How to Use RFpeptides Online for Macrocyclic Peptide Design

Designing a molecule that binds a specific protein surface is hard. Designing one that does it with nanomolar affinity, matches a crystal structure to within 1.5 Å, and starts from a diffusion model with no prior knowledge of the target binding mode is something that until recently could not be done computationally at all.

RFpeptides is a denoising diffusion pipeline introduced by Rettie et al. [1] that does exactly that, specifically for macrocyclic peptides. Tested against four diverse proteins with 20 or fewer designs per target, it produced medium to high-affinity binders against all four. The best anti-GABARAP macrocycle bound at Kd = 6 nM with a sub-nanomolar IC50. X-ray crystal structures for three of the four complexes matched the design models closely, with Cα RMSD below 1.5 Å [1].

On Neurosnap, the RFpeptides workflow runs through RFdiffusion v1 with the Cyclic Chains option enabled. This post walks through the full job setup, what each metric in the output means, and how to interpret results against what the published experimental validation actually showed.

RFdiffusion v1 service page on Neurosnap, the entry point for RFpeptides-style cyclic binder design.

Why GABARAP, and Why Macrocycles

GABARAP belongs to the ATG8 protein family, a set of ubiquitin-like proteins that are central to selective autophagy. ATG8 proteins recruit autophagy receptors and adaptors to the forming autophagosome through a short linear sequence motif: the LC3-interacting region, or LIR. The LIR consensus ([W/F/Y]-X-X-[I/L/V]) docks into two hydrophobic pockets on the ATG8 surface (HP1 and HP2), with the surrounding residues determining whether binding is selective for the LC3 subfamily or the GABARAP subfamily [2][3].

This LIR docking site (LDS) is the binding surface that RFpeptides targets. It is a well-defined, solvent-exposed groove on a relatively flat protein surface, which makes it challenging for small molecules but suitable for a cyclic peptide with enough contact area to engage both hydrophobic pockets simultaneously.

Macrocyclization matters here for two reasons. First, linear peptides in this size range are conformationally flexible and easy for proteases to cleave, particularly since the exposed termini are accessible. Head-to-tail cyclization removes the free termini, reducing proteolytic susceptibility and restricting the conformational ensemble toward the binding-competent geometry [4][5]. Second, conformational preorganization reduces the entropic cost of binding: a peptide that is already shaped like its bound conformation loses less entropy upon target engagement, which translates directly into tighter Kd [6]. The best macrocyclic drugs on the market (cyclosporine, vancomycin, octreotide) exploit exactly this principle.

What the RFpeptides Paper Demonstrated

Before running a job, it is worth knowing what the published benchmark actually established, because it sets the baseline for what the Neurosnap workflow is capable of reproducing.

Rettie et al. [1] designed macrocyclic binders against MCL1, GABARAP, MDM2, and RbtA. The GABARAP campaign produced the strongest binder (Kd = 6 nM, sub-nM IC50) in a batch of fewer than 20 tested designs. Crucially, for RbtA, a target with no experimentally determined structure in the PDB at the time, the pipeline produced a Kd < 10 nM binder starting from a predicted structure alone [1]. This is a strong demonstration that the method is not simply interpolating from known binding modes.

The X-ray crystal structure of the GABARAP-macrocycle complex (PDB: 9HGD [8]) confirmed that the computational design model and the actual binding mode are nearly identical. The macrocycle occupies the LIR docking site as designed, with the expected hydrophobic contacts at HP1 and HP2. This validates both the backbone generation and the downstream sequence design and scoring steps.

The Three-Step Workflow

The Neurosnap RFdiffusion v1 pipeline runs three models in sequence when Backbone Only is disabled:

RFdiffusion generates the backbone geometry. Given the target structure and a binder length range, the diffusion model proposes a peptide backbone de novo, conditioned on hotspot contacts if provided. The cyclic peptide extension in RFpeptides adds a cyclization constraint so that the generated backbone closes into a ring. The original RFdiffusion paper demonstrated that this denoising diffusion approach achieves outstanding performance across a wide range of design challenges, including protein binder design, by fine-tuning the RoseTTAFold network on structure denoising tasks [7].

