From Mass to Class: Classification of Amphetamines MS/MS Spectra via Graph Neural Networks

Identifying
emerging novel psychoactive substances (NPS), using
liquid chromatography high-resolution tandem mass spectrometry (LC-HR-MS/MS),
is often hindered by their frequent absence from mass spectral reference
libraries. We developed a graph neural network (GNN) classification
model to detect amphetamine derivatives, a subclass of NPS, based
on their fragmentation patterns. The data set was generated by transforming
high-resolution mass spectra from mass/charge (m/z) values into the chemical compositions of the fragments,
from which fully connected graphs were constructed. Spectra were extracted
from the NIST23 database, using only protonated HCD spectra to ensure
consistency, and collision energy was included as a parameter to enhance
spectral informativeness. The network architecture and hyperparameters
were systematically optimized, including network layout and pooling
layer type. To address the class imbalance in the data set, where
amphetamine derivatives were outnumbered 100:1 by nonamphetamines,
the loss function was adjusted to penalize false negatives 50 times
more than false positives. Evaluation of model performance on a randomly
selected 10% test subset yielded a recall of 0.86 and a precision
of 0.11. The model successfully detected the spiked amphetamine derivative
in a human plasma sample, achieving a 1% false positive rate and demonstrating
its applicability to complex biological matrices.