ncAA-RepDistill: a chemistry-native SMILES representation for permeability prediction and local edit ranking of non-canonical cyclic peptides

Abstract

Non-canonical amino-acid (ncAA) cyclic peptides occupy a chemically rich design space for tuning permeability, stability, and bioavailability, but they remain difficult to model because local chemical edits are strongly scaffold-dependent. Existing representation strategies address this problem only partially: sequence-oriented models often blur atom-level modifications, whereas general molecular models do not explicitly preserve peptide-level context. Here we present ncAA-RepDistill, a graph-aware SMILES representation model that distills complementary protein-space and molecular-space priors into a single deployable Transformer student. The model combines dual-teacher alignment with masked SMILES modeling, randomized-SMILES consistency, and a staged training strategy consisting of canonical-peptide pretraining followed by unlabeled ncAA adaptation. Under a unified frozen-feature evaluation protocol, ncAA-RepDistill improves the primary non-canonical cyclic-peptide permeability task, preserves scaffold-constrained local-edit ranking through ncAA-Scan, and retains useful transfer in a structured semipeptidic macrocycle analog-series validation and an external ncAA-aware antimicrobial benchmark. These results indicate that ncAA-RepDistill is a chemistry-native SMILES representation for ncAA cyclic peptides, particularly when endpoint prediction and local-edit sensitivity depend on preserving local chemical edits and peptide context. We introduce a chemistry-native, graph-aware SMILES representation that encodes canonical peptide bonds, macrocyclization linkers, and non-canonical residues in one unified atom-and-bond space, shaped by distillation from complementary protein-space and molecular-space teachers. Unlike sequence-alphabet or generic molecular encoders, it preserves scaffold-dependent local chemical edits, enabling both permeability prediction and within-scaffold single-edit ranking of non-canonical cyclic peptides under a unified frozen-feature protocol.