Artificial Intelligence in Molecular Property Prediction: Advances in Machine Learning, Molecular Representations and Explainable Modelling

Abstract

Artificial intelligence has proven to be a significant way to predict molecular properties in drug discovery, materials science, toxicology, chemical engineering, and environmental research. This review covers the recent developments of machine learning techniques, molecular representations and explainable modelling for molecular property prediction. Structured data and interpretable analysis continue to benefit from the use of traditional algorithms such as linear regression, random forest, support vector machines, decision trees, and gradient boosting. But deep learning architectures, graph neural networks, graph convolutional networks, message-passing models, graph attention networks, transformers, and foundation models have greatly enhanced the learning of complex structure-property relationships. Major molecular representation approaches, such as descriptors, fingerprints, molecular graphs, 3-D structures and learned multimodal embeddings, are also analysed in the review. Explainable AI from Integrated Gradients, attention mechanisms, and saliency mapping are explained in detail, creating transparency and chemically meaningful interpretation. Also, public benchmark datasets and evaluation metrics for regression and classification are discussed. Although significant advances have been made, data quality and model generalization, data interpretability, computational burden, uncertainty, and lack of reproducibility are still limitations. Future developments are anticipated to be focused on physics-informed models, transferable foundation architectures, standardized benchmarking, and better integration with experimental validation, while multimodal learning is expected to be a key focus.