Nobel Prizes usually recognize decades-old achievements. But this year’s prize in chemistry was awarded in part last week for very recent work that is just beginning to bear fruit: using artificial intelligence (AI) to design proteins never seen before. Proteins are the workhorse molecules of life, with millions existing in nature, but novel ones could transform medicine and technology. The new tools have already enabled researchers to churn out designer proteins for vaccines and cancer treatment, artificial pollution-eating enzymes, and molecular assemblies capable of seeding the growth of minerals. “We are just at the beginning of learning what we can build,” says Possu Huang, a protein designer at Stanford University.
Half of this year’s $1.1 million prize went to John Jumper and Demis Hassabis of Google’s DeepMind for their work in devising AlphaFold, an AI program that has all but solved the notorious protein folding problem: predicting a protein’s shape—and hence the function—from its chemical sequence. In 2020, Jumper and Hassabis showed that AlphaFold 2, trained on huge databases of protein structures and their animo acid sequences, was in many cases as good at predicting protein shapes as techniques that directly image them, such as x-ray crystallography. The other half of the prize went to David Baker of the University of Washington (UW) for tackling the inverse problem: going from a protein’s desired function to the amino acid sequence that would fold up into a molecule able to do that job.
The idea of making a novel protein to order “was kind of out there on the lunatic fringe,” Baker says. But in 2003, he and his colleagues showed it was possible with software called Rosetta, which combed through databases of known protein structures to look for bits that could be useful in a new, hypothetical protein.
In an early demonstration, Rosetta spit out a sequence of 93 amino acids that would in theory make a protein called Top7, which had a shape not used by biology. To confirm the design, Baker’s team synthesized a gene coding for Top7 and spliced it into bacteria. They harvested the produced protein and bombarded it with x-rays to determine that its structure was nearly as predicted. Although Top7 didn’t carry out any important function, the implications were revolutionary. “We can now design almost any protein shape we want,” says Casper Goverde, a protein designer at the Swiss Federal Institute of Technology Lausanne.
Since Baker’s early experiments, protein design software has incorporated increasingly powerful AI techniques. In June, for example, Huang’s team reported a model, called Protpardelle, geared to designing not just a protein’s general “backbone,” but also the specific clusters of atoms along its fringes—the so-called “side chains” critical for function. And earlier this year researchers led by Bonnie Berger, a computer scientist at the Massachusetts Institute of Technology, unveiled software called OmegaFold that’s better at designing “orphan” proteins for which there are few close cousins in nature to guide the design process. “Things are moving very quickly nowadays,” Berger says.
Vaccines have been an early payoff. In 2020, shortly after the COVID-19 pandemic began, UW researchers designed proteins that attached to a specific part of SARS-CoV-2’s spike protein and blocked the virus from penetrating human cells. Identifying this part of the spike protein enabled them to design a vaccine that arrayed dozens of copies of the critical protein part around a protein core to train the immune system to recognize and inactivate the same structures on SARS-CoV-2. After successful human trials, the vaccine, called SKYCovione, was approved last year for use in South Korea and the United Kingdom, although production has now been shelved because of the pandemic’s decline. UW researchers are working on other vaccines including a broad-spectrum flu shot that might eliminate the need for annual boosters and a vaccine against respiratory syncytial virus, a major killer of infants and elderly people.
Designers are also developing proteins to seek out and bind to distinctive molecules on the surface of cancer cells, tagging them for destruction by chemotherapy drugs—delivered, naturally, by designed viruslike protein containers. But tumor cells, like all cells, are surrounded by a fatty membrane made of insoluble proteins. That makes it tough for researchers to test drugs in solution—whether chemotherapies or vaccine-induced antibodies—that can best attack them. In June, Goverde reported redesigning membrane proteins to make them soluble, while preserving all their customary functions. “We can then use those to find antibodies that target the real thing,” Goverde says.
Tumors are not the only medical target. In a May preprint, Baker and colleagues reported designer proteins that can attach to the toxins in the venom of snakes such as cobras, preventing them from binding to nerve receptors. When injected into mice, the proteins protected the animals from an ordinarily lethal dose of venom by neutralizing the toxins. The designed proteins are small, making them more stable than traditional large proteins, which decompose quickly if not refrigerated. The researchers envision a penlike injector to be carried for use immediately following a snake bite.
Nonmedical applications are also emerging. In 2018, for example, Parisa Hosseinzadeh, now at the University of Oregon, and her colleagues reported designing a protein catalyst that could safeguard food production against contaminants by helping capture atoms of toxic metals. Hosseinzadeh’s group is now working on enzymes to break down plastics in the environment. And last year, Sarel Fleishman, a protein designer at the Weizmann Institute of Science, and his colleagues tried to improve on nature by making new enzymes that could help convert agricultural waste into biofuels. They looked for the best components of natural enzymes called xylanases, which microbes use to break down plant cell walls, and mixed and matched them to produce thousands of new xylanases. “We’ll see more and more efforts to use protein design to tailor enzymes to do the jobs we want them to do,” Hosseinzadeh says.
AI protein design could benefit the environment in other ways. Baker’s team has already shown it’s possible to improve the efficiency of enzymes that capture carbon dioxide, an advance that could lead to better smokestack scrubbers to fight climate change. And he says they’re now gearing up to see whether they can design enzymes to capture methane, an even more potent greenhouse gas.
Even more distant, Huang’s group has begun to think about re-engineering a protein called myosin, which drives muscle contraction, to be powered with light instead of ATP, the body’s normal chemical fuel. If successful, the effort could eventually lead to light-powered artificial muscles.
“That’s more science fiction at this stage,” Huang says. At least for now. But at the rate protein design is progressing, perhaps not for long.
More: https://www.science.org/content/article/ai-designer-proteins-could-transform-medicine-and-materials
