**Tools to Study Kinases**

[64] Seger D, Seger R, Shaltiel S. The CK2 phosphorylation of vitronectin. Promotion of cell adhesion via the alpha(v)beta 3‐phosphatidylinositol 3‐kinase pathway. Journal of

[65] Stepanova V, Jerke U, Sagach V, Lindschau C, Dietz R, Haller H, Dumler I. Urokinase‐ dependent human vascular smooth muscle cell adhesion requires selective vitro‐ nectin phosphorylation by ectoprotein kinase CK2. Journal of Biological Chemistry.

[66] Schvartz I, Kreizman T, Brumfeld V, Gechtman Z, Seger D, Shaltiel S. The PKA phosphor‐ ylation of vitronectin: Effect on conformation and function. Archives of Biochemistry

[67] Wechsler‐Reya RJ. Caught in the matrix: How vitronectin controls neuronal differentia‐

Biological Chemistry. 2001;**276**:16998‐17006. DOI: 10.1074/jbc.M003766200

2002;**277**:10265‐10272. DOI: 10.1074/jbc.M109057200

152 Protein Phosphorylation

tion. Trends in Neurosciences. 2001;**24**:680‐682

and Biophysics. 2002;**397**:246‐252. DOI: 10.1006/abbi.2001.2699

**Provisional chapter**

#### **Computational Modeling of Complex Protein Activity Networks Networks**

**Computational Modeling of Complex Protein Activity** 

DOI: 10.5772/intechopen.69804

Stefano Schivo, Jeroen Leijten,

Stefano Schivo, Jeroen Leijten, Marcel Karperien and Janine N. Post Marcel Karperien and Janine N. Post Additional information is available at the end of the chapter

Additional information is available at the end of the chapter

http://dx.doi.org/10.5772/intechopen.69804

#### **Abstract**

Because of the numerous entities interacting, the complexity of the networks that regulate cell fate makes it impossible to analyze and understand them using the human brain alone. Computational modeling is a powerful method to unravel complex systems. We recently described the development of a user-friendly computational tool, Analysis of Networks with Interactive MOdeling (ANIMO). ANIMO is a powerful tool to formalize knowledge on molecular interactions. This formalization entails giving a precise mathematical (formal) description of molecular states and of interactions between molecules. Such a model can be simulated, thereby in silico mimicking the processes that take place in the cell. In sharp contrast to classical graphical representations of molecular interaction networks, formal models allow in silico experiments and functional analysis of the dynamic behavior of the network. In addition, ANIMO was developed specifically for use by biologists who have little or no prior modeling experience. In this chapter, we guide the reader through the ANIMO workflow using osteoarthritis (OA) as a case study. WNT, IL-1β, and BMP signaling and cross talk are used as a concrete and illustrative model.

**Keywords:** WNT, IL1β, BMP, cartilage, computational model, ANIMO, cell signaling, network modeling

#### **1. Introduction**

#### **1.1. Signal transduction networks**

At any given point in time, cells are exposed to many different signals from their environment. Cells will have to interpret this multitude of signals they receive. Signal transduction

networks relay and integrate signals from membrane-bound receptors, via protein activation, to the nucleus in order to regulate cellular processes such as gene transcription, metabolism, proliferation, differentiation, and apoptosis (programmed cell death).

Kinases play a key role in signal transduction by transferring phosphate groups to their substrates in a process called phosphorylation [1–4]. In this context, phosphorylation is basically a way to hand over a signal. In practice, kinases function by phosphorylating serine, threonine, or tyrosine residues on downstream substrates, thereby inducing conformational changes and/or charge alterations, resulting in modulation of protein activities [5].

Signal transduction pathways are connected to other signal transduction pathways in a mechanism we call cross talk. Due to this cross talk, signaling pathways are part of extensive signaling networks. Ultimately, dynamic changes in the signaling network determine cell fate. Insight into this network-regulating cell fate is important for controlling stem cell differentiation, understanding diseases such as cancer and osteoarthritis (OA), defining better diagnostics based on biomarker expression, and designing precision therapies.

#### **1.2. Network topology and dynamics**

To understand signaling networks, graphical representations are very useful (and widely used). In such graphical network representations, network topology and protein interactions are displayed in a static way. This is very useful for understanding network topology but fails to show the dynamics of the network interactions. In addition, as networks become large with many interactions between signaling molecules, it becomes harder to comprehend and predict the speed of the network interactions. Since we want to understand the dynamics of signaling networks, we need to incorporate quantitative aspects like activity levels and the timing of interactions. Understanding the interplay between the quantitative dynamics and the distributed and concurrent nature of networks with large numbers of components is a formidable task; this task can only be successfully undertaken by using methods and techniques that are adequately supported by software tools.

#### **1.3. Computational modeling of signaling networks**

The systems biology approach to understanding biological systems starts off from a scientific question and then follows an empirical cycle—or rather a positive spiral—of knowledge/ theory → model → hypotheses → experiments → observations → update and/or refinement of knowledge/theory, until an answer to the original question is found (**Figure 1**). The model plays a pivotal role in this cycle:


An in silico model is always a simplified representation of biological reality and is never the aim in itself. Rather, it is a powerful means in the process of gaining an understanding of the

**Figure 1.** The empirical spiral: applying the empirical cycle in successive rounds leads to a gradual buildup of knowledge.

biological system. Given its role in the empirical cycle, the process of modeling is especially effective when applied by the experts with respect to a certain biological system. Biologists usually have a good sense of cause-and-effect relationships of molecular interactions. In addition, they are the most knowledgeable on the network topology and the dynamics of the biological system they are studying. Since they also benefit most from the generation of hypotheses and from an efficient experimental design, biologists would be the primary candidates to construct models of their research topic.

As models are a formalization of knowledge or theories, an underlying formalism is needed to express this knowledge. Different formal methods have been successfully applied to construct representations of biological systems. Among these methods are Boolean logic [6, 7], ordinary differential equations (ODEs, reviewed in Ref. [8]), interacting state machines [9, 10], process calculi [11, 12], timed automata [13–15], and Petri nets [16, 17]. Most of these formal methods have been implemented into software tools to aid the process of modeling.

Mastery of most existing modeling tools requires training and experience in mathematical modeling. In this respect, a lack of tradition in quantitative reasoning and formal methods within the biological community at large is still a stumbling block for widespread application of modeling of biological systems. To overcome this, we built an intuitive method for the construction of formal in silico models of the dynamics of molecular networks, supported by a user-friendly modeling tool, (Analysis of Networks with Interactive MOdeling (ANIMO) [18]).

#### **1.4. ANIMO**

networks relay and integrate signals from membrane-bound receptors, via protein activation, to the nucleus in order to regulate cellular processes such as gene transcription, metabolism,

Kinases play a key role in signal transduction by transferring phosphate groups to their substrates in a process called phosphorylation [1–4]. In this context, phosphorylation is basically a way to hand over a signal. In practice, kinases function by phosphorylating serine, threonine, or tyrosine residues on downstream substrates, thereby inducing conformational

Signal transduction pathways are connected to other signal transduction pathways in a mechanism we call cross talk. Due to this cross talk, signaling pathways are part of extensive signaling networks. Ultimately, dynamic changes in the signaling network determine cell fate. Insight into this network-regulating cell fate is important for controlling stem cell differentiation, understanding diseases such as cancer and osteoarthritis (OA), defining better diagnostics based on

To understand signaling networks, graphical representations are very useful (and widely used). In such graphical network representations, network topology and protein interactions are displayed in a static way. This is very useful for understanding network topology but fails to show the dynamics of the network interactions. In addition, as networks become large with many interactions between signaling molecules, it becomes harder to comprehend and predict the speed of the network interactions. Since we want to understand the dynamics of signaling networks, we need to incorporate quantitative aspects like activity levels and the timing of interactions. Understanding the interplay between the quantitative dynamics and the distributed and concurrent nature of networks with large numbers of components is a formidable task; this task can only be successfully undertaken by using methods and techniques

The systems biology approach to understanding biological systems starts off from a scientific question and then follows an empirical cycle—or rather a positive spiral—of knowledge/ theory → model → hypotheses → experiments → observations → update and/or refinement of knowledge/theory, until an answer to the original question is found (**Figure 1**). The model

An in silico model is always a simplified representation of biological reality and is never the aim in itself. Rather, it is a powerful means in the process of gaining an understanding of the

changes and/or charge alterations, resulting in modulation of protein activities [5].

proliferation, differentiation, and apoptosis (programmed cell death).

biomarker expression, and designing precision therapies.

that are adequately supported by software tools.

plays a pivotal role in this cycle:

**1.** To organize data and store knowledge **2.** To structure reasoning and discussion

**1.3. Computational modeling of signaling networks**

**3.** To perform in silico experiments and derive hypotheses

**1.2. Network topology and dynamics**

156 Protein Phosphorylation

ANIMO is an activity network tool, built as a plug-in to the network visualization program Cytoscape [19] and founded on the formalism of timed automata [13–15], but does not require the user to have any previous training in formal methods [20]. This provides the advantages of formal models (in silico experiments and model checking) without renouncing to usability.

Nodes in an ANIMO network represent an activity level of any given biological entity, e.g., proteins directly involved in signal transduction (e.g., kinases, growth factors, cytokines, genes, and mRNA. An *activity level* is associated to each node, to represent, for example, the relative amount of phosphorylated kinase or the concentration of mRNA. The activity level of a node can be altered by *interactions* with other nodes. ANIMO networks can include activations (→) and inhibitions (─┤), which will increase (resp. decrease) the activity level of the target node if the source node is active. For example, A → B will increase the activity level of B if A is active. The speed at which an interaction occurs is defined by its *k* parameter, which can be estimated qualitatively by choosing among a predefined set of options (*very slow*, *slow*, *medium*, *fast* and *very fast*) or by directly inputting a numerical value. We note that using the qualitative choices already leads to useful models: choosing, for example, a *slow* interaction to represents the production of a protein, and a *fast* one for a posttranslational modification such as phosphorylation is already enough to provide a realistic behavior in a network with the proper node topology [18, 20, 21].

A finer control on the network dynamics can be obtained by choosing for each interaction an approximated scenario which allows to describe the interaction. A choice is available among three scenarios:


Additionally, the *k* parameter can be manually set to numerical values, expanding the default qualitative choices. Methods for parameter fitting are also present in ANIMO, which allow to automatically adapt the parameters to a given data set [22]. These features are useful when comparing a model to experimental data and allow to easily try different parameter settings before needing to extend a model with new nodes or interactions.

#### **1.5. Experimental requirements**

Biological events can often be interpreted as changes in activity. Activities could be defined as changes in concentration, phosphorylation, or localization of a protein or changes in gene expression because they are causal factors with respect to downstream effects. Therefore, the state or concentration of the molecules can be described in terms of an activity. The more active the molecule is, the stronger it will affect downstream processes.

Experimental design could be performed according to these guidelines:

the user to have any previous training in formal methods [20]. This provides the advantages of formal models (in silico experiments and model checking) without renouncing to usability.

Nodes in an ANIMO network represent an activity level of any given biological entity, e.g., proteins directly involved in signal transduction (e.g., kinases, growth factors, cytokines, genes, and mRNA. An *activity level* is associated to each node, to represent, for example, the relative amount of phosphorylated kinase or the concentration of mRNA. The activity level of a node can be altered by *interactions* with other nodes. ANIMO networks can include activations (→) and inhibitions (─┤), which will increase (resp. decrease) the activity level of the target node if the source node is active. For example, A → B will increase the activity level of B if A is active. The speed at which an interaction occurs is defined by its *k* parameter, which can be estimated qualitatively by choosing among a predefined set of options (*very slow*, *slow*, *medium*, *fast* and *very fast*) or by directly inputting a numerical value. We note that using the qualitative choices already leads to useful models: choosing, for example, a *slow* interaction to represents the production of a protein, and a *fast* one for a posttranslational modification such as phosphorylation is already enough to provide a realistic behavior in a network with

A finer control on the network dynamics can be obtained by choosing for each interaction an approximated scenario which allows to describe the interaction. A choice is available among

• Scenario 1 (default): the interaction rate is linearly dependent on the *k* parameter and the activity level of the upstream node. This is the simplest scenario and is advised for all inter-

• Scenario 2: the interaction rate depends on the *k* parameter and on the activity levels of both nodes. In particular, it is linearly dependent on the activity level of the upstream node and inversely dependent on the activity of the downstream node. This scenario is used to

• Scenario 3 (*AND* gate): the interaction rate depends on the *k* parameter and on the activity level of two user-defined nodes. The user can determine whether the dependency on a node's activity is linear or inverse. This scenario can be used to represent Boolean AND gates, such as "A *AND* B → C," where it is required that both A and B are active in order

Additionally, the *k* parameter can be manually set to numerical values, expanding the default qualitative choices. Methods for parameter fitting are also present in ANIMO, which allow to automatically adapt the parameters to a given data set [22]. These features are useful when comparing a model to experimental data and allow to easily try different parameter settings

Biological events can often be interpreted as changes in activity. Activities could be defined as changes in concentration, phosphorylation, or localization of a protein or changes in gene

model reactions where the availability of substrate is a limiting factor.

before needing to extend a model with new nodes or interactions.

the proper node topology [18, 20, 21].

actions when first building an ANIMO model.

three scenarios:

158 Protein Phosphorylation

for C to become active.

**1.5. Experimental requirements**


We can discern primary (or direct) effects or higher order (or indirect) effects after treatment. Indirect effects are those in which feedback is involved. For signal transduction, the primary effect occurs in time points up to 240/480 minutes. For gene expression, primary effects typically take 4–12 hours. Higher order effects occur at different time ranges, e.g., signal transduction could occur up to 24/48 hours; for gene expression involving higher order effects, for example, in the case of cell differentiation, effects can take up to several weeks.

#### **2. Case study: ANIMO modeling of inflammatory signals in osteoarthritis**

Many diseases are multifactorial, affected by many factors including genetic predisposition, age, trauma, sex, etc. These factors influence the network topology as well as its dynamics. This is hard to capture in static networks. To guide the reader through the ANIMO workflow, we use osteoarthritis as a case study.

