Insulin therapy is vital for millions of diabetes patients but faces ongoing challenges such as high production costs, limited stability, and strict storage requirements. Our team recently discovered a minimized insulin scaffold from a venomous animal that is smaller, more stable, and highly soluble compared to human insulin. Despite its promising structure, this natural peptide exhibits minimal activity on the human insulin receptor, limiting its therapeutic potential.
To enhance receptor binding and activation, I developed an iterative AI-driven pipeline that generates and optimizes insulin variants. Beginning with ProteinMPNN for variant sequence generation, the pipeline predicts 3D structures using large language models like ESMFold and other machine learning tools such as AlphaFold3, followed by receptor docking with ZDOCK. Correct receptor engagement is verified through PyMOL and FoldSeek, while the PyRosetta DDG protocol evaluates binding affinity. Finally, molecular dynamics simulations with GROMACS assess complex stability, ensuring robust interactions. This process iterates thousands of times to refine promising candidates.
For experimental validation, designed variants are expressed via yeast display as a high-throughput method to screen for receptor binding. Successful candidates from yeast display re-enter the computational pipeline for further optimization, establishing a design–test feedback loop. Variants demonstrating strong receptor affinity proceed to chemical synthesis and comprehensive in vitro and in vivo testing, paving the way toward cost-effective, stable, and bioavailable insulin therapeutics.