A pill bought on the street may contain a substance never seen before — one not in any database, not on any official list, and not detectable by standard tests. The drug supply evolves faster than the tools designed to catch it. A method reported on bioRxiv.org on April 27 might change this dynamic. [1] It does not rely on a catalog of known compounds. Instead, it measures the physical traits of a pill and compares them to a computer-generated library of billions of possible chemical variants. This approach could identify a fentanyl variant before authorities have ever cataloged it.
Illicit laboratories synthesize new forms of dangerous drugs. These variants slip under the radar because they are not in the reference libraries that forensic labs use. The only way to confirm a fentanyl variant is to compare it to a pure sample analyzed in a lab. But there are billions of possible forms of fentanyl. Experts know only about 60,000 of them. Biochemist David Wishart of the University of Alberta in Edmonton, Canada, who was not involved with the work, described it as a ‘whack-a-mole problem.’ [1] The system is always one step behind.
The new method breaks this cycle. Bioanalytical chemist Tom Metz of Pacific Northwest National Laboratory in Richland, Wash, and his colleagues set out to eliminate the need for traditional reference libraries. [1] In prior work, they used two customized instruments to identify chemical features shared by fentanyl compounds. They also learned to distinguish between many unrelated molecules that share fentanyl’s molecular mass. All fentanyls have a common core chemistry. Labs can vary the surrounding chemical groups. Metz compared it to a Christmas tree. The tree is nearly always a pine tree of some sort, but each household decorates it differently. The instruments give clues to the precise elements that make up molecules. They also reveal how the molecules are structured and what shape they adopt during analysis.
The researchers then created a database of hypothetical variants. They computationally broke apart each of the roughly 60,000 known fentanyl and fentanyl-like molecules into a few different fragments. Then they recombined them to create several billion molecules. They eliminated nonsensical and implausible molecules from the digital catalog. For example, they removed molecules unlikely to penetrate the brain’s protective barrier. Finally, with help from machine learning, they predicted what real-world chemical measurements of the dreamed-up structures would look like. They combined that data with the data from the 60,000 known structures. The final digital library contained over 1 billion analogs.
The researchers could not test street drugs. So they created a mock fentanyl pill. It contained traces of 12 commercially available fentanyl varieties. It also included a chemically similar nonopioid decoy. The pill was cut with typical street pill ingredients like caffeine. They completed the feature-identifying measurements. Then they handed the raw data and the computer-generated library to another analytical chemist who had never seen the mock pill. The prompt was: “We suspect there’s fentanyl in this sample. Can you tell us which, if any, analogs are in it?”
The answer was yes. After multiple cycles of narrowing down possible matches, the blinded chemist identified six of the mock pill’s fentanyl components perfectly. Another four were narrowed down to a few possible candidates each. No pure compound library was needed. The remaining two lacked the signatures used for flagging or could not be fully teased apart. The results have not yet been peer-reviewed.
The approach is a “tremendous first step,” according to chemist A. Way Fountain III of the University of South Carolina in Columbia, who was not part of the study. [2] But it relies on customized instruments unavailable to most forensic or national security laboratories. Fountain said the technique should be tested with other classes of drugs or molecules to show where improvements are needed. Such tests are underway. Metz and his colleagues are studying several classes of molecules. They have also identified common features in a new family of lab-made opioids called nitazenes. These nitazenes are becoming prevalent in overdose cases.
Wishart thinks the work will help modernize the forensic community’s approach for identifying unknown compounds. Relying on a reference library of pure compounds, he said, “is still very 19th century thinking.” Molecular pharmacologist Gary Miller of Columbia University, who was not involved with the research, agreed. “Reference-free identification could be revolutionary from a scientific standpoint,” he said. [2] “These data demonstrate that the approach can work.”
The structure of the problem is like a tide. The tide of new drugs rises. The old methods recede. The new method does not try to build a wall against the tide. Instead, it learns to measure the shape of the water.
The method is not yet ready for law enforcement. The instruments are customized. The process requires expertise. But the principle is sound. The digital library of 1 billion analogs is a map of possibilities — not a catalog of what has been seen, but a catalog of what could be seen.
The work is part of a broader effort. Other researchers are also working on reference-free identification. The approach could apply to other classes of drugs or any molecule that has a common core with variable decorations.
Fentanyl is up to 100 times as potent as morphine. Two milligrams is potentially lethal. The United States reported more than 72,000 overdose deaths in 2023 alone. Fentanyl contains no natural ingredients. Underground labs can tweak its structure just enough to avoid detection while retaining its heroinlike effect. This has made it profitable to lace pills with a powerful substance that many users do not know they are consuming.
The old method required a pure sample. The new method requires only a measurement, compared to a computer-generated prediction based on the core chemistry. The prediction is not a guess — it is a calculation using machine learning trained on known structures to predict the measurements of unknown structures.
The test was a mock pill containing 12 fentanyl varieties. The blinded chemist identified six perfectly. Four were narrowed down to a few candidates. Two were not identified. This is not perfect, but it is a start. The method can be improved, the instruments made more accessible, the library expanded, and the machine learning refined.
The beauty of the approach is its simplicity. It does not try to know everything — it tries to know the pattern. The pattern is the core chemistry. The decorations are variable. The instrument measures the decorations. The computer predicts the decorations. The match is the identification.
The work is a step toward a future where the drug supply is not a mystery, where a pill can be tested without a catalog, and where the system is not always behind. The method does not need to have seen a molecule to know it.