Almost five years ago, DeepMind, one of Google’s most prolific AI-focused research labs, debuted AlphaFold, an AI system that can accurately predict the structure of many proteins in the human body. Since then, DeepMind has improved the system and released the latest, more capable version of AlphaFold (AlphaFold 2) in 2020.
And the laboratory research continues.
Today, DeepMind revealed that its latest release of AlphaFold, the successor to AlphaFold 2, can generate predictions for nearly every molecule in the Protein Data Bank, the world’s largest open-access database of biomolecules.
Already, Isomorphic Labs, a spin-off of DeepMind focused on drug discovery, is reportedly applying the new co-designed AlphaFold model to the design of therapeutics. post The DeepMind Blog helps you characterize different types of molecular structures that are important for treating diseases.
New features
The new AlphaFold’s capabilities extend beyond protein prediction.
DeepMind says the model uses not only nucleic acids (molecules that contain important genetic information) and post-translational modifications (chemicals), but also ligands (molecules that bind to “receptor” proteins and cause changes in the way cells communicate). It is claimed that the structure of can also be predicted accurately. changes that occur after a protein is produced).
Image credits: deep mind
Predicting the structure of proteins and ligands can be a useful tool in drug discovery, DeepMind notes, as it helps scientists identify and design new molecules that have the potential to become drugs.
Currently, pharmaceutical researchers use computer simulations known as “docking methods” to determine how proteins and ligands interact. Docking methods require specifying a reference protein structure and a preferred location on that structure for the ligand to bind.
However, the latest AlphaFold does not require the use of reference protein structures or suggested positions. The model can predict previously “structurally characterized” proteins, as well as simulate how proteins and nucleic acids interact with other molecules. DeepMind says this level of modeling is not possible with today’s docking techniques.
“Early analysis also shows that our model is significantly better. [the previous generation of] “AlphaFold investigated several protein structure prediction problems related to drug discovery, such as antibody binding,” DeepMind wrote in the post. “The dramatic jump in the performance of our model demonstrates the potential of AI to significantly enhance our scientific understanding of the molecular machines that make up the human body.”
However, the latest AlphaFold isn’t perfect.
in white paper Researchers from DeepMind and Isomorphic Labs detail the strengths and limitations of the system, saying it is a best-in-class method for predicting the structure of RNA molecules (molecules in the body that carry instructions to make proteins). It became clear that it was not possible.
Without a doubt, both DeepMind and Isomorphic Labs are working to address this issue.