Background

Type 2 diabetes mellitus (T2DM) is characterized by insulin resistance at its onset. Persistence of insulin resistance leads to pancreatic β-cell dysfunction and in extreme cases to β-cell apoptosis [1–3]. Insulin resistance increases the load on β-cells to produce more insulin in order to maintain blood glucose at normal levels. This homeostasis is maintained as long as β-cells can meet the increased insulin demand. However, persistence of excessive nutrients could lead to hyperglycemia, elevated free fatty acids (FFA), and inflammation, which severely impair β-cell functions, leading to insulin resistance and β-cell apoptosis.

The ER in the β-cells is responsible for the production and secretion of insulin. The increased demand for insulin synthesis in the presence of high glucose and FFA levels triggers the accumulation of misfolded proteins in the ER, causing ER stress and the consequent activation of the unfolded protein response (UPR). UPR initially attempts to mitigate ER stress by degrading misfolded proteins and preventing their further accumulation. However, when ER stress is not mitigated, UPR activates the apoptosis signals [4–6]. 78 kDa glucose regulated protein (GRP78) serves as a sensor of protein misfolding [7]. Under non-stressed conditions, GRP78 binds to three UPR initiator proteins, i.e. inositol requiring 1 (IRE1), PKR-like ER kinase (PERK), and activating transcription factor 6 (ATF6), and maintains them in the inactive state [8]. Under stressed conditions, GRP78 dissociates from these three proteins, causing their activation and initiation of UPR.

When ER stress can be resolved, the UPR assists β cells in their survival. However, when ER stress cannot be resolved the UPR activates the pro-apoptotic signals [9]. Hyperglycemia causes oxidative stress through the generation of reactive oxygen species (ROS) [10]. In the absence of an appropriate antioxidant response, the system experiences redox imbalance, leading to the activation of oxidative stress-sensitive signaling pathways. Cytokines, including FasL, TNF α, and IL-6, play important roles in the induction of β-cell apoptosis [11–15] as well as insulin resistance [16, 17]. Caspases serve as the final mediators of apoptosis. The upstream apoptosis initiator caspases 8 and 9 are activated on receiving death signal from the death-inducing signaling complex (DISC) and apoptosome respectively, which in turn activate the downstream apoptosis effector caspases 3, 6 and 7, which ultimately execute apoptosis [18].

Computational modeling is necessary to consolidate information from various sources, such as listed above, in order to obtain a comprehensive understanding of the pathogenesis of T2DM and investigate possible interventions by performing in silico simulations. A few dynamic models of insulin resistance in T2DM have been proposed recently. For instance, Brannmark et al. [19] proposed an ordinary differential equation (ODE) model of insulin signaling in T2DM. Rajan et al. proposed an ODE model to study the contribution of Forkhead box protein O1 (FOXO1) to insulin resistance in T2DM [20]. Another paper [21] presented an ODE model to simulate the development of insulin resistance by hyperglycemia, FFA, ROS, and inhibition of glucose transporter type 1 (GLUT-1) and glucose transporter type 4 (GLUT-4). However, there exists no model of β-cell apoptosis occurring in the T2DM condition. Also, there is no existing work that attempts to integrate the insulin resistance and β-cell apoptosis pathways in order to obtain a comprehensive understanding of the molecular mechanisms underlying T2DM. To discover potential therapeutic interventions for T2DM, it is essential to have a more comprehensive model for the mechanisms causing T2DM pathogenesis.

Therefore, we propose a Boolean network model integrating the insulin resistance pathway and β-cell apoptosis pathway for the purpose of obtaining deeper insights into the mechanisms of development and progression of T2DM. The aforementioned existing models are ODE models, whereas we constructed a Boolean network model. The reason behind this selection is that ODE models require detailed kinetic knowledge and time-series data for accurate parameter estimation. However, the size of our proposed network is relatively big (consisting of 72 nodes) and hence obtaining time-series expression data for all the genes would be expensive as well as time-consuming. Also, estimating the parameters of the ODE model with the time-series expression data of only a small subset of genes would result in erroneous parameter values. Furthermore, in a Boolean network model, gene expression is represented by either TRUE (1) or FALSE (0). By simplifying the gene expression levels into binary states, Boolean networks are feasible for simulating the behaviour of large regulatory networks in a qualitative way.

