III: Small: High-Throughput Annotation of Cellular Functions of Intrinsic Disorder in Proteins

Virginia Commonwealth University, Richmond VA

Investigators

Abstract

One of fundamental problems in molecular biology is to decipher functions of millions of uncharacterized protein sequences that are rapidly generated by high-throughput genome sequencing. The sequence-to-structure-to-function paradigm was used for decades to determine functions of proteins. However, recent research has broadened this paradigm by adding new players, proteins with intrinsic disorder (ID). They are highly abundant and cannot be solved with the structure-driven approach. While there are many widely used computational methods that accurately predict ID in protein sequences, methods for the prediction of the many functions of ID are lacking. This project will develop a family of novel, accurate, and high-throughput computational methods that predict all major functions of ID in protein sequences. It will produce putative functional annotations on an unprecedented scale of thousands of species, addressing the problem of high rate acquisition of raw sequence data and contributing to the increase of the rate of scientific discovery. These results will advance our understanding of fundamental biological processes and human health given the high prevalence of ID in human diseases and attractiveness of proteins with ID as drug targets. This project will also contribute to training of STEM students and researchers via short workshops and undergraduate and graduate level lectures and mentoring, focusing on demonstrating relevance and value of education and research in the emerging areas of protein bioinformatics. This project will conceptualize, design, rigorously test and deploy predictors of all major functions of intrinsic disorder including protein-RNA, protein-DNA, protein-protein, protein-ligand and protein-lipid interaction, flexible linkers and spacers, regions that host post-translational modifications, and moonlighting regions. These methods will provide fast and accurate predictions in the absence of sequence similarity when the commonly used sequence alignment-based approaches fail. The design will include a novel, hybrid function-specific feature extraction, empirical feature selection, and optimization of predictive models generated with modern machine learning algorithms. The inputs to these predictive models will be quantified and aggregated from a comprehensive set of empirically selected and sequence-derived structural and biophysical characteristics of proteins. The resulting methods will be benchmarked and made freely available to the broad research community via a web-based, server-side portal. The results will be also deposited into relevant public databases.