danisetiawan / templatecorr

Hierarchical template correction for chemical reactions

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

templatecorr

GitHub Actions Build Status

A Python package to hierarchically correct a list of reaction templates. See our publication On the influence of template size, canonicalization and exclusivity for retrosynthesis and reaction prediction applications for further information and benchmarks.

Installation

Download templatecorr from Github:

git clone https://github.com/hesther/templatecorr.git
cd templatecorr

Set up a conda environment (or install the packages in environment.yml in any other way convenient to you):

conda env create -f environment.yml
conda activate templatecorr

This environment uses the RDChiral C++ package to extract and canonicalize templates. If you instead want to use the older RDChiral Python package for backwards compatibility reasons, use the file environment_scripts.yml instead of environment.yml. Also, RDChiral C++ is currently not available on Windows, in this case, also use environment_scripts.yml.

Install the template corr package:

pip install -e .

Extract and correct templates from a file

To extract hierarchically corrected templates from a set of reactions, use correct.py. Prepare a csv, pkl/pickle of a pandas dataframe, json/json.gz or hdf5 of a pandas dataframe file of reactions. The file must at least contain one column with atom-mapped reaction SMILES. The column can have any name, default rxn_smiles, which can be specified via --reaction_column. Other information (other columns) in the files will be conserved. In the following we will use data/uspto_50k.csv (extract the archive data.tar.gz to access it). You can use the default arguments:

python correct.py --path data/uspto_50k

where --path specifies the path to the reaction file without file ending. This will create the file data/uspto_50k_corrected.csv, which now contains an additional column template, holding the extracted, canonicalized and corrected template. The above command is the same as

python correct.py --path data/uspto_50k --reaction_column rxn_smiles --name template --nproc 20 --drop_extra_cols --data_format csv

where --reaction_column rxn_smiles specifies the name of the column containing reaction SMILES, --name template sets the name of the column for the extracted templates in the output file (here to "template"), --nproc 20 parallelizes the program over 20 processes, --drop_extra_cols causes additional helper columns during extraction (canonical reactant SMILES, templates at radius 0 and 1) to be dropped before saving the dataframe to file, and --data_format csv specifies the input format of the data, as well as the output format.

Use to retrain a template relevance model

If you want to use the template correction code together with the template-relevance GitLab repository, there is a simple drop-in replacement: In your workflow, instead of using bin/process.py from the template-relevance repository, use temprel_scripts/process.py (same usage, same arguments). NEW: Optional additional parameters to specify the radius and presence of special groups (default --radius 1 with special groups. To not use special groups in the templates, use --no_special_groups).

Reproduce our study

If you want to reproduce the results of the publication On the influence of template size, canonicalization and exclusivity for retrosynthesis and reaction prediction applications, you need to create another conda environment (since the newest rdchiral version (C++) used above automatically canonicalizes templates). We will use both environments in the following.

conda deactivate
conda env create -f environment_scripts.yml
conda activate templatecorr_nocan
pip install -e .

Extract the archived file data.tar.gz if you have not already done so, go to the scripts folder and run the scripts in consecutive order (from 01 to 05). To reproduce the exact results from the manuscript, run script 01, 03, 04 and 05 with the templatecorr_nocan environment, and script 02 with the templatecorr environment. Since the canonicalization functionality is now per default used in the C++ rdchiral package, non-canonical templates can otherwise not be obtained easily.

AiZynthFinder models

AiZynthFinder template and policy model files are available in the folder aizynthfinder_models for the canonical-corrected template sets of this study (USPTO-50k and USPTO-460k).

Contact

For questions, feedback, concerns or wishes, contact Esther at eheid@mit.edu.

Copyright

Copyright (c) 2021, Esther Heid

Acknowledgements

Project based on the Computational Molecular Science Python Cookiecutter version 1.5.

About

Hierarchical template correction for chemical reactions

License:BSD 3-Clause "New" or "Revised" License


Languages

Language:Python 99.2%Language:Shell 0.8%