科学家们一直难以在古老岩石中确定生命的迹象。传统方法如分析有机分子的碳同位素比率，以及寻找类似微生物的结构，在鉴别1.6亿年以上的样本时效果有限。新的方法通过机器学习解决了该难题。研究人员将样本爆破后分析其分子碎片的频率和质量，并在二维图上显示数据，形成独特的峰值图谱。机器学习算法通过比较生物、有化石记录、陨石及工业产物样本的峰值“地形”，实现了对具有生物起源与无生物起源样本的判别。用该方法鉴别出的最古老生物样本年龄达33亿年，是此前分子证据可追溯样本年龄的两倍。

该团队建立了四种模型，以提高识别准确率。下一步，研究人员计划在NASA的资助下构建更大的训练集，拓展样本类型范围，包括更多真菌、化石动物及经受高温高压处理的样本。通过训练集增强，模型的判别能力将进一步提升，并有望用于分析其他星球的样本。然而，最有价值的样本目前仍困于火星表面，NASA尚无足够技术与预算将其送回地球。未来任务可能携带该系统，以现场分析样本。

此外，部分科学家推测，生命体及其成分的功能性，可能是区分生物与非生物物质的基础。这意味着决定性“地形”特征，可能反映出由具有特定功能的分子组合而成的共同属性，而非偶然拼凑。这类分析不仅有望发现生命，更可能揭示生命为何独特的根本原因。

Identifying evidence of life in ancient rocks has proved challenging. Traditional approaches—such as analyzing carbon isotope ratios and searching for microbe-like structures—have failed to reliably identify biological samples older than 1.6 billion years. A new machine learning technique addresses this issue by blasting samples and mapping the frequencies and masses of their molecular fragments onto a two-dimensional grid, producing a unique pattern of peaks for each sample. Researchers trained the algorithm on samples from living organisms, fossils, meteorites, and industrial products to distinguish biosignatures from abiotic origins. This method successfully identified a biological sample 3.3 billion years old—double the age detected by earlier molecular means.

The team built four models to increase accuracy. The next phase, funded by NASA, will involve expanding the training dataset to include a broader array of samples: more fungi, fossil animals, and specimens exposed to extreme heat and pressure. Enlarging this dataset will improve the models’ capacity to discriminate and adapt them for extraterrestrial applications. However, the most relevant samples remain stranded on Mars for now, as NASA lacks the capability and funding to return them to Earth. Future missions could deploy this technology to analyse material in situ.

Additionally, some researchers propose that the functional properties of living things may be a fundamental distinguishing feature from non-living matter. If so, these diagnostic “landscapes” may share a common characteristic reflecting specific molecular functions, rather than random assembly. Thus, identifying life this way could also provide key insights into what makes life unique.