In a Boolean network model the state of each gene is represented by either 1 (TRUE), indicating the gene is highly expressed, or 0 (FALSE) when the gene is lowly expressed. An edge in a Boolean network can be either activating or inhibiting [22]. In this paper, we have used random asynchronous Boolean simulation [23, 24], which updates genes in a random order in each iteration. This random asynchronous update method is inspired by the stochastic nature of gene regulatory networks, where gene expression alteration occurs in a random order rather than simultaneously [24].

Due to the lack of experimental gene expression data, we validate our simulation results by comparing predicted patterns of gene expression levels with experimental observations reported in the literature. We also analyze the dynamical behaviors of the model by visualizing the state transition graphs under different combinations of input signals. Our results show that the simple Boolean network model can capture some qualitative trends of the genetic circuits regulating the cell fate decision of β-cells, and shed light on the causes and processes of dysfunctional insulin metabolism and loss of β-cell homeostasis that occur in T2DM.

Methods

In this paper, we propose a Boolean network model of β-cell fate in T2DM. The model was constructed by extracting information from the KEGG pathways [25] and literature. The gene interactions incorporated into the model with reference to the existing literature are listed in Table 1. In this model, we integrated the β-cell apoptosis pathway with the insulin resistance pathway, as shown in Fig. 1. The apoptosis pathway consists of the signaling pathways triggered by ER stress (UPR pathway), oxidative stress, and 3 cytokines, i.e. FasL, TNF α, and IL-6. The insulin resistance pathways consist of phosphatidylinositide 3-kinase (PI3K)-protein kinase B (PKB or AKT) (KEGG ID: hsa04151), mammalian target of rapamycin (mTOR) (KEGG ID: hsa04150), janus kinase (JAK)- signal transducer and activator of transcription (STAT) (KEGG ID: hsa04630), and insulin (KEGG ID: hsa04910) signaling pathways. T2DM first causes insulin resistance, i.e. insulin fails to bind to insulin receptors in cells, thereby blocking the uptake of blood glucose by cells. Sustained insulin resistance finally leads to β-cell failure and apoptosis.

Fig. 1
figure 1

Gene Regulatory Network. Insulin resistance and β-cell apoptosis pathways involved in the pathogenesis of Type 2 diabetes mellitus. The red nodes denote the five input signals and the purple node represents β-cell apoptosis. A → B indicates activation of gene B by gene A, and A −| B indicates inhibition of gene B by gene A

Full size image
Table 1 The gene interactions incorporated into the model with reference to the existing literature
Full size table

The Boolean update functions, listed in Table 2, for the target genes in the model are defined by combining activating input genes using OR functions and inhibiting input genes using AND functions. The reason behind this combination strategy is that a target gene will be expressed when at least one of its activating genes is expressed and all of its inhibiting genes are absent.

Table 2 Boolean functions for the Boolean model
Full size table

The proposed Boolean network consists of 72 nodes, of which five are input signals, one node represents Apoptosis, and the remaining 66 nodes represent genes. We employ the random asynchronous Boolean update [23, 24] method to perform the simulations. The random asynchronous Boolean method first generates a random permutation of the nodes at each time step and updates the states of the nodes in the order specified by the permutation. This allows us to capture the stochastic changes in gene expressions that occur in real gene regulatory networks. The random asynchronous Boolean simulations were performed using the Python code provided in [23] which is available at https://gitlab.com/stemcellbioengineering/garuda-boolean.

For example, suppose a gene regulatory network consists of 3 genes, {g1,g2,g3}. The Boolean update functions for the genes are as follows:

$$g_{1} = g_{3} $$
$$g_{2} = g_{1} \lor g_{3} $$
$$g_{3} = g_{2} $$

Suppose an iteration randomly generates a permutation of nodes as {3,1,2}. Then the asynchronous Boolean updates will be carried out as follows:

$$g_{3}(t+1) = g_{2}(t) $$
$$g_{1}(t+1) = g_{3}(t+1) $$
$$g_{2}(t+1) = g_{1}(t+1) \lor g_{3}(t+1) $$

From the above equations, we see that the nodes are updated in a randomly generated order as specified by the permutation, rather than simultaneously.

