VISAR: an interactive tool for dissecting chemical features learned by deep neural network QSAR models.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Quantitative structure-activity relationship (QSAR) models are crucial for predicting compound activity.
• Interpreting 'black box' models like deep neural networks (DNNs) in QSAR remains a significant challenge.
• Understanding learned chemical features is vital for model validation and knowledge extraction.

Purpose of the Study:

• To develop an interactive tool, VISAR, for interpreting and visualizing DNN-based QSAR models.
• To enable exploration of structure-activity relationships from both global and local perspectives.
• To provide insights for drug design and enhance QSAR model validation.

Main Methods:

• Development of VISAR, a web application for constructing and training DNN QSAR models.
• Generation of activity landscapes to visualize correlations between chemical features and experimental activity.
• Mapping atom-wise contribution weights and visualizing substructures influencing activity.

Main Results:

• VISAR provides interactive visualization of DNN QSAR model outputs.
• Activity landscapes offer a global view of structure-activity relationships.
• Atom contribution mapping highlights local substructures critical for activity prediction.

Conclusions:

• VISAR facilitates the interpretation and interactive analysis of DNN QSAR models.
• The tool aids in drug design by providing insights into learned chemical features.
• VISAR serves as a valuable addition for QSAR model validation.