Abstract
Background
Signal transduction pathways are usually modelled using classical quantitative methods, which are based on ordinary differential equations (ODEs). However, some difficulties are inherent in this approach. On the one hand, the kinetic parameters involved are often unknown and have to be estimated. With increasing size and complexity of signal transduction pathways, the estimation of missing kinetic data is not possible. On the other hand, ODEs based models do not support any explicit insights into possible (signal) flows within the network. Moreover, a huge amount of qualitative data is available due to highthroughput techniques. In order to get information on the systems behaviour, qualitative analysis techniques have been developed. Applications of the known qualitative analysis methods concern mainly metabolic networks. Petri net theory provides a variety of established analysis techniques, which are also applicable to signal transduction models. In this context special properties have to be considered and new dedicated techniques have to be designed.
Methods
We apply Petri net theory to model and analyse signal transduction pathways first qualitatively before continuing with quantitative analyses. This paper demonstrates how to build systematically a discrete model, which reflects provably the qualitative biological behaviour without any knowledge of kinetic parameters. The mating pheromone response pathway in Saccharomyces cerevisiae serves as case study.
Results
We propose an approach for model validation of signal transduction pathways based on the network structure only. For this purpose, we introduce the new notion of feasible tinvariants, which represent minimal selfcontained subnets being active under a given input situation. Each of these subnets stands for a signal flow in the system. We define maximal common transition sets (MCTsets), which can be used for tinvariant examination and net decomposition into smallest biologically meaningful functional units.
Conclusion
The paper demonstrates how Petri net analysis techniques can promote a deeper understanding of signal transduction pathways. The new concepts of feasible tinvariants and MCTsets have been proven to be useful for model validation and the interpretation of the biological system behaviour. Whereas MCTsets provide a decomposition of the net into disjunctive subnets, feasible tinvariants describe subnets, which generally overlap. This work contributes to qualitative modelling and to the analysis of large biological networks by their fully automatic decomposition into biologically meaningful modules.
Background
Signal transduction pathways are of special interest in biological and medical sciences. Many diseases are related to disturbances in signalling pathways. For example proteintyrosine kinases (PTKs) are important regulators of intracellular signal transduction pathways, mediating development and multicellular communication in metazoans. Their activity is tightly controlled and regulated, but perturbation of the normal autoinhibitory constraints on kinase activity can result in oncogenic PTK signalling [1]. Another example are human Gproteincoupled receptors, which mediate response to light, odour, taste, hormones, and neurotransmitters. Widely prescribed drugs, such as βadrenergic receptor blockers, antihistamines, and serotoninreuptake inhibitors bind to specific Gproteincoupled receptors [2]. There are signalling modules, e.g., this Gprotein and the mitogenactivated proteinkinase (MAPK) cascade, which occur in eukaryotic organisms from yeast to human. Budding yeast (Saccharomyces cerevisiae) has proven to be indispensable in understanding the mechanisms, interrelationships, and regulation of these components. On account of this crucial role of budding yeast we discuss our approach on the best understood signalling pathway in S. cerevisiae, the one of the mating pheromone response [3,4].
Signal transduction pathways are usually modelled by a set of ordinary differential equations to describe the dynamic changes in the concentrations of the involved biochemical species [58]. One of the main problems in using these quantitative methods is the incomplete knowledge of kinetic parameters. Often only 30–50% of them are known, such that existing methods for parameter estimation fail. In addition, the numerical values of concentrations of species may not be considered, if they are very small, because of rounding effects in the solver algorithms. No insight into possible signal flows and the connections between species, e.g., feedback loops, are given explicitly. Moreover, more qualitative than quantitative data became available in the last years. This all has initiated the development of qualitative methods, i.e., discrete models and related evaluation techniques, which are used as a complementary intermediate step for construction and understanding of larger reliable models.
In order to model and analyse biochemical pathways on a qualitative level, dedicated methods have been developed. An example for a well established concept are the elementary modes [9], which are based on the incidence (stoichiometric) matrix of the underlying directed graph. This approach is applied for analysing metabolic networks. They are considered to be in steady state in order to use convex cone analysis techniques.
Recently, interaction graphs and Boolean networks are used to model and analyse signalling and regulatory networks [10]. [11] shows the limits of logical modelling techniques for gene regulatory networks. The authors propose an integrative systematic approach of logical and Petri net modelling discussing the tryptophan biosynthesis in E. coli. However, analyses are mentioned only in the outlook. Opposite to [12], who hold the view that the concept of elementary modes or minimal tinvariants, respectively, cannot be applied to signalling networks, [10] uses this concept, as we had already proposed in [1315] to analyse an apoptosis model.
We have based our modelling approach of signalling pathways on Petri net theory, because it comprises several analysis techniques including those, which are based on the incidence matrix. In contrast to the above mentioned techniques, Petri nets provide useful unique visualisation techniques with possibilities for hierarchical modelling and for animation. Especially this aspect is of crucial importance for experimentalists and for the dialog between experimentalists and theoreticians.
Carl Adam Petri introduced the basic concepts of Petri nets in his dissertation in the context of technical systems [16]. He developed an approach for describing and studying models, which consist of concurrent, i.e., independent, and/or causally dependent components. Since this time, many theorems and algorithms have been developed and implemented to analyse Petri net models, see, e.g., [17,18]. First applications of Petri net theory to biological systems were published in [1921]. Meanwhile, metabolic pathways [22,23], signal transduction pathways [14,24], and generegulatory networks [25,26] have been modelled and analysed successfully using various classes of Petri nets, qualitative as well as quantitative ones. An overview is given, e.g., in [27,28]. For an earlier bibliography of related application papers see [29].
Signal transduction pathways exhibit some special properties. In contrast to metabolic networks, in signalling pathways there is no substance flow, but a signal flow, which is performed, for example, by phosphorylated and dephosphorylated protein forms. There are no stoichiometric reaction equations given, which dictate the net structure and define the arc weights. Modelling signalling pathways we have to work on another abstraction level than in modelling metabolic networks. Petri nets provide a generic description principle, applicable on any level of abstraction.
In this paper we present a method, which enables the systematic deduction of the net structure from the known signalling processes in the cell, which are translated into logical terms as implication, conjunction, inclusive or exclusive disjunction, and negation. These terms can be translated unambiguously to Petri net components. The resulting model is a qualitative, discrete Petri net, which can be analysed for certain system properties to validate the model.
