Google Deep Mind Unveils Alpha Fold 3
Google Deep Mind (London, UK), in collaboration with Isomorphic Labs (London, UK), has announced a next-generation AI model for protein structure prediction. Called Alpha Fold 3, it builds upon the breakthrough success of Alpha Fold 2, which took the scientific and computing communities by storm with its release in 2020. The new model offers a leap in protein prediction capabilities to many new biological molecules and has the potential to greatly accelerate developments for new drugs.
What’s new?
In an article published in Nature, the researchers behind Alpha Fold 3 describe a myriad of upgrades. Arguably the most important of these is the ability of the new model to predict a wider range of biomolecules including DNA, RNA, and ligands (small molecules, such as drugs). The list of molecules in fact extends to most of the Protein Data Bank. Alpha Fold 3 is also capable of predicting interactions between such biomolecules. Alpha Fold 3 is claimed to achieve at least a 50% improvement in predicting protein interactions, and up to 200% in some categories.
How does it work?
Where Alpha Fold 2 relied on establishing spatial relationships between amino acids in a protein sequence to predict its 3D structure, Alpha Fold 3 uses all the atoms in the biomolecule. It starts with a random distribution of these atoms and progressively ‘de-noises’ it over many iterations to achieve the most plausible biomolecular structure.
This process is powered by a diffusion network akin to those used in AI image generation models. It is this that allows Alpha Fold 3 to predict structures of far more complex molecules and their interactions.
The importance
As the first AI system to surpass physics-based tools for biomolecular structure prediction, Alpha Fold 3 is a profound development that serves to enhance research into new drugs and build a more comprehensive understanding of biological processes. Offered as a free and easy-to-use research tool it is also highly accessible to the research community, bolstering the potential for new discoveries. It is especially useful for researchers given it serves as a general model capable of outperforming others who are highly specialized in specific task categories.
