189-2 Pyrolysis-GC-MS and supervised machine learning suggest that 2.52 Ga organic matter may hold fragmentary biomolecular evidence of photoautotrophy

Session: Paleontology, Paleoecology/Taphonomy (Posters)

Presenting Author:

Authors:

Abstract:

The sedimentary rock record holds diverse clues about the history of life on Earth. Some have remained well-preserved for millions, if not billions, of years, but others have degraded with time. Diagnostic biomolecules, which can indicate the taxonomic affinities and metabolic processes of organisms that formed organic sediments, are particularly short-lived, and have not been found to persist in the rock record beyond ~1.65 billion years. However, it is possible that diagenetically altered biomolecular fragments retain signals of phylogenetic affiliation or physiological processes even after identifiable biomolecules have degraded. 

Here, we present evidence of such signal retention in the form of pyrolysis-GC-MS data analysed via supervised machine learning methods which supports the conclusion that 2.52 Ga organic matter preserved within fenestrate microbialites may contain biomolecular fragments indicative of photoautotrophy. The results are of special interest both because of the age of the samples and because, while the studied microbialites are morphologically consistent with a photoautotrophic microbial community, they are also morphologically consistent with a chemoautotrophic microbial community, and they contain isotope evidence for microbial sulfur cycling. 

The conclusion is based on analyses of pyrolysis-GC-MS data collected from 406 fossil, modern biogenic, meteoritic, and synthetic organic samples. First, 272 well-characterized modern and fossil biological samples, plus meteoritic and synthetic samples were divided into 9 categories, which were then randomly separated into training and test sets. Pairwise testing, using the random forest method, of all possible combinations of the 9 categories was performed to determine how well supervised machine learning could discriminate between them. The tests yielded 90% correct assignments in 25 of 36 pairings. 

Building on this success, additional comparisons were made, including a comparison of 107 non-photosynthetic samples to 151 samples of modern and fossil photosynthetic organisms. In this comparison, 92.3% of samples were correctly classified. When the model used for this comparison was applied to 131 Precambrian microbial samples of unknown affinities, the 2.52 Ga fenestrate microbialite samples had class assignment probabilities suggestive of oxygenic photosynthesis. Further investigation is necessary to fully explore the utility of this approach, but these results suggest that the analysis of biogeochemical data via supervised machine learning may allow researchers to access clues in the sedimentary rock record which were previously thought lost to time. 

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-6280

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.