III: Medium: Collaborative Research: Guiding Exploration of Protein Structure Spaces with Deep Learning

George Mason University, Fairfax VA

Investigators

Abstract

Decades of scientific enquiry beyond molecular biology have demonstrated just how fundamental form is to function. All chemical reactions in the living cell involve molecules bumping and sticking to one another, and molecular form or structure is a central determinant of complementarity and strength of molecular interactions. In particular, by assuming specific structures, proteins are able to regulate diverse processes that maintain and replicate the living cell. With structure determination in laboratories lagging desperately behind the rapidly-growing number of protein-encoding gene sequences by high-throughput sequencing technologies, computational approaches to the problem of protein structure prediction now have a central role in molecular biology. This project advances algorithmic research to address the current impasse in form-function related problems in molecular biology. In particular, the project develops advanced optimization methods to explore the vast protein structure space and leverages information-integration techniques under the deep learning framework to effectively guide the exploration towards biologically-active structures. This project will benefit researchers of diverse sub-communities in computational and biological sciences, result in open source, publicly-available software packages, and provide excellent training and mentoring opportunities for under-represented students at the interface of computational science and computational biology. This project advances algorithmic research in information integration and informatics to address the current impasse in structure-function related problems in computational structural biology. The main focus is on the de-novo protein structure prediction problem, which is central to inferring biological activities of a rapidly-growing number of protein-encoding gene sequences. The proposed research generalizes the problem of exploring and obtaining a comprehensive view of a protein's structure space as that of computing a diverse ensemble of constraint-satisfying structures and then leveraging information-integration techniques to guide the exploration to regions of the structure space relevant for biological activity. The research proposes hybrid stochastic optimization algorithms for comprehensive exploration of protein structure spaces, deep convolutional neural networks for better assessment of structure nativeness, and combines the two in an information-integration algorithmic framework to guide the exploration of a structure space towards native structures. By doing so, the proposed research investigates a direction complementary to physics-based treatments, proposing to supplant such treatments with machine-learned models of nativeness. The research will benefit researchers in machine learning, stochastic optimization, and information integration with application-driven interests in molecular modeling, protein structure prediction, and modeling of complex, dynamic systems. The research will be disseminated via various venues, including an open-source software package, and will provide training opportunities for under-represented students of all levels at the interface of optimization, deep learning, and computational biology. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.