Chemically transparent variable selection in NIR spectroscopy – PLS regression of paracetamol in pharmaceutical matrix aided by spectra simulation

Variable selection is widely applied in near-infrared (NIR) spectroscopy to improve model performance, but its chemical validity and robustness are rarely assessed. The complex nature of NIR spectra presents a fundamental challenge to conventional approaches to this problem. In this proof-of-concept study, we evaluated five variable selection methods, Variable Importance in Projection (VIP), Competitive Adaptive Reweighted Sampling (CARS), Moving Window PLS (MW-PLS), Monte Carlo Uninformative Variable Elimination (MCUVE), and Random Frog (RF), for quantifying paracetamol in a three-component pharmaceutical matrix using Partial Least Squares Regression (PLS-R). Two spectral pretreatments were compared: a minimal one, intended as a spectroscopically transparent reference in the form of a baseline offset and a Savitzky–Golay second derivative used as a performance-oriented but noise-exposing condition. Variable selections were interpreted via ab initio anharmonic vibrational simulations using Generalized Second-Order Vibrational Perturbation Theory (GVPT2), enabling direct assessment of chemical specificity. Only CARS and RF consistently concentrated on analyte-specific regions and yielded the most robust models under test-set validation; both reached R 2 VAL = 0.960 and RMSEP = 0.71. RF, in particular, selected strong but environment-sensitive polar vibrational features ( ν OH, ν NH) without overfitting; it achieved the lowest model complexity (2 latent variables) and highest alignment with known analyte bands. In contrast, VIP, MW-PLS, and MCUVE were more sensitive to pretreatment and/or amplified noise and produced less chemically specific selections. The results support a chemically grounded framework for comparing variable selection methods in quantitative NIR analysis of formulated products, while also illustrating that predictive accuracy alone is insufficient for their validation.