Liqun Luo Science Heavy Review System Interpretation ︱ Neural circuit structure-the system that makes the brain "computer" run at high speed

Written by Zheng Yuanjia, edited by Wang Sizhen, "Principles of Neurobiology" (Higher Education Press) has become an indispensable treasure tool for neurobiology research and application workers.
The author of this book is Luo Liqun
.

Liqun Luo is a neurobiologist, a member of the American Academy of Arts and Sciences, a member of the National Academy of Sciences, and a professor at Stanford University
.

On September 3, 2021, Professor Liqun Luo published the latest review article titled "Architectures of neuronal circuits" on Science, reviewing the use of tracers, physiological records, functional interference and calculations from the perspective of neural circuits over the past few decades.
Models and other different technologies study the different connection modes and functions between neurons, and discuss the possible situations of these loop structures in the process of development and differentiation, providing us with another comprehensive system of neural loop knowledge Summary
.

Introduced into the human brain, there are approximately 100 billion neurons, and each neuron has thousands of synaptic connections
.

Each neuron is a complex signal processing unit in itself, but the synaptic connection on the neuron enables the formation of a special neural circuit between the neuron and the neuron, which in turn makes the brain a truly powerful "computer" system
.

As shown in the figure, if you think of the brain as an article, then the neuron is like the "text" or "letter" in the article, and the microloop in a small area is like the "word" or "word".
"(Word), the neural connections on the cross-regional plane of a large area are like a "sentence" (sentence)
.

 Figure neurons and the brain (picture quoted from: L.
Luo, Science 2021; 373: eabg7285) At the level of "words", that is, the micro-neural circuit, it is mainly the specific connection between excitatory neurons and inhibitory neurons.
Form the most basic information processing functions
.

These micro-loops are the core unit that builds the entire complex signal processing and transmission system of the brain
.

At the "sentence" level, the neural circuits have diversified connections and functions due to the expansion of anatomical structures and regions
.

There are a large number of circulating neural circuits in the nervous system, which constitute neural activity dynamics
.

At present, many neuron loop structures at this level have yet to be discovered
.

 Although the brain is likened to a computer, the difference is that the computer is designed from top to bottom, but the neural circuit is the product of hundreds of millions of years of biological evolution, not designed by a "designer"
.

Some microcircuits in the brain may have originated a long time ago.
They were preserved in a certain branch during the evolutionary process, and then evolved to other neural regions.
Different neural subsystems developed independently during evolution, and the proliferation and proliferation of neurons Differentiation also plays an important role in the evolution of the brain, and these different factors may cause the interruption of brain connections
.

Background More than 100 years ago, Ramón y Cajal et al.
proposed that neurons are the basic unit of the nervous system, and information is transmitted from the dendrites of the neurons to the cell body, and then to the axons (Figure 1) [1]
.

However, individual neurons do not work in isolation, they work together in neural circuits to process information
.

Therefore, understanding how these connection patterns achieve specific calculations will enable us to decipher the principles of information processing in the nervous system and will promote new advances in artificial intelligence
.

   Figure 1 Information transmission of vertebrate neurons (picture quoted from: L.
Luo, Science 2021; 373: eabg7285) Progress Part One Common circuit motifs-"words" in brain "articles" 1.
Feedforward excitation (feedforward excitation) A series of continuous connections between excitatory neurons constitute feedforward excitation, which is the main way for signals to be transmitted from one neural area to another (Figure 2A)
.

At each stage, neurons usually receive signal input from multiple presynaptic connections (convergent excitation, Figure 2A, B), and through branch axons to multiple postsynaptic connections (divergent excitation, Figure 2A C) Output signal
.

Convergent excitation enables postsynaptic connections to selectively respond to certain neurons in presynaptic connections
.

When multiple input neurons carry the same but uncorrelated signals, the signal-to-noise ratio of the signal can also be improved
.

Divergent excitation is the process of processing the same signal by multiple downstream channels
.

For example, in the visual system of mammals, the signal goes from photoreceptors → bipolar cells → retinal ganglion cells → lateral geniculate nucleus (LGN) neurons → layer 4 primary visual cortex (V1) neurons → other layers of V1 neurons → Neurons in the upper cortex [1-3]
.

