Loading…
Cyclic peptide structure prediction and de novo macrocycle design, by wrapping a frozen structure predictor's positional encoding into a ring.
No providers recorded yet. Browse all providers
A structure predictor's relative positional encoding tells the network how far apart two residues sit along the chain, and for a linear peptide the N- and C-termini are the pair furthest apart. A head-to-tail cyclized macrocycle has no termini at all: those two residues are bonded neighbors. Nothing in the input representation of AlphaFold 2 can express that, and its training set excluded NMR structures and peptides shorter than 16 residues — exactly where cyclic peptides live.
AfCycDesign closes the gap at the input rather than in the weights. It replaces the standard offset with a custom N × N cyclic offset matrix in which sequence separation wraps around, so the terminal residues are one step apart and the encoding is invariant to circular permutation of the sequence. The pretrained network is otherwise untouched — no retraining, no fine-tuning — which matters because the reason a macrocycle-specific network does not exist is the absence of a training set large enough to build one. Stephen Rettie, Gaurav Bhardwaj and colleagues at the Institute for Protein Design posted the method in 2023 and published it in Nature Communications in 2025.
That single change unlocks three tasks the underlying network cannot otherwise perform: predicting the fold of a native cyclic peptide, redesigning the sequence of a macrocyclic backbone, and hallucinating new macrocycles from noise. Applying the offset to only the peptide chain of a two-chain input extends all three to protein–macrocycle complexes, which is how binders get designed.
The protocol is implemented inside the ColabDesign framework and runs from the released AlphaFold 2 parameters; it owns no checkpoint of its own. On 80 NMR cyclic peptide structures from the PDB held out of the network's training data, predictions reached median pLDDT 0.92 and median backbone RMSD 0.8 Å, with 58 of 80 at pLDDT above 0.7 and RMSD under 1.5 Å; of the 55 predicted at pLDDT above 0.85, 80% matched the experimental structure within 1.5 Å. Removing the cyclic offset degrades accuracy in both single-sequence and MSA modes. Prediction takes about two minutes on one GPU, where the Rosetta energy landscape for the same peptide required 28,042 sampled conformations and 120 compute hours.
Design uses categorical cross-entropy between predicted and target distograms for fixed-backbone redesign, and a combination of pLDDT, predicted aligned error and a contact term for hallucination. Sampling 48,000 models at each length from 7 to 13 residues yielded roughly 10,000 distinct torsion-bin clusters for 7mers rising to about 28,000 for the longer rings, and 10,681 high-confidence designs. Sequences from this route scored comparably to ProteinMPNN on the same backbones but succeeded on largely different ones, so the two are complementary.
Cyclic peptides are attractive therapeutics — protease-resistant, conformationally locked, and able to reach protein–protein interfaces that small molecules and antibodies cannot — and the scaffold library plus grafting pipeline turns that into a workflow. Grafting a five-residue p53 motif onto the hallucinated scaffolds and redesigning the rest produced an MDM2 inhibitor with an IC50 of 338 nM, whose 1.7 Å crystal structure matched the design model to 1.0 Å Cα RMSD; five substitutions then improved affinity roughly tenfold. Grafting the Nrf2 EETG hot loop produced Keap1 binders under 200 nM, and 798 of 1,014 catalogued hot loops match a scaffold in the set closely enough to graft.
Eight crystal structures — one redesigned backbone and seven hallucinated macrocycles — all came in under 1.0 Å Cα RMSD of their design models, including 11- to 13-residue rings that earlier physics-based methods could only stabilize with added crosslinks. The wider lesson, that a frozen structure predictor can be given a new capability through an input-side edit, recurs in later cyclic peptide work such as RFpeptides and CycleDesigner. The method handles canonical L-amino acids only, so non-canonical residues arrive afterwards through physics-based tools; validation rests entirely on crystallography, which cannot rule out alternative solution conformations; and with no AfCycDesign-specific checkpoint, its terms of use are those of the AlphaFold 2 parameters and the ColabDesign codebase.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.