ProteinMPNN then takes each generated backbone and designs amino acid sequences predicted to fold into that geometry. It does not use the target information directly at this stage; it optimizes for compatibility with the backbone. The original ProteinMPNN paper showed 52.4% native sequence recovery on protein backbones, far above the 32.9% achieved by Rosetta, and demonstrated broad utility across monomers, oligomers, and target-binding proteins including cyclic homo-oligomers [9]. Neurosnap generates 200 ProteinMPNN sequences per job and selects the five best-scoring ones.

AlphaFold2 then refolds each top sequence from scratch and evaluates whether the designed sequence actually adopts the designed binding pose. This step is the filter. Bennett et al. [10] showed that using AF2 or RoseTTAFold to assess whether a designed sequence folds into the intended monomer structure and binds as designed increases experimental success rates nearly ten-fold compared to energy-based design alone. The per-rank pLDDT, PAE, and pTM scores in the Neurosnap results table all come from this AF2 refolding step.

Running the Job on Neurosnap

Open RFdiffusion on Neurosnap and configure the job as follows for a GABARAP binder design run targeting the LIR docking site. Upload the target PDB with chain IDs assigned, using PDB 9HGD [8] with chain A as the target.

The Cyclic Chains = B field is the core setting. Without it, RFdiffusion generates a linear peptide binder. Setting it to B invokes the RFpeptides cyclization extension, which constrains the backbone sampling to produce closed ring topologies.

The completed public job used for this post:

https://neurosnap.ai/job/6a1f6e84dad598a779a8daf7?share=6a1f75cadad598a779a8db2b

Reading the Results

The Results Table

Completed job output. The results table shows five ranked designs sorted by ProteinMPNN score. Each row gives the backbone RMSD, AF2 mean pLDDT, AF2 mean PAE, AF2 iPAE, and AF2 pTM.

The top five designs produced these metrics:

The five designs are nearly identical in all metrics. This is expected: it means the diffusion model converged on a single well-defined backbone geometry that ProteinMPNN can satisfy with a small range of similar sequences. It does mean this run produced one structural solution rather than five diverse ones. For a real campaign, running multiple jobs and comparing diversity across runs is important.

The Structure Viewer

Interactive 3D structure viewer. The colored ribbon (AlphaFold2 pLDDT coloring) shows the target GABARAP (larger domain) and the generated cyclic peptide binder (shorter chain, lower right). High pLDDT regions are blue, lower-confidence regions shade toward orange.

The viewer confirms that the generated peptide sits close to the GABARAP surface. The most important visual check is that the binder chain (chain B) is physically docked against the target and not floating freely. The AlphaFold2 pLDDT coloring on the structure gives an immediate read on confidence. Most of the GABARAP body should be blue (very high confidence, pLDDT > 90), with the binder chain showing a mixed pattern that we unpack in the next section.

The published 9HGD crystal structure [8] shows the experimentally validated macrocycle binding at the GABARAP LIR docking site, engaging HP1 and HP2. The design model produced here is a new candidate for the same surface, generated without any prior knowledge of the 9HGD binding mode , which is what makes the comparison meaningful as a sanity check.

Understanding the pLDDT Plot

Per-residue pLDDT for all five ranked designs. The target GABARAP residues (positions 1–117) show consistently high confidence (80–96). The cyclic peptide binder residues (118+) show high confidence through most of the chain, then sharply low confidence at the termini.

The pLDDT plot has a clear structure worth understanding in detail.

Positions 1–117 (GABARAP target): The pLDDT is consistently high, in the 80–96 range across most of the domain. GABARAP is a well-characterized globular protein with many structural homologs in AlphaFold2's training set. The slight dips around positions 30–60 reflect loop regions with genuine flexibility, which is consistent with the known GABARAP structure.

Positions 118+ (cyclic peptide binder): The binder residues show high pLDDT through most of the peptide , in the 80–90 range, then a sharp drop to 30–50 at the very end. This drop is expected behavior for cyclic peptides validated with standard AlphaFold2, and it is not a quality signal to worry about.

The reason: AlphaFold2 was not trained on cyclized peptides. It has no positional encoding that encodes a covalent bond between the N- and C-terminus. From AF2's perspective, the termini of the binder chain are free chain ends, and the model correctly recognizes that free termini of short peptides are conformationally uncertain. Rettie et al. addressed this limitation in a companion paper by introducing AfCycDesign, which adds a cyclic offset to the AF2 positional encoding [11]. Using that approach, 36 of 49 native cyclic peptides were predicted at high confidence (pLDDT > 0.85) with RMSD below 1.5 Å. In the standard Neurosnap RFdiffusion pipeline, AF2 runs without the cyclic offset, so terminus confidence is structurally meaningless. What matters is the confidence of the binder residues that are not at the termini, and the interface between binder and target.

