SSNet: A Spectral Unmixing Framework for Enhancing Qualitative Sensitivity of SERS to Trace Targets in Complex Mixtures

Surface-enhanced Raman spectroscopy (SERS) is a powerful tool for spectrum-structure correlation across various fields. However, the qualitative and quantitative analysis of SERS to trace targets is often compromised, due to the co-adsorption and competitive adsorption from non-targets in complex systems. To unmix and identify the SERS signal of a target within a mixture SERS spectrum, we develop SSNet, an intelligent and self-supervised algorithm. Taking the trace detection of gelsemium phytotoxin in various food samples as an example, SSNet performs with high fidelity across the core qualitative and quantitative benchmarks: Raman peak intensity, peak position, and relative intensity. Without any prior knowledge of the matrix, SSNet achieved an expert level qualitative sensitivity. With the knowledge of similar matrices, the sensitivity was an order of magnitude higher than that of an expert, even when the SERS signal of the target is invisible to the naked eye. The ability of unmixing SERS signals of multiple targets is further demonstrated using the three structurally similar gelsemium phytotoxins. The exceptional performance and generalizability of SSNet enhance the on-site, in-situ, in-vivo, and operando applications of Raman/SERS. https://pubs.acs.org/doi/10.1021/jacs.5c16529