In the following, we give a brief introduction into the mating pheromone response pathway in S. cerevisiae and recall the necessary Petri net definitions. Then, we propose and discuss in detail a systematic modelling method with respect to signalling pathways. We focus on the analysis of signal flows and introduce for this purpose the new concepts of feasible tinvariants and of maximal common transition sets (MCTsets). Feasible tinvariants stand for minimal selfcontained subnets being active under a given input situation, while MCTsets can be interpreted as smallest biologically meaningful functional units. Both concepts define a fully automatic net decomposition into biologically meaningful modules. Whereas MCTsets provide a decomposition of the net into disjunctive subnets, feasible tinvariants describe subnets, which generally overlap. We apply our modelling technique to the analysis of the pheromone response pathway in S. cerevisiae and discuss the results. Finally, we give conclusions indicating also further projects.
The signal transduction pathway of mating pheromone response
The signalling pathway of the mating pheromone response in S. cerevisiae is reviewed, e.g., in [3,4,30]. In S. cerevisiae there are two different mating types of haploid cells, MATα and MATa. Cells of opposite mating type can mate, i.e., fuse to a diploid cell. This process is stimulated through small peptide pheromones, the αfactor and the afactor, respectively. A haploid cell secretes the own factor and carries receptors for detecting a cell of opposite mating type via the other factor. When a yeast cell is stimulated by pheromone, secreted by a nearby cell of the opposite mating type, it starts a complex signalling pathway to undergo a series of physiological changes in preparation for mating. These changes are controlled and regulated through the pheromone signalling pathway, which is nearly the same in MATα and MATa cells.
In the following, we refer to this pathway in MATa cells, so the considered cells have an αfactorspecific cell surface receptor STE2, which is coupled to a heterotrimeric Gprotein. This Gprotein consists of the three subunits Gα, Gβ, and Gγ, whereby the latter two can act as heterodimer Gβγ. A conformation change of the receptor catalyses an exchange of GDP to GTP in the Gα subunit leading to a dissociation of this monomer. If GTP is hydrolysed to GDP, the monomer reassociates with the dimer to the trimeric form of the Gprotein [30]. The Gprotein Gβγ subunits mediate a signal on the MAP kinase cascade by interacting with Cdc24 (Far1 transmitted), protein kinase Ste20, and scaffold protein Ste5. The Ste20, Ste11, and Ste7 kinases, which are activated serially, phosphorylate and activate itself the two MAP kinases Fus3 and Kss1. But, Fus3 attenuates the activation of Kss1 [31]. Fus3PP releases the transcription factor Ste12 of its suppression through Dig1 and Dig2 proteins. (Re)Forming such a repression complex necessitates inactive Fus3 or inactive Kss1 [32]. Pheromone induced genes prepare the mating of the two cells and the fusion of their nuclei. They are responsible for cell cycle arrest in the phase G1, and the synthesis of some signalling regulators, e.g., the protease Bar1, the regulator of GProtein Sst2, and the phosphatase Msg5. Other regulations of the signalling pathway are realised through receptor endocytosis [33] or degradation of involved kinases [34].
Petri nets
Let us first recall some basic definitions of Petri nets. More formal definitions are given, e.g., in [17,18,35]. Petri nets are bipartite directed multigraphs, i.e., they consist of two types of nodes, called places P = {p_{1},..., p_{n}} and transitions T = {t_{1},..., t_{m}}, and directed arcs, which are weighted by natural numbers and connect only nodes of different type. Places model typically passive system elements as conditions, states, or biological species, i.e., chemical compounds as proteins. Transitions stand generally for active system elements as events, or chemical reactions as de/activation. In graphical representations, places are depicted as circles and transitions as rectangles. Transitions without preplaces (postplaces) are called input (output) transitions and are drawn as flat rectangles. The arcs in the net describe the causal relation between active and passive elements. They are illustrated as arrows and labelled with their weight, if it is larger than one. Arcs connect an event with its preconditions, which must be fulfilled to trigger this event, and with its postconditions, which will be fulfilled, when the event takes place.
The fulfilment of a condition is realised via tokens residing in places. Principally, a place in a discrete net may carry any integer number of tokens, indicating different degrees of fulfilment. If all preplaces of a transition are marked sufficiently (corresponding to the arc weights) with tokens, this transition may fire. If a transition fires, tokens are removed from all its preplaces and added to all its postplaces, each corresponding to the given arc weights. Thus, the tokens are the dynamic elements of the system. If a condition must be fulfilled, but the firing of an adjacent transition does not remove any tokens from it, these nodes are connected via two converse arcs. In the following, these arcs are represented by bidirectional arrows as a shorthand notation and are called read arcs.
From a biological viewpoint, the tokens residing in places indicate whether the corresponding chemical species is present, i.e., its concentration is above a certain concentration level (threshold) in the cell. This presence enables the chemical reactions modelled by the place's posttransitions to take place.
A current distribution of the tokens over all places, usually given as m ∈ , describes a certain system state and is called a marking of the net. Accordingly, the initial marking m_{0 }of a net describes the system state before any transition has fired.
The incidence matrix C of a given Petri net is an (n × m)matrix (where n denotes the number of places and m the number of transitions, compare above). Every matrix entry c_{ij }gives the token change on the place p_{i }by the firing of the transition t_{j}. Thus, the incidence matrix does not reflect read arcs. A tinvariant is defined as a nonzero vector x ∈ , which holds the equation
C·x = 0. (1)
A tinvariant represents a multiset of transitions, which have altogether a zero effect on the marking, i.e., if all of them have fired the required number of times, a given marking is reproduced. The invariant property holds for an arbitrary initial marking. A tinvariant is called realisable, if a marking is reachable, such that all transitions of the tinvariant are able to fire in a suitable partial order. Analogously, a pinvariant is defined as a nonzero vector y ∈ , which holds the equation
y·C = 0. (2)
A pinvariant characterises a token conservation rule for a set of places, over which the weighted sum of tokens is constant independently from any firing, i.e., for a pinvariant y and any markings m_{i}, m_{j }∈ , which are reachable from m_{0 }by the firing of transitions, it holds
y·m_{i }= y·m_{j}. (3)
The nodes corresponding to the nonzero entries of an invariant x are called the support of x, written as supp(x). Considering Equations (1) and (2), it is obvious that a sum of tinvariants (pinvariants) gives again a tinvariant (pinvariant). An invariant x is called minimal, if its support does not contain the support of any other invariant z, i.e.,
∄ invariant z : supp(z) ⊂ supp(x), (4)
and the greatest common divisor of all nonzero entries of x is one.
A net is covered by tinvariants (pinvariants), if every transition (place) participates in a tinvariant (pinvariant). A tinvariant (pinvariant) defines a connected subnet, consisting of its support, the support's pre and postplaces (pre and posttransitions), and all arcs in between.
Methods
Modelling
First, we want to define general network components in the language of Petri nets. Considering substructures, which model logical terms as a Petri net, an ordinary implication is a trivial case. Figure 1 pictures two subnets representing conjunctioncoupled implications, i.e., A ⇒ (B ∧ C) and (D ∧ E) ⇒ F, respectively. If there is a token on place A, the precondition A is fulfilled, and the corresponding posttransition may fire. This event will fulfil the postconditions B and C getting each a token. This example illustrates that the number of tokens in a net is not generally conserved. The fulfilment of condition F in Figure 1 depends on both preconditions D and E.
