RareFold

RareFold is a protein structure prediction network that models the 20 canonical amino acids together with 29 noncanonical amino acids (ncAAs), paired with EvoBindRare, a generative framework that inverts the predictor to design peptide binders over that expanded alphabet. It was developed in Patrick Bryant's group at Stockholm University, with experimental work carried out at Karolinska Institutet, and posted as a bioRxiv preprint in May 2025.

Structure prediction and design have been effectively locked to the 20 proteinogenic residues, even though the Protein Data Bank contains 331 distinct amino acid types in single-chain proteins. That chemistry matters for engineering: ncAAs confer protease resistance, can reduce immunogenicity, and expand the catalytic repertoire available to designers. Two obstacles have stood in the way. Multiple sequence alignments cannot resolve modified residues, which appear as "X" in sequence databases, and atom-level frameworks such as AlphaFold 3 represent ncAAs by mapping them onto the nearest canonical analogue, discarding the very chemistry that defines them while making input and output dimensions impractical for iterative design.

RareFold's answer is tokenization. Each residue, canonical or modified, is one distinct token with its own set of local atomic frames — 49 frame types in total — so the model learns residue-specific structural and coevolutionary patterns while keeping a compact, fixed input space that a mutation-and-rescore design loop can call thousands of times. RareFoldGPCR, a later preprint from the same lab, transfers this model to G protein-coupled receptor peptide agonist design.

#Key Features

• 49-residue alphabet: The 20 canonical residues plus 29 of the most common ncAAs in the PDB, including MSE, SEP, TPO, PTR, MLY, M3L, HYP, and PCA — many of them ordinary post-translational modifications.
• EvoBindRare binder design: The inverted network designs linear and cyclic peptide binders of 10–20 residues from a target sequence alone, with no binding site specified and no protein complexes seen during training.
• Efficient design loop: A batched implementation cuts GPU memory use by 97% and speeds the design cycle 6-fold, running 1,000 mutation steps on a single 40 GB NVIDIA A100.
• Experimental validation: Designs against Ribonuclease A gave a linear binder (KD 2.13 µM) and a cyclic binder (KD 8.77 µM) by surface plasmon resonance, against 1.81 µM for the wild-type ligand, with HDX-MS confirming engagement at the predicted interface.
• Calibrated confidence: Predicted lDDT tracks true accuracy closely (Spearman ρ = 0.876), letting campaigns filter designs on plDDT > 85.

#Technical Details

RareFold extends the EvoFormer architecture from AlphaFold 2 to 49 token types, generating alignments with HHblits against uniclust30_2018_08 and using no structural templates. Training used 74,882 single-chain PDB structures — 62,530 for training, 12,175 for validation, and a 177-structure test set — of which 7,599 contain modified residues; the split follows AlphaFold 3's 2021-09-30 cutoff with 20% sequence-identity clustering, so the test set falls outside AlphaFold 3's training data. Training ran with batch size 24 on 256-residue crops across eight 80 GB A100 GPUs at a constant learning rate of 10⁻³ using the AlphaFold2 loss, followed by a 512-residue large-crop fine-tune and a further 5,000-step fine-tune penalizing inter-residue clashes for the design network.

Median Cα-lDDT reaches 0.96 on validation. On the test set, large-crop fine-tuning lifts median Cα-lDDT from 0.56 to 0.76 against 0.84 for AlphaFold 3 — a narrow gap given that RareFold saw about 0.04% of AlphaFold 3's training structures. Side-chain RMSDs differ by only 1–2 Å for most residue types, and RareFold is more accurate on SEP, SAH, and PCA. AlphaFold 3's lower RMSD on selenomethionine is an artifact of collapsed side-chain geometries with atoms closer than 1 Å; RareFold produces physically valid conformations and predicted all 177 targets within 40 GB, where AlphaFold 3 ran out of memory on three at 80 GB.

#Applications

RareFold serves peptide therapeutic discovery, where ncAAs and cyclization are established routes to protease resistance and oral availability. Because many supported residues are phosphorylated, methylated, or hydroxylated forms of canonical amino acids, the predictor is also useful for modelling post-translationally modified proteins. EvoBindRare gives groups without large compute budgets a practical de novo binder pipeline, and immunogenicity profiling of the RNase A designs in PBMC and tonsil organoid cultures detected no cytokine induction, peptide-specific antibodies, or shifts in immune cell subsets.

#Impact

RareFold shows that a token-per-residue representation, rather than atom-level modelling, is enough to place chemically diverse ncAAs accurately, and that the resulting network is cheap enough to invert into a design loop on a single GPU. The experimental results are candid about their limits: affinities match but do not exceed the wild-type ligand, ncAAs conferred no consistent affinity gain over canonical designs, and validation rests on a single target with one of seven linear designs and one cyclic design binding. The work remains a preprint awaiting peer review; code is released under Apache 2.0, while the parameters carry a CC BY-NC 4.0 license that restricts commercial use.