Along these feedforward excitement paths, the form of visual information transforms from light intensity to other forms such as contrast, edges, objects, and motion
.

This feed-forward excitement structure of the visual system inspired the development and development of "perceptrons" (that is, graphics recognition machines that simulate the human optic nerve control system), as well as "deep neural networks" for cognition and classification, and artificial intelligence is also used This technique solves problems beyond image analysis [4]
.

 2.
Feedforward and feedback inhibition (Feedforward and feedback inhibition) Although the long-range signals in the nervous system are mainly transmitted by excitatory neurons (with some exceptions, such as the basal ganglia and cerebellar circuits), inhibitory interneurons It also plays a key role locally [5-6]
.

Two widely used modes are feedforward suppression and feedback suppression (Figure 2B)
.

In feedforward inhibition, inhibitory neurons receive signal input from presynaptic excitatory neurons, so inhibitory (Figure 2 B, upper C neuron) and presynaptic excitability (Figure 2 B, upper A neuron) The input is concentrated in postsynaptic neurons
.

In feedback inhibition, the inhibitory neuron receives input from the excitatory neuron and projects it back to the excitatory neuron
.

In the above-mentioned visual pathways, almost every excitatory connection is accompanied by feedforward inhibition, feedback inhibition, or both
.

For example, LGN neurons directly activate V1 GABA releasing neurons to provide feedforward inhibition to layer 4 excitatory neurons, and layer 4 excitatory neurons also activate V1 GABA releasing neurons to provide feedback inhibition on themselves [7-8]
.

 Feedforward inhibition is faster than feedback inhibition because feedforward inhibition only passes through one synapse to reach the postsynaptic target cell after the excitatory signal, while feedback inhibition passes through two synapses (Figure 2B)
.

Feedforward inhibition is proportional to the input intensity, while feedback inhibition is proportional to the output intensity, both of which are used to adjust the duration and amplitude of the incoming excitement signal
.

For example, limiting the duration of activation in response to sensory input can allow loops to quickly return to their baseline activity levels to maximize their sensitivity to future environmental changes
.

Feedforward and feedback inhibition neuron networks usually work together to perform many interesting functions, such as adjusting the gain and dynamic range of the input signal, and promoting synchronization or oscillating discharge [6, 9]
.

Feedforward and feedback inhibition also play a vital role in maintaining the "balance" between excitement and inhibition (for example, strong excitement accompanied by strong inhibition) to prevent excessive excitement or inhibition
.

Figure 2 Common loop mode
.

 A Feed-forward excitation; B feed-forward inhibition and feedback inhibition; C side inhibition; D mutual inhibition (picture quoted from: L.
Luo, Science 2021; 373: eabg7285) 3.
Lateral inhibition side inhibition (Figure 2C) It is a widespread loop mode
.

It selects the information to be delivered to the downstream loop by amplifying the difference in activity between parallel pathways
.

For example, the photoreceptor neurons in the vertebrate retina activate horizontal cells to feedback inhibition of many nearby photoreceptor neurons
.

This behavior acts on the typical central-peripheral sensory area of ​​downstream ganglion cells, thereby enhancing the ability of these downstream neurons to extract spatial or color contrast information [10, 11]
.

  Fourth, mutual inhibition (Mutual inhibition) The communication between inhibitory neurons can make the loop have more characteristics
.

For example, if inhibitory neuron A directly inhibits inhibitory neuron B, then activation of A will release B's inhibition of the target neuron
.

If B also inhibits A, then they form a mutual inhibition mode (Figure 2D)
.

Mutual inhibition is widely used in circuits with rhythmic activities, such as those involving exercise [12]
.

In a longer time frame, mutual inhibition can also be used to regulate brain states, such as the sleep-wake cycle [13,14]
.

 The above discussion only involves "words" composed of two "characters": excitatory neurons and inhibitory neurons
.

In fact, neuron microcircuits are very rich
.

There are many differences between excitatory neurons and inhibitory neurons due to the heterogeneity of their dendritic morphology, ion channel properties, potential properties, impulse properties, subcellular distribution and intensity of input and output synapses
.

