« All posts

RFpeptides: De Novo Design of Protein-Binding Macrocycles via Deep Learning

RFpeptides extends RoseTTAFold2 and RFdiffusion to enable deep learning-based de novo design of protein-binding macrocyclic peptides for drug discovery.

Researchers have extended the RoseTTAFold2 and RFdiffusion architectures with cyclic relative positional encoding to create RFpeptides, a new computational pipeline for de novo design of protein-binding macrocyclic peptides.

Macrocyclic peptides sit between small-molecule drugs and biologics, offering potential for membrane permeability while binding challenging protein targets. Traditional approaches—natural product discovery and high-throughput screening—face limitations including synthetic difficulty, poor stability, and inability to simultaneously optimize multiple biophysical properties. Earlier physics-based design methods achieved only modest binding affinities, with experimental structures often failing to match design models.

By adapting RFdiffusion's proven protein-design success to macrocycles, the team implemented cyclic positional encoding that enables robust generation of diverse macrocyclic backbones. Testing showed thousands of structurally unique 10- and 12-residue macrocycle backbones emerging from 48,000 generated structures, with phi-psi angle distributions closely resembling natural protein Ramachandran plots—indicating the designs don't require extensive d-amino acid stabilization.

For engineers, the significance lies in compatibility: RFpeptides can leverage RFdiffusion's existing conditional generation features like epitope-specific targeting and motif scaffolding, and is designed to transfer directly to future networks such as RFdiffusion All-Atom, opening paths toward incorporating nucleic acids and other non-peptidic molecules into design calculations.

This synthesis was produced from its source by AI; there is no human editor or manual review step. How we work