Vector Construction

To construct the pGL4-Irx3 reporter, the promoter of the human Irx3 gene was amplified from genomic DNA extracted from the A431 epidermoid carcinoma cell line by PCR with primers ATACGGTACCGGTCCTCCCCAAACTTTCCC and GATTAGATCTCTCCGCGTTCGCCTATTGAT including KpnI and BglII restriction sites for directional cloning into the pGL4.12 luciferase reporter vector (Promega). Purified PCR products and a pGL4.12 vector plasmid preparation were digested with the two restriction enzymes as prescribed by the manufacturer (New England Biolabs), purified from agarose gel slices, and ligated at 4°C overnight in the presence of T4 DNA ligase (New England Biolabs). After transformation of competent Eserichia coli DH5 cells, transformants were grown in ampicilline containing Luria Bertani media and plasmids were purified from bacteria using the Qiagen Plasmid Mini Kit. The constructs were confirmed by bidirectional sequencing (Starseq GmbH).

To construct the pcDNA-Nkx2.2 expression vector, the cDNA of human Nkx2.2 (Source BioScience IRATp970C11116D) was amplified using PCR primers ATACAAGCTTCCACCATGTCGCTGACCAACACAAAG and GCGCCTCGAGTCACCAAGTCCACTGCTGG including HindIII and XhoI restriction sites and the Kozak sequence for optimal initiation of the translation. The PCR product was digested with the two enzymes and cloned in the eukaryotic expression vector pcDNA3 (Invitrogen) in a similar manner as described for pGL4-Irx3. All the constructs were confirmed by bidirectional sequencing (Starseq GmbH).

Cell Culture

HEK293T human embryonic kidney cells were cultured in high-glycose DMEM (Sigma) supplemented with 10% FCS and 1% Penicillin-Streptomycin (Gibco) at 37°C in a humidified atmosphere containing 5% CO₂. For transfections the calcium phosphate method was used as previously described [55]. Briefly, cells were seeded in pre-gelatinized 24 well plates and 24 hours after seeding, at 90% confluency, chloroquine was added to the medium. For each well of the 24-well plate, 0.83 µg plasmid was mixed with 6.37 µl 0.5 M CaCl₂ containing solution in a total volume of 12.73 µl, thereafter 12.73 µl of HEPES buffered solution was added to the mixture of DNA and CaCl₂, and the final solution was aerated with a pipette tip and incubated for 10 minutes before it was added to the cells dropwise. To prepare the HEPES buffered solution 28 ml 5 M NaCl solution, 11.9 g HEPES, 750 µl 1 M Na₂HPO₄ were dissolved in water to give a total volume of 500 ml and the pH was set exactly at 7.1. The plasmids used for each experimental condition (all of those repeated in quadruplicates) are indicated in table S6. The pMAX-GFP was included in varying quantities to attain a constant amount of DNA across the wells. The transfection medium containing chloroquine and calcium-phosphate precipitate was replaced by the original medium after 20 hours and the transfection efficiency was also visually controlled by observing GFP fluorescence.

Luciferase measurement

Cells were washed in phosphate buffered saline and harvested with passive lysis buffer (Promega) for luciferase and protein measurements. Extracts were distributed into the wells of a 96 plate and Luciferase Assay Reagent II (Promega) was injected onto the samples in a Varioskan (Thermo Scientific) plate reader. The integration time was set at 1000ms for each measurement. The light emission values were normalized against the protein content of the samples measured by BCA protein assay (Thermo Scientific): 200 µl working reagent (50 part BCA ‘A’ +1 part BCA ‘B’) was added to 25 µl of each sample and after 30 minutes incubation in the dark the absorbance of the samples were measured in a Varioskan plate reader at 562 nm.

RNA Extraction and Reverse Transcription

Total RNA from transfected cells (cells with expression of exogenous Nkx2.2: Nkx2.2_a and Nkx2.2_b) and non-transfected cells (cells which have only endogenous Nkx2.2: control_a and control_b) were isolated using the RNeasy Plus Mini kit (Qiagen, Cat. No. 74136). Following extraction, RNA Cleanup was performed with a DNase digestion step using RNeasy Mini kit (Qiagen, Cat. No. 74106) according to the manufacturers protocol. Reverse transcription was then performed by using the M-MLV reverse transcriptase (Invitrogen, Cat. No. 28025-013), with approximately 1.2 µg total RNA retrieved from each sample. The amplified cDNA was aliquoted and stored at −20°C.

