Second complete map of a fruit fly brain completed
Second complete map of a fruit fly brain completed
果蝇大脑的第二张完整图谱绘制完成
On Friday, researchers announced the completion of a map of every neuron in the brain of a male fruit fly. The “connectome” provides a tool that can accelerate neurobiology research. But it also provides an opportunity to do some science on its own, as the connectome of a female Drosophila had been completed earlier this year. 周五，研究人员宣布完成了雄性果蝇大脑中每一个神经元的图谱绘制。这张“连接组”（connectome）提供了一种可以加速神经生物学研究的工具。由于雌性果蝇的连接组已于今年早些时候完成，这项新成果也为开展独立科学研究提供了契机。
The work also provided the team behind it the opportunity to refine tools that are likely to be applied to ever-more complex nervous systems, including (potentially) those of vertebrates. The new work involved a collaboration between biologists at the Howard Hughes Medical Institute’s Janelia Research Campus and computer scientists at Google—both acknowledge that neither could have done the project without the other. 这项工作还为背后的团队提供了完善工具的机会，这些工具未来很可能被应用于更复杂的神经系统，包括（潜在的）脊椎动物神经系统。这项新工作由霍华德·休斯医学研究所（HHMI）珍妮莉亚研究园区的生物学家与谷歌的计算机科学家合作完成——双方都承认，离开对方，谁也无法独立完成这个项目。
Preparation of an entire brain for imaging at the necessary resolution requires a distinct set of skills, as does interpreting what those images indicate. But building a complete picture of the hundreds of millions of synapses in a brain as small as the fruit fly’s is a task that can’t be achieved by humans in a manageable amount of time. The people behind the effort expect that in the long term, the effort will be worth it, as the connectome could give neurobiologists a valuable tool for understanding how the brain works. 准备一个完整的大脑以达到所需的成像分辨率需要一套独特的技能，解读这些图像所代表的含义亦是如此。然而，要为像果蝇大脑这样微小的器官构建包含数亿个突触的完整图谱，是一项人类无法在可控时间内完成的任务。研究人员预计，从长远来看，这项努力是值得的，因为连接组可以为神经生物学家提供理解大脑运作机制的宝贵工具。
Establishing a connectome
建立连接组
Our interactions with the world begin with sensory input—the neurons that register sound, light, touch, and more. From there, most brain activity involves neurons communicating with each other. This communication transforms the inputs into signals the rest of the brain can interpret, routes information to relevant processing centers, and often produces some kind of output, from forming a memory to moving a muscle. 我们与世界的互动始于感官输入——即那些记录声音、光线、触觉等的神经元。在此之后，大脑的大部分活动涉及神经元之间的相互交流。这种交流将输入转化为大脑其他部分可以解读的信号，将信息引导至相关的处理中心，并通常产生某种输出，从形成记忆到移动肌肉。
All that processing is dictated by which neurons have connections to others. For example, the visual system does some basic recognition of its own before passing the results to the brain’s visual processing centers. If those centers detect something like text, they can use connections to the language centers to interpret it, and so on. To understand how a brain works, then, we need a catalog of the connections in the brain, since those dictate how information flows through its various systems. That catalog is a connectome. 所有这些处理过程都取决于哪些神经元与其他神经元相连。例如，视觉系统在将结果传递给大脑的视觉处理中心之前，会先进行一些基础的识别工作。如果这些中心检测到类似文字的内容，它们可以通过与语言中心的连接来解读它，以此类推。因此，要理解大脑是如何工作的，我们需要一份大脑连接的目录，因为这些连接决定了信息如何在各个系统中流动。这份目录就是连接组。
In practical terms, a connectome is the list of every neuron in a brain, including its location in three-dimensional space, and the connections (termed synapses) it forms with other neurons. That’s more complicated than it may sound. Each neuron can form multiple, branched processes called axons, allowing it to form hundreds of connections to other neurons. So while the nervous system of the fruit fly consists of only roughly 150,000 neurons, and the brain contains only a fraction of those, the new work discovered over 300 million synaptic connections in the fly brain. 从实际角度来看，连接组是大脑中每一个神经元的列表，包括其在三维空间中的位置，以及它与其他神经元形成的连接（称为突触）。这比听起来要复杂得多。每个神经元都可以形成多个分支过程，称为轴突，使其能够与数百个其他神经元建立连接。因此，尽管果蝇的神经系统仅由大约 15 万个神经元组成，而大脑仅包含其中的一小部分，但这项新工作在果蝇大脑中发现了超过 3 亿个突触连接。
