RareFold designs non-canonical peptides, organoids flag immune silence

What the work claims

Li et al. introduce RareFold, an AlphaFold2-style architecture that represents each of the 20 canonical amino acids and 29 non-canonical amino acids as a distinct token.¹ The authors claim this residue-as-token approach enables accurate structure prediction and sequence design across chemically diverse residues, and they apply it in EvoBindRare to design linear and cyclic peptide binders against Ribonuclease A. They further claim that the resulting peptides, which incorporate multiple non-canonical residues in their predicted interfaces, show no detectable added immunogenicity in human-derived organoid and PBMC models.

How it works

Conventional structure-prediction models are confined to the 20 canonical amino acids. RareFold extends the EvoFormer architecture by assigning each canonical or non-canonical residue its own token and its own local coordinate frame, yielding 49 distinct frame types. This lets the model learn sequence-structure relationships for modified chemistries without approximating them as their closest natural analogue.

The network was trained on roughly 75,000 structures, a tiny fraction of the data used by AlphaFold3. On a held-out validation set of 731 structures covering 24 non-canonical types, RareFold achieved a median Cα lDDT of 0.96. On a stricter test set of 174 structures covering 13 non-canonical types, the fine-tuned model reached a median Cα lDDT of 0.76, compared with 0.84 for AlphaFold3, though the authors note AlphaFold3 was trained on far more data and sometimes produced physically implausible collapsed side-chain geometries for selenomethionine. Predicted confidence (plDDT) correlated strongly with actual accuracy (Spearman rho = 0.876).

For design, the authors inverted RareFold into EvoBindRare and ran 1,000 mutation steps to generate 10-20 residue linear and cyclic peptides. Synthesis constraints limited the designs to 20 canonical plus 12 non-canonical amino acids. Seven linear and six cyclic candidates were selected; one linear binder (L17) and one cyclic binder (C14) were experimentally characterised by surface plasmon resonance, with affinities of 2.13 µM and 8.77 µM respectively, comparable to the 1.81 µM wild-type ligand. Hydrogen-deuterium exchange mass spectrometry confirmed that both designs protected regions of the target consistent with their predicted binding interfaces.

Where a skeptic should push

The binding success rate is modest: one of seven linear designs showed measurable binding, and only one of six cyclic designs could be synthesised at high purity and tested. This is consistent with the difficulty of de novo binder design, but it is not evidence that non-canonical residues confer a decisive advantage in affinity. The authors themselves note that canonical amino acids are often sufficient for affinity optimisation in well-characterised systems.

The immunogenicity experiments are the most organoid-relevant part, and they are also the most bounded. C14 and L17 were incubated at 0.25, 2.5, and 25 µM with PBMCs and with immune tonsil organoids from independent donors for seven days. No cytokine secretion (IL-6, TNF-alpha, IL-1beta, IFN-gamma, IL-12) above the wild-type baseline was detected, no peptide-specific IgG/IgM/IgA antibodies appeared, and flow cytometry showed no shift in T cell, B cell, or monocyte subsets. Positive controls behaved as expected. But this is a single seven-day exposure in two ex vivo systems; it does not capture biodistribution, repeated dosing, HLA diversity, tissue-resident immune cells, or the influence of disease state.

What this changes for engineered peptide immunogenicity in organoids

For organoid-based drug discovery, the immunogenicity readout is the important advance. Engineered peptides, especially those containing non-canonical amino acids or cyclisation, can evade standard immune surveillance, but that same property can make them unpredictable immunogens. RareFold pairs computational design with an ex vivo human immune organoid assay, giving a plausible early filter for anti-drug antibody risk and innate cytokine release before moving to animal models or patients.

The opportunity is to make immunogenicity profiling a standard checkpoint in peptide-engineering pipelines. Tonsil organoids retain organised lymphoid tissue with B and T cell compartments, so they can in principle detect both T-cell-dependent and T-cell-independent responses better than PBMC monocultures. If the assay proves reproducible across donors and peptide chemistries, it could reduce late-stage immunogenicity failures for cyclic peptides, stapled peptides, and other synthetic biologics.

The threat is false reassurance. A negative result in two donors after one week does not mean a peptide will be non-immunogenic in patients. The assay also measures only one kind of immune response; it does not address off-target pharmacology, manufacturing impurities, or the immunogenicity of the full formulation. And because the field lacks standardised immune-organoid protocols, cross-lab reproducibility remains unproven.

The bottom line

RareFold is a credible technical extension of structure-prediction models into non-canonical chemistry, and EvoBindRare demonstrates that the resulting designs can be experimentally validated. The immunogenicity data in PBMCs and tonsil organoids are encouraging but preliminary. What would move the needle is a larger design campaign in which RareFold produces a clinically relevant binder that also passes a longitudinal, multi-donor organoid immunogenicity study.