Osteoarthritis (OA) is a painful, disabling disease with a high prevalence, occurring in about 15% of the population. The lifetime risk of knee OA is over 40% for man and almost 50% for women (reviewed in Ref. [23]). The lack of insight into the intricate signaling network of the cartilage has prevented the identification of highly needed disease-modifying osteoarthritic drugs (DMOADs). We aim to solve this by generating a comprehensive computational model of the signaling network in the cartilage [24]. In this chapter, we describe three important pathways in cartilage and OA development as a case study.

#### **2.1. Osteoarthritis**

Articular cartilage (AC) is a highly resilient tissue that covers the surfaces at the ends of long bones and ensures the pain-free and supple movement of our joints. The cartilage is mainly composed of one single cell type, the chondrocyte, which secretes and shapes the cartilaginous matrix that is necessary for its load-bearing properties. The biomechanical properties of the cartilage are dependent mainly on the composition, as well as the integrity of its matrix [25]. Once damaged, articular cartilage has low self-repair and regenerative capabilities eventually resulting in OA. This is due to its avascular nature, lack of innervation, and the embedding of chondrocytes in a dense matrix preventing cell migration. In addition, abnormalities in the cartilage-specific matrix cause a variety of skeletal malformation syndromes as well as adult-onset degenerative disorders such as OA.

OA is the most common form of arthritis and a leading cause of mobility-associated disability. OA is characterized by degeneration of articular cartilage, typical bone changes, and signs of inflammation, particularly in end-stage disease. The current management of OA is symptomatic, aimed at reduction of pain and at the end stage of disease total joint replacement as a successful treatment option for large joints (e.g., knee and hip) [26]. These treatments, however, do not cure the disease. There are no systemic drugs that can modify the disease process. Once the cartilage is damaged, no treatment exists that can intervene effectively, and the affected joint enters a disease continuum toward osteoarthritis.

Cartilage tissue homeostasis depends on a fine balance between catabolic (breakdown) and anabolic (buildup) processes. Homeostasis is regulated by a number of signaling pathways, including BMP and WNT signaling [27–31]. The amplitude of the signaling can be fine-tuned via antagonists in the extracellular space (reviewed in Ref. [32]). Typical catabolic pathways include the inflammatory pathways, including TNFα and IL-1β.

#### **2.2. Osteoarthritis at the molecular level**

OA is a disease caused by loss of homeostasis, resulting in altered mechanical and biochemical signals. Some of the key biochemical signals are growth factors such as WNT, IL-1β, transforming growth factor beta, bone morphogenetic proteins (BMPs), and Indian hedgehog homolog [27, 33, 34]. Mechanical stress on the extracellular matrix (ECM) plays a key role in OA development [27, 33, 34]. Any changes in this complex biological system, such as those caused by injury or aging, can disrupt cartilage homeostasis and lead to either catabolism characterized by expression of, for example, matrix metalloproteases (MMPs), and aggrecanases (ADAMs), or anabolism characterized by expression of collagen II and aggrecans [33]. As OA progresses the chondrocytes start to lose their characteristic phenotype, which in some cases results in differentiating into hypertrophic chondrocytes [35], resulting in endochondral ossification, by destroying the surrounding collagen II and replacing it with collagen X. Eventually, the hypertrophic chondrocytes will recruit osteoblasts that will proceed to form an osteophyte [35]. It is important to note that OA is not a disease that will damage the whole joint evenly. Throughout the cartilage there will be cells in different stages of differentiation, ranging from seemingly healthy cells to osteophyte forming hypertrophic chondrocytes.

The direct control of chondrogenic differentiation and hypertrophy is believed to be tightly regulated by the transcriptional activity of two main transcription factors: RUNX2, a transcription factor important for the regulation of hypertrophic differentiation, and SOX9, master transcription factor for chondrogenic development [36, 37]. The exact activity of these factors seems a key in determining the outcome of the chondrocyte phenotype.

The first steps in any computational modeling workflow are to thoroughly investigate the signal transduction pathways involved in the disease and to choose which pathways will be focused on. In this example, we will show the BMP and WNT signaling pathways for their importance in cartilage development and IL-1β as an inflammatory signal involved in osteoarthritis.

#### *2.2.1. WNT signaling in the cartilage and osteoarthritis*

Osteoarthritis (OA) is a painful, disabling disease with a high prevalence, occurring in about 15% of the population. The lifetime risk of knee OA is over 40% for man and almost 50% for women (reviewed in Ref. [23]). The lack of insight into the intricate signaling network of the cartilage has prevented the identification of highly needed disease-modifying osteoarthritic drugs (DMOADs). We aim to solve this by generating a comprehensive computational model of the signaling network in the cartilage [24]. In this chapter, we describe three important

Articular cartilage (AC) is a highly resilient tissue that covers the surfaces at the ends of long bones and ensures the pain-free and supple movement of our joints. The cartilage is mainly composed of one single cell type, the chondrocyte, which secretes and shapes the cartilaginous matrix that is necessary for its load-bearing properties. The biomechanical properties of the cartilage are dependent mainly on the composition, as well as the integrity of its matrix [25]. Once damaged, articular cartilage has low self-repair and regenerative capabilities eventually resulting in OA. This is due to its avascular nature, lack of innervation, and the embedding of chondrocytes in a dense matrix preventing cell migration. In addition, abnormalities in the cartilage-specific matrix cause a variety of skeletal malformation syndromes as well as

OA is the most common form of arthritis and a leading cause of mobility-associated disability. OA is characterized by degeneration of articular cartilage, typical bone changes, and signs of inflammation, particularly in end-stage disease. The current management of OA is symptomatic, aimed at reduction of pain and at the end stage of disease total joint replacement as a successful treatment option for large joints (e.g., knee and hip) [26]. These treatments, however, do not cure the disease. There are no systemic drugs that can modify the disease process. Once the cartilage is damaged, no treatment exists that can intervene effectively, and

Cartilage tissue homeostasis depends on a fine balance between catabolic (breakdown) and anabolic (buildup) processes. Homeostasis is regulated by a number of signaling pathways, including BMP and WNT signaling [27–31]. The amplitude of the signaling can be fine-tuned via antagonists in the extracellular space (reviewed in Ref. [32]). Typical catabolic pathways

OA is a disease caused by loss of homeostasis, resulting in altered mechanical and biochemical signals. Some of the key biochemical signals are growth factors such as WNT, IL-1β, transforming growth factor beta, bone morphogenetic proteins (BMPs), and Indian hedgehog homolog [27, 33, 34]. Mechanical stress on the extracellular matrix (ECM) plays a key role in OA development [27, 33, 34]. Any changes in this complex biological system, such as those caused by injury or aging, can disrupt cartilage homeostasis and lead to either catabolism characterized by expression of, for example, matrix metalloproteases (MMPs), and aggrecanases (ADAMs), or anabolism characterized by expression of collagen II and aggrecans [33].

pathways in cartilage and OA development as a case study.

adult-onset degenerative disorders such as OA.

the affected joint enters a disease continuum toward osteoarthritis.

include the inflammatory pathways, including TNFα and IL-1β.

**2.2. Osteoarthritis at the molecular level**

**2.1. Osteoarthritis**

160 Protein Phosphorylation

The canonical Wnt pathway is crucial for cell survival and OA activation. The canonical pathway is characterized by the axin/glycogen synthase kinase-3β (GSK3-β) destruction complex which maintains low intracellular concentration of the key transcriptional regulator, β-catenin [33]. WNTs bind to the Frizzled (Fz) transmembrane receptors, resulting in the recruitment of the transmembrane protein LRP5/6. This complexation leads to the phosphorylation and dissociation of the destruction complex allowing β-catenin to accumulate and translocate to the nucleus [33]. In turn, β-catenin downregulates both collagen type 2A (*COL2A*) and SRYbox 9 (*SOX9*), leading to cell dedifferentiation and proliferation [38]. The WNT pathway can be activated by IL-1β through the phosphoinositide 3-kinase (PI3K) – protein kinase B (Akt) pathway [39, 40].

#### *2.2.2. BMP2 signaling in osteoarthritis*

BMP2 signaling is key pathway in the development of both the bone and cartilage. In endochondral bone formation, it is responsible for the clustering of the mesenchymal stem cells, the acquisition of the chondrocyte phenotype, and the final differentiation into hypertrophic chondrocytes [41, 42]. This final step is stopped in order to produce adult chondrocytes [41]. BMP2 is found in both healthy and OA adult chondrocytes [42]. BMP2 signaling occurs when BMP2 binds to its type 1 and type 2 receptors, which in turn phosphorylate mothers against decapentaplegic (SMAD) homologs 1, 5, and 8. Subsequently, SMAD 1, 5, or 8 dimers could bind to the ubiquitous SMAD-4 transcription factor [41]. BMP2 signaling can upregulate *Col2a*, *SOX9*, *ColX*, and *MMP13* gene expression [38, 41]. Once OA is advanced, BMP2 can cause hypertrophic differentiation of chondrocytes, leading to osteophyte formation [41].

#### *2.2.3. Interleukin 1β signaling in osteoarthritis*

IL-1β is a key pro-inflammatory cytokine that drives OA progression by inducing the expression of cartilage degrading enzymes such as matrix metalloproteinases (MMPs) [43, 44].

IL-1β signals by binding to the transmembrane IL-1 receptor I (IL-1RI), leading to the activation of multiple signaling pathways. The canonical IL-1β pathway signals through NF-κB, but IL-1β can also activate the p38-MAPK and JNK-MAPK pathways. The activated receptor then assembles two signaling proteins, myeloid differentiation primary response gene 88 (MYD88) and interleukin-1 receptor-activated protein kinase 4 (IRAK4). Together, the proteins form a stable IL-1-induced first signaling module and activate IRAK1 and IRAK2, which in turn activate TRAF6, PELI 1-3, TAK1, and MEKK3 [45]. IRAK1 also activates the inhibitor of nuclear factor B kinase (IKK) complex, which is necessary for the translocation of NF-κB (nuclear factor kappa-light-chain-enhancer of activated B cells) subunits to the core. The IKK complex consists out of IKK1 and IKK2 plus the regulatory subunit NF-κB essential modifier (NEMO) and phosphorylates IkB, the inhibitor of NF-κB, leading to its degradation [45, 46]. Due to the degradation of IkB, two NF-κB subunits, p50 and p65, are released and translocate into the nucleus, where they bind to conserved DNA motifs, which exist in many IL-1β responsive genes, including the genes for IL-1B [45, 46] and IκBα [45].

For the p38 MAPK and the JNK pathways, TAK1 and MEKK3 are mainly responsible, which activate MAPK kinase kinases (MKKs) 3, 4, 6, and 7 [45]. In addition, ERK1/ERK2 is also activated as a result of the activation of the IKK complex, which influences MKK1. All three MAPK pathways influence the activation protein 1 (AP-1), affecting the DNA expression of IL-1β response genes [45].

An increasing amount of data indicates the influence of IL-1β on the degeneration of extracellular matrix in the pathology of OA [47]. In addition, we have previously shown that WNT/βcatenin inhibits IL-1β-induced MMP expression in human articular cartilage [48]. Moreover, we showed that the WNT/β-catenin-regulated transcription factor TCF4 can bind to NF-κB, thereby enhancing NF-κB activity [30].

#### **2.3. Defining an a priori network**

The aim of computational modeling is not to provide the most complete representation of all the interactions in a signaling network, but to use as many interactions as needed to provide insight into cellular mechanisms. As such, models are indeed simplified representations of the real situation: one can choose a level of abstraction depending on the available information and the research question that is asked. The level of abstraction is always a trade-off between precision and feasibility.

In the case study presented here, we do not aim to build a precise model, but aim to show how building a model enables researchers of all levels of experience to visualize, summarize, and formalize models. Generating a relatively simple model in which key interactions are shown enables researchers to test and discuss various hypotheses quickly. With the obtained insight, one can then choose to validate only those hypotheses that the researchers will expect to truly yield new information. The model is then used as a backbone for the smart design of wet lab experiments rather than the trial-and-error methods that are traditionally used in the field.

#### *2.3.1. Defining a research question*

*2.2.3. Interleukin 1β signaling in osteoarthritis*

162 Protein Phosphorylation

IL-1β response genes [45].

thereby enhancing NF-κB activity [30].

**2.3. Defining an a priori network**

precision and feasibility.

IL-1β is a key pro-inflammatory cytokine that drives OA progression by inducing the expression of cartilage degrading enzymes such as matrix metalloproteinases (MMPs) [43, 44].

IL-1β signals by binding to the transmembrane IL-1 receptor I (IL-1RI), leading to the activation of multiple signaling pathways. The canonical IL-1β pathway signals through NF-κB, but IL-1β can also activate the p38-MAPK and JNK-MAPK pathways. The activated receptor then assembles two signaling proteins, myeloid differentiation primary response gene 88 (MYD88) and interleukin-1 receptor-activated protein kinase 4 (IRAK4). Together, the proteins form a stable IL-1-induced first signaling module and activate IRAK1 and IRAK2, which in turn activate TRAF6, PELI 1-3, TAK1, and MEKK3 [45]. IRAK1 also activates the inhibitor of nuclear factor B kinase (IKK) complex, which is necessary for the translocation of NF-κB (nuclear factor kappa-light-chain-enhancer of activated B cells) subunits to the core. The IKK complex consists out of IKK1 and IKK2 plus the regulatory subunit NF-κB essential modifier (NEMO) and phosphorylates IkB, the inhibitor of NF-κB, leading to its degradation [45, 46]. Due to the degradation of IkB, two NF-κB subunits, p50 and p65, are released and translocate into the nucleus, where they bind to conserved DNA motifs, which exist in many

For the p38 MAPK and the JNK pathways, TAK1 and MEKK3 are mainly responsible, which activate MAPK kinase kinases (MKKs) 3, 4, 6, and 7 [45]. In addition, ERK1/ERK2 is also activated as a result of the activation of the IKK complex, which influences MKK1. All three MAPK pathways influence the activation protein 1 (AP-1), affecting the DNA expression of

An increasing amount of data indicates the influence of IL-1β on the degeneration of extracellular matrix in the pathology of OA [47]. In addition, we have previously shown that WNT/βcatenin inhibits IL-1β-induced MMP expression in human articular cartilage [48]. Moreover, we showed that the WNT/β-catenin-regulated transcription factor TCF4 can bind to NF-κB,

The aim of computational modeling is not to provide the most complete representation of all the interactions in a signaling network, but to use as many interactions as needed to provide insight into cellular mechanisms. As such, models are indeed simplified representations of the real situation: one can choose a level of abstraction depending on the available information and the research question that is asked. The level of abstraction is always a trade-off between

In the case study presented here, we do not aim to build a precise model, but aim to show how building a model enables researchers of all levels of experience to visualize, summarize, and formalize models. Generating a relatively simple model in which key interactions are shown enables researchers to test and discuss various hypotheses quickly. With the obtained insight, one can then choose to validate only those hypotheses that the researchers will expect to truly yield new information. The model is then used as a backbone for the smart design of wet lab experiments rather than the trial-and-error methods that are traditionally used in the field.

