Code for: A Graph Neural Network Charge Model Targeting Accurate Electrostatic Properties of Organic Molecules
Authors: C. Adams, J.T. Horton, L. Wang, D.L. Mobley, D. W. Wright, D. J. Cole
Each subdirectory contains code and data related to a specific study:
-
benchmark_wf_analysis/
Scripts to run DFT and CCSD calculations, along with notebooks for data analysis and plotting. -
benchmarking_new_charge_models/
Code for generating training, validation, and test loss metrics for the new charge models. -
comparison_of_charge_models/
Tools to compare various charge models by evaluating partial charges, dipoles, and electrostatic potentials (ESPs). Includes scripts to generate QM ground truth data, apply charge models, and analyze results. -
compare_fda_drugs/
Comparison of half-polarised machine learning charges with AM1-BCC charges across FDA-approved drugs. -
conformer_test/
Code to evaluate how existing and new charge models capture changes in electronic properties across molecular conformations. -
exploring_the_dataset/
Scripts to explore chemical features in the dataset used for training, validation, and testing. -
medchem_explorations/
Exploratory analysis of partial charges from the new $NAGL_{MBIS} model for two biological targets. -
producing_esp_from_mbis/
Evaluation of the error introduced when reconstructing ESPs using the MBIS multipole expansion up to the quadrupole level. -
timing_test/
Benchmarking the runtime of the new$NAGL_{MBIS}$ model against AM1-BCC on carbon chains of increasing length.
Most scripts require a Conda environment called charge_model_env. You can install it using:
conda env create -n charge_model_env --file charge_model_env_minimal.ymlThen activate the environment and install the nagl-mbis model:
pip install git+https://github.com/bismuthadams1/nagl-mbis --no-build-isolationFor generating quantum mechanical data, we use QCFractal. Two environments are needed:
-
Database Environment
This environment is required to launch the QCArchive database. Setup instructions and the environment YAML can be found here:
QCArchive Admin Guide -
Psi4 Worker Environment
Needed to run Psi4 calculations and submit them to the database. The environment YAML is available at:
qcarchive-worker-openff-psi4-ddx.yaml
Notebooks involving ligand docking require the FeGrow software. Please follow the setup instructions provided in their repository to create a suitable environment.