Abstract
Artificial intelligence, especially deep learning, has revolutionized protein structure and function prediction, overcoming the limitations of traditional experimental methods. This review systematically examines key AI-driven advances in three areas: single-chain structure prediction, protein–protein interaction and complex structure prediction, and direct functional property prediction. These models have achieved experimental-level accuracy in many tasks, demonstrating the power of data-driven approaches in decoding protein folding and function. Despite these successes, challenges remain, including predicting dynamic conformational states, reducing reliance on evolutionary information, improving interpretability, and bridging molecular-scale predictions with cellular environments. Future directions involve integrating physics-based modeling, developing multimodal foundation models, and closing the loop between computation and experimentation. This progress not only accelerates basic biological research but also holds great promise for drug discovery, enzyme design, and synthetic biology, highlighting AI's transformative role in understanding life and addressing biomedical challenges.