PepBridge

Designing a peptide or mini-protein binder that docks against a specific target requires getting two things right at once: the shape of the binder's molecular surface must be complementary to the receptor interface, and the underlying backbone must fold into a structure that actually presents that surface. Most generative protein-design methods optimize backbone geometry or sequence directly and treat surface complementarity as an emergent afterthought. PepBridge inverts this framing by making the interface surface the primary design object and then recovering a consistent structure from it.

PepBridge, introduced in November 2025 by Guanlue Li, Xufeng Zhao, and Sören Laue at the University of Hamburg with Fang Wu at Stanford University, is a generative framework for the joint design of protein and peptide ligand surfaces and structures complementary to a target receptor. Its central idea is to use a denoising diffusion bridge model (DDBM) to learn a direct mapping between two coupled distributions, the receptor surface and the complementary ligand surface, rather than diffusing from pure noise. This lets the model generate a candidate binder surface that is conditioned on, and geometrically matched to, the surface it must bind.

The work positions itself alongside target-conditioned peptide and binder design methods such as DiffPepBuilder and RFdiffusion, but differs in treating molecular surface complementarity and hydrophobic packing at the interface as the explicit generative target. It was presented at NeurIPS 2025.

#Key Features

• Surface-first design via a diffusion bridge: A denoising diffusion bridge model interpolates between the paired receptor and ligand surface distributions, directly generating a complementary binder surface instead of sampling from unconditioned noise.
• Joint surface and backbone generation: Rather than designing surface and structure separately, PepBridge produces both together so that the predicted backbone is consistent with the designed interface surface.
• Multi-model structure diffusion: Backbone structure is recovered with a combination of an SE(3) diffusion model for frames, a torus diffusion model for torsion angles, and a logit-normal diffusion model for residue identity.
• Shape-Frame Matching Networks: Dedicated matching networks enforce alignment between the generated surface geometry and the backbone frames, bridging the surface and structure representations.
• Target-conditioned binder generation: Given a receptor surface, the model proposes structurally viable peptide or mini-protein ligands tailored to that specific interface.

#Technical Details

PepBridge couples a surface-generation stage with a structure-recovery stage. The surface stage uses a denoising diffusion bridge model, a diffusion variant that learns transport between two endpoint distributions, to map a target receptor surface to a complementary ligand surface while preserving physical and biochemical relevance at the interface. The structure stage is a multi-model diffusion system: an SE(3)-equivariant diffusion model generates backbone frames, a torus (wrapped) diffusion model handles backbone and side chain torsion angles, and a logit-normal diffusion model assigns residue identities. Shape-Frame Matching Networks tie the surface and backbone representations together so the recovered structure presents the designed surface. The model is trained on protein-peptide and protein-ligand complex data drawn from PepBDB and QBioLip, datasets of experimentally determined binding complexes that provide paired receptor-ligand surfaces for the bridge to learn from. Inference runs from a fixed released checkpoint via inference_pepbridge.py, followed by reconstruct.py to assemble full-atom PDB outputs.

#Applications

PepBridge targets de novo design of peptide and mini-protein binders against a chosen receptor, a core task in therapeutics and molecular tool development where surface complementarity at the interface largely determines binding. Computational biologists and protein engineers can use it to propose candidate binders conditioned on a target surface, generating structurally consistent backbones and sequences as starting points for downstream filtering, folding validation, and experimental testing. Its surface-centric formulation is particularly relevant when the binding mode is defined by interface shape and hydrophobic packing rather than by a known sequence motif.

#Impact

PepBridge contributes a distinctive surface-first perspective to the rapidly growing field of target-conditioned generative protein and peptide design, demonstrating that diffusion bridge models can learn the geometric coupling between a receptor surface and its complementary binder. By jointly generating surface and structure rather than treating them sequentially, it offers an alternative to backbone-first pipelines and was accepted at NeurIPS 2025. As a recent release, its real-world adoption and experimental validation remain to be established, and reported results to date are computational; the released code (MIT license) and checkpoint lower the barrier for the community to evaluate and build on the approach.