Physics-inspired explainable AI for mechanistic and decision support in drug discovery

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Drug Discovery

Background:

• Explainable artificial intelligence (XAI) is crucial for trustworthy drug discovery but often relies on post hoc, correlational methods.
• Current XAI approaches in drug discovery may lack mechanistic interpretability, limiting their utility in molecular design.

Purpose of the Study:

• To review physics-inspired XAI as a framework for enhancing mechanistic interpretability and decision support in molecular design.
• To evaluate the strengths and limitations of physics-inspired XAI in modeling molecular recognition and structure-property relationships.
• To propose 'molecular intelligence' as a decision-centric framework for drug discovery.

Main Methods:

• Review of current literature on physics-inspired XAI in drug discovery.
• Analysis of XAI's application in modeling molecular recognition and local structure-property relations.
• Examination of the role of biological and pharmacological processes in drug efficacy and toxicity.

Main Results:

• Physics-inspired XAI shows strengths in modeling molecular recognition and local structure-property relationships.
• Efficacy and toxicity are influenced by biological and pharmacological factors beyond molecular physics.
• Explainability currently improves prioritization and experimental efficiency but has limited impact on overall R&D outcomes.

Conclusions:

• Physics-inspired XAI offers a promising framework for mechanistic interpretability in molecular design.
• Integrating XAI with biological and pharmacological knowledge is essential for a comprehensive understanding of drug properties.
• The proposed 'molecular intelligence' framework aims to enhance decision-making by integrating prediction, validation, and domain knowledge in drug discovery.