DeepMind, a leading AI company, has recently announced the launch of AlphaFold 3, the latest version of its groundbreaking protein folding prediction tool. Derived from machine learning, AlphaFold projects the 3D structure of proteins based on amino acid sequences, essentially ‘folding’ them like origami. This is a crucial aspect of biochemistry and molecular biology, given that proteins, the building blocks of life, can have their functions determined by the way they fold. Misfolded proteins can lead to the development of diseases, including Parkinson’s and Alzheimer’s.
Predicting protein misfolding is an ongoing scientific challenge because of the vast number of configurations that proteins can take. The computational intensity required to determine the correct structure is mitigated by AlphaFold’s use of deep learning. This software trains neural networks on known protein structures and then applies this knowledge to predict the 3D shapes of proteins based on their amino acid sequences.
AlphaFold 3 includes an advanced iteration of the Evoformer module, a critical component of the deep learning system used in AlphaFold 2. Following processing by the Evoformer, AlphaFold 3 employs a unique diffusion network to construct the predicted structures. Similar to AI image generators like DALL-E, the process begins with a ‘cloud’ of atoms and repeatedly refines the structure, resulting in the final, most likely accurate, molecular configuration.
Importantly, AlphaFold 3’s model goes beyond proteins. It incorporates information on DNA, RNA, and small molecules, thus enabling evaluation of some of their intricate interactions. It’s capable of parsing over 99% of all known biomolecular complexes based on the Protein Data Bank data. Isomorphic Labs, which collaborated with DeepMind on AlphaFold 3, is already using the model for drug design applications in partnership with pharmaceutical companies.
DeepMind’s AlphaFold project, ongoing since 2016, made major strides with AlphaFold 1, which won the CASP13 protein structure prediction competition in 2018. AlphaFold 2, launched in 2020, delivered such high prediction accuracy that the scientific community considered the protein-folding problem essentially solved.
DeepMind has facilitated the use of this research by launching the AlphaFold Server, a free platform enabling researchers to use AlphaFold 3 without the need for extensive computational resources or machine learning expertise.
AlphaFold applications have contributed to new research pursuits, like studying proteins capable of degrading environmental pollutants and enhancing understanding of rare tropical diseases.
In collaboration with EMBL’s European Bioinformatics Institute, DeepMind released the AlphaFold Protein Structure Database in July 2021, with structure predictions for over 350,000 proteins, including the entire human proteome. The database now includes over 200 million structures, covering nearly all cataloged proteins known to science and has been accessed by more than one million users in over 190 countries.
AlphaFold 3 represents the latest development in DeepMind’s pioneering protein discovery and analysis system, positioning AlphaFold as an essential tool for today’s scientific discovery and the future of biotech development.
