Why Different Neuron Parts Learn Differently?

0:00 Your brain is constantly rewiring itself with every single ongoing experience. Like right now watching this video, the physical connections between millions of your neurons are being modified dynamically. This incredible ability is known as synaptic plasticity. And figuring out the principles behind it is one of the biggest goals in modern neuroscience. Historically, this has been notoriously hard to study. actually seeing and measuring the strength of these tiny synaptic connections is experimentally even tougher than recording the activities of neurons themselves. But in recent decades, thanks to some amazing advancements in experimental methods, we are finally getting a real glimpse into the world of synaptic plasticity even in awake behaving animals. And what we are

0:46 discovering isn't just some one universal rule. Instead, it appears that neurons have a whole toolkit of different mechanisms they use in different situations. As the study published last month in the journal science revealed and this is the paper I would like to explore today, it appears that the very same neuron actually uses different plasticity rules for its different signapses. So in this video we will talk about what these different plasticity mechanisms are, what methods researchers used to uncover it and what it might tell us about information processing in cortical neurons. Computation in our brain is carried out by neurons which are electrically excitable cells. This means that the outer membrane of a nerve cell can

1:31 amplify and send waves of electrical activity like brief zaps of charge moving between the inside and outside the cell. Typically, these electrical signals are collected by thin branching structures called dendrites. Think of them as neurons receiving antennas. The signals then travel towards the cell body called the soma. Here a crucial decision is made. If the incoming signal is strong enough to cross a certain threshold, the neuron fires off a large electrical impulse called the action potential. This action potential, often just called a spike, then zips along the axon, which is a specialized projection acting as the neuron's outgoing wire. But something really interesting happens

2:18 when this electrical signal needs to jump from one cell to another. At the point of contact called the sinapse, the information completely changes its form. Instead of staying electrical, it is converted into chemical messengers. These chemicals called neurotransmitters are released by the sending neuron, float across a tiny gap and when they are detected by the receiving neuron, they trigger new electrical signals there. This electrical tochemical to electrical conversion isn't just some weird biological quirk. It is a super sophisticated mechanism that allows for the incredible flexibility and control which make learning possible. When the spike arrives at the end of a sending

3:03 axon, it kicks off a cascade of events. First, this increased voltage at the axon terminal causes special protein channels to open, letting calcium ions rush into the terminal. This flood of calcium then causes tiny membranebound sacks called vesicles which are packed full of neurotransmitter molecules to fuse with the cell membrane and spill their contents into the synaptic clft that tiny gap between the neurons. When these molecules glutamate being the most important for us today bind to the receptor proteins on the dendrite of the receiving neuron they activate those receptors. Now, most of these excitatory sinapses are not located on the smooth

3:49 surface of the dendrite. Instead, they are often found on tiny mushroom-shaped protrusions sticking out from the dendrite called dendritic spines. Think of them as little specialized compartments for individual sinapses. So, it is on these spines that the glutamate binds to receptors allowing ions to flow into the post synaptic neuron and change its membrane voltage. The strength of this sinapse, basically how much the receiving neurons voltage changes in response to the spike of the center neuron depends on a lot of different factors. But for our purposes today, let's think of the synaptic strength mainly as the number of receptor proteins embedded in the membrane of that dendritic spine. So when plasticity

4:36 occurs, it would have to either grow that spine and its sinapse, increasing the surface area and adding additional receptor proteins leading to stronger signals, or shrink the spine and sinapse, diminishing the effect of its activation on the post synaptic membrane voltage. When the strength of a sinapse increases and stays that way for a while, it's called long-term potentiation or LTP. Think of it as turning the volume up on that specific connection. Now, I should note that there is also short-term potentiation, but this one is more fleeting. It doesn't involve these lasting structural changes. So, in this video, we'll focus only on the long-term kind, the one that truly rewires the brain, lasting from hours to days. The

5:24 brain uses specific learning rules to decide when to strengthen or weaken a connection, often based on the activity patterns of neurons involved. A famous guiding principle attributed to the psychologist Donald Hab is that neurons that fire together wire together. The idea is that if a sending neuron consistently helps to make the receiving neuron fire, the connection between them should get stronger. This is a cornerstone for understanding many forms of LTP. The key question then becomes, how does this tiny sinapse know when it's time to change its strength? To answer this, we need to zoom even more into the molecular machinery inside the

