PepBridge implements joint design of protein surface and backbone using a diffusion bridge model (NeurIPS 2025).

PepBridge is a novel framework for the joint design of protein surface and backbone structures. It leverages receptor surface geometry and biochemical properties to generate ligand structures that are both conformationally stable and chemically feasible. Starting from a receptor surface represented as a 3D point cloud, PepBridge employs a Denoising Diffusion Bridge Model (DDBM) to generate complementary ligand surfaces. A multi-modal diffusion model then predicts the corresponding backbone structures, while Shape-Frame Matching Networks ensure alignment between the surface geometry and the predicted backbone architecture. This integrated approach promotes both surface complementarity and structural plausibility in the design of peptide–receptor complexes.

We recommend using conda. From this folder (Pepbridge/):

conda env create -f environment.yml
conda activate pepbridge

# Optional: install extra wheels if needed for your CUDA/PyTorch

pip install torch-scatter -f https://data.pyg.org/whl/torch-2.0.0+cu117.html

pip install joblib lmdb easydict wandb

• PepBDB: http://huanglab.phys.hust.edu.cn/pepbdb/db/1cta_A/
• QBioLip: https://yanglab.qd.sdu.edu.cn/Q-BioLiP/Download

PepMerge/
├── 1a0n_A/
│   ├── peptide.pdb
│   ├── receptor.pdb
│   ├── pocket.pdb
│   ├── surface_1a0n_A_peptide.pdb.obj
│   └── surface_1a0n_A_pocket.pdb.obj
├── 1a1a_C/
└── ...
Process_Data/
├── names.txt
├── pep_pocket_train_surf_structure_cache.lmdb
└── pep_pocket_test_surf_structure_cache.lmdb

Use models_con/pep_dataloader.py to prepare inputs for training. Processed datasets and pretrained model weights are available here: link.

From this folder (Pepbridge/):

python train_pepbridge.py

From repo root:

python Pepbridge/train_pepbridge.py

Configuration lives in configs/learn_surf_angle.yaml (edit to change training settings).

Generate peptide structures from trained checkpoints:

1. Sample candidates

python inference_pepbridge.py

2. Reconstruct full PDBs from samples

python reconstruct.py

We appreciate the inspiration from DDBM, DiffAb, and PepFlow.