Analysis of network topological features for identifying potential drug targets

Identifying potential drug targets is a crucial task for drug discovery. Traditional in silico approaches utilize only pro-tein sequence or structural information to predict whether a protein can be a drug target, and achieve limited suc-cess. Since proteins function in the context of interaction networks by interacting with other cellular macromolecules, analysis of topological features of proteins in such networks can reveal important insights on whether a protein can be a potential drug target. In this paper, we introduced ten new topological features extracted from human protein in-teraction networks. When designing these new features, we specially emphasized the roles of three disease-related groups of proteins: known drug targets, disease genes, and essen-tial genes. Based on the topological feature set, we built supervised learning models using support vector machines, L1-regularized logistic regression, and k-nearest neighbors to predict whether testing proteins can be drug targets or not. We also analyzed the relevance of each feature to the probability of proteins being drug targets. We achieved up to 80 % classification accuracy using tenfold cross validation, and yielded very stable results with a large number of ran-dom samplings. Our method can also be used to prioritize multiple candidate proteins according to their probability of being drug targets