Smad2 and Smad3 cooperate and antagonize simultaneously in vertebrate neurogenesis

The transforming growth factor beta (TGF-β) pathway is one of the most conserved and prolific signaling cascades (Massagué, 2000; ten Dijke and Hill, 2004). Despite its simple architecture, this pathway is involved in the regulation of many processes, including growth, proliferation, differentiation and survival. Extracellular TGF-β and activin are ligands that bind to a type II transmembrane receptor, triggering the assembly of the tetrameric active ligand–receptor complex by binding to a type I receptor. Active ligand–receptor complexes then recruit and phosphorylate C-terminal residues of the downstream effectors Smad2 and Smad3 (R-Smad proteins). Phosphorylated R-Smad proteins form heterotrimeric complexes with Smad4 that enter the nuclei and recruit co-factors in order to ultimately regulate target gene expression (Fig. 1A–C) (Moustakas et al., 2001; Moustakas and Heldin, 2002). These DNA-binding complexes can contain two Smad2 and one Smad4 molecule (S224); two Smad3 and one Smad4 (S334); or even one Smad2, one Smad3 and one Smad4 molecule (S234) (Shi and Massagué, 2003).

Despite their 91% amino acid sequence similarity, Smad2 and Smad3 recruit different co-factors and target different regulatory sequences (Brown et al., 2007). However, they often share similar functions, playing redundant roles in ovarian granulosa cells (Li et al., 2008), in zebrafish mesoderm induction (Jia et al., 2008) and in other cellular contexts (Bernard, 2004). Moreover, when introduced into the Smad2 locus, Smad3 can rescue the Smad2-deletion phenotype in developing mouse embryos (Dunn et al., 2005). However, Smad2 and Smad3 fulfil opposing roles in breast cancer metastasis (Petersen et al., 2010) and regulation of growth and cell migration in pancreatic adenocarcinoma cells (Ungefroren et al., 2011), as well as in forkhead-dependent transcription (Labbé et al., 1998). Moreover, transcriptional profiling experiments showed that Smad3-activated genes are repressed by Smad2 (Yang et al., 2003). How such similar molecules can cooperate and/or antagonize distinct events remains somewhat unclear.

In this study, we characterized the roles of Smad2 and Smad3 in the context of vertebrate neurogenesis, where activation of the TGF-β pathway, in particular Smad3, has been shown to promote cell-cycle exit and differentiation (García-Campmany and Martí, 2007). Our results show a complex interplay between Smad2 and Smad3, where cooperation and antagonism occur simultaneously in the same cellular context. Through in vivo experiments in the chick embryo neural tube, we show that both knockdown and overexpression of Smad2 enhances neurogenesis. We developed a simplified mathematical model of the pathway in which all the experimental observations fit with a scenario in which neurogenesis is promoted by both S234 and S334 trimers.

To study the roles of Smad2 and Smad3 in the developing neural tube of the chick embryo in vivo, we first analyzed their expression in neural tube sections. Strong immunostaining for Smad2 and Smad3 was evident in the ventricular zone where neural progenitors reside, with some enrichment in the transition zone (Fig. 1D,E). Although Smad2 is expressed throughout the dorsal–ventral (DV) axis of the neural tube (Fig. 1D), Smad3 expression is restricted to particular DV domains (Fig. 1E), as reported previously (Garcia-Campmany, 2007). Differentiated neurons that were identified by the expression of RNA-binding proteins HuC and HuD (HuC/D) did not express Smad2 (Fig. 1D) or Smad3 (Fig. 1E), as confirmed by in situ hybridization, which showed the marginal zone of the neural tube where differentiated neurons reside to be devoid of Smad2 and Smad3 mRNA transcripts (supplementary material Fig. S1). To quantify Smad2 and Smad3 protein expression within the ventricular zone, fluorescence intensity was measured at the intermediate neural tube (white rectangles in Fig. 1D,E) as a function of the distance to the lumen. Both total and single-cell measurements of Smad2 and Smad3 inmunostaining intensity (black lines and crosses in Fig. 1F,G, respectively) indicated an enrichment in the transition zone. Moreover, single-cell measurements of the nuclear to cytoplasmic fluorescence intensity ratio as a function of distance to the lumen, showed that Smad2 and Smad3 are mainly located in the cytoplasm. However, as the nuclei move towards the transition zone, the average ratio of Smad2 and Smad3 within the nuclei increases (black lines in Fig. 1F,G, respectively), suggesting a potential transcriptional activation of Smad2 and Smad3 in the transition zone, where neurogenesis occurs.

