Introducing Deep-Learning–Designed Macrocycles: High-Affinity Protein-Binders

Members of the IPD: Gaurav Bhardwaj, Frank DiMaio, and 2024 Nobel laureate David Baker have unveiled a deep-learning framework capable of the de novo design of macrocyclic peptides that bind proteins with nanomolar affinity and remarkable precision. This work is based on a report from June 20, 2025, in Nature Chemical Biology
Accurate de novo design of high-affinity protein-binding macrocycles using deep learning and has been added to RFdiffusion. 

What are protein-binding macrocycles, and why do they matter?

Macrocycles are ring-shaped peptides known for combining the stability of small molecules with the specificity of biologics. Their constrained structure boosts binding affinity, cell permeability, and in vivo resilience, making them ideal for therapeutic targets that are difficult to modulate with traditional antibodies or small molecules.

Designing macrocyclic peptides that bind with high affinity to protein targets has traditionally relied on resource-intensive experimental screening. Traditional computational methods struggled to generate these molecules de novo: directly tailoring both their cyclic backbone and side chains for tight binding at atomic-level precision is a computationally expensive process.

New Approach: RFpeptides

The authors present RFpeptides, a computational pipeline powered by denoising diffusion models. This approach directly designs macrocyclic peptides by generating cyclic backbone structures that precisely fit the shape of a given protein binding pocket.

In the same step, RFpeptides also optimizes the orientation of amino acid side chains to strengthen the interaction between the peptide and the target protein. Unlike traditional methods that rely on large-scale screening, this model produces a small, targeted set of high-potential binders using only computation.

Key Results

The researchers tested RFpeptides on four diverse protein targets, each relevant to therapeutic or biological research. These include:

• MCL1: An anti-apoptotic protein involved in cancer pathways.
• MDM2: Another cancer-related protein that regulates the tumor suppressor p53.
• GABARAP: A protein important for autophagy and intracellular trafficking.
• RbtA: A bacterial protein used to demonstrate the method’s generalizability across species.

For each target, only about 20 designed macrocycles were synthesized and tested, which is extraordinarily small compared to the billions or trillions of randomly tested peptides in traditional screening workflows. This small number highlights the models’ precision in generating high-quality candidates.

Among the results, macrocycles targeting MCL1 and MDM2 showed binding affinities in the 1 to 10 micromolar range, which is moderate strength for peptide binders. Notably, some of the macrocycles designed to bind GABARAP and RbtA achieved sub-10 nanomolar dissociation constants (K<sub>D<\sub>), along with sub-nanomolar potency in inhibition assays. These results indicate high binding strength and functional activity.

Mechanism & Validation

RFpeptides works by first generating a cyclic peptide backbone that fits into a chosen pocket on the protein’s surface. It then places the side chains in a way that enhances the binding interface.

To confirm the designs, the team used high-resolution structural methods such as X-ray crystallography and cryo-electron microscopy. The results of these experiments showed that the actual structures of the macrocycle–protein complexes closely matched the computational predictions, validating the method’s accuracy down to the atomic level.