Peptide binder design by mimicking the binding interface of a known receptor or antibody, generating all-atom peptides through latent diffusion.
No providers recorded yet. Browse all providers
Peptides sit in a useful gap between small molecules and biologics: synthetically made, better at penetrating tissue than an antibody, and engineerable for high specificity. What they lack is a reliable route to discovery. Screening a random library against a new target returns hits at a rate low enough that most programmes never start, and generative models trained on general proteins tend to produce helical binders rather than the short, loop-heavy peptides that make good drugs.
PepMimic reframes the problem as imitation. For most therapeutic targets a binder already exists — an approved antibody, a nanobody, or the natural receptor the target evolved to engage — and that binder's interface encodes where and how to grip the protein. PepMimic extracts the reference interface and uses it to steer generation of a short peptide that reproduces the same contact pattern, moving binding knowledge from a large molecule into a small one. Where no binder exists, the authors first design a protein binder with RFdiffusion and then mimic that artificial interface, which is how they reached TROP2.
The model came from the Institute for AI Industry Research at Tsinghua University with collaborators at Peking University, Renmin University of China, and Helixon, and was published in Nature Biomedical Engineering in 2025.
PepMimic has three trained components. An all-atom variational autoencoder maps peptide geometry reversibly to a low-dimensional latent point cloud; a diffusion model in that latent space turns Gaussian noise into a point cloud conditioned on the binding site, jointly generating E(3)-invariant and E(3)-equivariant features that the decoder converts back to residue types and coordinates; and a contrastively trained latent interface encoder supplies the mimicry gradient. The latent space is chemically meaningful — amino acid representations cluster perfectly by identity, and their pairwise distances correlate with BLOSUM62 substitution scores at r = 0.47. Training uses PepBench protein-peptide complexes of 4-25 residues, with the ProtFrag set of peptide-like fragments from monomer structures pretraining the diffusion model before fine-tuning.
On a non-redundant benchmark of 93 complexes, with targets above 40% sequence identity to training removed, PepMimic beat RFdiffusion and the retrained co-design models HSRN and dyMEAN on Rosetta interface energy, Cα RMSD, all-atom DockQ, and amino acid recovery. On peptides with no secondary structure it recovered 40% of reference residues against 33% for RFdiffusion, and its amino acid composition tracked the reference far more closely (KL 0.06, versus 0.38 for RFdiffusion and 0.89 for dyMEAN). For PD-L1, CD38, BCMA, HER2, and CD4, the top 384 designs — ranked by Rosetta, FoldX, and AlphaFold-Multimer — were synthesised and measured by surface plasmon resonance imaging. Counting a hit as a dissociation constant below 100 nM, success rates exceeded 10% for CD38, HER2, and CD4; across the campaign 8% of peptides reached 10⁻⁸ M and 26 peptides reached 10⁻⁹ M, substantially above a random library screen run under identical conditions. FITC-labelled peptides injected by tail vein into breast, myeloma, and lung tumour mouse models accumulated at tumour sites.
The method suits any programme holding a structure of the target bound to something and wanting a peptide instead: a synthetic imaging agent, a tumour-homing carrier for a payload, or a competitive inhibitor of a protein-protein interaction. The mouse imaging results point most directly at diagnostic tracers and targeted delivery. Code is MIT-licensed, with checkpoints in GitHub releases and a Colab notebook for inference; the checkpoint archive carries no separate licence, and the FoldX evaluation step requires its own academic licence.
PepMimic's contribution is a design principle as much as an architecture: existing binders are a dense source of interface knowledge that latent-space guidance can transfer across molecular scales. The wet-lab campaign is unusually large for a peptide design paper — five targets plus a de novo case, hundreds of synthesised candidates each, and in vivo follow-up — which makes the reported hit rates more informative than the in-silico benchmarks. The main constraint is inherited from the premise: mimicry quality depends on a compact reference interface, and the paper's own analysis shows that interfaces with scattered residues, or dominated by beta strands, yield peptides that score well computationally but often fail experimentally.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.