Neurons in all sorts of different colors, some glowing
Neurons in all sorts of different colors, some glowing

How Fear Restructures the Mouse Brain

By combining deep learning and electron microscopy, researchers now have a more detailed understanding of how fear changes the brain.

A black and white headshot
Natalia Mesa

Natalia Mesa was previously an intern at The Scientist and now freelances. She has a PhD in neuroscience from the University of Washington and a bachelor’s in biological sciences from Cornell University.

View full profile.


Learn about our editorial policies.

Aug 15, 2022

ABOVE: © iStock.com, bestdesigns

Neurons communicate via synapses—tiny, button-like protrusions that sprout from one neuron and connect it to the next. These minuscule structures are thought to be the backbone of learning and memory, changing in strength and number as we learn. At about 1/5,000th the width of a human hair, synapses can be hard to visualize, and researchers are just beginning to develop the tools necessary to do so. 

In a study published in Cell Reports on August 2, researchers at the Chinese Academy of Sciences and Shanghai University used a combination of deep learning algorithms and high-resolution electron microscopy to map out how frightful experiences rearrange brain connections. They found that when mice learn to fear the sound of a buzzer, neurons in their hippocampus form more connections with other neurons downstream and shuttle more mitochondria to synaptic sites. This change in neural connectivity increases the information storage capacity of the brain without increasing synaptic density. 

“I think this paper is interesting . . . both from a technology perspective, and also from a biology perspective,” says neuroscientist Shu-Hsien Sheu, a senior scientist at the Howard Hughes Medical Institute’s Janelia Research campus, who was not involved in the study. “Their paper is trying to address the fundamental but still unresolved question of what underpins learning and memory.”

The study coauthors imaged the mouse brains using a technique called serial-section Electron Microscopy (ssEM), which shoots beams of high-speed electrons at thin slices of tissue to generate ultra-high-resolution images. This allows researchers to visualize the smallest of structures.

This technique captures images layer by layer, resulting in a detailed picture of everything that’s in each thin slice. The method generally produces massive amounts of data. In this case, the result is a messy constellation of the cells, their organelles, passing axons and dendrites, and synapses. 

See “Synapses on Stage

“The paper is replete with very state-of-the-art deep learning-based image analysis,” says Stephen Smith, a molecular neuroscientist at the Allen Institute for Brain Science in Seattle. “And it’s one of the first papers to apply high throughput computational electron microscopy with deep learning . . . to the problem of learning.”

Wading through this data can be extremely cumbersome. The 2D images ssEM produces need to be assembled into a 3D cube, in which synapses take shape among the tangle of structures that ssEM images also capture, allowing researchers to pick them out of the crowd. This laborious task is typically done by hand and can take hours, says Sheu.

Image of cells in prefrontal cortex in grey
Transmission electron microscopy image of cells in the prefrontal cortex
Mikaela Laine, University of Helsinki


Previous attempts to automate the process of identifying synapses through ssEM have used the full, reconstructed 3D EM images as opposed to individual 2D images, but have either been slow or not performed as well as humans do. The coauthors of the new paper, however, used a different approach. They trained an algorithm to pick out tell-tale markers of synapses—structures found within synaptic vesicles, the synaptic cleft, and the postsynaptic density—on each 2D image. They also trained a separate algorithm to look for mitochondria. 

Then, once the algorithms had picked out synapses or mitochondria in each image, they used another algorithm to automatically line the images up. This sped up the process of picking out synapses and mitochondria dramatically, allowing the researchers to study hundreds of thousands of synapses in parallel. The algorithm could pick out cells at high accuracy in one-third the time a human could. “Extracting EM data . . . is time-consuming,” says Sheu. “They were able to extract information on a very large scale. The sample size . . . is orders of magnitude higher than if one were to do it by hand.”

Synapses between two neurons are formed at specialized sites called presynaptic boutons, which form a one-way connection to one or sometimes multiple dendrites, the tree-like branches of neurons that make up the receiving end of synaptic transmissions. Typically, this is a one-to-one connection, but in mice, boutons can form connections with multiple dendrites, and dendrites can form connections with multiple boutons. “These multi-contact synapses are not that great in number,” says Sheu.

To look at how these particular synapses change in the context of fear learning, the researchers analyzed the synapses of mice that had learned to associate a tone with an electrical shock to their foot. The researchers found that while the number of synapses didn’t change, individual presynaptic boutons connected to more dendrites after fear conditioning. The team also found mitochondria in larger numbers gathered at the presynaptic boutons, even those without more connections. This implies that mitochondria are dividing quickly at the presynaptic site. Since mitochondria provide energy for cellular activity and help cells regulate calcium—the molecule that triggers the release of neurotransmitters—they’re likely important players in increasing synaptic strength. 

The researchers then used a machine learning approach and information theory—which quantifies a network’s capacity to store and relay digital information—to calculate the amount of information stored by a computational neural network modeled after the networks they imaged with ssEM. They found that a network with a larger proportion of multi-contact synapses was able to encode more information than a traditional network built with only one-to-one synapses. This implies that, although the number of synapses in the tissue they imaged doesn’t change, the number of bits of information that could potentially be stored in the mouse hippocampus still increases after fear conditioning.

Sheu says that, in the future, he’d like to see the researchers classify these multicontact synapses further. “With [ss]EM data . . . you can probably tell [the difference] between inhibitory and excitatory synapses, but you can’t really determine the types of neurotransmitters” they use. He’d like to see what types of neuromodulators, like serotonin and dopamine, the modified synapses use to excite their postsynaptic partners, since these molecules are likely very important for learning and memory. 

Both Sheu and Smith think that future studies are likely to benefit from this approach. “I think in the future if EM . . . can be done more easily . . . people will start to do this type of [study] in . . . different experimental conditions,” says Sheu.