Epitope-specific antibody design by a deep learning generative model

Abstract

The growing need for antibodies with customized specificity provides a rich environment for engineering efforts. In recent years, despite having streamlined experimental pipelines, the fundamental math requiring extensive libraries and screen campaigns to get an initial binding signal remains unchanged. A major advancement would be to directly design in silico an epitope-specific binder from scratch, providing a signal for potential optimization by artificial evolution. We have observed several key advantages in neural network approaches over existing methods. By leveraging the unique properties of neural networks, we developed a generative model for immunoglobulin 3D structures, with which diverse structures can be modeled with unprecedented speed. We extended it to a purely deep learning-based protein-protein interface design pipeline that optimize not only spatial orientations but fully-flexible protein structures on the fly to desired epitopes. This novel strategy explores neural network’s capabilities in modeling dynamic structures, and preliminary experimental results on multiple targets support the plausibility of in silico design of epitope-specific antibodies.