IL-1β responsive genes, including the genes for IL-1B [45, 46] and IκBα [45].

We base our research question on our analysis of the role of the individual pathways and their possible cross talk on OA development. Although the roles of WNT and BMP in cartilage and OA development are well described, the precise interactions between WNT, BMP, and IL-1β in regulating OA and chondrocyte hypertrophy are not fully understood (reviewed in Ref. [31]).

The model can be used for rapid and reiterative queries to derive and probe hypotheses such as whether IL-1β could influence cartilage homeostasis by modulating the activity of the cartilage and bone transcription factors SOX9 and RUNX2 and their downstream targets. Similarly, the role of BMP and WNT signalings, two important pathways in cartilage development and maintenance, can be explored on their potential modulatory roles on IL-1β expression and function.

To explore this example and provide a concrete guide through the ANIMO workflow, we will first draw a priori knowledge network, then formalize this network based on literature and our own data, and then perform a few simple in silico experiments that will be validated in the wet lab.

#### *2.3.2. Drawing a priori knowledge network*

To build an a priori network, we use KEGG (www.genome.jp/kegg) and WikiPathways (wikipathways.org) to decide on the topology of the proteins in the network.

We first draw an a priori network diagram that includes some of the most important factors in the signaling pathways of interest: IL-1β (based on: WikiPathways WP2637), WNT (e.g., the WNT homepage [49]), and BMP (e.g., WP1425). We include those intracellular molecules that we can actually measure in our experiments (see Sections 1.5 and 2.6), and added dashed lines to indicate other molecules important for the pathway were omitted (see Section 2.2).

In the WNT pathway, the inhibition of the destruction pathway leads to the inhibition of destruction of β-catenin, resulting in its upregulation and nuclear translocation [32]. So the net effect of WNT binding to its receptor is the increase of β-catenin activity. To simplify the model, we omit the many steps involving the double inhibition mechanisms in the WNT pathway, resulting in the simplified path WNT → WNTR → GSK3 → β-catenin → TCF/LEF.

SOX9 regulates expression of the matrix proteins collagen 2 and aggrecan. RUNX2 regulates transcription of collagens I and X and MMP13 [50, 51]. Since the activity of SOX9 and RUNX2 is key to the switch from the cartilage to hypertrophic cartilage, we included SOX9 and RUNX2 and some of their target genes in our diagram (**Figure 2**, **3A**).

#### **2.4. Adding dynamics to the network**

Once a priori knowledge network has been chosen, it needs to be drawn in ANIMO as a collection of nodes and interactions. For each node, a maximum number of activity levels can be chosen: unless a model is extremely large, it is safe to use 100 levels for all nodes. After providing the node with a name and an initial activity (which describes the state of the node at the start of a simulation), a description can optionally be added. Descriptions can be used as rationale for the presence of a certain node in the model, for example, a node description can contain citations to literature or references to own experimental results.

When adding an interaction between two nodes, a choice for an approximation scenario and a *k* parameter is necessary. The *k*-constants in our a priori knowledge network are not taken from literature as the strength of all interactions is assumed to be equal. This is a pragmatic decision, as many actual *k*-values are not described for most of the protein interactions in our network. In our initial models, we assume that there are in general two types of reactions: fast reactions for posttranslational modifications, such as phosphorylation, and slow reactions where gene transcription occurs. We therefore add reactions between nodes using these two types of reaction speed with a "Scenario 1" setting. We also add auto-inhibition to indicate inhibition as described in the literature for, e.g., receptor internalization, phosphatase activity, and, in the case of NF-κB, nuclear export as regulated by IκB. A more in-depth discussion on parameters, scenario choices, and network topology can be found in Refs. [21, 22].

Based on our experience, we expect proteins directly downstream of an activated receptor, such as p38 and JNK that are downstream of IL-1R, to be most activated by phosphorylation about 15–30 minutes after stimulation, and that the activity would be decreased to the starting situation between 60 and 240 minutes after stimulation. We therefore set the initial parameters to match these assumptions.

**Figure 2.** A priori knowledge network of WNT, BMP, and IL-1β pathways. IL-1β canonical (blue) and noncanonical signaling (green) showing cross talk with the transcription factors RUNX and SOX9 (both light blue), transcription factors from the WNT (red), and BMP2 (purple) pathways. Solid arrows indicate direct protein interaction, and dashed arrows indicate that intermediate protein interaction is omitted because no cross talk occurs between these proteins with other pathways. The colors indicate the canonical and noncanonical pathways corresponding to the external signals WNT, BMP2, and IL-1β.

#### **2.5. Testing network effects of different stimuli in silico**

When adding an interaction between two nodes, a choice for an approximation scenario and a *k* parameter is necessary. The *k*-constants in our a priori knowledge network are not taken from literature as the strength of all interactions is assumed to be equal. This is a pragmatic decision, as many actual *k*-values are not described for most of the protein interactions in our network. In our initial models, we assume that there are in general two types of reactions: fast reactions for posttranslational modifications, such as phosphorylation, and slow reactions where gene transcription occurs. We therefore add reactions between nodes using these two types of reaction speed with a "Scenario 1" setting. We also add auto-inhibition to indicate inhibition as described in the literature for, e.g., receptor internalization, phosphatase activity, and, in the case of NF-κB, nuclear export as regulated by IκB. A more in-depth discussion on

parameters, scenario choices, and network topology can be found in Refs. [21, 22].

to match these assumptions.

164 Protein Phosphorylation

WNT, BMP2, and IL-1β.

Based on our experience, we expect proteins directly downstream of an activated receptor, such as p38 and JNK that are downstream of IL-1R, to be most activated by phosphorylation about 15–30 minutes after stimulation, and that the activity would be decreased to the starting situation between 60 and 240 minutes after stimulation. We therefore set the initial parameters

**Figure 2.** A priori knowledge network of WNT, BMP, and IL-1β pathways. IL-1β canonical (blue) and noncanonical signaling (green) showing cross talk with the transcription factors RUNX and SOX9 (both light blue), transcription factors from the WNT (red), and BMP2 (purple) pathways. Solid arrows indicate direct protein interaction, and dashed arrows indicate that intermediate protein interaction is omitted because no cross talk occurs between these proteins with other pathways. The colors indicate the canonical and noncanonical pathways corresponding to the external signals Testing the effect of different stimuli in a computational network is performed in small steps. During each step the parameters in the model can be updated so that the dynamics of the various nodes in the network match our knowledge.

Step 1. What is the normal "steady-state" (=No Input) situation of the nodes represented in the network? For example, for articular chondrocytes it is known that the transcription factor SOX9 is active and that collagen 2 and aggrecan are expressed. It is also known that the WNT and IL-1β pathways are inactive and that BMP is active at a low level [52, 53]. We therefore can adjust our starting activities to these settings. This is generated in **Figure 4**. We display the activities of the proteins of which we plan to measure the phosphorylation, ERK1/2, GSK3, JNK, and p38 as well as the gene expression of AXIN2 and COL2A1. Initially, COL2A is expressed, indicating SOX9 activity.

Step 2. In a first test of the response of cells to various stimuli, we tested the presence of WNT starting from our "steady-state" model generated as described in Step 1. Since we do not starve our cells in the experiment, BMP2 is active at a low level of 20 activity units. After WNT addition we see that GSK3 becomes activated, and the activity peaks between 30 and 60 minutes after WNT addition and then trails off around 400 minutes. We see that AXIN2 becomes present between 2 and 4 hours after WNT treatment. This is probably faster than what can be expected from a newly synthesized mRNA. At the same time, due to the inactivation of SOX9 by β-catenin [54], we observe a reduction in the COL2A1 expression around 2–3 hours after WNT addition (not shown).

Step 3. We then tested the presence of IL-1β starting from our "steady-state" model. Again, BMP is active at a level of 20. We now observe that within 15 minutes of IL-1β addition, the three downstream kinases ERK1/2, p38, and JNK become active. In turn, these kinases activate RUNX2 [55], thereby activating its target genes. Due to the negative feedback loop from RUNX2 to JNK and p38, the activity curve is more narrow for these proteins than it is for ERK1/ERK2 [56]. GSK3 becomes slightly active via AKT activation by IL-1β. We see no reduction of COL2A mRNA expression and a transient activation of COL1 and COLX mRNA expression.

Step 4. Next, we want to see the effect of dual stimulation of WNT and IL-1β when starting from a healthy situation as described under Step 1. Addition of WNT and IL-1β, in the presence of 20% BMP, decreases the activity of SOX9 and therefore causes loss of COL2A1 expression. At the same time, RUNX2 becomes active, thereby activating MMP13, COL1, and COLX, ultimately leading to a hypertrophic phenotype.

Step 5. The next question was what is the effect of IL-1β and WNT in the presence of high BMP activity? In our model, the presence of high BMP activity is enough to prevent the loss of COL2A expression (not shown), while at the same time leading to induction of RUNX2 activity and the corresponding COL1 and COLX expression. However, we would have expected that the high levels of WNT and IL-1β would lead to reduced SOX9 activity as seen in articles describing OA (reviewed in Ref. [31]). We therefore need to adapt our network to match the literature data.

Step 6. It has been described that SOX9 and β-catenin influence each other's activity [54]. Also, RUNX2 suppresses SOX9 in bone formation [57], while SOX9 suppresses RUNX2 activity [36]. We added mutual inhibitions between SOX9 and RUNX2. This showed that when RUNX2 becomes active, it suppresses the SOX9-induced COL2 expression.

These few in silico experiments provide insight into the possible cross talk between these three pathways and their effect on SOX9 and RUNX2 activity and the possible effects of upstream signals. While asking the questions, we modified the parameters of the reactions, added inhibitory loops, and checked signaling cross talk in order to match the reaction speed in the model with our literature data or own experience.

This initial model now allows us to investigate the mechanism of IL-1β in inducing cellular hypertrophy by directly regulating SOX9 and RUNX2 activity. We hypothesize that IL-1β will increase expression of hypertrophic genes by upregulating RUNX2 activity and downregulating SOX9 activity (**Figure 3**).

#### **2.6. Testing hypothesis by wet lab experiments**

#### *2.6.1. Designing experiment*

The outcomes of the in silico experiments are used as guideline for the experimental setup for the wet lab validation. In this case, we questioned the effect of IL-1β on WNT signaling in the presence or absence of BMP signaling in the cartilage. For this we stimulated cells with IL-1β, WNT3A, or BMP2 either alone or in combinations. Parts of the data used in this chapter are published previously [21]. The other raw data can be obtained upon request.

#### *2.6.2. Wet lab data*

After the creation of the initial model with defined nodes, the next step was to obtain the data to fit into the model. This step was carried out mainly with wet lab data complemented with literature data. It is important that the analyses of wet lab data show consistency with well-known osteoarthritic cellular responses; these analyses are done prior to inclusion of experimental data into the model.

**Figure 4** shows the measured and predicted activity of the various proteins of our network. We observe attenuations in, for example, p38, JNK, and ERK phosphorylation, where WNT partially inhibited the effect of IL-β on the phosphorylation of these proteins. We have already described these data [21]. When we add BMP to the cells, in combination with IL-1β and WNT, we see that in addition to the lower activity of the proteins, there is also a delay in the time by which the maximum activity is reached. This results in a delay and a reduction of the level of gene expression of all genes tested.

#### *2.6.3. Comparing wet lab data to in silico data*

The wet lab data obtained from the experiment were normalized and rescaled from 0 to 100 in order to be comparable with ANIMO's simulation results. For the complete normalization

Computational Modeling of Complex Protein Activity Networks http://dx.doi.org/10.5772/intechopen.69804 167

Step 6. It has been described that SOX9 and β-catenin influence each other's activity [54]. Also, RUNX2 suppresses SOX9 in bone formation [57], while SOX9 suppresses RUNX2 activity [36]. We added mutual inhibitions between SOX9 and RUNX2. This showed that when

These few in silico experiments provide insight into the possible cross talk between these three pathways and their effect on SOX9 and RUNX2 activity and the possible effects of upstream signals. While asking the questions, we modified the parameters of the reactions, added inhibitory loops, and checked signaling cross talk in order to match the reaction speed

This initial model now allows us to investigate the mechanism of IL-1β in inducing cellular hypertrophy by directly regulating SOX9 and RUNX2 activity. We hypothesize that IL-1β will increase expression of hypertrophic genes by upregulating RUNX2 activity and downregulat-

The outcomes of the in silico experiments are used as guideline for the experimental setup for the wet lab validation. In this case, we questioned the effect of IL-1β on WNT signaling in the presence or absence of BMP signaling in the cartilage. For this we stimulated cells with IL-1β, WNT3A, or BMP2 either alone or in combinations. Parts of the data used in this chapter are

After the creation of the initial model with defined nodes, the next step was to obtain the data to fit into the model. This step was carried out mainly with wet lab data complemented with literature data. It is important that the analyses of wet lab data show consistency with well-known osteoarthritic cellular responses; these analyses are done prior to inclusion of

**Figure 4** shows the measured and predicted activity of the various proteins of our network. We observe attenuations in, for example, p38, JNK, and ERK phosphorylation, where WNT partially inhibited the effect of IL-β on the phosphorylation of these proteins. We have already described these data [21]. When we add BMP to the cells, in combination with IL-1β and WNT, we see that in addition to the lower activity of the proteins, there is also a delay in the time by which the maximum activity is reached. This results in a delay and a reduction of the

The wet lab data obtained from the experiment were normalized and rescaled from 0 to 100 in order to be comparable with ANIMO's simulation results. For the complete normalization

published previously [21]. The other raw data can be obtained upon request.

