(656i) Context-Aware Representations from Deep Learning for Antibody Design

Protein-protein interactions drive most protein function including antigen recognition and neutralization, cell signaling, catalysis, etc. Designing proteins and antibodies to bind protein targets of interest such as the spike protein of the SARS-Cov2 or a protein receptor implicated in a disease is an important task in developing therapeutics and diagnostic tools. We formulate the protein interface design problem as the task of predicting the interfacial sequence of a protein binder given the sequence and structural context of the interface. For this task, we developed an equivariant graph neural network that learns residue-level representations at protein interfaces. Since the number of protein interfaces with solved structures is limited, we train our model to learn from all available structures and fine-tune on protein and antibody-antigen interfaces. The model generalizes residue representations from proteins to protein-protein and antibody-antigen interfaces learning context-aware representations. On a test set of 400 protein interfaces (non-redundant from training sets), our model recovers over 38% of the native interface sequence with an average maximum recovery of 66.7% per target. We further fine-tune the model on antibody-antigen interfaces. On a test set of over 70 antibody-antigen interfaces including single-domain antibodies (non-redundant from training sets), the model recovers 37.8% of the native interface with an average maximum recovery of 72.3% per target. Furthermore, for heavy chain hypervariable loops, H1, H2, H3, the model recovers 46.5%, 48.9% and 39.3% respectively. Our model presents a fast and context-aware approach to design proteins and antibodies in the context of their binding partners.