The consistent emergence of novel psychoactive substances (NPS) and slightly altered versions of well-known drugs of abuse puts forensic toxicologists and law enforcement at a continual disadvantage. For many in the industry, it feels like just when they resolve the structure for a new group of NPSs, yet another cluster of unknowns present themselves.
Now, researchers at the University of British Columbia have taken a proactive, rather than reactive, approach to the problem—they have successfully trained AI to predict new drugs before they hit the market.
In a study published in Nature Machine Intelligence, UBC medical student Michael Skinnider and his team used a database of known psychoactive substances contributed by forensic laboratories around the world to train an artificial intelligence algorithm on the structures of these drugs. Based on the training, the deep neural network model generated about 9 million potential designer drugs.
The team then tested the molecules against 196 new designer drugs that emerged after the model was trained. Incredibly, more than 90 percent of the new drugs were present in the generated set, meaning the model was able to predict nearly all of the new drugs discovered since it was trained.
“The fact that we can predict what designer drugs are likely to emerge on the market before they actually appear is a bit like the sci-fi movie Minority Report, where foreknowledge about criminal activities about to take place helped significantly reduce crime in a future world,” said senior study author David Wishart, a professor of computing science at the University of Alberta. “Essentially, our software gives law enforcement agencies and public health programs a head start on the clandestine chemists, and lets them know what to be on the lookout for.”
Even with this foresight, forensic toxicologists are still left with the problem of identifying unknown substances. Typically, it takes months for chemists to elucidate the spectrum of a novel, unknown drug. Given the success of the AI system in forecasting new drugs, the researchers wondered if it could also be used to determine what an unknown drug is—based solely on its mass.
Using only the mass of the 196 new designer drugs, the AI model ranked the correct chemical structure of an unidentified designer drug among the top 10 candidates 72 percent of the time. When the researchers integrated tandem mass spectrometry data, this improved to 86 percent. With just one guess, the AI model demonstrated it could predict the correct structure more than half (51 percent) of the time.
“It was shocking to us that the model performed this well, because elucidating entire chemical structures from just an accurate mass measurement is generally thought to be an unsolvable problem,” said Skinnider. “Narrowing down a list of billions of structures to a set of 10 candidates could massively accelerate the pace at which new designer drugs can be identified by chemists.”
But the model doesn’t just have forensic implications. Skinnider says the same type of AI could be used to discover any kind of new molecule, including previously unknown kinds in human blood and urine that could lead to medical advancements.
“There is an entire world of chemical ‘dark matter’ just beyond our fingertips right now. I think there is a huge opportunity for the right AI tools to shine a light on this unknown world,” said Skinnider.
Photo credit: Raimond Klavins/Unsplash