The practical threshold: per-residue pLDDT > 70 is the standard cutoff for reliable local geometry [12]. In these designs, the majority of binder residues are well above that. The terminal drop should be noted but not used to downrank designs.

Understanding the ProteinMPNN Score Plot

ProteinMPNN score distribution across all 500 generated sequences (sorted ascending). Scores from 1.14 to ~1.25 show a long, shallow plateau; the curve steepens sharply past rank ~450.

The ProteinMPNN score is a negative log-likelihood: lower means the model assigns higher probability to that sequence given the backbone. The score plot for this job has two distinct regions.

The plateau (~ranks 1–450, scores 1.14–1.25): The curve rises slowly, almost flat. A large fraction of generated sequences are considered roughly equally compatible with the backbone by ProteinMPNN. This is a good sign: it means the backbone geometry is not overly constrained, and ProteinMPNN has many valid sequence solutions to choose from. Designs from this plateau are structurally diverse at the sequence level while being functionally equivalent from a backbone-compatibility standpoint.

The steep rise (~ranks 450–500, scores 1.25–1.38): The curve steepens sharply, indicating that the remaining sequences required the model to make sequence choices it considered unlikely for this backbone. These are lower quality and would not be passed to AF2 for validation.

The top five designs come from the lowest end of the plateau at MPNN scores 1.14–1.15. Their near-identical scores reflect convergence on a single backbone solution. For a production design campaign, one strategy is to sample more diverse backbones across multiple RFdiffusion runs and then compare ProteinMPNN score distributions between jobs to identify which backbone geometries admit the widest range of sequence solutions.

pTM and Interface PAE

The AF2 pTM scores here range from 0.81 to 0.83. pTM (predicted TM-score) estimates how well the predicted structure matches a hypothetical true structure at the full-complex level. Scores above 0.5 indicate a broadly correct fold; scores in the 0.8+ range for a protein-peptide complex indicate that AF2 has high confidence in the overall complex geometry.

For binder design, what matters most is not pTM but the interface PAE, the predicted aligned error for residue pairs that span the binder-target interface. A low interface PAE indicates AF2 is confident about the relative orientation of binder and target. The AF2 iPAE values here (5.36–5.99 Å) are moderate. They indicate that AF2 predicts the complex but with some uncertainty in the exact interface geometry. This is typical for short peptide binders and is not disqualifying. It reflects both the genuine flexibility of a small cyclic peptide and the fact that standard AF2 does not know the peptide is cyclic.

Comparing to the Experimental Result

The RFpeptides paper [1] reported that the best GABARAP macrocycle (GAB_D23) binds with Kd = 6 nM and that the crystal structure (9HGD) matches the design model with Cα RMSD < 1.5 Å. The macrocycle contacts the LIR docking site of GABARAP, with hydrophobic residues engaging HP1 and HP2 as designed.

The Neurosnap job described here generates new candidate macrocycles for the same surface. The workflow is identical: RFdiffusion backbone generation with cyclic constraint, ProteinMPNN sequence design, AF2 validation. What determines whether these new designs achieve similar affinity is primarily the quality of the interface contacts: the number and geometry of hydrogen bonds, the buried surface area, the hydrophobic complementarity at HP1 and HP2, and the degree to which the designed backbone actually matches the bound conformation in solution.

The computational metrics from the job (pLDDT, PAE, ProteinMPNN scores) filter for designs where AF2 is confident in the complex geometry, but they do not directly predict binding affinity. The 6 nM published binder was confirmed experimentally. For any new design from this workflow, experimental validation is the final step. The computational pipeline narrows a pool of thousands of possible sequences to five high-confidence candidates.

Practical Tips for Production Runs

Set hotspot residues when you know the binding site. GABARAP's LIR docking site involves residues in the HP1 and HP2 grooves. If you know the target epitope, providing hotspot residues in the ChainResidue format (e.g., A33,A50-64) directs RFdiffusion toward that region rather than exploring the full surface. Without hotspot guidance, the model may converge on a different surface with weaker contacts.

