Predicting potential drugs and drug-drug interactions for drug repositioning
The purpose of drug repositioning is to predict novel treatments for existing drugs. It saves time and reduces cost in drug discovery, especially in preclinical procedures. In drug repositioning, the challenging objective is to identify reasonable drugs with strong evidence. Recently, benefiting from various types of data and computational strategies, many methods have been proposed to predict potential drugs. Signature-based methods use signatures to describe a specific disease condition and match it with drug-induced transcriptomic profiles. For a disease signature, a list of potential drugs is produced based on matching scores. In many studies, the top drugs on the list are identified as potential drugs and verified in various ways. However, there are a few limitations in existing methods: (1) For many diseases, especially cancers, the tissue samples are often heterogeneous and multiple subtypes are involved. It is challenging to identify a signature from such a group of profiles. (2) Genes are treated as independent elements in many methods, while they may associate with each other in the given condition. (3) The disease signatures cannot identify potential drugs for personalized treatments. In order to address those limitations, I propose three strategies in this dissertation. (1) I employ clustering methods to identify sub-signatures from the heterogeneous dataset, then use a weighting strategy to concatenate them together. (2) I utilize human protein complex (HPC) information to reflect the dependencies among genes and identify an HPC signature to describe a specific type of cancer. (3) I use an HPC strategy to identify signatures for drugs, then predict a list of potential drugs for each patient. Besides predicting potential drugs directly, more indications are essential to enhance my understanding in drug repositioning studies. The interactions between biological and biomedical entities, such as drug-drug interactions (DDIs) and drug-target interactions (DTIs), help study mechanisms behind the repurposed drugs. Machine learning (ML), especially deep learning (DL), are frontier methods in predicting those interactions. Network strategies, such as constructing a network from interactions and studying topological properties, are commonly used to combine with other methods to make predictions. However, the interactions may have different functions, and merging them in a single network may cause some biases. In order to solve it, I construct two networks for two types of DDIs and employ a graph convolutional network (GCN) model to concatenate them together. In this dissertation, the first chapter introduces background information, objectives of studies, and structure of the dissertation. After that, a comprehensive review is provided in Chapter 2. Biological databases, methods and applications in drug repositioning studies, and evaluation metrics are discussed. I summarize three application scenarios in Chapter 2. The first method proposed in Chapter 3 considers the issue of identifying a cancer gene signature and predicting potential drugs. The k-means clustering method is used to identify highly reliable gene signatures. The identified signature is used to match drug profiles and identify potential drugs for the given disease. The second method proposed in Chapter 4 uses human protein complex (HPC) information to identify a protein complex signature, instead of a gene signature. This strategy improves the prediction accuracy in the experiments of cancers. Chapter 5 introduces the signature-based method in personalized cancer medicine. The profiles of a given drug are used to identify a drug signature, under the HPC strategy. Each patient has a profile, which is matched with the drug signature. Each patient has a different list of potential drugs. Chapter 6 propose a graph convolutional network with multi-kernel to predict DDIs. This method constructs two DDI kernels and concatenates them in the GCN model. It achieves higher performance in predicting DDIs than three state-of-the-art methods. In summary, this dissertation has proposed several computational algorithms for drug repositioning. Experimental results have shown that the proposed methods can achieve very good performance.
Druy repositioning, Gene signature, Drug-drug interaction, Human protein complex signature, Machine learning, Deep learning, Graph convolutional network
Doctor of Philosophy (Ph.D.)