Loading…
Pocket-conditioned peptide designer: twin diffusion models generate an inhibitor backbone from receptor pocket geometry, then predict its sequence.
No providers recorded yet. Browse all providers
The usual way to build a peptide inhibitor computationally is a relay: a structure generator proposes a backbone that fits a receptor pocket, then an inverse-folding model reads that backbone and writes a sequence for it. The handoff loses the pocket. Inverse folders such as ProteinMPNN and ESM-IF see only the peptide's own geometry, so a sequence that folds to the intended shape may bind many proteins with that shape rather than the one target. They are also trained on full-length chains, far outside a peptide's length regime.
This model, from Po-Yu Liang and Jun Bai in the Department of Computer Science at the University of Cincinnati, closes that gap by conditioning both stages on the receptor. It chains two diffusion models — one that generates a peptide backbone given the pocket, and a second that predicts the amino acid sequence given both the generated backbone and the pocket — so pocket identity informs the sequence directly instead of being inferred from shape. The paper gives the model no name, calling it only "OurModel".
The second idea is representational. Rather than diffusing over 3D coordinates and needing an equivariant network to handle rotations, the peptide is encoded as four backbone bond angles and four dihedrals per residue, following FoldingDiff. Angles do not change under rotation or translation, so E(3) invariance is a property of the encoding rather than something the architecture must learn; the authors add the backbone oxygen to the angle set, which carries the hydrogen-bonding geometry that secondary-structure assignment depends on. Posted to arXiv in October 2024 and published at ISBRA 2025, it is the structure-first counterpart to the same lab's later PepEDiff, which instead diffuses in a protein-embedding space with no explicit geometry.
Both modules are transformer-based. The structure model trains with the wrapped normal noising process and wrapped smooth L1 loss from FoldingDiff; the sequence model combines an ELBO term with cross-entropy. Training data comes from BioLiP, which holds 781,684 protein–ligand interactions of which 35,167 are protein–peptide; filtering out complexes above 5 Å resolution, duplicate PDB entries, peptides under three residues and structures Biopython or DSSP could not parse leaves 8,868 receptor–peptide complexes, split 80:10:10. Every peptide is under 30 residues. On held-out structure generation the ext ensemble reaches 2.49 Å average backbone RMSD and 0.42 average TM-score, against 4.54 Å / 0.15 for RFdiffusion and 3.28 Å / 0.14 for ESMFold, with AlphaFold2 at 3.82 Å and OmegaFold at 7.41 Å. Sequence recovery on native peptide backbones is 59.76% for the ensemble and 47.41% for the single ext-4 model, versus 19.92% for GraDe-IF, 19.04% for ProteinMPNN and 17.86% for ESM-IF.
The target use is de novo design of peptide inhibitors for a chosen binding site — immune checkpoint blockade is the motivating example, where a peptide that occupies a checkpoint protein's partner-binding pocket can restore immune cell activity. A user supplies a receptor pocket and receives complete candidates — backbone and sequence together — with no separate inverse-folding pass. Because it is trained specifically on short peptides in complex with receptors, it fits screening workflows that need many pocket-matched candidates to triage by docking before synthesis.
The contribution is a clean demonstration that pocket information belongs in the sequence stage, not just the structure stage, and that an angle-based encoding buys E(3) invariance cheaply enough to make a two-stage peptide pipeline practical. Validation is entirely computational — Rosetta binding energy, docking score and contact rate, with no experimental binding data — and the docking scores trail non-pocket-aware pipelines, which the authors attribute to the metric being dominated by the receptor's energy. The repository ships code and links to pretrained checkpoints for both stages plus the cleaned dataset, but carries no license file, and weights and data sit on a personal university OneDrive rather than a durable archive, the same fragility as PepEDiff.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.