EXPLAINABLE MACHINE LEARNING APPROACHES FOR MOLECULAR PROPERTY PREDICTION

Abstract

Molecular property prediction plays an essential role in drug discovery and material science, yet traditional machine learning models often act as black boxes, limiting insight into their decision-making. This paper reviews recent advances in explainable machine learning (XML) applied to molecular property prediction, focusing on methods that improve interpretability without sacrificing performance. Notable developments include graph neural network (GNN) approaches with integrated explainability benchmarks such as GradInput and Integrated Gradients arXiv, and transformer-based frameworks like Lamole, which align attention mechanisms with chemist-annotated structural concepts arXivOpenReview.

Empirical studies have shown that explainable models can uncover meaningful molecular features—such as substructures responsible for dual-target bioactivity Nature, or contributors to ADME (Absorption, Distribution, Metabolism, Excretion) profiles via SHAP values BioMed Central. More recent work quantifies how local and global explanations (e.g., SHAP, LIME) help interpret models across toxicity and property datasets NatureScienceDirect. Data in these studies come from open-source public chemical datasets, including MoleculeNet benchmarks and curated ADME property datasets (all are freely accessible online), ensuring reproducibility.

This paper aims to synthesize these advances and highlight best practices in explainable molecular ML—integrating model accuracy with interpretability. In doing so, it addresses the emerging demand for transparent AI in chemistry, enabling scientists to understand, trust, and act on machine-generated predictions. Our discussion outlines not only the strengths of current techniques but also their limitations and areas where interpretability remains a challenge. The goal is to guide future research toward models that are both high-performing and human-understandable—so we can build the strongest, most useful paper yet.