HTOCSP

This is a public repository that aims to automate the high-throughput organic crystals prediction by integrating various existing tools such as PyXtal, pyocse, RDKit, AmberTools, and CHARMM.

Python Setup

git clone this repository and then go to the root directory

conda env create -n htocsp conda activate htocsp

or UMA

conda env create -n htocsp-uma -f environment_UMA.yml
conda activate htocsp-uma
`

If you want to update the existing ost enviroment

conda activate htocsp
mamba env update --file environment.yml

CHARMM Setup

One can request a free academic version of CHARMM and then install it via the following commands. Note, make sure you compile charmm with the simplest option with qchem, openmm, quantum and colfft.

$ ./configure --without-mpi --without-qchem --without-openmm --without-quantum --without-colfft \
-D CMAKE_DISABLE_FIND_PACKAGE_OpenMP=TRUE \
-p ~/CHARMM
$ make -j 8
$ make install

After a few minutes, you should see the following messages

[ 0%] Built target charmm_c
[ 99%] Built target charmm_fortran
[100%] Built target charmm_cxx
[100%] Built target charmm
Install the project...
-- Install configuration: "Release"
Installing into $HOME/CHARMM
-- Up-to-date: $HOME/CHARMM/bin/charmm

Then add the path of charmm executable to your .bashrc file and source it.

export PATH=$HOME/CHARMM/bin:$PATH

To check if the installation is successful, go to /HTOCSP/tests/CHARMM and run the example:

$ charmm < charmm.in
NORMAL TERMINATION BY NORMAL STOP
MOST SEVERE WARNING WAS AT LEVEL 1
$$$$$ JOB ACCOUNTING INFORMATION $$$$$
ELAPSED TIME: 2.44 SECONDS
CPU TIME: 2.40 SECONDS

You should see quickly see the output of NORMAL TERMINATION.

MACE Setup

MACE setup does not require any additional installation beyond what is already included in the main htocsp environment. No extra configuration is needed—simply ensure you have completed the Python Setup and CHARMM Setup sections above.

UMA Setup

Create the UMA environment from the environment_UMA.yml file.

HuggingFace Token Login

This UMA workflow requires access to models hosted on HuggingFace, you'll need to authenticate with a HuggingFace token. Generate or retrieve your token from https://huggingface.co/settings/tokens and then log in:

conda activate htocsp-uma
python - << 'EOF'from huggingface_hub import loginlogin(token="YOUR_TOKEN_HERE")EOF

Replace YOUR_TOKEN_HERE with your actual HuggingFace API token. This will store your credentials locally for future use.

Quick Test

After the environment is correctly setup, you can run the follow script directly from your terminal. This will quickly run 2 generations of sampling with a total of 8 structures.

frompyxtal.optimizeimportWFS# Samplinggo=WFS(smiles="CC(=O)OC1=CC=CC=C1C(=O)O",
wdir="apirin-quick",
sg=[14],
tag='aspirin',
N_gen=2,
N_pop=4,
N_cpu=1,
ff_style='gaff',
)
go.run()

The output should look like the following

Method : Stochastic Width First Sampling
Generation: 2
Population: 4
Fraction : 0.60 0.40 0.00
Generation 0 starts
0 83 18.91 3.89 12.92 107.6 1 0 0.96 0.95 0.73 19.9 68.8 0.5 -111.4 -4.9 10.1 0 -85.745 Random 1 83 9.07 17.01 5.62 80.4 1 0 0.77 0.37 0.84 81.8 16.1 -56.3 -128.7 -4.1 -157.3 0 -87.927 Random 2 82 5.69 12.06 12.31 79.8 1 0 0.42 0.29 0.10 -5.6 -21.3 126.5 -65.3 -2.6 152.8 0 -88.225 Random 3 81 21.82 8.30 7.59 68.8 1 0 0.90 0.87 0.26 -102.3 -21.1 -147.9 90.1 0.0 -180.0 0 2500.000 Random Generation 0 finishes: 4 strucs
0 82 5.69 12.06 12.31 79.8 1 0 0.42 0.29 0.10 -5.6 -21.3 126.5 -65.3 -2.6 152.8 0 -88.225 Random Top
0 83 9.07 17.01 5.62 80.4 1 0 0.77 0.37 0.84 81.8 16.1 -56.3 -128.7 -4.1 -157.3 0 -87.927 Random Top
0 83 18.91 3.89 12.92 107.6 1 0 0.96 0.95 0.73 19.9 68.8 0.5 -111.4 -4.9 10.1 0 -85.745 Random Top
Gen 0 time usage: 47.7[Calc] 0.0[Proc]
Generation 1 starts
0 83 13.54 10.84 7.54 50.4 1 0 0.40 0.97 0.19 86.8 14.8 -105.8 -90.0 -0.0 180.0 0 2500.000 Random 1 82 14.33 8.12 7.21 98.3 1 0 0.51 0.06 0.27 92.9 16.8 1.7 155.8 9.4 160.8 0 -82.825 Random 2 83 13.30 5.61 11.48 96.0 1 0 0.99 0.05 0.73 -17.7 34.3 132.6 -120.7 -16.8 18.5 0 -89.025 Mutation
3 81 18.70 6.76 16.85 112.9 1 0 0.54 0.84 0.65 -61.9 62.4 -1.3 85.8 1.1 3.2 0 -77.307 Random Generation 1 finishes: 8 strucs
1 83 13.30 5.61 11.48 96.0 1 0 0.99 0.05 0.73 -17.7 34.3 132.6 -120.7 -16.8 18.5 0 -89.025 Mutation Top
1 82 14.33 8.12 7.21 98.3 1 0 0.51 0.06 0.27 92.9 16.8 1.7 155.8 9.4 160.8 0 -82.825 Random Top
1 81 18.70 6.76 16.85 112.9 1 0 0.54 0.84 0.65 -61.9 62.4 -1.3 85.8 1.1 3.2 0 -77.307 Random Top
Gen 1 time usage: 44.2[Calc] 0.0[Proc]

In this example, the structure ended with 2500.000 means an invalid structure. Make sure you don't see all structures ends up with 2500.000.

Productive Examples

Please ref to the examples folder to run more productive examples.

Citation

Zhu Q, Hattori S. (2024). Automated High-throughput Organic Crystal Structure Prediction via Population-based Sampling

@misc{zhu2024-htocsp,
title={Automated high-throughput organic crystal structure prediction via population-based sampling},
author={Zhu, Qiang and Hattori, Shinnosuke},
journal={Digital Discovery},
volume={4},
number={1},
pages={120--134},
year={2025},
publisher={Royal Society of Chemistry},
doi={10.1039/D4DD00264D},
url={https://doi.org/10.1039/D4DD00264D},
}

Contacts:

• Qiang Zhu (qzhu8@uncc.edu)
• Shinnosuke Hattori (shinnosuke.hattori@sony.com)