Predicting Types of Protein-Protein Interactions Using Various Multiple-Instance Learning Algorithms

Analysis of protein-protein interactions (PPI) is an important issue to understand the biological mechanism of a cellular process. Although large volumes of PPI data have been collected, only a few amounts of PPIs have been elucidated at the functional level. It is required to predict functional types of PPIs. A PPI described in existing pathways, as the KEGG pathways, often corresponds to a pair of complexes, each of which is composed of several subunits (proteins). On the other hand, functional annotations, as provided by the Gene Ontology, have been accumulated mainly for simple proteins. It is difficult for a usual supervised learning method to predict PPI types because the relationship between the input variables (annotations for subunits) and the target variable (PPI type) is ambiguous. With regard to this point, we assume that a subunit pair can determine the interaction type between complexes. Intuitively, this assumption means that an interaction between complexes can be reduced to an interaction between a subunit pair across those complexes. Based on this assumption, the PPI type prediction task can be formalized as a problem of the Multiple-Instance Learning (MIL), which is a kind of semi-supervised learning and has been applied to various fields including drug activity estimation: a complex pair with a PPI type is formulated as a labeled bag and a possible subunit pair across that complex pair as an instance. The goal is to predict labels of unseen bags based on labeled bags and the feature vectors of every instance in every bag. To solve that problem, we have already proposed a method called Voting Diverse Density (VDD)[2], which is a weighted voting system based on the Maron’s Diverse Density[1]. This paper compares the method with other MIL algorithms by solving a binary classification version of that problem.