人工智能凭借海量细胞数据，绘制出大脑新功能区图谱。
内容总结：
人工智能助力绘制大脑“细胞社区”精细图谱
长期以来，神经科学家一直致力于绘制大脑图谱，以理解不同脑区与特定功能或疾病的关联。传统绘制方法依赖显微镜观察和人工标注，耗时费力且存在主观性。近年来，单细胞RNA测序技术能识别成千上万种脑细胞类型，但海量数据反而使厘清细胞如何空间组合形成有功能的“邻里社区”变得异常困难。
近日，艾伦脑科学研究所的博西莉卡·塔西奇团队与加州大学旧金山分校的计算神经科学家合作，开发了一款名为“CellTransformer”的定制机器学习算法。该算法通过分析超过1000万个小鼠脑细胞的基因表达数据，能够以远超人类的速度和精度，识别出大脑中由不同细胞类型混合构成的、具有潜在功能意义的精细分区。
研究团队将算法应用于五个小鼠大脑的数据集。CellTransformer的工作原理类似于通过观察一个街区的建筑群来推断其中某栋建筑的类型。它通过学习细胞与其邻近细胞的基因表达关系，预测并划分出大脑的“细胞社区”。结果不仅与现有权威小鼠脑图谱高度吻合，还新发现了超过1000个以往未被描绘的精细亚区。例如，在大脑深处与运动、奖赏等功能相关、以往被视为单一结构的纹状体（小鼠中称为尾壳核）内，算法清晰地划分出了多个不同的亚区，这有望解释为何同一大脑区域能参与多种看似不相关的任务。
这项发表于《自然·通讯》的研究为脑图谱绘制提供了强大的新工具。科学家指出，更精细的脑分区图谱将有助于精准定位神经疾病的病变区域，并设计针对性干预措施。未来，团队计划将该方法应用于人类大脑，并整合神经连接追踪等其他技术，以构建更全面、立体的脑图谱。此外，该算法也有望用于绘制其他器官的细胞组织图谱，推动更广泛的生物医学研究。
专家表示，人工智能在此并非取代科学家，而是成为加速发现的得力助手，帮助人类探索仅凭自身难以洞察的复杂生物结构奥秘。
中文翻译：
以海量细胞数据为食，人工智能绘制大脑新“街区”地图
引言
房地产经纪人会告诉你，房子最重要的特征是“地段、地段、地段”。在神经科学领域，情况也类似。“在大脑中，位置就是一切，”自称“生物制图师”的博西尔卡·塔西奇说。大脑某一部位的损伤可能摧毁记忆；另一部位的损伤则可能干扰人格。没有一张好地图，神经科学家和医生就会迷失方向。
一个多世纪以来，研究人员一直在绘制大脑地图。通过追踪显微镜下可见的细胞模式，他们创造了彩色的图表和模型，勾勒出不同区域，并能将其与功能联系起来。近年来，他们增添了无比丰富的细节：现在可以逐个细胞进行分析，并根据其内部的基因活动来定义每个细胞。但无论他们切割得多么仔细，分析得多么深入，他们绘制的大脑地图似乎总是不完整、混乱且不一致。例如，一些大型脑区与许多不同的任务相关联；科学家怀疑它们应该被细分为更小的区域，每个区域各司其职。迄今为止，从庞大的基因数据集中绘制这些细胞“街区”的地图，既是一项挑战，也是一项繁琐的工作。
最近，艾伦脑科学研究所的神经科学家兼基因组学家塔西奇和她的合作者们招募了人工智能来协助分类和制图工作。他们将来自五个小鼠大脑的基因数据——总计1040万个单个细胞，每个细胞包含数百个基因——输入到一个定制的机器学习算法中。该程序输出了堪称“神经房产经纪人”梦想的地图，在较大的脑区内划分出了已知的和新颖的细分区域。人类即使花上几辈子也无法勾勒出这样的边界，但该算法在几小时内就完成了。作者于十月在《自然·通讯》上发表了他们的方法。
通过将同样的技术应用于其他动物并最终应用于人类，研究人员不仅希望详细描绘大脑更精细的结构布局，还希望生成并检验关于大脑各部分在健康和疾病状态下如何运作的假说。
“我们想了解细胞在三维空间中是如何组织的，”哥伦比亚大学的神经科学家克劳迪娅·多格说，她并未参与这项研究。“只有知道它们是如何组织的，我们才能弄清楚它们之间如何可能协同工作。”
神经制图学
大脑绘图是一门古老的科学，可追溯到20世纪初，当时德国神经科学家科比尼安·布罗德曼定义了大脑皮层——大脑外部负责思考的部分——的区域。他用一种能将遗传物质染成紫色的染料对人类大脑切片进行染色，然后在显微镜下进行研究，大脑细胞的密度和排列产生了不同的、可观察到的纹理。他勾勒边界，绘制了一张包含52个区域的地图，即布罗德曼分区，其中一些至今仍被认可。