6:09 sinapse. At the heart of these synaptic changes lies a truly remarkable molecule, the NMDA receptor. This receptor can be thought of as a tiny sophisticated coincidence detector sitting right there at the sinapse. Here is the general idea of how it works. When the presinaptic neuron spikes, it releases glutamate which binds to a couple of main types of receptors on the receiving neurons dendrite. First, we have what's called ampi receptors. These are like everyday workh horses of signal transmission. When glutamate latches onto them, they quickly pop open and let positive sodium ions rush into the receiving neuron. This causes a small brief electrical excitation, a little bit of

6:55 depolarization in that neuron. The second kind is our special NMDA receptors. When glutamate binds to these receptors, they don't necessarily open right away. This is because at the neuron's resting state, the channel of the NMDA receptor is physically blocked by a magnesium ion. Imagine it like a cork stuck in the neck of a bottle. Even if glutamate is present and the channel is open, other ions can't flow through it because the magnesium is blocking the entry. Now, you might be thinking, what is the point of these NMDA receptors if they are usually blocked anyway? Well, here's where the magic happens. It turns out that if the membrane of the receiving neuron becomes deolarized

7:41 enough, meaning it gets sufficiently positively charged, right where NMDA receptor is located, this positive charge can actually push the magnesium out of the channel. With the cork gone, the NMDA receptor can finally open wide and let the calcium ions in. Calcium is an incredibly versatile messenger molecule inside the cell. When it floods through an open NMDA receptor, it initiates the whole cascade of different molecular events. These ultimately signal the neuron to do things like insert more receptors into the membrane of the sinapse and even reorganize the cell's internal scaffolding, the actin cytokeleton, which can make the dendritic spine physically larger. So when a significant amount of calcium

8:28 enters through these NMDA receptors, it is essentially the neuron getting the message. This particular signapse was with a strong activation of the neuron. So let's crank it up. But what causes this crucial deolarization which expels the magnesium block in the first place? Won't sodium coming in through the open amper receptors automatically do the trick? Well, here is where it gets interesting. Usually the activation of amper receptors at a single sinapse on its own causes deolarization that is too small and too brief to reliably open NMDA receptors. This is where the coincidence detector part truly shines. For significant NMDA receptor activation and

9:13 that vital calcium influx, we generally need a stronger and more sustained deolarization. This can happen in a couple of main ways. First leading to classical headband plasticity is the effect of the post synaptic neuron actually firing its own spike. If the receiving neuron gets enough excitatory input from many different sinapses all summing up from different dendrites, its entire cell body might depolarize enough to fire its own action potential. Now that spike doesn't just travel down the axon to signal the next neuron. an echo of it can also spread backwards into the dendrites. We call this a back propagating action potential. As a quick

9:59 note, despite the name similarity, this is totally different from back propagation in machine learning setting which is about how artificial neural networks learn. Here a back propagating action potential is just a wave of depolarization traveling in reverse direction from the cell body back into the dendrites of the same neuron. it doesn't communicate anything to other neurons only to the sinapses of that neuron signaling that it just spiked. So if a presinaptic neuron releases glutamate and this glutamate is still present bound to the NMD receptors when the back propagating action potential sweeps through that part of the dendrite the conditions for long-term potentiation can be met. The

10:45 glutamate ensures that the NMDA receptor is primed and the strong deolarization from the back propagating spike then provides the necessary voltage kick to expel the magnesium block. This near simultaneous presence of glutamate and strong depolarization allows calcium to flood in triggering the strengthening pathways. It's this crucial pairing, the presinaptic activity preceding or coinciding with the post synaptic spike that embodies the fire together wire together principle for this kind of plasticity. But there is another way this can happen leading to what's called non-hibane plasticity. Neurons especially parameal cells we find in the cortex have these incredibly complex

