Declarative approaches and tools for constructing and analyzing biological networks
Context and goal
In a post-genomic perspective, understanding biological networks is a fundamental issue. Modelling such networks is becoming an essential part of the work of today’s biologists. Predictions deduced from such models help them to go further in the analysis, particularly in designing new experiments.
Our goal is to develop software tools to assist biologists for constructing such models. Difficulties mainly arise because that part of the information to construct a model is not available. The usual way to proceed is rather empirical: modelers make “reasonable guesses”, for example to value parameters, and if the resulting model does not verify an observation, they try to adjust this unique model so that it verifies the observation. On the contrary, our AI-oriented approach relies on specifying all available knowledge in terms of constraints (logical formulae). In this way we can get automatically the set of the models which are consistent with the current level of knowledge: they are the solutions of the constraints. The predictions which can be obtained from the modeling are theorems deduced from the constraints. If the set of models is empty, it means that the constraints are inconsistent, in which case the constraints can be revised: for example by removing automatically as few constraints as possible or by adding/suppressing interactions or species. Also, some special classes of models can be specified according to some optimality criterions like the number of interactions. Finally, according to the resulting predictions biologists could devise appropriate experiments. If a property is predicted (true for all models), checking it experimentally means verifying its validity. If not, it means going ahead in the analysis by expelling, according to the result of the experiment, a subset of the models.
They concern mainly three topics:
- Mathematical modeling of biological networks. We focus mainly on discrete modeling of genetic networks. From Thomas’s network formalism as a starting point, we want to extend it to take into account i) the notion of time (delay) and of spatiality, 2) the composition of models and of their properties. On the prediction side, in addition to check that already identified properties are theorems, it is essential, for obtaining new knowledge, to provide a framework for designing high level languages expressing biological meaningful properties. The design of these languages needs the help of biologists as they must be relevant to future experiments. Their framework could be based for example on propositional logic or on CTL (Computational Tree Language). Coll. with G. Bernot et al. (U. de Nice).
- Computer implementation. Consistency checking or consistency repairing could be very much time consuming. We study how to overcome this difficulty with new approaches like ASP (Answer Set Programming), a very efficient logical programming technology based on a non monotonous logic which allows to specify defaults. Also, we intend to study CHR (Constraint Handling Rules, for implementing specific constraints solver), global constraints (specific to certain classes of problems) or continuous/stochastic constraints (for continuous variables). Coll. with Y. Hamadi et al. (Microsoft, Cambridge), Jacques Cohen (Brandeis U., Boston), Jacques Nicolas et al. (IRISA, Rennes), T. Schaub et al. (Postdam U.).
- Biological applications. We have collaborations with biologists on different systems like the carbon starvation response in E. coli (Delphine Ropers et al., INRIA, Grenoble), the Drosophila Gap-gene system (Denis Thieffry et al., ENS, Paris) or the mammalian iron regulation system (Jean-Marc Moulis, CEA, Grenoble).