宾夕法尼亚州立大学医学院的神经解剖学家金永秀说，几十年来，大脑绘图科学家使用的工具并不比布罗德曼的先进多少。“解剖学家过去常做的是，他们拿着一支铅笔，在大脑图像上看起来不同的区域之间‘画线’，”他说。其中一张这样的地图是2020年发布的《艾伦小鼠大脑通用坐标框架》，它基于1675个小鼠大脑的数据，包含了1000多个不同的区域。这类地图无疑具有价值，但也难免带有主观性：当金永秀向资深科学家请教他们方法的秘诀时，他说答案常常是：“全在我脑子里。”
最近，更先进的分子技术使得神经制图师能够研究单个细胞。在这种框架下，一个细胞的身份由其数万个基因中哪些被“开启”来决定，这可以通过细胞中存在的RNA分子（活跃DNA区域的副本）序列来体现。因此，科学家可以切分大脑，测量每个细胞的RNA，然后将这些基因模式映射回细胞的原始位置。
这种方法已经区分出了数千种不同类型的大脑细胞，远超以往所知。艾伦研究所2023年发布的最新小鼠大脑图谱包含了超过5000种不同的细胞类型。基于三位逝者大脑中300万个细胞绘制的《人类脑细胞图谱初稿》则定义了3313种细胞类型。
但是，这些庞大的数据集并未产生塔西奇所寻求的那种大脑地图。她说，由此生成的地图所划分的区域并不总是“具有生物学意义”。这是因为大多数脑区并非由单一细胞类型定义，而许多细胞类型也不局限于一个区域。相反，每个区域都包含多种细胞类型的混合体，包括不同种类的神经细胞以及大脑的支持细胞和免疫细胞。
打个比方，想象一下飞机上的乘客望向窗外，试图识别下方城市的街区边界。如果乘客每次只关注一栋建筑，就无法辨别其周围环境。要识别街区，他们需要关注不同类型的建筑如何聚集在一起：一个街区可能挤满了褐砂石房屋和游乐场，另一个街区可能主要是大型公寓楼和便利店，第三个街区则可能满是高层写字楼和餐厅。
要绘制大脑子区域的地图，塔西奇需要分析不同细胞类型如何聚集在一起。这对于她的人脑来说，尽管其复杂精妙，也无法仅通过研究RNA数据独立完成。
塔西奇需要更好的计算工具——以及一位研究伙伴。
街区观察
塔西奇在加州大学旧金山分校的计算神经科学家雷扎·阿巴西-阿斯尔那里找到了完美的合作者。“我一直对如何利用人工智能来理解大脑中的细胞组织感到着迷和好奇，”他说。
为了定义细胞“街区”，阿巴西-阿斯尔和他的研究生亚历克斯·李从单个小鼠大脑的390万个细胞收集的RNA图谱开始。他们编写了一个机器学习算法，选择其中一个细胞，其身份和基因表达信息会被隐藏。然后，这个他们称之为CellTransformer的人工智能，会根据其邻近细胞的基因表达和类型来预测该细胞的基因表达和类型，检查是否猜对，并根据结果更新其算法。通过将这个过程重复数百万次，算法学会了不同类型的大脑细胞如何以及在哪里聚集在一起。由此，它可以构建这些细胞群的高分辨率地图。
回到那个空中城市观察者的比喻，CellTransformer所做的相当于用拇指挡住窗户上的一栋建筑，然后预测它的类型。周围环境提供了关于哪种结构适合该街区的线索。
阿巴西-阿斯尔说，将大脑绘图视为邻近细胞之间的关系，是允许算法绘制出有意义的神经“街区”（每个街区由不同细胞类型的混合体构成）的“秘制酱料”。根据科学家要求的精细程度，它可以在小鼠大脑中定义从25到1300个不等的“街区”，尽管这未必是大脑区域数量的上限。塔西奇说，借助人工智能，“我们看到了人眼无法看到的东西。”
使用来自另外四个小鼠大脑（包括雄性或雌性小鼠，以及从左到右或从前到后切片的大脑）的单细胞RNA数据，CellTransformer生成了相似的地图。多格说，这极好地证明了该技术的可靠性。
虽然该算法利用其预测来对细胞进行分组，但它并非生成全新的地图，因此不会像某些生成式AI模型那样产生“幻觉”。尽管如此，将CellTransformer的新颖输出与已知的大脑地图进行比较至关重要。作为一个可信的比较对象，研究团队回到了手绘的《艾伦小鼠大脑通用坐标框架》。CellTransformer的地图与之匹配良好，呈现了相似的结构，如皮层的分层。
