Cyclic peptide structure prediction for sequences carrying unnatural amino acids, adding atom-level features and cyclization-aware position encoding.
No providers recorded yet. Browse all providers
Cyclic peptides occupy a useful middle ground in drug discovery: large enough to bury the flat interfaces small molecules cannot address, small enough to synthesize chemically, and — once cyclized — far more protease-resistant than linear peptides. Medicinal chemists push those advantages further with unnatural amino acids: D-amino acids, N-methylated backbones, and non-canonical side chains that improve permeability, metabolic stability, and affinity. That combination is what structure prediction cannot handle. AlphaFold-Multimer and its relatives assume the twenty canonical residues and an implicitly linear chain; a head-to-tail cyclized peptide studded with non-standard residues violates both.
HighFold2 closes that gap. Developed by Cheng Zhu, Sen Cao, and colleagues in Hongliang Duan's group at Macao Polytechnic University with collaborators at Zhejiang University of Technology, Shenzhen Highslab Therapeutics, and Zhejiang University, it rebuilds three parts of the AlphaFold-Multimer stack: the residue geometry definitions, the feature representation, and the positional encoding. It predicts both cyclic peptide monomers and their complexes with protein targets, and was published in Briefings in Bioinformatics after a January 2025 bioRxiv preprint.
The training strategy works around a data problem. Solved cyclic peptide structures with unnatural amino acids are far too scarce to fine-tune on, so the model is fine-tuned entirely on linear peptides with unnatural residues and cyclization is introduced at inference time by rewriting the relative position encoding — making cyclic prediction a zero-shot capability rather than a learned one. HighFold2 sits alongside HighPlay, the same group's cyclic peptide binder design method, and complements all-atom successors such as AlphaFold 3 and RoseTTAFold All-Atom.
Atom features come from converting chains shorter than 50 residues to SMILES with RDKit. Element identity is one-hot encoded — ten common elements individually, rare elements sharing a bucket — passed through eight-head attention and an MLP projecting to 21 dimensions, then pooled from atoms to residues and added to AlphaFold-Multimer's residue-type one-hot matrix before the Evoformer. Bond types are encoded as an N x N x 5 tensor, reduced by a linear layer, pooled the same way, and added to the pair representation.
Training data came from the ModPep dataset: 501 linear peptide entries reduced to 419 unique PDB structures, then filtered by chain redundancy and by discarding unnatural residues appearing once, leaving 382 samples covering 23 distinct unnatural amino acids, split 7.5:1.5:1 into train, validation, and test. Features were generated with ColabFold, with templates excluded to prevent leakage. The cyclic test set is independent: 34 cPEPmatch structures restricted to the same 23 unnatural residues and to head-to-tail or disulfide cyclization, never seen in training. There the median Cα RMSD is 1.891 Å, median all-atom RMSD 2.872 Å, and median RMSD over unnatural-residue atoms 2.579 Å. Ablations removing the modified position encoding, the atomic-scale features, or the ensemble each degrade accuracy, and predicted peptide pLDDT correlates with observed RMSD on both test sets, giving a per-prediction confidence signal.
The target users are peptide therapeutics teams, where cyclic scaffolds bearing unnatural residues are standard practice and structural insight has been computationally unavailable. Typical uses are triaging synthesis candidates by predicted conformation, modeling how a macrocycle presents its side chains against a protein interface, and generating structures for docking or molecular dynamics. The code is on GitHub under GPL-3.0 with a ColabFold-based prediction script, while the fine-tuned parameters come from an external OneDrive link rather than a model hub.
HighFold2 shows that AlphaFold's residue-centric geometry can be extended to chemistry it was never trained on, without retraining from scratch, by editing the three places where the canonical-residue assumption is encoded. The zero-shot route to cyclization — train on linear, infer on cyclic — transfers to other prediction problems whose target class has almost no solved structures. The limits are scale and scope: 382 training structures over 23 unnatural residue types and a 34-structure cyclic test set are small by structure-prediction standards, cyclization is restricted to head-to-tail and disulfide forms, and residues outside the trained set are unsupported. Hosting weights on OneDrive rather than an archive also leaves long-term availability on a single link.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.