11:31 branching dendrites. Now imagine if several sinapses that are very close to each other on one of these fine dendritic branches all become active around the same time. Their individual amp mediated deolarizations can add up locally right there in this tiny segment of the dendrite. If this local sum is strong enough, it can push out those magnesium blocks from the NMDA receptors in that specific neighborhood, even if the neuron as a whole doesn't fire an action potential. This allows for highly specific compartmentalized plasticity. Through this kind of local coactivity, the neuron can strengthen a little cluster of inputs on one dendritic branch independently of what

12:17 the rest of the neuron is doing. Great. Now we understand precisely the question that the science paper was addressing. When a signapse under goes plasticity, one determines whether it's the heben mechanism driven by that post synaptic spike coinciding with the input or does it happen through the non-heb local coactivity instead? And crucially is there any pattern to this? Do certain kinds of synapses prefer one mechanism over the other? To find out, let's dive deeper into what exactly the researchers did. They focused their attention on parameal cells in the motor cortex of mice. These are the key neurons involved when we learn new movements. Parameal neurons have a very characteristic shape

13:04 with a main thick trunk-like dendrite called the aical dendrite. This epical dendrite reaches up towards the surface of the cortex and then branches out extensively like a crown of a tree. Closer to the cell body, there is also a bushier set of basil dendrites that spread out more horizontally. These different dendritic compartments, a versus basil, don't just look different. They also receive inputs from different sources and have distinct electrical properties. The authors hypothesized that these distinct dendritic compartments might actually employ different plasticity rules. Specifically, they suspected that the plasticity in the basal dendrites which are electrically closer to the cell body where action potentials are usually

13:51 generated might be more reliant on the classical hairband mechanism. The coincidence of synaptic input and the neurons own back propagating action potentials. But for a dendrites, especially those fine distant branches, they thought a different non-hebian rule might apply. Perhaps because these regions are so far away from the soma and more electrically isolated, plasticity there is driven more by local coactivity between nearby sinapses on the same dendritic branch independently of the cell body. While previous work in brain slices demonstrated these kind of compartment specific rules, it was still a big open question whether those processes actually happen this way in the brain of

14:36 an alive animal that is actively learning something new. And this is exactly what they set out to test. To study plasticity in a natural learning setting, the researchers trained mice on a motor task. Animals had to learn to press a lever beyond a certain point in response to an auditor cue to get a sugary reward. As mice got better at this task over several days, pressing more successfully and with quicker reaction times, their motor cortex was busy rewiring itself through synaptic plasticity. Now, during this learning process, the group was monitoring what was happening inside the mouse brain. If you just look at the brain under a normal microscope, you won't be able to see these subtle synaptic changes. So they had to employ

15:22 some very clever tricks using advanced imaging methods. First, to actually see the synaptic input, they genetically modified the neurons to express a special protein, a glutamate sensitive fluorescent reporter, which is colloquially pronounced as glue sniffer. This sensor sits at the sinapse and lights up when it binds to glutamate. This allowed them to see the activation of individual sinapses in real time as flashes of green light. At the same time, to monitor the neuron's own activity, they used another fluorescent protein, a red glowing calcium indicator called RCAMP. Calcium levels shoot up inside the neuron when it's active. The researchers established that large

16:08 widespread calcium signals that flooded entire segments of image dendrites were a reliable proxy for the neuron firing a spike. This was crucial because it allowed them to identify moments of heavy coincidence when a sinapse got input right around the time when the whole neuron fired. Finally, the authors also took structural images of the same exact dendritic spines over consecutive days before and after the learning sessions. By precisely quantifying how the size of each individual sinapse changed from one day to the next, they could identify which sinapses had undergone LTP as detected by the spine getting bigger. So with this impressive setup, they

16:53 could correlate the activity patterns of individual synapses and the whole neuron on one day with the physical changes in those sinapses on the next. This allowed them to directly ask what kind of activity predicts whether a sinapse will get stronger during learning. And what they found was pretty remarkable, beautifully confirming the hypothesis. The rules for synaptic plasticity were indeed strikingly different depending on whether they were looking at the epical or basal dendrites of the very same neuron. When they analyzed the synapses on aical dendrites, they discovered that their strengthening as seen by the spine growth was best predicted by local synaptic coactivity. In other words, an epical spine was much more likely to

