Machine learning joins the hunt for extraterrestrial life

A machine-learning framework trained on just 18 samples — eight carbonaceous meteorites and 10 terrestrial geological specimens — can separate biological from non-biological organic chemistry well enough to analyse material from future Mars and Enceladus sample-return missions. That is the measured claim behind LifeTracer, published in PNAS Nexus (4(11) pgaf334) by researchers at NASA's Goddard Space Flight Center and Georgia Institute of Technology, and it signals a wider shift: astrobiology is increasingly betting on statistics rather than predefined biomarkers.

The stakes are concrete. Astronomers have confirmed 6366 exoplanets, with more than 8100 further candidates awaiting verification, and missions from NASA's stalled Mars Sample Return campaign to JAXA's Martian Moons eXploration (MMX) probe — launched on 20 October 2026 — promise fresh material to interrogate.

How does LifeTracer work?

José Aponte, an astrochemist at Goddard, and colleagues analysed their samples using 2D gas chromatography coupled with high-resolution time-of-flight mass spectrometry. The system learned molecular patterns in meteorites (abiotic) and terrestrial geology (biotic), then classified unseen samples.

"LifeTracer searches for patterns in molecular abundances, distributions and mass-spectral features across complex organic mixtures, and uses those patterns to distinguish biotic from abiotic samples, which reduces the risk of overlooking unfamiliar biosignatures," Aponte says. He adds that the framework delivers "efficient unbiased analysis and high sensitivity to unknown chemistry", handling thousands of molecular features simultaneously — including low-abundance molecules that targeted analyses routinely filter out.

Traditional biomarker approaches start from a predefined list of candidate molecules — amino acids, lipids, pigments — built largely on terrestrial biology. Aponte notes these "inherently depend on our ability to predict what molecules life produces". LifeTracer does not. Aponte concedes the limitation plainly: "Our current dataset is small, the framework can continue learning as additional extraterrestrial and terrestrial datasets become available." His team's next steps are expanding the reference database and handling more complex mixtures.

What can diversity metrics add?

A separate international team, publishing in Nature Astronomy (10 664), applies "diversity metrics" borrowed from ecology. The method measures how many molecules are present and how evenly they distribute. "Life does not only produce certain molecules. It organizes molecular mixtures in characteristic ways," says Fabian Klenner, a planetary scientist at the University of California, Riverside.

The asymmetry is the finding. "We clearly see that biological systems produce broader and more even distributions of amino acids than abiotic systems do," Klenner says. "But for fatty acids, the trend is different" — biology selects a narrower set for membrane function, while abiotic chemistry spreads across chain lengths. The method has distinguished biological, non-biological and mixed samples in tests, works on degraded samples, and needs no new instrument — it can run on relative-abundance data from current missions. Klenner flags Europa Clipper as a priority target.

Can statistics catch terraforming aliens?

The most radical proposal comes from Harrison Smith's team at the Earth-Life Science Institute, Institute of Science Tokyo. Their agent-based simulation (ApJ 1001 102) models life spreading between star systems and terraforming planets via panspermia. Result: detectable statistical correlations between planet locations and observable traits — a signature with, they argue, very few false positives.

"Technosignatures have the potential to be statistically impossible to explain without life, and biosignatures — at least the ones we'd be able to measure on exoplanets, aren't," Smith says. The approach clusters planets in "composition" space rather than physical space. It echoes Edward Schwieterman's 2024 University of California, Riverside study (ApJ 969 20), which identified five artificial greenhouse gases — C2F6, C3F8, SF6, NF3 and CF4 — that could persist in atmospheres for millennia and may prove easier to detect than biosignatures.

The ELSI model has not yet been applied to real data. Smith says ESA's ARIEL mission, surveying hundreds of diverse planets, suits the method better than NASA's Habitable Worlds Observatory, which will deeply characterise only a handful of Earth-size worlds.

What are the limits?

Thomas Beatty, an astronomer at the University of Wisconsin, cautions: "The method needs something like a thousand well-characterized Earth-like atmospheres, and we'll be lucky to get a few dozen from HWO once it launches in the 2040s." It also assumes lifeless planets don't correlate compositionally with galactic position — an unverified premise.

All three approaches treat life as "a statistical departure" from what physics and chemistry alone would produce, Beatty observes. "That's a powerful approach, because it doesn't require alien life to resemble life on Earth. But every statistical biosignature is only as strong as its null hypothesis, and right now our knowledge of the abiotic universe is the limiting step." Building that abiotic baseline, Beatty suggests, is where observational effort should now go.