Machine learning model for detecting fentanyl analogs from mass spectra

In recent years, fentanyl and its analogs have been increasingly abused, leading to tragic outcomes. One way to tackle this problem is to rapidly detect fentanyl analog compounds and prevent them from reaching people who could be harmed by them. However, the emergence of novel fentanyl analogs that evade detection by classic mass spectral library matching has exacerbated the problem. We propose supervised machine learning classification models as a complementary approach to library matching for detecting fentanyl analogs from mass spectra. To develop and apply such models, we extract two dozen peak-based and similarity-based input features from each spectrum of interest. Using techniques such as random forests, neural networks, and logistic regression, we identify patterns within these features' values, resulting in strong detection performance. Within a crossvalidation framework, we achieve 99% probability of detection and 1% probability of false alarm on a representative set of several thousand mass spectra. These results suggest that machine learning models may offer a robust complement to library matching. Practitioners from diverse fields, including border security, law enforcement, and military may benefit from this capability to detect drugs of abuse.