RFPeptides: De novo design of protein-binding macrocycles using deep learning

Researchers introduced RFpeptides, a denoising diffusion-based deep learning pipeline for de novo design of protein-binding macrocycles, reporting medium to high affinity binders against all four diverse protein targets tested with 20 or fewer designed macrocycles per target. For the target Rhombotarget A (RbtA), the team obtained a high-affinity binder with a dissociation constant (Kd) below 10 nM starting from the predicted target structure, and X-ray structures of macrocycle-bound myeloid cell leukemia 1, γ-aminobutyric acid type A receptor-associated protein and RbtA complexes matched the computational models with a Cα root-mean-square deviation under 1.5 Å. The work, published in Nature Chemical Biology, offers a framework for rapid, custom design of macrocyclic peptides for diagnostic and therapeutic applications.

Abstract Developing macrocyclic binders to therapeutic proteins typically relies on large-scale screening methods that are resource intensive and provide little control over binding mode. Despite progress in protein design, there are currently no robust approaches for de novo design of protein-binding macrocycles. Here we introduce RFpeptides, a denoising diffusion-based pipeline for designing macrocyclic binders against protein targets of interest. We tested 20 or fewer designed macrocycles against each of four diverse proteins and obtained binders with medium to high affinity against all targets. For one of the targets, Rhombotarget A RbtA , we designed a high-affinity binder K <sub d</sub < 10 nM despite starting from the predicted target structure. X-ray structures for macrocycle-bound myeloid cell leukemia 1, γ-aminobutyric acid type A receptor-associated protein and RbtA complexes match closely with the computational models, with a Cα root-mean-square deviation < 1.5 Å to the design models. RFpeptides provides a framework for rapid and custom design of macrocyclic peptides for diagnostic and therapeutic applications. Similar content being viewed by others Main Macrocyclic peptides present a promising avenue for developing new therapeutics that bridge the gap between small-molecule drugs and large biologics<sup 1 https://www.nature.com/articles/s41589-025-01929-w ref-CR1 , 2 https://www.nature.com/articles/s41589-025-01929-w ref-CR2 </sup . Biologics, while capable of binding diverse therapeutic targets with high affinity and selectivity, are usually unable to cross cell membranes because of their large size and high polarity, limiting them to extracellular targets. Conversely, small molecules can access intracellular targets but are not ideal for targeting proteins lacking deep hydrophobic pockets. In principle, macrocyclic peptides with sizes between small molecules and proteins can be developed to modulate molecular targets inaccessible to traditional therapeutic modalities<sup 3 https://www.nature.com/articles/s41589-025-01929-w ref-CR3 </sup . The ability to develop custom protein-binding macrocycles for diverse protein targets would have many diagnostic and therapeutic applications. Traditionally, the development of peptide therapeutics has relied on natural product discovery or high-throughput screening of trillions of random peptides for target binding using display-based techniques<sup 1 https://www.nature.com/articles/s41589-025-01929-w ref-CR1 , 2 https://www.nature.com/articles/s41589-025-01929-w ref-CR2 </sup . However, natural product discovery has several challenges, particularly synthetic difficulties, marginal stability and low mutational tolerance of identified hits<sup 4 https://www.nature.com/articles/s41589-025-01929-w ref-CR4 </sup . While powerful, the high-throughput screening methods are time-intensive, cost-intensive and labor-intensive and only span a small fraction of the rich chemical and structural diversity accessible to macrocycles. Moreover, such approaches frequently fail to simultaneously optimize for multiple biophysical properties, such as target binding, selectivity and membrane permeability, because of the precise structural control required to achieve such functional properties<sup 5 https://www.nature.com/articles/s41589-025-01929-w ref-CR5 </sup . Structure-guided design methods offer a complementary approach to the library screening approaches, enabling rapid in silico exploration of a large chemical and structural diversity to design macrocycle binders for therapeutic targets. We previously developed physics-based methods for designing hyperstable constrained peptides, structured macrocycles and binders to protein targets by borrowing the motifs or interactions from previously described binding partners as anchors<sup 6 ref-CR6 , 7 ref-CR7 , 8 ref-CR8 , 9 https://www.nature.com/articles/s41589-025-01929-w ref-CR9 </sup . However, despite the high accuracy observed in the design of monomeric macrocycles with these methods<sup 7 https://www.nature.com/articles/s41589-025-01929-w ref-CR7 </sup , the design of protein-binding macrocycles has had limited success, achieving only modest binding affinities and, in many cases, with the experimentally determined structures not agreeing with the design models<sup 7 https://www.nature.com/articles/s41589-025-01929-w ref-CR7 , 8 https://www.nature.com/articles/s41589-025-01929-w ref-CR8 , 10 https://www.nature.com/articles/s41589-025-01929-w ref-CR10 </sup . The reliance on previously described binding partners for starting motifs also restricts such approaches to well-studied protein targets. In recent work, we described a pipeline for hallucinating and predicting the structures of macrocyclic peptide monomers by modifying AlphaFold2 AF2 to include cyclic relative positional encoding named ‘AfCycDesign’ <sup 11 https://www.nature.com/articles/s41589-025-01929-w ref-CR11 </sup . Other promising deep learning DL methods were described recently to predict the structures of macrocycles and macrocycle–target complexes<sup 12 https://www.nature.com/articles/s41589-025-01929-w ref-CR12 , 13 https://www.nature.com/articles/s41589-025-01929-w ref-CR13 </sup and to design peptide binders to protein targets<sup 14 ref-CR14 , 15 ref-CR15 , 16 https://www.nature.com/articles/s41589-025-01929-w ref-CR16 </sup . However, these methods have not been extensively structurally validated to date or shown to robustly perform atomically accurate de novo design of macrocyclic peptide structures in complexes with diverse protein targets. Computational methods that can accurately design high-affinity macrocycle binders de novo, using just the information of target structure or sequence, are required for wider therapeutic applications. We reasoned that recent breakthroughs in generative DL methods could be leveraged to develop a robust pipeline for the accurate and efficient design of macrocycle binders. Diffusion models for protein design, such as RFdiffusion<sup 17 https://www.nature.com/articles/s41589-025-01929-w ref-CR17 </sup , are trained to generate diverse protein structures from randomly initialized residues as starting points and have demonstrated remarkable success in designing protein monomers, binders and symmetric oligomers of medium-sized to large-sized proteins. However, despite considerable recent progress in DL-based protein design methods, these methods are not readily applicable to designing macrocyclic peptides. Developing analogous methods for peptide design from scratch has been challenging because of the limited availability of experimental data for training such models. To address these challenges, we set out to extend the RoseTTAFold2 RF2 <sup 18 https://www.nature.com/articles/s41589-025-01929-w ref-CR18 </sup structure prediction network and the RFdiffusion<sup 17 https://www.nature.com/articles/s41589-025-01929-w ref-CR17 </sup protein backbone generation framework to incorporate cyclic relative positional encoding and enable the generation of the macrocyclic peptide backbones. Extending RF2 and RFdiffusion for macrocycles We began by examining the ability of the RF2 ref. <sup 18 https://www.nature.com/articles/s41589-025-01929-w ref-CR18 </sup structure prediction network to model known macrocyclic peptide structures. We implemented a modified Methods https://www.nature.com/articles/s41589-025-01929-w Sec7 cyclic relative position encoding for RF2 Fig. 1a https://www.nature.com/articles/s41589-025-01929-w Fig1 and observed robust prediction of natural cyclic peptide structures Supplementary Fig. 1 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Given this success, we reasoned that the same relative positional encoding should enable RFdiffusion<sup 17 https://www.nature.com/articles/s41589-025-01929-w ref-CR17 </sup to generate macrocyclic peptide structures because of its similar network architecture. We added the cyclic positional encoding scheme to RFdiffusion and observed robust generation of diverse macrocyclic peptides Fig. 1b,c https://www.nature.com/articles/s41589-025-01929-w Fig1 and Supplementary Fig. 2 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Similar to the previously described work on designing monomeric cyclic peptides with physics-based methods<sup 7 https://www.nature.com/articles/s41589-025-01929-w ref-CR7 </sup and AfCycDesign<sup 11 https://www.nature.com/articles/s41589-025-01929-w ref-CR11 </sup , we observed 9,045 and 8,913 structurally unique 10-residue and 12-residue backbones, respectively, when 48,000 macrocycle backbones were generated for each size Supplementary Fig. 2 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . The distribution of phi and psi values in these generated backbones is similar to the standard Ramachandran plot for protein structures Supplementary Fig. 2 https://www.nature.com/articles/s41589-025-01929-w MOESM1 , suggesting that generated backbones do not require extensive d-amino acids to stabilize the generated structures<sup 7 https://www.nature.com/articles/s41589-025-01929-w ref-CR7 </sup . While we did not attempt