A Machine-Learning Approach to Biosignature Exploration on Early Earth and Mars Using Sulfur Isotope and Trace Element Data in Pyrite

We propose a novel approach to identify the origin of pyrite grains and distinguish biologically influenced sedimentary pyrite using combined <i>in situ</i> sulfur isotope (δ³⁴S) and trace element (TE) analyses. To classify and predict the origin of individual pyrite grains, we applied multiple machine-learning algorithms to coupled δ³⁴S and TE data from pyrite grains formed from diverse sedimentary, hydrothermal, and metasomatic processes across geologic time. Our unsupervised classification algorithm, K-means++ cluster analysis, yielded six classes based on the formation environment of the pyrite: sedimentary, low temperature hydrothermal, medium temperature, polymetallic hydrothermal, high temperature, and large euhedral. We tested three supervised models (random forest [RF], Naïve Bayes, k-nearest neighbors), and RF outperformed the others in predicting pyrite formation type, achieving a precision (area under the ROC curve) of 0.979 ± 0.005 and an overall average class accuracy of 0.878 ± 0.005. Moreover, we found that coupling TE and δ³⁴S data significantly improved the performance of the RF model compared with using either TE or δ³⁴S data alone. Our data provide a novel framework for exploring sedimentary rocks that have undergone multiple hydrothermal, magmatic, and metamorphic alterations. Most significant, however, is the demonstrated potential for distinguishing between biogenic and abiotic pyrite in samples from early Earth. This approach could also be applied to the search for potential biosignatures in samples returned from Mars.