Single-Molecule SERS Discrimination of Proline from Hydroxyproline Assisted by a Deep Learning Model

Abstract

Discriminating low-abundance hydroxylation is a crucial and unmet need for early disease diagnostics and therapeutic development due to the small hydroxyl group with 17.01 Da. While single-molecule surface-enhanced Raman spectroscopy (SERS) sensors can detect hydroxylation, subsequent data analysis suffers from signal fluctuations and strong interference from citrates. Here, we used our plasmonic particle-in-pore sensor, occurrence frequency histogram of the single-molecule SERS spectra, and a one-dimensional convolutional neural network (1D-CNN) model to achieve single-molecule discrimination of hydroxylation. The histogram extracted spectral features of the whole data set to overcome the signal fluctuations and helped the citrate-replaced particle-in-pore sensor to generate clean signals of the hydroxylation for model training. As a result, the discrimination of single-molecule SERS signals of proline and hydroxyproline was successful by the 1D-CNN model with 96.6% accuracy for the first time. The histogram further validated that the features extracted by the 1D-CNN model corresponded to hydroxylation-induced spectral changes.