Identifying drug-target and drug-disease associations using computational intelligence
Background: Traditional drug development is an expensive process that typically requires the investment of a large number of resources in terms of finances, equipment, and time. However, sometimes these efforts do not result in a pharmaceutical product in the market. To overcome the limitations of this process, complementary—or in some cases, alternative—methods with high-throughput results are necessary. Computational drug discovery is a shortcut that can reduce the difficulties of traditional methods because of its flexible nature. Drug repositioning, which aims to find new applications for existing drugs, is one of the promising approaches in computational drug discovery. Considering the availability of different types of data in various public databases, drug-disease association identification and drug repositioning can be performed based on the interaction of drugs and biomolecules. Moreover, drug repositioning mainly focuses on the similarity of drugs and the similarity of agents interacting with drugs. It is assumed that if drug D is associated or interacts with target T, then drugs similar to drug D can be associated or interact with target T or targets similar to target T. Therefore, similarity-based approaches are widely used for drug repositioning. Research Objectives: Develop novel computational methods for drug-target and drug-disease association prediction to be used for drug repositioning. Results: In this thesis, the problem of drug-disease association identification and drug repositioning is divided into sub-problems. These sub-problems include drug-target interaction prediction and using targets as intermediaries for drug-disease association identification. Addressing these subproblems results in the development of three new computational models for drug-target interaction and drug-disease association prediction: MDIPA, NMTF-DTI, and NTD-DR. MDIPA is a nonnegative matrix factorization-based method to predict interaction scores of drug-microRNA pairs, where the interaction scores can effectively be used for drug repositioning. This method uses the functional similarity of microRNAs and structural similarity of drugs to make predictions. To include more biomolecules (e.g., proteins) in the study as well as achieve a more flexible model, we develop NMTF-DTI. This nonnegative matrix tri- factorization method uses multiple types of similarities for drugs and proteins to predict the associations between drugs and targets and their interaction score. To take another step towards drug repositioning, we identify the associations between drugs and disease. In this step, we develop NTD-DR, a nonnegative tensor decomposition approach where multiple similarities for drugs, targets, and diseases are used to identify the associations between drugs and diseases to be used for drug repositioning. The detail of each method is discussed in Chapters 3, 4, 5, respectively. Future work will focus on considering additional biomolecules as the drug target to identify drug-disease associations for drug repositioning. In summary, using nonnegative matrix factorization, nonnegative matrix tri-factorization, and nonnegative tensor decomposition, as well as applying different types of association information and multiple types of similarities, improve the performance of proposed methods over those methods that use single association or similarity information.
Drug repositioning, drug-target interaction, matrix factorization, tensor decomposition.
Doctor of Philosophy (Ph.D.)