Zhi DongXi (public number: zhidxcom)
<b>Compile</b> the | Cheng Qian
<b>Edited by | Li Shuiqing</b>
On Monday, according to foreign media NewScientist, a team of researchers at Columbia University in Canada created a machine learning tool, DarkNPS, which can use AI to quickly screen for new psychoactive substances (NPS), which are stimulant drugs similar to cocaine and heroin.
Due to the chemical diversity of NPS and its short presence on the illicit market, the current detection measures for the drug face significant challenges. DarkNPS can greatly shorten the recognition time of emerging NPS and effectively shorten the search scope of its chemical structure.
The paper was published in Nature's sub-journal Nature Machine Intelligence.
Link: https://www.nature.com/articles/s42256-021-00407-x
New psychoactive substances, also known as "planned drugs" or "laboratory drugs", are drug analogues obtained by criminals from the chemical treatment of controlled drugs to evade crackdown, and have similar or stronger effects such as excitement, hallucinogenicity, and anesthesia to controlled drugs.
Michael Skinnider, a member of the research team at the University of British Columbia in Canada, and his colleagues created a machine learning tool called DarkNPS that performs simple analysis through AI tools to quickly extract possible molecular structures of the chemical structure of psychoactive "novel drugs."
The tool can rapidly develop laboratory testing services to screen for drug use that has similar effects to drugs such as cocaine and heroin and cannot be detected in current tests.
The team collected the chemical structures of about 1,700 known novelly designed drugs from forensic laboratories around the world and trained them. The training set includes tandem mass spectrometry results for each drug, that is, translating and analyzing some of its recombinant protein data, which can provide information about the molecular mass and the elements it contains. This allows the AI to recognize patterns between tandem mass spectrometry data and chemical structures.
Skinnider said: "Our method can shorten the time required to identify NPS from weeks or months to several hours. ”
Given tandem mass spectrometry data for a previously unknown NPS, DarkNPS can guess its molecular structure. The accuracy of the process is about 51%. If ai can predict the top 10 items of its molecular structure, its accuracy will increase to 86%, which means that the tool can effectively narrow the search for molecular structure.
Skinnider said: "This not only saves a lot of time, but also identifies new psychoactive substances that are newly listed more quickly. ”
The team said in the paper that DarkNPS can automatically elucidate the chemical structure of unknown NPS using only mass spectrometry data, and generate models based on the deep layers of chemical structures.
The team's models span the fields of chemistry and deep learning, as they have the potential to generate molecules with arbitrary physico-chemical or biological properties as needed, solving the so-called "reverse design" problem.
Much of this work focuses on analyzing the molecular structure possibilities of specific drug countermeasures that are active.
The research team said they sought to generate NPS-like molecules that matched the characteristics of one or more analytical measurements over the course of the study, and used strategies appropriate for dealing with low-level data stores to achieve this.
A robust generation model of designing a drug is learned from about 1700 examples and sampled from this model to randomly generate new molecular structures similar to existing NPS storage spaces.
The research process showed that the frequency with which new molecules were sampled from the model could be used to indicate the chemical structure that most likely explained the precise mass observed. The resulting structure is combined with tandem mass spectrometry data to further improve the accuracy of structure analysis.
DarkNPS was the first to elucidate the molecular structure of NPS, which first appeared in Europe in February 2021.
The end result of this model is the establishment of a deep generation model of novel psychoactive substances.
Many computational tools have been developed to automatically identify drugs and their metabolites in mass spectrometry data. However, all of these tools require a database of known chemical structures with which the observed mass spectrometry data can be compared. Therefore, these tools cannot be used to identify new synthetically designed drugs that are not found in existing databases.
The team reasoned that by generating a new, nuclear power source-like database of chemical structures, completely unknown nuclear power sources could be automatically identified. So the researchers set out to learn a deep generative model of the chemical structure of the nuclear power source, from which new nuclear power source structures were randomly sampled (Figures 1a-b).
The latest and most comprehensive resource database currently containing NPS structure is HighResNPS, a database for NPS screening, with contributors from dozens of forensic laboratories around the world submitting data to HighResNPS when they detect new substances in biological samples or items seized by law enforcement. Still, at the beginning of June 2020, the database contained only 1753 unique NPS structures.
The limited size of this dataset reflects the number of nuclear power sources that appear on the illicit market and are subsequently detected by forensic laboratories. However, the datasets typically used to train chemical structure generation models will contain hundreds of thousands of pieces of data.
After that, the research team obtained data on 194 NPS drugs and found that 176 of them appeared in the data generated by the AI. The researchers are also using AI to extrapolate 100 million possible chemical structures to study drugs that could be generated in the future.
In addition, researchers at Columbia University's Melman School of Public Health are investigating the use of machine learning to evaluate the law and its relationship to prescription opioid dispensing patterns.
"Machine learning methods are increasingly being applied to similar high-dimensional data problems and may provide a complementary approach to other forms of policy analysis, including as a screening tool to identify policy and legal provision interactions that require further attention." Silvia Martins, an associate professor of epidemiology at Columbia University, said.
The special chemical structure of NPS makes the illegal molecule have an opportunity, and the AI model can quickly identify its molecular structure, and use AI to pre-infer its possible situation from the molecular structure of the drug, preventing some drugs from flowing into the market in advance. This provides a new paradigm for the application of AI tools to assess similar drug molecular structures, drug dispensing patterns, and other issues.
Source: New Scientist, Columbia University website