After performing the simulations for a fixed number of iterations, a directed graph of states is obtained, where each state is a vector representing the expression levels of all genes at a particular time step. The strategy of strongly connected components (SCCs) is employed on this directed graph to capture the dynamic nature of the states [23]. An SCC of a directed graph is a sub-graph that is strongly connected, i.e., each node is reachable from every other node in the sub-graph. An illustration of SCC is given in Fig. 2. Each node is a state with the expression levels of all the genes in the network (for the example we assume a network with five genes) and there is a path between each pair of nodes in both directions. Let us consider that an SCC consists of a set of N states {S1,S2,...,SN}. The probability of state Si being one of the states of the SCC is given by:

$$P(S_{i}) = \frac{\text{ number of occurrences of } S_{i}}{\sum_{j=1}^{N}\text{ number of occurrences of } S_{j}}. $$
Fig. 2
figure 2

Strongly Connected Component. An example of a strongly connected component (SCC). Suppose the network consists of five genes. Then each node is a state which contains the expression levels of the five genes. An arrow from state S1 to state S2 indicates an update step. In an SCC all states can be reached from every other state

Full size image

We calculate the gene expression level of each gene in a particular SCC as the sum of probabilities of states where the gene is in the ON state. Therefore, the expression level of a gene, gi, with respect to an SCC is determined as follows:

$$Exp(g_{i}) = \sum_{S_{j} \in OnSt(g_{i})}P(S_{j}) $$

where

$$OnSt(g_{i}) = \{S_{j} \in SCC \mid g_{i}(S_{j})=1\}. $$

It is easy to see that

$$\sum_{j=1}^{N}P(S_{j}) = 1 $$

We use ER stress, oxidative stress, TNF α, FasL, and IL-6 as input signals. Also, based on the literature, some of the nodes are assigned specific values (Table 3) and the rest are set to random values as initial conditions. We performed simulations using different combinations of the input signals, as shown in Table 4. We carried out 1000 simulation runs and 1000 Boolean update steps per simulation for each input signal. The results of the simulations are presented and discussed in the following section.

Table 3 Initial conditions
Full size table
Table 4 Different combinations for the input signal nodes
Full size table

Due to the lack of experimental data, we validate our proposed Boolean network model using relevant literature (see Table 1). For each gene gi, we use the same symbol gi to represent its binary expression level.

$$ g_{i} \,=\, \begin{cases} 1 & \text{if }g_{i} \text{ is reported as expressed in the literature} \\ 0 & \text{if }g_{i} \text{ is reported as not expressed in the literature} \end{cases} $$

In our model, we determine the expression level of each gene with respect to a particular SCC. Thus the gene expression levels are in the range [0, 1]. We assume that if the expression value of a gene is greater than 0.50, then the gene is expressed, otherwise, it is not expressed.

For the purpose of validating our proposed model, we employ the performance metrics of precision, recall (sensitivity), specificity, and F1 score. The simulation result of our proposed model is verified against the literature as follows. For each gene gi,

$$ g_{i} \in \begin{cases} \text{True positive,} & \text{if }g_{i} = 1 \text{ (simulation result) and }g_{i} = 1 \text{ (literature)} \\ \text{True negative,} & \text{if }g_{i} = 0 \text{ (simulation result) and }g_{i} = 0 \text{ (literature)} \\ \text{False positive,} & \text{if }g_{i} = 1 \text{ (simulation result) and }g_{i} = 0 \text{ (literature)} \\ \text{False negative,} & \text{if }g_{i} = 0 \text{ (simulation result) and }g_{i} = 1 \text{ (literature)} \end{cases} $$

The four evaluation metrics are calculated using the following formulae:

$$Precision = \frac{\textit{True positive}}{\textit{True positive} + \textit{False positive}} $$
$$Recall or sensitivity = \frac{\textit{True positive}}{\textit{True positive} + \textit{False negative}} $$
$$Specificity = \frac{\textit{True negative}}{\textit{True negative} + \textit{False positive}} $$
$$F1\ \ {score} = \frac{2 \times precision \times recall}{precision + recall} $$

Results

Comparison with the literature

The expression levels of genes in the SCCs obtained by performing simulations with our proposed Boolean model are listed in Tables 5 and 6. Simulations performed using input signal cases 1, 2, 3, 4, 5, 7, and 8 (Table 4) result in two attractors (SCCs). Apoptosis is ON in both of the attractors. Simulations performed using input signal case 6 (Table 4) result in six attractors (SCCs). Apoptosis is ON in four attractors and OFF in the remaining two attractors. These observations are consistent with the literature where ER stress, oxidative stress, and cytokines have been shown to cause apoptosis of β-cells individually as well as together [4–6].

