Abstract
We introduce AbDiffuser, an equivariant and physics-informed diffusion modelfor the joint generation of antibody 3D structures and sequences. AbDiffuser isbuilt on top of a new representation of protein structure, relies on a novelarchitecture for aligned proteins, and utilizes strong diffusion priors toimprove the denoising process. Our approach improves protein diffusion bytaking advantage of domain knowledge and physics-based constraints; handlessequence-length changes; and reduces memory complexity by an order ofmagnitude, enabling backbone and side chain generation. We validate AbDiffuserin silico and in vitro. Numerical experiments showcase the ability ofAbDiffuser to generate antibodies that closely track the sequence andstructural properties of a reference set. Laboratory experiments confirm thatall 16 HER2 antibodies discovered were expressed at high levels and that 57.1%of the selected designs were tight binders.