Real-Time PCR

The levels of Nkx2.2 gene expression were determined using relative real-time PCR quantification. Amplifications were performed with TaqMan Gene Expression Master Mix (ABI), on a Rotor-Gene Q cycler (Qiagen). Each reaction consisted of 20 ng cDNA, 1x TaqMan Gene Expression Master Mix (final concentration), 1x PrimeTime® Mini Assay probe and primer mix (IDT assay Hs.PT.49a.94096.g with 0.5 µM (final concentration) of each primer pair and 0.25 µM (final concentration) of probe). Reactions were performed in triplicate. GAPDH was used as the reference gene (IDT assay Hs.PT.49a.2918858.g). The amplification conditions comprised an initial UDG incubation at 50°C for 2 min, then denaturation at 95°C for 10 min, followed by 40 cycles of denaturation at 95°C for 15 sec, then annealing and extension at 60°C for 1 min. Sequences of primers and probes used for expression analysis are listed in Table S1. The relative expression of Nkx2.2 gene was calculated from the average Ct values of each triplicate by using the method.

Experimental Statistics

Statistical analyses of the luciferase assay and the relative real-time PCR were performed. Differences in the levels of luciferase activity or gene expression in between samples were analysed using one-way analysis of variance (ANOVA).

Construction of GRNs

TFs downstream of Shh signalling were used as nodes of the network. Regulatory connections between them were based on published studies and possible additional links were tested using Boolean modelling (figure 2(a)).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Gene regulatory networks.

(A) GRN of ventralization assembled from published experimental data. TFs are denoted by ellipses. Transcriptional inhibitions between TFs and target genes are denoted by blunted arrowheads. Associations depicted by continuous lines were demonstrated experimentally; associations denoted by dotted lines were tested with Boolean simulations. (B) Minimal GRNs obtained from Boolean simulations. Inhibition represented by continuous lines were experimentally verified, additional dashed lines represent connections present in the minimal matching GRNs. The line between Nkx2.2 and Dbx2 contains a circle at both ends representing that the direction of inhibition is not defined in the minimal GRNs.

https://doi.org/10.1371/journal.pone.0111430.g002

Conversion of a GRN to a Boolean model

The regulatory links associated with each of the GRNs shown in figure 2(a) were transformed into Boolean update rules. Only inhibitory connections were considered and hence, the general update rule for each Boolean variable was defined as (1)where attains the values 0 or 1 and the set contains the Boolean values of the regulators (inhibitors) of . The set of inhibitors and, hence, the update rules were fixed during each Boolean simulation.

Boolean Simulations

An asynchronous random order updating (ROA) scheme was utilized as described in [24]. Briefly, a Boolean step of ROA consists of updating all variables in a random order. All Boolean simulations were conducted from a defined initial state and at each simulation step the binary value of each variable was updated according to the corresponding Boolean logic rule. The Boolean simulations were continued until an attractor state was obtained.

Test of alternative GRNs

For each possible GRN ( in total) shown in figure 2(a) a set of Boolean update rules were defined. For each set of update rules simulations were conducted from all possible initial states (each of the eight variables could take the value 0 or 1, so there were initial states). For each set of update rules a set of attractors were obtained and the fixed point attractors associated with a particular set of regulators were compared to the binary representation of the expression states of progenitor cells in the ventral neural tube (see figure 1(c)).

Simulating knockout or overexpression

During simulations to model knockout and ectopic expression experiments, the value of the specific TF was held fixed at 0 (knockout) or 1 (ectopic expression).

