Learning biological interactions from multiple data sources

Inferring gene and protein interactions allow us to understand the mechanism of biological processes and the events that lead to diseases. This research focuses on designing joint learning algorithms and frameworks for studying biological interactions, including gene regulatory networks, protein-protein interactions, gene-protein interactions, and gene-drug interactions for hit selection.;In the gene regulatory network study, a modified Inductive Causation (mIC) algorithm is presented. It combines steady state data with partial prior knowledge of topological ordering for joint learning the causal relationship among genes. In real data analysis, the mIC algorithm identified the important causal relations associated with WNT5A, a gene playing an important role in melanoma.;In the protein-protein interaction study, a Bayesian modeling framework is applied for learning high order logic relationships among proteins from multiple data sources. The method is able to identify several known disease-related interactions such as Mad Cow Disease-related prions interactions.;In the protein-DNA interaction study, a kernel-based learning method is presented for predicting DNA-binding proteins. The method applies a kernel function on each type of data source and combines the kernels for classification using Support Vector Machines (SVMs). The results demonstrate that the kernel method in combination with multiple data sources has great potential in accurately predicting DNA-binding proteins.;Large-scale RNAi screening is a powerful tool for discovering functional genomics and identifying drug targets. In the gene-drug interaction study, a computational analysis is presented for identifying key factors in assay optimization. A new method to account for the systemic effects from the siRNA transfection step is developed. Furthermore, hit selection processes for studying gene-drug interactions are performed for identifying drug targets.;The above four topics are centered around the key issue of revealing biological interactions. The work discusses the problems from different levels of interaction networks and performs disease-related analysis of gene, protein and drug interactions. The contributions are mainly the designing and improving of the computational algorithms and methods for joint learning interactions from various kinds of biological data. These methods can be used as efficient tools for understanding gene and protein interaction relationships and identifying potential drug targets.