Easy-to-use command line virtual screening tool utilizing machine learning to build activity-based model given a set of input compounds. The model is built over features of molecular fragments features and uses Naive Bayes classifier coupled with likelihood ratio scoring to prioritize the compound library to be screened.
- Python 2.7 or 3.5
- RDkit
- PaDEL (only when you want to use it to generate fragments features; otherwise RDKit will be used)
git clone https://github.com/davidhoksza/bayescreen.git
Bayescreen consists of three utilities:
- Build a model of activity based on features of fragments on known actives and inactives
- Screen a compound library against a model
- Analyze a model
To carry out the screening, one needs a set of known active and inactive molecules and a set of molecules (either in SDF or SMILES format) to screen. The following commands show an example usage of Bayescreen to screen the CDK2 enzyme. The dataset is part of the collections from the lbvs-environment project, but for convenience the data of the CDK2 dataset are distribued together with Bayescreen. An archive with the SDF files with the molecules are available in the test directory. To unzip them, run:
tar xzvf test/cdk.tar.gz -C test
To build a model based on known actives stored in test/a_cdk_train.sdf and inactives stored in test/i_cdk_train.sdf using RDKit for generating fragment features and store it in model.bm run:
python build_model.py -a test/a_cdk_train.sdf -i test/i_cdk_train.sdf -m test/model.bm -c
The model.bm is a JSON file which stores the probabilities used to build the Naive Bayes classifier, imputation and normalization values for every feature and other informations about the model such as which feature generator was used (RDKit or PaDEL) or binning information.
If the fragments and corresponding features are needed, remove the -c (--clean) option.
Screening of the compounds ranks the compounds based on their probability of being active against the same target as the a_train.sdf molecules, i.e. the molecules on which the model was built.
In order to also evaluate the perfromance of Bayescreen we use the splits provided by the lbvs-environment. Thus we separately screen the a_test.sdf molecules and i_test.sdf corresponding to the active and inactive molecules, respectively (obviously, this division will not be known in real VS campaign). both of the
To screen the compounds run:
python screen.py -m test/model.bm -d test/a_cdk_test.sdf -o test/a_test.out
python screen.py -m test/model.bm -d test/i_cdk_test.sdf -o test/i_test.out
The x_test.out contains the screening library sorted by decreasing probability of those molecules being active (average probability ratios of their fragments feature vectors).
To evaluate the performance, we provide a utility which utilizes RDKit to obtain the area under the ROC curve (AUC) and enrichment factors (EF) on 0.5%, 1%, 2% and 5% of the database. To get the performance, run:
python evaluate.py -a test/a_test.out -i test/i_test.out
To analyze the model run:
python analyze_model.py -m test/model.bm -f
The output includes the following information:
- General model information (list features, used fragment types, feature generator)
- Importance of feature values in decreasing order. If the
-foption is turned on, all the values are shown, otherwise only the top 50 are listed.
If you use Bayescreen, please cite it using the following citation (journal version is in process):
Hoksza, D., & Škoda, P. (2016, October). Using Bayesian modeling on molecular fragments features for virtual screening. In Computational Intelligence in Bioinformatics and Computational Biology (CIBCB), 2016 IEEE Conference on (pp. 1-6). IEEE.
We welcome any bug reports, enhancement requests, and other contributions. To submit a bug report or enhancement request, please use the GitHub issues tracker or contact directly David Hoksza.