RUNX2 becomes active, it suppresses the SOX9-induced COL2 expression.

in the model with our literature data or own experience.

**2.6. Testing hypothesis by wet lab experiments**

ing SOX9 activity (**Figure 3**).

166 Protein Phosphorylation

*2.6.1. Designing experiment*

*2.6.2. Wet lab data*

experimental data into the model.

level of gene expression of all genes tested.

*2.6.3. Comparing wet lab data to in silico data*

**Figure 3.** The ANIMO model built to represent the cross talk between the WNT, BMP, and IL-1β pathways. (A) The initial configuration of the model. (B) An example of simulation in ANIMO: activity levels of all nodes are shown after 120 minutes of treatment with WNT + IL-1β using color coding. The node colors are indicative of their activity at the indicated time points, with green being most active, via yellow to red, which is inactive as indicated in the figure, bottom left.

**Figure 4.** Comparing ANIMO's results with wet lab data: signal transduction. The results from two versions of the model are shown: the initial version with qualitative parameters (initial model) and the one obtained with ANIMO's automatic parameter fitting feature (fitted model). Colors are indicative of activity with green being most active and red, via yellow, inactive (see Legend).

procedure, we refer to Ref. [20] and ANIMO's manual. The resulting.csv tables, together with the model, can be found online, in the link below: http://fmt.cs.utwente.nl/tools/animo/ content/models/Phosphorylation.

To evaluate the accuracy of the model, we compared its simulation results against the wet lab data concentrating at first on the signal transduction part of the network (**Figure 4**, first two rows). The initial match was already quite close, even if the parameters used in the model were all of qualitative nature (**Table 1**). The heat map graphs we show in **Figure 4** were obtained in ANIMO, where wet lab data can be directly compared to simulation results.

Another feature provided by ANIMO is parameter fitting, which allows to automatically try different parameter values for the interactions in a model, comparing the simulations with a given data set. This lets the researcher check whether the model topology can be a plausible explanation of the reference wet lab data. In case no parameter set can satisfyingly match the given data, or if only a very narrow parameter choice fits well, it is likely necessary to try a different wiring of the network model.

#### *2.6.4. Optimizing model*

Our next step was to use ANIMO's automatic parameter fitter on the model, using the wet lab data as reference. We divided the model in two subnetworks roughly corresponding to


procedure, we refer to Ref. [20] and ANIMO's manual. The resulting.csv tables, together with the model, can be found online, in the link below: http://fmt.cs.utwente.nl/tools/animo/

**Figure 4.** Comparing ANIMO's results with wet lab data: signal transduction. The results from two versions of the model are shown: the initial version with qualitative parameters (initial model) and the one obtained with ANIMO's automatic parameter fitting feature (fitted model). Colors are indicative of activity with green being most active and red,

To evaluate the accuracy of the model, we compared its simulation results against the wet lab data concentrating at first on the signal transduction part of the network (**Figure 4**, first two rows). The initial match was already quite close, even if the parameters used in the model were all of qualitative nature (**Table 1**). The heat map graphs we show in **Figure 4** were obtained in

Another feature provided by ANIMO is parameter fitting, which allows to automatically try different parameter values for the interactions in a model, comparing the simulations with a given data set. This lets the researcher check whether the model topology can be a plausible explanation of the reference wet lab data. In case no parameter set can satisfyingly match the given data, or if only a very narrow parameter choice fits well, it is likely necessary to try a

Our next step was to use ANIMO's automatic parameter fitter on the model, using the wet lab data as reference. We divided the model in two subnetworks roughly corresponding to

ANIMO, where wet lab data can be directly compared to simulation results.

content/models/Phosphorylation.

via yellow, inactive (see Legend).

168 Protein Phosphorylation

different wiring of the network model.

*2.6.4. Optimizing model*


**Table 1.** Parameters for WNT, BMP, and IL-1β signaling in the initial model (k\*) and the fitted model that was optimized using experimental data (k\*\*; see Section 2.6.4).

the WNT and IL-1 pathways. The two subnetworks were independently fit, based on the wet lab data for the treatments with WNT and IL-1. Dividing the model allowed us to limit the parameter space for the automatic search, making it more rapid: it took less than 3 minutes to complete. In practice, to fit a part of the ANIMO model, we disabled the part of the network we were not focusing on. We then selected those interactions whose parameters we wanted to optimize and clicked on the "Optimize *k*-Values" command in ANIMO's interface. After providing the proper file with the wet lab data, we let the tool to automatically try different *k*-values for the interactions, comparing the model simulations with the data and determining the fitness. Once the tool could find a better fitting set of parameters, the process would terminate, showing the resulting match for the candidate parameter set. We repeated the same procedure on both WNT and IL-1 pathways, fitting the signal transduction parts to the data. We then simulated the WNT + IL-1 treatment in the model and compared it to the data, finding it was fitting already well enough (see **Figure 4**, first and last row).

**Interaction k\* Qualitative parameter Scenario k\*\***

170 Protein Phosphorylation

JNK --| JNK 0.004 Medium 1 0.00055435 MMP13 --| MMP13 0.002 Slow 1 0.002 NF-κB --> IkB 0.016 Very fast 1 0.016 NF-κB --> IL-1B 0.002 Slow 1 0.00156559 NF-κB --> MMP13 0.001 Very slow 1 0.001 p38 --> RUNX2 0.004 Medium 1 0.00399137 p38 --| p38 0.004 Medium 1 0.04260553 RUNX2 --> COL1 0.002 Slow 1 0.002 RUNX2 --> COLX 0.002 Slow 1 0.002 RUNX2 --> MMP13 0.001 Very slow 1 0.001 RUNX2 --| JNK 0.004 Medium 1 0.00606149 RUNX2 --| p38 0.004 Medium 1 0.00406615 RUNX2 --| RUNX2 0.004 Medium 1 0.01181898 RUNX2 --| SOX9 0.004 Medium 1 0.004 SMAD1/5/8 --> ID1 0.002 Slow 1 0.00270426 SMAD1/5/8 --> RUNX2 0.008 Fast 1 0.008 SMAD1/5/8 --> SMAD7 0.004 Medium 1 0.00671529 SMAD1/5/8 --> SOX9 0.004 Medium 1 0.004 SMAD7 --| SMAD1/5/8 0.008 Fast 2 0.02130379 SOX9 --> ACAN 0.001 Very slow 1 0.001 SOX9 --> COL2A 0.001 Very slow 1 0.001 SOX9 --| beta-catenin 0.001 Very slow 1 0.00028855 SOX9 --| RUNX2 0.004 Medium 1 8.7016E-05 SOX9 --| SOX9 0.002 Slow 1 0.00424899 SOX9-activating signal --> SOX9 0.016 Very fast 1 0.00255911 TCF/LEF --> AXIN2 0.001 Very slow 1 0.00021457 TCF/LEF --| TCF/LEF 0.004 Medium 1 0.004 WNT --> WNTR 0.008 Fast 1 0.02304019 WNTR --> AKT 0.016 Very fast 1 0.04626554 WNTR --> GSK3 0.008 Fast 1 0.00372738 WNTR --> WNTR internalization 0.004 Medium 1 0.00600663 WNTR internalization --| WNTR 0.016 Very fast 1 0.06497972

**Table 1.** Parameters for WNT, BMP, and IL-1β signaling in the initial model (k\*) and the fitted model that was optimized

using experimental data (k\*\*; see Section 2.6.4).

For BMP2 signaling we optimized SMAD activity based on phosphorylation of Western blot of SMAD1/5/8. SMAD1/5/8 was most active at 15 minutes posttreatment (data not shown).

Finally, we compared the model with the polymerase chain reaction (PCR) experimental data, repeating the fitting process only on that part of the network. The resulting model parameters can be found in **Table 1** (k\*\*), while the comparison between PCR data and ANIMO mRNA node activities is shown in **Figure 5**. We note that the general trends were captured in the model. Differences between activities in the model vs. the experimental data are especially visible in the longer time scales, around 24 hours. This can be expected because higher order effects, such as feedback loops, which take place on longer time scales, are not included in our model.

**Figure 5.** Comparing ANIMO's results with wet lab data: protein production. The activities of nodes representing mRNA in the ANIMO model (bottom panel) were compared against wet lab PCR data for actin, Col2a, ID1, and IL-1 (top panels).

#### *2.6.5. Validating hypothesis*

Our hypothesis that IL-1β will increase expression of hypertrophic genes by upregulating RUNX2 activity and downregulating SOX9 activity can now be tested in our optimized model. For this, we can investigate, in silico, the changes in activity of SOX9 and RUNX2 and their corresponding genes, even when no gene expression is measured in the wet lab.

Our model predicts that in the presence of IL-1β and WNT, SOX9 activity will be inhibited and RUNX2 will be activated with an initial peak activity in the first hour (**Figure 6**). This indicates loss of cartilage homeostasis and a slight increase in hypertrophy, which is sustained in time due to the permanent increase in RUNX2 activity (not shown). So even though the initial WNT and IL-1 signal are no longer present, an increase in RUNX2 activity results in a sustained expression of collagen 1 and collagen 10 as well as MMP13, albeit at a low level. Even high levels of BMP2 cannot prevent the loss of SOX9 activity, eventually leading to hypertrophy. These data validate our hypothesis, at least in silico.

**Figure 6.** In silico validation of hypothesis: IL-1β will increase expression of hypertrophic genes, *COL1*, *COL10*, and *MMP13*, by upregulating RUNX2 activity and downregulating SOX9 activity.

#### **3. Discussion and conclusion**

In this chapter we provide an example of a workflow for starting computational modeling based on literature and experimental data. The aim was not to make the most comprehensive model in terms of network topology, but to understand the dynamics of the network activity in terms of signaling cross talk and corresponding downstream effects.

We chose to use the software ANIMO as a plug-in in Cytoscape as it offers a user-friendly interface in which biologists can interactively create and explore computational models of signal transduction networks. This allows to gain intellectual control over the dynamic behavior of the network that is modeled. We showed that network topologies can be constructed, modified, and enhanced with a formal description of the associated dynamic behavior. The process of modeling biological network dynamics is a prerequisite for formally comparing experimental data to a priori knowledge. ANIMO can also be used in research groups to assist in the storage and transfer of knowledge on biological networks and as a guide in discussions.

Most of the plug-ins for Cytoscape are based on static analyses, for example, they make it possible to find the hubs in a network, to cluster nodes by specific features, or to associate external data sources to the network. This allows to effectively represent large quantities of information and obtain useful insight from them, but the focus is still on the "static picture." ANIMO concentrates instead on the network dynamics: applying an abstract representation of biochemical kinetics, it allows to represent how signaling networks evolve with time under different conditions. Graphs and node colors provide the user with useful representations of the network dynamics. Further analyses are enabled by the possibility to perform model checking on the underlying timed automata model, which can be used without the need to acquire additional training in formal methods.

ANIMO describes biological entities in the network in terms of their activity. This generalizes easily into most signal transduction processes. However, this can also be used to model any process that can be abstractly modeled as a variable activity. Examples are the inclusion of processes such as receptor internalization and phosphatase activity but also inhibition of an activated protein by proteosomal degradation or nuclear export. This flexibility helps the user in describing parts of the network for which the molecular details are unknown or of lesser importance.

In the model presented here, we show how a priori knowledge network based on three signaling pathways can be constructed and tested in silico by asking questions in small steps at a time. We then showed how experimental data of a limited number of proteins and genes, at a wide range of time points, aid to optimize topology and dynamics of the proteins and mRNAs in our network. In the next step, we can prioritize and design new experiments that can be validated in the wet lab. Seeing the role of a computational model in the empirical spiral in **Figure 1**, the work is never finished, but each step in the cycle aids to optimize the model and hence the molecular insight into the dynamics and topology of the cellular signal transduction network.

**Figure 6.** In silico validation of hypothesis: IL-1β will increase expression of hypertrophic genes, *COL1*, *COL10*, and

In this chapter we provide an example of a workflow for starting computational modeling based on literature and experimental data. The aim was not to make the most comprehensive model in terms of network topology, but to understand the dynamics of the network activity

We chose to use the software ANIMO as a plug-in in Cytoscape as it offers a user-friendly interface in which biologists can interactively create and explore computational models of signal transduction networks. This allows to gain intellectual control over the dynamic behavior of the network that is modeled. We showed that network topologies can be constructed, modified, and enhanced with a formal description of the associated dynamic behavior. The

Our hypothesis that IL-1β will increase expression of hypertrophic genes by upregulating RUNX2 activity and downregulating SOX9 activity can now be tested in our optimized model. For this, we can investigate, in silico, the changes in activity of SOX9 and RUNX2 and

Our model predicts that in the presence of IL-1β and WNT, SOX9 activity will be inhibited and RUNX2 will be activated with an initial peak activity in the first hour (**Figure 6**). This indicates loss of cartilage homeostasis and a slight increase in hypertrophy, which is sustained in time due to the permanent increase in RUNX2 activity (not shown). So even though the initial WNT and IL-1 signal are no longer present, an increase in RUNX2 activity results in a sustained expression of collagen 1 and collagen 10 as well as MMP13, albeit at a low level. Even high levels of BMP2 cannot prevent the loss of SOX9 activity, eventually leading to

their corresponding genes, even when no gene expression is measured in the wet lab.

