RFpeptides: Designing Protein-Binding Macrocycles

CD ComputaBio provides software-based computational services to support research and development. We do not offer free software packages.

Overview

Why Binding Mode Matters Alongside Affinity

Display-based screening explores very large peptide collections and can identify strong binders without first specifying a detailed binding geometry. Structure-guided design offers a complementary route: select a target region, generate compatible molecular shapes, and test candidates whose interfaces are explicitly modeled. This control could be useful when a particular interaction surface or functional site must be addressed.

Macrocycles are difficult design objects because ring closure constrains backbone geometry and small sequence changes can alter conformation. A confident predicted complex does not guarantee that a peptide can be synthesized, folds as intended, or binds experimentally. RFpeptides therefore evaluates three distinct questions: can a cyclic backbone be generated, can a sequence support that backbone, and can the resulting molecule recognize its target as designed?

The paper focuses on macrocyclic binder discovery rather than oral delivery. Its affinity and structural results do not establish membrane permeability, intracellular target engagement, or a clinically suitable route of administration.

How RFpeptides Makes Protein Design Aware of a Ring

The key adaptation is cyclic relative positional encoding. In a linear chain, the sequence endpoints are far apart in positional terms. In a head-to-tail macrocycle, those residues are neighbors. Adjusting the positional representation makes the network account for the circular relationship while retaining conventional representations for the protein target and peptide–target relationships.

The authors first applied cyclic encoding to RoseTTAFold2 structure prediction and then to RFdiffusion backbone generation. This repurposes knowledge learned from protein structures without requiring a large, dedicated experimental macrocycle training collection. It is an architectural adaptation of established networks, rather than a new model trained entirely from cyclic peptide data.

The complete workflow has three principal stages: RFdiffusion generates target-associated cyclic backbones; ProteinMPNN assigns compatible amino-acid sequences; and AfCycDesign and/or cyclically adapted RoseTTAFold predict the resulting complexes. Iterative ProteinMPNN and Rosetta Relax steps generate sequence alternatives. Additional filters assess structural confidence, agreement with the generated backbone, interface packing, calculated interaction energy, and aggregation propensity.

Hotspot conditioning can direct generation toward specified target residues. It was used for GABARAP and RbtA, whereas the selected MCL1 candidates converged on the BH3 interaction site without specified hotspots. The method therefore supports both less constrained exploration and targeted interface design, depending on the campaign.

Four Targets Reveal Different Performance Levels

The experimental campaigns span two predominantly helical binding pockets, a mixed secondary-structure interaction surface, and a flatter bacterial protein surface. Their outcomes should be compared target by target. A nanomolar result on one target is not a general affinity guarantee for the pipeline.

These denominators distinguish computational nomination from successful experimental production. For MCL1, 13 of the 27 selected designs did not yield enough correctly cyclized product for characterization. For MDM2, three of eleven were lost at the cyclization step. GABARAP testing used six high-purity products from thirteen selected candidates. Such attrition is directly relevant to planning synthesis budgets and timelines.

The abstract's statement that twenty or fewer macrocycles were tested per target describes the final experimental test sets. It should not be interpreted as twenty or fewer initial computational designs or synthesis nominations. Each campaign began with thousands of generated structures.

MCL1 and MDM2: Useful Binders Without Nanomolar Affinity

For MCL1, the authors generated 9,965 cyclic backbones before sequence design and filtering. The best experimentally measured binder, MCB_D2, reached a Kd of 2 μM. Its complex structure was determined at 2.1 Å resolution. After aligning target chains, the peptide Cα RMSD between model and crystal structure was 0.7 Å; aligning the macrocycles alone gave 0.4 Å.

The crystal structure confirmed contacts from both the helical segment and the connecting loop. Its helix had the opposite N-to-C orientation to natural MCL1 binders discussed in the study, while additional loop interactions contributed to recognition. This is stronger evidence of designed geometry than affinity measurement alone: the molecule binds through an interface closely resembling the computational proposal.

The MDM2 campaign generated 10,000 backbones and four sequences per backbone, producing 40,000 design models. Eight synthesized peptides were tested, three showed screening binding signals, and MDB_D8 reached 1.9 μM affinity. The modeled interaction reproduces recognizable features of p53-site recognition, but the paper does not provide an equivalent MDM2 complex crystal structure. Its structural evidence should therefore remain distinct from that for MCL1.

