Editor: Editorial Department
Just today, people were shocked by this 1 cubic millimeter nano-scale human cerebral cortex map. Google's decade of neuroscience achievements, the Human Brain Atlas, has also been featured in Science. Among them, scientists have discovered previously undiscovered cells and new patterns of connection.
The human cerebral cortex can now be modeled at nanoscale resolution!
The connectomics team at Google Research has been around for 10 years. As a commemoration, just today, the team released this 1.4PB map of human brain connectomes.
It was in this image that Google's scientists discovered features that had never been seen before.
The figure contains 57k cells and 150M synapses.
1 cubic millimeter of brain, drawn in amazing detail. The impact of this picture is really shocking.
This 3D map, which covers a volume of about one cubic millimeter, is one millionth of the entire brain and contains about 57,000 cells and 150 million synapses, which is a huge data of 1.4 petabytes
For 10 years, the Google Research Connectomics team has been working to improve our understanding of brain structure and function by enabling high-throughput methods to study neural network architectures in the brain.
And this commemorative article can be said to be the culmination of the research results of the past decade.
The article has been published in Science and reported by Nature.
Address: https://www.science.org/doi/10.1126/science.adk4858
What is the significance of studying the reconstruction of the human brain?
Scientists at Google believe that if we continue to study brain connections, we may one day understand how our memories are formed, and even find the cause of neurological disorders, autism, and Alzheimer's disease.
Netizens said: "These images are so shocking, it's like looking at complex structures in outer space. Maybe in eternal life, we won't be able to understand everything in our brains."
Some people also said: "The human brain has 100 billion neurons and 3,000 trillion synapses, while GPT-4 only has 2 trillion parameters, so we still have a lot of space."
Six brain nerve maps with amazing results
The six brain maps shown next are all drawn with the help of Google AI. In this way, we finally unveil the intricacies of the brain.
Visualization through six layers of the cerebral cortex
First, Harvard researchers collected thousands of extremely thin cross-sectional images from donated brain samples.
This small piece of healthy brain tissue was removed during surgery for a person with epilepsy so that doctors could reach the area that needed surgery.
Subsequently, Google developed advanced AI tools to build an interactive 3D model of this brain tissue.
As you can see in the image below, this 3D model highlights the high complexity of the human brain.
This small sample alone (about 3 millimeters long, one millionth of the human brain's capacity) would require more than 1 million gigabytes of data, or 1.4 petabytes.
This is the highest-resolution and largest dataset ever on the structure of the human brain.
The sample comes from a part of the cerebral cortex (gray matter) called the anterior temporal lobe (see figure). The cerebral cortex has six layers that color neurons according to their size and type. In the magnified view of all neurons, the layers of neurons are clearly visible. The surface of the brain is at the top edge of the image
Dense "map"
One cubic millimeter of tissue sample contains about 50,000 cells and about 150 million synapses.
Some neuronal pairs have amazing properties, they are very closely connected, with up to 50 synapses connected to each other.
The image below shows a close-up of excitatory neurons, colored by size, with the largest red and the smallest blue. The core diameter of these cells is about 15-30 microns.
Mirror Dance
In the process of reconstruction, the researchers also discovered a peculiar phenomenon, in which cell populations often appear in a "mirror symmetry" manner.
Just like the opening picture, this pair of cells is like dancing.
This layer contains so-called "triangular neurons", which have one basal dendrites that are much larger than the others. Seventy-seven percent of triangular neurons can be divided into two main categories: one with large basal dendrites inclined to one side of the sample, and one inclined to the other side at a mirror-symmetrical angle.
Statistical analysis showed that neurons with the same tilt type were more likely to be adjacent to each other. This statistical correlation suggests that there may be some unknown underlying function at work.
Swim in synapses
Neurons in the brain, closely connected.
As follows, this neuron (white) has more than 5,000 axons (blue) from other neurons that transmit signals.
And there are at least an equal number of synapses, and the signal is transmitted from the axon to the receiving neuron.
The synapses in the diagram are shown in green.
A curious find: axonal helix
One of the peculiar findings of this study is an "axonal helix", which is the blue part of the image below.
Axons (blue) are the filamentous parts of nerve cells that are responsible for transmitting signals out of the cell.
These annular axons are very rare in the sample, and in some cases, they will be on the surface of another cell (yellow).
As for the function of this blue "axonal helix", it is still unknown.
Serious networking
The white part of the diagram below is a single neuron.
It receives signals that determine whether the neuron is firing.
