RareFoldGPCR

RareFoldGPCR (RFG) is a G protein-coupled receptor (GPCR)-specialised AI model for structure prediction and peptide design that supports noncanonical amino acids (NCAAs). It was developed in Patrick Bryant's group at Stockholm University and released as a bioRxiv preprint in October 2025. NCAAs expand the chemical space of peptides beyond the twenty standard residues, opening new interaction modes and functional activities that are inaccessible to natural sequences. GPCRs, meanwhile, form one of the largest classes of therapeutic targets, yet designing peptides that selectively modulate their signalling remains difficult.

RFG adapts the RareFold model—which combines structure prediction with binder design over an expanded amino-acid alphabet—through transfer learning on high-resolution GPCR structures from GPCRdb. The result is a system that can model and rationally design both linear and cyclic peptides bearing NCAAs, despite never having been trained on NCAA-based GPCR modulators. This zero-shot generalization to new chemistry distinguishes RFG within the growing family of AlphaFold-derived design tools and the broader EvoBind/RareFold ecosystem.

#Key Features

• NCAA-aware design: RFG recognizes 49 amino-acid types—the 20 standard residues plus 29 rare variants such as MSE, TPO, MLY, PTR, and HYP—enabling peptides with chemistry beyond natural design.
• GPCR specialization via transfer learning: Transfer learning on high-resolution GPCRdb structures adapts the base RareFold model to accurately model and design GPCR-targeting peptides.
• Linear and cyclic agonists: RFG designs both linear and cyclic peptide agonists, including cyclic peptides with entirely novel sequences and topologies that create new agonist modes.
• Experimental validation: Designed agonists for the glucagon-like peptide-1 receptor (GLP1R) were validated for functional activity in cell-based assays.
• Pathway-biased signalling: Design metrics relate to pathway specificity, allowing agonists that activate the cAMP response without recruiting β-arrestin to reduce receptor desensitisation.

#Technical Details

RFG is built by transfer learning from RareFold, a structure-prediction-and-design model that operates over an amino-acid alphabet extended to include noncanonical residues. Training on high-resolution GPCR structures curated from GPCRdb specializes the model for this receptor class while preserving its ability to place NCAAs, all without exposure to NCAA-containing GPCR modulators during training. The authors probe RFG by expanding different regions of the native GLP-1 hormone to generate active agonists, and by designing cyclic peptides with novel topologies. The released code is licensed under Apache 2.0, while model parameters and the design protocol are distributed under CC BY-NC 4.0 for non-commercial use; the pipeline uses HHblits with Uniclust for sequence alignment and runs designs in minutes on an NVIDIA A100 GPU.

#Applications

RFG is aimed at peptide therapeutic discovery against GPCRs, a target class central to metabolic, neurological, and endocrine disease. By supporting NCAAs and cyclization, it lets researchers design agonists with improved properties and novel binding modes, and its pathway-specificity analysis enables biased agonism—for example, sustaining cAMP signalling while limiting β-arrestin-driven desensitisation. The GLP1R case study is directly relevant to metabolic drug development, and the general strategy extends to other GPCR targets and peptide modalities.

#Impact

RareFoldGPCR demonstrates that transfer learning on a specific target class can generalize a design model to new chemistry and molecular topology, producing experimentally active agonists that incorporate residues never seen in training. This bridges computational peptide design and functional pharmacology, showing that structure-based design can reach beyond the constraints of the twenty natural amino acids. As a preprint awaiting peer review, its conclusions rest on the reported cell-based assays, and the non-commercial license on parameters and protocol limits industrial reuse.