Figure 1. Two conjunction subnets. Two subnets representing conjunctioncoupled implications, i.e., A ⇒ (B ∧ C) and (D ∧ E) ⇒ F, respectively. If there is a token on place A, the precondition A is fulfilled, and the corresponding posttransition may fire. This event will fulfil the postconditions B and C getting each a token. If there is a token on the places D and E, the preconditions D and E are fulfilled, and the corresponding posttransition may fire. This event will fulfil the postcondition F getting a token.
Figure 2 depicts two subnets representing disjunctioncoupled implications, i.e., G ⇒ (H ∨ I) and (J ∨ K) ⇒ L, respectively. The two posttransitions of place G are in conflict, i.e., if there is one token on place G, both its posttransitions may fire, but only one of them can actually fire. This subnet represents an exclusive disjunction. Therefore, a nondeterministic behaviour results. In contrast to that, the pretransitions of place L are concurrent and can fire independently from each other, if any. With this subnet an inclusive disjunction is represented.
Figure 2. Two disjunction subnets. Two subnets representing disjunctioncoupled implications, i.e., G ⇒ (H ∨ I) and (J ∨ K) ⇒ L, respectively. While the subnet on the left hand side stands for an exclusive disjunction, the righthand subnet represents an inclusive disjunction. The two posttransitions of place G are in conflict, i.e., if there is one token on place G, both its posttransitions may fire, but only one of them can actually fire. This subnet represents an exclusive disjunction. In contrast to that, the pretransitions of place L are concurrent and can fire independently from each other, if one token is in each of the places J and K. With this subnet an inclusive disjunction is represented.
In all implications, presented in Figure 1 and Figure 2, the tokens in the preplaces are removed by the firing of their posttransitions. Thus, the precondition is no longer fulfilled when the connected event took place. The situation, where the state of a precondition fulfilment is preserved, is discussed below in a general form, see case 1.
A crucial point in signal transduction pathways are modifications of certain proteins, resulting in a functionally active or inactive form of these proteins. Many of them are activated by phosphorylation and deactivated by dephosphorylation, and act in different ways depending on their current state. A qualitative model allows by its structure to distinguish between different states of one protein. Modelling this situation, we use a subnet, which includes the different modifications (states) of one protein as places forming a pinvariant. A pinvariant y, which holds supp (y) = 2 may stand for a subnet, representing a negation by two places (A and ¬A), see Figure 3. The essential states are represented by the places p1 (A) and p2 (¬A) forming the pinvariant, i.e., the token depicted on place p1 circulates only between p1 and p2 by firing of t1 and t2, respectively. To preserve this invariant structure, the adjacent transitions t3 and t4, respectively, are connected via read arcs.
Figure 3. A Petri net example. A Petri net, whose places p1 and p2 form a pinvariant, representing for example a logical negation. p1 can be a condition A and p2 a condition ¬A, or vice versa, indicating two states of the system, for example an activated or deactivated protein. The place p1 carries one token to indicate in which state the system is. Transitions t3 and t4 are connected to this pinvariant by read arcs in order to hold the token on those places, which form the pinvariant.
Altogether, there are three cases in signal transduction models, which can be distinguished by their different use of read arcs:
• Case 1: A substance A does not lose its activity by interacting with a second substance B, see Figure 4.
Figure 4. Case 1 using read arcs. Case 1 of using read arcs is represented. Precondition (substance) A does not lose its fulfilment (activity) by triggering the fulfilment (activation) of a condition (substance) B.
• Case 2: A substance C triggers several events represented by transitions, which are independent of each other, see Figure 5.
Figure 5. Case 2 using read arcs. Case 2 of using read arcs is represented. Precondition (substance) C triggers several events represented by transitions, which are concurrent, i.e., independent from each other.
• Case 3: The above mentioned context of a pinvariant, see Figure 3.
The subnet presented in Figure 5 differs from the subnet on the left hand side of Figure 2 by the use of read arcs. In Figure 2 the two posttransitions of place G are in conflict. Thus this subnet represents an exclusive disjunction G ⇒ (H ∨ I). By applying read arcs in such a substructure, the posttransitions of the corresponding place are concurrent, i.e., independent from each other. Thus, the subnet in Figure 5 represents an inclusive disjunction C ⇒ (D ∨ E).
Before turning to the next step, the model validation, let us summarise the main steps of the systematic model development:
1. compilation of biological knowledge about the signalling pathway of interest, e.g., from the literature or by database search,
2. translation of the interactions of relevant biological species into basic logical terms,
3. transferring these logical terms into net components,
4. assembling these net components to one Petri net,
5. if necessary, include input and output transitions ensuring that there are no places without pre or posttransitions in the net.
The last but one point ensures that the developed Petri net is connected. The input and output transitions introduced in the last point make the net to a transitionbordered one. They represent the interface of the biological system to its surroundings.
Model validation
The input transitions of a transitionbordered Petri net cause unboundedness, i.e., there is no upper bound for the number of tokens in such a net. Therefore, the state space of the net, i.e., all reachable markings, is infinite, and the dynamic net properties cannot be decided by the computation of the reachability graph, which contains all reachable markings as nodes. But, we do get information about the model dynamics via the analysis of the net structure, represented by the incidence matrix. Thus, our approach to validate the model is mainly based on the net invariants. The definition of invariants, see Equations (1) and (2), was introduced in [36]. The application of a system's minimal tinvariants to analyse metabolic networks in the steady state was proposed in [9] by the definition of elementary modes. In the context of metabolic networks pinvariants are used to model substrate conservations, see, e.g., [23,37]. However, in order to apply these techniques, which have been proven to be useful in the context of metabolic networks, to signal transduction networks, we have to adjust them first to our special needs.
Pinvariants in signal transduction models represent also some kind of conservation relationship, however in the sense of the several modifications of a given species as introduced in Figure 3. A species cannot be consumed or produced, it can change its current state only, and it has to be always in exactly one state. The first step of model validation aims at checking the minimal pinvariants for their biological plausibility. Since the weighted sum of tokens over a pinvariant is constant (see Equation 3), there have to be always some tokens residing in at least one place of each minimal pinvariant ensuring that this part of the net may contribute to the system behaviour. Moreover, because a given species can always be in one state only, each pinvariant gets just one token, here. These tokens are typically placed in the pinvariants in such a way that the initial marking of the net represents an inactive state of the biological system.
The next step is to consider all possible signal flows through the network represented by realisable tinvariants, and to check their biological meaning. All possible tinvariants can be computed as nonnegative linear combination of the minimal ones, see Equation (1). In former case studies, which do not contain the special pinvariant construct, see Figure 3, it was sufficient to discuss the minimal tinvariants, which are always realisable by construction [14]. Including pinvariants, we have to process the minimal tinvariants to get realisable ones, defining minimal selfcontained subnets, which are active under a given input situation. For this purpose, we introduce the concept of feasible tinvariants.
Feasible tinvariants
The calculation of tinvariants does not take into account the initial marking. Moreover, read arcs are not reflected in the incidence matrix. Therefore, the question arises, whether a minimal tinvariant is really realisable. We assume that read arcs are the only source for nonrealisability, i.e., there is a sufficient token number in each of the minimal pinvariants in the initial marking.
