Positive unlabeled learning for deriving protein interaction networks

Binary classification is the process of labeling the members of a given data set on the basis of whether they have some property or not. To train a binary classifier, normally one needs two sets of examples from each group, usually named as positive and negative examples. However, in some domains, negative examples are either hard to obtain or even not available at all. In these problems, data consist of positive and unlabeled examples. An example to this kind of problems is derivation of protein-protein networks.Biological processes in a living organism depend on proteins and mostly interactions of proteins. It is important to determine which proteins interact to understand how an organism survives. While it is possible to derive by experiments that two proteins interact with each other, it is much harder to conclude that they do not. Even if we do not observe the interaction of two proteins during an experiment, they may interact in a different time or condition.In this thesis we first present a survey of algorithms which can handle such problems, and then provide a comparison of some of these algorithms on the protein-protein interaction derivation problem by using the available (positive) interaction information. Thus we identify which algorithms can be used or have potential to be used for deriving protein-protein interaction networks.