Abstract: Proteins are fundamental molecules central to life activity, with their three-dimensional structures determining their biological functions and mechanisms of action. Although experimental methods such as X-ray crystallography, nuclear magnetic resonance (NMR), and cryo-electron microscopy (cryo-EM) have achieved significant success in structural elucidation, they still face challenges such as high cost, long time consumption, and limited applicability. In recent years, rapid advancements in artificial intelligence, particularly deep learning techniques, have revolutionized protein structure prediction. From early statistical energy functions and homology modeling to cutting-edge deep learning models integrating multi-head attention mechanisms and large-scale parameterized networks, there have been substantial improvements in both prediction accuracy and efficiency. Algorithms represented by AlphaFold and RoseTTAFold have not only consistently achieved remarkable results in static structure prediction but also demonstrated broad application prospects in variant screening, drug design, and precision medicine. Furthermore, this review systematically presents structure-prediction methods based on protein language models (such as ESM-3), and further explores two categories of exploratory strategies－those based on spectral descriptors and those employing sampling-augmented deep learning with large-scale biophysical data－while assessing their research progress and application potential, thereby providing theoretical and practical guidance for future studies.

Key words: protein structure prediction, artificial intelligence, AlphaFold, RoseTTAFold, ESM-3, spectral descriptors