GABARAP: Binding and Competition Are Separate Readouts

GABARAP offered a binding surface with both α-helical and β-strand elements. The authors generated 20,000 backbones and 80,000 sequence-designed models, filtered these to 335 candidates, clustered them into 80 structural groups, and selected thirteen diverse macrocycles. This diversity-based selection matters because repeatedly testing closely related top-scoring designs may miss alternative productive geometries.

GAB_D8 and GAB_D23 bound GABARAP with SPR-derived Kd values of 6 nM and 36 nM. In a separate AlphaScreen competition assay measuring disruption of GABARAP binding to the K1 peptide, their IC50 values were approximately 0.7 nM and 2.5 nM. Kd and IC50 describe different experiments and cannot be substituted for one another.

Structural validation has an important qualification. GAB_D8 was crystallized with the homolog GABARAPL1, not GABARAP, because suitable GABARAP complex crystals were unavailable. The homolog has identical residues within 5 Å of the modeled peptide, and one observed conformation matched the design closely. However, the crystal also showed another bound conformation with a register shift. GAB_D23 was crystallized directly with GABARAP and showed a larger local backbone difference, with a target-aligned peptide RMSD of 1.7 Å. The strongest examples do not imply uniform sub-1.5 Å agreement for every solved complex.

RbtA: Designing Against a Predicted Target Structure

RbtA, a cell-surface protein from Acinetobacter baumannii, tested whether an experimental target structure was essential. The authors used AlphaFold2 and RoseTTAFold2 predictions, concentrating on regions where both methods agreed and confidence was high. They selected a relatively flat region in the N-terminal domain and specified hotspot residues for macrocycle generation.

From 20,000 generated backbones and four designed sequences per backbone, twenty-six candidates were nominated. Eleven yielded sufficient material and purity for testing; four showed binding signals in the 100 μM screen. RBB_D10 was then characterized at 9.4 nM affinity. Crystallography subsequently supported both the target fold and the designed peptide interface, with a target-aligned peptide RMSD of 1.4 Å.

This result shows that a predicted structure can provide a productive starting point in a carefully evaluated case. It does not establish that all high-confidence target predictions are reliable design substrates, nor does binding a bacterial surface protein demonstrate bacterial killing or therapeutic efficacy.

What Development Teams Can Use—and Still Need to Validate

The study supports a workflow in which computational design narrows the experimental search while preserving a structural hypothesis. It also shows why ranking must be calibrated locally. Filter distributions differed across targets, and the authors adjusted thresholds accordingly. The best MCL1 and GABARAP binders were not among the top three designs by the cited confidence ranking. A single universal cutoff or highest-ranked candidate is therefore not a substitute for a diverse experimental panel.

Synthetic feasibility remains a major bottleneck. Failed or low-yielding cyclization removed substantial fractions of selected candidates. Integrating precursor-route assessment could be a useful future extension, but this paper does not demonstrate an integrated synthesis-prediction solution. Likewise, predicted folding into a binding-compatible conformation without the target is supportive computational evidence, not direct measurement of the solution ensemble.

Finally, the work establishes binding and selected biochemical competition results, not clinical readiness. Cellular activity, selectivity across related proteins, proteolytic stability, exposure, delivery, and safety require further testing. Noncanonical residues, alternative crosslinkers, and other cyclization chemistries are identified as future extensions. These boundaries allow the paper's strongest conclusion to stand: de novo macrocycle design can produce experimentally confirmed recognition geometries with small final testing panels.

Overview of What CD ComputaBio Can Provide

The research connects to computational tasks in cyclic peptide discovery and evaluation. The service areas below address related research needs; they do not imply reproduction of RFpeptides performance or ownership of the authors' platform.

References

1. Rettie SA, Juergens D, Adebomi V, et al. Accurate de novo design of high-affinity protein-binding macrocycles using deep learning. Nature Chemical Biology. 2025;21:1948–1956. Published online June 20, 2025. https://doi.org/10.1038/s41589-025-01929-w.

* For Research Use Only.

Related Services