Neurogenesis in the neural tube involves the lateral migration of neural progenitors committed to differentiation, as well as the de novo expression of pan-neural markers such as HuC/D. Thus, monitoring the position of electroporated cells by evaluating EGFP expression together with that of HuC/D serves to assess neurogenesis in vivo (Fig. 2A) (García-Campmany and Martí, 2007; Xie et al., 2011). When analyzed 48 hours post electroporation (hpe) of either control (pCIG) or Smad2, EGFP⁺ cells were evenly distributed throughout the ventricular zone and the mantle zone (HuC/D⁺ cells: Fig. 2E,F), although overexpression of Smad2 resulted in an increase in the total protein detected in western blots (Fig. 2B) and through the increased fluorescence intensity at the cellular level (Fig. 2C). However, Smad3 promoted the lateral migration of EGFP⁺ cells; a phenotype that was enhanced by the co-electroporation of Smad2 and Smad3 (Fig. 2G,H). When quantified, a significantly higher proportion of differentiated cells was detected after the electroporation of Smad3 and Smad2/3 (% GFP⁺, HuC/D⁺ cells at 48 hpe: 33±1% controls, 35±2% Smad2, 49±3% Smad3, 69±1% Smad2/3; Fig. 2I).

To test the requirement of Smad2/3 in neurogenesis, we reduced their endogenous expression by electroporating short-hairpin RNAs specific for chick Smad2 and Smad3. Electroporation of shSmad2 strongly diminished Smad2 protein expression (Fig. 2D,K), without affecting the endogenous expression of Smad3 (supplementary material Fig. S2). Conversely, electroporation of shSmad3 provoked a strong reduction in Smad3 protein expression (Fig. 2D,M), without affecting endogenous Smad2 expression (supplementary material Fig. S2). However, whereas knockdown of Smad3 caused the expected reduction in neurogenesis, knockdown of Smad2 resulted in a significant increase in neurogenesis (% GFP⁺, HuC/D⁺ cells at 48 hpe: 40±3% control, 63±3% shSmad2, 25±2% shSmad3; Fig. 2N). Together, these results suggest that Smad2 cooperates with Smad3 in the activation of target genes involved in neurogenesis, whereas it also antagonizes Smad3 activity in the same cellular context.

In order to dissect out the transcriptional responses mediated by Smad2 and Smad3, we expressed vectors containing luciferase reporters driven by either Smad2-specific [the activin response element (ARE)] (Chen et al., 1996) or Smad3-specific (-CAGA-) (Dennler et al., 1998) response elements (Fig. 3A–D) in the chick neural tube. The ARE reporter was activated 24 hpe of Smad2 in the chick neural tube, but not after Smad3 electroporation. However, co-electroporation of Smad3 with Smad2 interestingly resulted in a strong decrease in ARE activity compared with that obtained with the same concentration of Smad2 alone (∼3× reduction), reflecting antagonism between Smad2 and Smad3 in the regulation of the Smad2-specific targets (Fig. 3B). By contrast, the CAGA reporter was strongly activated by Smad3 but not by Smad2 electroporation. However, co-electroporation of Smad3 with Smad2 enhanced CAGA activity up to fivefold that of Smad3 alone, reflecting the cooperation between the Smad2 and Smad3 in the regulation of Smad3-specific targets (Fig. 3C).

Because knockdown of Smad2 resulted in increased neurogenesis, we assessed the transcriptional responses in the absence of either Smad2 or Smad3. The weak endogenous responses to the ARE reporter did not allow its activity to be monitored following loss of function, whereas knockdown of Smad3 (24 hpe of shSmad3) did reduce the activity of the CAGA reporter (Fig. 3D). Interestingly, knockdown of Smad2 resulted in strong activation (∼20×) of the CAGA reporter (Fig. 3D). Because Smad3 expression levels were not increased after shSmad2 electroporation (supplementary material Fig. S2), these results suggest that the presence of Smad2 prevents Smad3 from activating its own targets, provoking an antagonistic effect.

