Deriving general structure-activity/selectivity relationship patterns for different subfamilies of cyclin-dependent...

Area of Science:

• Medicinal Chemistry
• Computational Biology
• Drug Discovery

Background:

• Cyclin-dependent kinases (CDKs) are crucial for cell cycle regulation and are key targets in cancer therapy.
• Developing selective CDK inhibitors is challenging but essential for minimizing off-target side effects.

Purpose of the Study:

• To derive general structure-activity relationship (SAR) patterns for modeling CDK inhibitor selectivity and activity.
• To utilize machine learning for predicting and identifying potent and selective CDK inhibitors.

Main Methods:

• Collected 8592 small molecules with binding affinities to CDK1, CDK2, CDK4, CDK5, and CDK9 from Binding DB.
• Calculated molecular descriptors and trained supervised Kohonen networks (SKN) and counter propagation artificial neural networks (CPANN) models.
• Validated models using tenfold cross-validation and external test sets; performed virtual screening on 2 million PubChem molecules.

Main Results:

• SKN models achieved prediction accuracies from 0.75 to 0.94 on external test sets.
• Identified molecular descriptors like hydrophilicity and total polar surface area for activity and selectivity mapping.
• Virtual screening yielded areas under the receiver operating characteristic curves from 0.72 to 1.00 for SKN models.

Conclusions:

• The study provides a robust machine learning framework for predicting CDK inhibitor properties.
• Developed models and identified chemical space regions for active and selective CDK inhibitors.
• Contributes to addressing the challenge of CDK selectivity, aiding in the development of safer cancer therapeutics.