Experimental workspace for espfit

This is a workspace to test and experiment espfit which is a refactored version of refit-espaloma to train and validate espaloma. Espaloma is an end-to-end differentibale graph neural network framework to parameterize classical molecular mechanics force fields.

Description

The motivation and final goal of this repository to refit espaloma using QC data and NMR experimental observables of RNA nucleosides to simulate RNA systems.

Manifest

• experiments/ - workspace for each espaloma model trained on different dataset and/or protocol
  • espaloma-0.3.2/ - experiments using espaloma-0.3.2 which is equivalent to the model created in Ref1.
  • openff-default/ - experiments using espaloma model trained with the same dataset as espaloma-0.3.2 but with espfit
  • spice-default/ - experiments using espaloma model trained with the SPICE-1.1.4 dataset
  • spice-openff-default/ - experiments using espaloma model trained with the SPICE dataset included in the espaloma-0.3.2 training dataset
  • spice2-default/ - experiments using espaloma model trained with the SPICE-2.0.0 dataset
• scripts/ - common scripts to run benchmark experiments
  • pl-benchmark/ - alchemical protein-ligand binding free energy calculations
  • rna-nucleoside/ - RNA nucleoside simulations

Note that espaloma-0.3.2, openff-default, and spice-openff-default uses QC data generated using B3LYP-D3BJ/DZVP level of theory whereas spice-default and spice2-default uses ωB97M-D3BJ/def2-TZVPPD. See Ref2 for the details about the impact of QM level of theory.

Reference

[1] Takaba, K et al., Machine-learned molecular mechanics force field for the simulation of protein-ligand systems and beyond, 2023, arXiv:2307.07085
[2] Behara, P. K. et al., Benchmarking QM theory for drug-like molecules to train force fields, 2022, OpenEye CUP XII, Santa Fe, NM. Zenodo