The Impact of Conformer Quality on Learned Representations of Molecular Conformer Ensembles

Code and data for reproducing the preprint: Adams, K. and C.W. Coley, The impact of conformer quality on learned representations of molecular conformer ensembles, 2025.

Environment

environment.yml contains the full conda environment. The primary dependencies are torch-geometric==2.1.0 (& supporting packages like torch-scatter==2.0.9), pytorch=1.13.0, and rdkit=2022.03.2.

Downloading data and trained model weights

Before reproducing any results, you will need to download the datasets (target properties and conformers) from this Dropbox link: https://www.dropbox.com/scl/fo/afrzp54g2wlhi1a4l0mdh/AOK1rmK9foRIqQn9e86BAUI?rlkey=epk8o0rse2ino7tble9xnxwzk&st=jicowj7m&dl=0

This Dropbox also includes all the trained models and inference results that were reported in the preprint. Due to the numerous models trained and evaluated, the total size of this data is ~30GB.

Retraining

train_acids_benchmark.py contains the training script used to train all models. It takes as input a list of command-line parameters (parsed by argparse) that define each model version as well as the composition of the corresponding training data. Each model is trained separately with separate calls to train_acids_benchmark.py.

job_submissions.py contains the list of all parameters used for each model that is reported in the preprint. Calling job_submissions.py will submit multiple, independent jobs (through a SLURM system) for training each model.

Evaluations

evaluate_acids_benchmark.py contains the evaluation script used to evaluate each model.

Calling submissions_evaluations_benchmark.py will submit multiple, independent jobs (through a SLURM system) in order to evaluate each model that is reported in the preprint.

Analyzing results

The notebooks analysis_benchmark_20241109_figure{*}.ipynb analyze all evaluation results (downloadable from the above Dropbox link) to reproduce the main-text figures that are reported in the preprint.

License

This project is licensed under the MIT License -- see LICENSE file for details.

Citation

@misc{adams2025impactconformerqualitylearned,
      title={The impact of conformer quality on learned representations of molecular conformer ensembles}, 
      author={Keir Adams and Connor W. Coley},
      year={2025},
      eprint={2502.13220},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2502.13220}, 
}