PepMirror

PepMirror is a generative model for designing D-peptide binders—peptides built from mirror-image (D-) amino acids—that target native L-proteins. D-peptides are attractive therapeutic candidates because their non-natural chirality makes them resistant to proteases and less immunogenic, but the scarcity of experimentally determined hetero-chiral (D-peptide to L-protein) structures has made data-driven design of such binders largely unexplored. PepMirror addresses this gap by transferring knowledge learned from abundant homo-chiral (L–L) data to the hetero-chiral (D–L) design setting.

The key idea, introduced in a February 2026 arXiv preprint from a team including Yanyan Lan at Tsinghua University, is to inject axial (pseudo-)vector features alongside the standard E(3)-equivariant polar vector features used in geometric deep learning. Because chirality is encoded by the distinction between polar and axial vectors under reflection, this added representation lets a model trained only on L–L examples generalize to designing D-peptide binders without requiring native hetero-chiral training structures.

Implemented within a latent diffusion framework, PepMirror is presented by its authors as the first wet-lab validated generative AI for de novo design of D-peptide binders, offering a new perspective on how to handle molecular chirality in structure-based protein design.

#Key Features

• Cross-chirality generalization: Learns from homo-chiral L–L binding data and transfers to the hetero-chiral D–L design task, sidestepping the lack of native D-peptide–protein structures.
• Axial vector features: Augments E(3)-equivariant polar vectors with axial vectors so the network can represent and reason about chirality, the geometric property that distinguishes L- from D-amino acids.
• Latent diffusion design: Generates target-conditioned peptide binders through a latent diffusion process operating over the augmented geometric representation.
• Wet-lab validated: A designed 10-mer D-peptide against CD38 was experimentally confirmed with a dissociation constant of roughly 10 μM, drawn from a small set of designs.

#Technical Details

PepMirror combines a latent diffusion generative model with an E(3)-equivariant geometric backbone in which polar vector features are supplemented by axial vector features to capture chirality. In silico, the authors report that PepMirror achieves the strongest overall interface-quality metrics against a broad panel of recent peptide and binder design baselines, including RFdiffusion, DiffPepBuilder, PepFlow, D-Flow, PPFlow, PepBridge, PepGLAD, and UniMoMo variants, and shows the smallest performance degradation when moving from L- to D-peptide tasks. Experimentally, the team identified a D-binder against the cancer antigen CD38 (a 10-mer with KD ≈ 10 μM) among a small set of designs. Code is released on GitHub, with pretrained weights distributed via Zenodo.

#Applications

PepMirror targets therapeutic peptide discovery, where D-peptides are valued for their metabolic stability and reduced immunogenicity. By generating candidate D-peptide binders directly against a protein target of interest, it can support hit generation for hard-to-drug surfaces—such as the CD38 antigen used in validation—and inform protease-resistant peptide therapeutics and chemical-biology probes. More broadly, the axial-vector approach is relevant to any structure-based design setting where chirality matters.

#Impact

As the first reported wet-lab validated generative model for de novo D-peptide binder design, PepMirror extends machine-learning-based binder design into the previously underserved hetero-chiral regime. Its axial-vector formulation offers a general recipe for incorporating chirality into equivariant networks, which may influence future work on mirror-image molecules and stereochemically aware design. As a recent preprint, its benchmarks and the single validated CD38 binder will benefit from independent replication and broader experimental testing across additional targets.