So how do you go about mapping something like that? Gerry Rubin, a senior group leader at the Janelia Research Campus and one of the senior authors on the new paper, described how things have changed considerably based on the complexity of the system. “I was a graduate student at the [UK’s Laboratory of Molecular Biology]… and when I got there in 71, they already bought this giant computer, and they had the idea that they were going to use machine vision and computers to assemble the C. elegans connectome,” Rubin said. 那么，该如何绘制这样的图谱呢？珍妮莉亚研究园区的高级组长、该论文的资深作者之一格里·鲁宾（Gerry Rubin）描述了随着系统复杂性的增加，情况发生了怎样的巨大变化。“我当时是（英国分子生物学实验室的）研究生……1971 年我到那里时，他们已经买了一台巨型计算机，并设想利用机器视觉和计算机来组装秀丽隐杆线虫（C. elegans）的连接组，”鲁宾说。
C. elegans is a small, transparent worm with just over 300 neurons and would seem to be a tractable system. “It took them about two years to realize that the computers were nowhere near powerful enough,” Rubin said, “and so they went with printing everything out on photographic prints and colored magic markers and circling neurons and tracing it by hand.” That level of attention would simply not work for something as complex as the fruit fly, which is also very much not transparent, making its nerves difficult to image. 秀丽隐杆线虫是一种微小的透明蠕虫，只有 300 多个神经元，看起来是一个易于处理的系统。“他们花了大约两年时间才意识到当时的计算机性能远远不够，”鲁宾说，“于是他们不得不把所有东西打印在照片上，用彩色记号笔圈出神经元，并用手进行追踪。”这种工作方式对于像果蝇这样复杂的生物来说根本行不通，而且果蝇并不透明，这使得其神经成像变得非常困难。
Fortunately, computers have advanced considerably, as have the algorithms we’re able to run on them. For the new work, the biologists took a dissected fruit fly brain (along with part of its ventral nerve cord) and cut it into a huge series of evenly spaced slices. By knowing the exact order of the slices, the researchers were able to maintain the three-dimensional architecture of the brain even while converting it into a series of roughly two-dimensional objects that could be imaged using electron microscopy, provided the resolution needed to identify small cellular structures. From there, computers become essential to processing the images. 幸运的是，计算机技术已经取得了长足进步，我们能够在上面运行的算法也是如此。在这项新工作中，生物学家将解剖后的果蝇大脑（连同部分腹神经索）切成了一系列巨大的、等间距的薄片。通过掌握切片的精确顺序，研究人员能够在将其转换为一系列大致二维的对象时，依然保持大脑的三维结构。这些对象可以通过电子显微镜成像，从而提供识别微小细胞结构所需的分辨率。从这里开始，计算机对于处理这些图像变得至关重要。
Putting AI to work
让 AI 发挥作用
Michal Januszewski, a staff scientist at Google Research, told Ars that the first step is to use a form of generative AI to ensure that the areas at the site of each slice are linked up properly. No matter how carefully you slice, there will be some distortions and a bit of material lost when you make a cut. “When you take those blocks and you stitch them back together computationally, there’s a little bit of a gap in between them so the tissue doesn’t completely smoothen,” Januszewski said. “We use [a generative] model to make the tissue look as if the seams were not there, and that then makes all the downstream processing easier because you can basically ignore the problem to a large degree.” 谷歌研究中心的专职科学家米哈尔·亚努谢夫斯基（Michal Januszewski）告诉 Ars，第一步是使用一种生成式 AI 来确保每个切片位置的区域能够正确连接。无论切片多么小心，切割时总会出现一些变形和少量材料损失。“当你通过计算将这些块拼接在一起时，它们之间会存在一点缝隙，导致组织无法完全平滑，”亚努谢夫斯基说。“我们使用（生成式）模型让组织看起来就像没有缝隙一样，这使得后续的所有处理变得更容易，因为你基本上可以在很大程度上忽略这个问题。”
A separate model acts by filling in the spaces defined by cellular membranes, allowing the system to track individual cells across 3D space. “It is different in a number of ways from what people commonly think when they talk about AI,” Januszewski told Ars. “One is that it is actually a recurrent process, so it literally moves through space as it makes the outline of the neurons, and it is a visual model, so it converts voxels out of the images from the microscope, [converting them] into a 3D presentation of the neurons.” Still, other models are used to recognize synapses and classify the type of synapses. 另一个模型通过填充细胞膜定义的空间来发挥作用，使系统能够追踪三维空间中的单个细胞。“这与人们在谈论 AI 时通常所想的有很大不同，”亚努谢夫斯基告诉 Ars。“首先，它实际上是一个循环过程，因此它在勾勒神经元轮廓时会真正在空间中移动；其次，它是一个视觉模型，它将显微镜图像中的体素（voxels）转换为神经元的三维呈现。”此外，还有其他模型被用于识别突触并对突触类型进行分类。