Raman spectroscopy coupled with PLS-CNN error-min fusion strategy for conformity discrimination and amino acid...

Yeast extract (YE) is a widely used bioproduct in the food and biofermentation industries, whose quality is determined by amino acid composition and conformity to industrial specifications. Conventional assessment relies on amino acid analyzers and microbial growth assays, which are time-consuming and unsuitable for in situ monitoring. Raman spectroscopy offers a non-destructive alternative, but its application to complex biomolecular matrices such as YE is hindered by strong fluorescence backgrounds, spectral overlap, and heterogeneous linear/nonlinear relationships across multiple analytes. In this work, 1064 nm excitation was selected over 830 nm to suppress fluorescence interference and yield clearer spectral features. For conformity classification, a one-dimensional convolutional neural network (CNN) with channel attention achieved 92% accuracy, outperforming conventional classifiers. For amino acid quantification, a variable-wise PLS-CNN error-min fusion strategy was developed, in which the CNN contribution weight for each amino acid is independently optimized by minimizing validation error, enabling adaptive linear-nonlinear balancing across heterogeneous targets. The unified fusion model achieved coefficients of determination of 0.989 for hydrolyzed amino acids and 0.956 for free amino acids across mixed process types, outperforming standalone PLS, CNN, and other baselines. The framework was validated across standard, low-hydrolysis, high-hydrolysis, and high-nucleotide YE variants, demonstrating feasibility for rapid spectroscopic analysis of complex biomolecular matrices pending broader validation.