Statistical Learning for biological network inference and biological data-mining

Unraveling biological networks from experimental data and prior knowledge can be expressed in the framework of statistical learning. We consider two kinds of networks : gene regulatory networks which are considered as dynamical systems, and protein-protein network which are not a biological object per se but a convenient representation of a set of possible physical relations between proteins. Targetting network inference in both cases, we present two approaches we have developed for a few years: modeling approaches and predictive approaches. Modeling approaches are based on a dynamical model of the behaviour of gene regulatory networks as dynamical systems. We have chosen the probabilistic framework of state-space models or hidden Markov models to represent the evolution of gene regulatory networks and we show how frequentist and Bayesian approches provide very good estimation of dynamics parameters. Results are provided on experimental data ( on microorganisms) as well as toy data such as DREAM data Predictive approaches only tackle link prediction (or edge prediction) without considering any model of the behaviour of the network. We show how these approaches, implemented in the novel context of output kernel regression, can provide accurate prédiction of edges, even when very few proteins are labeled. New results are presented on yeast data. Finally, we define a research project consistent with some of the activities of the lab. This project is built on two main directions : data-mining of structured biological data and (again) modeling and inferring large networks.