For example, in the neocortex of mammals, there are three inhibitory neurons: Martinotti cells, basket cells, and chandelier cells.
Their presynaptic ends point to excitability.
The distal dendrites, cell bodies, and initial segments of axons of pyramidal neurons control how the pyramidal neurons integrate synaptic input and generate impulse spikes
.

In the stomach ganglion (stomatogastric ganglion), neurons that inhibit each other have different ion channels and input-output synaptic strength, which is the basis for them to keep firing continuously in each rhythm cycle
.

In addition, the neuron microcircuit also includes many types of regulatory neurons that will be discussed next
.

The second part has a special loop structure with specific functions-the "sentences" in the brain "articles".
The loop levels to be discussed next are more diversified in scale and configuration, and are not so easy to summarize
.

The author attempts to generalize some high-order loop structure patterns that have been found in multiple neural regions and different species
.

1.
Continuous topographic mapping (Continuous topographic mapping) Continuous topographic mapping is a common organization method for presenting information in the nervous system
.

Adjacent input neurons are connected to adjacent target neurons through ordered projections of axons (Figure 3A)
.

For example, in retinal topology mapping, neighboring retinal ganglion cells synapsely connect to neighboring LGN neurons, LGN neurons are connected to neighboring V1 neurons, and V1 neurons are connected to neighboring higher-order visual cortex neurons
.

Retina topological mapping enables the spatial relationship of the external world captured by the retina to reproduce the captured information in the V1 and higher visual cortex regions
.

In the sensory and motor model, sensory stimuli from the neighboring body are roughly reproduced in the neighboring areas of the primary somatosensory cortex, and the motor output to neighboring body parts is also roughly controlled by the neighboring motor cortex
.

 Due to its robust development mechanism, topological mapping can provide a convenient way to process continuous hierarchical information
.

It has many advantages in information processing and calculation
.

For example, retinal topology mapping promotes the ability to extract local contrast through side suppression, thereby enhancing object recognition
.

In addition, by placing loop components adjacent to each other, the mapping can save energy by minimizing the loop length
.

The design of "convolutional neural networks" (convolutional neural networks (including convolutional calculations and deep structure feedforward neural networks, which is one of the representative algorithms of deep learning) is based on topological mapping, which greatly reduces the need to adjust artificial neural networks.
The number of variables needed, thereby speeding up the calculation speed [4, 15]
.

Figure 3 Specific loop structure for specific functions
.

A continuous brain map; B discrete parallel processing
.

(Picture quoted from: L.
Luo, Science 2021; 373: eabg7285) 2.
Discrete parallel processing Discrete parallel processing (Figure 3B) presents and processes signals in parallel through discrete information channels
.

A typical example is the vascular globule tissue of vertebrate olfactory bulbs and insect antennal lobes: olfactory receptor neurons expressing the same odorant receptor transmit their axons to the same vascular bulb, and then synapse connects to their corresponding secondary On the dendrites of projection neurons, discrete olfactory processing channels are formed [16,17]
.

Different axons converge on the same spheroid, which improves the signal-to-noise ratio
.

Different vascular spheres represent not continuous signals, but discrete olfactory receptor neurons, which reflect the nature of the chemical substances that activate these odor receptors
.

Discrete parallel processing is also a characteristic of the mammalian taste system
.

  Discrete parallel processing is often used in combination with continuous topological mapping
.

For example, in the retina, discrete layers are superimposed on the topological map of the retina.
Different bipolar and ganglion cell types form specific connections to process different types of visual signals, such as brightness, color, and motion in parallel
.

Compared with serial processing of information, parallel processing reduces the depth of calculation, thereby reducing the error rate and increasing the processing speed
.

A distinguishing feature of a complex nervous system (with a large number of neurons and a large number of connections between each neuron) is the ability to perform large-scale parallel processing, and this structure is increasingly used in the design of computer systems [ 18,19]
.

 3.
Dimensionality expansion In the dimensional expansion structure, signals from a relatively small number of input neurons diverge to a much larger number of output neurons (Figure 3C), allowing output neurons to present different signal input combinations
.

