Rettie et al. recently introduced AfCycDesign, an AlphaFold2-based method that accurately designs and predicts structures of cyclic peptides.
AfCycDesign is capable of generating protein sequences that fold into desired target structure by inverse folding of given backbone structures. The method can also hallucinate, from scratch, novel macrocyclic peptides.
Although there have been growing interests in the application of cyclic peptides as therapeutic agents, the development of deep learning models for their design has been faced with challenges resulting from the limited number of experimental structures for training.
One approach is to augment available structures with models generated by existing physics-based design methods and structures sampled by molecular dynamics simulation.
However, the accuracy of the models trained on such data will be contingent on the accuracy of the methods that produced the data.
Inspired by the remarkable success of AlphaFold2 (AF2) and RosettaFold in 3D protein structure prediction from 1D sequences, Rettie et al. repurposed the AF2 neural network for the structure prediction and design of cyclic peptides from single sequences, reducing the reliance on evolutionary and structural templates.
The information encoded by the AF2 model was adapted for the design of macrocyclic peptide facilitated by suitable cyclic constraints and a relative positional encoding invariant to cyclic permutation of the input sequences.
In a structure prediction benchmark assessment comprising 80 PDB test structures:
🎯 Structures predicted by AfCycDesign exhibited median pLDDT and RMSD of 0.88 and 1.13 Å.
🎯 50 predictions out of the 80 test cases showed good confidence, pLDDT > 0.7, and RMSD < 1.5 Å.
🎯 39 out of 49 test cases where AfCycDesign predicted structures with high confidence (pLDDT > 0.85) had RMSD < 1.5 Å.
🎯 Addition of MSAs during predictions further improved the accuracy, with 58 out of 80 test cases having RMSD < 1.5 Å and pLDDT > 0.7.
To evaluate AfCycDesign for de novo design, cyclic peptides with 7-13 residues were sampled to generate 10,000 unique sequences with high confidence.
🔥 A small subset of the designs was synthesized and crystallized, and the X-ray crystal structures of seven of the sequences having diverse sizes and topologies accurately matched the design structural models with RMSD < 1.0 Å.
🧮 AfCycDesign also outperformed Rosetta on macrocyclic peptide sequence design tasks. Of 3,274 unique structural clusters from large-scale backbone sampling runs, 1145 clusters had pLDDTs > 0.9 for sequences designed by AfCycDesign, while only 63 clusters had pLDDTs > 0.9 for Rosetta-designed sequences.
References
Paper: Cyclic peptide structure prediction and design using AlphaFold
Code: af_cyc_design
