Identification of novel small-molecule drugs often relies on high-throughput screening (HTS) experiments. The success of the HTS approach is largely dependent on the extent and diversity of compound libraries subjected to profiling. Although feasible, drastic expansion of compound sets inevitably increases the cost, time and labour of the drug discovery process, and requires high-end robotic instruments.
To that end, researchers at FIMM developed Virtual Kinome Profiler (VKP), a cost-effective computational platform that captures distinct representations of chemogenomic association of the druggable kinome for drug repurposing and lead identification applications. VKP is based on an ensemble support vector machine (eSVM) classifier to enable high-throughput virtual profiling of compound-kinase interactions.
A schematic illustration of VKP depicting the statistical model for chemogenomic analysis and the eSVM model for binding activity class predictions across the druggable kinome.
“I hope the platform will significantly enhance hit-lead optimization phase of future drug discovery process and would serve as a go-to resource for compound repositioning strategies”, says FIMM-EMBL PhD student Balaguru Ravikumar, the lead author of the study who developed the VKP platform.
In the first case study, the researchers profiled approximately 37 million compound-kinase pairs and made predictions for more than 151,000 compounds in terms of their repositioning and lead molecule potential, against 248 kinases simultaneously. The model resulted in 10-fold enrichment of true hits when compared to standard biochemical assays, making it possible to target the follow-up experiments on the most potent interactions only.
“This is again a great demonstration how machine learning can guide high-throughput experiments, leading to reduced costs and time, yet improving the accuracy of the hit selection”, says Tero Aittokallio, the corresponding author of the study.
The bioactivity information used for developing the model is one of the most comprehensive resources to date. The platform is implemented as an easy-to-use web-application, that requires minimal user requirements and compound description for the prediction tasks, hence providing chemical biologists model-informed suggestions regarding their compounds’ activity across the druggable kinome.
The web-application and the dataset for model development are freely available for academic use at https://virtualkinomeprofiler.fimm.fi/. The source-code implementing the data analysis and machine learning model is available under the Mozilla Public License 2.0. For commercial use, the platform and the source-code are available through yearly-licensing agreement from the Helsinki Innovation Services (HIS).
Reference: Balaguru Ravikumar, Sanna Timonen, Zaid Alam, Elina Parri, Krister Wennerberg, Tero Aittokallio. Chemogenomic Analysis of the Druggable Kinome and Its Application to Repositioning and Lead Identification Studies, Cell Chemical Biology (2019), https://doi.org/10.1016/j.chembiol.2019.08.007