A Hybrid Structure-Based Machine Learning Approach for Predicting Kinase Inhibition by Small Molecules.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Kinases are key drug targets, with over 70 inhibitors developed.
• Understanding precise kinase-compound interactions remains a challenge.
• Existing methods struggle with broad target spectrum prediction.

Purpose of the Study:

• To develop a computational approach for kinome-wide binding measurements.
• To enhance structure-based machine learning for kinase inhibitor discovery.
• To improve the prediction of qualitative and quantitative kinase-compound affinities.

Main Methods:

• Created the Kinase Inhibitor Complex (KinCo) dataset with predicted kinase structures and experimental binding constants.
• Developed a machine learning loss function integrating qualitative and quantitative binding data.
• Trained a structure-based machine learning model on the KinCo dataset.

Main Results:

• The developed approach surpasses methods using only crystal structures for predicting kinase-compound affinities.
• The model demonstrates superior performance compared to structure-free methods.
• The approach successfully generalizes to diverse kinase sequences and compound scaffolds, capturing known kinase biochemistry.

Conclusions:

• This computational strategy enhances the prediction of kinase-compound interactions.
• The method offers a powerful tool for structure-based drug discovery.
• The approach improves understanding of kinase target spectrum and binding characteristics.