Abstract
RNA structure prediction remains one of the most challenging problems in computational biology, with significant implications for understanding gene regulation, drug design, and synthetic biology. While deep learning has revolutionized protein structure prediction, RNA presents unique challenges including limited training data, complex noncanonical interactions, and conformational flexibility. This review examines the evolution from traditional physics-based methods to current deep learning approaches for RNA secondary and tertiary structure prediction. After briefly exploring traditional methods, like Direct Coupling Analysis and physics-based simulations, we systematically review three deep learning paradigms: language model-based methods, end-to-end structure predictors, and geometry-distance prediction approaches. Furthermore, we identify critical future research directions focusing on advanced tokenization strategies to address data scarcity and explainable artificial intelligence techniques to improve model interpretability. Despite significant progress, achieving transformative performance requires continued methodological innovation, specifically designed for RNA's unique characteristics, and a substantial expansion of high-quality structural datasets.