Calibrated geometric deep learning improves kinase-drug binding predictions.

Area of Science:

• Biochemistry
• Computational Biology
• Pharmacology

Background:

• Protein kinases are crucial cellular regulators with significant therapeutic potential, particularly in oncology.
• Kinase inhibitors represent a major class of approved drugs, yet many kinases remain undrugged.
• Current computational methods for predicting kinase-compound interactions often neglect 3D structural information.

Purpose of the Study:

• To develop a deep learning algorithm, KDBNet, that leverages 3D protein and molecule structures for accurate binding affinity prediction.
• To introduce a robust method for quantifying and calibrating prediction uncertainties to guide drug discovery.
• To improve the efficiency of identifying novel kinase-drug pairs with high binding affinity.

Main Methods:

• KDBNet employs graph neural networks to learn structural representations of protein binding pockets and drug molecules.
• The algorithm captures geometric and spatial features critical for binding interactions.
• An uncertainty quantification and calibration algorithm is integrated into the prediction framework.

Main Results:

• KDBNet significantly outperforms existing deep learning models in predicting kinase-drug binding affinities.
• The model's estimated uncertainties are well-calibrated and informative regarding prediction errors.
• Integration with Bayesian optimization facilitated data-efficient active learning for kinase-drug discovery.

Conclusions:

• KDBNet provides a powerful 3D-aware deep learning approach for predicting kinase-drug binding.
• Accurate uncertainty estimation enhances the reliability of KDBNet in guiding drug discovery efforts.
• The KDBNet framework accelerates the identification of novel and effective kinase inhibitors.