A Novel Deep Siamese Convolution Network for Detecting Fentanyl Analogs From Mass Spectra

Abstract

Mortality rates have risen dramatically in recent years due to the misuse of fentanyl and its analogs. Due to the easy synthesis and rapid emergence of various fentanyl analogs, an accurate detection model is particularly desirable. The existing classifiers cannot meet the requirements for their accurate detection. For the small sample size detection problem of fentanyl analogs of electron impact (EI) or electrospray ionization (ESI) mass spectra, a novel mass spectra classification model based on deep Siamese convolutional network (DSCN) was proposed. First, the input mass spectra are augmented to be the input mass spectral pairs. Second, 1D CNN is involved in the Siamese network to extract the spectral features. Finally, the classification network based on FC layers and Softmax layer is used to detect the fentanyl analogs. Contrastive loss function and cross-entropy loss function are combined to train the network parameters of DSCN. Experimental results show that, compared with other machine learning and deep learning methods, the proposed DSCN can achieve better performance on the detection of fentanyl analogs.