Improving Protein Interactions Prediction Using Machine Learning and Visual Analytics

By Mudita Singhal, Ph.D. Washington State University December 2007 Chair: John H. Miller The response of biological systems to external stimuli is ruled by their cellular interaction networks. This makes the problem of inferring cellular interaction networks essential to decipher the basic operational principles of biological systems. Knowing which proteins exist in a certain organism or cell type and how these proteins interact with each other are necessary for the understanding of biological processes at the whole cell level. The determination of the protein-protein interaction (PPI) networks has been the subject of extensive research and it has been shown that domain-domain interactions (DDIs) are good indicators of possible protein interactions, and can more accurately predict protein interactions than comparing full-length protein sequences. Despite the development of reasonably successful methods there is definite scope for improvement. This thesis is aimed at developing machine learning based computational techniques that utilize domain information in the proteins to predict PPI networks. This research aims to make four major contributions to the field of PPIs. The first two are the development of two new PPI prediction algorithms, DomainGA and DomainSVM. DomainGA is a genetic algorithm based multi-parameter optimization method which quantifies DDIs and uses them to predict PPI. The second method, DomainSVM utilizes v the DDI scores obtained from DomainGA in a Support Vector Machine (SVM) based learning system to improve PPIs prediction by overcoming the limitations of DomainGA. These two methods can be used as a two-step filtering process to validate experimentally detected PPI. The third contribution is score assignment to DDIs which is proven to be discriminatory between positive and negative PPI. Finally the fourth contribution is a visual analytic environment called CABIN (Collective Analysis of Biological Interaction Networks) which provides a one-of-its-kind tool to analyze, compare and integrate multiple predicted networks obtained from public data sources and/or inference algorithms such as DomainGA and DomainSVM. The predicted interactions accompanied by a confidence score and an exploratory visualization environment shall help researchers validate experimental observations and/or make an informed decision while generating hypothesis and models for designing new experiments.