Building a Foundation SERS Model for Lipids through Fatty Acid Pretraining for Annotation across Chemical Spaces.

Machine learning analysis of vibrational spectra is often closed-set, performing well for known molecular classes but degrading sharply when test molecules fall outside the training library. Herein, we introduce a SERS-based domain-informed foundation model for fatty acid-derived lipids that replaces categorical assignments with vector matching. The model is trained exclusively on primitive single-chain free fatty acids, with each structural attribute encoded by five orthogonal molecular vectors in a multidimensional space: (1) carbon number, (2) number of C═C bonds, (3) C═C position, (4) C═C geometry, and (5) number of carbon chains. This modular ensemble enables zero-shot prediction of all five vectors in parallel from previously unseen spectra, allowing untargeted molecular reconstruction by proximity in this space rather than by class labels. Despite being trained only on free fatty acids, the model generalizes to complex, multichain lipids absent from the training set, achieving accuracies of 91.7% for withheld free fatty acids, 85.5% for fatty acid esters of hydroxy fatty acids, and 80.0% for triglycerides. To enhance interpretability, we utilize density functional theory to provide a mechanistic basis for the spectral features underlying each prediction. We further demonstrate matrix-tolerant multiplex quantitation in artificial sweat and urine, recovering mixture ratios with 2-9% error across broad composition ranges. Collectively, this strategy enables extrapolative, interpretable spectral-to-structure prediction from SERS across adjacent chemical spaces.