Use CyclicMPNN Online
Commercially Available CyclicMPNN No-Code Web Server
CyclicMPNN: Stable Cyclic Peptide Sequence Design with a Fine-Tuned ProteinMPNN
CyclicMPNN is a version of ProteinMPNN fine-tuned for designing amino acid sequences for cyclic peptide backbones built from canonical L-alpha-amino acids. It was developed by Andrew C. Powers, Yanapat Janthana and Parisa Hosseinzadeh, with affiliations at the University of Oregon, the Flatiron Institute and Rice University. The authors report higher success than ProteinMPNN at folding into the target backbone, and demonstrate motif inpainting and de novo design.
Cyclic peptides are attractive scaffolds, but sequence design tools trained on proteins are not tuned to their constrained geometry, and only a small number of cyclic peptide structures exist in the PDB. CyclicMPNN addresses this by combining 80 PDB cyclic peptide structures with a large set of computationally generated cyclic backbones.
Attribute | Details |
|---|---|
Base model | ProteinMPNN, fine-tuned for 200 epochs (epoch-164 checkpoint used) |
Residues | Canonical L-alpha-amino acids |
Training data | 80 PDB cyclic peptides plus GenKIC-generated 6-, 8- and 10-residue backbones |
Validation | Folding with HighFold, a modified AlphaFold for cyclic peptides |
Input | A cyclic peptide backbone, optionally with a fixed motif |
How CyclicMPNN Was Trained on Cyclic Peptide Backbones
Backbone generation: Rosetta GenKIC sampled poly-alanine cyclic backbones, filtered by Ramachandran and backbone hydrogen-bond criteria.
Pseudo-labeling: ProteinMPNN designed 5 sequences per backbone, the best by model score was kept, and HighFold folded it.
Quality filter: designs were kept only if backbone RMSD was at most 0.35 angstrom (6- and 8-mers) or 0.5 angstrom (10-mers).
Fine-tuning: the filtered designs and PDB structures, clustered with MMseqs2 at 0.3 identity, were used to fine-tune ProteinMPNN.
CyclicMPNN Benchmark Results Versus ProteinMPNN
On de novo test backbones, designs from CyclicMPNN folded closer to the target than those from ProteinMPNN, with higher median pLDDT.
Test set | Median RMSD, CyclicMPNN | Median RMSD, ProteinMPNN | Median pLDDT (CyclicMPNN vs ProteinMPNN) |
|---|---|---|---|
6-mer | 0.97 angstrom | 1.17 angstrom | 80.13 vs 76.18 |
8-mer | 1.13 angstrom | 1.55 angstrom | 82.00 vs 74.86 |
10-mer | 1.79 angstrom | 1.99 angstrom | 71.91 vs 70.18 |
14-mer (RFPeptide) | 0.81 angstrom | 1.16 angstrom | 77.51 vs 75.12 |
In a motif inpainting test built on the Nrf2 EETG motif with 100 eight-residue scaffolds, CyclicMPNN produced 12 stable peptides by P_Near funnel analysis versus 6 for ProteinMPNN, and seven sequences with P_Near of at least 0.6 versus one.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with CyclicMPNN on Tamarind Bio
De novo cyclic peptide design: assign sequences to generated cyclic backbones, such as 6- to 14-residue scaffolds.
Motif inpainting: design the remaining residues around a fixed functional motif, as shown with the Nrf2 EETG motif.
Redesign of known structures: propose sequences for cyclic peptide backbones taken from the PDB.
Candidate triage: generate several sequences per backbone and rank them by folding agreement.
How to Use CyclicMPNN on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and open CyclicMPNN.
Provide a backbone: upload a cyclic peptide backbone structure, either poly-alanine or poly-glycine, or one derived from the PDB.
Fix residues if needed: to inpaint around a motif, mark the motif positions as fixed, if exposed.
Set sampling options: choose how many sequences to generate; the authors designed 5 per backbone and kept the best by model score.
Run the job: launch the design and wait for results.
Download sequences: retrieve the designed sequences and scores.
Validate: fold the designs with a cyclic-aware predictor such as HighFold and check RMSD and pLDDT before synthesis.
Things to Keep in Mind
Design is limited to canonical L-alpha-amino acids; non-canonical residues are not supported.
Evaluation relies on predicted structures from HighFold, not on experimental validation.
Several comparisons use small sets, for example four X-ray structures and 100 motif scaffolds.
Stability statistics were computed on subsets of the designs.
Source: Andrew C. Powers, Yanapat Janthana and Parisa Hosseinzadeh, CyclicMPNN: Stable Cyclic Peptide Sequence Generation, bioRxiv 2026, DOI 10.64898/2026.01.31.702993, https://www.biorxiv.org/content/10.64898/2026.01.31.702993v1. Code: ParisaH-Lab/CyclicMPNN on GitHub.