该算法还能够识别出新的“街区”，即包括《艾伦小鼠大脑通用坐标框架》在内的先前神经科学方法所遗漏的区域。以纹状体为例，这是大脑中部附近一个有条纹、大致呈C形的结构。在小鼠大脑地图中，纹状体被称为尾壳核，“你只看到一个巨大的结构，”加州大学洛杉矶分校的神经解剖学家洪里格·欣蒂里扬说，她未参与这项新研究。已知它参与运动、奖赏和整体大脑管理。大脑的一个部分如何能执行如此不同的任务？
CellTransformer的解释是，它毕竟不是一个统一的大脑区域。该地图证实，尾壳核实际上被细分为更小的区域，尽管研究人员尚未将每个区域与特定功能对应起来。此外，这些新的细分区域与欣蒂里扬及其同事在2016年基于完全不同的技术（追踪尾壳核与其他区域之间的连接）发表的地图吻合得很好。
欣蒂里扬说，在整个大脑中识别出这样的子区域，可以解决神经科学家之间关于同一大型脑区承担截然不同功能的争论。很可能“他们都是正确的，只是他们关注的是不同的区域，”她说。
阿巴西-阿斯尔和塔西奇对CellTransformer能够准确匹配已知大脑地图的能力感到兴奋，更令他们激动的是该算法绘制出了新颖的细分区域。例如，脑干的中脑网状核，参与启动运动，是一个研究相对不足的区域，阿巴西-阿斯尔说。CellTransformer在那里识别出了四个新的“街区”。每个“街区”都具有特别普遍的细胞类型和特异性激活的基因。它们还拥有几种在早期分析中被归入大脑完全不同部位的细胞类型。
地图在手
《自然·通讯》上的这篇论文主要是介绍CellTransformer方法并展示其能够发现新区域；这上千个新“街区”仍需验证。如同任何对新领域的探索，绘制地图仅仅是开始。最令人兴奋的是科学家们可能用它来做什么。“我们对结构的理解越精细，我们的探究和干预就能越具体，”欣蒂里扬说。
新出现的问题集中在所有这些神经“街区”的功能上。为了精确确定每个微小区域的作用，科学家可以在实验动物中消除或激活这些新识别的区域，然后观察行为变化。
真正的成果将是将CellTransformer应用于人类大脑。多格推测，一些“街区”在小鼠和人类之间会匹配得很好，而另一些则会存在差异。不幸的是，该算法做出准确预测所需的数据量目前尚无法从人类大脑获得——至少目前如此。小鼠大脑包含约1亿个细胞，而人类大脑约有1700亿个，并且这个庞大的细胞集合仍在进行基因分析。阿巴西-阿斯尔和塔西奇认为，当足够数量的此类数据可用时，CellTransformer将能够应对挑战。
他们还对将其他技术（如欣蒂里扬使用的连接追踪技术）整合到CellTransformer中感兴趣。这就像在城市街区地图上添加街道和高速公路。此外，除了大脑，同样的算法还可以提供其他器官的详细细胞地图，使科学家能够比较，例如，健康肾脏与糖尿病肾脏。
人类科学家单凭自己根本无法理清这些细节。“我把AI视为人类的一种助手，”金永秀说。“发现将以一种戏剧性的方式加速。”
英文来源：
Fed on Reams of Cell Data, AI Maps New Neighborhoods in the Brain
Introduction
Real estate agents will tell you that a home’s most important feature is “location, location, location.” It’s similar in neuroscience: “Location is everything in the brain,” said Bosiljka Tasic, a self-described “biological cartographer.” Brain injury in one spot could knock out memory; damage in another could interfere with personality. Neuroscientists and doctors are lost without a good map.