*MMP13*, by upregulating RUNX2 activity and downregulating SOX9 activity.

in terms of signaling cross talk and corresponding downstream effects.

**3. Discussion and conclusion**

hypertrophy. These data validate our hypothesis, at least in silico.

*2.6.5. Validating hypothesis*

172 Protein Phosphorylation

In ANIMO we proved our hypothesis that IL-1β will increase expression of hypertrophic genes by upregulating RUNX2 activity and downregulating SOX9 activity. For this we used a combination of literature and experimental data to optimize the model parameters. This allowed us to obtain insight into the order of events in the presence of WNT and/or IL-1β at the level of SOX9 and RUNX2 activity. In addition, it allowed insight into the complex interconnectivity of three individual pathways. Such models also yield high content data at high temporal resolution, a feat that is difficult to achieve using only wet lab approaches.

Interestingly, in one computational model, we are able to show a combination of protein activity (phosphorylation) and subsequent mRNA expression. This is a combination model of events at very different time lines. The advantage of our strategy, which included automatic parameter fitting, is the possibility to predict cell fate based on both changes in phosphorylation/protein activity and corresponding gene expression differences.

In the future, ANIMO and related tools may lead to a new paradigm for interactive representation of biological networks. Networks in digital textbooks and articles could be displayed as animations amenable to modifications by readers. Repositories of formal descriptions of signaling modules could be used to put together executable signaling networks. A more user-friendly way of interacting with dynamic network models will lead to a more thorough understanding of biological networks and will accelerate hypothesisdriven research.

#### **Acknowledgements**

We thank Patrick Kaufhold and Dr. Jacqueline Plass for the technical assistance. We thank Ricardo Urquidi Camacho for the technical assistance and valuable discussion during the time that he was an MSc graduate student in our lab in 2012 and 2013. We thank Kannan Govindaraj for critically reading the manuscript and for the valuable discussion.

#### **Author details**

Stefano Schivo1 , Jeroen Leijten<sup>2</sup> , Marcel Karperien2 and Janine N. Post<sup>2</sup> \*


#### **References**


[7] Shmulevich I, Zhang W. Binary analysis and optimization-based normalization of gene expression data. Bioinformatics. 2002;**18**(4):555-565

In the future, ANIMO and related tools may lead to a new paradigm for interactive representation of biological networks. Networks in digital textbooks and articles could be displayed as animations amenable to modifications by readers. Repositories of formal descriptions of signaling modules could be used to put together executable signaling networks. A more user-friendly way of interacting with dynamic network models will lead to a more thorough understanding of biological networks and will accelerate hypothesis-

We thank Patrick Kaufhold and Dr. Jacqueline Plass for the technical assistance. We thank Ricardo Urquidi Camacho for the technical assistance and valuable discussion during the time that he was an MSc graduate student in our lab in 2012 and 2013. We thank Kannan

and Janine N. Post<sup>2</sup>

\*

Govindaraj for critically reading the manuscript and for the valuable discussion.

, Marcel Karperien2

1 Formal Methods and Tools, CTIT, University of Twente, Enschede, The Netherlands

2 Developmental BioEngineering, MIRA, University of Twente, Enschede, The Netherlands

[1] Hunter T. Protein kinases and phosphatases: The yin and yang of protein phosphoryla-

[2] Cohen P. Targeting protein kinases for the development of anti-inflammatory drugs.

[3] Graves JD, Krebs EG. Protein phosphorylation and signal transduction. Pharmacology

[4] Fischer EH. Cellular regulation by protein phosphorylation: A historical overview.

[5] Pearson RB, Kemp BE. Protein kinase phosphorylation site sequences and consensus

[6] Mendoza L, Thieffry D, Alvarez-Buylla ER. Genetic control of flower morphogenesis in

specificity motifs: Tabulations. Methods in Enzymology. 1991;**200**:62-81

*Arabidopsis thaliana*: A logical analysis. Bioinformatics. 1999;**15**(7-8):593-606

driven research.

174 Protein Phosphorylation

**Author details**

Stefano Schivo1

**References**

, Jeroen Leijten<sup>2</sup>

\*Address all correspondence to: j.n.post@utwente.nl

tion and signaling. Cell. 1995;**80**(2):225-236

and Therapeutics. 1999;**82**(2-3):111-121

Biofactors. 1997;**6**(3):367-374

Current Opinion in Cell Biology. 2009;**21**(2):317-324

**Acknowledgements**


[36] Cheng A, Genever PG. SOX9 determines RUNX2 transactivity by directing intracellular degradation. Journal of Bone and Mineral Research: The Official Journal of the American Society for Bone and Mineral Research. 2010;**25**(12):2680-2689

[21] Scholma J, Schivo S, Urquidi Camacho RA, van de Pol J, Karperien M, Post JN. Biological networks 101: Computational modeling for molecular biologists. Gene.

[22] Schivo S, Scholma J, Karperien M, Post JN, Van De Pol J, Langerak R, editors. Setting parameters for biological models with ANIMO. Electronic Proceedings in Theoretical

[23] Johnson VL, Hunter DJ. The epidemiology of osteoarthritis. Best Practice & Research.

[24] Scholma J, Schivo S, Karperien M, Langerak R, van de Pol J, Post JN. An echo in biology: Validating the executable chondrocyte. Osteoarthritis and Cartilage. 2014;**22**:S157-S157

[25] Ahmed TA, Hincke MT. Strategies for articular cartilage lesion repair and functional

[26] Sarzi-Puttini P, Cimmino MA, Scarpa R, Caporali R, Parazzini F, Zaninelli A, et al. Osteoarthritis: An overview of the disease and its treatment strategies. Seminars in

[27] Leijten JC, Emons J, Sticht C, van Gool S, Decker E, Uitterlinden A, et al. Gremlin 1, frizzled-related protein, and Dkk-1 are key regulators of human articular cartilage homeo-

[28] Ling W, Leijten JC, Post JN, Karperien M. Fibroblast growth factor-1 is a mesenchymal stromal cell secreted factor stimulating proliferation of osteoarthritic chondrocytes.

[29] Ma B, Landman EB, Miclea RL, Wit JM, Robanus-Maandag EC, Post JN, et al. WNT signaling and cartilage: Of mice and men. Calcified Tissue International. 2013;**92**(5):399-411

[30] Ma B, Zhong L, van Blitterswijk CA, Post JN, Karperien M. T cell factor 4 is a pro-catabolic and apoptotic factor in human articular chondrocytes by potentiating nuclear factor kappaB signaling. The Journal of Biological Chemistry. 2013;**288**(24):17552-17558 [31] Zhong L, Huang X, Karperien M, Post JN. The regulatory role of signaling crosstalk in hypertrophy of MSCs and human articular chondrocytes. International Journal of

[32] Clevers H, Nusse R. Wnt/beta-catenin signaling and disease. Cell. 2012;**149**(6):1192-1205 [33] Wang M, Shen J, Jin H, Im HJ, Sandy J, Chen D. Recent progress in understanding molecular mechanisms of cartilage degeneration during osteoarthritis. Annals of the New

[34] Thomas RS, Clarke AR, Duance VC, Blain EJ. Effects of Wnt3A and mechanical load on cartilage chondrocyte homeostasis. Arthritis Research and Therapy. 2011;**13**(6):R203

[35] Umlauf D, Frank S, Pap T, Bertrand J. Cartilage biology, pathology, and repair. Cellular

Computer Science, EPTCS; Open Publishing Association. 2014

restoration. Tissue Engineering. Part B, Reviews. 2009;**3**:305-329

Arthritis and Rheumatism. 2005;**35**(1 Suppl 1):1-10

Osteoarthritis and Cartilage. 2013;**21**:S273-S274

Molecular Sciences. 2015;**16**(8):19225-19247

York Academy of Sciences. 2011;**1240**:61-69

and Molecular Life Sciences. 2010;**67**(24):4197-4211

stasis. Arthritis and Rheumatism. 2012;**64**(10):3302-3312

Clinical Rheumatology. 2014;**28**(1):5-15

2014;**533**(1):379-384

176 Protein Phosphorylation


**Provisional chapter**

#### **FRET-Based Biosensors: Genetically Encoded Tools to Track Kinase Activity in Living Cells Track Kinase Activity in Living Cells**

**FRET-Based Biosensors: Genetically Encoded Tools to** 

DOI: 10.5772/intechopen.71005

Florian Sizaire and Marc Tramier Florian Sizaire and Marc Tramier Additional information is available at the end of the chapter

Additional information is available at the end of the chapter

http://dx.doi.org/10.5772/intechopen.71005

#### **Abstract**

[49] Nusse R. The WNT Homepage. 2017. Available from: http://web.stanford.edu/group/

[50] Ducy P, Zhang R, Geoffroy V, Ridall AL, Karsenty G. Osf2/Cbfa1: A transcriptional acti-

[51] Komori T. Regulation of bone development and extracellular matrix protein genes by

[52] Goldring MB, Tsuchimochi K, Ijiri K. The control of chondrogenesis. Journal of Cellular

[53] Zhong L, Huang X, Karperien M, Post JN. Correlation between gene expression and osteoarthritis progression in human. International Journal of Molecular Sciences.

[54] Akiyama H, Lyons JP, Mori-Akiyama Y, Yang X, Zhang R, Zhang Z, et al. Interactions between Sox9 and beta-catenin control chondrocyte differentiation. Genes &

[55] Wu M, Chen G, Li YP. TGF-beta and BMP signaling in osteoblast, skeletal development,

[56] Wuelling M, Vortkamp A. Transcriptional networks controlling chondrocyte proliferation and differentiation during endochondral ossification. Pediatric Nephrology.

[57] Smith N, Dong Y, Lian JB, Pratap J, Kingsley PD, van Wijnen AJ, et al. Overlapping expression of Runx1(Cbfa2) and Runx2(Cbfa1) transcription factors supports cooperative induction of skeletal development. Journal of Cellular Physiology. 2005;**203**(1):133-143

and bone formation, homeostasis and disease. Bone Research. 2016;**4**:16009

vator of osteoblast differentiation. Cell. 1997;**89**(5):747-754

RUNX2. Cell and Tissue Research. 2010;**339**(1):189-195

nusselab/cgi-bin/wnt/

178 Protein Phosphorylation

Biochemistry. 2006;**97**(1):33-44

Development. 2004;**18**(9):1072-1087

2016;**17**(7):1126

2010;**25**(4):625-631

Fluorescence microscopy is widely used in biology to localize, to track, or to quantify proteins in single cells. However, following particular events in living cells with good spatio-temporal resolution is much more complex. In this context, Forster resonance energy transfer (FRET) biosensors are tools that have been developed to monitor various events such as dimerization, cleavage, elasticity, or the activation state of a protein. In particular, genetically encoded FRET biosensors are strong tools to study mechanisms of activation and activity of a large panel of kinases in living cells. Their principles are based on a conformational change of a genetically encoded probe that modulates the distance between a pair of fluorescent proteins leading to FRET variations. Recent advances in fluorescence microscopy such as fluorescence lifetime imaging microscopy (FLIM) have made the quantification of FRET efficiency easier. This review aims to address the different kinase biosensors that have been developed, how they allow specific tracking of the activity or activation of a kinase, and to give an overview of the future challenging methods to simultaneously track several biosensors in the same system.

**Keywords:** kinase, biosensor, FRET, multiplex, protein conformation, fluorescence microscopy

#### **1. Introduction**

Investigating kinase activity in living cells remains a challenge, and usual methods are limited when one wishes to study cellular dynamic events. For a large panel of kinases, the phosphorylation state of the kinase or its substrate has become the main indicator of its activity [1]. One of the most common methods to study this activity is to perform Western Blot analysis on cell extracts by targeting the phosphorylated kinase residue or the phosphorylated substrate residue with an antibody. However, this semi-quantification of the activity state of the kinase

© 2016 The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. © 2017 The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

applies only for the whole population of the cells [2]. Thus, the other most frequent way to investigate kinase activity is to perform immunofluorescence by targeting the phosphorylated kinase or the phosphorylated substrate with a fluorescent antibody for microscopy observation. This method allows localization of the proteins phosphorylation state in a single cell. However, these two methods suffer from one major limitation: the inability to track this activation state both in space and in time to track dynamic events in living cells. Indeed, this requires lysing the cells or fixing them and permeabilizing them, which prevents sufficient spatio-temporal resolution to investigate intracellular events [3].

To overcome this limitation, new tools have been developed including Forster resonance energy transfer (FRET) biosensors [4]. FRET is a nonradiative transfer of energy of one donor fluorophore to an acceptor fluorophore and relies on (i) an overlap of the emission spectrum of the donor with the excitation spectrum of the acceptor, (ii) an adequate orientation between the two fluorophores, and (iii) a distance less than 10 nm between the two [5]. This feature has been used to investigate various cellular events such as protein–protein interactions by genetically tagging the two proteins of interest with a donor and an acceptor fluorescent protein [6], the intra-cellular Ca2+ signal by using calmodulin biosensor [7], proteases activity where the substrate is flanked by two fluorescent proteins, and the decrease of FRET indicates a cleavage of the protein [8], Rho GTPases for cytoskeleton dynamics [9, 10], and mechanical forces at adherent junctions [11, 12, 13]. The first kinase biosensor has been developed for cAMP-dependant protein kinase A (PKA) [14]. From this example and by taking advantage of the FRET characteristics and the conformational modifications of the phosphorylated sensors, several tools to monitor the kinase activity in space and time in living cells have been developed.

By conception, these tools are genetically encoded providing an invaluable advantage to endogenously producing the biosensor in live samples. In this review, we will first present genetically encoded FRET biosensors to monitor kinase activities based on phosphorylated peptide substrate. We will then introduce an alternative way of designing biosensors based on a conformational change of the kinase itself. Finally, we will present new methodological challenges such as multiplex FRET measurements in the same cell, thus allowing simultaneous monitoring of several kinase activities in time and space.

