Peptide design stops picking among twenty letters: AtomWeaver packs side-chain atoms first and names the residue afterwards, so a new building block costs minutes
Synopsis
The inverse-folding models used to design peptides emit a probability over exactly twenty natural residues at each position, and adding an unnatural one means supplying its structural data and retraining the whole model. AtomWeaver instead grows side chains as atoms in three dimensions under a fixed backbone and a target protein, then matches each finished atom cloud against a reference library of residue templates to read off identity. On PeptideArena, 353 backbones against twelve targets, it stayed competitive with twenty-residue methods and held the same performance regime with a 300-residue vocabulary.
Interpretation
Peptide design no longer ends in a probability over twenty residue labels; the model lays down atoms in three dimensions and decides the residue name last. ProteinMPNN and its descendants, including PeptideMPNN, which finetunes ProteinMPNN to design peptide sequences against a given target binding site, emit a distribution over exactly twenty labels at every position, so a residue outside that layer cannot be proposed however well it would fit the pocket. On 353 backbones from PeptideArena, which sets de novo peptide binders against twelve protein targets, designs were refolded with an independent structure-prediction model and scored on whether they returned to the intended pose; AtomWeaver was competitive with established canonical methods, and performance stayed in the same regime when the vocabulary moved from 20 residues to 300.
Each side-chain atom starts on a nested shell at roughly the distance where that atom usually sits from the backbone, strung along the ray the side chain would project at that site, toward the receptor when buried and into solvent when exposed. Atom count resolves together with geometry: every atom slot can be a ghost that is not there, and ghosts flow back onto the backbone and vanish, so glycine's zero atoms and tryptophan's ten come out of the same machinery. Residue identity is read off only at the end by matching each finished atom cloud against a reference library of residue templates, and across 3,500 designs the model used 168 distinct non-canonical amino acids, about 60% of what was available to it, four of which it had never seen in training.
Perspective
The evaluation here is computational: 353 backbones and twelve protein targets, with designs refolded by an independent structure-prediction model and judged on whether they return to the intended binding pose. Stereochemistry and backbone extension were also demonstrated on the Hirulog-3 complex with thrombin, recovering 69.6% of D sites while mislabeling under 3% of L residues. Extending the vocabulary has a stated route: supply reference side-chain conformations for the residue, refit the read-out head on CPU in a few seconds, and design.
The joint redesign of Hirulog-3 swaps β-homoarginine for β-homotryptophan, trades a salt bridge for pocket-filling contacts through an asparagine–arginine–aspartate motif, and adds a D-alanine, refolding back into essentially the native pose; these are described as structural hypotheses rather than validated analogues.