And this graph shows all the axons that can tell it to emit (green) and all the axons that can tell it not to emit (blue).
Imagine how many of these neurons there are throughout the brain, that's a lot of information!
Building a "brain map" at the cellular level
Although the functions of most vital organs of the human body are not much different from those of other animals, the peculiarities of the human brain set us apart from other creatures on Earth.
The human brain, made up of billions of interconnected neural networks, is probably the most computationally complex machine in existence, capable of outperforming many power-hungry human computing systems, but it consumes only about 12 watts, about the same as an incandescent light bulb.
At present, what we know about the human brain stops at which region is responsible for what function. To further explore how it works, such as how memories are formed and the mechanisms of neurological disorders, we need to go down to the cellular level.
This is the emerging field of "connectomics", which aims to understand and accurately reproduce the connections between brain cells and build a "neuronal map" in the brain.
The connectomics imaging method used nanoscale resolution to reconstruct petavoxel-level fragments in the cerebral cortex, including 1,600 neurons, 32,000 glial cells, 8,000 vascular cells, and 150 million synapses, but the actual brain tissue involved was only one cubic millimeter, equivalent to the size of half a grain of rice.
To reconstruct the work, it is first necessary to collect image data from the real sample, i.e., a sample of brain tissue from the epilepsy patient mentioned above.
Using this sample, Professor Lichtman's team in the Department of Molecular and Cellular Biology at Harvard University made more than 5,000 sections about 30 nanometers thick and collected high-resolution images using a device called a multibeam scanning electron microscope, which took 326 days to acquire alone.
The brain tissue samples used in the study were from what was thought to be located in the left anterior temporal lobe
Based on this, the team aligned and stitched the image data to reconstruct the three-dimensional structure of each cell, including axons and dendrites, to identify synaptic connections between cells and classify cells.
Using a small piece of human brain tissue, the researchers constructed a three-dimensional image of nearly all the neurons and their connections. Excitatory neurons are shown in the upper figure, and inhibitory neurons are shown in the lower figure. The tissue samples used are approximately 3 mm wide, where the diameter of the neuronal cell body ranges from 15-30 μm
New discoveries in neuroscience
These brain reconstruction efforts have revealed some never-before-seen structures that could change our understanding of how the brain works.
For example, studies have discovered a rare but very powerful synaptic connection in which there may be more than 50 individual synaptic connections between a pair of neurons.
96.5% of the connections between axons and target cells contained only one synapse, but 0.092% of connections contained four or more synapses. The study found the morphology of these connections, which, combined with the results of statistical analysis, can show that these powerful connections are not accidental.
Further study of these connections may reveal the functions they assume in the brain, such as as as a mechanism for a rapid neural response or a way of encoding important memories.
In extremely rare cases, a single axon (blue) forms multiple synaptic connections (yellow) with the target neuron (green), the purpose of which is unknown
Given that the brain tissue samples were from epilepsy patients, although the researchers did not observe obvious pathological signs under the light microscope, it cannot be ruled out that these special structures are related to the patient's disease or the drugs they are taking, and perhaps more samples can be analyzed to clarify the cause.
These findings may be just the tip of the iceberg, but the research team said that the dataset is very large and complex, and it is believed that there are many more new brain structures and features to be discovered in the future.
AI empowers brain science
Due to the rise of AI and the development of various software tools, connectomics is becoming more and more powerful.
Before the advent of AI tools, the first linkome, published in 1986, contained only 302 neurons of Caenorhabditis elegans, and it took researchers 16 years to manually color the cells on microscopic photographs of all nematode cross-sections.
When the Google Connectomics team was founded a decade ago, one of their visions was to use the cutting-edge results of AI to process huge datasets in biology from 302 neurons to complex biological tissues of tens of billions of cells.
The development of connectomics since the 70s of the 20th century
With the help of AI, the researchers no longer need to manually colorize 1.4 petabytes of EM data, and they have developed an RNN model called "flood-filling" that can automatically segment EM images and reconstruct nerve cells with high accuracy and without the need for extensive manual proofreading.
这篇文章在2018年就发表于Nature Methods。
Address: https://www.nature.com/articles/s41592-018-0049-4
On top of this, the team also developed an automatic recognition algorithm, SegCLR, which is used to identify and classify different parts of cells in the brain neural network.
In order to store and manage massive cubes, the Google team also launched TensorStore, an open-source software library based on C++ and Python, which has been widely used and the project has received 1.3k stars on GitHub.
Project address: https://github.com/google/tensorstore