Researchers have been mapping the brain for more than a century. By tracing cellular patterns that are visible under a microscope, they’ve created colorful charts and models that delineate regions and have been able to associate them with functions. In recent years, they’ve added vastly greater detail: They can now go cell by cell and define each one by its internal genetic activity. But no matter how carefully they slice and how deeply they analyze, their maps of the brain seem incomplete, muddled, inconsistent. For example, some large brain regions have been linked to many different tasks; scientists suspect that they should be subdivided into smaller regions, each with its own job. So far, mapping these cellular neighborhoods from enormous genetic datasets has been both a challenge and a chore.
Recently, Tasic, a neuroscientist and genomicist at the Allen Institute for Brain Science, and her collaborators recruited artificial intelligence for the sorting and mapmaking effort. They fed genetic data from five mouse brains — 10.4 million individual cells with hundreds of genes per cell — into a custom machine learning algorithm. The program delivered maps that are a neuro-realtor’s dream, with known and novel subdivisions within larger brain regions. Humans couldn’t delineate such borders in several lifetimes, but the algorithm did it in hours. The authors published their methods in Nature Communications in October.
By applying the same technique to other animals and eventually to humans, researchers hope not only to detail the brain’s finer-grained layout but also to generate and test hypotheses about how the organ’s parts operate in health and disease.
“We want to understand how the cells are organized in 3D space,” said Claudia Doege, a neuroscientist at Columbia University who wasn’t involved in the study. “Only if we know how they are organized can we figure out how they can potentially work with each other.”
Neural Cartography
Brain mapping is an old science, dating back to the early 1900s when the German neuroscientist Korbinian Brodmann defined regions of the cerebral cortex — the outer, thinking part of the brain. He stained human brain slices with a dye that turned genetic material violet and then studied them under the microscope, where the densities and arrangements of brain cells produced different, observable textures. He traced the borders to create a map of 52 regions, known as Brodmann areas, some of which are still recognized today.
For decades, brain-mapping scientists wielded tools little more advanced than Brodmann’s, said Yongsoo Kim, a neuroanatomist at Penn State College of Medicine. “What anatomists used to do is, they have a pencil, and they draw the line” between different-looking regions on brain images, he said. One such map, the Allen Mouse Brain Common Coordinate Framework, which was published in 2020, was based on data from 1,675 mouse brains and includes more than 1,000 different areas. Such maps are undeniably valuable but also inevitably subjective: When Kim asked senior scientists to impart the secrets of their methods, he said the answer was often, “It’s all in my head.”
Recently, more advanced molecular techniques have allowed neuro-cartographers to investigate individual cells. Under this framework, a cell’s identity is determined by which of its tens of thousands of genes are turned on, something that can be represented by the sequences of RNA molecules (copies of active DNA regions) present in the cell. Thus, scientists can slice up a brain, measure the RNAs from each cell, and then map those genetic patterns back to the cells’ original locations.
This approach has distinguished thousands of individual types of brain cells, many more than previously known. The Allen Institute’s latest mouse brain atlas, published in 2023, includes more than 5,000 different cell types. The first-draft Human Brain Cell Atlas, based on 3 million cells from the brains of three deceased people, defines 3,313 cell types.
But those massive datasets didn’t yield the kind of brain cartography that Tasic sought. The resulting maps generated regions that weren’t always “biologically meaningful,” she said. That’s because most brain regions aren’t defined by a single cell type, and many cell types aren’t limited to one region. Instead, each area contains a mixture of cell types, including different kinds of nerve cells plus the brain’s support and immune cells.
For comparison, imagine an airplane passenger looking out the window and trying to identify neighborhood boundaries within a city below. If the passenger focuses on just one building at a time, they can’t discern its surroundings. To identify neighborhoods, they need to focus on how different building types group together: One neighborhood might be crowded with brownstones and playgrounds, another could be populated by mostly larger apartment buildings and bodegas, a third might be full of high-rise office complexes and restaurants.
To map the brain’s subregions, Tasic needed to analyze how different cell types grouped together. That’s not something her human brain, for all its glorious complexity, could do on its own by studying the RNA data.
Tasic needed better computational tools — and a research partner.
Neighborhood Watch
Tasic found the perfect collaborator in Reza Abbasi-Asl, a computational neuroscientist at the University of California, San Francisco. “I have always been fascinated and intrigued by how we can leverage AI to understand cellular organization in the brain,” he said.
To define cellular neighborhoods, Abbasi-Asl and his graduate student Alex Lee started with RNA profiles collected from 3.9 million cells in a single mouse brain. They programmed a machine learning algorithm to choose one cell; its identity and gene expression would be masked. Then the AI, which they called CellTransformer, would predict that cell’s gene expression and type based on those of its neighbors, check if it had guessed right, and update its algorithm based on the result. By repeating this process millions of times, the algorithm learned how and where different types of brain cells group together. From there, it could build a high-resolution map of those groups.
