Computational antibody structure prediction and antibody-antigen docking

Antibodies are proteins that are the recognition elements of the immune system and increasingly used as drugs. Antibodies bind tightly and specifically to antigens to block their activity or to mark them for destruction. Three-dimensional structures of the antibody-antigen complexes are useful for understanding their mechanism and for designing improved antibody drugs. Experimental determination of structures is laborious and not always possible, so I have developed computational tools to predict the structures of antibody and I have used them for predicting antibody-antigen interaction complexes. In this thesis, I describe the development and implementation of RosettaAntibody, a new protocol for high-resolution homology modeling of antibody variable domains (FV). RosettaAntibody is the first FV homology modeling protocol that optimizes the relative orientation of the FV light and heavy chains. I have also developed a homology modeling protocol specific for camelid heavy chain antibody variable regions (VHH) which are unique in their absence of a light chain. VHHs have higher stability and solubility, and are easier to manufacture compared to classical antibodies yet exhibit similar diversity in antigen recognition with high binding affinities, making them ideal candidates for alternate scaffolds in the development of therapeutic monoclonal antibodies. The camelid VHHs have long complementarity determining region (CDR) loops which are extremely difficult to model, but I have discovered structural features in VHHs that have improved loop modeling for these long loops. To enable the scientific community to benefit from my research I have created the RosettaAntibody server where a user can provide the heavy and light chain FASTA amino acid sequences of a classical antibody to obtain ten homology models for the query sequence. Computer-predicted models of antibodies, or homology models, typically have errors which can frustrate algorithms for prediction of protein-protein interfaces (docking), and result in incorrect prediction of structures of the bound complexes. To compensate for such errors, I have developed SnugDock, a new docking algorithm which incorporates flexibility to overcome structural errors in the antibody structural model. The algorithm allows both intramolecular and interfacial flexibility in the antibody during docking, resulting in improved accuracy approaching that when using experimentally determined antibody structures. I have also successfully tested my technique in CAPRI (Critical Assessment of PRediction of Interactions), the world-wide blind protein docking challenge. Finally I have generalized my protocol and used it to model the ternary complex of a K-48 linked di-ubiquitin with an Ube2g2 enzyme using nuclear magnetic resonance (NMR) experimental observations to bias my simulations. My protocol has converged to a low-energy structure that explains most of the NMR experimental observations. The homology modeling and docking approaches that I have developed will be useful for new therapeutic antibodies for which experimentally determined coordinates are not available, enabling structural analysis required to make rational choices for better antibody drug designs.