Proteins, often referred to as molecular machines, are essential biomolecules that perform a wide range of cellular functions, typically by forming complexes. Understanding their three-dimendional (3D) structures is key to deciphering their functions. However, a significant gap exists between the vast number of known protein sequences and the relatively limited number of experimentally determined protein structures. Unraveling the mechanisms of protein folding remains a central challenge in understanding the sequence-structure/dynamics-function relationship. In recent years, machine learning (ML) has become a transformative tool across many scientific fields, and structural biology is no exception. Proteins have benefited substantially from advances in artificial intelligence (AI), as numerous ML-based methods have emerged for modeling the structures of both individual proteins and their complexes. Recent breakthrough in ML have marked a major leap forward in tackling the protein folding problem. ML-based AI algorithms for protein structure prediction —most notably AlphaFold—use protein sequence information to accurately predict 3D structures of monomers and multimeric protein complexes, achieving unprecedented levels of precision. Following the success of AlphaFold, recognized with the 2024 Nobel Prize in Chemistry, researchers worldwide have intensified efforts to leverage AI for unraveling complex biological challenges—from drug discovery to protein-protein interactions. This review highlights ML-based approaches, with a primary focus on AlphaFold and its derivatives, while also covering other notable methods such as the hybrid deep-learning based RoseTTAFold and protein language model-based ESMFold. These tools have diverse applications in protein structure modeling and significantly advance our understanding of the intricate relationships between sequence, structure, dynamics, and function. While ML-based methods still face limitations in certain cases —such as membrane proteins, which are underrepresented in experimental structural databases, or antibody–antigen interactions, which involve highly diverse and difficult-to-model hypervariable regions—advances in computational techniques and the incorporation of new experimental data are steadily improving the accuracy of these algorithms in tackling such challenges. Overall, the implementation of ML in the study of molecular machines represents a promising direction, with the potential to bridge the sequence-structure gap and address longstanding questions in structural biology and medicine.
Çelik et al. (Wed,) studied this question.