To confirm these results at the cellular level, we co-electroporated Smad2 and Smad3 with the CAGA reporter driving EGFP expression (Fig. 3E–K). Although Smad2 was unable to activate CAGA-EGFP above control levels (Fig. 3E,F), Smad3 induced strong cell-autonomous activation of the reporter (Fig. 3G). Co-electroporation of Smad2 and Smad3 produced strong EGFP expression similar to that of Smad3 alone, which was probably due to saturation of EGFP expression (data not shown). Moreover, knockdown of Smad2, but not Smad3, induced cell-autonomous expression of CAGA-EGFP (Fig. 3H–K), which is consistent with the results from the luciferase experiments.

The experimental data reflect a complex interplay between Smad2 and Smad3 in the regulation of their specific targets: a relationship that is reproduced in neurogenesis. In gain-of-function (GOF) experiments, Smad2 and Smad3 cooperate to regulate Smad3 targets and neurogenesis while antagonizing the regulation of Smad2-specific targets. Conversely, in loss-of-function (LOF) conditions Smad2 and Smad3 antagonize the regulation of Smad3-specific targets and neurogenesis. To understand the mechanisms driving such cellular responses, we developed a simplified theoretical model of the TGF-β pathway (see the Materials and Methods). Models have been developed previously to study nucleo-cytoplasmic shuttling (Clarke et al., 2006; Schmierer et al., 2008; Clarke and Liu, 2008), signal processing (Cellière et al., 2011), the generation of transient responses (Melke et al., 2006; Vilar et al., 2006; Zi et al., 2011) and the role of endocytosis (Zi and Klipp, 2007) in the TGF-β signaling pathway. However, to our knowledge, no model has been developed to predict how transcriptional responses are regulated.

We developed such a model that simulates the transcription of two arbitrary genes specific for Smad2 (Fig. 4A, black lines) and Smad3 (Fig. 4A, red lines). Increasing Smad2 levels (∼7×, based on experimental data) enhances Smad2 target transcription (∼7× increase) without significantly affecting direct Smad3 targets (Fig. 4A). However, increasing Smad3 (∼12×, based on experimental data) enhances transcription of Smad3 (∼3× increase) targets while reducing transcription of specific Smad2 targets (∼7× decrease; Fig. 4A). A simultaneous increase in Smad2 and Smad3 reduces Smad2-specific targets compared with the transcription provoked by Smad2 alone (∼1.2× decrease), whereas it enhances Smad3-specific transcription (∼1.5× increase) compared with Smad3 alone. Moreover, the model predicts that a reduction in Smad2 levels (∼9× decrease) increases Smad3-specific targets (∼3× increase; Fig. 4A, LOF). Together, these in silico experiments qualitatively reproduce in vivo responses for Smad2- and Smad3-specific target genes (Fig. 4B–D).

Overall, the mathematical model reproduced the experimental data when we consider that Smad4 is not limiting for the formation of the heterotrimers, and that the transcriptional trimers are formed reflecting the concentration of Smad2 and Smad3. In the wild-type state, R-Smad proteins interact with one another and with Smad4 to form each type of possible heterotrimer (S₂₂₄, S₂₃₄ and S₃₃₄). However, Smad2 targets are activated solely by the S₂₂₄ complex, whereas Smad3 targets are activated both by S₂₃₄ and S₃₃₄. Augmenting the amount of Smad2 favors the formation of S₂₂₄, and reduces the formation of S₂₃₄ and S₃₃₄ trimers. Conversely, increasing Smad3 favors the formation of the S₃₃₄, and reduces that of S₂₃₄ and S₂₂₄ complexes. Similarly, a reduction in Smad2 would therefore reduce the formation of S₂₂₄, and increase the formation of S₂₃₄ and S₃₃₄, thereby resulting in the activation of Smad3-specific targets and enhanced neurogenesis (Fig. 4B–D).

