Iterative Refinement Graph Neural Network for Antibody Docking and Design

Tuesday April 19th, 4-5pm EST | Wengong Jin, Broad Institute & Massachusetts Institute of Technology

Antibodies are versatile proteins that bind to pathogens like viruses and stimulate the adaptive immune system. The antibody binding affinity is determined by complementarity-determining regions (CDRs) which closely interact with antigen residues (epitopes). In this talk, I will present RefineGNN, a new generative model for automatic design of antibody CDRs. Different from standard inverse folding methods, we seek to co-design a CDR sequence and 3D structure rather than assuming the 3D structure is given. Specifically, RefineGNN unravels a sequence autoregressively while iteratively refining its predicted 3D structure. To further model antibody-antigen interaction, I will extend RefineGNN for antibody CDR docking. Given a CDR sequence and an epitope 3D structure, the model predicts the CDR backbone and side chain structure around the epitope. RefineGNN represents an antibody-antigen complex as a dynamic graph and updates it in a hierarchical, equivariant manner. On a standard antibody design benchmark, RefineGNN outperforms existing baselines in terms of both contact prediction and sequence recovery.

Preprint: https://arxiv.org/abs/2110.04624

Recording: https://youtu.be/px5iC79jtfc