Advanced spectral modeling for bacterial strains: A MARS-PLS2 approach with Lasso regularization and baseline...

The natural characteristics of the infrared spectroscopic data are that it tends to distort the baseline, there is high-dimensionality and non-linear correlation that hinder reliable prediction of biochemical properties. To overcome these obstacles, this study introduces an integrated MARS-PLS2-Lasso framework that incorporates the effective baseline correction, non-linear regression, latent variable extraction, and sparse variable selection to promote the chemometric modeling accuracy and interpretability. Out of four baseline correction methods, viz. Asymmetric Least Squares (ALS), AirPLS, Polynomial fitting, and Wavelet baseline correction, the Wavelet method (sym8, Level 5) was found to be the most successful, in that it was able to represent local spectral variation with low-frequency noise. This technique achieved high predictive accuracy with RMSE = 0.2846-0.6857, MAE = 0.2371-0.5445 and MSE = 0.0810-0.4705 specifying both high model fit and minimal residual error across bacterial spectra. The Wavelet-corrected spectra revealed six key functional regions that contributed most significantly to bacterial differentiation: 720cm^-1 to 750cm^-1 (C-Cl stretching, CH bending), 1000cm^-1 to 1300cm^-1 (C-O stretching, esters, carboxylic acids), 1500cm^-1 to 1650cm^-1 (CC stretching), 1687cm^-1 to 1793cm^-1 (CO stretching, conjugated carbonyls), 2771cm^-1 to 3143cm^-1 (CH stretching, alkanes, alkenes), 3290cm^-1 to 3595cm^-1 (O-H and NH stretching ). Vibrational domains of interest are biochemical components of lipids, proteins, amides and polysaccharides that determine the structural integrity and metabolic activity of bacteria. The proposed MARS-PLS2-Lasso model leverages Multivariate Adaptive Regression Splines (MARS) to capture nonlinear relationships through adaptive basis functions, while Partial Least Squares (PLS2) extracts latent components that maximize covariance between spectral predictors and multiple bacterial responses. Lasso regularization adds sparsity to the model and reduces the complexity of the model, as well as penalizes less interesting basis functions, which overfit the model. Such a combination is used to provide a reasonable approximation of the parameter even in high-dimensional spectral data. In general, MARS-PLS2-Lasso provides a sound, interpretable, and chemically consistent way of high dimensional infrared spectral modeling. It is highly predictive, less noisy and has a more adequate manner of interpreting spectral-biochemical interactions, and thus, a bright way of bacteria modeling, spectral diagnostics and further use in bio-analytical spectroscopy.