0.1 Defining the TFs to be represented in the model

A set of TFs downstream of Shh signalling are known to promote the development of specific cell types (p3, pMN, p2, p1 and p0 neurons) in the vertebrate neural tube (see [3], [34] and references therein) as shown in figure 1(a). Combinatorial expression of eight specific TFs (namely Nkx2.2, Pax6, Olig2, Irx3, Nkx6.2, Dbx1, Nkx6.1, Dbx2) depicted in figure 1(b) determines the five distinct cell types generated in the ventral spinal cord. In our mathematical model, these TFs were chosen as variables to account for the different progenitors. Other TFs such as Nkx2.9, FoxA2 or Pax7 are also targets of Shh signalling but were not included in our model, either because of overlapping functions with other TFs or because their spatial signalling range (as defined by the location along the dorsoventral axis of the embryo) was outside of the region considered in our model.

Patterning in the neural tube is thought to be achieved via transcriptional repression [35]: Nkx2.2, Nkx6.1, Nkx6.2, Dbx1 and Dbx2, but not Pax6 or Irx3 contain a domain that has repressive effects. The inhibitory role of these TFs was demonstrated by reporter assays with full length and truncated proteins lacking the repression domain [35]. During the simulations we assumed that all factors including Pax6 and Irx3 act as transcriptional repressors, based on studies uncovering their inhibitory roles [36]–[38]. It is of interest to note that most of the TFs present in our model have been shown to act both as repressors and activators [36]–[39], their status being dependent on the presence of co-regulators [35].

In our model none of the TFs have activating inputs and hence assumed to be expressed constitutively in the absence of repressing signals. This assumption correlates well with the observation that Pax6, Irx3, Dbx1 and Dbx2 are expressed in the absence of Shh, but their expression persist even in p0 cells, when Shh is present. The other TFs in our model: Olig2, Nkx2.2, Nkx6.1 and Nkx6.2 can be activated by Shh [36]–[38], which was assumed to be present throughout the simulations. Importantly, even though Shh was not explicitely incorporated into our model, its action was represented via the assumption that the TFs activated by Shh were constitutively expressed in the absence of their inhibitors.

Interactions between the eight candidate TFs in the context of ventralization have been studied experimentally (see figure 2(a) and table S1), using a range of techniques that include analysis of protein expression patterns in either chick embryos after misexpression of TFs [37], [38] or mutant mouse embryos [38], [40].

0.2 Boolean network model predicts regulatory connections necessary to obtain fixed point attractors that match specific cell types in the ventral spinal cord

Developmental GRNs should be robust to perturbations [29], [41] and hence their topology is expected to define its fixed point attractors (hereafter referred to as attractors) which here correspond to experimentally observed cell phenotypes (p0, p1, p2, pMN and p3 progenitors) as shown in figure 1(c). The GRN assembled from experimentally derived regulatory connections (see the network defined by the continuous arrows in figure 2(a)) was converted into a Boolean model and asynchronous simulations were conducted (as described in Materials and Methods) to identify attractors of the system. However only cell types p0, p1 and p2 were represented and four of the obtained attractors did not correspond to any cell types (see table S2). Attractor limit cycles were not obtained in any of the simulations.

In order to obtain a GRN that can account for the different progenitor cell types that emerge during ventralization, additional regulatory connections were tested in a Boolean framework. When modelling transcriptional regulation, experimental findings (summarized in table S1) informed our modelling assumptions as follows:

1. only transcriptional inhibition was considered
2. transcriptional inhibition was always incorporated when there was supporting experimental evidence of regulation
3. transcriptional inhibition was never incorporated when there was supporting experimental evidence of absence of regulation
4. transcriptional inhibition was not incorporated if two TFs were coexpressed in at least one cell type as shown in figure 1

Using the above assumptions, the wiring diagram shown in figure 2(a), representing the interactions between the TFs, was constructed. In figure 2(a), continuous lines represent experimentally verified connections while dashed lines represent possible additional transcriptional regulations. During the simulations, the continuous lines were always present and each dashed line was either present or absent, so that in total different GRNs were tested as described in Materials and Methods.

In order to obtain matching GRNs, attractors of each Boolean model were compared to the binary states that correspond to the different cell types of the ventral neural tube (see figure 1(c)). If all five cell types were represented and no additional fixed points were found, then the GRN was considered matching. The GRN containing all possible additional connections was matching, but GRNs with fewer connections were also sufficient. The two simplest GRNs, i.e. the GRNs containing the smallest number of regulatory connections are depicted in figure 2(b). The minimal number of additional connections compared to default was three: in all matching GRNs Nkx6.1 was predicted to inhibit Dbx1 and Nkx2.2 was predicted to inhibit Irx3; additionally, in the two simplest matching GRNs, either Nkx2.2 inhibited Dbx2 or Dbx2 inhibited Nkx2.2.

