A Perspective on Explanations of Molecular Prediction Models

Area of Science:

• Chemistry
• Artificial Intelligence
• Computational Chemistry

Background:

• Deep learning (DL) models offer powerful predictive capabilities in chemistry but often function as "black boxes", hindering trust and adoption.
• Lack of interpretability in DL models creates skepticism among chemists regarding their use in critical decision-making processes.
• Explainable artificial intelligence (XAI) emerges as a crucial field to bridge this gap by providing transparency into AI predictions.

Purpose of the Study:

• To review the principles and emerging methods of XAI within the chemical domain.
• To present novel XAI techniques developed by the authors for interpreting DL models in chemistry.
• To demonstrate the application of XAI in predicting key molecular properties like solubility, blood-brain barrier permeability, and scent.

Main Methods:

• Review of fundamental XAI principles and evaluation strategies relevant to chemical applications.
• Development and application of specific XAI methods, including chemical counterfactuals and descriptor explanations.
• Utilizing a two-step approach: first, building a predictive DL model, then applying XAI for interpretation.

Main Results:

• XAI methods successfully explain DL predictions for molecular properties.
• Techniques like chemical counterfactuals and descriptor explanations reveal critical structure-property relationships.
• The study validates the utility of XAI in enhancing the interpretability of DL models for chemists.

Conclusions:

• XAI is essential for overcoming the "black-box" nature of DL in chemistry, fostering wider adoption.
• Interpretable DL models, facilitated by XAI, provide valuable insights into molecular behavior and design.
• A combined approach of DL modeling and XAI analysis is a powerful strategy for discovering novel structure-property relationships.