17:39 grow if it and its immediate neighboring spines on the same branch were frequently active together during the learning session on the previous day. Interestingly, whether or not the whole neuron fired in action potential at the same time did not seem to be the primary driver for plasticity. It was all about that local neighborhood coactivity. To really nail this down, the authors performed a crucial experiment. They used a genetic tool to specifically block the neuron from firing action potentials. Even when the postsaptic neuron could not spike, plasticity in aical dendrites was largely unaffected. This was powerful evidence showing that a dendrites indeed use a non-habin spike independent mechanism

18:25 for plasticity. But when they shifted their focus to basal dendrites, the story changed completely. Here the strengthening of synapses was best predicted by the classical heben mechanism. The coincidence of synaptic input and the neurons own action potentials. This suggests that the functional clustering of synapses might be a less prominent feature for basal dendrites compared to aical ones. And just as they did previously, when the researchers blocked the neuron from firing action potentials, the ability of these basil sinapses to undergo LTP was significantly diminished. So the evidence was incredibly clear. The same neuron was using two distinct plasticity rules in

19:10 its different dendritic compartments. Epical sinapses seem to prioritize local teamwork and could modify themselves based on patterns of input within their small neighborhood. Basil sinapses on the other hand were more tuned to whether their activity contributed to the overall output of the neuron. But why might such functional diversity within a single neuron be important? Isn't it just a consequence of the way neurons are shaped and their electrical properties? for example that the distal apical branches are electrically isolated and so the back propagating action potentials just don't effectively invade them. Well, that's certainly a part of the picture. However, the difference between a and basal dendrites goes beyond basic biohysics. These compartments are wired

19:57 up to receive different kinds of information. For example, a dendrites are known to receive so-called feedback connections. Those are signals sent from higher order cortical areas back to the lower level processing regions often carrying contextual information or if you've watched my previous video on predictive coding they might carry the brain's predictions about upcoming sensory input. In contrast basal dendrites typically receive feed forward inputs. This is information flowing in the other direction along the processing hierarchy carrying the results of earlier computations. In the context of predictive coding, basal dendrites are prime locations for receiving prediction errors. Mismatch between what was predicted and what was

20:42 actually observed. In fact, there is a whole very fascinating line of research which we might explore in future videos suggesting that the very anatomy of the parameal neuron with two distinct input zones might be the way of the brain physically implement the computation for comparing predictions and reality. At this stage, however, we don't have a definitive answer for why learning rules need to be different in these compartments. Perhaps the non-hib and local co-activity at the aical branches is particularly useful for binding together different patterns of contextual input, allowing the neuron to learn complex associations. Meanwhile, the hein mechanisms at basal dendrites might be crucial for strengthening the pathways that reliably drive the neuron to fire, effectively wiring different

21:29 cells into specialized computational circuits. But one thing is crystal clear from studies like this. We keep uncovering evidence that biological neurons are vastly more complex and sophisticated information processing devices than we previously thought. Such insights are still largely waiting to make their way into the field of artificial intelligence. But perhaps this kind of clever compartmentalized learning using different rules for different components will turn out to be one of the keys to understanding the brain's incredible computational capabilities and efficiency. In the meantime, if you're interested in exploring deeper how computation can be carried out by single neurons, check out our today's sponsor, brilliant.org. Brilliant is one of the best places to master STEM topics online. They offer a

22:15 wide range of engaging courses combining interactive visualizations with hands-on problem solving. This approach allows you to learn by doing and develop intuitive understanding in byite-size lessons. Especially relevant to today's topics is their new course on digital circuits which covers the fundamentals of computation from logic gates which are ubiquitous in neuroscience and engineering to truth tables and equivalent circuits. Brilliant also offers a huge collection of other courses in math programming and related fields. Whether you're just starting out or looking to expand your knowledge into new frontiers, Brilliant has something for everyone. If you're ready to embrace your curiosity, head to brilliant.org/ardam kursenov to get a

23:01 30-day free trial of everything Brilliant has to offer, plus a 20% discount on annual membership. If you enjoyed the video, share it with your friends, subscribe to the channel if you haven't already, and press like button. Stay tuned for more neuroscience and machine learning topics coming up. [Music]