Antibody-SGM, a Score-Based Generative Model for Antibody Heavy-Chain Design

Area of Science:

• Computational biology
• Protein engineering
• Artificial intelligence in drug discovery

Background:

• Traditional antibody design relies on random mutagenesis and energy assessments.
• Recent diffusion models excel at generation but often neglect full structural or sequence details.
• Existing models require additional steps to predict missing structural or sequence components.

Purpose of the Study:

• To introduce Antibody-SGM, a joint structure-sequence diffusion model for comprehensive antibody design.
• To address limitations of current models by integrating sequence-specific attributes and functional properties.
• To generate native-like, full-atom antibody heavy chains with valid sequence-structure pairs.

Main Methods:

• Development of Antibody-SGM, a joint structure-sequence diffusion model.
• Refinement of generation process to ensure valid sequence-structure pairings.
• Integration of sequence-specific attributes and functional properties into the generative process.
• Application of active inpainting for simultaneous sequence and structure optimization.

Main Results:

• Successful generation of full-atom, native-like antibody heavy chains.
• Demonstrated versatility in applications: full-atom antibody design, antigen-specific CDR design, and heavy chain optimization.
• Validation of generated antibodies using AlphaFold3.
• Identification of critical antibody sequences and structural features.
• Optimization of protein function through active inpainting learning.

Conclusions:

• Antibody-SGM represents a significant advancement in protein design, offering a versatile and powerful tool.
• The model successfully integrates sequence and structure for enhanced antibody generation and optimization.
• This approach holds promise for revolutionizing protein engineering and antibody design.