DMol

Environment installation

This code was tested with PyTorch 2.0.1, cuda 11.8 and torch_geometrics 2.3.1

• Download anaconda/miniconda if needed

• Create a rdkit environment that directly contains rdkit:

  conda create -c conda-forge -n dmol rdkit=2023.03.2 python=3.9

• conda activate dmol

• Check that this line does not return an error:

  python3 -c 'from rdkit import Chem'

• Install graph-tool (https://graph-tool.skewed.de/):

  conda install -c conda-forge graph-tool=2.45

• Check that this line does not return an error:

  python3 -c 'import graph_tool as gt'

• Install the nvcc drivers for your cuda version. For example:

  conda install -c "nvidia/label/cuda-11.8.0" cuda

• Install a corresponding version of pytorch, for example:

  pip3 install torch==2.0.1 --index-url https://download.pytorch.org/whl/cu118

• Install other packages using the requirement file:

  pip install -r requirements.txt

• Navigate to the ./src/analysis/orca directory and compile orca.cpp:

  g++ -O2 -std=c++11 -o orca orca.cpp

Note: graph_tool and torch_geometric currently seem to conflict on MacOS, I have not solved this issue yet.

Run the code

• All code is currently launched through python3 main.py. Check hydra documentation (https://hydra.cc/) for overriding default parameters.

• To run the discrete model with the default QM9 configuration:

  python3 main.py

• To select a different dataset, override the dataset.name parameter with Hydra. For example, to use Guacamol:

  python3 main.py dataset=guacamol

  See the YAML files under configs/dataset for the list of supported datasets and their options.

Prepare datasets and ring compression configuration

Before running any experiment, you must:

• 1. Download raw datasets

  • QM9
    • The code will automatically download from the PyG/DeepChem URLs if files are missing (see QM9Dataset.raw_url and QM9Dataset.raw_url2 in src/datasets/qm9_dataset.py):
      • https://deepchemdata.s3-us-west-1.amazonaws.com/datasets/molnet_publish/qm9.zip
      • https://ndownloader.figshare.com/files/3195404
    • If you prefer manual download, place the extracted files under
      data/qm9/raw/ (e.g. gdb9.sdf, gdb9.sdf.csv, uncharacterized.txt, train.csv, val.csv, test.csv).
  • MOSES
    • Automatic download URLs are defined in MOSESDataset (src/datasets/moses_dataset.py):
      • Train: https://media.githubusercontent.com/media/molecularsets/moses/master/data/train.csv
      • Test: https://media.githubusercontent.com/media/molecularsets/moses/master/data/test.csv
      • Scaffolds: https://media.githubusercontent.com/media/molecularsets/moses/master/data/test_scaffolds.csv
    • To download manually, save them as train_moses.csv, val_moses.csv, test_moses.csv under
      data/moses/raw/.
  • Guacamol
    • Download URLs are defined in GuacamolDataset (src/datasets/guacamol_dataset.py):
      • Train: https://figshare.com/ndownloader/files/13612760
      • Test: https://figshare.com/ndownloader/files/13612757
      • Valid: https://figshare.com/ndownloader/files/13612766
      • (All: https://figshare.com/ndownloader/files/13612745, used only if you need the full set)
    • Save them as guacamol_v1_train.smiles, guacamol_v1_valid.smiles, guacamol_v1_test.smiles under
      data/guacamol/raw/.
  • Make sure the folder structure matches what is expected in the corresponding dataset classes under src/datasets/.

• 2. Collect ring statistics (optional but recommended)

  • Use ring_statistics.py at the project root to analyze which ring structures are most frequent in your dataset:

    # Example: compute ring statistics for a SMILES file
    python3 ring_statistics.py data/qm9/raw/qm9_train.smiles

  • The script will output a text file (e.g. *_ring_stats.txt) containing all observed rings and their frequencies.

  • Inspect this file and decide which ring SMILES you want to compress into supernodes (e.g. C1CCC1, C1CC1, N1CCC1).

• 3. Configure ring compression in the dataset config

  • Open the corresponding YAML file under configs/dataset/, for example:
    • QM9: configs/dataset/qm9.yaml
    • MOSES: configs/dataset/moses.yaml
    • Guacamol: configs/dataset/guacamol.yaml
  • In each of these files there is a ring compression section with the keys:
    • ring_types: list of ring SMILES that will be compressed into supernodes.
    • ring_weights: list of molecular weights (one per entry in ring_types), used to compute total molecular mass after compression.
    • statistics_after_4 (or statistics_after_compression): path to the JSON file where compressed-dataset statistics are stored.
  • Replace the default ring_types and ring_weights with the ring structures you selected from the statistics file, making sure the lengths of both lists match.
  • After changing these parameters, re-run the dataset preprocessing (by running python3 main.py once) so that the compressed graphs and updated statistics are recomputed.

In summary, the full workflow is:

1. Download the raw dataset files into data/<dataset_name>/raw/.
2. (Optional but recommended) Run ring_statistics.py on a SMILES file from that dataset to obtain ring frequency statistics.
3. Choose a small set of ring SMILES to compress and put them into the ring_types and ring_weights fields of the corresponding YAML config under configs/dataset/.
4. Launch experiments with python3 main.py (and override dataset=... if you want a dataset other than the default QM9).