DualGPT-AB: a dual-stage generative optimization framework for therapeutic antibody design

Area of Science:

• Biotechnology
• Computational Biology
• Immunology

Background:

• Optimizing therapeutic antibody properties like specificity and immunogenicity is crucial but challenging with current methods.
• Existing techniques are time-consuming and struggle to balance multiple antibody characteristics effectively.

Purpose of the Study:

• To introduce DualGPT-AB, a novel dual-stage conditional generative pre-trained transformer (GPT) framework for efficient therapeutic antibody design.
• To leverage artificial intelligence for simultaneous optimization of multiple antibody properties, enhancing development efficiency.

Main Methods:

• Utilized a conditional GPT to model antibody sequence-property relationships using learnable embeddings for desired properties.
• Incorporated a reinforcement learning strategy to improve antibody sequence exploration and optimization efficiency.
• Generated and computationally validated antibody heavy chain complementarity-determining region 3 (CDRH3) sequences.

Main Results:

• DualGPT-AB successfully generated antibody CDRH3 sequences with multiple desired properties.
• 8% of designed antibodies showed excellent HER2-binding affinities.
• Wet-laboratory validation confirmed enhanced tumoricidal activity compared to Herceptin in HER2-positive cancer models.

Conclusions:

• DualGPT-AB is a powerful AI-driven approach for therapeutic antibody development.
• The framework demonstrates potential for accelerating the design of antibodies with improved efficacy and safety profiles.
• This method offers a promising solution for overcoming limitations in traditional antibody engineering.