在 2026 年 2 月 9 日一份关于 Nature Communications 研究的报导中，研究人员介绍了 CellTransformer，这是一种机器学习方法，可摄取大规模小鼠脑细胞资料，以预测细胞类型如何聚集，并生成细致的大脑图谱。利用这些预测的「邻域」（neighborhoods），系统产出一张约有 1,300 个亚区域的地图，旨在超越粗略的解剖标签，并解释为何单一被命名的结构看起来能支撑多种看似迥异的功能。

CellTransformer 不仅重现了已知的脑部制图，还标示出先前方法未捕捉到的新型细分，包括 Allen Mouse Brain Common Coordinate Framework。在纹状体（caudoputamen）中，它以多个更小的区域取代「一个巨大结构」的观点，并与 2016 年发表的一份独立连结追踪（connectivity-tracing）地图高度一致。在脑干的中脑网状核（midbrain reticular nucleus）中，它辨识出 4 个新的邻域，这些邻域以富集的细胞类型与基因活化模式为特征，其中包含若干先前分析曾指派到不同脑部位置的细胞类型。

该论文主要展示此方法；新提出的 1,000-plus 个邻域仍需验证，例如在动物中选择性地活化或沉默特定亚区域，并量测行为变化。下一个重要步骤是转译到人类，但资料需求是瓶颈：小鼠大脑约有 100,000,000 个细胞，而人类大脑约为 170,000,000,000，且目前尚无法取得达到该规模的完整人类基因体（genetic）剖析资料。若在未来能取得足够的资料集，作者预期此方法可延伸到跨物种比较、整合连结追踪资料，以及在其他器官中进行同样细致的制图（例如对比健康与糖尿病肾脏），其核心主张是：当人类无法以这种解析度手动分辨结构时，AI 能加速发现。

In a February 9, 2026 report on a Nature Communications study, researchers introduced CellTransformer, a machine-learning method that ingests large-scale mouse brain cell data to predict how cell types cluster and to generate a fine-grained brain atlas. Using these predicted “neighborhoods,” the system produced a map with about 1,300 subregions, aiming to move beyond coarse anatomical labels and address why single named structures can appear to support multiple, seemingly disparate functions.

CellTransformer not only reproduced known brain cartography but also flagged novel subdivisions missed by prior approaches, including the Allen Mouse Brain Common Coordinate Framework. In the striatum (caudoputamen), it replaced the “one huge structure” view with multiple smaller areas, aligning well with an independent connectivity-tracing map published in 2016. In the brainstem’s midbrain reticular nucleus, it identified 4 new neighborhoods characterized by enriched cell types and gene activation patterns, including several cell types that earlier analyses had assigned to a different brain location.

The paper primarily demonstrates the method; the 1,000-plus newly proposed neighborhoods still need validation, for example by selectively activating or silencing specific subregions in animals and measuring behavioral changes. A major next step is translation to humans, but the data demands are a bottleneck: the mouse brain has about 100,000,000 cells versus roughly 170,000,000,000 in the human brain, and comprehensive human genetic profiling at that scale is not yet available. If and when sufficient datasets arrive, the authors expect the approach to extend to cross-species comparisons, integration of connection-tracing data, and similarly granular mapping in other organs (for example, contrasting healthy and diabetic kidneys), with the core claim that AI can accelerate discovery when humans cannot manually resolve structure at this resolution.