A Deep-Learning-Aided Drug Screening Based on Visualization of a Hidden Layer as Chemical Space

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Deep learning models are increasingly used in drug discovery but often lack refinement, leading to unreliable drug leads.
• Current deep learning approaches in drug screening can be opaque, hindering the prioritization of promising compounds for experimental validation.

Purpose of the Study:

• To develop an improved drug screening method using graph convolutional networks (GCNs) and hidden layer visualization.
• To enable prioritization of candidate compounds and elucidate structure-activity relationships (SAR) in drug discovery.
• To identify novel therapeutic leads, specifically for histone deacetylase (HDAC) inhibitors.

Main Methods:

• Utilized a graph convolutional network (GCN) deep learning architecture for drug candidate prediction.
• Implemented visualization techniques for a hidden layer within the GCN's output process.
• Applied the method to screen for potential histone deacetylase inhibitors.

Main Results:

• The proposed method effectively visualizes hidden layers of the deep learning model.
• Successfully prioritized compounds for experimental testing from a large pool of predicted active molecules.
• Identified novel lead compounds with potential activity as histone deacetylase inhibitors.
• Provided insights into the relationships between chemical structures and their biological activity.

Conclusions:

• The developed deep learning-aided screening method enhances the efficiency and interpretability of drug discovery.
• Visualization of hidden layers offers a valuable tool for prioritizing compounds and understanding SAR.
• This approach demonstrates significant potential for identifying novel drug leads in medicinal chemistry.