zeobind

This repository contains code for the computational screening workflow used in the paper "Exhaustive mapping of the zeolite-template chemical space" by Xie et. al. If you use our code and data, please cite our cite our manuscript:

@misc{xie_exhaustive_2024,
	title = {An exhaustive mapping of zeolite-template chemical space},
	copyright = {https://creativecommons.org/licenses/by-nc/4.0/},
	url = {https://chemrxiv.org/engage/chemrxiv/article-details/66f8658812ff75c3a1cb235d},
	doi = {10.26434/chemrxiv-2024-d74sw},
	language = {en},
	urldate = {2024-10-10},
	author = {Xie, Mingrou and Schwalbe-Koda, Daniel and Semanate-Esquivel, Yolanda M. and Bello-Jurado, Estafanía and Hoffman, Alexander and Santiago-Reyes, Omar and Paris, Cecilia and Moliner, Manuel and Gómez-Bombarelli, Rafael},
	month = oct,
	year = {2024},
}

Installation

To install the code, clone the repository and install the dependencies using the following commands:

mamba create -f environment.yml
mamba activate zeobind

For Apple Silicon systems, install mamba first with homebrew, then run bash environment_m1_install.sh. Note that there might be some libraries with different versions as this repo was originally developed on Linux. Furthermore, when running the bash scripts described below for various tasks, source ~/.zshrc should be added to the top of the bash script, and the device flag should be changed to mps.

The data used in the paper, dataset of hypothetical molecules, as well as predictions across all the IZA framework - hypothetical molecule chemical space, can be found at Materials Data Facility. Download the data and place it in the zeobind/data directory.

Usage

This repo can be used to screen molecules from pre-generated predictions over the entire hypothetical OSDA library and experimentally known zeolites, predict for new molecules, and train models for each of the three prediction tasks described in the manuscript. Each functionality is described below. To learn more about the available user-specified inputs, read the argparse portion of the python script that is called in the corresponding bash script.

Screening molecules from predictions

We provide an example bash script for screening for molecules for a targeted framework using the pre-generated predictions stored inside data/predictions.

Example bash scripts, which are also the ones used in the paper's CHA and ERI case studies, can be found at run_scripts/inference/screen_{cha,eri}.sh. The example script takes as input the target framework, desired charge of the molecule and down selection filters, and screens the already generated predicted binding energies of hypothetical molecules and saves the final pool of selected molecules. Run

bash run_scripts/inference/screen_cha.sh

Predicting on new molecules

If you have a new set of molecules you would like to predict on, an example bash script can be found at run_scripts/predict.sh. It takes as input the list of molecules and list of zeolites, generates conformers and computes descriptors based on them for each molecule, reads the already generated descriptors for the zeolite, and then predicts and saves the binding energies. Change the task flag depending on which of the three tasks you would like to predict for.

bash run_scripts/inference/predict.sh

Training models

To train models, example bash scripts can be found at run_scripts/training/. It requires the training dataset, data splits, molecule and zeolite input files, and model architecture and training parameters.

bash run_scripts/training/binary/template.sh

Reproducing results and figures

The notebooks in the notebooks directory generate metrics and figures reported in the paper.