General information
Protein kinases contribute to the regulation of almost all aspects of cell function. Pathological kinase activity is often associated with the cell switching from normal to pathological state and disease manifestation at the organismal level. Thus, kinases are of particular interest as therapeutic targets, and more than ten kinase inhibitors were approved for the usage in clinical practice in the last four years. The primary difficulty associated with the discovery of novel kinase inhibitors is their selectivity. Thus, to find the chemical entity with the appropriate balance between safety and efficacy, it is necessary to study the activity of compounds against as many kinases as possible in the early stage of development. KinScreen is aimed to optimize this process.
Key features
► Predict kinase targets with PASS software to plan experimental testing and/or unravel molecular mechanism of action of compound
► Visualize results on the kinome tree to assess the distribution of targets across kinase families
► Search for analogous compounds across ChEMBL database to find experimental data on them
Our web-service is based on the PASS (Prediction of Activity Spectra for Substances) software (Filimonov, D. A., et al. "Prediction of the biological activity spectra of organic compounds using the PASS online web resource." Chemistry of Heterocyclic Compounds 50.3 (2014): 444-457.). PASS was developed to analyze arrays of chemical compounds with associated biological activities using naïve Bayes approach and elucidate «structure-activity» relationships for the sake of the discovery of novel compounds with desired activity or activities. To train PASS we used data on chemical structures and activities of compounds which were tested against kinases according to the ChEMBL database (21-st version). We provided extensive filtration for the extracted data to eliminate inaccurate or ambiguous molecular representations and unreliable activity data in accordance with practices adopted for building reliable and predictive (Q)SAR models. Initially, we extracted 71 733 chemicals, only 69 454 of them we used for training. As a result, the software was prepared and became able to provide users with the computer assessment of inhibitory activity for chemical compounds against more than 300 of protein kinases. The quality of prediction was assessed as ROC AUC calculated using leave-one-out cross-validation which exceeded 0.85
Predicted protein kinases are mapped on the dendrogramm representing human kinome and its families. Illustration reproduced courtesy of Cell Signaling Technology, Inc. (www.cellsignal.com) and is based on the original work of Manning et al. (Manning, Gerard, et al. "The protein kinase complement of the human genome." Science 298.5600 (2002): 1912-1934.)
The search is carried out in the space of the predicted kinase targets as opposed to more common approaches using chemical similarity. Thus, analogous compounds may be more or less structurally dissimilar to the query compounds, but share common profile of the predicted kinase targets with it. Using Locality Sensitive Hashing Forest (LSH) method (Bawa, Mayank, Tyson Condie, and Prasanna Ganesan. "LSH forest: self-tuning indexes for similarity search." Proceedings of the 14th international conference on World Wide Web. ACM, 2005.) from sci-kit learn library (Pedregosa, Fabian, et al. "Scikit-learn: Machine learning in Python." Journal of Machine Learning Research 12.Oct (2011): 2825-2830.) our web-service provides five nearest neighbors for the query compound.
Results of prediction are provided in tabular form. Results include kinase identifiers and corresponding confidence scores. Confidence score represent the diffenrence between the probability for compound of interest to inhibit particular kinase and probability to not to inhibit this kinase. The higher score means higher confidence in the predicted activity.
Content coming soon
Input chemical structure first, please