to comprehensively enumerate the structural space of cyclic peptide monomers, RFpeptides can readily be scaled up to comprehensively cover the structural space accessible to macrocyclic peptides. Encouraged by the transferability of the cyclic positional encoding, we set out to use RFdiffusion for the de novo design of protein-binding macrocycles. We chose RFdiffusion for several reasons. Firstly, we expected the high experimental success rate of RFdiffusion<sup 17 https://www.nature.com/articles/s41589-025-01929-w ref-CR17 , 19 https://www.nature.com/articles/s41589-025-01929-w ref-CR19 </sup for protein binder design to carry over to macrocycle binder design. Secondly, de novo binder design with AfCycDesign as is would be far more computationally expensive and has not been successfully implemented or experimentally validated. Thirdly, the method can still take advantage of the current built-in conditional generation functionalities of RFdiffusion, such as epitope-specific targeting and ‘motif’ scaffolding. Lastly, the method should be directly transferable to other current and future RoseTTAFold-based design networks, such as RFdiffusion All-Atom<sup 20 https://www.nature.com/articles/s41589-025-01929-w ref-CR20 </sup , for incorporating nonpeptidic molecules nucleic acids, ions, etc. during design calculations. We modified the RFdiffusion protein binder design pipeline to use cyclic relative position encodings for the generated chain and standard positional encodings for the target and interbinder target indices Fig. 1d https://www.nature.com/articles/s41589-025-01929-w Fig1 . We then completed our design pipeline by using ProteinMPNN<sup 21 https://www.nature.com/articles/s41589-025-01929-w ref-CR21 </sup to design amino acid sequences compatible with the backbones generated by RFdiffusion Fig. 1e https://www.nature.com/articles/s41589-025-01929-w Fig1 . We chose ProteinMPNN for its improved performance in sequence design and ability to generate sequences with better solubility profiles over the sequences generated by traditional physics-based methods<sup 22 https://www.nature.com/articles/s41589-025-01929-w ref-CR22 </sup . This pipeline readily generated macrocycles with diverse secondary structure content against target proteins Fig. 1f https://www.nature.com/articles/s41589-025-01929-w Fig1 and the inclusion of standard RFdiffusion hotspot features clearly shifted the distribution of generated binders toward desired residues Supplementary Fig. 3 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . We refer to this integrated pipeline as ‘RFpeptides’ throughout the remainder of the text. De novo design of macrocyclic binders to myeloid cell leukemia 1 and MDM2 We selected myeloid cell leukemia 1 MCL1 as our first target protein, given the availability of multiple high-resolution X-ray crystal structures available to initiate the design calculations. MCL1 is also a promising target for anticancer therapeutics because of its roles in autophagy, cell survival, DNA repair and cellular proliferation<sup 23 https://www.nature.com/articles/s41589-025-01929-w ref-CR23 </sup . For targeting MCL1, we used RFpeptides to generate 9,965 diverse cyclic peptide backbones, followed by four iterative rounds of ProteinMPNN and Rosetta Relax to design four amino acid sequences for each generated backbone. We expected the local changes to the generated backbone during the Rosetta Relax steps to allow for improved amino acid sequence diversity from the ProteinMPNN steps. While there are other ways to achieve increased sequence diversity, including generating multiple sequences per backbone from ProteinMPNN or adding noise during ProteinMPNN sequence generation, we did not explicitly try or compare them in this study. For downselecting the design candidates for experimental testing, we used AfCycDesign to repredict the designed macrocycle–target complexes from the macrocycle sequence and the target structure as a template. We selected the designs on the basis of the confidence metric interface predicted aligned error iPAE and the similarity between the original design model and the protein–macrocycle complex predicted by the AfCycDesign Supplementary Fig. 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . For further stringency in the design selection process, we also used RF2 to repredict the complex structures, reasoning that the design models predicted identically by two orthogonal structure prediction networks AfCycDesign and RF2 should have a higher likelihood of binding to the target as designed. However, the 1,984 selected designs at this stage were still more than the number of designs we could reasonably synthesize and test experimentally. Therefore, we next used Rosetta<sup 24 https://www.nature.com/articles/s41589-025-01929-w ref-CR24 </sup to calculate the ‘physics-based’ metrics of interface and macrocycle quality, such as calculated binding affinity ddG , spatial aggregation propensity SAP of the designed macrocycle and the molecular surface area of the interface contacts CMS Supplementary Fig. 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . After strictly filtering the designed candidates on DL-based and physics-based metrics, we selected 27 designs for synthesis, biochemical and biophysical characterization. Despite specifying no hotspots to guide the generation process to a specific patch on the MCL1 structure, all selected designs bound to the functionally relevant MCL1–BH3 interaction site Supplementary Fig. 5 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . While all selected designs include an α-helical segment, they feature different sequences, macrocycle placement and target interactions Supplementary Fig. 5 https://www.nature.com/articles/s41589-025-01929-w MOESM1 and Supplementary Table 1a https://www.nature.com/articles/s41589-025-01929-w MOESM1 . In addition to the common helical motifs, the loop regions of the selected macrocycles also contribute extensive side-chain-mediated and backbone-mediated interactions to the binding interface. During the chemical synthesis using Fmoc-based solid-phase synthesis Methods https://www.nature.com/articles/s41589-025-01929-w Sec7 , the yields for the correctly cyclized product for 13 designs were low and insufficient for further characterization. We tested the remaining 14 macrocycles for binding to biotinylated MCL1 using surface plasmon resonance SPR single-cycle kinetics experiments Supplementary Fig. 6 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Three macrocycles showed binding to the MCL1, with the best binder, MCB D2 MCL1 binding design 2 Fig. 2a https://www.nature.com/articles/s41589-025-01929-w Fig2 , demonstrating a binding affinity of 2 µM Fig. 2b https://www.nature.com/articles/s41589-025-01929-w Fig2 . To confirm whether the designed macrocycle adopts the designed structure and engages MCL1 in the designed binding mode, we determined the X-ray crystal structure of MCB D2 bound to MCL1 at 2.1 Å resolution. The crystal structure was nearly identical to the design model, with a root-mean-square deviation r.m.s.d. of 0.7 Å over all of the Cα atoms of the macrocycle with target chains aligned Fig. 2c https://www.nature.com/articles/s41589-025-01929-w Fig2 and Cα r.m.s.d. of 0.4 Å within the macrocycles when aligned Fig. 2d https://www.nature.com/articles/s41589-025-01929-w Fig2 . The side-chain rotamers of the interacting residues in the crystal structure also closely matched the design model Fig. 2d https://www.nature.com/articles/s41589-025-01929-w Fig2 . The crystal structure also confirmed that the binding interactions are not restricted to the helix region of the designed macrocycle but are also contributed by the loop regions Fig. 2e,f https://www.nature.com/articles/s41589-025-01929-w Fig2 . While several hydrophobic interactions from the MCB D2 helical segment are similar to those seen in natural MCL1 binders for example, BH3 peptide , Supplementary Fig. 7 https://www.nature.com/articles/s41589-025-01929-w MOESM1 , the N-to-C orientation of the helix is flipped in the case of MCB D2. The loop region of MCB D2 makes additional hydrophobic contacts and a cation–π interaction with MCL1 Fig. 2e https://www.nature.com/articles/s41589-025-01929-w Fig2 and Supplementary Fig. 7d https://www.nature.com/articles/s41589-025-01929-w MOESM1 that we did not observe in previously reported natural MCL1 binders and their analogs. All three hits with an observable binding signal at 100 µM featured this cation–π interaction. Encouraged by the experimental validation of the MCL1 binding macrocycles, we next sought to design binders to MDM2, an E3 ligase that interacts with tumor suppressor protein p53 and has multiple critical roles in tumor growth and survival<sup 25 https://www.nature.com/articles/s41589-025-01929-w ref-CR25 </sup . We generated 10,000 macrocycle backbones spanning diverse lengths amenable to chemical synthesis 16–18 residues and designed four amino acid sequences for each generated backbone using iterative rounds of ProteinMPNN and Rosetta Relax protocols Methods https://www.nature.com/articles/s41589-025-01929-w Sec7 . Design models were filtered on the basis of the confidence metrics and similarity of the AfCycDesign predictions to the designed complexes and the interface quality metrics calculated using Rosetta Supplementary Fig. 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . AfCycDesign predicted 7,495 of the 40,000 design models to bind MDM2 with high confidence normalized iPAE < 0.3 Supplementary Fig. 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . In contrast to our approach for MCL1, we chose not to do any additional filtering with RF2 as the results between AfCycDesign and RF2 were fairly consistent. We also adjusted the filter thresholds for in silico filters as their overall distribution differed substantially from the distribution observed for MCL1 Supplementary Fig. 