PROTAC-STAN

This is the official codebase of the paper: "Interpretable PROTAC degradation prediction with structure-informed deep ternary attention framework"

Overview

This study introduces PROTAC-STAN, a structure-informed deep ternary attention network (STAN) framework for interpretable PROTAC degradation prediction. It's the first study to coherently model the three-body interactions of PROTAC therapeutics with a tailored deep learning architecture. PROTAC-STAN represents PROTAC molecules across atom, molecule, and property hierarchies and incorporates structure information for POIs and E3 ligases using a protein language model infused with structural data. Furthermore, it simulates interactions among three entities via a novel ternary attention network tailored for the PROTAC system, providing unprecedented insights into the degradation mechanism.

Datasets

PROTAC-DB

The original data can be accessed at PROTAC-DB.

PROTAC-fine

We enrich degradation information to the PROTAC-DB 2.0 and construct a refined PROTAC dataset named PROTAC-fine. The data are stored in data/PROTAC-fine folder.

The data cleaning script is provided in clean.ipynb.

Directory instructions

Demo

.
├── config_demo.toml
├── data
│   └── demo
└── demo.ipynb

Training and inference

.
├── config.toml
├── data
├── data_loader.py
├── data.py
├── inference.py
├── main.py
├── model.py
├── saved_models
└── tan.py

Custom data preparation

.
├── data
│   └── custom
├── esm_embed
│   ├── get_embed_s.py
│   ├── model
│   └── README.md
└──  prepare_data.ipynb

Pretrained models

# PROTAC-STAN
.
└── saved_models
    └── protac-stan.pt

# ESM-S
.
└── esm_embed
    └── model
        ├── esm2_t33_650M_UR50D.pt
        └── esm_650m_s.pth

Download the ESM model weights esm2_t33_650M_UR50D.pt from here, and the structure-informed ESM model weights esm_650m_s.pth from here.

Demo

We provide PROTAC-STAN running demo through a Jupyter notebook demo.ipynb. Note it is based on a small demo dataset of PROTAC-fine. This demo only takes about 5 minutes to complete the whole pipeline. For running PROTAC-STAN on the full dataset, we advise GPU ram >= 8GB and CPU ram >= 16GB.

System requirements

PROTAC-STAN has been tested on Linux operating systems (Ubuntu 20.04.1).

Python Dependencies:

• Python (version >= 3.11.5)
• PyTorch (version >= 2.1.0)
• RDKit (version >= 2023.9.2)
• pyg (version >= 2.5.1)

The full requirements are provided in protac-stan.yml

Installation guide

It normally takes about 10 minutes to install on a normal desktop computer (based on your network).

1. Create Conda environment

conda create -n protac-stan python=3.11.5
conda activate protac-stan

2. Install Pytorch

conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=11.8 -c pytorch -c nvidia
conda install numpy==1.26.4

3. Install other essential packages

pip install torch_geometric==2.5.1
pip install rdkit==2023.9.2
pip install pandas==2.1.1
pip install toml==0.10.2
pip install wandb
# [Optional] Install torch-scatter to accelerate pyg computing
wget https://data.pyg.org/whl/torch-2.1.0%2Bcu118/torch_scatter-2.1.2%2Bpt21cu118-cp311-cp311-linux_x86_64.whl
pip install torch_scatter-2.1.2+pt21cu118-cp311-cp311-linux_x86_64.whl

4. Download repository

git clone https://github.com/PROTACs/PROTAC-STAN.git
cd PROTAC-STAN

Tip

See protac-stan.yml for full requriements.

Training to reproduce results

We have prepared the PROTAC-fine dataset in directory data/PROTAC-fine.

To train the PROTAC-STAN model from scratch, run the following script:

python main.py

Evaluation results of PROTAC-STAN and baselines on test set considering data leakage:

Inference on your data

inference.py leverage PROTAC-STAN as a powerful tool to perform interpretable PROTAC degradation prediction.

1. Prepare your customed data following prepare_data.ipynb
2. Predict your data:

# Usage: python inference.py [-h] [--root ROOT] [--name NAME] [--save_att]
python inference.py --root 'data/custom' --name 'custom'

You may use --save_att argument to save attention maps to take further anaysis, here are our examples:

Tip

You may use Python packages like matplotlib, RDKit, Visualization software like Maestro, PyMOL and so on.

3D and 2D attention map visualization

Molecule and complex visualization

We also provide a simple PyMOL script pymol_visualize.py to assist with 3D visualization, see here.

Citation

@article{chen2025Interpretable,
  title = {Interpretable {{PROTAC Degradation Prediction With Structure}}-{{Informed Deep Ternary Attention Framework}}},
  author = {Chen, Zhenglu and Gu, Chunbin and Tan, Shuoyan and Wang, Xiaorui and Li, Yuquan and He, Mutian and Lu, Ruiqiang and Sun, Shijia and Hsieh, Chang-Yu and Yao, Xiaojun and Liu, Huanxiang and Heng, Pheng-Ann},
  year = {2025},
  month = sep,
  journal = {Advanced Science},
  pages = {e08138},
  issn = {2198-3844, 2198-3844},
  doi = {10.1002/advs.202508138},
  langid = {english},
}