The use of read arcs corresponding to the first case described above is discussed in [15]. For an analysis, these read arcs are replaced by unidirectional ones in the direction of the main token (signal) flow. The same rationale holds true for the replacement of the read arcs corresponding to the second case. But those ones, which are connected with (see the third case), have to remain in the net in order to keep the tokenpreserving structure of the pinvariant. As mentioned above, read arcs are not reflected in the incidence matrix. Therefore, depending on a given marking, some of the minimal tinvariants may not be realisable considering them in a net with read arcs. For example, the pinvariantadjacent transitions t3 and t4 in Figure 3 are treated as apparent input transitions, because their connection with the places p1 and p2, respectively, is not reflected in the incidence matrix. If this kind of transitions is part of a tinvariant, whereby their preplaces (connected via read arcs) carry no tokens, they are actually not able to fire. Therefore, the minimal tinvariants have to be processed to get realisable tinvariants for a suitable marking.
In accordance with our objective of model validation, we are looking for minimal selfcontained tinvariants, which are realisable in the initial marking, i.e., minimal multisets of transitions, which can fire in an appropriate order without the firing of the other transitions, not belonging to that tinvariant. Those tinvariants are called feasible tinvariants. To make a tinvariant feasible, all of its involved transitions have to be able to fire, i.e., for each of them it is true either that it is an input transition or that its preplaces enable it to fire. The latter case means that either the considered preplaces carry already sufficient tokens in the initial marking or these preplaces get sufficient tokens via the firing of other transitions involved in this tinvariant. We distinguish between realisable and feasible tinvariants, because the notion of realisability of a tinvariant is a general term of Petri net theory, referring to the reachability of a marking, where the tinvariant's transitions can fire. Feasible tinvariants are special realisable tinvariants, which are minimal ones concerning their realisability in the initial marking. They are introduced because of a special net structure, which causes the interruption of the tinvariant, when it leads for example to a pinvariant, which carries a special marking.
Table 1 lists all feasible tinvariants of the net in Figure 3, consisting of all minimal tinvariants and one additional processed one. The processing of minimal tinvariants looks for minimal (nonnegative) linear combinations, resulting into feasible tinvariants z. Let us formulate this processing of the minimal tinvariants in a general form. For each read arc between a transition t_{j }and an empty place p_{i }of a pinvariant the following procedure has to be applied. Let t_{1},...,t_{k }be all the pretransitions of p_{i}, which are not connected via a read arc with p_{i}, see Figure 6. If there are no tokens residing in p_{i }in the initial marking, a tinvariant is not realisable if it includes t_{j }and excludes t_{l }∀ l ∈ {1,..., k}. The following combinations z of feasible tinvariants x and one nonfeasible tinvariant y reestablish the given read arc between transition t_{j }and place p_{i}
Table 1. The tinvariants of the example in Figure 3.
Figure 6. A subnet representing the feasibility of minimal tinvariants. The feasibility of minimal tinvariants is represented to illustrate Equation (5). Place p_{i }belongs to a pinvariant. It is adjacent to the pretransitions t_{1}, t_{2},...t_{k }and connected to the transition t_{j }by a read arc. X and Y stand for the sets of tinvariants including the pre and the posttransitions, respectively, which are contained in the corresponding dashed rectangles. Since p_{i }does not carry tokens, tinvariants are not feasible if they contain transition t_{j }without containing at least one of the transitions t_{1}, t_{2},...t_{k }. Thus, the read arc has to be bridged as described in Equation (5) to get feasible tinvariants.
z = x + y ∀x ∃ l ∈ {1,...,k} : x_{l }≠ 0
with x_{j }= 0
and y_{j }≠ 0 ∧ ∀ l ∈ {l,...,k} : y_{l }= 0. (5)
Note that for each feasible minimal tinvariant x providing a token on the place p_{i }one linear combination z is constructed. Appropriate combinations have to be performed iteratively, if there occur several read arcs in the net, which have to be bridged. The combined tinvariants do not hold the criterion to be minimal, see Equation (4). But, since only inevitable combinations of minimal tinvariants are constructed, see the restrictions in Equation (5), the resulting combined tinvariants are minimal with respect to their realisability in the initial marking. Moreover, they only have nonnegative entries, because they are build as a sum of two tinvariants, which are defined as vectors in , see Equation (1).
Feasible tinvariants define, similarly to the tinvariants, connected subnets, characterised by its support. Generally, these subnets overlap. They represent minimal selfcontained subnets being active under a given input situation. So, each of these subnets stands for a possible signal flow in the net. The examination of the feasible (minimal or combined) tinvariants for their biological plausibility is a central step of model validation. It has to be checked, whether each feasible tinvariant has a biological meaning and whether every modelled part of the considered signalling pathway is reflected in a corresponding feasible tinvariant [14].
Finally, it has to be checked, whether the net is covered by tinvariants. This property ensures that every transition participates in a tinvariant, i.e., every biological atomic action in the model may take place as part of the basic behaviour of the net.
While minimal tinvariants are computed based on the net structure only, feasible tinvariants are constructed with respect to the marking of the net. Summarising the discussion above, the following situations are possible: the minimal tinvariants can reach (1) from input to output transitions, (2) from input transitions to minimal pinvariants, (3) from minimal pinvariants to minimal pinvariants, or (4) from minimal to output transitions. The transitions of a tinvariant influence a minimal pinvariant in such a way that altogether its initial state is reproduced. Not every possible combination is considered to bridge a read arc and to connect minimal tinvariants. Such a procedure would lead to tinvariants reaching from input transitions to output transitions, only. As described above, those minimal tinvariants are combined, which meet at a minimal pinvariant with a corresponding place not carrying a token.
Maximal common transition sets (MCTsets)
In order to support the check of feasible tinvariants for their biological meaning, the transitions are grouped into socalled maximal common transition sets (MCTsets) by their occurrence in the minimal tinvariants: ∀i, j ∈ {1,...,m} the transitions t_{i }and t_{j }are grouped into the same MCTset, if and only if they participate in exactly the same minimal tinvariants, i.e., all tinvariants x hold
χ_{{0}}(x_{i}) = χ_{{0}}(x_{j}), (6)
whereas χ_{{0} }denotes the characteristic function, binary indicating if an argument is equal to zero. This supportoriented grouping leads to maximal sets of transitions, where each set of transitions ϑ holds
∀x ∈ X : ϑ ⊆ supp(x) ∨ ϑ ∩ supp(x) = ∅, (7)
whereas X denotes the set of all minimal tinvariants x.
The grouping according to Equation (6) represents an equivalence relation in T, the set of transitions, which leads to a partition of T. The equivalence classes ϑ are the MCTsets. Transitions, not contained in any tinvariant, form their own MCTset. MCTsets define subnets, but not necessarily connected ones. An example of a disconnected MCTset will be given below in the presented model.
The resulting MCTsets are disjunctive and represent a possible decomposition of large biochemical networks into rather small subnets, which can be read as functional units. Because of the way, in which they are generated, each of the MCTsets may represent a signalling unit or a building block with its own biological meaning. In the following, only those MCTsets are considered, which contain more than one transition.
