Quantifying Components in a Model Vaccine with Machine-Learning-Augmented Raman Spectroscopy

Vaccination is a highly efficient strategy in controlling infections. Aluminum-containing adjuvants have long been used to enhance immunogenicity, but quantification of adsorbed antigens remains analytically challenging. This study uses Raman spectroscopy, a powerful, nondestructive technique, augmented by machine learning to characterize model vaccines comprising Bovine Serum Albumin adsorbed to aluminum hydroxide. Spectral fingerprints of the pure components, their contributions to the vaccine mixtures, and the resulting concentrations were estimated using autoencoder and benchmarked against Multivariate Curve Resolution (MCR), a state-of-the-art linear deconvolution method. We implement a custom autoencoder (AE) architecture featuring a scale-insensitive reconstruction loss and a concentration-anchored latent space, which we compare against Multivariate Curve Resolution (MCR) to evaluate linear versus nonlinear decomposition efficacy. By integrating synthetic spectra based on the Contextual Out-of-Distribution Integration (CODI) method, we achieved a concentration prediction accuracy that aligns with standards for biological reference methods. While MCR proved more robust for recovery of component spectra, the AE demonstrated superior performance in estimating concentrations. These results highlight the potential for characterization of adsorbed vaccines and offer a pathway for improved quality control in biopharmaceutical formulations.