Can chemical AI help solve the fentanyl crisis?

The notion that computer programs with an understanding of chemistry could be used for nefarious purposes has troubled Bartosz Grzybowski even since he watched the 1996 action film The Rock more than a dozen years ago.

The film centers on the threat of a chemical weapons attack from rogue military agents based at Alcatraz. Just how difficult, Grzybowski wondered, would it be to make a nerve agent from readily available chemicals?

Grzybowski, who holds faculty positions in chemistry at the Polish Academy of Sciences and UNIST, is well placed to work that out. He has long been a leader in using computers to plan chemical syntheses and was behind the development of the chemical retrosynthesis platforms Chematica—which is now known as MilliporeSigma’s Synthia—and Allchemy.

Grzybowski’s chemical weapons thought exercise did not put him at ease. When he and his team used their algorithm to study the synthesis of VX, they found that the deadly nerve agent could be made from simple household products like clarifiers, garden fertilizers, and kitchen salt (Angew. Chem., Int. Ed. 2012, DOI: 10.1002/anie.201202210).

Back when Grzybowski did that work in 2012, his program could only plan syntheses using published chemistry. Now, he says, programs like Allchemy can use artificial intelligence to find new ways to make known targets. And, perhaps more significantly, Grzybowski and his Allchemy coworkers have used encyclopedias, handbooks, and the patent literature to train their algorithms to generate scalable syntheses. These syntheses aren’t limited to the laboratory creation of a few milligrams, as commonly reported in the literature. Instead, they can be done on kilogram or metric ton scale.

That ability to search for scalable routes got Grzybowski thinking about another deadly chemical: fentanyl, the synthetic opioid that’s been linked to more than 70,000 annual overdose deaths in the US in recent years, according to the National Center for Health Statistics. What would happen if he asked Allchemy to look for scalable syntheses of fentanyl using inexpensive, readily available starting materials? Would it come up with routes beyond what’s already used to make this opioid?

“This is such a simple molecule, after all,” he says.

The results were worrisome. Within 15 min, the program identified 254 scalable syntheses of fentanyl that used 234 unique molecules. When asked to bypass any routes that use chemicals regulated by the US Drug Enforcement Administration (DEA), the program still found 166 fentanyl syntheses (Chem 2024, DOI: 10.1016/j.chempr.2024.03.025).

“Many of these things are popular chemicals, and you cannot just ban them,” Grzybowski says. For example, the algorithm found a method for making fentanyl that starts with 4-hydroxypyridine. Although this is an aromatic heterocycle, whereas fentanyl’s core heterocycle is saturated, the program suggested that a simple salt formation followed by reduction can produce the desired fentanyl intermediate. In fact, the method is described in a Chinese patent from 2015.

Rick A. Bright, who was director of the US Department of Health and Human Services’ Biomedical Advanced Research and Development Authority from 2016 to 2020 and is now CEO of Bright Global Health, says reading Grzybowski’s Chem article made him think about the power of chemical AI and who should be using it.

Although it’s tempting to want to lock down such technology, Bright says that would be a mistake. “By locking and handcuffing the good people, you don’t stop the bad people,” he says.

Gordon Broderick, a chemical engineer at the University of Saskatchewan who uses AI to study immune dysfunction, agrees. AI, he says, is like a hammer. “You can use it to build a house, or you can use it to bludgeon somebody.”

Bright and Broderick wrote reactions to Grzybowski’s article that appeared alongside it in Chem (2024, DOI: 10.1016/j.chempr.2024.03.028 and 10.1016/j.chempr.2024.03.026). Both argue that so much good can come from chemical AI that the technology shouldn’t be constrained.

For example, Bright thinks the technology could be used to help ease drug shortages. And Broderick notes that he wants to use chemical AI for cancer research. “If you prohibit the use of that algorithm, maybe there’s a problem I can’t solve that would have helped thousands of families,” he says.

The fentanyl problem is multifaceted, Bright says, and requires a multipronged approach. It’s possible, he says, that the DEA and other government agencies could use chemical AI to track more chemicals that could be used to make fentanyl.

In fact, Bright, Broderick, and Grzybowski all point out that chemical AI technology could help the government track fentanyl starting materials. AI won’t eliminate the problem, Grzybowski says, but it will give the government more information to monitor who might be making the drug.

In addition to alerting authorities to possible fentanyl starting materials, chemical AI can indicate what byproducts each synthesis might create, providing a synthetic fingerprint for each route. This sort of fingerprinting is “potentially tremendously helpful,” says Edward Sisco, a chemist at the US National Institute of Standards and Technology who uses trace analysis to study illicit drugs, including fentanyl.

Such fingerprinting could point to chemical sources and suppliers, Sisco says. And this technology could be translated to any other drug.

Sisco wonders how effective the fingerprinting would be with confiscated street drugs that have been cut with other chemicals. “Can we pick up those signatures even in the presence of all those other compounds?” he asks.

Back in 2012, when Grzybowski published his work on the synthesis of chemical weapons like VX, he says he met with US military officials, but ultimately, nothing happened. Grzybowski says he’s been in contact with some US government agencies about his recent work on fentanyl but can’t go into details. The DEA and the US Department of Homeland Security did not respond to multiple requests for comment on Grzybowski’s work.

Now that we know what chemical AI can do, Grzybowski says, “The question is this: Is the intelligence community going to act?”