Automated identification of class-separating principal component subspaces in biomedical Raman and Fourier Transform...

This study presents PCA AutoExplorer, an open-source tool for automated identification of three-component Principal Component subspaces (hereafter referred to as "PCA triplets") that maximize class separation in clinical vibrational spectroscopy. The algorithm exhaustively evaluates all Principal Component triplets, combining Mahalanobis distance (unsupervised) and Linear Discriminant Analysis accuracy (supervised) to rank subspaces and compare preprocessing modes and spectral ranges. It introduces a marker strength plot - summing absolute loadings from Principal Components in top triplets - and integrated Principal Component loading heatmaps to globally prioritize diagnostic bands. Class separation is additionally visualized with 2D/3D t-distributed stochastic neighbor embedding. The method was validated on spectra acquired from dried blood serum samples of acute coronary syndrome patients and controls: Fourier Transform Infrared Spectroscopy (n=364) and Raman Spectroscopy (n=254), across 800-1800 and 2700-3500 cm^-1, in intensity, first derivative, and second derivative modes. For 50 Principal Components, 19 600 triplets were assessed, with significance determined via permutation testing. Raman second derivative most often yielded the highest separation, while Fourier Transform Infrared Spectroscopy favored intensity or first derivative depending on spectral range. Top subspaces frequently involved higher-order Principal Components (e.g., PC1-PC4-PC26), achieving Mahalanobis distances up to ∼2.03 and Linear Discriminant Analysis accuracies of 96%-100% (p<0.001). Marker strength highlighted diagnostic bands such as 1522 cm^-1 (Raman Spectroscopy, first derivative), 1649 cm^-1 (Fourier Transform Infrared Spectroscopy, second derivative), 2951 cm^-1 (Raman Spectroscopy, second derivative), and 3479 cm^-1 (Fourier Transform Infrared Spectroscopy, second derivative). T-distributed stochastic neighbor embedding projections confirmed strong separation and revealed within-group heterogeneity. PCA AutoExplorer provides a reproducible, statistically rigorous framework for identifying diagnostically relevant Principal Component subspaces and prioritizing spectral biomarkers, enhancing the reliability of biomarker discovery in clinical vibrational spectroscopy and adaptable to other omics-related spectral analyses.