Mineral biosignature identification from Raman spectroscopy using machine learning.

Biosignature detection remains a key challenge in astrobiology, yet robust mineral biosignatures remain limited. Raman spectroscopy is increasingly applied in planetary exploration, but its high-dimensional spectral information has not yet been fully exploited for biosignature discrimination using data-driven approaches. Here, we integrate Raman spectroscopy with interpretable machine learning to distinguish biotic from abiotic apatite, a ubiquitous phosphate mineral in terrestrial and extraterrestrial environments. We compile 331 apatite Raman spectra from abiotic and biotic sources and extract 21 band-resolved spectral features. Principal component analysis reveals systematic separation between abiotic and biotic endmembers. A random forest classifier achieves 96.8% accuracy on an independent test set. Robustness is confirmed by multiple validation schemes, including leave-one-source-out cross-validation across 60 independent data sources, indicating that model performance generalizes beyond source- or instrument-specific artifacts. Feature importance identifies two dominant controls: phosphate-band broadening as a structural indicator of disorder and the carbonate-band intensity as a chemical signature of substitution. Density-functional calculations reproduce these features in simulated spectra and indicate that carbonate substitution doubles phosphate-tetrahedral distortion and increases formation energies by two orders of magnitude. Mechanically, higher carbonate contents during biomineral apatite formation reduce crystallinity and broaden Raman bands. We propose that the trained machine-learning model and a two-feature decision map enable the rapid probabilistic discrimination of unknown apatite samples. Our Raman-based machine-learning framework establishes a broadly applicable and mission-relevant strategy for deep-time archives and future planetary missions.