Computational Models For Domain-Peptide Mediated Protein-Protein Interactions

Peptide recognition module (PRM) mediated protein-protein interactions (PPIs) are critical for better understanding the relationship between genomes and networks. High-throughput experimental screens, such as phage display, can be used to identify their binding motifs for use in computationally predicting high confidence PPIs with binding site information. Computational approaches for predicting protein interactions are either limited by their inability to predict peptide recognition module mediated interactions or do not consider many known constraints governing these interactions. A novel ensemble method for predicting in vivo SH3 domain-peptide mediated PPIs in S. cerevisae and H. sapiens using phage display data is presented. As with similar methods, this method uses position weight matrix models of protein linear motif preference in combination with a range of evidence sources related to binding site and cellular constraints on protein interactions. The novelty of this approach is the large number of evidence sources used and the method of combination of peptide based and protein pair based evidence sources. A novel semi-supervised training framework is used to train peptide and protein Gaussian naïve Bayes models using both labeled and unlabeled datasets. Research into the different evidence sources led to the development of state-of-the-art algorithms for predicting PPIs using semantic similarity in the Gene Ontology (GO), gene or protein expression, and network topology. A novel method to compute semantic similarity between GO terms annotated to proteins in interaction datasets which considers the unequal depths of the ontology is developed (TCSS). Most PPI prediction methods rely on observations from a single gene expression profile (GEP) to predict novel interactions. Correlation coefficients from multiple GEPs are combined into a single model to improve PPI prediction. Protein expression is proposed as a new evidence source for predicting PPIs. A machine learning model is developed for predicting high confidence PPIs using graph descriptors, such as edge density, transitivity, and mutual clustering coefficient of known interaction networks (NTOP). All the algorithms developed during this research are open source and freely available for community use.Ph.D