henhaobio
Sign in

Nature Biomedical EngineeringBiochemistryPeer-reviewed journal article

Peptide Design through Binding Interface Mimicry

Xiangzhe KongiD, Rui JiaoiD, Haowei LiniD, Ruihan GuoiD, Wenbing HuangiD, Wei‐Ying MaiD, Zihua WangiD, Yang LiuiD, Jianzhu MaiD

PDFDOIPreprintData
In-depth full-text digest available · Jump to digest ↓

Abstract

Peptides offer distinct advantages for targeted therapy, including oral bioavailability, cellular permeability, and high specificity, which set them apart from conventional small molecules and biologics. In this work, we developed an AI algorithm, named PepMimic, to transform a known protein receptor or an existing antibody of a target into a short peptide drug by mimicking the binding interfaces between targets and known binders. The structural root mean square deviation and interface DockQ with reference binders were 61% and 75% better than the best existing methods on the PepBench datasets. We then applied this novel peptide-design methodology to five drug targets: PD-L1, CD38, BCMA, HER2, and CD4. SPRi results show that 8% of the peptides exhibited dissociation constant (KD) values at the 10-8M level, and 26 peptides achieving KD values as low as 10-9M. This success rate was 20,000 times higher than that observed in a random library screening conducted under identical conditions. PepMimic was applied to target proteins lacking available binders by first utilizing AI algorithms to design protein binders, followed by the generation of peptides through simulation of these artificial interfaces. The top-ranked peptides underwent extensive cellular validation and in vivo testing through tail vein injections in breast, myeloma, and lung tumor mouse models. Experimental results demonstrated effective membrane binding and highlighted the strong potential of these peptides for clinical diagnostic imaging and targeted therapeutic applications.

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (21 Jan 2025) · CC BY 4.0

PepMimic, an all-atom latent diffusion model guided by a contrastive interface encoder, designs peptides mimicking known binders; 8% of tested peptides across five targets reached 10^-8M KD.

Designing peptide mimics of known binders normally needs expert knowledge of interaction patterns and handles discontinuous interface residues poorly. PepMimic learns interface patterns in a latent space and generalizes across reference interfaces without retraining, while modeling side-chain atoms explicitly.

Key results

  1. ShownOn a non-redundant peptide-binder test set, PepMimic beat RFDiffusion, HSRN and dyMEAN on RMSD, DockQ, AAR and interface energy, and its amino acid usage matched the reference distribution more closely.

    Abstract: RMSD and DockQ 61% and 75% better than best existing methods on PepBench. 93 test complexes, 40 samples per target. AAR 40% vs 33% for RFDiffusion on peptides without secondary structure (p ≤ 0.05). KL 0.06 vs 0.89 (dyMEAN), 0.38 (RFDiffusion). · Abstract; Performance on target-specific peptide design

  2. ShownLatent gradient guidance steers generated peptides toward the reference interface, and the Interface Hit metric tracks true residue overlap. Mimicry was harder for antibody and nanobody interfaces than for PD1 or a designed mini-protein.

    Interface hit vs overlapping residues Pearson 0.99 (P < 0.001). Guided vs unguided interface distances reduced (P < 0.01), notably after 50 of 100 diffusion steps, on four PD-L1 binding partners (PD1, mini-protein, nanobody, durvalumab). · Performance of interface-mimicking

  3. ShownSPRi screening of 384 designed peptides across five targets gave a much higher hit rate than random-library screening. Hit rate varied by target and was lowest for PD-L1 and BCMA.

    8% of peptides at the 10^-8M level and 26 peptides at 10^-9M. Success (KD below 100 nM) above 10% for CD38, HER2 and CD4. A random library gave 8 positive PD-L1 peptides; PepMimic success rate reported as 20,000 times higher. · Abstract; Results on Mimicking Existing Binders for Target…

  4. ShownPeptides mimicking RFDiffusion-generated mini-binders for CD38 and TROP2 also gave high-affinity hits, and for CD38 this AI-binder route typically gave higher success rates than natural reference binders.

    Success (KD below 100 nM) approx. 16% for CD38 and 8% for TROP2 (384 peptides: 290 CD38, 94 TROP2). 14 peptides at or stronger than the 10^-8M level. Random library gave 9 TROP2 peptides at 10^-8M; reported 90,000 times higher success rate. · Introduction; Results on Mimicking AI-Generated Binders; Ranking Strategy for SPRi Experiments

  5. SuggestedGallium-68-labeled PD-L1 and TROP2 peptides showed selective tumor uptake in dual-xenograft mice at 30 minutes, supporting imaging use for the two selected peptides.

