AtomWeaver™
AtomWeaver™
AtomWeaver™ is our inverse folding model for peptide binders. Given a fixed peptide backbone and the target protein it sits against, it proposes the side chains that should fill the pocket. Unlike the inverse folding methods it is measured against, it is not confined to the twenty canonical amino acids.
Conventional models pick a residue type first and place its atoms afterwards, which permanently fixes the alphabet they can reach. AtomWeaver™ reverses that order. It generates side-chain atoms as unlabelled points in space, then reads identity off the finished atom cloud. Vocabulary becomes a choice made at inference time rather than a property of the trained weights.
Inverse
Folding
Folding
Fixed-backbone sequence discovery is a recurring step in building a peptide therapeutic. Once a promising backbone is established for a target pocket, inverse folding accelerates the search for sequences that will hold it. That task is mature for the canonical twenty and unsolved for the thousands of non-canonical amino acids.
Non-canonical amino acids are much of what makes a peptide a drug. They confer protease resistance, improve membrane permeability and tune pharmacokinetics, and many thousands are commercially available. A method restricted to the canonical twenty therefore reaches only a fraction of the chemistry that peptide synthesis allows, which is why non-canonical design still depends on slow medicinal chemistry campaigns.
Atoms
First
First
The model is conditioned only on the fixed peptide backbone and its target protein. A multi-component flow carries side-chain atoms from a nested shell prior to a final cloud of variable size. Each atom starts on a shell at an empirical radius along a pseudo-Cβ direction derived from the backbone alone, so the prior carries no hint of residue identity.
A discrete element track runs alongside the coordinate flow. Its vocabulary includes a token for atoms that are simply absent, so the model decides how large a side chain is instead of being told. Ten SE(3)-equivariant blocks handle the atom track, with the target protein's own atoms participating directly in the graph.
Identity is assigned only once the cloud is complete. A decode-time discretizer scores each predicted cloud against a library of canonical and non-canonical reference templates, blending a distance-matrix correlation with a learned classifier and penalties for atom count, element composition and chirality.
Training spans bound chemistry of every kind: small molecule–protein complexes, antibody–protein complexes and peptide–protein complexes, together with a large body of synthetic peptide–protein complexes carrying non-canonical amino acids.
Vocabulary
Because identity is decided at decode time, the addressable vocabulary is a property of the library rather than of the trained weights. Adding a new non-canonical residue costs one reference structure and no retraining.
Our Monomer Database catalogues 4,719 monomers. AtomWeaver™ addresses 450 of them today, and reaching the rest is a question of adding reference structures rather than training a new model.
In evaluation it selected residues it had never been prompted for. The clouds it produces reproduce real side-chain geometry, and reach chemistry no canonical-only method can express at all, among them cysteine staples, phosphorylated serine and tyrosine, and cyclohexylalanine.
Benchmarks
PeptideArena is our open benchmark for peptide binder design, comprising de novo designed peptide backbones against twelve protein targets, each supplied as a two-chain complex of backbone and receptor. Because those backbones are designed rather than observed, no native sequence exists to recover, and evaluation rests on structural self-consistency: whether a design refolds onto the backbone and binding pose it was given. Sequence recovery is undefined outside the canonical twenty, so this is the criterion that makes non-canonical design measurable.
On PeptideArena we designed 353 de novo peptide backbones across twelve targets and refolded every design. AtomWeaver™ returned self-consistent structures and the highest interface confidence, scored by pDockQ2, of the methods compared.
Across those twelve targets the open-vocabulary readout emitted 168 distinct non-canonical codes while holding that accuracy, so the breadth of chemistry did not cost design quality. Among them were β-amino acids and N-methylated amino acids.
Against deep mutational scanning data for two peptide-target systems, covering 39 positions and 1,548 measured mutants, the canonical readout showed high agreement with experiment. In the mixed canonical and non-canonical setting, which canonical-only baselines cannot address at all, it kept its ranking signal on both systems.
In
Practice
Practice
Designing with non-canonical amino acids is the core of our peptide engineering practice rather than an extension of it. It runs through PQ203, our Sortilin-targeting peptide–drug conjugate now in Phase 1 trials for advanced solid tumours, whose peptide carries non-canonical residues.
AtomWeaver™ widens the chemistry that practice can reach. Proposing non-canonical side chains directly, instead of designing a canonical sequence and substituting into it afterwards, shortens the path from a validated binding mode to a sequence worth synthesising. That applies in the same way to partnered programs and to our own pipeline.
A second-generation model, trained on a substantially larger dataset, is already in use internally. We are always glad to bring this capability to collaborations with industry partners; if that is of interest, get in touch.
Paper
The full method, benchmarks and supplementary material are published on bioRxiv: AtomWeaver: Multi-Component Flow Matching with a Structured Geometric Prior Facilitates Non-Canonical Peptide Design (Kitaygorodsky et al., 2026)
The release includes the trained weights, inference code and the reference library, along with the tools to extend it. A design run takes a peptide backbone and its target in a single PDB. Sampling uses a GPU; the read-out runs on CPU.