An explainable deep learning platform for molecular discovery

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Explainable AI (XAI)

Background:

• Deep learning models accelerate novel compound discovery but often act as black boxes, limiting chemical insight.
• Explainable deep learning (XDL) aims to provide understandable reasoning behind AI predictions.
• Identifying active structural classes, not just individual compounds, can significantly improve drug discovery efficiency.

Purpose of the Study:

• To present an explainable deep learning platform for mining vast chemical spaces and identifying key substructures linked to desired activity.
• To enable the discovery of active structural classes of molecules, focusing initially on antibiotics.
• To provide a user-friendly protocol for data generation, model implementation, and evaluation of explainability.

Main Methods:

• Utilized Chemprop, a software package employing graph neural networks (GNNs) for molecular property prediction.
• Developed a protocol for experimental data generation, model training, and explainability assessment.
• Focused on identifying structural classes of antibiotics with desired activity.

Main Results:

• Demonstrated an explainable deep learning platform capable of mining large chemical spaces and pinpointing active chemical substructures.
• Successfully applied the platform to discover structural classes of antibiotics.
• The protocol requires no coding proficiency or specialized hardware, executable within 1-2 weeks.

Conclusions:

• The developed platform effectively integrates explainable deep learning for enhanced molecular discovery.
• It facilitates the identification of active structural classes, guiding hypothesis generation and optimizing chemical space exploration.
• The platform's broad applicability extends to discovering various small molecules (anticancer, antiviral, senolytic) and inorganic molecules with specific properties.