Google DeepMind announces 'AlphaFold 3,' an AI model that can predict the structure and interactions of all biological molecules with extreme accuracy

Google DeepMind has collaborated with Isomorphic Labs, an AI drug discovery company, to release AlphaFold 3, an AI model that can predict the structures and interactions of biological molecules such as proteins, DNA, RNA, and ligands with extremely high accuracy.

Google DeepMind and Isomorphic Labs introduce AlphaFold 3 AI model
https://blog.google/technology/ai/google-deepmind-isomorphic-alphafold-3-ai-model/

Accurate structure prediction of biomolecular interactions with AlphaFold 3 | Nature
https://www.nature.com/articles/s41586-024-07487-w

All life is made up of molecules such as proteins and DNA, but these molecules do not function alone; they form life through the interactions of millions of combinations of molecules. AlphaFold 3 is an AI model that predicts the structure and interactions of all biological molecules, and has achieved 50% higher accuracy in predicting interactions between proteins and other molecular types than previous models, as well as 100% higher accuracy in important categories of interactions.

AlphaFold 3 builds on the foundation of AlphaFold 2, released in 2020. The previous model, AlphaFold 2, succeeded in making fundamental advances in protein structure prediction, helping to lead to numerous discoveries in areas such as malaria vaccines, cancer treatments, and enzyme design. The new model, AlphaFold 3, can handle a wide range of biological molecules in addition to proteins, potentially driving innovation in a wide range of fields, from research into biorenewable materials and resilient crops to drug design and genomics research.

The figure below shows AlphaFold 3's prediction of the spike protein '7PNM' of the coronavirus , one of the causes of the common cold. The AlphaFold 3 structure prediction is colored and the actual structure is displayed in gray, indicating an accurate match.


Given an input list of molecules, AlphaFold 3 generates a bonded 3D structure and infers how the molecules will fit together. The core of AlphaFold 3's model is an improved version of the Evoformer module, AlphaFold 2's deep learning architecture, which processes the input and then assembles predictions using a diffusion network similar to image generation. The diffusion process goes from a scattered atomic state to a precise molecular structure through a number of steps.

The figure below shows the predicted structure of the DNA-binding protein '7R6R.' The DNA double helix is shown in pink, and the protein in blue. The true molecular structure shown in grey was finally determined after extensive experimentation, and AlphaFold 3's prediction matches the true molecular structure almost perfectly.


AlphaFold 3 has a drug design function that predicts molecules commonly used in medicines, such as ligands and antibodies that bind to proteins and cause various effects. AlphaFold 3 can predict drug interactions with greater accuracy than all previous models, and is the first AI model to predict biomolecular structure better than physics-based tools.

The figure below shows the structure of the RNA-modifying protein '8AW3' predicted by AlphaFold 3. The protein is shown in blue, the RNA chain in purple, and the two ions in yellow. You can see that it matches the actual structure shown in gray.


AlphaFold 3 can be accessed through the newly released research tool ' AlphaFold Server ', and most of its functions can be used free of charge for non-commercial use. Isomorphic Labs, which developed AlphaFold 3, is already working with pharmaceutical companies to use AlphaFold 3 to research new treatments.