Next, we studied the dynamics of the simplest matching GRNs. Figure 3 shows the results obtained with ROA updating (see Materials and Methods) and using the Boolean logic rules corresponding to the GRN depicted in figure 2(b) when Nkx2.2 inhibits Dbx2. Unlike figure 2, where nodes represent TFs, the nodes in figure 3 represent the joint expression states of TFs, i.e. the binary value of each TF in the order (Nkx2.2, Nkx6.1, Olig2, Pax6, Irx3, Dbx2, Nkx6.2, Dbx1). Red nodes indicate attractor states and are denoted by the name of the corresponding neural progenitor (p0, p1, p2, pMN and p3 cells), other nodes are labelled by the expression level of each TF. Because of the asynchronous updating, one state can have more than one successor states, but attractors of the system persist. In figure 3, edges starting from a node represent the most likely update route: in the case of two or more arrows starting from the same node, more than one successor states were equally possible.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Boolean simulation results of the minimal GRN where Nkx2.2 inhibits Dbx2.

In contrast to figure 2, the nodes in this figure represent the TF expression states as a vector of zeros and ones in the order (Nkx2.2, Nkx6.1, Olig2, Pax6, Irx3, Dbx2, Nkx6.2, Dbx1). Red nodes stand for attractor states and are denoted by the name of the corresponding neural progenitor (p0, p1, p2, pMN and p3 cells), other nodes are labelled by the expression level of each transcription factor. As a result of the asynchronous update in the Boolean model, one state can have more than one successor states, but the figure only shows the most probable update route(s) for each node.

https://doi.org/10.1371/journal.pone.0111430.g003

The full transition tables for both of the simplest Boolean logic rules are shown in tables S3 and S4. In the case of the simplest logic rule, when Nkx2.2 inhibits Dbx2, transient oscillations never occur and starting from any of the initial states, an attractor state may be reached in a single Boolean step. When Dbx2 inhibits Nkx2.2, transient oscillations may occur and starting from any of the initial states, the minimum number of Boolean steps to reach a steady state is one or two.

For each attractor the possible basin of attraction [42] is defined as the set of input nodes from which at least one possible path leads into the attractor state. The percentage of the basin of attraction was calculated for the nodes representing p0, p1, p2, pMN and p3 cells as a percentage of the initial nodes that can attain the attractor. The weighted percentage of the basin of attraction was also calculated: the contribution of each initial state was divided by the number of attractors it could attain. In the simplest GRN when Nkx2.2 inhibits Dbx2, the basin of attraction (see table S5) is largest for the most ventral p3 state and gradually decreases towards attractors corresponding to more dorsal cell phenotypes, whereas when Dbx2 inhibits Nkx2.2 the basin of attraction is largest for the pMN and only the second largest for p3. This result suggests, that the p3 or pMN phenotype is the most robust and the p0 and p1 are the least robust cell types.

The regulatory connections hypothesized by the model were examined further. Each of the inhibition Nkx2.2Dbx2 and Dbx2Nkx2.2 was only present in one of the simplest GRNs found by Boolean modelling (figure 2(b)) and as such their likelihood of being a genuine link is smaller than the others. The inhibitions Nkx6.1Dbx1 and Nkx2.2Irx3 both represent regulatory connections between TFs expressed in distant progenitor domains and, to the best of our knowledge neither link has been validated experimentally yet. Even so, some experimental evidence is consistent with the predicted Nkx6.1Dbx1 link: the findings that Dbx1 is repressed by Nkx6.2 [40] and that Nkx6.1 and Nkx6.2 show overlapping functions [40] can be taken together to indicate a possible Nkx6.1Dbx1 inhibition. We note that this regulatory link was hypothesized by Vallstedt et al. (see figure 8(b) in [40]). To the best of our knowledge not even such an indirect evidence for Nkx2.2Irx3 yet exists and so in section 0.4 we focused our attention to verify the hypothesis that Nkx2.2 inhibits Irx3.