For example, insect mushroom bodies (olfactory projection neurons → mushroom body Kenyon cells → mushroom body output neurons) and vertebrate cerebellum (mossy fibers → cerebellar granule cells → Purkinje cells)
.

In both cases, a relatively small number of input neurons (projection neurons or mossy fibers, respectively) synapsely connect to a large number of output neurons (Kenyon cells or granular cells, respectively)
.

Therefore, the output neuron level information can be presented in a higher-dimensional space, and each dimension represents the firing rate of a cell
.

The firing patterns of postsynaptic neurons can be used to distinguish small differences in input patterns (for example, different projection neuron groups represent different odor combinations)
.

Using this structure, learning can be done by adjusting the synaptic strength of the output neuron.
After training, the same input can produce different output patterns (Figure 3C)
.

Another example of dimensional expansion is the entorhinal cortex→dentate gyrus granular cells→CA3 pyramidal neural circuit (Figure 3D, upper)
.

A large number of granular cells in the dentate gyrus can perform pattern separation of information about space and objects from the entorhinal cortex, and further processing by the hippocampus downstream loop [20,21]
.

But unlike the mushroom body and cerebellar cortex, this circuit cannot be used for learning and training
.

This may be because the microcircuit in the hippocampus conducts unsupervised learning, while the cerebellum and mushroom microcircuit conducts supervised reinforcement learning
.

 Figure 3 Specific loop structure for specific functions
.

C dimension expansion; D circulatory circuits at the level of mammalian brain neuron groups (top) and the level of mammalian brain regions (bottom) (picture quoted from: L.
Luo, Science 2021; 373: eabg7285) 4.
Circulatory circuits ( Recurrent loops) The nervous system is full of cyclic circuits in which neurons are usually connected to each other through interneurons
.

These circulatory circuits are not uniform in scale from specific nerve areas to most areas of the brain
.

For example, in the mammalian visual system, in addition to the "bottom-up" projection from LGN→V1→higher cortex, the "top-down" projection from higher cortex→V1→LGN also has attention control And many other functions
.

Long-distance loops can include continuous topology mapping or discrete parallel processing architecture (Figure 3D)
.

Circulatory circuits usually support abundant neural activity, but in most cases, their exact role is unclear and may vary from case to case
.

Understanding the principles and functions of information processing in cyclic circuits is also a major challenge in modern neuroscience
.

 5.
Biased input-segregated output (Biased input-segregated output) In addition to the loop composed of excitatory and inhibitory neurons, the nervous system also uses modulatory neurons to complete important functions
.

Since regulating neurons transmit signals through neurotransmitters (such as monoamines and neuropeptides), compared with fast excitatory and inhibitory neurotransmitters (binding to ionic receptors), regulating neurons are Posttouch neurons act more slowly
.

In addition to acting through the synaptic cleft, regulatory neurotransmitters can also be released at non-specific postsynaptic sites, the so-called "volume release", which can be used farther than the typical synaptic cleft.
Affect the target in distance
.

 Using mouse virus tracing technology, it can be observed that the midbrain dopamine, midperitoneal serotonin, and hypothalamic neuropeptide galanin systems all adopt a "biased input-separated output" structure (Figure 4A) [22,23]
.

Each system can be divided into multiple parallel subsystems to separate output projections for different target areas of different behaviors and functions
.

Each output subsystem receives a certain amount of deviation input from similar areas, so that these subsystems can be adjusted differently by external and internal stimuli
.

But there is another situation, such as the locus coeruleus norepinephrine system.
Although the branching pattern of a single neuron may be specific, its axons project to one brain area and also widely project to other areas
.

These observations indicate that the locus coeruleus norepinephrine system uses an integrated delivery architecture (Figure 4B), which may be suitable for regulating the overall state of the brain (such as wakefulness)
.

Figure 4 The input-output of the neuromodulation circuit
.

A.
Biased input-separated output structure; B.
Integration-propagation structure (picture quoted from: L.
Luo, Science 2021; 373: eabg7285) In addition to the loop architecture patterns summarized above, some other architectures, such as mammals, have also been found Neocortex, glial circuits, etc.
do not completely meet the above categories.
In the future, we still need to continue to explore the depth and breadth of neural circuits.
Only when we know more "sentences" and understand the changes and complex interactions in them, we can better In-depth understanding of their composition (such as "paragraphs") and even the final entire "article", that is, the entire nervous system
.

