Antibodies play a vital role in aiding the immune system to identify and subdue harmful foreign antigens by binding to specific portions of antigens (i.e. epitopes) with high specificity and affinity. These properties have led to them becoming a $100 billion+ global market as therapeutics and extensive work has been done to develop experimental and
in-silico methods to design desired antibody-antigen interactions. Despite the high specificity exhibited by most naturally-occurring antibodies, a concern of experimentally and computationally developed antibodies is off-target binding. Predicting whether or not an antibody will bind to a given protein is an example of the protein-protein docking problem, which is a long term topic of interest in the field of computational protein science. Examples of docking methods include Hex, Z-dock, and RosettaDock, among others. Z-dock uses Fourier transforms to perform a step-wise, three-dimensional, grid-based, blind search over all six degrees of freedom to identify poses with high shape complementarity. Hex modifies the Z-dock approach by using spherical harmonic functions and searching over five degrees of freedom. RosettaDock follows an iterative search procedure starting with the random placement of low resolution structures followed by a low resolution Monte-Carlo search and high resolution refinement. While successful, these methods follow a trial and error based approach that attempts to search the entire solution space iteratively. Such an approach is time consuming, especially for larger molecules, and can have precision-related issues depending on step size and number of random poses explored. Currently, it is not feasible to predict antibody – protein binding interactions for many proteins in a reasonable period of time. Hence there is a need for faster and more precise approaches to simulate antibody-protein docking.
Previously, we demonstrated that ~5 antibody residues contribute 70-80% of the total binding energy upon antigen binding, consistent with previous definitions of “hotspot†residues. Based on this finding, we developed a de novo binding protein design tool, AUBIE, which identified strong binding interactions around binding loops as part of its procedure. In this work, we have extended the AUBIE methodology and developed a computational algorithm for predicting antibody - protein binding interactions. The approach uses pairwise distance matching of strong binding interaction locations around protein surfaces to rapidly identify antibody – protein complexes. We will present on the details of the algorithm’s development and preliminary results, demonstrating the feasibility of scanning thousands of proteins for off-target antibody binding sites. We believe that this work can be a useful tool for screening possible therapeutic antibodies.
