that persistent mTORC1 activity leads to β-cell dysfunction in βFurKO mice, which could serve as a new genetic model to study mTORC1 hyperactivity in β-cells.

Loss of FURIN activity in β-cells potentially affects cleavage of several substrates that contribute to the observed phenotype to different extents. We investigated candidate substrates based on a potential link with the mTORC1-ATF4 axis and focused on the V-ATPase accessory subunits Ac45 and PRR. FURIN is known to be critical for V-ATPase function in mice (13) and yeast (38), and interestingly, cryo-EM models of the rat (39) and bovine (40) V-ATPase complex revealed that Ac45 and PRR are present as cleaved forms. Here, we show almost complete lack of Ac45 cleavage and reduced PRR shedding in Furin-deficient β-cells. Previous work in Xenopus intermediate pituitary melanotrope cells has shown that deletion of the endoproteolytic processing site in Ac45 causes it to accumulate in the ER (41). As such, disrupted trafficking of Ac45 and potentially V-ATPase subunits to the secretory pathway might cause disrupted lysosomal acidification and mTORC1 hyperactivity in βFurKO cells. In this context, osteoclasts in which V-ATPase function is disrupted show increased lysosomal pH, which in turn leads to elevated mTORC1 activity in this cell type (30). Our results suggest that in β-cells, mTORC1 activity is dependent on lysosomal pH as well. Alternatively, uncleaved Ac45 might modify β-cell mTORC1 activity by relaying AA sufficiency from the lysosomal lumen or alter interactions between V-ATPase and mTORC1 components, as was hypothesized for patients with follicular lymphoma carrying ATP6AP1 mutations (42). Similar to Ac45, PRR is highly expressed in murine and human islets, and its expression is reduced in islets from donors with diabetes (28). Interestingly, knockdown of Atp6ap2/Prr in a β-cell line resulted in an increase in the percentage of ISGs and reduced granular acidification (28), similar to β-cells lacking Furin, as seen in the current study and that by Louagie et al. (13). Mice lacking PRR in β-cells exhibit accumulation of large vacuoles that consume insulin content (43). Although we did not observe these structures, it does not rule out a contribution of PRR to the phenotype in βFurKO mice, which might exhibit a milder phenotype as a result of reduced PRR shedding. The exact timing and mechanism of PRR cleavage and shedding are still matters of debate and might be cell type dependent (27,29). We did not observe significant upregulation of mTORC1 activity upon Ac45 or PRR knockdown, which might reflect the fact that the efficiency of gene knockdown is insufficient to obtain a phenotype or that the cause is multifactorial. The phenocopy obtained with bafilomycin suggests a crucial role for the V-ATPase.

Finally, we did not test the efficiency of Cre-mediated Furin recombination in RIPHER neurons (22), so we cannot entirely rule out any phenotypic contribution by hypothalamus-specific Furin inactivation. However, the phenocopy observed in a β-cell line generated by CRISPR-Cas9 editing strongly suggests a defect at the β-cell level.

In conclusion, we describe the importance of the proprotein convertase FURIN in the regulation of β-cell mass and function in vivo, likely related to its function in regulating the V-ATPase activity. Furthermore, we report that lack of Furin results in activation of the stress-induced transcription factor ATF4, mediated by mTORC1. As such, this study sheds light on a novel molecular mechanism for the regulation of β-cell function by FURIN and provides evidence that it is an important candidate in the PRC1 susceptibility locus for T2D.

Acknowledgments. The authors would like to thank Sandra Meulemans (KU Leuven), Elisa Cauwelierand (KU Leuven), and Cindy Baldwin (Institut de Recherches Cliniques de Montréal) for technical assistance.

Funding. This work was supported by the Fonds Wetenschappelijk Onderzoek (FWO) Vlaanderen (grant G0B0617N) and by the Deutsche Forschungsgemeinschaft (German Research Foundation) under Germany’s Excellence Strategy within the framework of the Munich Cluster for Systems Neurology (EXC 2145 SyNergy–ID 390857198). J.L.E. is a Chercheur-boursier from the Fonds de recherche Santé and was supported by the Canadian Institutes of Health Research (grant PJT-148771). B.R.-M. was supported by the Miguel Servet Type I program (CP19/00098) from the Institute de Salud Carlos III and by the European Regional Development Fund. B.B. and I.C. were supported by predoctoral fellowships from the FWO Vlaanderen.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. B.B., I.C., B.R.-M. and J.W.M.C. designed the research. B.B., I.C., and B.R.-M. performed a majority of in vivo and cell culture experiments, with help from C.S. B.B., I.C., K.V. and N.V.G. performed the electron microscopic analysis. N.J. and J.L.E. performed the experiments with human islets. B.D., S.F.L., and J.D. performed the proteomic analysis. L.V.L. and F.S. performed the microarray analysis. B.B., L.T., and J.D. performed bioinformatic analysis of microarray and proteomic data. B.B., I.C., B.R.-M., and J.W.M.C. collected and analyzed the data and wrote the manuscript, with input from the rest of the authors. All authors approved the final version of the manuscript. J.W.M.C. is the guarantor of this work and, as such, has full access to all the data in the study and takes responsibility for the integrity of the data.