The Application of Explainable AI (XAI) in Predicting Molecular Properties and Extracting Hidden Patterns in Chemical Data

The rapid advancement of artificial intelligence (AI) in chemistry has resulted in the development of highly accurate predictive models for molecular properties. However, the opaque nature of these AI systems limits their practical application in drug discovery and materials science. Explainable AI (XAI) overcomes this challenge by identifying the structural features that drive predictions. This transparency allows chemists to extract actionable insights, validate model decisions, and guide the rational design of molecules. This article reviews state-of-the-art XAI techniques applied to molecular property prediction, including model-agnostic methods (e.g., SHAP, LIME), model-specific approaches (e.g., attention mechanisms, integrated gradients), and interpretability methods based on graph neural networks (GNN). We explore their applications across fields such as drug discovery, materials science, toxicity assessment, and quantum chemistry, demonstrating how XAI reveals hidden structure-activity relationships (SAR), identifies novel pharmacophores, and uncovers non-obvious chemical patterns. Recent advancements have integrated attribution with uncertainty quantification, and experimental validation of XAI-guided molecular design has shown promising results. Despite these advances, challenges such as computational scalability, validation of explanations, and standardization remain. The article concludes by discussing emerging trends in XAI, including AI-guided counterfactual design, multi-scale explanations, and the integration of XAI into closed-loop molecular discovery platforms.

کلیدواژه ها:

Explainable AI ، Molecular property prediction ، Graph neural networks ، Interpretable machine learning ، Drug discovery ، Structure-activity relationships ، SHAP ، Attention mechanisms