Characterization of protein interactions

Abstract. Available information on molecular interactions between pro-teins is currently incomplete with regard to detail and comprehensive-ness. Although a number of repositories are already devoted to capture interaction data, only a small subset of the currently known interac-tions can be obtained that way. Besides further experiments, knowledge on interactions can only be complemented by applying text extraction methods to the literature. Currently, information to further character-ize individual interactions can not be provided by interaction extraction approaches and is virtually nonexistent in repositories. We present an approach to not only confirm extracted interactions but also to characterize interactions with regard to four attributes such as ac-tivation vs. inhibition and protein-protein vs. protein-gene interactions. Here, training corpora with positional annotation of interacting proteins are required. As suitable corpora are rare, we propose an extensible cura-tion protocol to conveniently characterize interactions by manual anno-tation of sentences so that machine learning approaches can be applied subsequently. We derived a training set by manually reading and an-notating 269 sentences for 1090 candidate interactions; 439 of these are valid interactions, predicted via support vector machines at a precision of 83 % and a recall of 87%. The prediction of interaction attributes from individual sentences on average yielded a precision of about 85 % and a recall of 73%.