oa Untargeted Analysis of Geologic Samples with Machine Learning

“Geolipidomics” describes the recent application of analytical techniques and data processing pipelines of the biological science’s “-omics” fields to answer environmental and Earth science questions. Untargeted approaches allow for a complete view of compositional changes across samples, which can identify new biomarkers. As untargeted “-omics” approaches are transferred from biology and applied to sediment, soil, and rock samples, they must be adapted for the more chemically complex matrices. We identified three prospective machine learning feature extraction methods, DeepResolution2, AutoRes, and GNPS MSHub, which were designed for the untargeted analysis of uncomplicated biological samples and one more traditional, signal processing workflow (ADAP-GC). We compared targeted data from hydrous pyrolysis experiments of Woodford Shale heated at temperatures up to 343°C over 15–150 minutes to results from each approach. Only MsHub was able to recreate the targeted data results robustly and in particular the MPI-4 (methyl phenanthrene) thermal maturity index that is measured with traditional chemical analysis techniques. Therefore, only MsHub results were used for untargeted analysis. The untargeted analysis of these samples showed that a wide range of molecular features are affected by temperature rise temperature and are potential new targets for examining rapid heating in samples with unknown thermal histories.