Could the chemical fingerprints of life really persist for billions of years, hidden in rocks reshaped by eons of heat and pressure? A new study says yes and the story of Earth’s biosphere begins far earlier than scientists once believed.
Using a combination of high-resolution analytical chemistry and artificial intelligence, researchers from the Carnegie Institution for Science have for the first time detected biomolecular traces in rocks up to 3.3 billion years old. The work extends the molecular record of oxygen-producing photosynthesis by more than 800 million years, tracing photosynthetic signals in 2.5- billion-year-old formations where none had previously been found. Until now, reliable chemical evidence for these processes had only been recovered from rocks younger than 1.7 billion years.
They analyzed 406 samples ranging from modern plants, animals, fungi, ancient fossils, sedimentary rocks ranging in age from 3.8 billion to 10 million years, and meteorites. The authors decomposed organic and inorganic material into molecular fragments by using pyrolysis gas chromatography–mass spectrometry chemical “echoes” of long‑vanished biomolecules. Three‑quarters of the dataset were used to train a random forest classifier a supervised machine‑learning model to discriminate between biological and non‑biological sources and to identify photosynthetic signatures. During testing, the AI achieved up to 98 percent accuracy for distinguishing living from non‑living sources and 93 percent accuracy in the detection of photosynthesis.
This approach sidesteps the need for intact molecules such as lipids or DNA, which are rarely preserved beyond 1.7 billion years. Instead, it focuses on the distribution patterns of hundreds to thousands of degraded fragments, a method that-as lead author Robert Hazen put it-allows scientists to “tease out whispers of ancient life from highly degraded molecules.” The AI’s sensitivity revealed biosignatures in rocks that had undergone intense metamorphism, where traditional geochemical proxies often fail.
The implications for early Earth science are profound. Geochemical and isotopic studies have long debated the timing of oxygenic photosynthesis, with some molecular clocks placing its origin in the Palaeoarchaean, 3.5-3.2 Ga, and others in the Mesoarchaean, 3.2-2.8 Ga. The new chemical evidence aligns with interpretations that oxygenic phototrophy may have been established well before the Great Oxidation Event (∼2.4 Ga), reshaping models of atmospheric evolution and the co‑development of microbial metabolisms. It also complements biosignatures such as stromatolites and microbially induced sedimentary structures which are often complicated by abiotic look‑alikes.
The method represents a powerful tool for life detection beyond Earth from an astrobiology perspective, given it will not depend on pristine biomolecules. It could be applied to the heavily altered extraterrestrial samples from Mars, for instance, where radiation, oxidation, and mineral adsorption degrade organic matter. All ongoing missions to Mars, including Perseverance, collect regolith and rock cores for eventual return to Earth. Applying AI‑enabled fragment analysis to such samples could reveal subtle biosignatures otherwise invisible, much as it did with ancient terrestrial rocks.
The training set for the study even included meteorites, which allowed the model to distinguish abiotic organic chemistry typical of carbonaceous chondrites from biotic patterns. This is an important distinction in planetary science, where extraterrestrial organics may produce biological isotopic fractionations. As pointed out in research on geochemical biosignatures, early Earth’s record is complicated by high fluxes of abiotic organic matter both from endogenous and extraterrestrial sources. AI pattern recognition provides a complementary line of evidence that can reinforce assessments of biogenicity.
The researchers are now trying to extend the dataset to over 1,000 well-documented specimens from various geological contexts, including Archean terrains in Australia, South Africa, Greenland, and Canada. Greater diversity will improve the model’s ability to resolve fine distinctions-such as separating photosynthetic from non-photosynthetic plant tissues, where current accuracy dips below 80 percent.
For Hazen and colleagues, the work is not just a breakthrough in paleobiology, but a template for astrobiology. My dream is that this approach becomes a new standard in both fields, he said. By reading the molecular ghosts locked in stone, scientists may soon extend the timeline of life’s chemical record even further and perhaps one day, hear the whispers of biology from other worlds.
