Home  > News from Ibis > Talks, seminars, poster presentations,...

Workshop on Identification and Control of Biological Interaction Networks - Presentation abstracts


Global optimization approaches for the identification and control of biological systems

Julio R. Banga, IIM-CSIC, Vigo (Spain)

Abstract : Mathematical optimization aims to make a system or design as effective or functional as possible, computing the quality of the different alternatives using a mathematical model. This talk is focused on applications of mathematical optimization in computational systems biology, highlighting the need of using global optimization methods in order to obtain proper solutions. Examples will be given where suitable optimization methods are used for topics ranging from model building and optimal experimental design to optimal manipulation (control) in metabolic engineering and synthetic biology. Finally, several perspectives for future research will be outlined.

Back to the workshop main page

Quantitative temporal logic for systems biology

Alexandre Donzé, Verimag, Grenoble (France)

Abstract : The Signal Temporal Logic (STL) is adapted to describe properties which constrain real-valued signals such as timed evolutions of quantities involved in a biological system. Recently, we extended STL with a quantitative (robust) interpretation which provides a numerical margin by which a simulation trace satisfies or violate a property. Moreover, we can estimate in some cases the sensitivity of this margin to a parameter change. By combining this information with different parameters exploration strategies, we get an efficient methodology to investigate which properties are satisfied by a model, how robustly these properties are satisfied and how to find parameters values which guarantees a robust satisfaction. I will describe this methodology, the tool implementing it, and an illustration of its application on an enzymatic network involved in angiogenesis.

Back to the workshop main page

Models for cancer data : from simple to complex and multiscale formalism

Benjamin Ribba, INRIA (EPI NUMED), ENS Lyon (France)

Abstract : The rate of success in the development of new drugs in oncology is particularly low. Even for “old” and widely used chemotherapeutic compounds, controversial views persist on the best efficient way to deliver them to patients. As a possible tool to optimise the development and the use of those drugs, quantitative modelling and simulation initiatives are more and more recommended by regulatory agencies. Often, in clinical settings, available data are sparse and in consequence, models must be simple. A standard analysis method used in the field of pharmaceutical research is mixed-effect nonlinear regression where model parameters are considered as random variables to account for the different sources of variability within the analyzed population. More recently, the development of new drugs - targeting specific molecular processes involved in tumour growth - has generated complementary data such as microscopic evaluation of biomarkers activity. This offers the possibility to shift from simple models to more complex ones, expected to integrate the main biological mechanisms occurring in cancer, including complex regulatory networks. Obviously, this shift comes with open methodological issues regarding parameter estimation and model validation.

In my talk, I will present recent results we obtained in modelling cancer data and will emphasize the methodological issues rising from the development of complex and multiscale models.

Back to the workshop main page

Towards real-time control of gene expression : in silico analysis

Gregory Batt, INRIA Contraintes, Rocquencourt (France)

Abstract : We consider here the problem of the real-time control of the expression of a single gene at the cellular level in yeast. We provide a description of the biological problem and its mathematical formulation. We develop a model predictive control strategy tailored to the specificities of the biological problem and assess in silico its effectiveness and its robustness to biological variability.

Back to the workshop main page

Statistical relational learning for network inference

Florence d’Alché-Buc, CNRS IBISC, Université d’Evry-Val d’Essonne (France). In collaboration with Céline Brouard, Christel Vrain et Marie-Anne Debily.

Abstract : Biological network inference from experimental data has received much attention from the machine learning community during last years. In terms of machine learning, two main approaches now prevail : modeling approaches and predictive approaches.

In modeling approaches, the object of interest is the behavior of the dynamical system underlying the network and the goal is to identify the system by estimating parameters of a given model. Both experimental data and prior knowledge can be introduced into the estimation process either in a Bayesian framework or in a regularization framework.

In predictive approaches, the focus is on the relations between components (the edges of the network). Given the assumption that part of the network is known, a classifier able to predict if one gene regulates another one, is built from data. While less developed in bioinformatics that in other fields such as social networks, such approaches can provide powerful results when the amount of prior knowkedge is sufficient.

In this talk, I will mainly present a new contribution in the field of predictive approaches, using a new framework for statistical relational learning, called Markov Logic Network introduced by Domingos et al. A Markov Logic Network encodes a set of weighted first order logic rules and encapsulates it into a probabilistic framework that allows to use frequentist or Bayesian estimation methods. Here, the goal is to learn logical rules that conclude on the existence of a regulation. The available experimental data (gene expression) and different properties of genes such as GO labels, positions of genes on chromosome into a knowledge database have first been encoded into first order logic. The known graph, extracted from the text-mining software Ingenuity, was encoded using the predicate "regulates". A classic Inductive Logic Program was first used to acquire a set of candidates rules that conclude on this predicate. The, penalized conditional likelihood maximization was used to estimate the weights of each rule. The approach was successfully applied to a gene regulatory network involved in the switch proliferation / differentiation of keratinocytes.

Finally at the end of the talk, I will suggest how supervised predictive and modeling approaches can be linked.

Back to the workshop main page

Identification of genetic network dynamics : a model invalidation approach

E.Cinquemani, IBIS, INRIA Grenoble - Rhône-Alpes, Montbonnot (France)

Abstract : We address the problem of reconstructing the structure and the parameters of ODE models of genetic network dynamics from quantitative time-course gene expression data. We consider a family of ODE models whose structure resembles unate functions, a class of Boolean functions that were argued to capture the majority of the gene activation rules. We investigate the properties of this class of ODE models and exploit them to develop a computationally effective approach to identification. This is based on the idea of eliminating modelling hypotheses inconsistent with the data prior to searching for the best fitting model, thus restricting parameter optimization to the family of model structures compatible with the data. Performance of the resulting algorithm is tested on a simulated example as well as on challenging experimental data from IRMA, a synthetic network engineered in yeast cells that has been proposed in the literature as a benchmark for network inference methods.

Back to the workshop main page