Interpretable Machine Learning Biosignature Detection From Ocean Worlds Analogue CO2 Isotopologue Data

Abstract

Future missions to icy ocean worlds (OW) such as Europa and Enceladus will evaluate the habitability and potential for biosignatures on these worlds. These missions will benefit from autonomous science and machine learning (ML) methods to process high volumes of data and prioritize signals of interest for the first available downlink. Mass spectrometers (MS) are suitable instruments for implementing science autonomy due to their rich spectral data products and potential for biosignature detection. Light stable isotopes are strong candidates for biosignatures due to the large fractionations promoted by biological activity. However, complex abiotic geochemistry may obscure or mimic biogenic isotope fractionations. ML may accurately disentangle biosignatures from abiotic mimicry in MS data; however, ML model predictions can be inscrutable to human interpretation, compromising trust in scientifically significant detections. We develop and test a new biosignature detection ML model using a novel, laboratory-generated, CO₂ isotopologue data set of analogue OW samples. These data include various potential OW seawater chemistries and biotic mimicry. Our ML approach includes feature (variable) construction, providing mathematical and geochemical context for biosignatures, and a feature selection method called Nearest-neighbors Projected Distance Regression (NPDR) that identifies important predictors. Our Random Forest biosignature model predicts the presence of biosignatures with 87.3% mean accuracy regardless of the sample brine chemistry. We add network visualization of main effects and statistical interactions for interpretation of model prediction mechanisms. We use single-sample (local) variable importance scores to diagnose false predictions for individual samples, which is crucial for trust in astrobiology ML biosignature models.