AlphaFold is an artificial intelligence (AI) system developed by DeepMind, a subsidiary of Alphabet Inc., that predicts the 3D structure of proteins from their amino acid sequences.1 It uses deep learning algorithms trained on vast amounts of genomic and structural data to model the complex folding patterns of proteins.2 AlphaFold's breakthrough performance in the Critical Assessment of protein Structure Prediction (CASP) competition showcased its ability to determine protein structures with accuracy rivaling experimental methods.1 By making accurate protein structure predictions widely accessible, AlphaFold has the potential to accelerate research in fields such as structural biology, drug discovery, and disease understanding.2

In May 2024, Google DeepMind unveiled AlphaFold 3, a groundbreaking advancement in AI-driven protein structure prediction. This latest iteration of the AlphaFold system represents a significant leap forward, capable of predicting the structure and interactions of "all of life's proteins"1. AlphaFold 3 builds upon the success of its predecessor, AlphaFold 2, which revolutionized the field in 2020 by accurately predicting nearly every protein in the human body2.

The new model's capabilities extend beyond individual protein structures, offering potential breakthroughs in drug discovery, biotechnology, and genomics3. AlphaFold 3's improved accuracy and expanded scope could accelerate research into treatments for diseases such as cancer and malaria4. This advancement marks a new era in scientific discovery, with AI playing an increasingly crucial role in unraveling the complexities of biological systems and potentially transforming our understanding of life at the molecular level5.

AlphaFold has been applied to predict the structures of SARS-CoV-2 proteins, the virus responsible for COVID-19. In early 2020, when experimental structures were still pending, AlphaFold was used to model these proteins. The predicted structures were first examined by scientists at the Francis Crick Institute in the UK before being released to the broader research community. The team validated AlphaFold's predictions against the experimentally determined structure of the SARS-CoV-2 spike protein, which was shared in the Protein Data Bank, prior to releasing the computationally predicted structures of other under-studied viral proteins.

One notable example is AlphaFold 2's prediction of the ORF3a protein structure, which closely matched the structure independently determined by researchers at the University of California, Berkeley using cryo-electron microscopy. The ORF3a protein is thought to help the virus escape from host cells after replication and may also contribute to triggering the inflammatory response to infection. While these predicted structures may not be the immediate focus of therapeutic research, they contribute to the scientific community's understanding of the SARS-CoV-2 virus and its biology.

Despite its groundbreaking achievements, AlphaFold faces several challenges and limitations in protein structure prediction. One significant limitation is its difficulty in accurately predicting the impact of single mutations on protein stability, as demonstrated by high school students in a bioinformatics challenge1. This highlights the need for caution when using AlphaFold for mutation-related studies.

AlphaFold also struggles with predicting protein-protein interactions and multimeric structures, as it was initially designed to model individual proteins2. Additionally, the AI has limitations in modeling disordered regions of proteins and predicting structures for proteins with limited experimental data3. These challenges underscore the importance of combining AlphaFold's predictions with experimental validation and other computational methods to gain a comprehensive understanding of protein structures and functions in real-world applications such as drug discovery4.