Scripts and notebooks shown here are used to generate the dataset and plots for the work in [pkasolver]
All functions, scripts and notebooks must be run from an environment with an installed version of pkasolver and its dependencies.
The scripts in the scripts folder can be used to generate the full dataset.
The pKa values for the ChEMBL dataset are estimated using Epik.
For each molecule one microstate per protonated and deprotonated state and a corresponding pKa is suggested by Epik and used for training without considering any other tautomers.
-
prepare_chembl_data.sh--contains the pipeline to download molecular data from the CHEMBL-database, predict its pka-data with Schrödinger's Epik and finally convert it the pytorch geometric graph data, to be used in the initial training. Epik is a closed-source programm for pKa-prediction and requires an active license. Feel free to replace Epik with any other tool for pKa-prediction. Be aware that this probably requires you to change some parts of the scripts, as other tools will probably not use the same output format as epic, and also you probably won't need to convert the data to the maesto (.mae) format. -
prepare_test_data.sh-- prepares the Novartis and Literature test data. -
training_model.sh--script for training and fine tuning a model with the ChEMBL dataset and the experimental dataset.
The plotting.ipynb Jupyter notebook can be used to generate all plots shown in /plots.
To do so you first have to unzip 04_chembl_dataset_filtered.sdf.gz and then run
python 04_1_split_epik_output.py --input ../04_chembl_dataset_filtered.sdf.gz --output ../04_chembl_dataset_pyg.pkl
and
python 05_data_preprocess.py --input ../04_chembl_dataset_pyg.pkl --output ../05_chembl_dataset_pyg.pkl.
This generates all the data needed to rerun the notebook.
00_download_mols_from_chembl.py:
--input: None
--output: path to output file (sdf.gz, sdf)
--filters the molecules of the chembl database by the specified criteria (e.g. max number of rule of five violation = 1) and outputs them to a gzipped sdf file.
01_convert_sdf_to_mae.py
--input: path to input file (sdf.gz, sdf)
--output: path to output file (mae.gz, mae)
--takes sdf file (can be gzipped) and converts it to Schrödinger maestro (mae) file. Schrödinger "ligprep" CLI-binary must be installed and path must be specified inside the script.
02_predict_pka_with_epik.py
--input: path to input file (mae.gz, mae)
--output: path to output file (mae.gz, mae)
--takes molecules from Schrödinger maestro (mae) file and returns new mae file containing Epik pka prediction data for each molecule. Schrödinger "Epik" CLI-binary must be installed and path must be specified inside the script.
03_convert_mae_to_sdf.py
--input: path to input file (mae.gz, mae)
--output: path to output file (sdf.gz, sdf)
--takes Schröndiger maestro (mae) file (can be gzipped) and converts it to sdf file. Schrödinger "convert" CLI-binary must be installed and path must be specified inside the script.
04_0_filter_testmols.py
--input: path to input file (mae.gz, mae)
--output: path to output file (sdf.gz, sdf)
--filter: path to output file (sdf, sdf.gz)
--takes sdf file of initial training molecules and sdf file of training molecules (both optionally gzipped) and returns only those initial training molecules not contained in the training molecules file as sdf file.
04_1_split_epik_output.py
--input: path to input file (sdf.gz, sdf)
--output: path to output file (pkl)
--takes sdf file with molecules containing Epik pka predictions in their properties and outputs a new sdf where those molecules containing more than one pka get duplicated so that every molecules only contains one pka value. The molecule associated with each pka is the protonated form of the respective pka reaction
04_2_prepare_rest.py
--input: path to input file (sdf.gz, sdf)
--output: path to output file (pkl)
--takes sdf of molecule set containing pka data and returns it as a pkl file.
05_data_preprocess.py
--input: path to input file (pkl)
--output: path to output file (pkl)
--takes pkl file of molecules containing pka data and returns pytorch geometric graph data containing protonated and deprotonated graphs for every pka
06_training.py
--input: set of training molecules as pyg graphs (pkl)
--output: path to output file (pkl)
--model_name: name of model architecture used (string)
Optional parameters: --model: path for saving model or containing model for retraining (pkl) --val: set of validation molecules as pyg graphs (pkl) --epochs: set number of training epochs (default == 1000) --reg: regularization dataset -r: flag for retraining model at path given by --model
--takes training set as pkl file and trains new model or retrains existing one.
This project is licensed under the MIT License - see the LICENSE.md file for details.
Copyright (c) 2021, Fritz Mayr, Oliver Wieder, Marcus Wieder, Thierry Langer