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . After filtering on interface metrics Methods https://www.nature.com/articles/s41589-025-01929-w Sec7 , we identified 17 designs with iPAE < 0.3, ddG < −50 kcal mol<sup −1</sup , CMS 300 Å<sup 2</sup and SAP < 35. We selected 11 top-ranked designs by ddG for biochemical and biophysical characterization. The 11 selected designs had diverse sizes, shapes and sequences Supplementary Fig. 8 https://www.nature.com/articles/s41589-025-01929-w MOESM1 and Supplementary Table 2 https://www.nature.com/articles/s41589-025-01929-w MOESM1 ; however, they were all predicted to bind the same site as the p53 transactivation domain Supplementary Fig. 8 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Three of the selected designs had poor yields during the cyclization step of the chemical synthesis, preventing further experimental characterization with them. We tested the remaining eight peptides for binding to the biotinylated MDM2 by SPR and identified three binders with observable binding signals at 100 µM Supplementary Fig. 9 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . The best design, MDB D8 Fig. 2g https://www.nature.com/articles/s41589-025-01929-w Fig2 , demonstrated a binding affinity of 1.9 µM in the SPR single-cycle kinetics experiment Fig. 2h https://www.nature.com/articles/s41589-025-01929-w Fig2 . The computational model for this design makes several key contacts at the interface that are similar to interactions observed in native MDM2–p53 complex structures Fig. 2i https://www.nature.com/articles/s41589-025-01929-w Fig2 and Supplementary Fig. 10 https://www.nature.com/articles/s41589-025-01929-w MOESM1 <sup 25 https://www.nature.com/articles/s41589-025-01929-w ref-CR25 </sup . Despite different overall structures, all three hits from the SPR screen had a similar binding motif composed of phenylalanine, tryptophan and either leucine or methionine from the helical segment of the macrocycle. Together, these data highlight the promising accuracy of the RFpeptides pipeline to design diverse macrocyclic binders for selected targets of interest. De novo design of macrocyclic binders to γ-aminobutyric acid type A receptor-associated protein We next set out to design binders against a target with a binding site that is structurally different from MCL1 and MDM2, formed by a mix of α-helices and β-strands in contrast to all α-helical pockets of MCL1 and MDM2 . We selected γ-aminobutyric acid type A receptor-associated GABARAP as the target, a protein responsible for mediating autophagy through its role in autophagosome biogenesis and recruitment of cargo, resulting in lysosomal degradation of damaged or surplus proteins and organelles<sup 26 https://www.nature.com/articles/s41589-025-01929-w ref-CR26 </sup . Peptide modulators against GABARAP could have therapeutic applications in the treatment of late-stage cancers<sup 27 https://www.nature.com/articles/s41589-025-01929-w ref-CR27 </sup or as chimeric peptides for autophagy-mediated targeted protein degradation<sup 28 https://www.nature.com/articles/s41589-025-01929-w ref-CR28 </sup . Our target binding site for GABARAP, which is also the binding site for the native LC3-interacting region or Atg8-interacting motif<sup 29 https://www.nature.com/articles/s41589-025-01929-w ref-CR29 </sup , is formed by a mix of β-strand and α-helix secondary structures Fig. 3 https://www.nature.com/articles/s41589-025-01929-w Fig3 and Supplementary Fig. 13 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . For designing macrocyclic binders against the human GABARAP, we used a similar pipeline as described above for MCL1 and MDM2 Methods https://www.nature.com/articles/s41589-025-01929-w Sec7 but we doubled the number of generated designs and defined six hotspot residues Lys46, Lys48, Tyr49, Leu50, Phe60 and Leu63 to guide the macrocycle backbone generation to a specific site on the target Fig. 3a,d https://www.nature.com/articles/s41589-025-01929-w Fig3 . We generated 20,000 macrocycle backbones and designed the amino acid sequences using ProteinMPNN and Rosetta Relax protocols. Of the resulting 80,000 design models, we selected 335 macrocyclic designs on the basis of AfCycDesign iPAE < 0.13 and Rosetta ddG < −30 kcal mol<sup −1</sup , SAP < 35 and CMS 300 Å<sup 2</sup interface metrics Supplementary Fig. 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Instead of trying to synthesize and characterize all 335 cyclic peptides which would have required substantial time and experimental resources , we clustered the 335 designs into 80 different clusters on the basis of their three-dimensional structures and selected representative designs from diverse clusters for further biochemical characterization. We selected 13 diverse macrocycles of 12–17 residues for synthesis and experimental validation Supplementary Table 3 https://www.nature.com/articles/s41589-025-01929-w MOESM1 and Supplementary Fig. 11 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Unlike the design candidates described above for MCL1 and MDM2, several of the selected macrocycles for GABARAP showed cyclic β-sheet structures with several edge–strand interactions with the target Supplementary Fig. 11 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . We successfully synthesized six designs with high purity 90% and tested them for binding to GABARAP using SPR Supplementary Fig. 12 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Two designs, GAB D8 and GAB D23, showed binding affinities of 6 nM and 36 nM, respectively Fig. 3b,e https://www.nature.com/articles/s41589-025-01929-w Fig3 . To further characterize the binding of GAB D8 and GAB D23, we tested the ability of these designs to disrupt the interaction of GABARAP with linear peptide K1 a previously described binder to this site<sup 30 https://www.nature.com/articles/s41589-025-01929-w ref-CR30 </sup in AlphaScreen assays. GAB D8 and GAB D23 demonstrated a half-maximal inhibitory concentration IC<sub 50</sub of 0.7 nM and 2.5 nM in the AlphaScreen assay, respectively Fig. 3i https://www.nature.com/articles/s41589-025-01929-w Fig3 . To our knowledge, GAB D8 is the most potent macrocyclic GABARAP binder to date. In crystallization trials, we did not obtain crystals of sufficiently high quality for GAB D8 bound to GABARAP. We instead crystallized GAB D8 bound to GABARAPL1, a homolog of GABARAP with 86% overall sequence identity and 100% sequence identity for residues within 5 Å of GAB D8 in the design model. The X-ray crystal structure for GAB D8 bound to GABARAPL1 matched very closely with the design model, with a Cα r.m.s.d. of 1.2 Å over the macrocycle when aligned by the target protein to the closest of the four copies in the asymmetric unit Fig. 3c https://www.nature.com/articles/s41589-025-01929-w Fig3 and Supplementary Fig. 13 https://www.nature.com/articles/s41589-025-01929-w MOESM1 and a Cα r.m.s.d. of 0.47 Å when aligned by macrocycle alone Fig. 3g https://www.nature.com/articles/s41589-025-01929-w Fig3 . Notably, the X-ray structure of the GAB D8–GABARAPL1 complex showed two different bound conformations of GAB D8, one that closely matched the design model and a second one that partially deviated from the design model Supplementary Fig. 14 https://www.nature.com/articles/s41589-025-01929-w MOESM1 , with a register shift nucleated by Thr10 from the macrocycle forming main-chain-mediated and side-chain-mediated hydrogen bonds with Lys48 on the target. GAB D23 crystallized readily with GABARAP and also closely matched the design model with a Cα r.m.s.d. of 1.7 Å when aligned by the target Fig. 3f https://www.nature.com/articles/s41589-025-01929-w Fig3 and Cα r.m.s.d. of 0.74 Å across the macrocycle alone Fig. 3g https://www.nature.com/articles/s41589-025-01929-w Fig3 . The X-ray crystal structure confirmed the key designed interactions, such as Trp5 and Ile8, with the main difference between the design model and the X-ray structure being the switch from a type I β-turn from Leu1 to Gly4 in the design model to a less regular conformation in the crystal structure, with a tendency for a type I′ β-turn from Glu2 to Trp5. While our original design models were predicted with single sequences as inputs to AF2, we retrospectively predicted the GAB D8–GABARAPL1 and GAB D23–GABARAP complex structures with multiple-sequence alignment MSA inputs. These MSA-based predictions of the designs matched even more closely with the X-ray crystal structures, with a Cα r.m.s.d. of 0.5 Å and 0.9 Å for the GAB D8–GABARAPL1 and GAB D23–GABARAP complexes, respectively, when aligned by the target structure Supplementary Fig. 15 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Overall, these data demonstrate the ability of our de novo design pipeline to identify high-affinity binders against targets with diverse pocket shapes and surfaces without requiring library-scale screening. Design of macrocyclic binders to predicted structures Given the high accuracy and binding affinity of macrocycles designed against selected targets, we next set out to design macrocyclic binders against targets without any experimentally determined structure. We reasoned that the high accuracy of RFpeptides could mitigate the inherent risk of designing against a predicted target structure. We designed macrocycles against Rhombotarget A RbtA , a recently identified cell surface protein from the ESKAPE pathogen, Acinetobacter baumannii . There are no experimentally determined structures available for this protein and sequence-based searches against the Protein Data Bank PDB did not return notable matches to other protein structures. We predicted the structure of the 617-aa full-length protein using AF2 and RF2; both methods predicted similar overall structures Cα r.m.s.d. of 0.4 Å over 509 residues excluding the signal peptide and transmembrane domain with high confidence predicted local distance difference test pLDDT 90 Supplementary Fig. 16 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . AF2 and RF2 both predicted two distinct extracellular domains: an N-terminal β-helix domain and a C-terminal Ig-like domain Supplementary Fig. 16 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . While there were some differences in the predicted structures from AF2 and RF2, we decided to focus our binder design calculations on regions that were predicted nearly identically and with high confidence by AF2 and RF2. On the basis of our preliminary design runs without hotspots to guide the diffusion, we identified a patch in the N-terminal domain to pursue in our large-scale design calculations against this target and defined hotspots Leu144, Phe202, Phe204, Tyr206, Val208, Leu231 and Ala269 for peptide backbone generation Fig. 4a https://www.nature.com/articles/s41589-025-01929-w Fig4 . In contrast to the concave pockets targeted for MDM2 and MCL1, this selected patch for RbtA is considerably flatter and difficult to target with conventional computational and experimental approaches Supplementary Fig. 17 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . We generated 20,000 backbones for macrocycle binders and designed four amino acid sequences for each backbone using iterative rounds of ProteinMPNN and Rosetta Relax. Designs were filtered using AfCycDesign confidence metrics and Rosetta interface metrics, as described in earlier sections Supplementary Fig. 