Interpretable Machine Learning for Molecular Biosignatures: a Novel Single-Sample Feature Importance Method That Is Sensitive To Statistical Interactions

Isotope ratio mass spectrometry (IRMS) of volatiles (e.g., CO2) promises to be a powerful tool for potential biosignature detection for future missions to ocean worlds (OW) such as Europa and Enceladus. Machine learning (ML) methods for IRMS data could enable science autonomy by onboard prediction of seawater chemistry and biosignature presence. However, ML models are likely to be complex and involve statistical interactions between features (variables), which can make predictions seem opaque and enigmatic. For ML predictions as significant as extraterrestrial biosignatures, we must place extraordinary confidence in models. It is therefore essential that these models make interpretable predictions (i.e., human-understandable) and include false-prediction diagnostics. We achieve high accuracy and interpretability in ML biosignature and seawater chemistry models for OW through a nearest-neighbors feature selection tool that detects statistical interactions between predictors, constructs interaction networks for visualization of selected features working together to make a prediction, and reports single-sample feature importance scores for false-detection diagnostics. Here we develop a novel single-sample nearest-neighbors projected distance regression(ssNPDR) feature selection method that improves upon existing single-sample algorithms through the inclusion of statistical interactions while providing false-prediction diagnostics for ML models.