Computational Analyses On Some Association Problems In Network Medicine

ABSTRACT:After the completion of sequencing whole genome for human beings and many other species, we came to the post-genome era. We expect in the era to disclose the function of genes and proteins, the products of genes, to resolve some medical problems which are related to our health and diseases.Research indicates that in organisms, cellular components exert their functions through interactions with other cellular components. The whole interactions represent an interactome, and the potential complexity of this network is huge and complicated. The disease phenotype is reflected by various pathobiological processes that interact in a complex network. Network theory based approaches can deepen our understanding about cellular interactome, disease-related genes, disease-related pathway, and so on. Consequently, diseases will be more effectively cured by better produced drugs. Network medicine which applies network theory to addressing medicine-related problems emerges and it not only offers a supporting platform to systematically explore the molecular complexity of a disease, but also helps to identify new disease-related genes and drug targets.Network medicine covers a wide range of research area. We expect to broaden our understanding about the mechanism of diseases by addressing three association problems in network medicine. The main research in this dissertation includes:(1) Compared to a smaller number of experimentally verified drug-target interactions, there are many more unlabelled interactions. We therefore proposed a semi-supervised algorithm NetCBP, which integrates network consistency and several biological features including chemical and genomic similarity measures, to predict potential interactions between drugs and their targets. Analyses reveal that our algorithm has a lower time complexity. Experiments discover that prediction accuracy of our algorithm is also improved. Excellent prediction results are received in a case study. Moreover, a comprehensive drug-target interaction prediction is conducted using our algorithm. Some strongly predicted interactions have been confirmed by existing drug databases.(2) Increasing evidences show that microRNA-related mutation and expression malfunction often lead to various diseases; however, little research has been conducted to disclose the associations between microRNAs and OMIM diseases. Most diseases curated in OMIM database contain no information about associated microRNAs. We therefore apply random walk to predicting microRNA-OMIM disease associations. OMIM disease similarity network and microRNA-disease association network are first constructed. Random walk is then applied on the disease similarity network to prioritize candidate diseases. Experimental results suggest that good prediction accuracy can be received. We finally make a comprehensive prediction and some strongly predicted associations have been confirmed, which indicates the usefulness of our method.(3) Only disease similarity information is used in random walk to predict microRNA-disease associations. To improve prediction accuracy, we introduced to integrate microRNA function similarity information, disease similarity information and known microRNA-disease associations for association prediction. Both similarity information is widely analyzed. Experimental results indicate that prediction accuracy can be improved by integrating both similarity information. The best-preformed algorithm NetCBI is chosen for microRNA-disease association prediction. Some of the strongly predicted associations have been verified by existing databases.(4) Single nucleotide polymorphisms (SNPs) play a vital role in explaining the difference of disease susceptibility and resistibility among populations and people. Linkage disequilibrium (LD) exists among SNPs, therefore there is no need to assay all SNPs. Only some tagging SNPs (tagSNPs) are necessary for covering all the SNPs. We comprehensively analyzed the advantages and disadvantages of the existing LD tagSNPs selection algorithms in a single population. Meanwhile a refined algorithm MinTag is presented. Experiments suggest that MinTag can select the fewest number of tagSNPs and receive the highest reduction ratio. Experimental results also indicate that precinct partitioning can dramatically reduce runtime.Three association problems in network medicine have been studied in this disseration and effective algorithms have been proposed to address the three problems. Our research will provide guide for future biological experiments. Moreover, the presented algorithms may also be suitable to resolve similar association problems in network medicine.