Explainable Molecular Property Prediction: Aligning Chemical Concepts With Predictions via Language Models

Area of Science:

• Computational chemistry and cheminformatics.
• Artificial intelligence in scientific discovery.
• Molecular modeling and property prediction.

Background:

• Accurate molecular property prediction is vital for drug discovery and materials science.
• Existing transformer models lack chemically meaningful explanations and fail to reveal structure-property relationships.
• Need for explainable AI (XAI) in molecular sciences.

Purpose of the Study:

• To develop an explainable molecular property prediction framework, Lamole, providing chemical concept-aligned explanations.
• To improve the faithfulness of explanations to molecular structure-property relationships.
• To demonstrate Lamole's utility in interpretable molecular optimization and discovery.

Main Methods:

• Utilizing Group SELFIES as input tokens for language model pre-training and fine-tuning.
• Analyzing self-attention weights and gradients to quantify substructure impact.
• Implementing a marginal loss function to align explanations with chemical annotations and the data manifold.
• Integrating Lamole with an evolutionary algorithm for interpretable molecular editing.

Main Results:

• Lamole achieves comparable prediction accuracy to existing models.
• Explanation accuracy is boosted by up to 14.3%, establishing a new state-of-the-art in explainable prediction.
• Demonstrated actionable utility through an interpretable molecular optimization pipeline.

Conclusions:

• Lamole offers a powerful framework for explainable molecular property prediction.
• The approach provides chemically meaningful and faithful explanations, advancing AI in molecular sciences.
• Lamole serves as a practical guide for molecule discovery and optimization beyond post-hoc analysis.