Deep Learning for Protein Structure Prediction: From AlphaFold to the Next Frontier

Abstract

Protein structure is closely linked with protein function. For many decades, scientists tried to predict the three-dimensional structure of a protein from its amino acid sequence. Experimental methods such as X-ray crystallography, nuclear magnetic resonance spectroscopy, and cryo-electron microscopy provide reliable structures. However, these methods can be expensive and time-consuming. Computational methods therefore became important alternatives. Early prediction methods were based mainly on sequence similarity, physical energy functions, and structural templates. Their accuracy was limited for many proteins. Deep learning changed this field. AlphaFold2 showed that artificial intelligence can predict the structures of many proteins with near-experimental accuracy (Jumper et al., 2021). RoseTTAFold provided another powerful deep-learning approach (Baek et al., 2021). Protein language models such as ESMFold later showed that structural information can also be learned directly from very large collections of protein sequences (Lin et al., 2023). AlphaFold3 further expanded the field by predicting interactions among proteins, DNA, RNA, ions, small molecules, and modified residues (Abramson et al., 2024). Generative models such as RFdiffusion and ESM3 are now moving the field from structure prediction toward protein design (Watson et al., 2023; Hayes et al., 2025). Recent open models are also increasing access to advanced biomolecular modelling. Despite this progress, major challenges remain. Proteins are dynamic molecules. They may adopt several conformations. Their structures are influenced by ligands, membranes, modifications, and the cellular environment. Future models therefore need to predict not only one structure, but also molecular dynamics, interactions, binding strength, and biological function. This review describes the development of deep learning for protein structure prediction and discusses the major directions that may define the next frontier.