ABI Innovation: DeepStruct: Learning representations of protein 3-d structures and their interfaces using deep architectures

Colorado State University, Fort Collins CO

Investigators

Abstract

Proteins perform many cellular functions, made possible by complex networks of interactions; knowing the location of the interaction sites on the proteins is key for understanding exactly how they work. Important applications include designing drugs and therapeutic agents. Experimental techniques for determining the interfaces between proteins are expensive and time consuming, so computational structural biologists seek to predict these mathematically. Current prediction methods use a limited number of features hand-crafted by an expert. An alternate approach is to learn the important features directly from all of the data, using a method called deep neural networks. This proposal explores a combined approach: use expert intuition for some features but add the power of unsupervised learning with deep neural networks to learn additional, novel features. The results will enrich the way protein structural features are understood in terms of their functional properties, whether those are catalytic sites, protein-binding sites or other sites important to the protein structure. Certainly in the prediction of protein structure itself machine learning scoring methods are showing great promise. Aspects of the research will be used in courses offered through a recently awarded NSF-NRT training grant, The training grant establishes an interdisciplinary program at the interfaces of biology, engineering, math/statistics and computer science. The program prepares students for a variety of career paths. Research and education experiences will provide students with valuable expertise in a computational area that is highly valued by top technology firms, such as Google and Facebook, which have research teams exploring the possibilities of deep neural networks. This work proposes a paradigm shift in the field of protein interface prediction and scoring: from hand-crafted features and standard off-the-shelf classifiers to an approach that augments existing features with automatic learning of the features that characterize the 3-d structures of proteins, combined with the use of learning algorithms that are specifically designed for the characteristics of the problem. The proposed approach has multiple novel aspects: the proposed learning approach leverages information contained in the entire protein data bank (PDB) to learn features that characterize protein structures at multiple scales and levels of abstraction. It introduces a novel neural network architecture and regularization terms that constrain the solution towards biologically relevant results. The primary alternative to this machine learning-based interface prediction uses docking simulations; however, current docking energy functions are not accurate enough, so that a near-native solution is often not ranked high enough on the list of outcomes to be useful. Extensions of the proposed architectures for interface prediction will be employed for re-scoring docking solutions to improve their predictive success. A workflow that integrates docking and machine learning-based interface prediction and scoring is proposed to explore the synergism between these tasks. Information on the progress made on the project is available through the project website: http://www.cs.colostate.edu/~asa/projects.html.