Machine Learning-Based SERS Chemical Space for Two-Way Prediction of Structures and Spectra of Untrained Molecules

Abstract

Identifying unknown molecules beyond existing databases remains challenging in surface-enhanced Raman scattering (SERS) spectroscopy. Conventional SERS analysis relies on matching experimental and cataloged spectra, limiting identification to known molecules in databases. With a vast chemical space of >10⁶⁰ molecules, it is impractical to obtain the spectra of every molecule and rely solely on <i>in silico</i> techniques for spectral predictions. Here, we showcase an ML-based SERS chemical space that leverages key spectra-structure correlations to achieve two-way spectra-to-structure and structure-to-spectra predictions for untrained molecules with a >90% average accuracy. Using a SERS chemical space comprising 38 linear molecules from four classes (alcohols, aldehydes, amines, and carboxylic acids), our experimental and <i>in silico</i> studies reveal underlying spectral features that enable the prediction of untrained molecules represented by two molecular descriptors (functional group and carbon chain length). For forward spectra-to-structure predictions, we devise a two-step "classification and regression" ML framework to sequentially predict the functional group and carbon chain length of untrained molecules with 100% accuracy and ≤1 carbon difference, respectively. In addition, using an eXtreme Gradient Boosting (XGBoost) regressor trained on the two molecular descriptors, we attain inverse structure-to-spectra prediction with a high average cosine similarity of 90.4% between the predicted and experimental spectra. Our ML-based SERS chemical space represents a shift in molecular identification from traditional spectral matching to predictive modeling of spectra-structure relationships. These insights could motivate the expansion of SERS chemical spaces and realize demands for present and future SERS technologiesfor accurate unknown identification across diverse fields.