Antibody-SGM: Antigen-Specific Joint Design of Antibody Sequence and Structure using Diffusion Models

Abstract:
This study builds upon the promising diffusion models for protein backbone generation, addressing their limitation in guiding the generation with sequence-specific attributes and functional properties. To overcome this, we present AntibodySGM, a novel joint structure-sequence diffusion model that enables the joint generation of protein sequences and structures. Our model starts from random sequences and structural features, and iteratively denoises to generate valid pairs of sequences and structures, resulting in full-atom native-like antibodies. Antibody-SGM demonstrates its versatility by designing full-atom antibodies, antigen-specific CDR design, antibody optimization, validation with Alphafold2, and key antibody sequence and structural features. By allowing simultaneous optimization of both sequence and structure, Antibody-SGM opens new possibilities for designing functional proteins with precise sequence and structural attributes, providing a pathway for protein function optimization through active inpainting learning. These advancements showcase the potential of our approach in protein engineering and expand the capabilities of protein design models.

Suggested citation:
Xie, Xuezhi et al. “Antibody-SGM: Antigen-Specific Joint Design of Antibody Sequence and Structure using Diffusion Models.” ICML - Computational Biology (2023)</i>.