Use CyclicMPNN Online
Commercially Available CyclicMPNN No-Code Web Server
CyclicMPNN: Next-Generation Sequence Generation for Stable Cyclic Peptides
Discover CyclicMPNN, the state-of-the-art deep learning model tailored specifically for cyclic peptide sequence design. By bridging the gap between traditional protein design models and the unique structural demands of macrocycles, CyclicMPNN enables researchers to rapidly generate energetically stable sequences optimized to fold precisely into target backbones.
What is CyclicMPNN?
Cyclic peptides represent a highly promising class of therapeutics due to their exceptional stability, cell permeability, and resistance to proteolytic degradation. However, designing sequences using only canonical amino acids that stably fit a specified cyclic backbone has historically been a significant challenge.
CyclicMPNN is a specialized variant of the revolutionary ProteinMPNN architecture. Because the original ProteinMPNN was trained predominantly on larger macromolecular structures from the Protein Data Bank (PDB), it lacks native representation for small cyclic peptides. CyclicMPNN solves this limitation by fine-tuning ProteinMPNN weights on a meticulously engineered combination of native crystal structures and extensive in silico generated cyclic poly-alanine peptide backbone ensembles.
The result is a highly specialized model that dramatically outpaces general protein sequence architectures, producing stable, diverse, and accurate sequences in a single round of design.
Key Features & Capabilities
Unrivaled Structural Accuracy: CyclicMPNN achieves superior structural reconstruction accuracy (lower RMSD) and significantly higher confidence metrics (pLDDT scores) compared to both ProteinMPNN and alternative models like HighMPNN across all evaluated lengths (6, 8, 10, and 14 residues).
De Novo Peptide Generation: Seamlessly pair CyclicMPNN with synthetic, diffused, or machine-learning-generated backbones (such as GenKIC or RFPeptide) to quickly evaluate and assign high-probability sequences.
Advanced Motif Inpainting: Lock down specific functional or binding segments (motifs) and let CyclicMPNN conditionally generate the surrounding scaffold. CyclicMPNN produces up to twice as many stable, high-affinity peptides in constrained design tasks than standard models.
Enhanced Sequence Diversity: Avoid repetitive sequence spaces. CyclicMPNN explores a broader, alternative landscape of compatible sequence variations, driving novelty and better hit rates in downstream therapeutic discovery.
Performance & Benchmarks
CyclicMPNN demonstrates exceptional benchmarks across both de novo scaffolds and experimentally validated X-ray structures:
Recapitulating Experimental Structures: In direct evaluations against PDB crystal structures (such as 9CDT, 9HGC, 9CDU, and 9HGD), CyclicMPNN provided a highly consistent sequence recovery rate and consistently achieved sub-angstrom structural reconstruction errors (RMSD <= 1.0 A) in a single round of sequence design.
Sub-Angstrom De Novo Accuracy: For short macrocycles (de novo 6-mers), CyclicMPNN achieved a median RMSD of just 0.97 A (compared to 1.17 A for ProteinMPNN) and a dramatically elevated stability metric (PNear value of 0.6 vs. 0.08).
Scalability to Large Scaffolds: On larger 14-mer backbones generated via RFPeptide, CyclicMPNN improved structural recapitulation by 22.78% relative to ProteinMPNN and by 42.85% relative to HighMPNN.
What is Tamarind Bio?
Tamarind Bio is a leading computational biology platform that provides user-friendly web access to cutting-edge deep learning utilities for molecular design and structure prediction. By turning complex command-line bioinformatic tools into accessible, high-throughput cloud applications, Tamarind Bio empowers researchers, structural biologists, and therapeutics developers to accelerate their discovery pipelines without needing dedicated local hardware or complex environments.
How to Use CyclicMPNN on Tamarind Bio
Designing stable, high-affinity cyclic peptide sequences on Tamarind Bio requires only a few straightforward steps:
Upload or Input Your Target Backbone: Provide the 3D coordinates of your cyclic peptide backbone in standard standard format (such as a PDB file) generated from structural suites or generation tools like GenKIC, RFPeptide, or CyclicCAE.
Configure Your Generation Parameters:
Choose De Novo Sequence Design to sample complete sequences for your scaffold unconditionally.
Choose Motif Inpainting if you want to lock specific amino acid indices in place (e.g., a specific target-binding segment) and conditionally generate the rest of the macrocyclic ring.
Run Inference: Submit your job to generate the optimized sequences. The underlying platform samples sequences according to CyclicMPNN's specialized weights.
Analyze the Output: Download the generated sequences alongside their respective model scores. You can seamlessly feed these results directly into structure prediction modules (like HighFold) right on the platform to check your design's predicted RMSD, pLDDT scores, and structural convergence.