From Mass Spectra to Biosignatures: Interpretable Machine Learning for Life Detection on Icy Ocean Worlds

NASA's Europa Clipper mission, and its MAss Spectrometer for Planetary EXploration (MASPEX), will arrive at Jupiter's moon, Europa, in 2030 and characterise its chemical conditions through mass spectrometry (MS). A critical challenge for probing habitability is abiotic mimicry, whereby geological processes produce organic compounds that mimic or obscure biogenic signatures. This work develops an interpretable binary machine learning (ML) framework for distinguishing between biotic and abiotic pyrolysis-gas chromatography-MS (py-GC-MS) samples, and identifying the mass-to-charge ratio ($m/z$) fragment ions driving predictions. 154 py-GC-MS datasets, comprising 140 biotic and 14 abiotic samples, were transformed into py-MS features analogous to MASPEX's intermediate data products. Four classifiers were evaluated, with XGBoost achieving the highest F1 (0.982) and ROC-AUC (0.988) scores. Predictions were interpreted using permutation importance, coefficient analysis, and SHapley Additive exPlanations (SHAP), which identified lighter aliphatic fragment ions ($m/z$ $60$, $79$, and $81$) as associated with biotic predictions at high abundances, while the heaviest ion analysed ($m/z$ $265$) contributed at near-zero abundance. SHAP also revealed non-linear relationships between interacting fragment ion abundances that were not captured by linear classifiers. These findings demonstrate the potential of this proof-of-concept framework for translating predictions into chemically meaningful insights, supporting interpretable ML as a tool for biosignature interpretation on icy ocean worlds.