    PD-L1 peptide RIWAYRKFMD (KD = 43.0 nM): 6.02 ± 0.4 vs 0.68 ± 0.1 %ID/cc, MDA-MB-231 vs A549, 8.85-fold. TROP2 peptide PDMWTEAQLRFL: 6.15 vs 0.38 %ID/cc, BxPC3 vs A549, 16.18-fold. · Results on Mimicking Existing Binders for Target…; Results on Mimicking AI-Generated Binders; In vivo imaging of tumor

Methods, data and reproducibility

An all-atom autoencoder maps peptide residues to latent point clouds, and a latent diffusion model generates them conditioned on the binding site. Gradients from a contrastively trained interface encoder steer generation toward a reference interface. Candidates were ranked with Rosetta, FoldX and AlphaFold Multimer, then tested by SPRi.

  • PDB protein-peptide complexes (before 8 Dec 2023) for training and validation, clustered at 40% receptor identity to avoid leakage · n = 6,105 unique complexes; 4,157 train and 114 validation after cluster filtering
  • Large Non-Redundant (LNR) protein-peptide test set for in silico benchmarking · n = 93 complexes
  • Unsupervised pre-training set of peptide-like fragments extracted from PDB monomers · n = 70,645 fragments
  • Reference binder complexes (antibodies, nanobodies, natural receptors) for PD-L1, CD38, BCMA, HER2, CD4; peptides tested by SPRi · n = 384 designed peptides
  • RFDiffusion mini-binders for CD38 and TROP2, filtered by AlphaFold Multimer interface pLDDT, used as mimicry templates · n = 12,000 candidates per target; 384 peptides tested
  • Random mix-and-split peptide library screened against PD-L1 and TROP2 as experimental baseline · n = 8 positive (PD-L1), 9 positive (TROP2)
  • Subcutaneous human cancer cell xenografts (MDA-MB-231, A549, BxPC3) for PET/CT and fluorescence imaging · mouse (NOD/SCID) bearing human cell line xenografts · n = n=3 per group in Methods; n=5 in figure legends
Methods & tools
PepMimic, RFDiffusion, HSRN, dyMEAN, Rosetta (REF2015, FastRelax), FoldX, AlphaFold Multimer, MMseqs2, FreeSASA, DSSP, Plexera PlexArray HT SPRi, microPET/CT (Mediso)
Data & accessions
https://www.rcsb.org/
Reproducibility gaps
Code is stated to be at github.com/kxz18/PepMimic, but the printed link has a stray space, so it is not listed as a URL. Datasets are said to be on Zenodo with no link or record ID printed. Hyperparameters are in Supplementary Notes.
Limitations (5)
  • Interface hits on strands rose in silico but likely reflect artifacts, since linear peptides rarely form stable strands alone. Success rates were lower for PD-L1 and BCMA, whose reference interfaces are richer in helices and strands. · stated by the authors
  • Antibody and nanobody interfaces, which are larger and more complex, were harder to mimic than PD1 or mini-protein interfaces. The flat PD-L1 binding site is also known to be hard for small binders. · stated by the authors
  • This work focused on imaging peptides for tumors; therapeutic peptides for cancer treatment are left to future work. · stated by the authors
  • In vivo validation covers only a few selected peptides in xenografts. Group size differs between Methods (n=3) and figure legends (n=5), and no statistical tests are given for tracer uptake. · noted in this summary
  • The 20,000 and 90,000 fold gains compare computationally ranked peptides with a bead-based random library screen. Methods give few details on how library hits were quantified, so the multipliers are hard to audit. · noted in this summary

Read this if you design peptide binders and want to see how interface-guided latent diffusion performed in silico, by SPRi against random-library screening, and in xenograft PET imaging.

Citations
30
FWCI
9.06
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMao '26Filius '25Bryant '22Wei '25Charih '22Zhang '25
64% semantic matchBioinformatics
Qingyi Mao, Silong Zhai, Sen Cao+6 more·Journal of Chemical Information and Modeling·Apr 4, 2026· 2 citations