0.3 The Boolean model reproduces experimental results involving over- and underexpression of specific genes

Although the dynamic development of different cell types as a function of Shh signalling could not be studied by our model, it was possible to predict the response to changes in TF expressions. In order to check the validity of our model predictions, under- and over-expression studies were replicated in silico. In the simulations the binary representation of the different neuronal cell phenotypes (see figure 1 (c)) determined the initial values of each TF and the variables were updated at each time step according to the Boolean rules corresponding to the simplest GRNs shown in figure 2(b) (both rules yielded the same results). Under-(over-)expression was simulated by holding the value of the relevant variable fixed at zero (one) and the simulation was continued until an attractor was reached.

We considered experiments in which motoneuron (MN) progenitors were converted from or to different cell types. We chose this cell type for further study because of its outstanding clinical significance as the progenitor which gives rise to MNs that directly innervate muscles. The in silico simulation results are summarized in table 1 and discussed below. The first columns in table 1 show the simulation details: initial state, the list of variables kept fixed at 0 or 1 and the attractor obtained. As a result of the asynchronous updating, more than one attractor is possible. The last two columns show the reference paper and compare the simulation results with the published experimental results.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Boolean simulation results conducted with the given initial state and values kept fixed at 0 (knockout genes) or at 1 (overexpressed genes) to reproduce experimental results.

https://doi.org/10.1371/journal.pone.0111430.t001

Briscoe et al. studied the results of misexpression of Nkx6.1 [37] in cells dorsal to the MN domain. MN subtype determinant markers were detected when Nkx6.1 was misexpressed in the absence of high-level Irx3 expression (in cells where expression of Nkx6.1 occurred earlier than Irx3), whereas V2 neuron markers were detected when co-expression of Nkx6.1 and Irx3 occurred. They also studied Irx3 and Nkx2.2 misexpression in MN progenitors [37]. In Irx3 expressing cells MN markers were absent and V2 neuronal markers were detected more ventral to normal positions. Similarly, MN markers were repressed in Nkx2.2 expressing cells and V3 neuronal markers appeared.

In silico Boolean simulations could reproduce the effect of Nkx6.1 overexpression (simulation 1 in table 3) when Irx3 was absent (Irx3 = 0 throughout the simulation). Conversely, cells were predicted to attain the p2 and not the pMN state, when Irx3 expression levels were allowed to change (simulation 2 in table 3). Although this result did not explain how Irx3 is inhibited during the initial phase of progenitor specification, it showed that in order to obtain pMN cells, the final progenitor cell phenotype had to be attained before Irx3 expression started, as otherwise cells attained the p2 neuronal phenotype. This result is consistent with the experimental observation reported in [37]. Simulations yielded p2 and p3 cell types when Irx3 and Nkx2.2 were fixed respectively at 1 (see simulations 3, 4) in agreement with the experimental results [37].

When Sander et al. studied expressions of neural progenitor markers in Nkx6.1 mutant mouse embryos, they found that in these mice p2 progenitor cells switched fate to p1 cells [43] and MN cell development was also disrupted [43]. The Boolean model could partially reproduce these findings: when we fixed Nkx6.1 = 0, pMN cells turned into an unidentified cell type and p2 cells into p1 or p0 (simulations 5, 6).

In the case of Nkx6.1, Nkx6.2 double mutants, ectopic expression of p0 marker was found in p2 and pMN domains, suggesting that some of the cells that would have originally developed to p2 or pMN, turned into p0 [40]. In Boolean simulations, with Nkx6.1 = Nkx6.2 = 0, pMN cells turned into an unidentified cell type and p2 cells into p1 (simulation 7, 8), so that once again the experimental findings were partly reproduced.

Novitch et al. studied the effect of misexpressed Olig2 in the spinal cord and found MN markers in cells that expressed Olig2 [38]. This experiment was simulated by keeping the value of Olig2 fixed at 1 in p2, p1 and p0 cells. In all cases, the obtained attractor corresponded to the pMN cell type as summarized in simulation 9.

