Guided Docking as a Data Generation Approach Facilitates Structure-Based Machine Learning on Kinases.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Structural bioinformatics

Background:

• Machine learning models are crucial for exploring chemical spaces in drug discovery.
• The integration of 3D structural information in models is beneficial but limited by the scarcity of experimental protein-ligand complex structures.
• Kinase drug discovery is a key area where structural data is often lacking.

Purpose of the Study:

• To address the challenge of limited protein-ligand complex structures in drug discovery.
• To generate synthetic kinase-ligand complex data for training machine learning models.
• To evaluate the impact of using generated complex data on binding affinity prediction accuracy.

Main Methods:

• Generated synthetic kinase-ligand complex data using template docking for a subset of ChEMBL assay data.
• Utilized an E(3)-invariant graph neural network for training.
• Compared the performance of structure-based models with synthetic poses against ligand-only or drug-target interaction models.

Main Results:

• Models trained with synthetic binding poses demonstrated significantly higher precision in predicting binding affinities.
• The inclusion of generated 3D structural information improved predictive performance compared to models lacking such data.
• The study validates the utility of computationally generated complex data for enhancing predictive models.

Conclusions:

• Generating synthetic protein-ligand complex data is a viable strategy to overcome data limitations in structure-based drug discovery.
• Structure-based machine learning models incorporating synthetic poses offer superior binding affinity prediction.
• This approach can advance the efficiency and accuracy of drug discovery pipelines, particularly for targets like kinases.