Analysis of SERS spectra of plasmonic nanoprobes for multiplexed biomarker detection using machine learning
Surface-enhanced Raman spectroscopy (SERS) has wide applications in chemical and biosensing as well as imaging. Raman spectra obtained from SERS exhibit characteristic narrow peaks that allow higher degrees of multiplexing than possible with fluorescence imaging. The nanorattle is a bimetallic nanoparticle which can be loaded with different dyes to produce SERS for multiplexed mRNA detection assays and in vivo imaging. But as multiplexing degree increases, so does spectral complexity, making analysis difficult. Machine learning has been applied for SERS-based chemical recognition and quantification. However, multiplexed, assays using SERS labels or imaging using SERS-labeled materials rarely utilize machine learning. Since the spectral shapes of each multiplexed label is known, analysis is easy when multiplexing <4 dyes given the computational tradeoff. Here we demonstrate and compare the use of spectral decomposition, support vector regression, and convolutional neural network (CNN) for "spectral unmixing"of SERS spectra obtained from a highly multiplexed mixture of 7 SERS-active nanorattles. Training data was simulated by combining individual nanorattle spectra by linear scaling and addition. We show that CNN performed the best in determining relative contributions of each distinct dye-loaded nanorattle.