Results and discussion
Derivation of the Petri net model
In accordance with the elucidation above we developed a Petri net, modelling the signal transduction pathway of the mating pheromone response in S. cerevisiae. The Petri net in Figure 7 presents – to the best of our knowledge – for the first time a qualitative model of the pheromone pathway, and it extends the ODE model given in [6]. The net consists of 42 places and 48 transitions, which are listed by their name and biological meaning in Tables 3 and 4, respectively.
Figure 7. The Petri net model of the mating pheromone response pathway of S. cerevisiae. The Petri net, modelling the signal transduction pathway of the mating pheromone response in S. cerevisiae. The meaning of the places is listed in Table 3 and of the transitions in Table 4. The logical nodes are coloured in grey. A logical node is identified by its name and exists in multiple copies in the net, which are logically identical. This construct is mainly used to avoid immoderate arc crossings. The transitions, which are contained in the seven MCTsets of Table 5, are coloured differently.
Table 3. The places of the model.
Table 4. The transitions of the model.
Proteins, protein complexes, and other chemical compounds are represented as places, and complex formation/cleavage, protein de/phosphorylation, and other reactions by transitions. Tables 3 and 4 show in their right column, which chemical compounds and which biological events are considered in the model. Generally formulated, the model contains a complete pheromone response pathway, i.e., activation and composition of the pheromone receptor complex, composition of the MAP kinase cascade complex and MAP kinase cascade, transcription factor activation, and transcription of genes, which response to the mating pheromone. Moreover, there are several positive and negative feedback regulations included in the model, i.e., regulation of the ligand αfactor (via Bar1), of the receptor Ste2 (endocytosis induced by Yck1 and Yck2), of Gprotein subunits (via Sst2), of the MAPK cascade (via labelling for degradation), and of the regulation of transcription (via repression of the transcription factor Ste12).
We take into account the above mentioned substructures, e.g., the place complex2 (p17) does not lose its activity by firing of transition Ste11_phos_Ste7 (t18), i.e., in the MAPKcomplex the protein Ste11 remains in its active form after activating the protein Ste7, see case 1. The place Fus3PP (p20) triggers several transitions Ste12_inhibit_phos t22), Ste12_phos (t24), Far1phos (t31), and Sst2_phos (t34), which are independent of each other, see case 2. The place pair Kss1_phos (p38) and unphos_Kss1 (p39) forms a pinvariant as discussed in case 3.
With one exception, there exist only arcs weighted by 1 because no substance flow with stoichiometric relations takes place, but signal flow is described by the model. This special arc is weighted by 2 to reduce the token number in the structural cycle, which is formed by the places complex3 (p18), complex4 (p19), compl_without_Fus3 (p21), and the transitions Ste7_phos_Fus3 (t19), Fus3PPrelease (t20), and binding_free_Fus3 (t21). By using this arc weight, it is ensured that the activation of the transition Ste7_phos_Fus3 (t19) depends on firing of transition Ste11_phos_Ste7 (t18), see Figure 7. In the biological context, this means that the crucial phosphorylation of the last kinases in the MAPK cascade depends on the activation of the last but one.
In the initial marking there is one token residing in the place inact_Far1 (p31) and Sst2_in_nucleus (p34), respectively, which symbolise that these proteins are already present in the cell in low concentration independently of a signal.
Application of Petri net concepts
Pinvariant analysis
There are three minimal pinvariants in the net. Table 2 contains these pinvariants as well as the corresponding biological meaning of them. Pinvariants 1 and 2, respectively, stand for the modification of the dimeric and monomeric, respectively, subunits of the receptor coupled Gprotein, which occur either in the trimeric form represented in the place trimer_bound_to_receptor (p5) or which is dissociated in its subunits G_beta_gamma_dimer (p7) and G_alpha_GTP (p6). Therefore, both of these pinvariants include the place p5 in addition to one of the places p6 and p7, respectively. The places Kss1_phos (p38) and unphos_Kss1 (p39) indicate the modification of the involved kinase Kss1 and form the third minimal pinvariant according to case 3. Thus, all these minimal pinvariants can be interpreted in a biologically reasonable meaning.
Table 2. The minimal pinvariants and their biological meaning.
In the initial marking of the net there are some tokens already residing in special places. According to the elucidation in section Model validation, the places trimer_bound_to_receptor (p5) and unphos_Kss1 (p39) carry tokens ensuring that the minimal pinvariants contribute to the system behaviour.
Tinvariant analysis
For our model we got ten minimal tinvariants covering all transitions. As mentioned above, the transitions are grouped into MCTsets, if and only if they participate in exactly the same minimal tinvariants, see Equation (6). If a huge number of tinvariants arises for a network, the concept of MCTsets provides additional structure information. Due to their definition, MCTsets represent sets of events, which occur only together. These sets, together with their pre and postplaces and all arcs in between, should symbolise signalling functional units of the model with an own biological meaning, see Table 5. For example, MCTset 1 consists of the transitions MATalpha_cell(surrounding) (t1), receptor_synthesis (t3), binding_factor_to_receptor (t2), and receptor_conformation_change (t4). Therewith, this MCTset represents the synthesis of Ste2, its binding to the present αfactor and its resulting conformation change. In Figure 7 the MCTsets, which contain more than one transition, are shown in different colours. As mentioned above, the MCTsets are not necessarily connected subnets. MCTset 5 gives an example for a disconnected one because it contains techn_input (t44,), compare Table 5 and Figure 7.
Table 5. The maximal common transition sets.
The necessarily in the net remaining read arcs endanger the feasibility, i.e., the realisability in the initial marking, of the minimal tinvariants. Following our modelling approach, minimal tinvariants are not feasible, if they involve a transition, which is connected via a read arc with an empty preplace, which does not get tokens by firing of other transitions involved in this tinvariant. To ensure the feasibility of all tinvariants, these nonfeasible ones are joint with some others, which provide tokens on the critical preplaces. The read arcs adjacent to the place G_beta_gamma_dimer (p7) remain in the net because of the minimal pinvariant 1. Therefore, the minimal tinvariants 3, 4, 5, and 6 do not fulfil the feasibility criterion, because they contain the MCTset 3, but do not contain any transitions, which provide tokens at place p7, which is connected with the MCTset 3 via read arcs. These nonfeasible tinvariants have to be processed in the way explained above, see Equation (5). Only the minimal tinvariant 1 involves transition division (in_alpha_subunit:GDP→GTP) (t5), which provides tokens on place p7 and does not involve the MCTset 3. Therefore, each of the four nonfeasible tinvariants is combined with tinvariant 1. These combinations provide four feasible, nonminimal tinvariants. Table 6 lists the feasible tinvariants and indicates if they are still minimal, i.e., nonprocessed.
Table 6. The tinvariants.
