This repository contains a method to generate 3D conformer ensembles directly from the molecular graph as described in our paper.

• python (version>=3.7.9)
• pytorch (version>=1.7.0)
• rdkit (version>=2020.03.2)
• pytorch-geometric (version>=1.6.3)
• networkx (version>=2.5.1)
• pot (version>=0.7.0)

Download and extract the GEOM dataset from the original source:

1. wget https://dataverse.harvard.edu/api/access/datafile/4327252
2. tar -xvf 4327252

Run make conda_env to create the conda environment. The script will request you to enter one of the supported CUDA versions listed here. The script uses this CUDA version to install PyTorch and PyTorch Geometric. Alternatively, you could manually follow the steps to install PyTorch Geometric here.

This should result in two different directories, one for each half of GEOM. You should place the qm9 conformers directory in the data/QM9/ directory and do the same for the drugs directory. This is all you need to train the model:

python train.py --data_dir data/QM9/qm9/ --split_path data/QM9/splits/split0.npy --log_dir ./test_run --n_epochs 250 --dataset qm9

Use the provided script to generate conformers. The test_csv arg should be a csv file with SMILES in the first column, and the number of conformers you want to generate in the second column. This will output a compressed dictionary of rdkit mols in the trained_model_dir directory (unless you provide the out arg):

python generate_confs.py --trained_model_dir trained_models/qm9/ --test_csv data/QM9/test_smiles.csv --dataset qm9

However, note that to reproduce the numbers in our paper, one needs additionally to run scripts/clean_smiles.py to account for inconsistent molecules in the dataset. See also count_geomol_failures.ipynb . You can use the provided visualize_confs.ipynb jupyter notebook to visualize the generated conformers.

To train the model, our code randomly samples files from the GEOM dataset and randomly samples conformers within those files. This is a lot of file I/O, which wasn't a huge issue for us when training, but could be an issue for others. If you're having issues with this, feel free to reach out, and I can help you reconfigure the code.

Currently, the model is hardcoded for atoms with a max of 4 neighbors. Since the dataset we train on didn't have atoms with more than 4 neighbors, we made this choice to speed up the code. In principle, the code can be adapted for something like a pentavalent phosphorus, but this wasn't a priority for us.

We can't deal with disconnected fragments (i.e. there is a "." in the SMILES).

This code will work poorly for macrocycles.

To ensure correct predictions, ALL tetrahedral chiral centers must be specified. There's probably a way to automate the specification of "rigid" chiral centers (e.g. in a fused ring), which I'll hopefully figure out soon, but I'm grad student with limited time :(

Code like this doesn't improve without feedback from the community. If you have comments/suggestions, please reach out to us! We're always happy to chat and provide input on how you can take this method to the next level.