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

Realizing the therapeutic potential of antibodies requires simultaneously optimizing multiple properties, such as antigen-binding specificity, viscosity, clearance and immunogenicity. However, existing methods used for this task are time consuming and resource intensive, often struggling to balance these properties. Here we propose DualGPT-AB, a dual-stage conditional generative pre-trained transformer (GPT) framework for therapeutic antibody design. DualGPT-AB leverages a conditional GPT to model sequence-property relationships by representing the desired properties as learnable embeddings, while incorporating a reinforcement learning strategy to promote antibody sequence exploration and improve optimization efficiency. Computational experiments show that DualGPT-AB is capable of generating antibody heavy chain complementarity-determining region 3 (CDRH3) sequences fulfilling multiple desired properties. Notably, 8 out of 100 randomly selected antibodies from our designed candidate library exhibit excellent HER2-binding affinities. Wet-laboratory validation confirms that DualGPT-AB identifies antibodies with enhanced tumoricidal activity compared with Herceptin, a widely used antibody drug for treating HER2-positive cancers. Overall, DualGPT-AB is a promising approach for advancing artificial intelligence-driven therapeutic antibody development.