Towards an Integrated Protein-protein Interaction Map

As protein-protein interaction plays a major role in most cellular processes, identifying the full repertoire of interacting protein pairs in the cell is of great importance. This challenge has been addressed both experimentally and computationally. Large-scale experimental studies provide data sets of interacting proteins. These data sets, while only partial and noisy, allow us to characterize properties of interacting proteins and lead to the development of predictive algorithms. Most algorithms, however, ignore possible dependencies between the various interacting pairs, and predict them independently of one another. In this study, we present a computational approach that overcomes this drawback by predicting all interactions simultaneously. For this, we build an integrated probabilistic model that involves all attributes we wish to predict, as well as the observations we have on these attributes. We use the language of relational Markov random fields to build such an integrated probabilistic model, and to learn its properties efficiently. We then use this to predict all interactions simultaneously. We show that by modeling dependencies between interactions, we achieve a more accurate description of the protein interaction network, and gain new insights into the properties of the interacting pairs.