Prediction of Cytochrome P450 Substrates Using the Explainable Multitask Deep Learning Models.

Area of Science:

• Biochemistry
• Pharmacology
• Computational Chemistry

Background:

• Cytochromes P450 (P450s or CYPs) are critical Phase I metabolic enzymes, processing approximately 75% of therapeutic drugs.
• CYP-mediated metabolism is linked to toxic metabolite generation and drug-drug interactions, necessitating predictive tools.

Purpose of the Study:

• To develop and evaluate multitask learning models for simultaneously predicting substrates of five major drug-metabolizing P450 enzymes (CYP3A4, 2C9, 2C19, 2D6, 1A2).
• To enhance early-stage drug development by accurately identifying potential P450 substrates.

Main Methods:

• Construction of multitask learning models utilizing fingerprints and graph neural networks.
• Training and validation on collected substrate datasets for multiple CYP enzymes.
• Application of Shapley additive explanation and attention mechanisms for substructure identification.

Main Results:

• The multitask model achieved superior performance, with an average AUC of 90.8% on the test set, outperforming single-task and conventional machine learning models.
• The model showed robust performance even with limited substrate data for enzymes like CYP1A2, 2C9, and 2C19.
• Key substructures associated with P450 substrates were identified and validated.

Conclusions:

• Multitask learning, particularly with graph neural networks, offers a powerful approach for predicting P450 drug metabolism.
• The developed models provide valuable insights for assessing drug-drug interactions and metabolic liabilities early in drug discovery.
• Explainability methods aid in understanding the structural basis of P450-substrate interactions.