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . On the basis of these in silico metrics, we selected 26 designs for biochemical and structural characterization with AfCycDesign iPAE < 0.4, ddG < –30 kcal mol<sup −1</sup , r.m.s.d. between the design model and AfCycDesign prediction < 1.5 Å and CMS 300 Å<sup 2</sup Supplementary Fig. 18 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . The selected designs covered diverse sizes 13–18 aa , sequences, shapes and secondary structures Supplementary Fig. 18 https://www.nature.com/articles/s41589-025-01929-w MOESM1 and Supplementary Table 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . We expressed the Avi-tagged version of the RbtA N-terminal domain residues 20–458 and used it for binding screens using SPR. Four of 11 designs that were synthesized in sufficient quantity and purity showed a binding signal at 100 µM in our screens Supplementary Fig. 19 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . On the basis of further binding experiments with SPR, we determined the dissociation constant K <sub d</sub of the best binder, RBB D10, to be 9.4 nM Fig. 4b https://www.nature.com/articles/s41589-025-01929-w Fig4 . The design model for RBB D10 showed extensive contacts to the target with several side-chain-mediated polar contacts and hydrophobic interactions Fig. 4f–h https://www.nature.com/articles/s41589-025-01929-w Fig4 . To confirm the structures of RbtA and RBB D10 and the binding mode between them, we determined the high-resolution X-ray crystal structure of apo and macrocycle-bound RbtA using X-ray crystallography at 2 Å and 2.6 Å resolution, respectively. The apo structure of the RbtA N-terminal domain, which is also the first experimentally determined structure from this class of bacterial proteins, matched our AF2 and RF2 predictions for this target very closely, with an overall Cɑ r.m.s.d. of 1.2 Å and 1.1 Å between the X-ray structure of the RbtA N-terminal domain and the AF2-predicted and RF2-predicted structures, respectively Fig. 4c https://www.nature.com/articles/s41589-025-01929-w Fig4 . The complex structure also confirmed the structure and binding mode of our designed macrocycle, RBB D10, with the X-ray structure matching the design model with an r.m.s.d. of 1.4 Å Fig. 4d https://www.nature.com/articles/s41589-025-01929-w Fig4 . Notably, the conformation adopted by the macrocycle in the X-ray structure, including the side-chain rotamers involved in interactions with the target, was almost identical to the design model with an r.m.s.d. of 0.4 Å Fig. 4e–h https://www.nature.com/articles/s41589-025-01929-w Fig4 . Together, these data highlight the high accuracy and success rates provided by RFpeptides even while designing macrocycles against targets without deep pockets or targets with no known structures. Overall, these data show that RFpeptides can sample extensive structural and chemical diversity of macrocycles during the backbone and sequence generation steps against selected targets and, finally, select the shapes and sequences ideally suited for binding the target surface or pockets. The highest-affinity binders against each target are also predicted to fold into the bound conformations even in the absence of the target Supplementary Fig. 20 https://www.nature.com/articles/s41589-025-01929-w MOESM1 , suggesting that macrocycles are designed to fold into binding-competent conformations. For all four design campaigns described here, selected designs demonstrated good solubility in aqueous buffers despite not imposing any particular sequence constraints related to solubility during the sequence design step using ProteinMPNN<sup 21 https://www.nature.com/articles/s41589-025-01929-w ref-CR21 </sup . Notably, combining DL-based and physics-based in silico filters helps to select medium to high-affinity binders. However, we note that the distribution of such metrics varies substantially across the four selected targets and adjustments to filtering thresholds were required on the basis of the shape and chemical composition of the target pocket. While in silico metrics enrich well for binders, the relative ranking within the selected designs does not perfectly match the experimental binding affinities. The highest-affinity binders for MDM2 and RbtA had the best or second-best iPAE values among the designs chosen for those targets Supplementary Tables 2b https://www.nature.com/articles/s41589-025-01929-w MOESM1 and 4b https://www.nature.com/articles/s41589-025-01929-w MOESM1 ; however, the hit peptides against MCL1 and GABARAP were not among the top three ranked designs Supplementary Tables 1b https://www.nature.com/articles/s41589-025-01929-w MOESM1 and 3b https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Integration with high-throughput methods in the future should enable testing of more designs and inform absolute threshold values and filtering schemes for the single-shot design of peptide binders to any arbitrary target. Discussion Here, we describe RFpeptides, a generative DL pipeline for precise de novo design of macrocycle binders against a wide range of protein targets. The power of the approach is highlighted by the high affinities K <sub d</sub < 10 nM of the designed macrocyclic binders to GABARAP and RbtA and the nearly identical X-ray crystal structures and design models of the macrocycle-bound MCL1, GABARAP and RbtA Cα r.m.s.d. of 0.7 Å, 1.2 Å and 1.4 Å, respectively . The RFpeptides approach offers several advantages over traditional methods. Firstly, the design approach should enable faster and more efficient discovery of macrocyclic binders. Despite testing fewer than 20 designed candidates per target in contrast to trillions of peptides tested in traditional library-based approaches , we achieved high-affinity binders for two targets without requiring any further experimental optimization; to our knowledge, this is a considerably higher success rate than achieved with any previous method. Secondly, in contrast to the untargeted nature of the random library-based approaches, RFpeptides can be used for designing custom binders to specific patches and sites, as demonstrated for GABARAP and RbtA. Lastly, the atomically accurate nature of the design models enables structure-guided optimization for properties beyond target binding as well as further increases in affinity , bypassing the bottleneck of complex structure determination, which has hindered the optimization of leads from library screening. Combined with the design principles for membrane traversal, RFpeptides could enable the design of peptides simultaneously optimized for target binding and cell permeability or oral bioavailability. RFpeptides also has considerable advantages over previous computational peptide design methods. Information on known ligands and/or binding partners is not required to initiate design. RFpeptides can design macrocycles completely de novo from just the structure or sequence as in the case of RbtA of the target, enabling design against molecular targets intractable with previous methods. RFpeptides is not limited to generating macrocycles with particular motifs or topologies; the diffusion process generates macrocycles with diverse shapes and sizes and selects the topologies appropriate for the protein being targeted. Among the four targets tested here, binders for MCL1 and MDM2 have helical motifs, binders for GABARAP have a β-sheet topology and binders for RbtA sample looplike conformations that make extensive contacts with the flat surface of this target. We anticipate that RFpeptides will enable the rapid design of custom macrocyclic binders against a wide range of molecular targets, accelerating efforts to develop peptides for diverse functional applications. With the rapid advances in DL methods and frameworks, including the recent development of all-atom diffusion models, we aim to extend the approach to generative design of macrocycles with noncanonical amino acids, crosslinkers and cyclization chemistries. Methods Computational methods for cyclic peptide binder design Macrocyclic peptide monomers and binders were designed with RFpeptides using a three-stage pipeline: backbone generation using RFdiffusion with the cyclic offset applied to the peptide chains, followed by sequence design using ProteinMPNN and, finally, structure prediction of the designed peptide–target complexes using either AfCycDesign and/or RoseTTAFold with the cyclic offset applied to the peptide. Designs were further filtered and downselected using Rosetta metrics and, in some cases, clustered on the basis of Cα r.m.s.d. Detailed computational methods, including example scripts, can be found in Supplementary Section 2.2 