To experimentally test the outcome of this in silico model, we quantified the trimer formation in each experimental condition (Fig. 4E,F). In the wild-type state, the chick embryo neural tube contains the S₂₃₄ trimer (Fig. 4F), whereas when Smad2 is overexpressed, the S₂₂₄ trimer preferentially forms, without affecting the Smad3-specific transcription of neurogenic targets (Fig. 4F,G). Smad3 overexpression results in the preferential formation of the S₃₃₄ trimer, which enhances Smad3-specific transcription and neurogenesis (Fig. 4F,G). Similarly, co-expression of Smad2 and Smad3 enhances transcription and neurogenesis of Smad3 by promoting the formation of the S₂₃₄ trimer (Fig. 4F,G).

Here we show in vivo and in silico that the formation of transcriptionally active Smad2/3 hetero-trimers dictates the dynamics of Smad2- and Smad3-specific transcription. Although clearly modulated by Smad2, the role of TGF-β in neurogenesis follows the trend of Smad3 direct targets: it is enhanced after Smad3 GOF alone or in combination with Smad2, diminished after Smad3 LOF and increased after Smad2 LOF. This correlates with the binding of Smad3, but not Smad2, to the stress response factor ATF3 in response to TGF-β, which mediates the repression of transcription of Id1 (inhibitor of DNA binding 1) (Kang et al., 2003). Additional direct targets controlling proliferation and differentiation might include p21CIP1 and p15INK4b, because their regulatory regions have binding sites for Smad3 (Gomis et al., 2006).

Our model predicts an increase in Smad2-driven transcription after Smad3 LOF, which we could not detect in our in vivo experiments owing to the poor sensitivity of the ARE-luciferase reporter. This prediction is consistent with the results in tissue culture experiments where the activity of the ARE reporter is reduced in Smad2-null cells, whereas it is activated in Smad3-null cells (Piek et al., 2001). It also fits the data showing that TLP, a TRAP-1-like protein that does not phosphorylate R-Smad proteins, represses the ability of Smad3 to form a transcriptionally active complex while increasing Smad2-driven transcription (Felici et al., 2003).

We propose that the mechanism reported in this study is responsible for the divergence between studies that focused on comparing the functions of Smad2 and Smad3 in other biological contexts. Indeed, this dual co-operation/antagonism not only influences neurogenesis but also affects other biological events regulated by the TGF-β pathway. A deeper understanding of how R-Smad proteins interact will be crucial to understand the dynamics and function of the TGF-β pathway in order to develop more efficient treatments for diseases in which this pathway is compromised, because impairment of the growth inhibitory function of TGF-β is considered to be a hallmark of cancer (Dennler et al., 1999; Hanahan and Weinberg, 2000).

We consider a Michaelis–Menten type of interaction for the phosphorylation of the R-Smad proteins (Eqs. 1,2). Michaelis–Menten constants and phosphorylation rates of each of the R-Smad proteins by the TGFβR receptor were chosen to be equal. Transcription was computed as a Hill function of the amount of active complex with no cooperativity. Parameter values (supplementary material Table S1) and parameter selection are discussed in detail in the selection of model parameters section.

Several authors (e.g. Labbé et al., 1998) hypothesize that the antagonism emerges from competition of the R-Smad proteins to bind to Smad4, but the fact that the pathway can get activated up to 20 times above the endogenous levels (Fig. 3C,D) reflects that the amount of Smad4 is not the limiting reactant in the reaction in the wild-type situation. In addition, this hypothesis does not explain the strong cooperativity in Smad3 direct targets when combining Smad2 with Smad3, explained by our model.

Experimental measurements of total concentration of Smad2, Smad3 and Smad4 have been performed in a specific cell type (Clarke et al., 2006), reporting values of 100,000, 20,000 and 100,000 total molecules per cell of Smad2, Smad3 and Smad4, respectively. Assuming a cell volume of 1×10⁻¹² l, this gives us a value of  = 166 nM,  = 33 nM,  = 166 nM. Because wild-type values of Smad2, Smad3 and Smad4 are likely to be very cell-type dependent, and our three semi-quantitative estimations (immunofluorescence, western blot and co-immunoprecipitation) reflect similar values for Smad2 and Smad3, we decided to use for the model a standard value of 100 nM for all three Smad proteins, which is close to the experimental measurements reported previously (Clarke et al., 2006). Our model predictions are robust against slight changes in Smad concentration (within the same order of magnitude), and it also works when using the exact values measured previously (Clarke et al., 2006).

