ConGraSyn: A Conformation Enhanced Graph Attention Framework for Predicting Synergistic Drug Combinations

This repository contains the official implementation of our paper:
ConGraSyn is a scale conformation-aware framework for predicting synergistic drug combinations by integrating 3D-enhanced molecular graphs with PPI-informed and drug-induced cell line features.

1. Introduction

ConGraSyn represents molecules as heterogeneous molecular graphs and predicts drug combination synergy using graph neural networks.
It provides substructure-level attention and integrates multi-source data, including PPI and cell lines omics profiles.

Paper Link: Coming soon...

1.1 Features

This method explicitly embeds 3D atomic coordinates and interatomic distances in 2D molecular graphs, introduces a multiscale conformational learning (MCL) module to capture local and global structural semantics, and fuses with molecular fingerprints to supplement global structural information. At the same time, the pre-trained TranSiGen model is used to generate drug-induced transcriptome features, which are fused with baseline omics data to characterize the dynamic response of cells to drugs.

1.2 File Structure

text ConGraSyn/ ← Project root directory ├── setup.py ← Packaging and installation configuration ├── requirements.txt ← Dependency management ├── README.md ← Project description └── src/ └── ConGraSyn/ ← Python package (contains all core source code) ├── __init__.py ├── Datas ← Data folder ├── Model ← Model script file ├── PrepareData ← Data preprocessing script file ├── ProcessorData ← Data processing script file └── const.py ├── config.py ├── train.py └── utils.py

1.3 Citation

If you find this repository helpful, please cite our work:

```bibtex

```

2. Usage

2.1 Requirements

We recommend the following Python environment: ```bash

---- Core Deep Learning Framework ----

torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0

⚠ torch-scatter must match your PyTorch and CUDA version.

Manual installation is recommended (see notes below).

---- GNN Packages ----

torch-geometric==2.4.0 dgl==1.1.2 # or dgl==1.1.2+cu118 depending on your CUDA version

---- Chemistry Toolkit ----

rdkit==2022.9.5 # from conda or RDKit wheels

---- ML + Data Processing ----

scikit-learn>=1.2.0 numpy>=1.24.0 pandas>=1.3.0 scipy>=1.7.0

---- Optional Utilities ----

tqdm matplotlib ```

Install core dependencies using:

bash pip install -r requirements.txt

Notes on Specific Dependencies

⚠ torch-scatter torch-scatter requires a PyTorch- and CUDA-matching build. Use the following command to install a compatible version: bash pip install torch-scatter -f https://data.pyg.org/whl/torch-2.1.0+cu118.html You can find more options at: PyG Installation Guide

⚠ rdkit rdkit is not available on PyPI; it is recommended to install via conda: bash conda install -c rdkit rdkit==2022.9.5

2.2 Preprocessing

Before cloning the code, please download the data in our Datas folder at https://doi.org/10.5281/zenodo.16210069 and put it into Datas. You can also download the full code here.

To preprocess the drug combination dataset:

bash python prepare_data.py

Processed files will be saved to ConGraSyn\datas\processed.

2.3 Train the Model

To train the ConGraSyn model: bash python train.py Results will be saved to the ConGraSyn\datas\results directory.