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…
In the pursuit of new polypeptide therapeutics, researchers often engage in fixed-backbone sequence discovery, or inverse folding, to establish promising backbones for target pockets. While existing computational methods excel at generating candidate sequences with the traditional twenty-letter canonical vocabulary, they fall short when it comes to incorporating non-canonical amino acids (NCAAs).
This limitation becomes even more pronounced in the realm of peptide binders, which increasingly rely on NCAAs, often necessitating time-consuming medicinal-chemistry campaigns to devise non-canonical designs. To address this challenge, researchers have developed AtomWeaver, a novel extension of inverse folding that utilizes a joint all-site, atom-level generative scheme.
This innovative approach does not restrict side-chain categorical assignment by predetermined or co-resolving residue identity, allowing for a more expansive exploration of peptide design possibilities. AtomWeaver's multi-component flow guides side-chain atoms as unlabeled points in a three-dimensional space, starting from a nested shell prior to a variable-count final atom cloud.
The model then reads the predicted identity by matching each cloud against a reference library of canonical and non-canonical templates. Crucially, identity is determined only during the decoding phase, enabling the model to access an addressable vocabulary that extends beyond the initial training weights. This means that incorporating new non-canonical amino acids is as simple as adding a reference structure without the need for retraining.
AtomWeaver's performance was validated through a deep mutational scan of two peptide-target systems, where it demonstrated high observed mean agreement with experimental values when compared to other inverse-folding methods. Moreover, AtomWeaver maintained ranking signal across both systems in the mixed canonical-noncanonical setting, where traditional canonical-only baselines struggle.
The method also displayed self-consistent designs on de novo binder backbones, outperforming other methods in terms of interface confidence. Perhaps most impressively, AtomWeaver achieved these metrics while demonstrating broad empirical coverage of their 300-residue vocabulary, including four non-canonical types that had never been encountered during training.
By transforming residue vocabulary into an expandable inference-time choice, AtomWeaver proves to be a versatile tool for both canonical and non-canonical peptide design, revolutionizing the field with its ability to adapt to new design possibilities without the need for extensive retraining.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