Run multiple jobs before filtering. A single job produces one backbone solution, often with five nearly identical designs as seen here. Running 5–10 jobs gives a set of structurally diverse backbone solutions to compare. The best campaign strategy is to generate many backbones first, then filter on ProteinMPNN score and AF2 metrics across the full set.

Do not over-interpret the terminal pLDDT drop. For cyclic peptide designs validated with standard AF2, the pLDDT drop at the peptide termini is an artifact of the model's lack of cyclic topology encoding. Score designs on the pLDDT of the non-terminal binder residues and on the interface PAE, not on the full mean pLDDT across the whole chain.

Short binder lengths are a practical starting point. The validated GABARAP binder in the published RFpeptides work was a 10–16 residue macrocycle. Starting with a binder length range of 10–20 covers this space. Shorter rings have fewer rotatable bonds, stronger conformational preorganization, and are easier to synthesize.

Compare ProteinMPNN score distributions across runs. A flat, slowly rising score plateau indicates a backbone that admits many good sequences. A steep, rapidly rising curve indicates a backbone with high sequence specificity, which may reflect a tight design but also increases the risk that the chosen sequence is near a local energy minimum that does not transfer to solution.

Ready to Try RFdiffusion?

Run your own RFpeptides-style macrocyclic binder design on RFdiffusion on Neurosnap. Upload a target structure, set Cyclic Chains to B, and review the complete RFdiffusion, ProteinMPNN, and AlphaFold2 pipeline output in one place.

References

1. Rettie, S., Juergens, D., Adebomi, V., et al. Accurate de novo design of high-affinity protein-binding macrocycles using deep learning. Nature Chemical Biology (2025). https://doi.org/10.1038/s41589-025-01830-2
2. Johansen, T. & Lamark, T. Selective autophagy: ATG8 family proteins, LIR motifs and cargo receptors. Journal of Molecular Biology 432, 80–103 (2020). https://doi.org/10.1016/j.jmb.2019.07.016
3. Wirth, M., Zhang, W., Bertoglio, F., et al. Molecular determinants regulating selective binding of autophagy adapters and receptors to ATG8 proteins. Nature Communications 10, 2055 (2019). https://doi.org/10.1038/s41467-019-10059-6
4. Vinogradov, A. A., Yin, Y. & Suga, H. Macrocyclic peptides as drug candidates: recent progress and remaining challenges. Journal of the American Chemical Society 141, 4167–4181 (2019). https://doi.org/10.1021/jacs.8b13178
5. Bechtler, C. & Lamers, C. Macrocyclization strategies for cyclic peptides and peptidomimetics. RSC Medicinal Chemistry 12, 1325–1351 (2021). https://doi.org/10.1039/D1MD00083G
6. Rowland, C. E. et al. Rational design of cyclic peptides, with an emphasis on bicyclic peptides. Current Opinion in Structural Biology 91, 102971 (2025). https://doi.org/10.1016/j.sbi.2025.103025
7. Watson, J. L., Juergens, D., Bennett, N. R., et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089–1100 (2023). https://doi.org/10.1038/s41586-023-06415-8
8. PDB 9HGD — GABARAP–macrocycle complex (GAB_D23). https://www.rcsb.org/structure/9HGD
9. Dauparas, J., Anishchenko, I., Bennett, N., et al. Robust deep learning–based protein sequence design using ProteinMPNN. Science 378, 49–56 (2022). https://doi.org/10.1126/science.add2187
10. Bennett, N., Coventry, B., Goreshnik, I., et al. Improving de novo protein binder design with deep learning. Nature Communications 14, 2625 (2023). https://doi.org/10.1101/2022.06.15.495993
11. Rettie, S., Juergens, D., Anand, R., et al. Cyclic peptide structure prediction and design using AlphaFold2. Nature Communications 16, 2465 (2025). https://doi.org/10.1038/s41467-025-59940-7
12. Agarwal, V. & Bhatt, D. L. The power and pitfalls of AlphaFold2 for structure prediction beyond rigid globular proteins. Nature Chemical Biology 20, 950–959 (2024). https://doi.org/10.1038/s41589-024-01638-w
13. RFdiffusion on Neurosnap: https://neurosnap.ai/service/RFdiffusion
14. Public RFdiffusion example job: https://neurosnap.ai/job/6a1f6e84dad598a779a8daf7?share=6a1f75cadad598a779a8db2b