Global Analysis of Deep Learning Prediction Using Large-Scale In-House Kinome-Wide Profiling Data

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Predicting compound activity and ADMET properties is crucial in drug discovery.
• Deep learning and non-deep learning methods are increasingly used for these predictions.

Purpose of the Study:

• To compare the performance of deep learning (single-task and multitask graph neural network - GNN) and non-deep learning (LightGBM) models for kinase activity prediction.
• To evaluate the extrapolative validity and analyze feature importance of the developed models.

Main Methods:

• Activity prediction using single-task GNN, multitask GNN, and LightGBM on in-house kinase assay data.
• Validation of the multitask GNN model on known kinase ligands.
• Analysis of feature importance to identify critical protein-ligand interaction sites.

Main Results:

• Single-task GNN models showed lower prediction accuracy than LightGBM.
• Multitask GNN models, integrating data from multiple kinases, significantly outperformed LightGBM.
• The multitask model demonstrated strong extrapolative validity and identified important interaction sites.

Conclusions:

• Multitask GNN models offer superior performance for kinase activity prediction compared to non-deep learning methods.
• Ligand-based prediction models can aid in both activity prediction and drug design by analyzing feature importance.