Interpretation of Structure-Activity Relationships in Real-World Drug Design Data Sets Using Explainable Artificial...

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Artificial Intelligence

Background:

• Deep Neural Networks (DNNs) are powerful tools for predicting molecular activities and properties.
• The 'black-box' nature of DNNs limits understanding of structure-activity relationships (SARs), hindering molecular optimization.
• Explainable Artificial Intelligence (XAI) methods are emerging to address this interpretability challenge.

Purpose of the Study:

• To apply and compare various XAI methods for interpreting DNN models in lead optimization.
• To investigate the effectiveness of XAI in uncovering key structural features influencing molecular activity.
• To develop novel visualization techniques for understanding SARs derived from DNN models.

Main Methods:

• Application and comparison of multiple XAI techniques to lead optimization datasets.
• Utilizing DNN models trained on molecular activity and property data.
• Development of an atom-based heatmap visualization scheme for SAR interpretation.

Main Results:

• XAI methods, particularly SHAP-based approaches, provide understandable and comprehensive interpretations of DNN models.
• Atom-based heatmaps offer valuable insights into the underlying SARs.
• Combined DNN and XAI approaches facilitate the identification of crucial structural features for molecular activity.

Conclusions:

• XAI methods, especially SHAP, effectively bridge the interpretability gap in DNN-based molecular modeling.
• Novel visualization techniques enhance the understanding of SARs derived from complex models.
• Interpreting DNN models in the context of their data and associated models is crucial for reliable insights.