AtomWeaver: Multi-Component Flow Matching with a Structured Geometric Prior Facilitates Non-Canonical Peptide Design

Fixed-backbone sequence discovery, or inverse folding, is a critical recurring task in the development of new polypeptide therapeutics. Once promising backbones are established for a target pocket, computational inverse folding methods greatly help accelerate generation of candidate sequences. Such methods are mature for the traditional case of limiting to the fixed twenty-letter canonical vocabulary; however, they cannot access the broader space of non-canonical amino acids (NCAAs). This design constraint is exacerbated for peptide binders, a fast-growing modality that readily incorporates NCAAs, though in practice non-canonical design frequently depends on laborious medicinal-chemistry campaigns. An extension of inverse folding to NCAAs is thus critical to accelerating design of novel therapeutic peptides. AtomWeaver uses a joint all-site, atom-level generative scheme that does not restrict side-chain categorical assignment by either predetermined or co-resolving residue identity. Conditioned only on a fixed peptide backbone and its target protein, its multi-component flow guides side-chain atoms as unlabeled points in R3 from a nested shell prior to a variable-count final atom cloud. Identity is then read by matching each predicted cloud against a reference library of canonical and non-canonical templates. Since identity is decided only at decode time, the addressable vocabulary is a property of the library rather than of the trained weights: a new NCAA costs one reference structure and no retraining, and the model can select residues it was never prompted for and never saw in training. On a deep mutational scan of two peptide-target systems, AtomWeaver's canonical readout shows high observed mean agreement with experimental values among the compared inverse-folding methods. In the mixed canonical-noncanonical setting that canonical-only baselines cannot support at all, it likewise retains ranking signal across both systems. AtomWeaver also displayed self-consistent designs on de novo binder backbones, with the highest interface confidence among compared methods. Notably, it reached these metrics while achieving broad empirical coverage of our 300-residue vocabulary, including four non-canonical types never visible in training. AtomWeaver thus serves canonical and non-canonical peptide design alike, while transforming residue vocabulary to an expandable inference-time choice.