Teasing out biochemical information from ancient organic-rich sediments, notably the timing of the emergence of photosynthesis relative to the inferred oxygenation of Earth’s atmosphere, remains a challenging opportunity. To tackle this problem, scientists analyzed 406 diverse ancient and modern samples and used supervised machine learning to discriminate samples of biogenic vs. abiogenic origin, as well as photosynthetic vs. nonphotosynthetic physiology. They found chemical evidence for biogenic molecular assemblages in Paleoarchean rocks (3.51 billion years ago) and for photosynthetic life in Neoarchean rocks (2.52 billion years ago).
An AI impression of the Early Earth. Image credit: Gemini AI.
Earth’s earliest life left behind little in the way of molecular traces.
The few fragile remnants such as ancient cells and microbial mats were buried, crushed, heated, and fractured within Earth’s restless crust before being thrust back to the surface.
These transformations all but obliterated biosignatures holding vital clues to the origins and early evolution of life.
Paleobiologists who search for signs of Earth’s most ancient life have long relied mainly on fossil organisms, including microscopic fossils of single cells and filaments, and the mineralized remains of cellular structures such as microbial mats and mound-like stromatolites, which provide convincing evidence of life as far back as 3.5 billion years ago. However, such remains are few and far between.
A second line of evidence relies on the preservation of diagnostic biomolecules in ancient rocks.
Life’s hardiest organic molecules — those derived from cell membranes or some metabolic processes — have been found in sediments as old as 1.7 billion years, while much older carbon-rich rocks preserve isotopic signatures that hint at a vibrant biosphere 3.5 billion years ago.
However, most ancient rocks preserve neither fossil cells nor any surviving biomolecules.
The vast majority of ancient carbon-bearing sediments have been heated and altered in ways that break every diagnostic biomolecule into countless small fragments.
Those fragments have proven too small and too generic to provide any clues about ancient life — until now.
“Ancient rocks are full of interesting puzzles that tell us the story of life on Earth, but a few of the pieces are always missing,” said Michigan State University researcher Katie Maloney, co-author of the study.
“Pairing chemical analysis and machine learning has revealed biological clues about ancient life that were previously invisible.”
Organic matter extracted from samples of 2.5-billion-year-old rock containing fossilized microorganisms like the one in this photomicrograph still contains biomolecular fragments that may have been produced via photosynthesis. Image credit: Andrew D. Czaja.
The researchers used high-resolution chemical analysis to break down organic and inorganic materials into molecular fragments, then trained an AI system to recognize the chemical ‘fingerprints’ left behind by life.
They examined a total of 406 fossil, modern biological, meteoritic, and synthetic samples.
The AI model distinguished biological from non-biological materials with over 90% accuracy and detected the earliest biomolecular evidence for:
(i) the photosynthetic origins of organic molecules in the 2.52-billion-year-old Gamohaan Formation, Campbellrand Group, South Africa, and the 2.30-billion-year-old Gowganda Group, Ontario, Canada;
(ii) the biogenicity of organic molecules preserved in the 3.51-billion-year-old Singhbhum Craton, India; the 3.33-billion-year-old Josefsdal Chert of the Barberton Greenstone Belt, South Africa; and the 2.66-billion-year-old Jerrinah Formation, Fortescue Group, Pilbara Craton, Australia;
(iii) and the apparently non-photosynthetic origin of organic species in the 3.5-billion-year-old Theespruit Formation, Barberton Greenstone Belt, South Africa, and the 3.48-billion-year-old Dresser Formation, Pilbara Craton, Australia.
“Ancient life leaves more than fossils; it leaves chemical echoes,” said senior author Dr. Robert Hazen, a researcher at the Carnegie Institution for Science.
“Using machine learning, we can now reliably interpret these echoes for the first time.”
“This innovative technique helps us to read the deep time fossil record in a new way,” Dr. Maloney added.
“This could help guide the search for life on other planets.”
“Understanding when photosynthesis emerged helps explain how Earth’s atmosphere became oxygen-rich, a key milestone that allowed complex life, including humans, to evolve,” said first author Dr. Michael Wong, also from the Carnegie Institution for Science.
“This represents an inspiring example of how modern technology can shine a light on the planet’s most ancient stories and could reshape how we search for ancient life on Earth and other worlds.”
“In future, we plan to test materials like anoxygenic photosynthetic bacteria — possible analogs for extraterrestrial organisms. This is a powerful new tool for astrobiology.”
“These samples and the spectral signatures they produce have been studied for decades, but AI offers a powerful new lens that allows us to extract critical information and better understand their nature,” added Carnegie Institution for Science’s Dr. Anirudh Prabhu, co-author of the study.
“Even when degradation makes it difficult to spot signs of life, our machine learning models can still detect the subtle traces left behind by ancient biological processes.”
“What’s exciting is that this approach doesn’t rely on finding recognizable fossils or intact biomolecules.”
“AI didn’t just help us analyze data faster, it allowed us to make sense of messy, degraded chemical data.”
“It opens the door to exploring ancient and alien environments with a fresh lens, guided by patterns we might not even know to look for ourselves.”
The team’s results appear this week in the Proceedings of the National Academy of Sciences.
_____
Michael L. Wong et al. 2025. Organic geochemical evidence for life in Archean rocks identified by pyrolysis-GC-MS and supervised machine learning. PNAS 122 (47): e2514534122; doi: 10.1073/pnas.2514534122
