Chemi-Net: A Molecular Graph Convolutional Network for Accurate Drug Property Prediction

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence

Background:

• Absorption, distribution, metabolism, and excretion (ADME) studies are vital for successful drug discovery.
• Current ADME prediction methods often rely on domain-specific features, which can be limiting.
• Deep learning offers a promising alternative for data-driven property prediction.

Purpose of the Study:

• To develop and evaluate Chemi-Net, a novel deep learning method for ADME property prediction.
• To compare the performance of Chemi-Net against Cubist, a widely used machine learning program.
• To assess the potential of Chemi-Net to accelerate drug discovery through enhanced prediction accuracy.

Main Methods:

• Chemi-Net was developed as a completely data-driven, domain knowledge-free deep learning model.
• A large-scale study was conducted at Amgen to compare Chemi-Net with the Cubist benchmark.
• Performance was evaluated across 13 different ADME property datasets.

Main Results:

• Chemi-Net consistently outperformed Cubist across all 13 datasets, achieving higher R² values.
• The median increase in R² value provided by Chemi-Net over Cubist was 26.7%.
• The deep learning approach demonstrated superior accuracy in predicting ADME properties.

Conclusions:

• Chemi-Net represents a significant advancement in data-driven ADME property prediction.
• The enhanced accuracy of Chemi-Net is expected to substantially accelerate the drug discovery process.
• This deep learning method offers a powerful, knowledge-free alternative to conventional approaches.