Official pytorch implementation of TiDE-Ab (ECCB 2026)

---

Title:       De Novo Epitope-Specific Antibody Design via Time-Dependent Guidance [Paper]

Authors:   Yunji Kim, Minkyung Baek

Abstract

We propose TiDE-Ab, a conditional SE(3) flow matching framework for de novo epitope-specific antibody design. By conditioning on unpaired antigen and antibody structures without any pre-aligned frame, TiDE-Ab inherits the structural stability of flow matching while enabling global binding pose search from scratch. To further improve epitope targeting, we introduce Time-Dependent Classifier-Free Guidance (TD-CFG), which replaces static conditioning with an adaptive schedule: strong guidance early to establish the global binding pose, followed by gradual relaxation for precise local CDR refinement. On 55 non-redundant benchmark complexes, TiDE-Ab outperforms RFantibody with higher epitope recall (0.935 vs. 0.878) and over 95% fewer steric clashes. In therapeutic case studies on TGF-β and IL-17A, TiDE-Ab reproduced the binding profiles of clinical antibodies across isoform-selective and cross-reactive epitopes, whereas RFantibody consistently failed to produce viable candidates.

---

💊 Therapeutic Case Studies

TiDE-Ab demonstrates high practical utility by navigating complex therapeutic design scenarios that require programmable binding specificity. By successfully targeting a wide range of user-defined epitopes on the same antigen —from highly selective patches to broad conserved surfaces— TiDE-Ab ensures generated backbones strictly adhere to precise geometric constraints, transforming de novo design into a highly controllable and programmable engineering process.

Case Study 1: TGF-β TGF-β3 Selective vs. Pan-TGF-β

Case Study 2: IL-17A IL-17A/F Dual vs. IL-17A Selective

⚙️ Installation

Select an installation method based on your environment. Conda is highly recommended for a clean setup. These guides assume you are using CUDA 12.4.

Option 1: Quick Install (Conda)

Best if your system supports CUDA 12.4. This sets up everything in one command.

# Create and activate conda environment
conda env create -f environment.yml
conda activate tideab

Option 2: Step-by-Step Install (Pip)

Best for custom setups or different CUDA versions.

1. Environment Setup

conda create -n tideab python=3.10 -y
conda activate tideab

2. Install PyTorch & Specialized Packages (Change cu124 to cu118 or other versions if necessary)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
pip install torch-scatter -f https://data.pyg.org/whl/torch-$(python -c "import torch; print(torch.__version__.split('+')[0])")+cu124.html

3. Install Remaining Dependencies

pip install -r requirements.txt

📂 Data Preparation

The model is trained on antibody-antigen complexes from the SAbDab database (cutoff: April 30, 2020). The dataset is managed through metadata files in data/splits/.

├── data
│   └── splits
│       ├── metadata_train.csv
│       ├── metadata_val.csv
│       └── metadata_test.csv

The metadata files in data/splits/ follow this schema:

Column	Description
pdb_name	Unique identifier for the complex (e.g., 1yy9_D_C_A).
num_chains	Total number of chains in the structure.
seq_len	Total sequence length of the complex.
cluster	Interaction cluster ID used for balanced sampling.

🚀 Running the Code

Training

Run with default settings or override parameters via command line:

python train.py experiment.optimizer.lr=0.0005  # (Optional) parameter override

Inference

1. Download Pre-trained Weights Download weights.pt and place it in any directory of your choice (e.g., checkpoints/).

2. Run Inference on Test Set Provide the path to your weights and the desired output directory as arguments:

python inference.py weight_path=<PATH_TO_WEIGHTS> inference_dir=<OUTPUT_DIR>

Demo

We provide an easy‑to‑run demo for epitope‑specific antibody design in the run_demo.ipynb.
In this demo, you are required to provide a PDB file containing your target antigen. The antibody design starts from the trastuzumab Fv fragment, from which the model resamples the CDR loops while preserving the framework structure, enabling de novo redesign of antibodies that bind the antigen in diverse binding modes.

🙏 Acknowledgements

This codebase is developed based on the FrameFlow repository. We thank the original authors for their pioneering work on $SE(3)$ flow matching for protein structures.

This project is licensed under the MIT License - see the LICENSE file for details.