Accurate de novo design of high-affinity protein binding macrocycles using deep learning

Area of Science:

• Computational biology
• Biochemistry
• Drug discovery

Background:

• Traditional methods for designing macrocyclic binders to therapeutic proteins are resource-intensive and offer limited control over binding modes.
• Existing physics-based and deep-learning methods lack robust approaches for *de novo* design of protein-binding macrocycles.

Purpose of the Study:

• To introduce RFpeptides, a novel denoising diffusion-based pipeline for the *de novo* design of macrocyclic peptide binders against protein targets.
• To demonstrate the efficacy of RFpeptides in generating high-affinity binders for diverse therapeutic targets.

Main Methods:

• Utilized a denoising diffusion model (RFpeptides) for *de novo* design of macrocyclic peptide binders.
• Tested designed macrocycles against four diverse protein targets (MCL1, MDM2, GABARAP, RbtA).
• Validated binding affinities (K_D, IC50) and determined complex structures via X-ray crystallography.

Main Results:

• RFpeptides successfully generated medium to high-affinity binders against all tested targets.
• Achieved K_D values between 1-10 μM for MCL1 and MDM2 binders.
• Obtained a 6 nM K_D and sub-nanomolar IC50 for an anti-GABARAP macrocycle, and <10 nM K_D for an anti-RbtA binder.
• X-ray structures confirmed high accuracy of computational design models (Ca RMSD < 1.5 Å for 3/4 structures).

Conclusions:

• RFpeptides provides a powerful and rapid framework for custom design of macrocyclic peptides.
• The *de novo* design approach with known binding modes facilitates downstream optimization for diagnostic and therapeutic applications.
• This method overcomes limitations of traditional screening approaches, accelerating the development of novel protein-binding agents.