RFPeptides: De novo design of protein-binding macrocycles using deep learning (www.nature.com)

A new deep learning pipeline called RFpeptides has been developed for the de novo design of protein-binding macrocycles, which are potential therapeutics that combine the advantages of small molecules and large biologics. Traditional methods for designing these binders typically involve extensive screening and lack control over binding modes, making RFpeptides significant by offering a more streamlined and customizable approach. By incorporating a denoising diffusion model within the RoseTTAFold2 framework, RFpeptides successfully generated diverse, high-affinity macrocyclic binders against various proteins, including a binder for the target Rhombotarget A that achieved a Kd of less than 10 nM. This advancement allows researchers to explore a vast structural diversity of macrocycles with improved binding affinity and stability, crucial for therapeutic applications. The successful integration of cyclic positional encoding into the design process enhances the reliability of the generated peptide structures, evidenced by close correlations between computational models and experimental X-ray structures. RFpeptides not only represents a leap forward in peptide design methodologies but also has the potential to accelerate the development of macrocyclic peptide therapeutics for various diagnostic and treatment applications in the biomedical field.