Contrastive explanations for machine learning predictions in chemistry.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Chemistry

Background:

• Contrastive explanations, derived from human reasoning, are crucial in explainable artificial intelligence (XAI).
• In machine learning, these explanations highlight features driving opposing model decisions by comparing alternative prediction outcomes.
• Applying contrastive explanations to chemistry requires methods for navigating high-dimensional chemical spaces.

Purpose of the Study:

• To introduce a systematic methodological framework for generating contrastive explanations for machine learning models in chemistry.
• To enable intuitive explanations of predictions, particularly in complex, high-dimensional feature spaces relevant to chemical applications.
• To facilitate a deeper understanding of how molecular features influence machine learning model predictions.

Main Methods:

• Developed the Molecular Contrastive Explanations (MolCE) methodology.
• MolCE generates virtual molecular analogues by substituting building blocks to explore alternative model decisions.
• Quantifies 'contrastive shifts' in model probability distributions resulting from these structural modifications.

Main Results:

• Successfully applied MolCE to explain the selectivity predictions of ligands for D2-like dopamine receptor isoforms.
• Demonstrated the framework's ability to provide intuitive explanations for complex machine learning predictions in a chemical context.
• Validated the effectiveness of MolCE in identifying key molecular features influencing model outcomes.

Conclusions:

• MolCE offers a powerful and systematic approach to generating contrastive explanations for machine learning models in chemistry.
• The methodology enhances the interpretability of chemical predictions, aiding in drug discovery and molecular design.
• This framework bridges the gap between complex machine learning models and actionable chemical insights.