#### **2. Substrate-based kinase activity biosensors**

The first genetically encoded FRET biosensor for kinase activity was called A-kinase activity reporter (AKAR) and was designed to investigate the activity of PKA [14]. The idea was to follow a conformational change by FRET in a fusion protein composed of a substrate peptide sequence and a phosphorylated recognition domain. In this foundational work, the biosensor was composed of two fluorescent proteins, CFP and YFP (cyan and yellow fluorescent proteins). Between them sits a first domain, "the peptide substrate," containing a sequence phosphorylated by PKA, followed by a second domain, "the phosphorylation recognition domain," that binds to the peptide substrate when phosphorylated, these two domains are separated by an elastic linker. In the presence of active PKA, the peptide substrate becomes phosphorylated, triggering its affinity for the phosphorylation-binding domain. This association between the two domains induces a conformation change of the biosensor that brings closer both fluorophores and increases FRET efficiency between CPF donor and YFP acceptor (**Figure 1**). The efficiency of FRET can be detected by ratiometric measurements between the intensity signal of the donor and the acceptor. This biosensor can be expressed in cell and is able to provide a response to cell treatment such as forskolin that raises the level of cAMPactivating PKA [14]. This tool has then been improved several times by using a better reversible phospho-binding domain called FHA1 [15] or by changing the fluorophore couple to improve the ratiometric measurements [16, 17]. The AKAR biosensor has been used to report the activity of PKA in neurons of mouse brain slices, showing its value in neurosciences [18].

applies only for the whole population of the cells [2]. Thus, the other most frequent way to investigate kinase activity is to perform immunofluorescence by targeting the phosphorylated kinase or the phosphorylated substrate with a fluorescent antibody for microscopy observation. This method allows localization of the proteins phosphorylation state in a single cell. However, these two methods suffer from one major limitation: the inability to track this activation state both in space and in time to track dynamic events in living cells. Indeed, this requires lysing the cells or fixing them and permeabilizing them, which prevents sufficient

To overcome this limitation, new tools have been developed including Forster resonance energy transfer (FRET) biosensors [4]. FRET is a nonradiative transfer of energy of one donor fluorophore to an acceptor fluorophore and relies on (i) an overlap of the emission spectrum of the donor with the excitation spectrum of the acceptor, (ii) an adequate orientation between the two fluorophores, and (iii) a distance less than 10 nm between the two [5]. This feature has been used to investigate various cellular events such as protein–protein interactions by genetically tagging the two proteins of interest with a donor and an acceptor fluorescent protein [6], the intra-cellular Ca2+ signal by using calmodulin biosensor [7], proteases activity where the substrate is flanked by two fluorescent proteins, and the decrease of FRET indicates a cleavage of the protein [8], Rho GTPases for cytoskeleton dynamics [9, 10], and mechanical forces at adherent junctions [11, 12, 13]. The first kinase biosensor has been developed for cAMP-dependant protein kinase A (PKA) [14]. From this example and by taking advantage of the FRET characteristics and the conformational modifications of the phosphorylated sensors, several tools to monitor the kinase activity in space and time in living cells have been

By conception, these tools are genetically encoded providing an invaluable advantage to endogenously producing the biosensor in live samples. In this review, we will first present genetically encoded FRET biosensors to monitor kinase activities based on phosphorylated peptide substrate. We will then introduce an alternative way of designing biosensors based on a conformational change of the kinase itself. Finally, we will present new methodological challenges such as multiplex FRET measurements in the same cell, thus allowing simultane-

The first genetically encoded FRET biosensor for kinase activity was called A-kinase activity reporter (AKAR) and was designed to investigate the activity of PKA [14]. The idea was to follow a conformational change by FRET in a fusion protein composed of a substrate peptide sequence and a phosphorylated recognition domain. In this foundational work, the biosensor was composed of two fluorescent proteins, CFP and YFP (cyan and yellow fluorescent proteins). Between them sits a first domain, "the peptide substrate," containing a sequence phosphorylated by PKA, followed by a second domain, "the phosphorylation recognition domain," that binds to the peptide substrate when phosphorylated, these two domains are separated by an elastic linker. In the presence of active PKA, the peptide substrate becomes

spatio-temporal resolution to investigate intracellular events [3].

ous monitoring of several kinase activities in time and space.

**2. Substrate-based kinase activity biosensors**

developed.

180 Protein Phosphorylation

Based on this concept, several new kinase FRET biosensors were developed. The E-kinase activity reporter (EKAR) biosensor is a FRET-based probe to study ERK activity [19]. The fluorophore pair is composed of the green donor eGFP and the red acceptor mRFP1. The consensus substrate peptide originates from Cdc25c, a member of the MAPK family. As other kinases from this family could phosphorylate the substrate, an ERK binding domain has been inserted to ensure ERK specificity. A WW domain (containing 2 tryptophans separated by around 20 aa) was used to bind the phosphorylated substrate [20], and a flexible linker allows a conformational change, when the Cdc25C peptide substrate is phosphorylated. This biosensor has also experienced several steps of optimization by modifying the fluorophores or the flexible linker [21, 22]. Among all these biosensors, the reversibility of the conformational modification is a major feature to study variations of kinase activation states [23].

Kinase biosensors are such powerful tools to investigate the dynamic of kinase activity events in cells that several of these biosensors have been created to study mitotic kinases activity through the cell cycle, including cyclin B1-Cdk1 [24]. A kinase biosensor has also been used to study PKC (protein kinase C) activation which is involved in tumor promotion. CKAR (C-kinase activity reporter) is composed of the CFP/YFP fluorophore pair, a specific peptide substrate for PKC, and the FHA2 domain of Rad53p that can bind to the phosphorylated substrate [25]. In this particular case, the unphosphorylated biosensor harbors a maximum

**Figure 1.** Mechanism of a substrate-based FRET biosensor. When the biosensor is not phosphorylated, it adopts an opened conformation keeping away the donor fluorophore D, and the acceptor fluorophore, A. After the peptide substrate phosphorylation by the kinase, a phospho-binding domain can bind to it gathering the fluorophore pair and allowing FRET.

FRET efficiency conformation, and FRET signal decreases once phosphorylated. By rapidly acquiring FRET efficiency, oscillations of PKC phosphorylation in a range of a minute were highlighted [25].

Other biosensors have been derived from CKAR. Polo-like kinase-1 (PLK1) is a major mitotic kinase that activates Cdc25C phosphatase, which abrogates the inhibitory phosphorylation of proteins controlling the entry to mitosis. A FRET-based biosensor has been created by replacing the peptide substrate of PKC with a peptide substrate of PLK1, the use of which revealed that the timed-control activation of PLK1 depends on Aurora A [26]. The choice of the kinase peptide substrate to construct the biosensor is a key point to improve its specificity. For Plk1, a c-Jun substrate-based biosensor was developed [27], since the previous version based on Myt1 substrate sequence was also sensitive to Mts1 activity [28]. The c-jun–based version was then used to demonstrate that Plk1 activity is required for commitment to mitosis during cell cycles [29].

A biosensor to study Aurora B activation has also been developed [30]. But in this work, authors wanted to monitor the kinase activity at a specific location, since it has been postulated that an activity gradient of Aurora B at the mitotic spindle may play a role for mitotic progression. If one considers a conventional version of the FRET biosensor, its diffusion throughout the cell is too fast, and it is not possible to reveal a precise localization of the activation. Thus, Fuller and co-workers have added different localization sequences to target the biosensor either to the centromere using a peptide from CENP-B or to the chromatin using histone H2B [30].

FRET-based substrate kinase biosensors are good tools to investigate kinase activity, but they have some limitations and present three major challenges: (i) the biosensor relies on the endogenous kinase phosphorylating the substrate peptide, and thus, FRET variation is observed only when the kinase is particularly abundant or heavily stimulated, (ii) the sequence flanking the phosphorylation residue(s) targeted by the kinase must be known and selective for the kinase under study, and (iii) these biosensors only explore the catalytic activity of the kinase toward a specific substrate at once and not the activation process of the kinase itself. To solve this last issue, a new set of kinase FRET biosensors has been developed based on conformational changes of the kinase when active.

#### **3. Conformational kinase-based biosensor**

An alternative way of genetically encoded FRET-based substrate kinase biosensors has been developed by directly using the full-length kinase peptide sequence. Activation of a kinase frequently relies on a conformational opening of the enzymatic pocket. The idea is then to tag the whole kinase at its N- and C-terminus with a FRET pair of fluorescent proteins to be able to monitor this kinase activation related to the conformational change (**Figure 2**). To our knowledge, the first kinase FRET biosensor using this concept was developed to study c-Raf conformation [31]. This biosensor called Prin-c-Raf uses the CFP/YFP pair to flank c-Raf. A flexible linker has been added between the acceptor fluorophore and the kinase to enhance

FRET efficiency conformation, and FRET signal decreases once phosphorylated. By rapidly acquiring FRET efficiency, oscillations of PKC phosphorylation in a range of a minute were

Other biosensors have been derived from CKAR. Polo-like kinase-1 (PLK1) is a major mitotic kinase that activates Cdc25C phosphatase, which abrogates the inhibitory phosphorylation of proteins controlling the entry to mitosis. A FRET-based biosensor has been created by replacing the peptide substrate of PKC with a peptide substrate of PLK1, the use of which revealed that the timed-control activation of PLK1 depends on Aurora A [26]. The choice of the kinase peptide substrate to construct the biosensor is a key point to improve its specificity. For Plk1, a c-Jun substrate-based biosensor was developed [27], since the previous version based on Myt1 substrate sequence was also sensitive to Mts1 activity [28]. The c-jun–based version was then used to demonstrate that Plk1 activity is required for commitment to mitosis during cell

A biosensor to study Aurora B activation has also been developed [30]. But in this work, authors wanted to monitor the kinase activity at a specific location, since it has been postulated that an activity gradient of Aurora B at the mitotic spindle may play a role for mitotic progression. If one considers a conventional version of the FRET biosensor, its diffusion throughout the cell is too fast, and it is not possible to reveal a precise localization of the activation. Thus, Fuller and co-workers have added different localization sequences to target the biosensor either to the centromere using a peptide from CENP-B or to the chromatin using

FRET-based substrate kinase biosensors are good tools to investigate kinase activity, but they have some limitations and present three major challenges: (i) the biosensor relies on the endogenous kinase phosphorylating the substrate peptide, and thus, FRET variation is observed only when the kinase is particularly abundant or heavily stimulated, (ii) the sequence flanking the phosphorylation residue(s) targeted by the kinase must be known and selective for the kinase under study, and (iii) these biosensors only explore the catalytic activity of the kinase toward a specific substrate at once and not the activation process of the kinase itself. To solve this last issue, a new set of kinase FRET biosensors has been developed based on conforma-

An alternative way of genetically encoded FRET-based substrate kinase biosensors has been developed by directly using the full-length kinase peptide sequence. Activation of a kinase frequently relies on a conformational opening of the enzymatic pocket. The idea is then to tag the whole kinase at its N- and C-terminus with a FRET pair of fluorescent proteins to be able to monitor this kinase activation related to the conformational change (**Figure 2**). To our knowledge, the first kinase FRET biosensor using this concept was developed to study c-Raf conformation [31]. This biosensor called Prin-c-Raf uses the CFP/YFP pair to flank c-Raf. A flexible linker has been added between the acceptor fluorophore and the kinase to enhance

highlighted [25].

182 Protein Phosphorylation

cycles [29].

histone H2B [30].

tional changes of the kinase when active.

**3. Conformational kinase-based biosensor**

**Figure 2.** Mechanisms of a conformational-based FRET biosensor. (A) The auto-inhibiting domain can bin the catalytic domain of the kinase bringing closer the donor fluorophore D and the acceptor fluorophore A, allowing FRET. When the kinase is activated, the auto-inhibiting domain unbinds and the kinase adopts an opened and active conformation with a FRET decrease. (B) A lot of proteins adopt a new conformation when activated that modulates the distance between the pair of fluorophores.

FRET efficiency. Mutation of the residues Ser259 and Ser261 preventing c-RAF phosphorylation and mimicking the active state of the kinase leads to an open conformation of Prin-c-RAF as FRET ratio is decreased. When a constitutively active mutant of AKT that negatively regulates c-RAF is expressed, the wild-type version of c-RAF shows high FRET signal consecutive to a closed inactive conformation, while the mutated version S259A and S261A stays open with a lower FRET ratio. By using this biosensor, authors were able to show that the constitutively active H-RasV12 localized at the plasma membrane binds and opens the wild-type biosensor in an active conformation, inducing the recruitment of MEK at the plasma membrane.

A biosensor for PKCγ consisting of the kinase flanked by the donor super cyan fluorescent protein 3 (SCFP3A) and the acceptor YFP has also been developed [32]. The kinase displays a pseudosubstrate domain that is able to bind and inhibit catalytic activity. In this work, they compared a different mutated form of the biosensor PKCγ-A24E, where the pseudosubstrate cannot bind to PKCγ and observed a decrease of FRET that they are associated with an opened active conformation.

Another conformational biosensor has been developed to study FAK (Focal Adhesion Kinase) activity by taking advantages of the conformational changes associated with the activation state of the kinase being controlled by an inhibitory domain [33]. For that, a biosensor was constructed with the full-length kinase containing a FERM domain (F for 4.1 protein, E for ezrin, R for radixin, and M for moesin) for membrane localization and a kinase domain using the CFP/YFP FRET pair, the donor at the N-terminus of the protein and the acceptor directly between the two domains. When the FERM domain binds to the catalytic domain of FAK, it inhibits the kinase activity and FRET occurs. On the contrary, the absence of FRET corresponds to an active and thus open conformation. It is then possible to monitor FAK activity at the focal adhesion of living cells by expressing the biosensor transiently in living cells.

A biosensor of maternal embryonic leucine zipper kinase (MELK) has been created consisting of the MELK sequence flanked by the CFP/YFP pair [34]. As well as FAK, MELK has an autoinhibited domain at the C-terminus that can bind to the catalytic domain of the kinase. This biosensor was expressed in *Xenopus* embryos, and conformational changes were monitored in dividing cells. It has been demonstrated that the biosensor exhibits a closed conformation in the cytosol and an open conformation at the cleavage furrow. But here again, as for previous conformational sensors, only conclusions on the conformational change of the kinase could be made. Its direct link to the kinase activation (and activity) was not tackled.

