Train and evaluate espaloma

This repository includes scripts to retrain and generate espaloma-0.3.0 forcefield. espaloma-0.3.0 force field is a Class I force field where the valence parameters are assigned and optimized via machine learning framework. This repository is part of espaloma-0.3.0-manuscript.

Note that there is a refactored repository to train espaloma (espfit) and an example workspace (espfit_workspace) to use espfit. However, please be aware that these repositories are still under development.

Description

We first convert the HDF5 files obtained from download-qca-dataset to DGL graphs. Here, we compute the AM1BCC-ELF10 partial charges using OpenEye toolkit as a reference.

Molecules with a gap between minimum and maximum energy larger than 0.1 Hartree (62.5 kcal/mol) are excluded from the dataset prior to the refitting experiment, similar to the original paper of Wang et al.. Since the van der Waals parameters affect the physical property prediction, which is computationally challenging to optimize, we focus on optimizing the valence parameters and use openff-2.0.0 force field (details can be found here) for the van der Waals paremeters.

Espaloma was trained to minimize the quantum mechanics energies and forces, and also applied L2 regularization to improper and proper torsion force contants. The electronegativity and hardness of atoms were predicted to determine the atomic partial charges, following the same protocol described in the original paper by Wang et al., which used the AM1BCC-ELF10 partial charges as a reference.

Manifest

• openff-default/
  • 01-create-dataset/ - Convert HDF5 files into DGL graphs
    • script/ - Stores scripts to convert HDF5 files into DGL graphs
    • Dataset/ - Collection of Dataset from QCArchive
      • spice-des-monomers/
      • spice-dipeptide/
      • spice-pubchem/
      • rna-diverse/
      • rna-trincleotide/
      • rna-nucleoside/
    • OptimizationDataset/ - Collection of OptimizationDataset from QCArchive
      • gen2/
      • pepconf-dlc/
    • TorsionDriveDataset/ - Collection of TorsionDriveDataset from QCArchive
      • gen2-torsion/
      • protein-torsion/
  • 02-train/ - Refit and evaluate espaloma
    • baseline/ - Scripts used to calculate baseline energies and forces using other forcefields
    • joint-improper-charge/charge-weight-1.0/ - Scripts used to train and evaluate espaloma
    • merge-data/ - Scripts used to preprocess dgl graphs prior to training
• envs/ - Stores conda environment files
  • environment-create-dataset.yaml - Conda environment used to convert HDF5 files into DGL graphs in 01-create-dataset/
  • environment-refit.yaml - Conda environment to train and evaluate espaloma in 02-train/

Note

Please refer here to find more details about the actual origin of the dataset described above.

Dependencies

Espaloma ver. 0.3.0 was used to create the DGL graphs in 01-create-dataset/. Note that version 0.3.0 is no longer compatible with the 0.2.x models, and vice versa. A fixed version of 0.3.0 (commit hash:4c6155b72d00ce0190b3cb551e7e59f0adc33a56) was used for the refitting experinment and model evaluation which allows improper torsions to be fit to n=1,2 phase multiplicity.

Note

For a quick start, the preprocessed data in openff-default/02-train/merge-data/ is available here on Zenodo for training espaloma-0.3.0.

Citation

If you find this helpful please cite the following:

@misc{takaba2023machinelearned,
      title={Machine-learned molecular mechanics force field for the simulation of protein-ligand systems and beyond}, 
      author={Kenichiro Takaba, Iván Pulido, Pavan Kumar Behara, Chapin E. Cavender, Anika J. Friedman, Michael M. Henry, Hugo MacDermott Opeskin, Christopher R. Iacovella, Arnav M. Nagle, Alexander Matthew Payne, Michael R. Shirts, David L. Mobley, John D. Chodera, Yuanqing Wang},
      year={2023},
      eprint={2307.07085},
      archivePrefix={arXiv},
      primaryClass={physics.chem-ph}
}