Machine learning for RNA-small molecule interactions: methodologies, evaluation, and future directions

RNA is emerging as a compelling therapeutic target due to its rich structural diversity and central involvement in gene regulation, splicing, translation, and cellular signaling. Alongside advances in structural biology, computational approaches, especially machine learning and deep learning, have transformed the systematic analysis of RNA-small molecule interactions. In this paper, we provide a comprehensive survey of computational methods for RNA-small molecule interaction prediction, with a particular emphasis on three core tasks: binding site prediction, binding preference prediction, and binding pose prediction. We first summarize the data resources commonly used to support these tasks and outline molecular representations that capture RNA structural features and the chemical properties of small molecules. We then review a spectrum of computational approaches for the above three tasks, ranging from statistical methods and traditional machine learning to deep learning. We further evaluate representative methods for each task, providing insights into their respective strengths and limitations. Finally, we highlight open challenges in current research on RNA-small molecule interactions, including limited data availability, the black-box nature of the models, and the growing need for generative modeling. We discuss emerging directions such as data augmentation, model interpretability, and RNA-targeted small-molecule design as promising pathways to address these challenges and advance rational RNA-targeted drug discovery.