ConfRank+: Extending Conformer Ranking to Charged Molecules

A High-Throughput Machine Learning Model for Charged Molecular Conformers

Preprint | Dataset | ConfRank 1.0

Installation

You can install the python code for running the ConfRankPlus model via pip:

pip3 install git+https://github.com/grimme-lab/confrankplus.git

It is highly recommended to install torch_cluster for significant faster computation of atomistic graphs, e.g., by running:

pip install git+https://github.com/rusty1s/pytorch_cluster.git

For further installation guidelines, see here.

Inference

Label xyz files from command line

You can label (multiple) xyz files (or other formats supported by the ase.io module) with the confrankplus command in your command line:

 confrankplus --files *.xyz --output_path confrank_output.xyz --total_charge 0  --fidelity r2SCAN-3c --batch_size 20

The results will be written to the path specified via --output_path. The units of the energies written to the output path will be in kcal/mol.

Furthermore, in the directory evaluate there are example scripts for running inference on the test datasets of our paper.

ASE Calculator

In addition to the confrankplus command line interface, we provide an ASE calculator.

Example:

import numpy as np

np.random.seed(0)
from ase.collections import s22
from ConfRankPlus.inference.calculator import ConfRankPlusCalculator

# load the ConfRankPlus calculator with r2SCAN-3c fidelity
calculator = ConfRankPlusCalculator.load_default(fidelity="r2SCAN-3c",
                                                 compute_forces=False)

molecule = s22['Adenine-thymine_Watson-Crick_complex']
molecule.set_calculator(calculator)
# energy before displacements:
energy_1 = molecule.get_potential_energy()
# Apply random displacements:
displacements = np.random.uniform(-0.01, 0.01, molecule.positions.shape)  # in Angstrom
molecule.set_positions(molecule.get_positions() + displacements)
# energy after displacement:
energy_2 = molecule.get_potential_energy()

# ASE uses eV as unit for energy:
print(f"The energy difference is {energy_2 - energy_1:.2E} eV.")

Note: our model has not been tested for geometry optimization or MD simulations.

Loading datasets

You can load the datasets from Zenodo as follows:

from ConfRankPlus.data.dataset import HDF5Dataset

filepath = ...  # path to .h5 file 
dataset = HDF5Dataset.from_hdf5(filepath=filepath, precision=64)

dataset will be an instance of the InMemoryDataset class from PyTorch Geometric.

Citation

When using or referring to ConfRankPlus, please cite it as follows:

@misc{ConfRankPlus25, 
author={Oerder, Rick and H{\"o}lzer , Christian and Hamaekers, Jan}, 
title={ConfRank+: Extending Conformer Ranking to Charged Molecules},
year={2025},
doi={10.26434/chemrxiv-2025-xkwk6}, 
note={ChemRxiv Preprint}}

and

@article{ConfRank24,
author = {H{\"o}lzer, Christian and Oerder, Rick and Grimme, Stefan and Hamaekers, Jan},
title = {ConfRank: Improving GFN-FF Conformer Ranking with Pairwise Training},
journal = {Journal of Chemical Information and Modeling},
volume = {64},
number = {23},
pages = {8909-8925},
year = {2024},
doi = {10.1021/acs.jcim.4c01524},
note ={PMID: 39565928},
URL = {https://doi.org/10.1021/acs.jcim.4c01524},
eprint = {https://doi.org/10.1021/acs.jcim.4c01524}}

License

The content of this repository is licensed under CC BY-NC-SA 4.0