When rule-based models need to count

08/07/2017 ∙ by Pierre Boutillier, et al. ∙ Harvard University 0

Rule-based modelers dislike direct enumeration of cases when more efficient means of enumeration are available. We present an extension of the Kappa language which attaches to agents a notion of level. We detail two encodings that are more concise than the former practice.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 When enumeration is necessary

The following motivating example [4] demonstrates a typical problem in which levels are necessary:

KaiC proteins have 6 independent phosphorylation sites. (De)phosphorylation of every site is independent. The more sites are phosphorylated, the bigger the probability that KaiC binds KaiA is.

A typical way to deal with this example consist in explicitly encoding rules for the internal states of the sites of interest. However, doing so induces an exponential blow-up in the number of rules (in the number of levels).

The BNGL [3] language introduces a notion of indistinguishable sites. i.e. one can define an agent with sites that all have the same name. Consequently, a single rule to specify that sites (out of ) are phosphorylated is enough. Moreover, the number of species is also reduced. Still, there are exponentially many ways to go from one level to another and enumeration is necessary to faithfully respect the dynamics of the system.

We now show that adding a syntactic layer to Kappa that offers support for counting will avoid the explosion in the number of rules.

2 Encoding counters

In Kappa, agents are typed and their signature is given. In the extension we are presenting, counters are part of an agent’s signature and the upper bound of the counter has to be specified.

Extended rules can test counter values (“equals” and “bigger than”). They can also modify their value (increase, decrease or assign a new value). Lastly, levels can be used in the rates of the rules.

We now present two possible ways of encoding counters.

2.1 Unary numbers

We define a new agent type succ with 2 sites p and n. A chains of succ agents denotes a counter with value . Agents equipped with counters are encoded as agents with an extra site for each counter, which is used to bind to a chain

of succ agents. For example, in the figure on the right, an agent A has a counter c set to 111Representing level by a chain of length means that we do not have to make a special case for it in the rules!.

Testing whether a counter is equal to consists in checking whether there is a chain of succ agents. A greater than test checks whether there is a chain succ agents, where the site n of the last succ agent does not matter.

A counter is incremented by adding succ agents between the agent and its succ chain. The rule depicted on the right increments c of A by one. Removing the beginning of the succ chain decreases the level.

It is important to stress that counter modifications are independent of its value as the encoding only manipulates the beginning of a succ chain.

In our encoding, a rule whose rate depends on levels is expanded into several rules, one for each level. This operation imposes the user-defined upper bound on the levels.

Creation of agents with levels also creates the necessary chain of succ to represent the levels.

Deletion is more problematic: the chain of succ is disconnected from the deleted agent, but not deleted. A possible solution is to collect free chains of succ by using a rule that says that succ agents with their sites p free are deleted “at infinite speed”.

2.2 Ruler

A second encoding allows tests “smaller than” in addition of “equal” and “greater than”, but is also more verbose and increases the size of states.

As before, counters are encoded as a chain of succ agents. In this encoding however, a succ agent has 3 sites: p and n to form chains but also a, which is where the other agents bind. Every agent with a counter bounded by has attached a chain of (always) succ. The value of the counter is given by which succ agent it is bound to on site a. For example, in the figure on the right, an agent A has a counter c set to .

Since site p and n are distinct, a direction can be attached to the chain of succ. The level is if the agent is bound to the “bottom” of the chain, and it is if it is bound to the “top”.

Incrementing/decrementing the levels are implemented by sliding on the succ chain. Testing whether a level is equal or greater than consists of inspecting the chain of succ “below” the connection to the agent. Whereas inspecting the length of the chain “above” the connection to the agent enables to test if the level is smaller than a value .

3 Conclusions

While fairly trivial, encoding counters in the Kappa simulator has greatly simplified some of the models written by the Kappa team. However, it comes with two drawbacks. First, there is a computational cost (mainly in term of memory management) to synthetize and degrade succ agents. Secondly, the simulator works with a plain Kappa model. There is an implementation cost to go back from the plain Kappa model to the user-written model, which is necessary when giving feedback to the user (for example, dumping the current state of the system). As a result, we plan to implement a native treatment of the counters in the Kappa simulator.

Another direction for future work is the static detection of overflow for counters. Currently, the maximal level of the counter has to be declared by the user. Nevertheless, this means the rules have to be written such that they can never increase a level beyond the defined boundary. The level of the counter can be checked by static analysis using interval computation, but for the meantime we have simply added watchdogs that dynamically raise the alarm if the site of a level becomes free, or if a chain of succ that is too long appears.

References

  • [1] Pierre Boutillier, Jérôme Feret, and Jean Krivine. Kappa and KaSim development page. http://kappalanguage.org.
  • [2] Vincent Danos, Jérôme Feret, Walter Fontana, Russell Harmer, and Jean Krivine. Rule-based modelling, symmetries, refinements. In Formal Methods in Systems Biology, First International Workshop, FMSB 2008, Cambridge, UK, June 4-5, 2008. Proceedings, pages 103–122, 2008.
  • [3] James R. Faeder, Michael L. Blinov, and William S. Hlavacek. Rule-Based Modeling of Biochemical Systems with BioNetGen, pages 113–167. Humana Press, Totowa, NJ, 2009.
  • [4] Susan S Golden. Integrating the circadian oscillator into the life of the cyanobacterial cell. In Cold Spring Harbor symposia on quantitative biology, volume 72, pages 331–338. Cold Spring Harbor Laboratory Press, 2007.