Diffusion models get better at cyclic peptides
A 2026 paper retrains an all-atom diffusion model to generate accurate conformers for small cyclic peptides, including noncanonical amino acids and complex linkages, with a practical stereochemistry correction step.
A lot of “AI for peptides” headlines implicitly assume the hard part is choosing a sequence.
In practice, peptide projects often stall on something less glamorous: what shape does this molecule actually adopt, and how many shapes does it sample? That question matters for binding, permeability, and even basic experimental interpretation.
It is also where cyclic peptides become uniquely annoying. Close a ring, sprinkle in a few noncanonical residues, add an unusual linkage, and many general-purpose structure tools stop being trustworthy.
A new Journal of Chemical Information and Modeling paper describes a targeted attempt to fill that gap. The authors retrained an all-atom diffusion model, AGDIFF, to generate conformer ensembles for small cyclic peptides, explicitly supporting noncanonical amino acids and “nonstandard” cyclization chemistries (PubMed).
Why cyclic peptides break a lot of defaults
Cyclic peptides are attractive because the ring can stabilize a bioactive conformation and resist proteases. But that same ring closure creates failure modes for modeling:
- you need the ring geometry to be correct,
- you need stereochemistry to be handled reliably,
- and you often care about an ensemble rather than a single best structure.
The paper’s framing is blunt: existing generalized biomolecular models are improving, but high-precision prediction for small cyclic peptides with noncanonical amino acids remains hard.
What the model actually does
AGDIFF was originally built for small-molecule conformer generation. In this work, it is retrained on a dataset of macrocyclic peptide conformers (CREMP, 36,198 members), using a 2D molecular graph representation.
That representation choice matters because it naturally accommodates:
- noncanonical amino acids,
- unusual connectivities,
- and complex linkages that do not look like “protein residues in a chain.”
They also add a stereochemical correction step, aimed at fixing a practical issue: models that are “good enough” geometrically but insensitive to chirality can quietly give you the wrong antipode.
On their benchmark, they report an average RMSD of 0.79 Å and a ring torsion fingerprint deviation of 6.55°, with analyses suggesting the generated conformers land in plausible regions of conformational space.
The real value: an unsexy layer for peptide design
This is not a clinical story. It is infrastructure.
If peptide design is moving toward “on-demand reagents”, as reviews are increasingly arguing, then the bottleneck shifts from generating ideas to vetting ideas with realistic constraints: specificity, stability, manufacturability, and yes, geometry (PubMed).
A diffusion model that reliably handles noncanonical residues and cyclic linkages is the kind of component that can make downstream workflows less wasteful. Not by guaranteeing a hit, but by reducing the number of “nice concept, impossible shape” molecules that make it into synthesis.
What would make this convincing beyond benchmarks
The obvious next step is prospective work.
A strong follow-on would be a case series where model-generated ensembles help a team pick between cyclic-peptide designs that are otherwise hard to distinguish, and then show that the chosen designs:
- bind better,
- permeate better,
- or translate into clearer structure-activity relationships.
That is where “accurate conformers” becomes more than a modeling win.