 The third part is the view of progress and development.
Computer circuits are the product of top-down design, and complex neuron circuits have evolved for hundreds of millions of years
.

Neural circuits also use evolutionary selection and experience to self-assemble or fine-tune during development
.

Observing a neuron circuit alone may not tell us which elements are functionally important, but by observing what is selected, expanded, reduced, eliminated, or repeated in the evolutionary process, it can hint which elements should be focused on in functional research
.

 1.
Evolution of neuronal circuits (Evolution of neuronal circuits) The existing bilaterally symmetrical nervous system may gradually evolve and mature; first, neurons with only muscle cells, and then a series of evolution of sensory motor neurons, starting from independent From sensory and motor neurons to interneurons and central neuron networks, the central nervous system and brain are finally produced [24, 25]
.

Some core modes, such as feedforward excitement and feedforward/feedback inhibition, may have started at an early stage when animals have interneurons and central nervous system.
Due to their effects, these modes are preserved in different species, and are used every time.
In different neural regions of each species
.

Other architectures also developed independently.
For example, the vascular globule organization of the olfactory system of insects and vertebrates is likely to be the result of convergent evolution, because other biological branches that evolved from their common ancestors did not have such organizations, and they used different Types of molecules act as odor receptors
.

 The gradual complexity of the nervous system requires the number of neurons, the types of neurons and their connections, and the expansion of brain regions
.

All these processes are caused by changes in DNA
.

A key mechanism of evolutionary innovation is gene duplication and differentiation.
For example, the duplication and differentiation of cone opsin genes make certain primates have trichromatic vision [26]
.

Replication and differentiation are also used in the evolution of neuron types and brain regions [27, 28]
.

In principle, the repetition and divergence of the evolution of brain regions can make neuronal circuits modular, that is, the abundant connections within the repeating units and the sparse connections between the units
.

In turn, the modular nature of neuronal circuits may accelerate evolution, because different modules can evolve independently of each other
.

 2.
Development of neuronal circuits (Development of neuronal circuits) Evolution mainly affects neural circuits by modifying genes involved in circuit connections during development
.

A very critical question is how a limited number of genes construct a nervous system with a large number of synaptic connections with specific patterns and structures
.

Extracellular signals and their cell surface receptors can identify specific targets through axons and dendritic growth cones.
They are the main molecules that establish the initial nervous system organization.
These molecules can also specify processes very precisely in certain circuits and organisms.
Touch connection【29, 30】
.

In order to establish the specificity of a large number of connections with a limited number of genes, one of the strategies is to use different expression levels of the same protein to specify different connections
.

This strategy can be used to construct a continuous topology map (Figure 5A), which may also be the reason for the rich loop structure
.

The gradient expression of cell surface molecules is also used to construct early discrete maps
.

However, discrete parallel processing needs to distinguish discrete cell types, and often use combined cell surface protein coding, so that a small number of proteins can specify more connections (Figure 5B, left)
.

The way to achieve combined coding is to divide the connection process into different spatio-temporal steps (Figure 5B, right)
.

In addition to saving molecules, this strategy can also enhance stability
.

Through the complicated time-conditioning control of its expression mode, the same connecting molecule can be used at different times and places, or in different parts of the same loop [31, 32]
.

 Figure 5 Connecting the neural circuit
.

A protein gradient constructs a continuous topological map; B combination strategy constructs a specific connection; C Herbert's law constructs a specific connection
.

(Picture quoted from: L.
Luo, Science 2021; 373: eabg7285) Spontaneous and experience-driven neuronal activity also refines synaptic connections
.

Activity-dependent synaptic connections have been shown to be produced through competition between neurons of different activity levels
.

An important mechanism for neuronal activity to influence connections is through Hebb's rule: the firing of presynaptic neurons leads to intensified firing of postsynaptic neurons, that is, "fire together and connect together" (Figure 5C)
.

