https://www.nature.com/articles/s41467-025-59940-7

Cyclic peptide structure prediction and design using AlphaFold2

Key figures

• [Figure 1]: Establishes that cyclic relative positional encoding enables AlphaFold2-based prediction of diverse native cyclic peptides, with 58/80 PDB cases predicted at RMSD < 1.5 Å and pLDDT > 0.7.
  
• [Figure 2]: Shows that AfCycDesign can redesign a 13-residue cyclic peptide backbone into a sequence whose X-ray structure matches the model at 0.3 Å Cα RMSD.
  
• [Figure 3]: Validates de novo hallucination of 7-10 residue cyclic peptides by X-ray crystallography, with selected designs matching models at 0.3-1.0 Å Cα RMSD.
  
• [Figure 4]: Extends hallucination to larger 11-13 residue macrocycles without extra disulfide crosslinks, with X-ray structures matching models at 0.3-0.8 Å Cα RMSD.
  
• [Figure 5]: Demonstrates functionalization of hallucinated scaffolds into MDM2 inhibitors, including RMG_14 at 338.4 nM IC50 and an optimized variant with 294 nM SPR affinity.
  
• [Figure 6]: Shows hot-loop grafting onto hallucinated scaffolds to generate Keap1 inhibitors, with KC4 reaching 87.7 nM IC50 and outperforming the linear Nrf2 peptide.
  

1) Thesis (one sentence)

To address the lack of rapid and accurate deep-learning methods for cyclic peptide structure prediction, redesign, and functional scaffold generation, in head-to-tail cyclic peptides and motif-grafted macrocycle binders to MDM2 and Keap1, AfCycDesign causes accurate structure prediction, high-confidence scaffold hallucination, and nanomolar target inhibition by imposing cyclic relative positional encoding in AlphaFold2/ColabDesign and coupling AF confidence losses with scaffold grafting, ProteinMPNN/Rosetta redesign, and biophysical filtering, supported by PDB benchmark prediction, X-ray crystallography, AlphaLISA, SPR, and fluorescence polarization.

2) Evidence card (three bullets only)

• Strongest result: (Figure 3; Figure 4; Figure 6) Eight redesigned or hallucinated cyclic peptides were structurally validated by X-ray crystallography with Cα RMSD values below 1.0 Å, and grafted Keap1 binders reached nanomolar IC50 values, with KC4 at 87.7 nM.
  
• Method enabler: (Figure 1; Figure 2; research type + tools) Computational protein design using cyclic relative positional encoding in AlphaFold2/ColabDesign enabled cyclic peptide prediction, fixed-backbone redesign, and hallucination, then Rosetta energy landscapes, ProteinMPNN/Rosetta redesign, X-ray crystallography, AlphaLISA, SPR, and fluorescence polarization validated structure and function.
  
• Critical limitation: (Figure 5; Figure 6) Protein-binding designs were generated by grafting known p53 or Nrf2 interface motifs onto hallucinated scaffolds rather than by fully motif-free target-surface design, so success still depends on having a compatible starting interaction motif.
  

Optional

Quote bank (2–4 short excerpts)

• Quote 1: “Here, we introduce AfCycDesign, a deep learning approach for accurate structure prediction, sequence redesign, and de novo hallucination of cyclic peptides.” (Abstract, page 1)
  
• Quote 2: “Using AfCycDesign, we identified over 10,000 structurally-diverse designs” (Abstract, page 1)
  
• Quote 3: “Information encoded in the AlphaFold2 network would be adequate to accurately predict and design macrocycles” (Introduction, page 2)
  
• Quote 4: “We did not measure protease or serum stability of reported designs in this work” (Discussion, page 11)
  

Key comparisons (1–3 lines)

• Compared to: Rosetta KIC-based macrocycle design, natural product discovery, and display-library screening.
  
• Win: Produces structurally diverse, atomically accurate cyclic peptide scaffolds with far less brute-force sampling and demonstrates motif-grafted nanomolar inhibitors from small test sets.
  
• Tradeoff: Functional binder design still relies on known motifs, canonical amino acids, in vitro assays, and crystallographic snapshots rather than direct evidence of intracellular stability, permeability, or unseeded target engagement.
  

Methods I might copy (protocol hooks)

• Construct design / Models: Encode N-to-C cyclization with a custom N x N cyclic offset matrix; for PDB benchmark prediction use ColabDesign with MSA, 6 recycles, random masking, and all 5 models; evaluate highest-pLDDT models by backbone heavy-atom RMSD against NMR ensembles; for fixed-backbone design use distogram categorical cross-entropy; for hallucination use 1 – pLDDT + PAE/31 + contact/2 with binary contact loss, cutoff 21.6875, num = peptide length, seqsep = 0; use random Gumbel initialization, 50 softmax priming steps at temperature 1.0, then a 50/50/10 stage schedule.
  
• Conditions / Instruments: Manual Fmoc synthesis used 0.2 mmol scale, 300 mg 2-Cl-Trt resin, 20% piperidine in DMF for 20 min deprotections, 5 eq Fmoc amino acid + 5 eq PyAOP + 10 eq DIEA for 20 min couplings, 2% TFA in DCM cleavage, 2 eq PyAOP + 3 eq DIEA overnight cyclization in 50 mL DCM, and 92.5:2.5:2.5:2.5 TFA:water:TIPS:DODT deprotection for 3 h; purified on Agilent Infinity 1260 HPLC with 1% per min gradient on Agilent ZORBAX 300SB-C18, 5 µm, 9.4 x 250 mm column.
  
• Readout / Analysis: For MDM2 competition, use PerkinElmer HTRF Human MDM2 Binding kit, 50 µM peptide and 5% DMSO for single-point screening, 10 mM DMSO peptide stocks and 2% assay DMSO for IC50 determination, and GraphPad Prism 10; for Keap1 fluorescence polarization, dilute samples from 100 µM to 1.5 nM, keep final DMSO below 1%, mix 5 µL sample with 0.5 µL of 1 µM Nrf2 peptide and 20 µL of 15 ng/µL Keap1, incubate 30 min at room temperature, read on BioTek Synergy Neo2 at 485 nm excitation and 528 nm emission with G-factor 0.87, and fit IC50 in GraphPad Prism 8 using a four-parameter logistic model.
  

Open questions / Theoretical implications (2–5 bullets)

• Can cyclic relative positional encoding be extended from monomeric macrocycles to ligand-conditioned ternary assemblies?
  
• Can fully motif-free AfCycDesign binder hallucination replace p53/Nrf2 hot-loop grafting for targets with no known binding motif?
  
• Will these scaffolds remain folded, soluble, protease-stable, and cell-permeable outside crystallography and purified in vitro assays?
  
• Can all-atom models incorporate noncanonical amino acids, side-chain crosslinks, and alternative cyclization chemistries directly into the hallucination objective?
  
• Can pLDDT, PAE, Rosetta Pnear, and interface scores be calibrated prospectively to predict functional binding rather than structural self-consistency alone?