Recently, we have developed an Aurora A biosensor based on conformational changes [35]. It is composed of the full-length kinase flanked by a GFP donor and a mCherry acceptor. To be functional, Aurora A undergoes a conformational change following autophosphorylation on the T288 residue [36, 37]. By exploiting this mechanism, we designed a biosensor that directly associates the conformational change of the kinase with its state of activation. Indeed, *in vitro* treatment with ATP leads to a closed conformation when treatment with phosphatase leads to an opened conformation. The activation state is also monitored by fluorescence lifetime imaging microscopy (FLIM) in living cells. Through this work, we show that the biosensor was able to functionally replace the endogenous Aurora A depleted by siRNA. Thus, by replacing the endogenous kinase, this biosensor is a direct reporter of the activation state of Aurora A at endogenous levels in stable cell lines with a good spatio-temporal resolution. With this tool, by dissociating the quantity and the activation state of the kinase, we were able to highlight a new nonmitotic role of Aurora A in G1 phase [35].

This kind of biosensor can also be adapted to other enzymatic activities. As an example, a BRET (bioluminescence resonance energy transfer)-based biosensor of the PTEN (phosphatase and tensin homolog) phosphatase has been developed and is composed of the fulllength PTEN protein flanked by a donor Rluc and the acceptor YFP [38]. PTEN biosensor immunoprecipitated from cells displays the same phosphatase activity on PIP3 (phosphatidylinositol 3,4,5 trisphosphate) and AKT (or Protein kinase B) as the wild-type PTEN. This biosensor can also be expressed at endogenous levels in human embryonic kidney (HEK) cells. The mutation of four residues Ser380, Thr382, Thr383, and Ser385 favoring a closed conformation leads to a strong decrease of BRET signal. The association between conformational changes and the activity state has allowed the monitoring of PTEN regulation in living cells. Authors have thus been able to correlate inhibition of the known activation pathway of PTEN using a CK2 (Casein Kinase 2) inhibitor, with its change of conformation or inversely by co-expressing S1PR2, an activator of PTEN. Once the biosensor was validated, the authors used it to identify new GPCRs (G protein–coupled receptors) activating PTEN.

The reliability of these biosensors, consisting of the full-length protein flanked by a pair of fluorophores, was recently applied to study any protein function associated with a conformational change. For example, a study of the conformational change of the Tau protein has been tackled using the protein flanked by a CFP/YFP pair [39]. The use of this biosensor led to the demonstration that the binding of Tau to the microtubules induces a switch to a hairpin conformation of Tau. In addition, it has been shown that mutations of Tau responsible of Frontotemporal dementia with parkinsonism-17 (FTDP-17) disorder alter this conformation change. This method has also been used to study vinculin conformation by using a biosensor consisting of the protein flanked by an mTurquoise donor and a NeonGreen acceptor [40]. Vinculin displays an auto-inhibited state, when the tail domain and the head domain are binding, increasing FRET signal. A mutated version of the biosensor that is unable to bind to talin showed a decreasing FRET signal and a disruption in the vinculin localization at focal adhesions. In contrast, paxillin knock-down or mutations leading to a decrease in actin binding did not modify FRET signal.

protein, E for ezrin, R for radixin, and M for moesin) for membrane localization and a kinase domain using the CFP/YFP FRET pair, the donor at the N-terminus of the protein and the acceptor directly between the two domains. When the FERM domain binds to the catalytic domain of FAK, it inhibits the kinase activity and FRET occurs. On the contrary, the absence of FRET corresponds to an active and thus open conformation. It is then possible to monitor FAK activity at the focal adhesion of living cells by expressing the biosensor transiently in

A biosensor of maternal embryonic leucine zipper kinase (MELK) has been created consisting of the MELK sequence flanked by the CFP/YFP pair [34]. As well as FAK, MELK has an autoinhibited domain at the C-terminus that can bind to the catalytic domain of the kinase. This biosensor was expressed in *Xenopus* embryos, and conformational changes were monitored in dividing cells. It has been demonstrated that the biosensor exhibits a closed conformation in the cytosol and an open conformation at the cleavage furrow. But here again, as for previous conformational sensors, only conclusions on the conformational change of the kinase could be

Recently, we have developed an Aurora A biosensor based on conformational changes [35]. It is composed of the full-length kinase flanked by a GFP donor and a mCherry acceptor. To be functional, Aurora A undergoes a conformational change following autophosphorylation on the T288 residue [36, 37]. By exploiting this mechanism, we designed a biosensor that directly associates the conformational change of the kinase with its state of activation. Indeed, *in vitro* treatment with ATP leads to a closed conformation when treatment with phosphatase leads to an opened conformation. The activation state is also monitored by fluorescence lifetime imaging microscopy (FLIM) in living cells. Through this work, we show that the biosensor was able to functionally replace the endogenous Aurora A depleted by siRNA. Thus, by replacing the endogenous kinase, this biosensor is a direct reporter of the activation state of Aurora A at endogenous levels in stable cell lines with a good spatio-temporal resolution. With this tool, by dissociating the quantity and the activation state of the kinase, we were able to highlight a

This kind of biosensor can also be adapted to other enzymatic activities. As an example, a BRET (bioluminescence resonance energy transfer)-based biosensor of the PTEN (phosphatase and tensin homolog) phosphatase has been developed and is composed of the fulllength PTEN protein flanked by a donor Rluc and the acceptor YFP [38]. PTEN biosensor immunoprecipitated from cells displays the same phosphatase activity on PIP3 (phosphatidylinositol 3,4,5 trisphosphate) and AKT (or Protein kinase B) as the wild-type PTEN. This biosensor can also be expressed at endogenous levels in human embryonic kidney (HEK) cells. The mutation of four residues Ser380, Thr382, Thr383, and Ser385 favoring a closed conformation leads to a strong decrease of BRET signal. The association between conformational changes and the activity state has allowed the monitoring of PTEN regulation in living cells. Authors have thus been able to correlate inhibition of the known activation pathway of PTEN using a CK2 (Casein Kinase 2) inhibitor, with its change of conformation or inversely by co-expressing S1PR2, an activator of PTEN. Once the biosensor was validated, the authors used it to identify new GPCRs (G protein–coupled receptors) activating

made. Its direct link to the kinase activation (and activity) was not tackled.

new nonmitotic role of Aurora A in G1 phase [35].

living cells.

184 Protein Phosphorylation

PTEN.

Thus, these genetically encoded biosensors are efficient at monitoring protein activation at cellular levels when expressed in living cells. A lot of proteins are known to adopt different conformation states according to their activation, and this is why FRET or BRET biosensors are best suited for tracking their activity in space and in time in living cells. It is likely that this tool will be used intensively to study protein conformation linked to activity in the next few years.

One can thus follow the activation of a kinase by following its conformational change using a conformation-based biosensor and follow its catalytic activity using a substrate-based biosensor. It would be of great interest to simultaneously follow activation and activity in a single living cell, a pursuit that calls for methods able to monitor the two different FRET biosensors simultaneously.

#### **4. New methodological insights for multiplexing kinase biosensors**

Owing to complex crosstalk between signaling pathways, multi-parameter biosensing experiments have become essential to correlate biochemical activities without lag time during a dedicated cellular process. A very exciting challenge has thus been to follow several FRET biosensors on the same sample at the same time and in the same location [41]. Commonly, FRET is measured by the fluorescence intensity ratio of the acceptor to the donor. In that case, whatever the two fluorescent protein FRET pairs chosen, CFP/YFP and mOrange/mCherry [42], mTFP1/mCitrine and mAmetrine/tdTomato [43, 44], mTagBFP/sfGFP and mVenus/ mKok [45], the multiplex approach suffers from two limitations: (i) a spectral bleed-through of the first acceptor in the second donor emission band that depends directly on the respective quantities of the two biosensors and (ii) the multiple excitation wavelength which requires sequential acquisition that does not adequately follow fast signal dynamics or signal changes in highly motile samples.

To overcome the first limitation, a meroCBD (merocyanine–Cdc42-binding domain) biosensor modified with a far-red organic fluorophore (Alexa750) was used for probing Cdc42 simultaneously with a genetically encoded CFP/YFP FRET-based biosensor for Rho A [46]. This approach prevents spectral bleed-through but cannot be generalized to all genetically encoded FRET biosensors, where organic fluorophores cannot easily replace fluorescent proteins. The same team also developed an environment-sensing dye called mero199 [47]. This dye can bind to the active form of Cdc42 leading to a shift of its excitation/emission ratio. In combination with a Rac1 biosensor, they were able to simultaneously monitor activation of both proteins and to correlate it with retraction or velocity of migrating MEF (mouse embryonic fibroblasts) cells. Very recently, an elegant method based on linear unmixing of 3D excitation/emission fingerprints applied to three biosensors simultaneously was published [48]. This type of approach based on image calculation is often limited by the different biosensors expression levels and a poor signal-to-noise ratio after complex image corrections.

To overcome the second limitation, the two FRET pairs CFP/YFP and Sapphire/RFP in combination with a single violet excitation were used [49], resulting in no lag time in biochemical activity recording. But again, in this case, the spectral bleed-through and excitation crosstalk necessitate linear unmixing. Another interesting approach for simultaneously multiplexing two FRET activities was developed using a "Large Stokes Shift" orange fluorescent protein, LSSmOrange [50]. The authors used a CFP-YFP together with LSSmOrange-mKate2 biosensors enabling imaging of apoptotic activity and calcium fluctuations in real time using intensity-based methods. Other studies were carried out utilizing FLIM instead of ratio imaging to measure FRET. When FRET occurs, donor fluorescence lifetime decreases. This method requires measurement of the donor fluorescence only and is independent of emission from the acceptor. By using CFP and YFP as donor and the same red acceptor (tHcRed), FLIM of CFP and YFP donors allow the two different FRET signals to be distinguished [51]. Combination of FLIM-FRET of a red-shifted TagRFP/mPlum pair with ratio imaging of a CFP/Venus pair allows maximal the spectral separation while, at the same time, overcoming the low quantum yield of the far-red acceptor mPlum [52]. The two last examples alleviated the spectral bleedthrough but not the limitation associated with multiple excitations.

To overcome both limitations, a novel red-shifted fluorophore mCyRFP1 has been developed with a high Stokes shift [53]. This fluorophore has an excitation spectrum in the range of the GFP emission spectrum (around 500 nm), but its emission spectrum is shifted compared to GFP. An emission dichroic filter allows simultaneous detection of the GFP fluorescence lifetime and the mCyRFP1 fluorescence lifetime. The authors were able to perform two-photon fluorescence lifetime imaging by using only one excitation laser at 920 nm with a Rhoa biosensor and a CaMKIIα biosensor. Furthermore, while the RhoA biosensor uses the pair mCyRFP1/ mMaroon1, the CaMKIIα biosensor uses mEGFP and dimVenus which is a dark fluorophore preventing bleed-through with mCyRFP1.

Recently, our team has developed a similar method by taking advantages of the LSSmOrange (Large Stoke Shift) and the dark fluorophore ShadowG [54]. We modified two substrate kinase biosensors, EKAR2G and AKAR4 (E-Kinase Activity Reporter type 2G for ERK and A-Kinase Activity Reporter type 4 for PKA), with a new pair of fluorophores mTFP1/ShadowG and LSSmOrange/mKate2, respectively. LSSmOrange and mTFP1 are both excitable by using a single 440 nm wavelength. By single excitation wavelength dual-color FLIM, we are able to simultaneously monitor the activity of ERK and PKA in living cells at the same location. Thus, the activity of each kinase in response to forskolin or EGF treatment can be imaged simultaneously. This approach overcomes the limitations of the multiple excitation wavelengths and bleed-through.

simultaneously with a genetically encoded CFP/YFP FRET-based biosensor for Rho A [46]. This approach prevents spectral bleed-through but cannot be generalized to all genetically encoded FRET biosensors, where organic fluorophores cannot easily replace fluorescent proteins. The same team also developed an environment-sensing dye called mero199 [47]. This dye can bind to the active form of Cdc42 leading to a shift of its excitation/emission ratio. In combination with a Rac1 biosensor, they were able to simultaneously monitor activation of both proteins and to correlate it with retraction or velocity of migrating MEF (mouse embryonic fibroblasts) cells. Very recently, an elegant method based on linear unmixing of 3D excitation/emission fingerprints applied to three biosensors simultaneously was published [48]. This type of approach based on image calculation is often limited by the different biosensors expression levels and a poor signal-to-noise ratio after complex

To overcome the second limitation, the two FRET pairs CFP/YFP and Sapphire/RFP in combination with a single violet excitation were used [49], resulting in no lag time in biochemical activity recording. But again, in this case, the spectral bleed-through and excitation crosstalk necessitate linear unmixing. Another interesting approach for simultaneously multiplexing two FRET activities was developed using a "Large Stokes Shift" orange fluorescent protein, LSSmOrange [50]. The authors used a CFP-YFP together with LSSmOrange-mKate2 biosensors enabling imaging of apoptotic activity and calcium fluctuations in real time using intensity-based methods. Other studies were carried out utilizing FLIM instead of ratio imaging to measure FRET. When FRET occurs, donor fluorescence lifetime decreases. This method requires measurement of the donor fluorescence only and is independent of emission from the acceptor. By using CFP and YFP as donor and the same red acceptor (tHcRed), FLIM of CFP and YFP donors allow the two different FRET signals to be distinguished [51]. Combination of FLIM-FRET of a red-shifted TagRFP/mPlum pair with ratio imaging of a CFP/Venus pair allows maximal the spectral separation while, at the same time, overcoming the low quantum yield of the far-red acceptor mPlum [52]. The two last examples alleviated the spectral bleed-

To overcome both limitations, a novel red-shifted fluorophore mCyRFP1 has been developed with a high Stokes shift [53]. This fluorophore has an excitation spectrum in the range of the GFP emission spectrum (around 500 nm), but its emission spectrum is shifted compared to GFP. An emission dichroic filter allows simultaneous detection of the GFP fluorescence lifetime and the mCyRFP1 fluorescence lifetime. The authors were able to perform two-photon fluorescence lifetime imaging by using only one excitation laser at 920 nm with a Rhoa biosensor and a CaMKIIα biosensor. Furthermore, while the RhoA biosensor uses the pair mCyRFP1/ mMaroon1, the CaMKIIα biosensor uses mEGFP and dimVenus which is a dark fluorophore

Recently, our team has developed a similar method by taking advantages of the LSSmOrange (Large Stoke Shift) and the dark fluorophore ShadowG [54]. We modified two substrate kinase biosensors, EKAR2G and AKAR4 (E-Kinase Activity Reporter type 2G for ERK and A-Kinase Activity Reporter type 4 for PKA), with a new pair of fluorophores mTFP1/ShadowG and LSSmOrange/mKate2, respectively. LSSmOrange and mTFP1 are both excitable by using a single 440 nm wavelength. By single excitation wavelength dual-color FLIM, we are able to

through but not the limitation associated with multiple excitations.

preventing bleed-through with mCyRFP1.

image corrections.