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Peptide synthesis Macrocyclic peptides described here were either purchased from Wuxi AppTec at greater than 90% purity or synthesized in-house using Fmoc-based solid-phase peptide synthesis. Peptides were typically synthesized on preloaded CTC resin. The resin was swollen in DCM followed by iterative deprotection with 20% piperidine in DMF and coupling with either HBTU Sigma or PyAOP Novabiochem and DIEA Sigma . The linear peptides were cleaved from the resin using either 2% TFA in DCM or 20% HFIP Oakwood Chemical in DCM. The solvent was removed by rotary evaporation and linear protected peptides were cyclized in either DCM, DMF or a mixture of both depending on the solubility of the peptide, using two equivalents of PyAOP and five equivalents of DIEA overnight. The protecting groups were removed using a cocktail of 95:2.5:2.5, TFA, water and TIPS for 2.5 h. The crude peptides were precipitated using cold diethyl ether. The precipitate containing the crude cyclization reaction was dissolved in a mixture of water and acetonitrile for purification using reverse-phase high-performance liquid chromatography LC . Peptide identities were confirmed by mass spectrometry MS . Purities for all synthesized and tested macrocyclic peptides are also summarized in Supplementary Tables 7 https://www.nature.com/articles/s41589-025-01929-w MOESM1 – 10 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . The mass spectrograms and analytical LC chromatograms for all purified peptides are shown in Supplementary Section 4 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . Protein expression and purification MDM2 and MCL1 The amino acid sequences of MCL1 PDB 2PQK https://doi.org/10.2210/pdb2PQK/pdb <sup 37 https://www.nature.com/articles/s41589-025-01929-w ref-CR37 </sup and MDM2 PDB 4HFZ https://doi.org/10.2210/pdb4HFZ/pdb <sup 38 https://www.nature.com/articles/s41589-025-01929-w ref-CR38 </sup were retrieved from the PDB. The optimized genes were then cloned into a Novogen pRSF-DUET plasmid Sigma, 71341-3 , incorporating a 6xHis-tag at the N terminus, followed by an Avi-tag and a tobacco etch virus TEV protease cleavage site. The resulting constructs were codon-optimized for Escherichia coli expression and synthesized by Genscript. For propagation, the plasmids were transformed into E. coli NEBα cells New England Biolabs, C2987 ; for protein expression, the plasmids were transformed into E. coli BL21 DE3 cells New England Biolabs, C2527 . A single sequence-verified colony was cultured in 50 ml of kanamycin 50 µg ml<sup −1</sup selective Luria Broth LB medium. This culture was incubated at 37 °C with shaking at 200 rpm for 16 h overnight. Subsequently, 50 units of optical density at 600 nm OD<sub 600</sub of the overnight culture were transferred to 1 L of fresh kanamycin 50 µg ml<sup −1</sup selective LB medium. The culture was grown at 37 °C with shaking at 200 rpm for 2 h until it reached an OD<sub 600</sub of 0.4–0.5 , at which point the temperature was decreased to 20 °C. The culture was grown until an OD<sub 600</sub of 0.7–0.8; protein expression was induced by adding 1 mM IPTG and the culture was left to grow overnight for 14 h. Cells were harvested by centrifugation at 5,000 g for 10 min at 4 °C, resulting in a cell pellet with a density of 5 g L<sup −1</sup . The pellet was immediately flash-frozen and stored at −20 °C for later use. For lysis, the pellet was thawed on ice and resuspended in 5 ml of lysis buffer per gram of pellet. This lysis buffer contained 50 mM Tris-HCl, 300 mM NaCl and 10 mM imidazole and was supplemented with 1× BugBuster protein extraction reagent Sigma-Aldrich, 70921 , 200 µg ml<sup −1</sup lysozyme Sigma-Aldrich, L6876 , 25 U per ml benzonase nuclease Sigma-Aldrich, E8263 and 1× cOmplete EDTA-free protease inhibitor cocktail Sigma-Aldrich, 11836170001 . The buffer was filter-sterilized using a 0.2 µm filter before the addition of benzonase, mixed by inversion and kept on ice until use. Cells were completely resuspended in the lysis buffer using a homogenizer at low speed and incubated for 30 min at room temperature 22–25 °C . Following incubation, the suspension was sonicated using a Q500 Sonicator equipped with a four-tip probe. Sonication was conducted for 2–3 min using pulses of 10–15 s on followed by 10–15 s off at 70% amplitude. The lysate was clarified by centrifugation at 16,000 g for 20 min. Ni-NTA agarose resin Qiagen, 30210 was equilibrated with 20 column volumes CV of ultrapure water, followed by 20 CV of equilibration buffer 50 mM Tris-HCl, 300 mM NaCl and 10 mM imidazole . Then, 4 ml of 50% resin suspended in equilibration buffer was used to bind His-tagged proteins from 25 ml of clarified lysate. All immobilized metal affinity chromatography IMAC steps were conducted at 4 °C. The lysate–resin mixture was incubated for 60 min on a rotary shaker set to a slow speed. After incubation, the resin was transferred to a 20-ml gravity column and allowed to completely settle. The resin was first washed with 20 CV of wash buffer 1 20 mM Tris-HCl, 250 mM NaCl, 10 mM imidazole and 5 mM β-mercaptoethanol , followed by another 20 CV of wash buffer 2 20 mM Tris-HCl, 500 mM NaCl and 35 mM imidazole . The bound proteins were then eluted with 8 ml of elution buffer 20 mM Tris-HCl, 250 mM NaCl, 350 mM imidazole and 2 mM DTT . Aliquots of the eluate were collected and analyzed using SDS–PAGE gels. The eluate was loaded onto a pre-equilibrated Superdex 75 10/300 GL column 25 mM Tris-HCl, 250 mM NaCl and 2 mM DTT and run at a flow rate of 0.6 ml min<sup −1</sup using an ÄKTA pure system for size-exclusion chromatography SEC . Then, 1 ml fractions were collected from the elution volume of 8–16 ml and those corresponding to peaks in the absorbance at 280 nm between an elution volume of 10 and 13 ml were assessed with SDS–PAGE gels. Fractions confirming the expected molecular weight were pooled and concentrated by centrifugation at 4,000 g for 30 min at 4 °C using Amicon Ultra-4 concentrators with a 3 kDa cutoff Millipore Sigma, UFC800308 to a final volume of 500 µl. The identity of the eluted proteins were confirmed by MS using an Agilent 6230 LC–MS time-of-flight system. Verified protein samples were processed for further applications: biotinylation for SPR analysis or tag removal by TEV protease cleavage for crystallography. Biotinylation was performed using the BirA biotin–protein ligase standard reaction kit Avidity, BirA-500 according to the manufacturer’s recommended conditions. The reaction was carried out at 4 °C overnight on a slowly shaking platform. For TEV protease cleavage, the proteins were treated with a 25:1 protein to TEVd enzyme ratio<sup 39 https://www.nature.com/articles/s41589-025-01929-w ref-CR39 </sup . Similarly, the mixture was incubated at 4 °C overnight on a slowly shaking platform. Following these treatments, samples underwent a cleanup step using 1 ml of Ni-NTA resin per 20 mg of protein. The resin was pre-equilibrated with 10 CV of ultrapure water and 10 CV of a buffer containing 25 mM Tris-HCl, 250 mM NaCl and 10 mM imidazole. The pre-equilibrated resin was added to the protein mixture and incubated for 30 min on a rolling platform at 4 °C. Subsequently, the mixtures were filtered through a 0.45 µm PVDF centrifugal filtering unit to remove the Ni-NTA-bound substrates. The eluate was collected and dialyzed in 2 L of 25 mM Tris-HCl, 250 mM NaCl and 2 mM DTT using a Slide-A-Lyzer G3 dialysis cassettes with a 3.5 kDa molecular weight cutoff Thermo Scientific, A52966 overnight for 18 h at 4 °C stirring. The dialyzed protein was concentrated to 0.2–0.5 ml as required for downstream assays , using the Amicon ultra concentrators as above , aliquoted and flash-frozen. Fractions were analyzed by mass spectroscopy for the efficacy of the biotinylation and TEV protease cleavage treatments, as previously described. GABARAP for SPR A synthetic complementary DNA was designed on the basis of the amino acid sequence of GABARAP UniProt O95166 https://www.uniprot.org/uniprot/O95166 and optimized for expression in E. coli using Benchling software. The construct was devised to include an N-terminal Avi-tag and TEV protease cleavage site and was cloned into the Novogen pET-50b + plasmid. This plasmid configuration introduced a tandem arrangement of protein tags at the N terminus: a 6xHis-tag, followed by a NusA solubility tag, another 6xHis-tag and a human rhinovirus HRV 3C protease cleavage site. Therefore, the final construct sequence was as follows: 6xHis–NusA–6xHis–HRV 3C–Avi–TEV–GABARAP. NusA was specifically chosen as a solubility tag because of its known effectiveness in enhancing protein solubility in E. coli <sup 40 https://www.nature.com/articles/s41589-025-01929-w ref-CR40 , 41 https://www.nature.com/articles/s41589-025-01929-w ref-CR41 </sup . The construct was synthesized and cloned by Genscript. As described above for MCL1 and MDM2 protein expression, the plasmids were introduced into E. coli NEBα cells and BL21 DE3 cells. A single sequence-verified colony was cultured in 50 ml of kanamycin 50 µg ml<sup −1</sup selective LB medium for 16 h at 37 °C, shaking at 200 rpm. Then, 50 OD<sub 600</sub units of this culture were transferred to 1 L of fresh kanamycin 100 µg ml<sup −1</sup selective autoinduction medium TBM-5052: 1.2% w/v tryptone, 2.4% w/v yeast extract, 0.5% v/v glycerol, 0.05% w/v d-glucose, 0.2% w/v d-lactose, 25 mM Na<sub 2</sub HPO<sub 4</sub , 25 mM KH<sub 2</sub PO<sub 4</sub , 50 mM NH<sub 4</sub Cl, 5 mM Na<sub 2</sub SO<sub 4</sub , 2 mM MgSO<sub 4</sub , 10 μM FeCl<sub 3</sub , 4 μM CaCl<sub 2</sub , 2 μM MnCl<sub 2</sub , 2 μM ZnSO<sub 4</sub , 400 nM CoCl<sub 2</sub , 400 nM NiCl<sub 2</sub , 400 nM CuCl<sub 2</sub , 400 nM Na<sub 2</sub MoO<sub 4</sub , 400 nM Na<sub 2</sub SeO<sub 3</sub and 400 nM H<sub 3</sub BO<sub 3</sub . The culture was grown at 37 °C with shaking at 200 rpm for 2 h, at which point the temperature was decreased to 22 °C and the culture was left to grow for 