SinCAA

A medicinal chemist who swaps one residue of a peptide for a close structural analogue expects the molecule's shape, and therefore its binding, to shift only slightly. Machine-learning models have no such intuition about non-canonical amino acids. A sequence model trained on the standard twenty-letter alphabet has no token for an N-methylated leucine or a fluorinated phenylalanine, and the usual workaround collapses every such residue onto one placeholder character — discarding exactly the chemistry that was added on purpose. SinCAA replaces that placeholder with a learned atomic representation, and it acquires one by training a graph transformer on the chemist's intuition directly: residues that adopt similar three-dimensional conformations should sit close together in embedding space.

It comes from the group of Xin Gao at the King Abdullah University of Science and Technology, with Syneron Opal and collaborators at Tongji University, Yonsei University and Dezhou University, published in Advanced Science.

The model is a representation learner, not a structure generator. Where NCFlow places a non-canonical residue into a pocket as explicit coordinates, SinCAA emits per-atom embeddings of a whole peptide graph for downstream predictors — having been pretrained only on isolated amino acids and small molecules, never on peptides.

#Key Features

• Conformational similarity metric: Each residue is capped with glycines, embedded in 3D with RDKit's ETKDG, aligned to a Cα-centred reference frame, and discretised into a voxel set labelled with element and partial charge. Two residues are compared by Earth Mover's Distance over the Jaccard similarities of their sampled conformers, matched with the Hungarian algorithm.
• Dual pretraining objectives: Contrastive learning teaches relative similarity — for each residue, two candidates are drawn from its twenty nearest neighbours and the closer one is the positive — while masked node reconstruction forces the model to recover each residue's specific atoms and bonds.
• Chained building blocks: Concatenating one- and two-residue chains during pretraining raises the overlap between subgraph patterns seen in training and those in real peptides, which is what lets an amino-acid-scale model reach peptide-scale graphs.
• Zero-shot substitution ranking: Distances between SinCAA embeddings correlate with measured changes in dissociation constant across single-position substitution series, so candidate swaps can be ranked without any labelled data for the target.
• Drop-in peptide encoder: Swapping SinCAA embeddings into PepNN's peptide module, at atom rather than residue granularity, extends protein–peptide binding-site prediction to cyclic and chemically modified peptides.

#Technical Details

The backbone is a GPS graph transformer with a Graph Isomorphism Network as the local message-passing operator, using two GIN layers per block rather than the customary one; the released checkpoint is 512 channels wide with 16 attention heads and four blocks. Pretraining used 387,482 non-canonical amino acids filtered from PubChem — split 309,986 for training and 77,496 for validation — with the masked-node task augmented by 10 million ZINC15 molecules. On held-out validation pairs the pretrained model separates conformationally similar residues from random ones at AUROC above 0.996.

Downstream evaluation centred on 2,761 KRAS-binding macrocyclic peptides curated from a public patent, 1,966 of them carrying more than one ring. Against small-molecule pretraining baselines and the peptide model PepLand, SinCAA improved Spearman's ρ by more than 10% on both Kd and IC50 prediction, and reached ρ = 0.492 on IC50 for test peptides containing residues absent from training, a setting where most baselines went negative. On cell permeability from CycPeptMPDB it reached an MSE of 0.251, edging a specialised cell-penetrating-peptide model built on hand-engineered descriptors (0.253). Zero-shot substitution ranking averaged ρ = 0.452 against 0.421 for Morgan fingerprint similarity. Feature extraction follows OpenFold, which the environment requires.

#Applications

The intended use is triage in therapeutic peptide programmes: ranking macrocyclic candidates by affinity, classifying NRAS-versus-KRAS selectivity, estimating membrane permeability, and locating binding sites on a target protein — all before synthesis. Because the embeddings carry signal without fine-tuning, it also supports in-silico residue scanning: proposing conformationally similar substitutions and ordering them by predicted property change. It extends to non-ribosomal peptides too, where ridge regression on SinCAA embeddings led six arylomycin antibacterial assays.

#Impact

SinCAA argues a useful point: a model never shown a peptide can still represent one, if it is taught the right relationships between the parts. That matters in a domain where labelled peptide data is scarce and many non-canonical residues appear only once across all public datasets. The evaluation is retrospective throughout — no designed peptide was synthesised or assayed, and the framework does not generate sequences — and the residue-scanning examples illustrate a workflow rather than recommend molecules. Code and the pretrained checkpoint are published on GitHub together with the curated amino acid list, though the repository carries no licence file, and the KRAS peptide dataset is available only on request from the corresponding author.