VISAR

VISAR visualizes and interprets deep neural network (DNN) quantitative structure–activity relationship (QSAR) models to map chemical feature gradients to compound- and atom-level contribution weights for predicted activities.

Key Features:

• Model construction and training: Build, train, and test deep neural network QSAR models to generate predictive relationships between chemical structures and activities.
• Activity landscape visualization: Construct activity landscapes that relate chemical feature space to experimental activity values across series of compounds.
• Gradient-to-atom mapping (global perspective): Compute gradients of model outputs with respect to input features and map those gradients to compounds as contribution weights per atom to evaluate global model behavior.
• Local contributor visualization: Identify and visualize positive and negative contributor substructures by assigning atom-level contribution weights and representing their influence on individual predictions.
• Exploratory visualization outputs: Generate visual representations of activity landscapes and atom contribution mappings to support interpretation and knowledge extraction from DNN QSAR models.

Scientific Applications:

• Drug design and lead optimization: Reveal key structural elements and substructures that influence biological activity to inform optimization of lead compounds.
• Model validation and interpretability assessment: Assess whether DNN QSAR models learn meaningful structure–activity patterns by visualizing feature contributions and activity mappings.

Methodology:

Build and train DNN QSAR models; generate activity landscapes mapping chemical feature space to experimental activities; compute gradients of model outputs with respect to input features and map those gradients to compounds to assign atom-level contribution weights; produce visualizations of global and local contribution mappings.