Returning to that airborne city observer, what CellTransformer does is the equivalent of holding up a thumb to the window to block one building, and then predicting its type. The surroundings provide clues as to what kind of structure fits into the neighborhood.
Approaching brain mapping as relationships between nearby cells was the “secret sauce,” Abbasi-Asl said, that allowed the algorithm to map out meaningful neural neighborhoods, each made of a blend of different cell types. Depending on the level of granularity the scientists asked for, it could define anywhere from 25 to 1,300 neighborhoods in the mouse brain, though that’s not necessarily the upper limit of brain regions. With AI, “we see things that a human eye cannot see,” Tasic said.
Using single-cell RNA data from four additional mouse brains — including ones from male or female mice, and ones sliced from left to right or front to back — CellTransformer produced similar maps. This, Doege said, is excellent evidence that the technique is reliable.
While the algorithm used its predictions to group cells, it wasn’t generating wholly new maps, so it couldn’t hallucinate as some generative AI models can. Nonetheless, it was essential to compare CellTransformer’s novel output to known brain maps. As a trusted comparator, the team returned to the hand-drawn Allen Mouse Brain Common Coordinate Framework. The CellTransformer map was a good match, laying out similar structures such as the layers in the cortex.
The algorithm was also able to identify new neighborhoods, regions that previous neuroscience methods, including the Allen Mouse Brain Common Coordinate Framework, had missed. Take the striatum, a striped, vaguely C-shaped structure near the middle of the brain. In maps of the mouse brain, where the striatum is called the caudoputamen, “you just see one huge structure,” said Hourig Hintiryan, a neuroanatomist at the University of California, Los Angeles who wasn’t involved in the new project. It’s known to participate in movement, reward, and overall brain management. How could one piece of brain perform such disparate tasks?
CellTransformer’s explanation is that it’s not one uniform brain region after all. The map confirmed that the caudoputamen is, in fact, subdivided into smaller areas, although researchers have not yet matched each region to a function. Moreover, the new subdivisions corresponded nicely to a map that Hintiryan and colleagues published in 2016 based on an entirely different technique, which traced connections between the caudoputamen and other regions.
Identifying such subregions across the brain, Hintiryan said, could resolve debates between neuroscientists who assign vastly different functions to the same large brain region. It seems likely that “they’re both correct, they’re just looking at different areas,” she said.
Abbasi-Asl and Tasic were thrilled with CellTransformer’s ability to accurately match known brain cartography, and even more excited that the algorithm mapped novel subdivisions. For example, the brainstem’s midbrain reticular nucleus, which is involved in initiating movement, is a fairly underexplored region, Abbasi-Asl said. CellTransformer picked out four new neighborhoods there. Each of those neighborhoods featured particularly prevalent cell types and specifically activated genes. They also had several cell types that earlier analyses had placed in an entirely different part of the brain.
A Map in Hand
The Nature Communications paper serves mainly to introduce the CellTransformer method and show that it can find novel regions; the thousand-plus new neighborhoods still require validation. As with any exploration of new territory, drawing the map is just the beginning. What’s most exciting is what scientists may be able to do with it. “The more granular our understanding of structure, the more specific we can get with our interrogations and interventions,” Hintiryan said.
Emerging questions center on the functions of all these neural neighborhoods. To pinpoint what each bit does, scientists could eliminate or activate these newly identified regions in lab animals and then check for behavioral changes.
The real prize will be to apply CellTransformer to human brains. Doege suspects that some neighborhoods will match well between mice and people, while others will diverge. Unfortunately, the quantity of data the algorithm needs to make accurate predictions isn’t available from human brains — at least, not yet. While the mouse brain contains about 100 million cells, the human brain has around 170 billion, and that menagerie is still undergoing genetic analysis. When sufficient amounts of that data become available, Abbasi-Asl and Tasic think CellTransformer will be up to the challenge.
They are also interested in incorporating other technologies, such as the connection tracing used by Hintiryan, into CellTransformer. This would be like adding streets and highways to the city neighborhoods. And beyond the brain, the same algorithm could offer detailed cell maps of other organs, allowing scientists to compare, for example, healthy versus diabetic kidneys.
Human scientists simply can’t sort out these details on their own. “I see AI as kind of a helper for the human,” Kim said. “Discovery will be accelerated in a dramatic way.”