NCFlow

Non-canonical amino acids (ncAAs) — halogenated aromatics, N-methylated backbones, unnatural side chains sold by peptide synthesis vendors — are central to macrocyclic peptide drugs and engineered enzymes, but sit outside what most protein design software can represent. AlphaFold2- and RoseTTAFold-based tools such as RFdiffusion cannot encode a residue outside the standard alphabet, and all-atom predictors in the AlphaFold 3 family need per-molecule conditioning inputs — atom counts, RDKit conformers — that differ for every ncAA, which makes a gradient-descent design loop awkward when the modification is not fixed in advance.

NCFlow, from Jin Sub Lee and Philip M. Kim at the University of Toronto, narrows the task instead. Rather than co-folding a whole complex, it treats ncAA modeling as single-residue structure prediction: given a protein backbone, a residue position, the surrounding pocket, and a SMILES description of any ncAA, it generates that residue's 3D conformation in place while repacking neighboring side chains. Because the ncAA enters as atom types and a bond graph rather than a vocabulary index, chemistry never seen in training is handled the same way as chemistry that was.

Fewer than 0.02% of PDB residues are non-canonical, many of them crystallographic aids such as selenomethionine, so NCFlow borrows data from general chemistry and protein-ligand structures in a three-stage curriculum before ever seeing an ncAA. The concurrent RareFold approach — whose GPCR successor is catalogued as RareFoldGPCR — extends AlphaFold2 to a fixed pool of 29 ncAAs; NCFlow leaves the vocabulary open.

#Key Features

• Arbitrary ncAA chemistry from SMILES: Any residue RDKit can parse can be placed, including the 939 synthesizable Fmoc-protected ncAAs curated from a commercial peptide-synthesis catalog.
• Three-stage pretraining ladder: Small-molecule conformer generation, then protein-ligand pocket modeling, then ncAA finetuning — over three orders of magnitude more training data than the PDB's ncAA records alone.
• Per-atom confidence estimates: A downsized confidence module predicts per-atom pLDDT, which correlates strongly and negatively with RMSD and is used to discard poor samples.
• In silico deep mutational scanning: Every peptide position is mutated to every compatible ncAA, then triaged by pLDDT, the AEV-PLIG affinity predictor, and alchemical relative binding free energies from the Alchemical Transfer Method.
• Lightweight inference: Sampling needs under 4 GB of VRAM and roughly 5 seconds per sample on a consumer RTX 3060, with 8 samples per ncAA enough to recover correct stereochemistry.

#Technical Details

NCFlow uses the independent-coupling conditional flow matching formulation, integrating a learned velocity field with a 10-step Euler solver from Gaussian noise centered on the target residue's Cα. The network reuses AlphaFold 3's Pairformer and Atom Transformer modules as implemented in Boltz-1, but drops the tokenization scheme for a purely atom-level representation: 16 Pairformer layers and 4 Atom Transformer layers with single and pair dimensions of 128 and 64, totaling 8.7 million trainable parameters. Pair features are interatomic distances plus one-hot bond orders. The architecture is not equivariant, so training uses rotation and translation augmentation, and samples with inverted chiral centers are filtered post hoc against the input SMILES.

Training ran on 4 NVIDIA A100 40GB GPUs for roughly seven days per stage: 13.8 million PubChem3D conformers, then 293,591 Plinder v2 protein-ligand complexes (mean symmetry-corrected ligand RMSD 2.23 Å, 62% under 2.0 Å), then 38,618 ncAA-protein environments from a 2023 PDB snapshot, holding out 231 rare ncAAs for testing. Pretraining improves mean ncAA RMSD by 0.21 Å over training from scratch (1.58 Å), and buried residues are predicted more accurately (1.08 Å) than surface ones (1.79 Å). In single-chain mode against backbone-superimposed predictions, NCFlow reaches 1.43 Å mean RMSD versus 3.29 Å for Boltz-1 and 3.24 Å for Boltz-2.

#Applications

The intended use is late-stage optimization of peptide binders: take an experimentally determined protein-peptide complex, enumerate single ncAA substitutions across the peptide, and rank them for predicted affinity gain before committing to synthesis. The authors ran this on four complexes spanning linear helical peptides, a head-to-tail cyclic peptide, and one cyclized both head-to-tail and by a disulfide, recovering variants predicted to improve binding by up to -7.0 kcal/mol through added halogen contacts, hydrogen bonds, or solvent-facing polar groups. Conditioning on a fixed backbone and a local pocket, it slots in after de novo backbone generation.

#Impact

NCFlow is a preprint awaiting peer review, and its design results are entirely computational — no ncAA variants were synthesized or assayed. The paper is candid about where the pipeline is fragile: the scoring stack works well on some systems (Spearman 0.76 for alchemical free energies on MDM2-p53) and not at all on others, alchemical simulations often fail on peptides longer than about 30 residues, and only single mutants were explored. Its contribution is the demonstration that a small, cheap, single-residue generative model can outperform far larger co-folding predictors at placing unfamiliar chemistry into a known pocket, without a curated residue list. Training and inference code is MIT-licensed, with checkpoints distributed separately.