We want to discuss the biological plausibility of the feasible tinvariants. Table 7 contains the biological meaning of the tinvariants in Table 6. They all include the receptor synthesis and the binding of the pheromone factor of a MATα cell in the surroundings, causing a receptor conformation change (MCTset 1). In tinvariant 1 the activated receptor causes the dissociation of the Gprotein (transition t5), whose subunits reassociate to a trimeric form (transition t6). In the tinvariant 2 the proteins Yck1 and Yck2 are bound through Akr1 to the plasma membrane and phosphorylate the activated receptor. This leads to a receptor endocytosis (MCTset 7) via an ubiquitination. The tinvariants 7 to 14 include a signal transduction via the Gprotein (transition t5) and the MAP kinase cascade (MCTset 3). A release and phosphorylation of the transcription factor Ste12 (MCTset 4) is involved in the tinvariants 7 to 10. This leads in these tinvariants to a transcription of the pheromone response genes preparing the cell for mating, including a feedback via Bar1 and Sst2 (MCTset 5). Far1 is transported out of the nucleus (MCTset 2) in tinvariants 8 and 10, while it initiates the cell cycle arrest (transition t32) in the tinvariants 7 and 9. By containing MCTset 6, the tinvariants 9 and 10 include a Ste12 repression via activated Fus3. In tinvariants 7 and 8, kinase Kss1 is activated and deactivated (transitions t40 and t42, respectively). This pinvariant 3 contributes also to the processes described in tinvariant 14, where Kss1 is phosphorylated and dephosphorylated (through transitions t40 and t41, respectively). Generally, the tinvariants 11 to 14 represent signalling pathways, which do not lead to a mating of the cell. All these tinvariants consist only of the receptor activation (MCTset 1) and the MAP kinase cascade (MCTset 3). Only tinvariant 13 involves an additional activation of Ste12 (MCTset 4), which is repressed via Kss1 (transition t43). The tinvariants 11 and 12 include a feedback inhibition via a protein degradation (transition t38), which is initiated through a phosphorylation of these proteins (transitions t37 and t39, respectively).
Table 7. The biological meaning of the feasible tinvariants.
Recapitulating, tinvariants 7 to 10 lead to a mating of the cell, while the other tinvariants enable the cell to regulate, and modulate a pheromone induced signal by attenuated response or response desensitisation.
Altogether it can be summarised that the known signalling processes are represented in the feasible tinvariants.
Theoretical knockout experiments
The Petri net model can serve as basis of theoretical knockout experiments. They can be constructed by deleting the corresponding place and adjacent transitions, modelling the absence of this compound in a null mutant. Animating the resulting net shows already where some token accumulations arise under the new situation. These affected compounds can characterise a null mutant. Theoretical knockout experiments by the Petri net model could point out, which compounds indicate a mutant and nominate possible indicators for biological exploration.
Furthermore, a deletion of transitions leads to a deletion of feasible tinvariants, at least in a net, which is covered by tinvariants. The remaining feasible tinvariants show, which events still take place in this new cell type. For example, it is known [30] that Ste5 null mutant cells are unable to mate or to arrest the cell cycle. This experiment can be reconstructed in our Petri net model by deleting the place Ste5(scaffold) (p11) and its adjacent transitions binding_to_Ste5 (t12) and Ste5_binds_Ste11 (t13). This results in the deletion of the MCTset 3, see Table 5. The feasible tinvariants 7 to 14 do not exist in the new mutant net because they include MCTset 3, compare Table 6. The transitions cell_fusion (t28) (i.e., MCTset 5), and celI_cycle_arrest_in_G1 (t32) occur only in the feasible tinvariants 7 to 10, which were deleted. Thus, the Ste5 mutant cells modelled by that mutant net are neither able to arrest the cell cycle nor to undergo a cell fusion, because the corresponding transitions do not occur in the feasible tinvariants of the net.
Conclusion
In this paper we describe a systematic approach to model and analyse signal transduction pathways using qualitative Petri nets. We propose a stepwise model development via the translation of the biological interactions into logical terms, which then in turn are transferred into net components.
We explain the model validation step with a strong focus on the invariant analysis. In our modelling approach, we use minimal pinvariants, e.g., for the representation of the switching behaviour representing a compound in its activated and deactivated form. These pinvariants are embedded into the whole network by read arcs. Therefore, minimal tinvariants have generally to be processed to get feasible tinvariants. According to the construction principle, feasible tinvariants correspond to minimal selfcontained subnets active under a given input situation. So, feasible tinvariants have to be checked for their biological meaning.
Furthermore, we define maximal common transition sets, which provide a decomposition of the net into not necessarily connected subnets. These structures can be interpreted as the smallest biologically meaningful functional units or building blocks of a given network.
The new concepts of feasible tinvariants and MCTsets have been proven to be useful for model validation and the interpretation of the biological system behaviour by their fully automatic approach to decompose a given network into biologically meaningful modules. Please note that it is possible to build the MCTsets based on the feasible tinvariants instead of the minimal ones. That makes sense in Petri nets, which contain more substructures of minimal pinvariants. MCTsets provide a decomposition of the net into disjunctive subnets, and feasible tinvariants describe subnets, which can overlap.
We illustrate this approach using the mating pheromone pathway in S. cerevisiae. The validated model denotes the known signal flows through the net, which are obtained by the invariant analysis. We are able to assign a reasonable biological meaning to all feasible tinvariants. These reflect correctly the signal transduction in such a system, i.e., the signal response behaviour. Thus, Petri net theory is also useful to model and analyse signalling pathways considering not a substance flow, but a flow of information (signals) through the net.
The Petri net model may serve as basis for theoretical knockout experiments. By deleting appointed places and adjacent transitions, null mutants can be modelled. The analysis results of the original net can be easily modified to get some information about the derived mutant cell.
Having once validated the structure of the qualitative discrete Petri net, several applications and extensions are possible. The net may be refined and extended to a quantitative one by including known or estimated kinetic parameters (e.g., as concentrations, reaction rates, equilibrium constants), using continuous or hybrid Petri nets, respectively [38,39].
In this way, the resulting quantitative Petri net preserves the structure of the underlying net topology. Continuous Petri nets may be considered as a structured ODEs description. They can be evaluated by standard differential equation solvers. For a related case study see [40]. Hybrid Petri net comprise generally discrete as well as continuous parts. Their evaluation requires dedicated simulation approaches. For a related case study see [24]. In both cases, the quantitative simulation will allow the observation of signal flows, already revealed by the underlying discrete model. However, contrary to our qualitative approach, where signal flows are represented explicitly by the computed feasible tinvariants, signal flows in the quantitative model are given only implicitly by the simulation results.
In this paper, we have chosen deliberately a rather small running example to be able to present all steps of the analysis procedure in full detail, which should help the reader to mimic the approach using his/her own case studies. It would be worth applying our approach to larger networks as presented in [41,42].
Software tools and supplementary material
The development of the net has been done using the graphical Petri net editor and animator Snoopy [43] described in [44]. The analyses have been performed using the Integrated Net Analyzer INA [45]. Both tools are freely available. They are running under Windows as well as under Unix/Linux. We provide the Petri net model and the INA analysis result file in the supplementary material [46].
Authors' contributions
The paper is a product of a strong cooperation between all three authors. The key concepts and case study has been elaborated by AS during the work on her diploma thesis. MH gave substantial and valuable contributions in the application of Petri net techniques. IK defined the topic and gave substantial contributions to the conceptional design of the research project. The editorial work has been done by all three authors. All authors read and approved the final manuscript.
Acknowledgements
This work has been partly supported by the Federal German Ministry of Education and Research (BMBF), BCB project 0312705D. We would like to thank Bente Kofahl, Edda Klipp, and Günter Czichowski for the fruitful discussions.
References