16 h. Cells were harvested, lysed and purified following the protocol outlined earlier for MCL1 and MDM2, with some modifications. The cultures yielded a cell pellet amounting to 15 g L<sup −1</sup . Lysis was completed using an IKA T18 microfluidizer at 450 psi, followed by lysate clarification by centrifugation at 16,000 g for 15 min. All IMAC steps were conducted at 22 °C, except for the incubation of the lysate–resin mixture, which was performed at 4 °C. Proteins bound to the resin were eluted with 5 ml of elution buffer 50 mM Tris-HCl pH 8, 250 mM NaCl and 300 mM imidazole . SEC was then performed using a Superdex 200 Increase 10/300 GL column Cytiva equilibrated with TBS 50 mM Tris-HCl pH 8 and 250 mM NaCl . Fractions confirmed by SDS–PAGE were pooled and concentrated using Amicon Ultra-15 concentrators with a 30 kDa cutoff Millipore Sigma, UFC9030 to a final volume of 1 ml. Downstream processing for SPR analysis was performed as described previously, with one modification. For biotinylation, the protein was first cleaved using HRV 3C protease with the reagents and protocol provided by the Pierce HRV 3C protease solution kit Thermo Scientific, 88946 . The digested samples were subsequently purified and verified, as outlined in earlier sections. GABARAP and GABARAPL1 for crystallography GABARAP and GABARAPL1 were expressed as glutathione S -transferase GST fusion proteins after transforming E. coli BL21 DE3 T1 cells with pGEX4T2-GABARAP and pGEX4T2-GABARAPL1 plasmids, respectively. Bacteria were cultivated in LB medium containing 100 µg ml<sup −1</sup ampicillin; gene expression was induced with 1 mM IPTG at an OD<sub 600</sub of 0.6–0.8 and allowed to proceed for 20 h at 25 °C. Afterward, cells were harvested by centrifugation at 3,000 g for 30 min at 4 °C. The bacterial pellet was washed with PBS 137 mM NaCl, 2.7 mM KCl, 1.8 mM KH<sub 2</sub PO<sub 4</sub and 10 mM Na<sub 2</sub HPO<sub 4</sub and resuspended in lysis buffer PBS supplemented with 5% v/v glycerol, 0.01% v/v β-mercaptoethanol, 10 µg ml<sup −1</sup DNase AppliChem, A3778 and cOmplete EDTA-free protease inhibitor cocktail Roche, 11836170001 before application to the cell disruptor Constant Systems, model TS1.1 for three cycles with 1.9 kbar at 4 °C. Lysates were cleared by centrifugation at 4 °C with 45,000 g for 45 min. The GST fusion proteins were purified from the supernatant by affinity chromatography using glutathione Sepharose 4B Cytiva, 1705605 . Cleavage with thrombin Sigma-Aldrich, 1.12374 during dialysis against 10 mM Tris-HCl and 150 mM NaCl pH 7.0 at 4 °C overnight yielded 119 amino acid proteins carrying an N-terminal Gly-Ser extension in addition to the native residues of GABARAP and GABARAPL1. Subsequently, samples were applied to a Hiload 26/60 Superdex 75 preparatory-grade size-exclusion column GE Healthcare equilibrated with 10 mM Tris-HCl and 150 mM NaCl pH 7.0 . Protein purity was assessed by SDS–PAGE and Coomassie staining. Fractions containing the eluted proteins were concentrated to 3–5 mg ml<sup −1</sup using Vivaspin 20 concentrators with a 3 kDa cutoff Sartorius , flash-frozen in liquid N<sub 2</sub and kept at −80 °C for long-term storage. RbtA β-helix domain For heterologous expression of the β-helix domain of RbtA residues A20–I459 in E. coli , the gene was amplified and fused with a SNAC tag GSHHWGS at the C terminus using the following primers: forward, GCTGCCCAGCCGGCGATGGCCATGGGCGCTGATATTGAAGTCACAACTAC; reverse, CAGTGGTGGTGGTGGTGGTGCTCGAGGCTGCCCCAATGATGGCTGCCGATATATTCAATTGCGCCTAAAT<sup 42 https://www.nature.com/articles/s41589-025-01929-w ref-CR42 </sup . The fragment was inserted into NcoI-digested and XhoI-digested pET-22b + by Gibson assembly to generate a construct with a C-terminal 6xHis fusion. The construct was confirmed by sequencing and transformed into E. coli Rosetta DE3 cells. To purify the β-helix domain of RbtA, an overnight culture of Rosetta DE3 cells carrying the construct was back-diluted 1:300 in 2× YT broth and grown at 37 °C with shaking at 200 rpm until the OD<sub 600</sub reached 0.4. The incubation temperature was reduced to 18 °C, IPTG was added to a final concentration of 0.3 mM and the culture was incubated for a total of 18 h. Cells were then collected by centrifugation and resuspended in lysis buffer containing 200 mM NaCl, 50 mM Tris-HCl pH 7.5, 10% glycerol v/v , 5 mM imidazole, 0.5 mg ml<sup −1</sup lysozyme and 1 mU of benzonase. Cells were then lysed by sonication and cellular debris was removed by centrifugation at 35,000 g for 30 min at 4 °C. The protein was purified from lysates using a 1 ml HisTrap HP column on an ÄKTA fast protein LC FPLC system. Column-bound protein was eluted using a linear imidazole gradient from 5 to 500 mM. Protein purity was assessed by SDS–PAGE and Coomassie staining. The fractions with high purity were concentrated using a 30 kDa cutoff Amicon filter and then further purified by FPLC using a HiLoad 16/600 Superdex 200 preparatory-grade column GE Healthcare equilibrated with sizing buffer 500 mM NaCl, 50 mM Tris-HCl pH 7.5 and 10% glycerol v/v . The fractions with high purity were concentrated and used for evaluation of macrocyclic binders or determination of X-ray structure. For determination of the X-ray crystal structure of RbtA, the C-terminal 6xHis-tag was removed by chemical cleavage at the SNAC tag. In brief, the buffer of the concentrated protein was exchanged to cleavage buffer 0.1 M CHES, 0.1 M NaCl, 0.1 M acetone oxime and 5 mM Fos-choline-12, pH 8.6 . The protein solution was diluted to 1 mg ml<sup −1</sup , followed by the addition of 1 mM TCEP and 1 mM NiCl<sub 2</sub . The mixture was vortexed and incubated at room temperature for 16 h. The precipitation was removed by centrifugation at 35,000 g for 30 min at 4 °C. The supernatant was concentrated and exchanged to Tris buffer 50 mM Tris-HCl pH 7.5 and 200 mM NaCl . The protein solution was incubated with a 1 ml bed volume of Ni-NTA beads to extract the cleaved 6xHis-tag. The resulting fraction was concentrated and then further purified by FPLC using a HiLoad 16/600 Superdex 200 preparatory-grade column. Crystallization of protein–cyclic peptide complexes MCL1 with cyclic peptide MCL1 18.5 mg ml<sup −1</sup and macrocycle MCB D2 were mixed in 1:2 molar ratio and incubated for 30 min at room temperature. Upon addition of the MCB D2 to the protein, we observed some precipitation. This precipitant was removed by centrifugation before crystallographic screening. Crystallization experiments for the MCL1–MCB D2 complex were conducted using the sitting-drop vapor diffusion method. Initial crystallization trials were set up in 200 nl drops using 96-well crystallization plates. Crystal drops were imaged using the UVEX crystal plate hotel system by JANSi. Diffraction-quality crystals for the complex appeared in 0.2 M sodium chloride, 0.1 M Bis–Tris pH 6.5 and 25% w/v polyethylene glycol 3350 Hampton Research in 2 weeks. GABARAP and GABARAPL1 with cyclic peptides Cyclic peptides GAB D8 and GAB D23 were dissolved in 10 mM Tris-HCl and 150 mM NaCl pH 7.0 and each mixed with both GABARAP and GABARAPL1, targeting a peptide-to-protein molar ratio of 3:2. After incubation for 10 min at room temperature, any insoluble components were removed by centrifugation 10 min at 20,000 g and 4 °C . The protein–peptide complexes were concentrated using Amicon Ultra-0.5 centrifugal filter units with a 3 kDa cutoff Merck until a final protein concentration of 6–8 mg ml<sup −1</sup GABARAPL1–GAB D8 or 13–15 mg ml<sup −1</sup GABARAP–GAB D23 was reached. Samples were once again cleared of particles by centrifugation 30 min at 20,000 g and 4 °C before application in crystallization experiments. Search for crystallization conditions was performed by the sitting-drop vapor diffusion method using robotic systems Freedom Evo Tecan and Mosquito LCP SPT Labtech with commercially available screening sets. Experiments were set up by combining 200 nl of protein–peptide complex with 100 nl for GABARAPL1–GAB D8 or 200 nl for GABARAP–GAB D23 of reservoir solution and plates were incubated at 20 °C. Crystals appeared for a number of conditions, which were subjected to optimization as appropriate. Diffraction-quality samples used for X-ray structure determination developed with reservoir solutions containing 0.17 M ammonium sulfate, 25.5% w/v PEG 4000 and 15% v/v glycerol for GABARAPL1–GAB D8 and 0.1 M MES pH 5.0 and 30% w/v PEG 6000 in the case of GABARAP–GAB D23. Diffraction data https://doi.esrf.fr/10.15151/ESRF-DC-1966164200 https://doi.esrf.fr/10.15151/ESRF-DC-1966164200 and https://doi.esrf.fr/10.15151/ESRF-DC-1979522808 https://doi.esrf.fr/10.15151/ESRF-DC-1979522808 for GABARAPL1–GAB D8 and GABARAP–GAB D23, respectively were collected at 100 K on beamline BM07 of the European Synchrotron Radiation Facility ESRF tuned to an X-ray wavelength of 0.9795 Å, using a Pilatus 6M detector DECTRIS . Data processing was carried out with XDS and XSCALE<sup 43 https://www.nature.com/articles/s41589-025-01929-w ref-CR43 </sup and included reflections up to a diffraction limit of 1.5 Å for GABARAP–GAB D23 and 2.5 Å for GABARAPL1–GAB D8. The GABARAP–GAB D23 structure featuring space group C2 was determined by molecular replacement MR using MOLREP<sup 44 https://www.nature.com/articles/s41589-025-01929-w ref-CR44 </sup with the structure of GABARAP from its K1 peptide complex PDB 3D32 https://doi.org/10.2210/pdb3D32/pdb <sup 30 https://www.nature.com/articles/s41589-025-01929-w ref-CR30 </sup as a template. For the GABARAPL1–GAB D8 complex, initial evaluation suggested tetragonal symmetry but with strong indications of twinning. Data integration in maximal translationengleiche subgroups followed by MR search using MoRDa<sup 45 https://www.nature.com/articles/s41589-025-01929-w ref-CR45 </sup revealed P2<sub 1</sub 2<sub 1</sub 2<sub 1</sub as the true space group, with near-perfect pseudomerohedral twinning accounting for apparent Laue group 4/mmm. To avoid bias in cross-validation, this pseudosymmetry