The artificial intelligence (AI) revolution in protein structure prediction continues. Only 1 year ago, software programs first succeeded in modeling the 3D shapes of individual proteins as accurately as decades-old experimental techniques can determine them. This summer, researchers used those AI programs to assemble a near-complete catalog of human protein structures. Now, researchers have upped the ante once again, unveiling a combination of programs that can determine which proteins are likely to interact with one another and what the resulting complexes— crucial engines of the cell—look like.
“It’s a really cool result,” says Michael Snyder, a systems biologist at Stanford University. “Everything in biology works in complexes. So, knowing who works with who is critical.” Those relationships were hard to reach with previous techniques. The new ability to predict them, he says, should yield an array of insights into cell biology and possibly reveal new targets for the next generation of therapeutic drugs.
Mapping proteins’ shapes down to the atomic scale has until recently required costly and slow experimental techniques, such as x-ray crystallography and nuclear magnetic resonance spectroscopy. Those experimental techniques, if they work at all, typically only produce individual protein structures.
Computer modeling experts have worked for decades to speed things up. Their recent success has depended on deep learning algorithms, which use databases of experimentally solved protein structures to train software programs how to predict structures for proteins based on their amino acid sequences.
Last year, two groups, one from a U.K. company called DeepMind and the other led by David Baker at the University of Washington, Seattle, created rival AI programs that both now churn out predicted protein structures by the thousands. The software also produced structures for a handful of known protein complexes, mostly in bacteria. But in eukaryotes—organisms from yeast to people—the interacting partners are often unknown. Identifying them and predicting how they come together in a complex was too high a bar for the original programs.
Now, both research groups have tweaked their programs so they can solve structures of protein complexes by the hundreds. Online today in Science, Baker and his colleagues use a combination of AI techniques to solve the structures of 712 complexes in eukaryotes.
To find proteins that may form complexes together, the team began by comparing the amino acid sequence of all 6000 yeast proteins to those from 2026 other fungi and 4325 other eukaryotes. The comparisons allowed the researchers to track how those proteins changed over the course of evolution and identify sequences that appeared to change in tandem in different proteins. The researchers reasoned that those proteins might form complexes, and that they changed in step to maintain their interactions. Then the team used its AI program, called RoseTTAFold, along with DeepMind’s AlphaFold, which is publicly available, to attempt to solve the 3D structures of each set of candidates. Out of 8.3 million identified coevolving yeast protein pairs, the AI programs identified 1506 proteins that were likely to interact and successfully mapped the 3D structures of 712, or about half.
“These interactions span all processes of eukaryotic cells,” says team member Qian Cong, a biomedical informatics expert at the University of Texas Southwestern Medical Center. Among the highlights, Cong and Baker say, are structures for protein complexes that allow cells to repair damage to their DNA, translate RNA into proteins in ribosomes, pull chromosomes apart during cell reproduction, and ferry molecules through the cell membrane.
“It’s a great example of the promise” of 3D structures, says DeepMind’s John Jumper, one of AlphaFold’s lead developers. By revealing precisely how proteins interact with one another, the models should help biologists visualize how previously unknown complexes carry out a multitude of jobs within the cell.
“These models give hypotheses for experimentalists to test,” Cong says. And because disrupting these interactions could offer new ways to intervene in a wide variety of diseases, she adds, “it also gives you a lot of potential new drug targets.”
More are likely on the way. Last month, Jumper and his colleagues posted a preprint on the bioRxiv server describing a new version of their AI, dubbed AlphaFold-Multimer, which mapped structures of 4433 protein complexes. Analyses within the AI program that gauge the confidence level of each fold suggest the structures were accurate up to 69% of the time. The bottom line, Baker says: “It’s really an exciting time for structural biology.”