Identification of biological networks such as signalling pathways, gene regulatory networks, protein-protein interaction networks and metabolic networks is considered as a key challenge in computational biology. Using machine learning framework, this problem can be addressed using different points of view, depending of course on the nature of the biological interactions to be inferred but also on the level of abstraction of the chosen modeling and the amount of prior knowledge available. Since 2000, research in statistical learning of biological networks have given rise to a rich panel of approaches whose interest overcomes the field of computational biology. Network identification has been tackled using large scale data-mining approaches, supervised predictive approaches and reverse-modeling approaches. In this sole last family, it is very instructive to focus on the numerous graphical models that have been proposed so far such as Graphical Gaussian Models, Bayesian networks, Dynamical Bayesian networks and state-space models. I will present a short review of these methods discussing among other issues model complexity, relevance to biology, ability to deal with hidden variables and scalability. I will also plead for the construction of a benchmark repository devoted to examples of relevant test problems even if the true relevant test has always to be made in vivo or in vitro.