Fast and accurate decoding of Raman spectra-encoded suspension arrays using deep learning

Abstract

A deep learning network called "residual neural network" (ResNet) was used to decode Raman spectra-encoded suspension arrays (SAs). With narrow bandwidths and stable signals, Raman spectra have ideal encoding properties. The different Raman reporter molecules assembled micro-quartz pieces (MQPs) were grafted with various biomolecule probes, which enabled simultaneous detection of numerous target analytes in a single sample. Multiple types of mixed MQPs were measured by Raman spectroscopy and then decoded by ResNet to acquire the type information of analytes. The good classification performance of ResNet was verified by a t-distributed stochastic neighbor embedding (t-SNE) diagram. Compared with other machine learning models, these experiments showed that ResNet was obviously superior in terms of classification stability and training convergence to different datasets. This method simplified the decoding process and the classification accuracy reached 100%.