Docking-Informed Machine Learning for Kinome-wide Affinity Prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Kinase inhibitors are crucial anticancer drugs, but achieving selectivity is challenging due to ATP-binding site competition.
• Off-target effects and toxicity are significant risks associated with current kinase inhibitors.
• Experimental assessment of kinase inhibitor binding across the kinome is costly and time-consuming.

Purpose of the Study:

• To develop a reliable and interpretable computational method for predicting kinase inhibitor selectivity.
• To facilitate the drug discovery and optimization process for kinase inhibitors.
• To provide a tool for assessing potential toxicity and off-target effects.

Main Methods:

• Aggregated known inhibitor-kinase affinities and generated a 3D interactome by docking inhibitors to X-ray structures.
• Trained a neural network using docked poses as a kinase-specific scoring function.
• Automated the entire pipeline from molecule to 3D-based affinity prediction.

Main Results:

• The neural network achieved a performance of R^2 = 0.63-0.74 on unseen inhibitors across the kinome.
• The developed computational method provides reliable predictions of kinase selectivity.
• The prediction pipeline is fully automated and accessible.

Conclusions:

• Machine learning on docked poses offers a powerful approach for predicting kinase inhibitor selectivity.
• The developed automated pipeline can be readily adopted in medicinal chemistry practice.
• This tool can significantly aid in the discovery and optimization of safer and more effective kinase inhibitors.