186 Protein Phosphorylation

Because FLIM is now a widely used microscopy approach, the decrease of the donor lifetime is sufficient to quantify FRET, and fluorescence of the acceptor is not mandatory, as it is still the case when one uses ratiometric FRET. Changing the fluorescent acceptor with a nonfluorescent acceptor leads to the development of a new kind of single-color FRET biosensor (**Figure 3A**). It is then perfectly adapted for simultaneous monitoring of kinase FRET biosensing.

Another method to get a single-color biosensor to perform multiplex could be based on homo-FRET measured by anisotropy [55, 56]. HomoFRET occurs when a fluorophore transfers its energy to a closely identical fluorophore. However, it is impossible to measure homoFRET by ratiometric or fluorescence lifetime measurements. Fluorescence anisotropy can be measured by detecting the parallel and the perpendicular light emitted by a fluorophore excited with a polarized light [57], and this anisotropy decreases when FRET occurs between nonparallel fluorescent dipoles. This method was already used to study protein oligomerization [58]. For example, a study has used fluorescence anisotropy to determine the degree of clustering of proteins such as GPI or EGFR fused to GFP in living cells [59]. This approach has been investigated to multiplex at the same time a conventional calcium heteroFRET biosensor using FLIM with the oligomerization of pleckstrin homology domains of Akt (Akt-PH) labeled

**Figure 3.** Single-color genetically encoded FRET biosensor. (A) When a fluorophore F is excited, it can transfer its energy to a dark acceptor DA by FRET. Even excited, the dark acceptor emits no detectable light, and FRET is measured by the measurement of donor fluorescence lifetime. It constitutes a single-color FRET biosensor. (B) When a fluorophore F is excited by a polarized light, it emits polarized fluorescence. When homoFRET occurs between two identical fluorophores, it can leads to the depolarization of fluorescence emission decreasing the anisotropy. Again, it constitutes a single-color FRET biosensor.

with mCherry [60]. From our knowledge, the development of an intramolecular homoFRET biosensor to follow a biochemical activity was not yet developed but has very interesting potential. Adapting this method to kinase biosensors would provide a new methodology to simultaneously follow multiple biosensors.

#### **5. Concluding remarks**

Kinases have multiple functions in cells, and their mechanisms are very dynamic in both space and time. We have focused our review on two types of kinase biosensors. The substrate-based kinase biosensors are good tools to specifically monitor the activity of a kinase, but it requires to have a good knowledge of the substrate peptide sequence, particularly for its specificity, and a biological system where the activity of the kinase is sufficient to detect FRET. The conformationbased biosensors provide information about the activation state of the kinase itself; however, they do not provide information about its catalytic activity that can be further regulated by other post-translational modifications. Gathering these different tools with a multiplex methodology by using the approaches of single-color FRET biosensor would provide new mechanistic insight to investigate kinase functions with an adequate spatio-temporal resolution.

#### **Acknowledgements**

We wish to apologize to the authors whose works were not cited in this review. We wish to thank G. Bertolin, G. Herbomel, and C. Demeautis for their many helpful discussions about FRET biosensors and D. Fairbass for critical reading the manuscript. Work in the laboratory is supported by the CNRS and the University of Rennes 1 and grant from the "Ligue Nationale Contre le Cancer" (LNCC, region Grand Ouest). F. S. is fellow of the Région Bretagne and University of Rennes 1.

#### **Author details**

Florian Sizaire and Marc Tramier\*

\*Address all correspondence to: marc.tramier@univ-rennes1.fr

Institut de Génétique et Développement de Rennes (IGDR), CNRS, Université de Rennes, Rennes, France

#### **References**

[1] Manning G, Whyte DB, Martinez H, Sudarsanam S. The protein kinase complement of the human genome. Science. 2002;**298**:1912-1934

[2] Yang P-C, Mahmood T. Western blot: Technique, theory, and trouble shooting. North American Journal of Medical Sciences. 2012;**4**:429

with mCherry [60]. From our knowledge, the development of an intramolecular homoFRET biosensor to follow a biochemical activity was not yet developed but has very interesting potential. Adapting this method to kinase biosensors would provide a new methodology to

Kinases have multiple functions in cells, and their mechanisms are very dynamic in both space and time. We have focused our review on two types of kinase biosensors. The substrate-based kinase biosensors are good tools to specifically monitor the activity of a kinase, but it requires to have a good knowledge of the substrate peptide sequence, particularly for its specificity, and a biological system where the activity of the kinase is sufficient to detect FRET. The conformationbased biosensors provide information about the activation state of the kinase itself; however, they do not provide information about its catalytic activity that can be further regulated by other post-translational modifications. Gathering these different tools with a multiplex methodology by using the approaches of single-color FRET biosensor would provide new mechanistic insight to investigate kinase functions with an adequate spatio-temporal resolution.

We wish to apologize to the authors whose works were not cited in this review. We wish to thank G. Bertolin, G. Herbomel, and C. Demeautis for their many helpful discussions about FRET biosensors and D. Fairbass for critical reading the manuscript. Work in the laboratory is supported by the CNRS and the University of Rennes 1 and grant from the "Ligue Nationale Contre le Cancer" (LNCC, region Grand Ouest). F. S. is fellow of the Région Bretagne and University of Rennes 1.

Institut de Génétique et Développement de Rennes (IGDR), CNRS, Université de Rennes,

[1] Manning G, Whyte DB, Martinez H, Sudarsanam S. The protein kinase complement of

simultaneously follow multiple biosensors.

**5. Concluding remarks**

188 Protein Phosphorylation

**Acknowledgements**

**Author details**

Rennes, France

**References**

Florian Sizaire and Marc Tramier\*

\*Address all correspondence to: marc.tramier@univ-rennes1.fr

the human genome. Science. 2002;**298**:1912-1934


[29] Gheghiani L, Loew D, Lombard B, Mansfeld J, Gavet O. PLK1 activation in late G2 sets up commitment to mitosis. Cell Reports. 2017;**19**:2060-2073

[15] Zhang J, Hupfeld CJ, Taylor SS, Olefsky JM, Tsien RY. Insulin disrupts β-adrenergic sig-

[16] Allen MD, Zhang J. Subcellular dynamics of protein kinase A activity visualized by FRET-based reporters. Biochemical and Biophysical Research Communications.

[17] Depry C, Allen MD, Zhang J. Visualization of PKA activity in plasma membrane micro-

[18] Gervasi N, Hepp R, Tricoire L, Zhang J, Lambolez B, Paupardin-Tritsch D, Vincent P. Dynamics of protein kinase A signaling at the membrane, in the cytosol, and in the nucleus of neurons in mouse brain slices. Journal of Neuroscience. 2007;**27**:2744-2750

[19] Harvey CD, Ehrhardt AG, Cellurale C, Zhong H, Yasuda R, Davis RJ, Svoboda K. A genetically encoded fluorescent sensor of ERK activity. Proceedings of the National

[20] Lu P-J, Zhou XZ, Shen M, Lu KP. Function of WW domains as phosphoserine-or phos-

[21] Vandame P, Spriet C, Riquet F, Trinel D, Cailliau-Maggio K, Bodart J-F. Optimization of ERK activity biosensors for both ratiometric and lifetime FRET measurements. Sensors.

[22] Komatsu N, Aoki K, Yamada M, Yukinaga H, Fujita Y, Kamioka Y, Matsuda M.Development of an optimized backbone of FRET biosensors for kinases and GTPases. Molecular Biology

[23] González-Vera J, Morris M. Fluorescent reporters and biosensors for probing the

[24] Gavet O, Pines J. Activation of cyclin B1–Cdk1 synchronizes events in the nucleus and

[25] Violin JD, Zhang J, Tsien RY, Newton AC. A genetically encoded fluorescent reporter reveals oscillatory phosphorylation by protein kinase C. The Journal of Cell Biology.

[26] Macůrek L, Lindqvist A, Lim D, Lampson MA, Klompmaker R, Freire R, Clouin C, Taylor SS, Yaffe MB, Medema RH. Polo-like kinase-1 is activated by aurora A to promote

[27] Liu D, Davydenko O, Lampson MA. Polo-like kinase-1 regulates kinetochore–microtubule dynamics and spindle checkpoint silencing. The Journal of Cell Biology.

[28] Bruinsma W, Macůrek L, Freire R, Lindqvist A, Medema RH. Bora and aurora-A con-

tinue to activate Plk1 in mitosis. Journal of Cell Science. 2014;**127**:801-811

nalling to protein kinase A in adipocytes. Nature. 2005;**437**:569-573

domains. Molecular BioSystems. 2011;**7**:52-58

Academy of Sciences. 2008;**105**:19264-19269

phothreonine-binding modules. Science. 1999;**283**:1325-1328

dynamic behavior of protein kinases. Proteomes. 2015;**3**:369-410

checkpoint recovery. Nature. 2008;**455**:119-123

the cytoplasm at mitosis. The Journal of Cell Biology. 2010;**189**:247-259

2006;**348**:716-721

190 Protein Phosphorylation

2014;**14**:1140-1154

2003;**161**:899-909

2012;**198**:491-499

of the Cell. 2011;**22**:4647-4656


[57] Chan FTS, Kaminski CF, Kaminski Schierle GS. HomoFRET fluorescence anisotropy imaging as a tool to study molecular self-assembly in live cells. Chemphyschem. 2011; **12**:500-509

[43] Ai H, Hazelwood KL, Davidson MW, Campbell RE. Fluorescent protein FRET pairs for

[44] Ding Y, Ai H, Hoi H, Campbell RE. Förster resonance energy transfer-based biosensors for multiparameter ratiometric imaging of Ca2+ dynamics and caspase-3 activity in

[45] Su T, Pan S, Luo Q, Zhang Z. Monitoring of dual bio-molecular events using FRET biosensors based on mTagBFP/sfGFP and mVenus/mKOκ fluorescent protein pairs.

[46] Machacek M, Hodgson L, Welch C, Elliott H, Pertz O, Nalbant P, Abell A, Johnson GL, Hahn KM, Danuser G. Coordination of Rho GTPase activities during cell protrusion.

[47] MacNevin CJ, Gremyachinskiy D, Hsu C-W, Li L, Rougie M, Davis TT, Hahn KM. Environment-sensing merocyanine dyes for live cell imaging applications. Bioconjugate

[48] Woehler A. Simultaneous quantitative live cell imaging of multiple FRET-based biosen-

[49] Niino Y, Hotta K, Oka K. Simultaneous live cell imaging using dual FRET sensors with

[50] Shcherbakova DM, Hink MA, Joosen L, Gadella TWJ, Verkhusha VV. An orange fluorescent protein with a large stokes shift for single-excitation multicolor FCCS and FRET

[51] Peyker A, Rocks O, Bastiaens PIH. Imaging activation of two Ras isoforms simultane-

[52] Grant DM, Zhang W, McGhee EJ, Bunney TD, Talbot CB, Kumar S, Munro I, Dunsby C, Neil MAA, Katan M, et al. Multiplexed FRET to image multiple signaling events in live

[53] Laviv T, Kim BB, Chu J, Lam AJ, Lin MZ, Yasuda R. Simultaneous dual-color fluorescence lifetime imaging with novel red-shifted fluorescent proteins. Nature Methods.

[54] Demeautis C, Sipieter F, Roul J, Chapuis C, Padilla-Parra S, Riquet FB, Tramier M. Multiplexing PKA and ERK1&2 kinases FRET biosensors in living cells using single

[55] Gautier I, Tramier M, Durieux C, Coppey J, Pansu RB, Nicolas J-C, Kemnitz K, Coppey-Moisan M. Homo-FRET microscopy in living cells to measure monomer-dimer transi-

[56] Tramier M, and Coppey-Moisan M. (2008). Fluorescence anisotropy imaging microscopy for homo-FRET in living cells. Methods in Cell Biology, (Elsevier), pp. 395-414.

excitation wavelength dual colour FLIM. Scientific Reports. 2017;**7**:41026

tion of GFP-tagged proteins. Biophysical Journal. 2001;**80**:3000-3008

imaging. Journal of the American Chemical Society. 2012;**134**:7913-7923

ratiometric imaging of dual biosensors. Nature Methods. 2008;**5**:401-403

single cells. Analytical Chemistry. 2011;**83**:9687-9693

Biosensors & Bioelectronics. 2013;**46**:97-101

a single excitation light. PLoS One. 2009;**4**:e6036

ously in a single cell. Chembiochem. 2005;**6**:78-85

cells. Biophysical Journal. 2008;**95**:L69-L71

2016;**13**:989-992

Nature. 2009;**461**:99-103

192 Protein Phosphorylation

Chemistry. 2013;**24**:215-223

sors. PLoS One. 2013;**8**:e61096

