Fig. 1.

Deep learning guided antibody optimization platform. (A) Overview of the pipeline. It demonstrates the computational/experimental feedback loop to refine antibody design. (B) Geometric deep learning model. The WT complex and the mutated complex structures are encoded using a shared geometric attention network. The effect of mutation measured by ΔΔG is then predicted by a network that compares features of the two complexes. (C) P36-5D2 antibody optimization. Given the complex structure, we first simulate different variants and then evaluate potential CDR mutations that will improve binding by predicted ΔΔG values. Mutants with top ΔΔG scores are examined in laboratory experiments, and those with neutralizing potency are combined for the next round of optimization. (D) Optimization improves neutralization ability against SARS-CoV-2 and Delta variant. (E) The log fold changes of IC₅₀ relative to the original antibody.