Protein content variation in milk can impact the quality and consistency of dairy products, necessitating access to in-line real time monitoring. Here, we present a chemometric approach for the qualitative and quantitative monitoring of β-lactoglobulin and α-lactalbumin, using mid-infrared spectroscopy (MIR). In this study, we employed Hotelling T2 and Q-residual for outlier detection, automated preprocessing using nippy, conducted wavenumber selection with genetic algorithms, and evaluated four chemometric models, including partial least squares, support vector regression (SVR), ridge, and logistic regression to accurately predict the concentrations of β-lactoglobulin and α-lactalbumin in milk. For the quantitative analysis of these two whey proteins, SVR performed the best to interpret protein concentration from 197 MIR spectra originating from 42 Cornell University samples of preserved pasteurized modified milk. The R² values obtained for β-lactoglobulin and α-lactalbumin using leave one out cross-validation (LOOCV) are 92.8% and 92.7%, respectively, which is the highest correlation reported to date. Our approach introduced a combination of preprocessing automation, genetic algorithm-based wavenumber selection, and used Optuna to optimize the framework for tuning hyperparameters of the chemometric models, resulting in the best chemometric analysis of MIR data to quantitate β-lactoglobulin and α-lactalbumin to date.

The preprocessing of infrared spectra can significantly improve predictive accuracy for protein, carbohydrate, lipid, or other nutrition components, yet optimal preprocessing selection is typically empirical, tedious, and dataset specific. This study introduces a Bayesian optimization-based framework designed for the automated selection of optimal spectral preprocessing pipelines within a chemometric modeling context. The framework was applied to mid-infrared spectra of milk to predict compositional parameters for fat, protein, lactose, and total solids. A total of 385 averaged spectra corresponding to 198 unique samples was split into a 70/30 ratio (training/test) using a group-aware Kennard-Stone algorithm, resulting in 269 averaged spectra (135 unique samples) for training and 116 spectra (58 unique samples) for testing. Six regression models: Elastic Net, Gradient Boosting Machines (GBM), Partial Least Squares (PLS), RidgeCV Regression, LassoLarsCV, and Support Vector Regression (SVR) were evaluated across three preprocessing conditions: (1) no preprocessing, (2) literature-derived custom preprocessing (e.g., MSC, SNV, and first and second derivatives), and (3) optimized preprocessing via the proposed Bayesian framework. Optimized preprocessing consistently outperformed other methods, with RidgeCV achieving the best performance for all components except lactose, where PLS slightly outperformed it. Improvements in predictive accuracy, particularly in terms of RMSEP were observed across all milk components. The best RMSEP results were achieved for protein (RMSEP = 0.054, $R^2=0.981$) and lactose (RMSEP = 0.026, $R^2=0.917$), followed by fat (RMSEP = 0.139, $R^2=0.926$) and total solids (RMSEP = 0.154, $R^2=0.960$). Literature-based pipelines demonstrated inconsistent effectiveness, highlighting the limitations of transferring preprocessing methods between datasets. The Bayesian optimization approach identified relatively simple yet highly effective preprocessing pipelines, typically involving few steps. By eliminating manual trial and error, this data-driven strategy offers a robust and generalizable solution that streamlines spectral modeling in dairy analysis and can be readily applied to other types of spectroscopic data across various domains.