Predicting Drug Mode of Action Using Drug-Drug Similarity Measures

Abstract

During the past two decades, the rate of drug compounds released for public use has rapidly decreased, owing to the unexpected side e ects and lack of e cacy displayed by drugs during clinical trials. Consequently, the traditional \one drug-one target" paradigm has been replaced by the concept of \polypharmacology", which considers all the possible drug-target bindings that contribute to the overall drug mode of action (MoA). Computational techniques focusing on predicting unknown drug MoA are given prominence recently, owing to the contributions made by such predictions in decreasing the drug attrition rates. Most of the recent computational models developed for the purpose of large-scale drug action prediction, rely on the hypothesis that similar drugs tend to share similar e ects, where the similarity between drug pairs play a major role in the process. Although several models have been proposed with diverse measures of similarity, each metric bearing its own advantage, all the models su er from missing data when restricted by a speci c similarity measure. In our study, we address the problem of missing data with the use of several similarity measures, while exploiting their individual advantage. For this purpose, we propose the Friends of Friends approach for similarity approximation by electing a mediator drug to indirectly nd the similarity between a given drug pair. As a pilot study, we used the pair-wise chemical similarity to approximate the missing similarity distances between drug-induced gene expression pro les. Taking a step further, we built a drug- drug network from approximated similarity distances to obtain communities, in which, drugs that share therapeutic actions are grouped together using a clustering algorithm. Although the approximations produced an accuracy of 37%, we observed a precision rate of 95% for the drug communities we have obtained by replacing the top 20% of the pair-wise similarity scores, which proves that our approximations work best for drug pairs that display highly similar therapeutic actions.