In addition, non-Hebbian mechanisms, such as homeostasis synaptic plasticity, also contribute to activity-dependent loop connections [33].
These activity-dependent mechanisms continue to be maintained in the adult nervous system, allowing animals to follow their Life experiences change their synaptic connection patterns
.

 But many synaptic connections are not completely specific
.

For example, in the neuromuscular system of vertebrates, the connection between the motor neuron population (pool) and the muscle is specific, but the specific connection pattern between the motor neuron and the muscle fiber is highly variable [34]
.

Similarly, in the Drosophila olfactory circuit, the synaptic connections between specific olfactory projection neuron types and mushroom body Kenyon cells are mostly random [35, 36]
.

 In short, there are two common mechanisms for establishing the connection mode of neuronal circuits: first, molecules connect to the nervous system, and second, neuronal activity and adjust connections with experience
.

There are also interactions between neuronal activity and molecular connections.
For example, neuronal activity can regulate the expression of molecular connections or supplement the role of molecular connections [37, 38]
.

However, in addition to the several examples discussed above, most of the current developmental studies have failed to solve the problem of how the loop pattern and architecture are established, and most studies on the function of the loop have not considered the limitations of development
.

Therefore, the cross-study of the development and function of neuronal circuits will help to answer these questions.
Summary and application of research tools in the scope of the circuit, such as the use of electron microscopy and transsynaptic tracing in different biological and neural regions, can be Obtaining a lot of data to explore the common laws of neural circuit architecture, recording neuronal activity and key factors in interference circuits can help understand their functions in information processing and animal behavior
.

The author believes that a key challenge in the future is how different loop modes and structures operate across scales? In the cross-species key loop structure, it is more valuable to study how "words" and "words" are combined into "sentences", and the use of single-cell transcriptomics to compare different neuron types in homologous brain regions is also the beginning A feasible method for related research
.

Comprehensive research on the structure, function, development and evolution of neural circuits will enable us to go beyond the level of individual neurons and gain insights into the nervous system, which will also inspire new artificial neural networks, which may one day realize general artificial intelligence
.

Original link: https:// doi .
org/ 10.
1126/science.
abg7285 Selected articles from previous issues [1] New discovery of EMBO J︱! AGHGAP11B promotes the expansion of the neocortex into adulthood and improves cognitive ability [2] Cell Death Differ︱ Qi Yitao/Wu Hongmei and others cooperate to reveal the molecular mechanism of SUMO modification regulating neurogenesis in adult mice [3] Cereb Cortex︱A2A receptor antagonist can Reversing the sequence learning impairment induced by abnormal aggregation of α-Syn [4] Neuron︱Nicotine promotes the new discovery of anxiety-the important role of inhibiting the ventral tegmental area-amygdala dopamine pathway [5] Int J Mol Sci︱ Frontier review Interpretation: Pathophysiological response and role of astrocytes in traumatic brain injury [6] Cereb Cortex | Wang Lang's research group revealed that astrocytes have experience-dependent steady-state plasticity [7] New discovery in Nature︱! The social spread of maternal behavior caused by oxytocin neurons [8] Genome Biol︱ Ding Junjun’s team systematically draws a three-dimensional structure map of chromatin during phase separation, dissolution and reconstruction [9] New Brain︱ method! Plasma astrocyte proliferation GFAP or a new potential marker for Alzheimer’s disease detection [10] Autophagy︱Zhang Zhidong’s team reveals a new mechanism for STING1 to induce autophagy to regulate RNA virus infection [11] Nature︱ Astrocytes Source IL-3 regulates the function of microglia, relieves AD pathological changes and cognitive impairment [12] JCI︱ Gao Tianming’s research group reveals that the prefrontal cortex has an opposite role in regulating anxiety and fear in the neural circuit [13] eLife︱ Single cell sequencing and neural circuit analysis jointly reveal the molecular genetic coding mechanism of brain initiation of attack/defense instincts [14] Nature︱Frontier! GluDs transduce different pre-synaptic signals to different post-synaptic receptor responses.
References (Swipe up and down to view) [1] SR Cajal, Histology of the Nervous System of Man and Vertebrates (Oxford Univ.
Press, 1995).
[2] CD Gilbert, TN Wiesel,