0.4 Nkx2.2 represses the transcriptional activity of the Irx3 promoter in a cellular context

Boolean modelling of the GRN of ventralization predicts that an inhibitory link between Nkx2.2 and Irx3 is necessary to attain all attractors associated with the different neural progenitor cell phenotypes. To validate the predicted Nkx2.2Irx3 link experimentally, we have constructed a reporter plasmid (hereafter referred to as pGL4-Irx3) based on the pGL4.12 reporter vector (Promega) containing an improved luciferase reporter gene. The promoter of the Irx3 gene (chr16:54321583-54323661) was amplified by PCR from genomic DNA extracted from the A431 [44] human epidermic carcinoma cell line and inserted upstream of the cDNA encoding luciferase. To express Nkx2.2, its full length cDNA was ordered from SourceBioscience (IRATp970C11116D), the appropriate restriction sites were added by PCR amplification and it was cloned it into the eukaryotic expression vector pcDNA3 (Invitrogen).

Calcium-phosphate mediated transient co-transfection experiments were performed in HEK293T [45] cells. The pGL4-Irx3 construct or a promoterless pGL4.12 was introduced into the cells with or without different amounts of the pcDNA3-Nkx2.2 plasmid (see figure 4(a)). Green fluorescent protein (GFP) expressing control plasmid, that does not bind to or regulate the expression of Irx3, was spiked onto every condition to verify the transfection efficiency (see a representative photo of transfected cells in figure S1) while the total amount of DNA and the amount of pGL4-Irx3 or promoterless pGL4.12 reporter plasmid were kept constant. Additionally, qPCR was used to confirm that the transfected cells expressed exogenous Nkx2.2 (see figure 5). The amount of cotransfected plasmids in each condition for qPCR and luciferase measurements are listed in table S6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Relative luciferase activity.

The light emission of luciferase was measured after addition of Luciferase Assay Reagent II, normalized against the protein content in samples transfected by pGL4-Irx3 (Irx), promoterless pGL4.12 (pB), pGL4-Irx3 plus increasing molar ratio of Nkx2.2 (N1, N2, N3) and promoterless pGL4.12 plus Nkx2.2. In the promoterless reporter vectors there is virtually no activity. The luciferase activity in samples N1, N2, N3 significantly differ from that of sample Irx3 (significance level 0.05). The reduction of the luciferase activity shows that Nkx2.2 negatively regulates Irx3 promoter activity, however no dose-dependent inhibition was observed.

https://doi.org/10.1371/journal.pone.0111430.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. Relative expression of Nkx2.2 in the transfected cells (cells with exogenous expression of Nkx2.2: Nkx2.2_a and Nkx2.2_b) and non-transfected cells (cells which have only endogenous expression of Nkx2.2: control_a and control_b).

Relative expression was calculated using the method. Ct values were normalized to the reference gene GAPDH and the average expression level of the non-transfected cells (control_a, control_b) were normalized to 1. Expression level of Nkx2.2 in transfected cells significantly differ from expression level of Nkx2.2 in non-transfected cells (significance level 0.05).

https://doi.org/10.1371/journal.pone.0111430.g005

We found that the transcriptional activity of the Irx3 promoter was significantly reduced in the presence of Nkx2.2 (figure 4). This observation provides compelling evidence that Nkx2.2 can regulate the Irx3 promoter in a cellular context. Of note, TLE-family proteins can act as co-repressors for Nkx2.2 [35], and TLE-like activity has already been reported in HEK293T cells [46]. Therefore the finding that Nkx2.2 acted in these experiments as a repressor and not as an activator was in line with our expectations. Importantly, we could not observe any effect of Nkx2.2 on the promoterless pGL4.12 vector that itself had a very low level of activity (figure 4), presumably due to the removal of cryptic TF binding sites from its backbone by the manufacturer. The lack of Nkx2.2 effect on the empty pGL4.12 further corroborates that the transcriptional change we have measured is attributable to an effect on the Irx3 promoter sequence and not on any other element in the pGL4-Irx3 plasmid.

Of note, the degree of inhibition induced by different amounts of Nkx2.2 did not differ significantly (see figure 4). The absence of a graded response suggests that the maximal effect was shown, since higher Nkx2.2 concentrations did not induce increased inhibition, indicating that in this concentration range of Nkx2.2 the system may be accurately modelled using a Boolean approach in which Irx3 expression levels are either on or off.