An artificial intelligence (AI)-powered software program released today by Google DeepMind offers scientists a potent new tool to predict how proteins work. Whereas earlier versions of the company’s software could model how the strands of amino acids making up a protein fold into its final 3D shape, the new version reveals how folded proteins bind and interact with a host of other molecules, including DNA, RNA, and other proteins. That interplay determines a protein’s role in the cell. Understanding it will help scientists design drugs that can block or boost a protein’s function.
The new AI, known as AlphaFold 3, follows closely on the heels of RoseTTAFold All-Atom, a related AI-guided software package for predicting interactions between proteins and other biomolecules developed by researchers led by David Baker at the University of Washington and described on 7 March in Science. Baker says DeepMind’s new software, reported today in Nature, “is very impressive” and produces more accurate predictions than his team’s software.
AlphaFold 3 (AF3) is the successor to AlphaFold 2, which was released in 2021 and learns to predict protein folding patterns by training on huge databases of known structures. According to DeepMind, in just 3 years, AF2 has been used by 1.8 million researchers to map out some 6 million different protein structures. But those maps are images of individual static proteins, ignoring the chemical communication going on inside cells. “Biology is a dynamic system,” says DeepMind CEO Demis Hassabis. “You have to understand how properties of biology emerge due to the interactions between different molecules in the cell.”
RoseTTAFold All-Atom led the way in mapping those interactions by using an approach called diffusion, common in AI image generators such as DALL-E, which refines an AI’s output by adding and removing statistical noise. John Jumper, one of AF3’s chief architects, says DeepMind, too, adopted a diffusion approach, but also made additional core changes to its software, in part by deemphasizing code that predicts similar behavior when proteins have close evolutionary ties in favor of predictions based on the physical behavior of amino acid building blocks. The upshot, Jumper says, is that “we’ve seen enormous advances in accuracy over other tools, and even AlphaFold 2.”
As Jumper and his colleagues report today, AF3 could correctly model known interactions between proteins and small, druglike molecules in 76% of the more than 400 cases tested, compared with roughly 40% for RoseTTAFold All-Atom. And for interactions between proteins and antibodies, AF3 was correct 62% of the time compared with 30% for AlphaFold Multimer, the company’s previous software package for modeling protein interactions with other biomolecules.
Julien Bergeron, a biologist at King’s College London who was given early access to test the new AF3 software, calls it “transformative” in its ability to speed up research. Rather than spend years in the lab studying a protein, they can get a result in minutes. “We can start testing hypotheses in silico,” Bergeron says. “I’m pretty certain that every structural biology and protein biochemistry research group in the world will immediately adopt this system.”
To encourage such widespread adoption, DeepMind researchers today also released AlphaFold Server, a free online platform that enables users to create AF3 models of proteins interacting with almost any other biomolecule. It’s a change from its approach with AF2, for which DeepMind researchers released a database containing some 200 million protein structures. But because the combinatorial associations between all these proteins and other biomolecules that bind to them is so vast, “it’s really not feasible to precompute everything and put it in a database,” says DeepMind’s Dhavanthi Hariharan. Instead, the server allows trained users to simply input the amino acid sequence of interest along with the nucleic acid sequence of a DNA or RNA strand or the formula of a small molecule drug. Within minutes, the software spits out images of how they likely interact, as well as confidence scores that rate the likelihood that the model is correct.
The tool is also able to incorporate the effects of “posttranslational” modifications to finished proteins, as well as changes to DNA and RNA known as epigenetic markers. That should allow it to predict how these biochemical tweaks alter protein functions, in some cases causing disease. Those predictions could in turn help scientists develop medicines to prevent or reverse these changes.
Max Jaderberg, chief AI officer of Isomorphic Labs, says of a DeepMind spinoff aiming to use modeling tools to discover drugs: “Understanding more about [protein interactions] will translate to much more effective drugs in the clinic and into the hands of patients.”
