JUBioactivities

===================

This library helps in easily loading datasets into python for further use in machine learning.

A) Install in a conda environment

The library can be installed by downloading the latest release and installing it in a conda environment (note: rdkit has to be installed manually)

For this:

1. navigate into the unzipped folder and activate the environment

2. Installation then can be done via

pip install .

The library will be installed in the site-packages.

B) Without installation:

Just add the libraries directory to the PYTHONPATH and use it without installation

Dependencies are pandas, standardiser, scikit-learn and rdkit

Use of the library

Each datasource contains an init file which (in most cases) provides two functions:

read_data() function: imports the results with the InchiKey as Index and the activity/assay result as a column with the according name

read_structure() function: imports the structures (as rdkit molecules) in a table with the InchiKey as Index (Note although for the InChI Key generation structures are standardised, the read_structure() call returns non standardised structures so the user is able to apply their custom procedures)

To use a dataset import the read_structures() or the read_data() function from the respective folder and use the dataset. An example notebook can be found under docs/examples

An overview of the available datasets or endpoints can be found in docs/Datasets_Endpoints.csv In there all available datasets with the respective endpoint are stored. Additional information is given about the number of compounds per endpoint and dataset (if a dataset contains more than one endpoint).

Internal File processing

The processing of the original files is done as follows:

1. The Matrix of Assay results For the assay results the assay data is read in and converted to a standard format: Continuous values are kept as given, binary values are converted to 1 if the assay outcome is positive, 0 if negative and NA if not available. Multiclass assays are encoded as one hot labels. Standardised InChI Keys are used as index to be able to match the structures from the structure matrix with the respective labels.

2. The chemical structures For the table the structures are read in and converted to an RDKit Molecule. This molecule is standardised via the Aktinson standardiser(https://github.com/flatkinson/standardiser). Afterwards a standardised InChIKey is generated to serve as index. The SMILES are either used directly from the data source or are generated via RDKit from the available (non- SMILES) structure type.To keep the structure file aligned with the Assay results file molecules where the structure cannot be imported to RDKit were kept (currently 1212 structures) these have to be filtered via preprocessing.

Important note: The generated SMILES are not standardised to allow the user to apply a sandardisation workflow of their choice.

Datasources

The datasources used for the library are:

The following webpages were used to accumulate the datasets:

• http://www.cheminformatics.org/datasets/index.shtml

• http://cdb.ics.uci.edu/cgibin/LearningDatasetsWeb.py

• https://www.epa.gov/chemical-research/distributed-structure-searchable-toxicity-dsst

• https://qsardb.jrc.ec.europa.eu/qmrf

• http://qsardb.org/repository/

• https://www.nlm.nih.gov/databases/download/data_distrib_main.html

• https://wiki.nci.nih.gov/display/NCIDTPdata/DTP+NCI+Bulk+Data+for+Download

• https://eurl-ecvam.jrc.ec.europa.eu/databases

• ftp://ftp.ics.uci.edu/pub/baldig/learning

Publications which are not included in the above sources are:

Benchmark Data Set for in Silico Prediction of Ames Mutagenicity Katja Hansen, Sebastian Mika, Timon Schroeter, Andreas Sutter, Antonius ter Laak, Thomas Steger-Hartmann, Nikolaus Heinrich, and Klaus-Robert Müller Journal of Chemical Information and Modeling 2009 49 (9), 2077-2081 DOI: 10.1021/ci900161g

• Computational Models for Human and Animal Hepatotoxicity with a Global Ap- plication Scope Denis Mulliner, Friedemann Schmidt, Manuela Stolte, Hans-Peter Spirkl, Andreas Czich, and Alexander Amberg Chemical Research in Toxicology 2016 29 (5), 757-767 DOI: 10.1021/acs.chemrestox.5b00465

• Cheminformatics Analysis of Assertions Mined from Literature That Describe Drug-Induced Liver Injury in Different Species Denis Fourches, Julie C. Barnes, Nicola C. Day, Paul Bradley, Jane Z. Reed, and Alexander Tropsha Chemical Research in Toxicology 2010 23 (1), 171-183 DOI: 10.1021/tx900326k

• Mixed learning algorithms and features ensemble in hepatotoxicity prediction Chin Yee LiewYen Ching LimChun Wei Yap Comput Aided Mol Des (2011) 25: 855. https://doi.org/10.1007/s10822-011-9468-3

• In silico Prediction of Drug Induced Liver Toxicity Using Substructure Pattern Recognition Zhang, C. , Cheng, F. , Li, W. , Liu, G. , Lee, P. W. and Tang, Y. Method. Mol. Inf., 2016 35: 136-144. DOI: 10.1002/minf.201500055

All references can also be found in docs/references.bib

To cite this dataset use:

@misc{JUBioactivities, author = {Jennifer Hemmerich and Guenter Klambauer},
title = {JUBioactivities: Datasets for Toxicity prediction},
year = {2019},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/JenniferHemmerich/JUBioactivities/}}
}

Future changes:

Instead of standardiser use RDKit standardisation

Use custom one hot encoder (remove scikit-learn dependency)

Get stream redirection to work with Notebooks