D4: Distance Diffusion for a Truly Equivariant Molecular Design

This is the code which was used to run the experiments for the paper "D4: Distance Diffusion for a Truly Equivariant Molecular Design". The paper has been accepted at ESANN 2025 - European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Bruges, Belgium, 2025

Requirements

The code was tested with Python 3.11.7 and with CUDA 11.4.0. The requirements can be installed by executing

• Download anaconda/miniconda if needed
• Create a new environment through the given environment files with the following command:

  conda env create -f <env_file>.yml

  where <env_file> is the name of the environment file to use. It is possible to install dependencies for CPU with environment_cpu.yml or for GPU with environment_cuda.yml.
• Install this package with the following command:

  pip install -e .

  which will compile required cython and c++ code.

Running experiments

The experiments can be run with different modality, the basic command is:

    python main.py +preset=<preset> seed=<seed> mode=<mode>

where the value <preset> could be chosen between qm9_distance and gdb13_distance. After the <preset> it's possible to choose the <seed>, the modality of running <mode> between <eval>, <train+eval> and <train>. The default values are mode=eval and seed=0. All the parameters relative to the training and evaluation phase could be seen inside ./config/.

Datasets

QM9 will be downloaded automatically to a new directory ./datasets when running an experiment. Regarding GDB13 only a random subset of the entire dataset is used, for reproducibility, this could be found directly inside ./dataset/GDB13/raw/ .

Checkpoints

Checkpoints are present in a shared folder on google drive (https://drive.google.com/drive/folders/1SO7mIvRe7mZjPZ8WF14Ygo6pyK5OY3-_?usp=sharing). To run them they should be downloaded and placed in the ./checkpoints folder.