Roughly one-third of all U.S. Food and Drug Administration–approved medicines, with collective annual sales of nearly $200 billion, target a single family of cell membrane proteins, called the G protein–coupled receptors (GPCRs), which deliver chemical messages in and out of cells. A growing number of the drugs are antibodies, which can lock tightly to specific proteins. But GPCRs are a tough target to hit, even for antibodies, because they barely protrude beyond the cell membrane.
Increasingly, antibody designers are using artificial intelligence (AI) to take aim. Researchers at Nabla Bio, a Massachusetts-based biotech company, report today that their new AI has, in a matter of months, conjured up dozens of GPCR-targeting antibody candidates, which promise to work as well as drugs that have spent years in the traditional pipeline. One marks a first for an AI-designed protein: It can turn on cell membrane signaling rather than blocking it.
“If the results are solid, it’s a breakthrough,” says Wei Wang, a biochemist at the University of California San Diego. David Baker, a protein design expert at the University of Washington and co-founder of Xaira Therapeutics, a California-based antibody design startup, says it’s an example of how quickly the field is accelerating. “It’s a very exciting time,” Baker says.
Antibodies, like all proteins, consist of chains of amino acids that fold up into fantastically complex 3D shapes, which govern what they bind to. In the body they block pathogens from entering cells or tag diseased cells for the immune system. But they can also be used to block disease-related proteins or deliver drugs that are attached as payloads. More than 160 engineered antibodies are approved for treating cancer, infections, and immune-related diseases. And with thousands of new versions in development, the market is expected to balloon to $455 billion a year by 2028, according to a 2022 analysis in Antibody Therapeutics. As Baker puts it, “Antibodies are the coin of the realm for the pharmaceutical industry.”
But designing antibodies is laborious. Candidates typically go through multiple rounds of improvements to ensure they have all the properties of good drugs, such as the ability to dissolve in bodily fluids and a specificity that prevents them from binding to unwanted sites and causing side effects. AI represents a way to speed up the discovery process. “Now, you can build all those [properties] in from the beginning,” Baker says.
An early example of the promise came in November 2024, when researchers led by Surge Biswas, Nabla Bio’s CEO, reported that their AI discovered antibodies that bind to CXCR7, one in the family of some 800 GPCR membrane proteins. That was an accomplishment, Biswas says, but the AI’s creations were “not that competitive with traditional antibodies.”
To improve its designs, the Nabla team tweaked its AI approach, borrowing an idea known as “test-time scaling,” a process of reasoning and iteration that OpenAI developed for its large language model ChatGPT. The AI begins by generating multiple proposed solutions to a problem, then sifts them through multiple reasoning steps until a final answer is produced. Now, Biswas says, “We’re bringing this idea to biology for the first time.”
Better results followed quickly. In a preprint posted today on the company’s website and expected soon on the bioRxiv server, the Nabla Bio team reports its AI designed tens of thousands of GPCR-binding antibodies. Lab studies showed dozens of them had “affinities”—a measure of their binding strength—as high or higher than existing antibody drugs that took years to develop. For example, some could target CXCR7 while not affecting a closely related GPCR called CXCR4, a level of discrimination most GPCR drugs struggle with. But perhaps most impressive was one that didn’t block its target, but instead turned it on. “If you have the ability to turn [GPCRs] on or off, you basically can control cellular biology and the disease state,” Biswas says. The AI was able to learn from that single surprise success to create hundreds more candidates that flick switches on.
“That’s impressive if it’s validated,” says Andrew Bradbury, chief scientist at Specifica, a biotech company that designs antibody libraries for pharma companies. When it comes to AI antibody drug design, he says, “there is a lot of froth and a lot of money.”
But AI is proving adept at designing antibodies aimed at other targets. In February, Baker and his colleagues reported the AI-aided discovery of antibodies that bind to an influenza protein common to all strains, an achievement that could open the way to a universal flu drug. The team also reported antibodies that block a potent toxin produced by the bacteria Clostridium difficile, a common and deadly hospital acquired infection.
Last fall, researchers at Absci, a biotech firm based in Washington state, reported designing the first antibody capable of binding to a protein target on HIV known as the caldera region. The region exists in all HIV strains, meaning it could lead to an all-purpose antibody drug against HIV. The company is also designing antibodies to treat endometriosis, inflammatory bowel disease, and even hair loss, says company co-founder Sean McClain.
The progress “is incredible,” McClain says. With so many possible antibody targets and therapies, there’s room for all of the companies now jumping in, McClain says. “It takes a village.”
More: https://www.science.org/content/article/ai-conjures-potential-new-antibody-drugs-matter-months