BlumeJensen P, Hunter T: Oncogenic kinase signalling.
Nature 2001, 355365. PubMed Abstract  Publisher Full Text

Wang Y, Dohlman HG: Pheromone signaling mechanisms in yeast: a prototypical sex machine.
Science 2004, 306(5701):15081509. PubMed Abstract  Publisher Full Text

Bardwell L: A walkthrough of the yeast mating pheromone response pathway.
Peptides 2004, 26(2):339350. Publisher Full Text

Gustin MC, Albertyn J, Alexander M, Davenport K: MAP kinase pathways in the yeast Saccharomyces cerevisiae.
Microbiol Mol Biol Rev 1998, 62(4):12641300. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Ciliberto A, Novak B, Tyson JJ: Mathematical model of the morphogenesis checkpoint in budding yeast.
J Cell Biol 2003, 163(6):12431254. PubMed Abstract  Publisher Full Text

Kofahl B, Klipp E: Modelling the dynamics of the yeast pheromone pathway.
Yeast 2004, 21(10):831850. PubMed Abstract  Publisher Full Text

Qu Z, Weiss JN, MacLellan WR: Coordination of cell growth and cell division: a mathematical modeling study.
J Cell Sci 2004, 117(18):41994207. PubMed Abstract  Publisher Full Text

Yi TM, Kitano H, Simon MI: A quantitative characterization of the yeast heterotrimeric G protein cycle.
Proc Natl Acad Sci 2003, 100(19):1076410769. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Schuster S, Hilgetag C, Schuster R: Determining elementary modes of functioning in biochemical reaction networks at steady state.

Klamt S, SaezRodriguez J, Lindquist JA, Simeoni L, D GE: A methodology for the structural and functional analysis of signaling and regulatory networks.
BMC Bioinformatics 2006., 7(56) PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Simao E, Remy E, Thieffry D, Chaouiya C: Qualitative modelling of regulated metabolic pathways: application to the tryptophan biosynthesis in E. Coli.