To estimate the concentration after GOF, we performed experiments where we electroporated Smad2 and Smad3 and immunostained neural tube sections 24 hpe (supplementary material Fig. S3) This way, we can measure the change in immunofluorescence intensity between not electroporated and electroporated cells. The average increase in pixel intensity for Smad2 was 7.6×, and for Smad3 was 12.1×. This gives us a final values for  = 1216 nM,  = 403 nM. Quantification of this experiment is shown in Fig. 2H.

To estimate the concentration after LOF, we performed measurements of the immunofluorescence intensity in cells electroporated with shRNA against Smad2 and Smad3 compared with control (supplementary material Fig. S3). The average reduction in pixel intensity for Smad2 was 9.6×, and for Smad3 was 10.1×. This gives us a final values of  = 17.4 nM,  = 3.3 nM. Quantification of this experiment is shown in Fig. 2I.

The balance between phosphorylation and de-phosphorylation of the R-Smad proteins has been shown to be a key mechanism to regulate the dynamics of nuclear–cytoplasmic shuttling (Hill, 2009), so nuclear localization of the R-Smad proteins can be used as a readout of the pathway activity. This way, the balance between phosphorylation and de-phosphorylation in our model in conditions of pathway activation, should mimic the experimental (i.e. most R-Smad phosphorylated after pathway activation, and most R-Smad un-phosphorylated when the pathway is inactive). Owing to the lack of experimental data regarding the un-phosphorylation rate of R-Smad, for our model we use a standard value used in generalized models involving phosphatase activity (Kholodenko et al., 2000; Lipkow and Odde, 2008) and other models of  = 1 ⁠.

Previous experimental data (Wakefield et al., 1987) and other previous TGFβR models (Zi et al., 2011) use a value of total TGFβR−R−I = 1000. Our model does not take into account receptor complex assembly and multimerization of receptor subtypes (Clarke and Liu, 2008), so we assume 20% of these values of total [TGFβR−TypeI−R] as active after ligand stimulation. Because these are membrane proteins, to calculate the concentration, we assume that they are distributed in a volume equal to the cell area (given a cell radius value of 13 μm for a spherical cell volume of 1×10⁻¹² l) times the typical height of a ligand–receptor complex (Allard et al., 2012) to obtain a value for [TGFβ] = 149 nM.

For the value of we used the median value used previously (Clarke et al., 2006). The value for the Michaelis–Menten constant has been selected within the range listed (Clarke et al., 2006) (the median 2.89×10⁵ molecules resulted in very low vales of R-Smad phosphorylation, so for lack of experimental data we assume the minimum value listed (Clarke et al., 2006) of  = 633 molecules = 1.03 nM. Values around ten times above or below the selected value also produce simulation results in agreement with our experimental data.

Due to lack of experimental values for association and dissociation of the R-Smad proteins after phosphorylation, we have used values where after phosphorylation, steady state is achieved in the order of seconds, and most of the R-Smad proteins are in a dimer and trimer configuration when phosphorylated, as shown experimentally. Experiments in Fig. 3C,D show that removing Smad2 increases activation of Smad3-specific targets, suggesting that the presence of Smad2 prevents Smad3 from activating its own targets. To take this into account, Smad3 affinity towards Smad3 was set lower than towards Smad2. This configuration induces more Smad2–Smad3 than Smad3–Smad3 complexes.

To our knowledge, experimental measurements of transcription rates of the different heterotrimer complexes in specific targets for Smad2 and Smad3, have not been reported. Previous systematic studies on mRNA production (Schwanhäusser et al., 2011) have shown an average of 10 mRNA molecules per hour. In addition, the same studies have reported an average of 140 protein molecules per mRNA. So values for  = 0.38 was chosen based on these data.

Fig. 3C shows that Smad2, when co-electroporated with Smad3, plays a role in Smad3 transcription. Our model fits the experimental data when considering that this role of the Smad2–Smad3–Smad4 complex in Smad3-specific targets is lower that the role of the Smad3–Smad3–Smad4 complex, so we set  = 19 generically as 50× higher than ⁠. Differences between and above 25× also reproduce the experimental data. Values higher than 50× difference result in higher differences between the different conditions, also in agreement with the experimental data, where fold change in transcription compared with the wild type is qualitatively higher than the model prediction.

