Google DeepMind today detailed a new version of AlphaFold, an artificial intelligence model it has built to help researchers study biological molecules such as proteins.

The Google LLC unit says the latest iteration of the model provides more accurate data about proteins’ structure than its predecessor. Furthermore, AlphaFold can now study a broader range of biological molecules. DeepMind believes that the model’s enhanced capabilities could help advance research in areas such as drug discovery. 

Proteins, the building blocks of life, are complex molecules that twist and fold into various shapes. The structures into which a protein configures itself directly influence its behavior. As a result, studying those structures is a major priority for scientists. 

Manually determining the shape of a protein can require years of work in some cases. As a result, scientists have long sought to develop software that can automate the task. Three years ago, DeepMind became the first to achieve that goal with its AlphaFold model, which demonstrated the ability to predict the structure of proteins in a matter of days.

The new version of AlphaFold that the Google unit debuted today offers enhanced prediction capabilities. According to DeepMind, it can estimate the shape of not only proteins but also other biological molecules.

The upgraded iteration of AlphaFold lends itself to, among other tasks, predicting the structure of so-called ligands. Those are molecules that can bind to a protein and cause changes in the way that protein functions. Ligands play an important role in cell signaling, a key biological process through which cells influence one another’s behavior. 

AlphaFold can estimate the structure of other molecules as well. Among those molecules are nucleic acids, a family of compounds that includes DNA and RNA. Moreover, DeepMind claims that AlphaFold can now not only calculate the shape of more molecules but also do so with increased accuracy.

When a ligand attaches, or binds, to a protein, the combined structure is known as a protein-ligand complex. Researchers historically evaluated the shape of such complexes with a method called docking. This method can only be used if there’s a significant amount of data available about the protein component of a protein-ligand complex.

According to DeepMind, the new version of AlphaFold can predict the shape of protein-ligand complexes more accurately than the best docking methods. Furthermore, it does so while requiring significantly less data than those methods. As a result, AlphaFold could potentially make it easier for scientists to study newly discovered protein-ligand complexes on which little information is available. 

DeepMind says the AI model provides increased accuracy in other areas as well. According to the Google unit, AlphaFold can predict the structure of nearly all the molecules in Protein Data Bank, a widely used scientific database. DeepMind claims the model often generates those predictions with “atomic accuracy.”

“Early analysis also shows that our model greatly outperforms AlphaFold2.3 on some protein structure prediction problems that are relevant for drug discovery, like antibody binding,” DeepMind researchers wrote in a blog post. “Additionally, accurately predicting protein-ligand structures is an incredibly valuable tool for drug discovery, as it can help scientists identify and design new molecules, which could become drugs.”

Scientists are already using AlphaFold to support research projects. DeepMind disclosed today that more than 1.4 million users have accessed the AlphaFold Protein Structure Database, which contains protein structures generated by the AI model. Furthermore, DeepMind spinoff Isomorphic Labs has incorporated AlphaFold into its drug discovery efforts. 

Image: DeepMind