Machine learning reveals 3.3‑billion‑year‑old chemical fingerprints of early life

New method detects ancient biological chemical signatures in South African rocks

Scientists have used a new machine-learning method to identify some of the oldest chemical traces of life on Earth in rocks from South Africa dated to about 3.3 billion years ago. Published this week in Proceedings of the National Academy of Sciences, the study demonstrates a way to tease biological patterns from highly degraded organic fragments, and the approach could help guide the search for life on other worlds.

The team concentrated carbon-rich material from ancient rock samples, fragmented the organic matter into thousands of tiny molecular pieces, and then applied a machine-learning algorithm to analyze the statistical distribution of those fragments. The method distinguishes suites of organic fragments that originated from once-living sources (microbes, plants or animals) from those produced by nonbiological processes with better than 90% accuracy.

"The remarkable finding is that we can tease out whispers of ancient life from highly degraded molecules," said Robert Hazen, a mineralogist and astrobiologist at the Carnegie Institution for Science and co-lead author of the study. He noted that while human analysts see hundreds or thousands of small molecular peaks, the machine-learning approach detects subtle patterns that reveal a biological origin.

The researchers report two key findings from South African samples: molecular evidence consistent with general microbial life in rocks about 3.3 billion years old, and molecular traces indicating oxygen-producing (oxygenic) photosynthesis in rocks dated to roughly 2.5 billion years ago. The latter finding is consistent with other lines of evidence that Earth’s atmosphere had become oxygenated by about 2.5 billion years ago, and it extends the organic-molecular record of oxygenic photosynthesis by some 800 million years compared with previous organic-molecule records.

All original biomolecules such as sugars and lipids are long degraded in these ancient rocks; what remains are fragments containing only a handful of carbon atoms. Nevertheless, the distribution patterns of these fragments differ noticeably between material derived from biological processes and material produced by nonliving chemistry, enabling the new approach to recover biosignatures far deeper in time than before.

Co-lead author Anirudh Prabhu of the Carnegie Institution said the work has three major implications: it roughly doubles the age at which organic-molecular evidence of life can be identified (from about 1.6 billion to ~3.3 billion years); it can discriminate different modes of life, such as photosynthetic versus non-photosynthetic organisms; and it demonstrates how machine learning can recover biosignatures even when original biomolecules are severely degraded.

The researchers have received a NASA grant to adapt and test the method for astrobiology. Hazen and Prabhu said they are excited about applying this technique to Martian samples returned to Earth and to samples that might be obtained by future rover missions. They also envision applications to the organic-rich plumes of Saturn’s moon Enceladus and the surfaces or materials from Titan or Europa.

Why it matters: This technique provides a complementary route to conventional fossil searches, opening a new window into Earth’s earliest ecosystems and offering a practical, machine-assisted tool for detecting ancient biological chemistry on other planetary bodies.

Reporting based on: Will Dunham, Reuters; study published in Proceedings of the National Academy of Sciences; quotes from study co-leads Robert Hazen and Anirudh Prabhu.