Non-overlapping sequences of chicken Smad3 (bp 550–569) and Smad2 (bp 264–283) were targeted following a previously published protocol. These fragments were cloned into the pSHIN vector (Kojima et al., 2004) to generate short hairpin RNAs (shRNAs) that could be electroporated into the neural tube. Controls were performed by electroporating the empty pSHIN vector. Human N-terminal FLAG-tagged Smad2, Smad3 and Smad4 were cloned into a pCIG vector.

Chick embryos were staged according to Hamburger and Hamilton (Hamburger and Hamilton, 1951). Plasmid DNA was injected into the lumen of HH stage 12 neural tubes, electrodes were placed either side of the neural tube and electroporation was performed using an Intracel Dual Pulse (TSS-100) electroporator delivering 50 ms square pulses of 40 V. Transfected embryos were allowed to develop for 24–48 hours, dissected and processed for immunohistochemistry, in situ hybridization, luciferase assays or protein extraction.

For in situ hybridization, embryos were fixed overnight at 4°C in 4% paraformaldehyde diluted in PBS, rinsed and processed for whole-mount RNA in situ hybridization using probes against chick Smad2 and Smad3 from the chicken EST project following standard procedures. Immunofluorescence was performed on transverse sections (40 µm) after fixation in 4% paraformaldehyde diluted in PBS for 2–4 hours at 4°C using the following monoclonal antibodies: Smad2 (Cell Signaling, 5339), Smad3 (Abcam, ab28379), HuC/D (Molecular Probes, A-21271). Anti-HA and anti-FLAG antibodies were generated in-house (Sebastian Pons Laboratory).

Embryos were electroporated with the DNAs indicated with a firefly-luciferase reporter construct driven by the ARE and CAGA promoter sequences, and a Renilla luciferase reporter construct for normalization (Promega). Embryos were harvested after 24 hours in ovo and only the interlimb sections of GFP-positive neural tubes were dissected and homogenized on ice with a Dounce homogenizer in passive lysis buffer. Firefly and Renilla luciferase activities were measured using a dual luciferase reporter assay system (Promega).

For the study of the S234 complexes by co-immunoprecipitation, embryos were co-electroporated at HH14 with the DNAs indicated (the total amount of DNA electroporated was normalized with the empty vector pCIG, n = 5 per condition) and the neural tube was dissected at 24 hpe. The cells were lysed by sonication pulses in lysis buffer (20 mM Tris-HCl, pH 7.4, 137 mM NaCl, 10% glycerol, 1% NP-40, 1 mM CaCl₂, 1 mM MgCl₂, 0.05% sodium azide) supplemented with protease inhibitors (1 mM PMSF, 1 µg/ml aprotinin and 1 µg/ml leupeptin, Sigma). The insoluble material was removed by centrifugation and one tenth of the resulting supernatant was reserved as the input material; the remaining volume was immunoprecipitated for 4 hours at 4°C with a monoclonal anti-HA agarose conjugate (Sigma). The beads were washed three times and boiled with 2× SDS loading buffer and the samples were resolved by 8% SDS-PAGE. After western blotting, the membranes were incubated with the indicated primary antibodies and infra-red-tagged secondary antibodies (anti-Rabbit IgG DyLight 800 conjugate and anti-rabbit IgG DyLight 680 conjugate; Thermo). The membranes were scanned using the Odyssey infrared imager (Li-Cor).

The ratio of differentiated electroporated cells was calculated by first quantifying the number of EGFP⁺ cells and then by quantifying the HuC/D⁺ cells among those that were electroporated. At least ten different confocal transversal sections of the neural tube were quantified from at least three different electroporated embryos. The change in immunofluorescence staining intensity was quantified in 50 electroporated cells versus 50 non-electroporated cells in three different neural tube sections for each condition (supplementary material Fig. S3). The bars represent the mean value and the error bars represent the s.e.m. Statistical significance in terms of P-values was calculated using the Z-test function in Numbers (Macintosh).

We are grateful to Susana Usieto for her invaluable technical assistance.