Discovering Domain-Domain Interactions toward Genome-Wide Protein Interaction and Function Predictions

Abstract

To fully understand the underlying mechanisms of living cells, it is essential to delineate the intricate interactions between the cell proteins at a genome scale. Insights into the protein functions will enrich our understanding in human diseases and contribute to future drug developments. My dissertation focuses on the development and optimization of machine learning algorithms to study protein-protein interactions and protein function annotations through discovery of domain-domain interactions. First of all, I developed a novel domain-based random decision forest framework (RDFF) that explored all possible domain module pairs in mediating protein interactions. RDFF achieved higher sensitivity (79.78%) and specificity (64.38%) in interaction predictions of S. cerevisiae proteins compared to the popular Maximum Likelihood Estimation (MLE) approach. RDFF can also infer interactions for both single-domain pairs and domain module pairs. Secondly, I proposed cross-species interacting domain patterns (CSIDOP) approach that not only increased fidelity of existing functional annotations, but also proposed novel annotations for unknown proteins. CSIDOP accurately determined functions for 95.42% of proteins in H. sapiens using 2,972 GO `molecular function' terms. In contrast, most existing methods can only achieve accuracies of 50% to 75% using much smaller number of categories. Additionally, we were able to assign novel annotations to 181 unknown H. sapiens proteins. Finally, I implemented a web-based system, called PINFUN, which enables users to make online protein-protein interaction and protein function predictions based on a large-scale collection of known and putative domain interactions.