Table 5 Gene expressions of the significant genes in the model for input signal cases 1-5 and 7-8
Full size table
Table 6 Gene expressions of the significant genes in the model for input signal case 6. Here A1-A6 denotes SCC1-SCC6
Full size table

From our simulation results, we observe that Caspases 3, 6, 7, 8, and 9, which serve as the final mediators of apoptosis [18] are TRUE in the attractors, even though in the initial condition they were set to FALSE. The ER stress sensor IRE1 and its downstream gene X-box protein binding 1 (XBP1) are TRUE in some attractors, and FALSE in others [26]. Another ER stress sensor, PERK is observed to be FALSE in all the attractors. Also, eukaryotic translation initiation factor 2 subunit 1 (EIF2S1), activating transcription factor 4 (ATF4), and C/EBP homologous protein (CHOP) are TRUE in some attractors and FALSE in the others. PERK phosphorylates and inactivates EIF2S1, which inhibits protein synthesis. Phosphorylated EIF2S1 increases the translation of ATF4 [8], which in turn activates pro-apoptotic CHOP, causing β-cell dysfunction and death [27]. The attractors where IRE1, XBP1, EIF2S1, ATF4, and CHOP have expression levels of 0 may denote the transition states when these genes are not contributing to apoptosis.

While associating with TNF-receptor-associated factor 2 (TRAF2) and apoptosis signal-regulating kinase 1 (ASK1), IRE1 activates jun N-terminal kinase (JNK) [28, 29], which in turn inhibits the anti-apoptotic protein B-cell lymphoma 2 (BCL2) [30]. Oxidative stress activates ASK1 [31, 32], JNK and p38 [33]. Activated p38 phosphorylates and elevates the expression of pro-apoptotic CHOP [34]. From the simulation results, we observe that the pro-apoptotic genes, TRAF2, ASK1, JNK, p38, BAX, and BAK are TRUE and the anti-apoptotic gene BCL2 is FALSE in one attractor, while the reverse states are observed in the other. X-linked inhibitor of apoptosis protein (XIAP), which inhibits Caspases 3, 7, and 9 [35, 36], has an expression level of 0, whereas direct IAP-binding protein with low pI (DIABLO) and high temperature requirement protein A2 (HtrA2), which inhibit XIAP [37], have expression levels of 1.

JNK phosphorylates and inhibits insulin receptor substrate (IRS) [38, 39]. IRS gene is FALSE in both of the attractors. PI3K has an expression level of around 0.50 in all the attractors. Tribbles homolog 3 (TRB3) is induced by ER stress through the ATF4-CHOP pathway [40]. Over-expression of TRB3 inhibits AKT and decreases glucose uptake [41]. TRB3 is TRUE in one attractor and FALSE in the other. AKT has an expression level of 0.50 in both of the attractors. Thus, from the results, we observe that ER stress inhibits the PI3K-AKT signaling pathway and promotes insulin resistance.

Insulin promotes conversion of glucose to glycogen by inhibiting glycogen synthase kinase-3 β (GSK3 β) through the PI3K-AKT signaling pathway, which leads to the activation of glycogen synthase (GS) [42]. From the simulation results, we observe that the expression level of GSK3 β, which inhibits glycogen synthesis through inhibition of GS [42, 43] is approximately 0.49 and that of GS is approximately 0.50. From these simulation results, we can infer that glycogen synthesis is reduced which contributes to insulin resistance.

In T2DM, the mammalian target of rapamycin complex 1 (mTORC1)/ S6 kinase (S6K) signaling is activated [44] leading to the inhibition of IRS [45, 46]. We observe from the simulation results that mTORC1 and S6K have expression levels of 1 thus inhibiting IRS which has an expression of 0. These events cause PI3K and AKT to have low expression levels of approximately 0.50, which in turn reduces glucose uptake through GLUT4 whose expression level is around 0.65.

FOXO1 increases the expression of phosphoenolpyruvate carboxykinase (PEPCK) and glucose-6-phosphatase (G6PC) and thus promotes glucose synthesis [47]. Insulin inhibits the expression of FOXO1 through the activation of the PI3K/AKT signaling pathway, which in turn suppresses PEPCK and G6PC, and thereby reduces glucose synthesis [47–49]. From our simulation results, we observe that FOXO1, PEPCK, and G6PC are FALSE. This could be due to the fact that PI3K and AKT are not completely inactive, though they may have low expression levels, and hence is still able to inhibit the expressions of FOXO1, PEPCK, and G6PC.