of the data was explicitly accounted for in flag assignment. The solution obtained for GABARAPL1–GAB D8 was subjected to a round of automated rebuilding in phenix.autobuild<sup 46 https://www.nature.com/articles/s41589-025-01929-w ref-CR46 </sup . In either case, model refinement was performed with phenix.refine<sup 47 https://www.nature.com/articles/s41589-025-01929-w ref-CR47 </sup , alternating with interactive rebuilding in Coot<sup 48 https://www.nature.com/articles/s41589-025-01929-w ref-CR48 </sup , which included stepwise introduction of cyclic peptides GAB D8 and GAB D23. According to validation using MolProbity<sup 49 https://www.nature.com/articles/s41589-025-01929-w ref-CR49 </sup and the wwPDB validation system https://validate-rcsb-2.wwpdb.org/ https://validate-rcsb-2.wwpdb.org/ , both models featured good geometry. Detailed statistics of data collection and refinement can be found in Supplementary Table 6 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . RbtA with cyclic peptide and apo RbtA RbtA 10 mg ml<sup −1</sup and RBB D10 were mixed in a 1:5 molar ratio and incubated for 30 min at room temperature. Initial crystallization trials were set up in 200 nl drops using 96-well crystallization plates and the experiments were conducted by the sitting-drop vapor diffusion method. Crystal drops were imaged using the UVEX crystal plate hotel system by JANSi. Diffraction-quality crystals for the RbtA–RBB D10 complex appeared in 0.2 M lithium sulfate, 0.1 M Tris pH 8.5 and 40% v/v PEG 400 JCSG Plus, Hampton Research . Additionally, we soaked the crystals in 22.32 mg ml<sup −1</sup RBB D10 for 5 min before flash-freezing. Crystals for RbtA alone 18.7 mg ml<sup −1</sup were grown in 0.1 M Bis–Tris pH 6.5 and 20% v/v PEG 5,000 MME SG1, Molecular Dimensions . All crystals were flash-cooled in liquid nitrogen before shipping to the synchrotron for data collection. Diffraction data were collected at the NSLS2 beamline AMX/FMX 17-ID-1/17-ID-2 . X-ray intensities and data reduction were evaluated and integrated by XDS<sup 43 https://www.nature.com/articles/s41589-025-01929-w ref-CR43 </sup and merged and scaled by Pointless and Aimless in the CCP4i2 program suite<sup 50 https://www.nature.com/articles/s41589-025-01929-w ref-CR50 </sup . The X-ray crystal structure was determined by MR using the designed model for phasing by Phaser<sup 51 https://www.nature.com/articles/s41589-025-01929-w ref-CR51 </sup . Next, the structure obtained from the MR was improved and refined by Phenix<sup 47 https://www.nature.com/articles/s41589-025-01929-w ref-CR47 </sup . Model building was performed by Coot<sup 48 https://www.nature.com/articles/s41589-025-01929-w ref-CR48 </sup in between the refinement cycles. The final model was evaluated by MolProbity<sup 49 https://www.nature.com/articles/s41589-025-01929-w ref-CR49 </sup . Data collection and refinement statistics are reported in Supplementary Table 5 https://www.nature.com/articles/s41589-025-01929-w MOESM1 . SPR SPR experiments were performed using a Cytiva Biacore 8K in HBS-EP+ buffer from Cytiva. Measurements were obtained by immobilization of biotinylated target protein using the biotin capture kit from Cytiva. Binding screens were performed by single-cycle kinetics experiments using the standard protocol in the Biacore 8K control software at 30 µl min<sup −1</sup with serial injections of 10 nM, 100 nM, 1 µM, 10 µM and 100 µM, an association time of 60 s and a dissociation time of 120 s. For MCL1 designs, a dissociation time of 150 s was used. To evaluate the affinity of successful designs, a nine-point single-cycle kinetics experiment was performed with an association time of 90 s and dissociation time of 300 s. The dilution series for MCB D2 was twofold starting at 20 µM, that for MDB D8 was fivefold starting at 50 µM, and those for GAB D8, GAB D23 and RBB D10 were fivefold starting at 20 µM. Reported measurements were analyzed using Biacore Insight evaluation software; sensorgrams were double-referenced and fit with a 1:1 binding kinetics fit model. AlphaScreen assay We used the AlphaScreen assay as described by Leveille et al.<sup 52 https://www.nature.com/articles/s41589-025-01929-w ref-CR52 </sup to measure inhibition of the GABARAP–K1 interaction by the computationally designed macrocycles. K1 is a previously described GABARAP binder with a K <sub d</sub of 10 nM ref. <sup 27 https://www.nature.com/articles/s41589-025-01929-w ref-CR27 </sup . Biotin-labeled peptide K1 was used at a final concentration of 10 nM and incubated with 10 nM final concentration of 6xHis–GABARAP in a final reaction volume of 50 μl. Computationally designed inhibitor peptides were serially diluted with 1:3 dilutions using the highest final concentration of 50 μM and added to the reaction mixture. The buffer used was 25 mM HEPES pH 7.3, 150 mM NaCl, 0.01% Tween, 1 mg ml<sup −1</sup BSA and 0.5% DMSO. The plate was covered in foil, centrifuged at 1,500 rpm for 2 min and incubated for 150 min at room temperature with shaking. Then, 20 μg ml<sup −1</sup final concentration of the streptavidin donor beads and nickel chelate acceptor beads were added in the dark before incubating for another 45 min. Data were collected on a Tecan plate reader using excitation at 680 nm and emission at 520–620 nm. Data were normalized to 0% buffer only and 100% protein and tracer peptide, no inhibitor controls. IC<sub 50</sub values were obtained from curve fits using GraphPad Prism 9 software, using the equation $Y=\frac{100}{ 1+{ \frac{X}{{{IC}} {50}} }^{h} }$, where X is the concentration of inhibitor and h is the Hill coefficient. At least three independent replicates were used to calculate the average IC<sub 50</sub and the s.e.m. Statistics and reproducibility No statistical method was used to predetermine sample size. One trial from the AlphaScreen that was used to determine the IC<sub 50</sub of GAB D8 was repeated and the repeated value is what was used. All data are included in the Source Data https://www.nature.com/articles/s41589-025-01929-w Sec25 . The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary https://www.nature.com/articles/s41589-025-01929-w MOESM2 linked to this article. Data availability The design models and sequences are available in Supplementary Information https://www.nature.com/articles/s41589-025-01929-w Sec24 . Crystal structures of MCB D2 bound to MCL1, GAB D8 bound to GABARAPL1, GAB D23 bound to GABARAP, RBB D10 bound to RbtA and apo RbtA were deposited to the PDB under accession codes 9CDT https://doi.org/10.2210/pdb9CDT/pdb , 9HGC https://doi.org/10.2210/pdb9HGC/pdb , 9HGD https://doi.org/10.2210/pdb9HGD/pdb , 9CDU https://doi.org/10.2210/pdb9CDU/pdb and 9CDV https://doi.org/10.2210/pdb9CDV/pdb , respectively. Source data https://www.nature.com/articles/s41589-025-01929-w Sec25 are provided with this paper. Code availability The code and scripts for running RFpeptides are available from Zenodo https://doi.org/10.5281/zenodo.15264344 https://doi.org/10.5281/zenodo.15264344 <sup 53 https://www.nature.com/articles/s41589-025-01929-w ref-CR53 </sup . The code and scripts for RFpeptides are also available from RFdiffusion GitHub repository https://github.com/RosettaCommons/RFdiffusion https://github.com/RosettaCommons/RFdiffusion . References 1. Vinogradov, A. A., Yin, Y. & Suga, H. Macrocyclic peptides as drug candidates: recent progress and remaining challenges. J. Am. Chem. Soc. 141 , 4167–4181 2019 . 2. Muttenthaler, M., King, G. F., Adams, D. J. & Alewood, P. F. Trends in peptide drug discovery. Nat. Rev. Drug Discov. 20 , 309–325 2021 . 3. Tsomaia, N. Peptide therapeutics: targeting the undruggable space. Eur. J. Med. Chem. 94 , 459–470 2015 . 4. Atanasov, A. G., Zotchev, S. B. & Dirsch, V. M. International Natural Product Sciences Taskforce & Supuran, C. T. Natural products in drug discovery: advances and opportunities. Nat. Rev. Drug Discov. 20 , 200–216 2021 . 5. Bhardwaj, G. et al. Accurate de novo design of membrane-traversing macrocycles. Cell 185 , 3520–3532 2022 . 6. Bhardwaj, G. et al. Accurate de novo design of hyperstable constrained peptides. Nature 538 , 329–335 2016 . 7. Hosseinzadeh, P. et al. Comprehensive computational design of ordered peptide macrocycles. Science 358 , 1461–1466 2017 . 8. Mulligan, V. K. et al. Computationally designed peptide macrocycle inhibitors of New Delhi metallo-β-lactamase 1. Proc. Natl Acad. Sci. USA 118 , e2012800118 2021 . 9. Hosseinzadeh, P. et al. Anchor extension: a structure-guided approach to design cyclic peptides targeting enzyme active sites. Nat. Commun. 12 , 3384 2021 . 10. Muratspahić, E. et al. Design and structural validation of peptide–drug conjugate ligands of the κ-opioid receptor. Nat. Commun. 14 , 8064 2023 . 11. Rettie, S. A. et al. Cyclic peptide structure prediction and design using AlphaFold2. Nat. Commun. 16 , 4730 2025 . 12. Grambow, C. A., Weir, H., Cunningham, C. N., Biancalani, T. & Chuang, K. V. CREMP: conformer–rotamer ensembles of macrocyclic peptides for machine learning. Sci. Data 11 , 859 2024 . 13. Zhang, C. et al. HighFold: accurately predicting structures of cyclic peptides and complexes with head-to-tail and disulfide bridge constraints. Brief. Bioinform 25 , bbae215 2024 . 14. Brixi, G. et al. SaLT&PepPr is an interface-predicting language model for designing peptide-guided protein degraders. Commun. Biol. 6 , 1081 2023 . 15. Xie, X., Valiente, P. A., Kim, J. & Kim, P. M. HelixDiff, a score-based diffusion model for generating all-atom α-helical structures. ACS Cent. Sci. 10 , 1001–1011 2024 . 16. Li, Q., Vlachos, E. N. & Bryant, P. Design of linear and cyclic peptide binders of different lengths from protein sequence information. Preprint at bioRxiv https://doi.org/10.1101/2024.06.20.599739 https://doi.org/10.1101/2024.06.20.599739 2024 . 17. Watson, J. L. et al. De novo design of protein structure and function with RFdiffusion. Nature 620 , 1089–1100 2023 . 18. Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373 , 871–876 2021 . 