FRACTAL: A Ratiometric-based Fingerprinting Framework for Robust Spectral Classification and Ratio-based Spectroscopic Marker Discovery
Year: 2026
Type: preprint
Abstract: Spectral classification is conventionally performed on band intensities, yet ratios between bands have long served as diagnostic indices in vibrational spectroscopy, from lipid-to-protein markers of membrane composition to amide ratios reporting protein secondary structure. Their use, however, has remained confined to a handful of expert-selected pairs, leaving most of the ratio information contained in a spectrum unexamined. Here I present FRACTAL (Full Ratio Combinatorial Analysis for Classification and Analytical Interpretation), which computes every pairwise log-intensity ratio among detected peaks to form a complete ratiometric fingerprint, grounded in compositional data analysis. Each component is a contrast between two assigned bands and is invariant to the multiplicative factors that scale spectral intensities. Applied to serum Raman spectra from a total of 103 donors (53 with COVID-19, 50 healthy controls), FRACTAL screened all 105 ratios among 15 reproducible peaks and identified 38 discriminant contrasts; a ten-ratio fingerprint then matched full-spectrum performance across six classifier architectures spanning linear, kernel-based and ensemble methods (mean internal AUC 0.89–0.93) and generalized to a withheld 20-donor cohort, where every classifier exceeded an AUC of 0.92 and all six converged on the same donor-level prediction accuracy at the median-AUC seed. The retained ratios map onto aromatic amino acid, acute-phase protein and protein–lipid contrasts independently reported in COVID-19. FRACTAL establishes exhaustive ratiometric fingerprinting as a compact, robust and interpretable representation for spectral classification and spectroscopic marker discovery.
Source: ChemRxiv
Author: Lamyaa M. Almehmadi
Institution: Massachusetts Institute of Technology
Language: English
Subfield: Biophysics
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7221724290