ZevedeiOancea I, Schuster S: A theoretical framework for detecting signal transfer routes in signalling networks.
Computers and Chemical Engineering 2005, 29:597617. Publisher Full Text

Heiner M, Koch I, Will J: Model validation of biological pathways using Petri nets – demonstrated for apoptosis.
Proc First Int Workshop on Computational Methods in Systems Biology, (CMSB 2003) Rovereto, LCNS 2602 2003, 173.

Heiner M, Koch I, Will J: Model validation of biological pathways using Petri nets – demonstrated for apoptosis.
Biosystems 2004, 75(1–3):1528. PubMed Abstract  Publisher Full Text

Heiner M, Koch I: Petri net based model validation in systems biology.
Proceedings of the 25th International Conference on Applications and Theory of Petri Nets, Bologna, LCNS 3099 2004, 216237.

Petri CA: Communication with automata (in German). Institut für instrumentelle Mathematik, Bonn: Schriften des IIM Nr. 3; 1962.

Murata T: Petri nets: Properties, analysis and applications.
Proceedings of the IEEE 1989, 541580. Publisher Full Text

Starke PH: Analysis of Petri net models (in German). Stuttgart: Teubner Verlag; 1990.

Reddy VN, Mavrovouniotis ML, Liebman MN: Petri net representation in metabolic pathways.
Proc Int Conf Intell Syst Mol Biol 1993, 328336. PubMed Abstract

Hofestädt R: A Petri net application of metabolic processes.
Journal of Systems Analysis, Modelling and Simulation 1994, 16:113122.

Reddy VN, Liebman MN, Mavrovouniotis ML: Qualitative analysis of biochemical reaction systems.
Comput Biol Med 1996, 26(2):924. PubMed Abstract  Publisher Full Text

Koch I, Junker BH, Heiner M: Application of Petri net theory for modelling and validation of the sucrose breakdown pathway in the potato tuber.
Bioinformatics 2005, 21(7):12191226. PubMed Abstract  Publisher Full Text

Voss K, Heiner M, Koch I: Steady state analysis of metabolic pathways using Petri nets.
In Silico Biol 2003, 3(3):367387. PubMed Abstract

Matsuno H, Tanaka Y, Aoshima H, Doi A, Matsui M, Miyano S: Biopathways representation and simulation on hybrid functional Petri net.
In Silico Biol 2003, 3(3):389404. PubMed Abstract

Chen M, Hofestädt R: A medical bioinformatics approach for metabolic disorders: Biomedical data prediction, modeling, and systematic analysis.

Doi A, Fujita S, Matsuno H, Nagasaki M, Miyano S: Constructing biological pathway models with hybrid functional Petri nets.
In Silico Biol 2004, 4(23):271291. PubMed Abstract

Hardy S, Robillard PN: Modelling and simulation of molecular biology systems using Petri nets: modelling goals of various approaches.
J Bioinform Comput Biol 2004, 2(4):595613. PubMed Abstract  Publisher Full Text

Pinney JW, Westhead DR, McConkey GA: Petri net representations in systems biology.
Biochem Soc Trans 2003, 31(6):15131515. PubMed Abstract  Publisher Full Text

Will J, Heiner M: Petri Nets in biology, chemistry, and medicine – bibliography. In Computer Science Reports 04/2002. Brandenburg University of Technology at Cottbus; 2002.

Dohlman HG, Thorner JW: Regulation of G proteininitiated signal transduction in yeast: paradigms and principles.
Annu Rev Biochem 2001, 70:703754. PubMed Abstract  Publisher Full Text

Elion EA, Qi M, Chen W: Signal transduction. Signaling specificity in yeast.
Science 2005, 307(5710):687688. PubMed Abstract  Publisher Full Text

Bardwell L, Cook JG, Voora D, Baggott DM, Martinez AR, Thorner J: Repression of yeast Ste12 transcription factor by direct binding of unphosphorylated Kss1 MAPK and its regulation by the Ste7 MEK.
Genes Dev 1998, 12(18):28872898. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Hicke L, Zanolari B, Riezman H: Cytoplasmic tail phosphorylation of the alphafactor receptor is required for its ubiquitination and internalization.
J Cell Biol 1998, 141(2):349358. PubMed Abstract  Publisher Full Text

Esch RK, Errede B: Pheromone induction promotes Ste11 degradation through a MAPK feedback and ubiquitindependent mechanism.
Proc Natl Acad Sci 2002, 99(14):91609165. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Baumgarten B: Petri nets basics and applications (in German). Heidelberg, Berlin, Oxford: Spektrum Akademischer Verlag; 1996.

Lautenbach K: Exact liveness conditions of a Petri Net class (in German). Bonn: GMD Report 82; 1973.

ZevedeiOancea I, Schuster S: Topological analysis of metabolic networks based on Petri net theory.
In Silico Biol 2003, 3(3):323345. PubMed Abstract

David R, Alla H: Discrete, continuous, and hybrid Petri nets. Berlin: Springer Verlag; 2005.

Lonitz K: Hybrid systems modelling in engineering and life sciences. Master's Thesis, Universität KoblenzLandau; 2005.

Gilbert D, Heiner M: From Petri nets to differential equations – an integrative approach for biochemical network analysis.
Proceedings of the 27th International Conference on Applications and Theory of Petri Nets, Turku, LCNS 4024 2006, 181200.

Schoeberl B, EichlerJonsson C, Gilles ED, Muller G: Computational modeling of the dynamics of the MAP kinase cascade activated by surface and internalized EGF receptors.
Nat Biotechnol 2002, 20(4):370375. PubMed Abstract  Publisher Full Text

Oda K, Matsuoka Y, Funahashi A, Kitano H: A comprehensive pathway map of epidermal growth factor receptor signaling.
Mol Syst Biol 2005., 1 PubMed Abstract  Publisher Full Text

Snoopy – Petri net editor and animator [http://wwwdssz.informatik.tucottbus.de/] webcite

Fieber M: Design and implementation of a generic and adaptive tool for graph manipulation (in German). Master's Thesis, Brandenburg University of Technology at Cottbus; 2004.

INA – The Integrated Net Analyzer [http://www2.informatik.huberlin.de/~starke/ina.html] webcite

Supplementary material [http://www.tfhberlin.de/bi/pheromone/] webcite