AI-Based Machine Learning Approach for Detection of Novel Amphetamine Derivatives in LC-HR-MS/MS analysis

Nir Cohen, Department of Analytical Chemistry, Bar-Ilan University, Ramat-Gan, Israel (nirc@iibr.gov.il)

The emergence of novel psychoactive substances continues to challenge forensic toxicology and public health. While NPS are poorly documented in mass spectral (MS) libraries, machine Learning can fill-in gaps and extend the range of current databases. We present a graph neural network (GNN)-based deep learning model that rapidly detects untargeted amphetamine derivatives, using high-resolution tandem mass spectrometry data without the need for a reference spectral library.

A data set of high-resolution MS/MS spectra in varying HCD energies of 135 amphetamine derivatives and over 12,000 non-amphetamines with amphetamine-like chemical formula was prepared from the NIST23 library. The spectra were transformed to chemical composition of the fragments, from which a fully connected graph was created and processed by a GNN. The spectra of 80% of the amphetamines and non-amphetamines in the data set, randomly chosen, were used for the training of our model. After the model learned the characteristics of ESI-MS/MS spectra belonging to amphetamine derivatives (without any human intervention), the model was tested on the other 20% of the compounds. The model achieved performance of recall of 0.83 and precision of 0.22, showing its capabilities for detection of novel amphetamine derivatives even with major structural changes compared to its training data.