DeepMind has announced AlphaFold 3, the latest version of its machine learning-assisted protein folding prediction tool. This cutting-edge tool forecasts the three-dimensional structure that chains of amino acids will create when they fold, a process that has a significant impact on the protein’s function. Understanding protein folding is essential to comprehend many biological processes and diseases since misfolded proteins can disrupt normal function and contribute to severe diseases like Alzheimer’s and Parkinson’s.
The complexity and variety of potential protein configurations make predicting protein folding computationally challenging. AlphaFold uses deep learning to predict protein structures in this vast space. It uses neural networks, which have been trained on known protein structures, to predict the 3D shape of proteins.
AlphaFold 3 has an improved version of its Evoformer module, which, after processing input molecules, uses a diffusion network to assemble predicted structures. This diffusion network is similar to AI image generators like DALL-E and begins with a structure that is refined iteratively until a likely accurate molecular model is reached. AlphaFold 3 also includes DNA, RNA, and small molecules information and can capture some of their complex interactions.
The model, trained using Protein Data Bank data, can process over 99% of all known biomolecular complexes in the database. DeepMind has collaborated with Isomorphic Labs on the AlphaFold 3 project, and already, the model is being applied to real-world drug design challenges in collaboration with pharmaceutical companies.
DeepMind first launched the AlphaFold project in 2016. By 2018, the first version of the AI system, AlphaFold 1, won the CASP13 (Critical Assessment of protein Structure Prediction) challenge, demonstrating the tool’s efficiency. The accuracy of AlphaFold 2, introduced in 2020 at CASP14, saw the scientific community largely consider the protein-folding problem as solved. AlphaFold 2’s methods paper has received over 20,000 citations since then.
AlphaFold’s evolution has enabled several novel research projects, such as studying proteins that degrade environmental pollutants and enhancing our understanding of rare tropical diseases. In partnership with EMBL’s European Bioinformatics Institute, DeepMind released the AlphaFold Protein Structure Database, providing access to over 350,000 protein structure predictions. It now includes over 200 million structures, encompassing almost all cataloged proteins known to science, and has seen over a million users from 190+ countries.
The AlphaFold Server, a user-friendly platform with free access, has also been launched, permitting researchers to utilize AlphaFold 3 without requiring extensive computational resources or machine learning experience. AlphaFold 3 signals more advancements for this top-tier protein discovery and analysis system.
