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

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Machine learning for vibrational spectra analysis is typically closed-set, limiting performance on unseen molecular classes.
• Existing methods struggle with out-of-distribution samples, hindering broad applicability.

Purpose of the Study:

• To develop a SERS-based foundation model for fatty acid-derived lipids that enables zero-shot spectral-to-structure prediction.
• To move beyond categorical assignments towards vector matching for enhanced molecular reconstruction.
• To demonstrate the model's generalizability to complex lipids and its utility in multiplex quantitation.

Main Methods:

• A SERS-based foundation model trained exclusively on primitive single-chain free fatty acids.
• Encoding structural attributes (carbon number, C=C bonds/position/geometry, chain number) into five orthogonal molecular vectors.
• Utilizing density functional theory for mechanistic interpretation of spectral features.
• Testing matrix-tolerant multiplex quantitation in artificial biofluids.

Main Results:

• The model achieved high accuracy for unseen free fatty acids (91.7%), fatty acid esters of hydroxy fatty acids (85.5%), and triglycerides (80.0%).
• Demonstrated zero-shot prediction capabilities by reconstructing molecules based on vector proximity.
• Recovered mixture ratios in artificial sweat and urine with 2-9% error, showcasing matrix tolerance.
• Provided mechanistic insights into spectral features via DFT calculations.

Conclusions:

• The developed domain-informed foundation model enables extrapolative, interpretable spectral-to-structure prediction from SERS.
• This approach successfully generalizes across adjacent chemical spaces, overcoming limitations of closed-set models.
• The model holds promise for untargeted lipid analysis and quantitative measurements in complex matrices.