AI Uncovers Oldest-Ever Molecular Evidence of Photosynthesis

AI Uncovers Oldest-Ever Molecular Evidence of Photosynthesis

By Lee Billings edited by Claire Cameron

While much of the history of life on Earth is written, the opening chapters are murky at best. On our ever-changing world, the older a rock is, the more it has changed, obscuring or even erasing evidence of ancient life. Beyond a hazy boundary of circa two billion years, in fact, this interference is so total that no pristine, unaltered Earth rocks are known to exist, making any potential sign of biology as clear as mud.

At least until now. In a study published on November 17 in the Proceedings of the National Academy of Sciences, a group of researchers say they’ve leveraged artificial intelligence to follow life’s trail further back in time than ever before, using machine learning to distinguish the echoes of biology from mere abiotic organic molecules in rocks as old as 3.3 billion years.

The results could more than double how far back in time scientists can convincingly claim to discern molecular signs of life in ancient rocks, the study authors say, citing previous record-setting measurements involving 1.6-billion-year-old rocks.

On supporting science journalism

If you’re enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

The study also flags signs of photosynthesis in 2.5-billion-year-old rocks—some 800 million years earlier than any other confirmed biomolecular evidence. The authors suggest that in the not-too-distant future similar techniques may be used to seek signs of alien life on Mars or the icy ocean moons of the outer solar system.

And such astrobiological applications wouldn’t necessarily demand the extremely costly task of retrieving material from Mars or any other extraterrestrial locale for in-depth study in labs back on Earth. “Our approach could run on board a rover—no need to send samples home,” says the study’s first author, Michael Wong, an astrobiologist at the Carnegie Institution for Science.

According to Karen Lloyd, a biogeochemist at the University of Southern California uninvolved with the study, the technique holds promise as an “agnostic” way of looking for life, independent of Earth-bound assumptions.

“This allows for the possible extrapolation from an extremely varied and diverse dataset of biomolecules in known living matter, extending to matter that may or may not have come from living things,” Lloyd says. “This is really helpful in the search for life on rocks that come from ancient Earth—as well as rocks that come from extraterrestrial bodies.”

Rocks containing familiar fossils—dinosaurs, ferns, fish, trilobites and so on—may seem creakingly ancient, but in fact represent less than the latest 10 percent of Earth’s 4.5-billion-year history. Put another way, for each of the circa 500 million years that make up the ongoing Phanerozoic (Greek for “visible life”) Eon, there exists nearly a decade of underlying planetary time in which early life flourished almost imperceptibly, scarcely registering in the fossil record beyond trace molecules such as lipids and amino acids.

The trouble, says lead author and Carnegie geologist Robert Hazen, is that those molecules degrade and disappear over time. “Our method looks for patterns instead, like facial recognition for molecular fragments,” he explains. “Think of the burnt Herculaneum scrolls that AI helped ‘read.’ You and I just see dots and squiggles, but AI can reconstruct letters and words.”

The team began by gathering more than 400 samples—some modern, some ancient, some from known abiotic sources like meteorites, others filled with fossils or living microbes, and several containing organic molecules but no obvious indicators of life. They fed them into an instrument called a pyrolysis gas chromatograph mass spectrometer (Py-GC-MS), which vaporized each sample to release and then categorize their constituent molecular fragments by mass and other properties. This yielded a rich “chemical landscape” for each sample, filled with tens of thousands to hundreds of thousands of peaks denoting different possible compounds and ripe for the AI’s pattern-spotting scrutiny.

After training the AI on about 75 percent of the sample data, the researchers unleashed it on the remaining 25 percent. The system correctly distinguished between biotic and abiotic samples for more than 90 percent of that material, but its certainty dwindled as a rock’s age and level of degradation increased; for samples older than 2.5 billion years, the AI flagged less than half as having a biotic origin, and with lower overall confidence.

Even so, it was very old samples from South Africa that led to the team’s most spectacular conclusions—signs of biogenic molecules in 3.3-billion-year-old specimens from a formation called the Josefsdal Chert, and evidence of ancient oxygen-producing photosynthesis in 2.5-billion-year-old rocks from the Gamohaan Formation. Preexisting geochemical evidence meant neither result was a surprise, but being backed up by biomolecular data is a true breakthrough. “The key is that our validation set included truly unknown samples—some debated for decades,” says paper co-author Anirudh Prabhu, who studies geoinformatics at Carnegie. “And the model made independent predictions that sometimes confirmed existing suspicions.”

The most surprising finds came from the AI outsmarting its human tenders. The system flagged a dead seashell as photosynthetic—an error, it seemed, until the researchers realized the system had picked up algae growing on the shell. A similar photosynthesis “false alarm” arose for a wasp’s nest, which the AI correctly linked to the chewed-up wood from which the nest was made. “The model was right—just for the wrong reason,” Prabhu says.

Linda Kah, a geochemist at the University of Tennessee in Knoxville who was not part of the study, calls it a “magnificent effort.” Its “big data” approach offers a roadmap for scientists seeking even more ancient biosignatures, she says—and poses questions that demand further investigation. For example: Does the AI’s diminishing returns for the most ancient and degraded samples mean the technique is approaching a fundamental limit of what can be recognized as biotic? Or might older samples instead simply contain more abiotic material because life had yet to fully infiltrate the available environments on the early Earth?

Answers could come soon. The team is already planning to test its AI on a broader, more diverse set of samples, including ones from even deeper in Earth’s history and from a wider range of extraterrestrial sources. And some interplanetary robotic explorers—NASA’s Curiosity rover among them—already carry Py-GC-MS instruments onboard, potentially offering chances for otherworldly ground-truthing of the technique.

“Studies such as this one take us one step closer in learning about the origin and evolution of life on Earth,” says Amy J. Williams, a geobiologist at the University of Florida who was also not part of the work. “They prepare us to address that most fundamental question of whether we are alone in the universe.”