The latest issue of "Nature" published on May 8th reports that AlphaFold 3 can predict the structures of protein interactions with other biological molecules with high accuracy. AlphaFold 3 is the newest iteration of this artificial intelligence model, developed by teams from Google DeepMind and Isomorphic Labs, showing significantly improved accuracy compared to previous specialized tools.
This cutting-edge model can predict the structures of complexes containing almost all types of molecules within the Protein Data Bank. This ability to computationally analyze the complex interactions between proteins and other molecules will expand researchers' understanding of biological processes and potentially advance drug development.
AlphaFold was first introduced in 2020, and both the iterative version AlphaFold 2 could predict the 3D structure of proteins based on their amino acid sequences. The subsequent AlphaFold-Multimer further advanced the prediction of protein-protein complexes. However, due to the vast differences in interactions among different types of proteins, predicting complexes using a single deep learning model has been challenging.
In the new study, John Jumper and colleagues from DeepMind reported that with significant improvements in the deep learning architecture and training system of the AlphaFold 2 model, it is now possible to make more accurate predictions of the structures of a wide range of biological molecular systems within a unified framework.
AlphaFold 3 can predict complexes of proteins with other proteins, nucleic acids, small molecules, ions, modified protein residues, as well as antibody-antigen interactions. Its prediction accuracy surpasses current prediction tools, including AlphaFold-Multimer.
However, researchers also highlighted some limitations, such as around 4.4% of structures exhibiting incorrect chirality (a type of symmetry property). They added that further improvements in simulation accuracy would require generating a large prediction set and ranking the predicted structures, incurring additional computational costs.
For more information, refer to the related paper: https://doi.org/10.1038/s41586-024-07487-w
