Targeted iPLS for the prediction of cheese-making traits from individual milk spectra

This study investigates the potential of the Targeted interval Partial Least Squares (iPLS) models for predicting cheese-making traits from individual milk mid-infrared spectra through high-resolution Fourier-transform infrared spectroscopy. Traditional full-spectrum PLS regression models demonstrated limited predictive ability for all traits, with R2 for validation (R2VAL) ranging from 0.00 to 0.15, largely due to spectral redundancy and multicollinearity. In contrast, the Targeted iPLS method markedly improved prediction accuracy. The three %CY traits achieved R2VAL ranging from 0.59 to 0.91, and %REC reached up to 0.87, with optimal prediction requiring selection of specific spectral regions related to protein, fat, and lactose, and exclusion of water regions. Tailored pre-treatment methods further enhanced trait-specific model performance. These findings underscore the importance of spectral region selection and pre-treatment customization in maximizing predictive performance. The Targeted iPLS framework offers a practical and effective tool for real-time monitoring and optimization in dairy processing.