19. Vázquez Torres, S. et al. De novo design of high-affinity binders of bioactive helical peptides. Nature 626 , 435–442 2024 . 20. Krishna, R. et al. Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Science 384 , eadl2528 2024 . 21. Dauparas, J. et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science 378 , 49–56 2022 . 22. Wicky, B. I. M. et al. Hallucinating symmetric protein assemblies. Science 378 , 56–61 2022 . 23. Widden, H. & Placzek, W. J. The multiple mechanisms of MCL1 in the regulation of cell fate. Commun. Biol. 4 , 1029 2021 . 24. Leman, J. K. et al. Macromolecular modeling and design in Rosetta: recent methods and frameworks. Nat. Methods 17 , 665–680 2020 . 25. Kussie, P. H. et al. Structure of the MDM2 oncoprotein bound to the p53 tumor suppressor transactivation domain. Science 274 , 948–953 1996 . 26. Szalai, P. et al. Autophagic bulk sequestration of cytosolic cargo is independent of LC3, but requires GABARAPs. Exp. Cell Res. 333 , 21–38 2015 . 27. Brown, H. et al. Structure-based design of stapled peptides that bind GABARAP and inhibit autophagy. J. Am. Chem. Soc. 144 , 14687–14697 2022 . 28. Ji, C. H. et al. The AUTOTAC chemical biology platform for targeted protein degradation via the autophagy–lysosome system. Nat. Commun. 13 , 904 2022 . 29. Popelka, H. & Klionsky, D. J. Analysis of the native conformation of the LIR/AIM motif in the Atg8/LC3/GABARAP-binding proteins. Autophagy 11 , 2153–2159 2015 . 30. Weiergräber, O. H. et al. Ligand binding mode of GABA <sub A</sub receptor-associated protein. J. Mol. Biol. 381 , 1320–1331 2008 . 31. van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. J. Mach. Learn. Res. 9 , 2579–2605 2008 . 32. Pedregosa, F. et al. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12 , 2825–2830 2011 . 33. Zhang, Y. & Skolnick, J. TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic Acids Res 33 , 2302–2309 2005 . 34. Dauparas, J. et al. Atomic context-conditioned protein sequence design using LigandMPNN. Nat. Methods 22 , 717–723 2025 . 35. Sternberg, A. MaxCluster: a tool for protein structure comparison and clustering. http://www.sbg.bio.ic.ac.uk/maxcluster/index.html http://www.sbg.bio.ic.ac.uk/maxcluster/index.html accessed 1 November 2024 . 36. Siew, N., Elofsson, A., Rychlewski, L. & Fischer, D. MaxSub: an automated measure for the assessment of protein structure prediction quality. Bioinformatics 16 , 776–785 2000 . 37. Fire, E., Gullá, S. V., Grant, R. A. & Keating, A. E. Mcl‐1–Bim complexes accommodate surprising point mutations via minor structural changes. Protein Sci. 19 , 507–519 2010 . 38. Anil, B., Riedinger, C., Endicott, J. A. & Noble, M. E. M. The structure of an MDM2–Nutlin-3a complex solved by the use of a validated MDM2 surface-entropy reduction mutant. Acta Crystallogr. D 69 , 1358–1366 2013 . 39. Sumida, K. H. et al. Improving protein expression stability and function with ProteinMPNN. J. Am. Chem. Soc. 146 , 2054–2061 2024 . 40. Bhandari, B. K., Gardner, P. P. & Lim, C. S. Solubility-Weighted Index: fast and accurate prediction of protein solubility. Bioinformatics 36 , 4691–4698 2020 . 41. Davis, G. D., Elisee, C., Newham, D. M. & Harrison, R. G. New fusion protein systems designed to give soluble expression in Escherichia coli . Biotechnol. Bioeng 65 , 382–388 1999 . 42. Dang, B. et al. SNAC-tag for sequence-specific chemical protein cleavage. Nat. Methods 16 , 319–322 2019 . 43. Kabsch, W. XDS. Acta Crystallogr. D 66 , 125–132 2010 . 44. Vagin, A. & Teplyakov, A. Molecular replacement with MOLREP. Acta Crystallogr. D 66 , 22–25 2010 . 45. Vagin, A. & Lebedev, A. MoRDa, an automatic molecular replacement pipeline. Acta Crystallogr. A Found. Adv. 71 , s19 2015 . 46. Terwilliger, T. C. et al. Iterative model building structure refinement and density modification with the PHENIX AutoBuild wizard. Acta Crystallogr. 64 , 61–69 2008 . 47. Liebschner, D. et al. Macromolecular structure determination using X-rays neutrons and electrons: recent developments in Phenix. Acta Crystallogr. D Struct. Biol. 75 , 861–877 2019 . 48. Emsley, P., Lohkamp, B., Scott, W. G. & Cowtan, K. Features and development of Coot. Acta Crystallogr. D 66 , 486–501 2010 . 49. Williams, C. J. et al. MolProbity: more and better reference data for improved all‐atom structure validation. Protein Sci. 27 , 293–315 2018 . 50. Winn, M. D. et al. Overview of the CCP 4 suite and current developments. Acta Crystallogr. D 67 , 235–242 2011 . 51. McCoy, A. J. et al. Read Phaser crystallographic software. J. Appl. Crystallogr. 40 , 658–674 2007 . 52. Leveille, A. N. et al. Exploring arylidene–indolinone ligands of autophagy proteins LC3B and GABARAP. ACS Med. Chem. Lett. 16 , 271–277 2025 . 53. Rettie, S. et al. Accurate de novo design of high-affinity protein-binding macrocycles using deep learning. Zenodo https://doi.org/10.5281/zenodo.15264344 https://doi.org/10.5281/zenodo.15264344 2025 . 54. Hunter, J. D. Matplotlib: a 2D graphics environment. Comput. Sci. Eng 9 , 90–95 2007 . 55. Waskom, M. seaborn: statistical data visualization. J. Open Source Softw 6 , 3021 2021 . 56. Pettersen, E. F. et al. UCSF ChimeraX: structure visualization for researchers, educators, and developers. Protein Sci 30 , 70–82 2021 . Acknowledgements We thank L. Stuart, L. Stewart, K. van Wormer, L. Goldschmidt, M. Kennedy, I. Haydon, J. Woodward, X. Li, Z. Taylor, H. Osterstock, G. Zhou, G. Gökçe, J. Palmer, K. Lindenauer, K. Campbell, M. Gloegl, R. Ragotte and M. Sadilek for their helpful feedback, guidance and support. We also thank the IPD Peptide, Crystallography and Biologics, Vaccines and Process Development core labs and the University of Washington Chemistry MS Facility for providing instrumentation support and expertise. This work was supported by funds from the DARPA Harnessing Enzymatic Activity for Lifesaving Remedies program HR001120S0052 contract HR0011-21-2-0012 to G.B., D.B., J.D.M. and M.L. , the Defense Threat Reduction Agency HDTRA1-19-1-0003 to D.B., G.B. and S.A.R. , the National Institutes of Health NIH 5R21AI178088-02 to G.B. and S.A.R. , the Howard Hughes Medical Institute HHMI Emerging Pathogens Initiative to J.D.M., G.B. and V.A. , startup funds from the University of Washington’s Department of Medicinal Chemistry and Institute for Protein Design to G.B. , the Audacious Project to G.B., A.K.B. and A.K. , the C19 HHMI Initiative grant to A.K.B. and A.K. , NIH R35-GM148407 to J.A.K. and A.N.L. , the European Union’s Horizon Europe research and innovation program under the Marie Skłodowska-Curie 101059124 to Y.F.B. , NIH R01-R0AI160052 to A.K.B. and A.K. , Deutsche Forschungsgemeinschaft project ID 267205415-SFB 1208 to D.W. and the Bill and Melinda Gates Foundation GR047983 to D.J. . All plots in this paper were generated using matplotlib or seaborn<sup 54 https://www.nature.com/articles/s41589-025-01929-w ref-CR54 , 55 https://www.nature.com/articles/s41589-025-01929-w ref-CR55 </sup . Peptide structures were rendered using ChimeraX 1.9 ref. <sup 56 https://www.nature.com/articles/s41589-025-01929-w ref-CR56 </sup or PyMOL 2.5.4. Data were analyzed and plotted with Pandas versions 1.4.3 and plotted using matplotlib version 3.7.0 and seaborn version 0.12.2. All figures were created with BioRender.com http://BioRender.com . This research used resources FMX/AMX of the National Synchrotron Light Source II, a US Department of Energy DOE Office of Science User Facility operated for the DOE Office of Science by Brookhaven National Laboratory under contract No. DE–SC0012704. The Center for BioMolecular Structure is primarily supported by the NIH National Institute of General Medical Sciences through a Center Core P30 Grant P30GM133893 and by the DOE Office of Biological and Environmental Research KP1607011 . This publication resulted from data collected using the beamtime obtained through Northeastern Collaborative Access Team BAG proposal no. 311950. Lastly, we would like to thank the staff of the ESRF and European Molecular Biology Laboratory for assistance and support in using beamline BM07 under proposal number MX–2587. Ethics declarations Competing interests D.W. is a cofounder of Priavoid and Attyloid. D.B. and G.B. are cofounders, advisors and shareholders of Vilya. The other authors declare no competing interests. Peer review Peer review information Nature Chemical Biology thanks Hyun Ho Park, Francesca Peccati and the other, anonymous reviewer s for their contribution to the peer review of this work. Additional information Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information Supplementary Information download PDF https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41589-025-01929-w/MediaObjects/41589 2025 1929 MOESM1 ESM.pdf Supplementary Sections 1–5, Figs. 1–59 and Tables 1–10. Source data Source Data Fig. 1 download XLSX https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41589-025-01929-w/MediaObjects/41589 2025 1929 MOESM3 ESM.xlsx Source data used in plots. Source Data Fig. 2 download XLSX https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41589-025-01929-w/MediaObjects/41589 2025 1929 MOESM4 ESM.xlsx Source data used in plots. Source Data Fig. 3 download XLSX https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41589-025-01929-w/MediaObjects/41589 2025 1929 MOESM5 ESM.xlsx Source data used in plots. Source Data Fig. 4 download XLSX https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41589-025-01929-w/MediaObjects/41589 2025 1929 MOESM6 ESM.xlsx Source data used in plots. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author s and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/licenses/by/4.0/ . About this article Cite this article Rettie, S.A., Juergens, D., Adebomi, V. et al. Accurate de novo design of high-affinity protein-binding macrocycles using deep learning. Nat Chem Biol 21 , 1948–1956 2025 . https://doi.org/10.1038/s41589-025-01929-w - Received: - Accepted: - Published: - Version of record: - Issue date: - DOI: https://doi.org/10.1038/s41589-025-01929-w