MolMix

Official repository for "MolMix: A Simple Yet Effective Baseline for Multimodal Molecular Representation Learning", accepted at the Machine Learning for Structural Biology Workshop, NeurIPS 2024.

Installation

To install the required dependencies, use the following commands:

conda create -n molmix python=3.10
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118
pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.4.0+cu118.html
pip install ogb rdkit schedulefree wandb multimethod matplotlib
pip install flash-attn --no-build-isolation

MARCEL Experiments

To run the MARCEL experiments:

1. Download the Drugs.zip and Kraken.zip archives from the official MARCEL repo.
2. Place them in the following locations:
   • ./datasets/drugs/raw/Drugs.zip
   • ./datasets/kraken/raw/Kraken.zip

Running Experiments

Experiment configuration files are available in the ./cfgs folder. To run an experiment using a specific configuration file, execute the following:

python main.py --cfg ./cfgs/THE_SPECIFIC_CONFIG.yaml

Codebase

This repository includes modified implementations from:

• GemNet
• SchNet

Additionally, various parts of the code are adapted from the MARCEL repository.

Citation

If you use this work, please cite the following:

@misc{manolache2024molmix,
      title={MolMix: A Simple Yet Effective Baseline for Multimodal Molecular Representation Learning}, 
      author={Andrei Manolache and Dragos Tantaru and Mathias Niepert},
      year={2024},
      booktitle={Machine Learning for Structural Biology Workshop, NeurIPS 2024},
      url={https://arxiv.org/abs/2410.07981}, 
}

Contact

If you have any issues with running the code or the implementation, please either open a GitHub issue, or contact me at

echo "manogm.irli.tto-nurei.stturtttdg.de" | sed 's/-/@/1; s/nurei.stturtttdg/ki.uni-stuttgart/; s/manogm.irli.tto/andrei.manolache/'