Fig. 1: GearBind-based in silico antibody affinity maturation pipeline.
From: Pretrainable geometric graph neural network for antibody affinity maturation
a Pipeline Overview: The pipeline features the geometric neural encoder, GearBind, which undergoes self-supervised pretraining on CATH and supervised learning on SKEMPI v2. The GearBind-based ensemble model is employed to perform in silico affinity maturation on a target antibody, given its bound structure to the native antigen. Guided by the model predictions, antibodies with improved binding affinity can be found after testing one or two dozen mutant candidates. NN neural network, Ab Antibody, CDR Complementarity-determining region. Designed using resources from Flaticon.com. b GearBind Model: GearBind employs a shared graph neural network to encode both the wild-type and mutant complex structures. For each structure, a relational interface graph is constructed. A geometric graph neural network, GearNet, then performs multi-relational and multi-level message passing on the graph to extract rich interface representations. The mutational effect ΔΔGbind is predicted by an antisymmetric predictor given the GearNet-extracted representations of the two complexes. c Self-supervised Pretraining: GearBind+P leverages mass-scale unlabeled protein structures via self-supervised, contrastive pretraining. The model is trained to contrast between native structures and randomly mutated structures with side-chain torsion angles sampled from a rotamer library. Pretraining helps GearBind+P explore the energy landscape of native protein structures and results in improved performance in ΔΔGbind prediction.
