The new AlphaFold 3 AI system for protein structure prediction has been released by Google DeepMind and its spinout Isomorphic, which is built around the system. According to the companies, the breakthrough system now shows at least a 50 percent improvement compared with existing prediction methods, and for some important categories of interaction prediction accuracy have been doubled.
AlphaFold 3 aims to go beyond proteins to a broad spectrum of biomolecules. The system and some results are described in a paper published in Nature, and the lead author is Google DeepMind’s Josh Abramson.
The system uses a novel, end‐to‐end, deep neural network trained to produce protein structures from amino acid sequence, multiple sequence alignments, and homologous proteins. In their paper, the researchers write, the AlphaFold 3 model has “A substantially updated diffusion-based architecture, which is capable of joint structure prediction of complexes including proteins, nucleic acids, small molecules, ions, and modified residues.”
The Isomorphic team reports, “To build on AlphaFold 3’s potential for drug design, we at Isomorphic Labs are already collaborating with pharmaceutical companies to apply it to real-world drug design challenges and, ultimately, develop new life-changing treatments for patients.”
Since its launch about two years ago, Isomorphic has announced two agreements worth $3B worth with each Eli Lilly and Novartis to apply AI to discover new drugs.
The new model builds on the foundations of AlphaFold 2, which in 2020 made a fundamental breakthrough in protein structure prediction. The companies report that so far, millions of researchers globally have used AlphaFold 2 to make discoveries in areas including malaria vaccines, cancer treatments, and enzyme design. The system has been cited more than 20,000 times and it has received many accolades, most recently the Breakthrough Prize in Life Sciences.
Given an input list of molecules, AlphaFold 3 generates their joint 3D structure, revealing how they all fit together. It models large biomolecules such as proteins, DNA, and RNA, as well as small molecules. Furthermore, AlphaFold 3 can model chemical modifications to these molecules that control the healthy functioning of cells.
Its developers say AlphaFold 3’s capabilities come from its next-generation architecture and training that now covers all of life’s molecules. At the core of the model is an improved version of the Evoformer module—a deep learning architecture that underpinned AlphaFold 2’s performance.
After processing the inputs, AlphaFold 3 assembles its predictions using a diffusion network, akin to those found in AI image generators. The diffusion process starts with a cloud of atoms, and over many steps converges on its final, most accurate molecular structure.
The companies report that AlphaFold 3’s predictions of molecular interactions surpass the accuracy of all existing systems.
They write in their report, “The new AlphaFold model demonstrates significantly improved accuracy over many previous specialized tools: far greater accuracy on protein-ligand interactions than state of the art docking tools, much higher accuracy on protein-nucleic acid interactions than nucleic-acid-specific predictors, and significantly higher antibody-antigen prediction accuracy than AlphaFold-Multimer v2.3.”
Scientists can access the majority of AlphaFold 3’s capabilities, for free, through the newly launched AlphaFold Server research tool.