In Case 6 where only signal IL6 is active, we observe six attractors (Table 6), of which four indicate apoptosis and two do not. For the attractors where apoptosis is observed, the expression levels of the genes are similar to those mentioned above for the other input signal cases. When apoptosis is not observed, i.e. in the two remaining attractors, the caspases, JNK, BAX, and BAK are FALSE. In one of these two attractors, BCL2 is FALSE and CHOP is TRUE. In the other attractor we observe the reverse expression pattern. Thus, in the presence of only IL-6, apoptosis may or may not be activated.

We further assessed the performance of our proposed Boolean network model by comparing model predictions of gene expressions against the literature. Considering the simulation results obtained using the 8 input signals listed in Table 4, the average precision, recall (sensitivity), specificity, and F1 score obtained for our model are 0.9524, 0.8, 0.875, and 0.8696, respectively. We observe that the validation scores for our model are not very high, maybe because our model is sensitive to some missing interactions.

State transition graphs

Figure 3 shows the state transition graph of the state space generated by simulations conducted using input signal combination given in case 8 (Table 4). The two dense red regions represent the two SCCs where apoptosis is ON. The blue nodes represent states where apoptosis is OFF. Thus from the state transition graph, we observe that, in the presence of all input signals, apoptosis is eventually activated, even though in the initial condition it is set to FALSE.

Fig. 3
figure 3

State Transition Graph 1. State transition graph obtained by simulating our proposed Boolean network model using input signal condition given in Case 8 of Table 4. Simulations generate 2 attractors, both having the Apoptosis node activated. Apoptosis is ON in the red coloured states and OFF in the blue colored states

Full size image

Figure 4 shows the state transition graph of the state space generated by simulations conducted using input signal combination given in case 6 (Table 4). The four dense red regions represent the four SCCs where apoptosis is ON. The two dense blue regions represent the two SCCs where apoptosis is OFF. Thus from the state transition graph, we observe that, in the presence of only IL-6, apoptosis may or may not be activated.

Fig. 4
figure 4

State Transition Graph 2. State transition graph obtained by simulating our proposed Boolean network model using input signal condition given in Case 6 of Table 4. Simulations generate 6 attractors. In four of the attractors Apoptosis is ON, denoted by red colour, and in the remaining two attractors Apoptosis is OFF, denoted by blue colour

Full size image

Comparison with random Boolean networks

We also compared our Boolean network model with random Boolean network models using the 8 input signal combinations given in Table 4. For cases 1, 2, 3, 4, 5, 7, and 8 we found that the number of attractors obtained by simulating the random Boolean networks ranges from 28 to 177, whereas for our Boolean network model the number of attractors is 2. Similarly, for case 6, the number of attractors obtained by simulating the random Boolean networks ranges from 25 to 180, whereas for our Boolean network model the number of attractors is 6. Thus, from the results we observe that the random Boolean networks typically have large numbers of attractors.

Conclusion

In this paper, we proposed a Boolean network model of the integrated insulin resistance and β-cell apoptosis pathways. Such a model, which explores the combined mechanism and interplay between insulin resistance and β-cell apoptosis in the pathogenesis of T2DM, has not been proposed before. We used the model to simulate the dynamics of gene expression induced by different combinations of the five input signals, i.e. ER stress, oxidative stress, and cytokines (TNF α, FasL, IL-6), which serve as triggers for insulin resistance and β-cell apoptosis.

The random order asynchronous update method was employed to perform the simulations, i.e. all nodes were updated in a random order at each update step. We assessed the performance of our model using the metrics of precision, recall (sensitivity), specificity, and F1 score, when validating our model against the literature. The precision score obtained is high, but sensitivity, specificity, and F1 scores are not so. One possible reason may be that some missing interactions affect the predictions of our model. We also compared our Boolean network model with random Boolean network models and observed that random Boolean networks typically have large numbers of attractors ranging from around 25 to 180, whereas our model shows small numbers of attractors ranging from 2 to 6.

As a future step, we can use this model to perform virtual gene knockout experiments to determine genes that play pivotal roles in insulin resistance and/or β-cell apoptosis, and these genes could be further investigated for possible disease interventions.