Data and Theory
August 12, 2021
August 10, 2021
All Captioned Videos Brains, Minds and Machines Summer Course 2021
Jeff Lichtman, Harvard University
PRESENTER: I'm going to talk today. I'm giving a talk I've not given before and I'm going to try out some new ideas. And this is a talk about something that people who do connectomics are very aware of, which is the relation between information or data and ideas. Theories that explain data.
And many of you may know this already but I want to just iterate this right at the very beginning here. That there are two ways information-- by information I mean just data, data that is reliable-- and theories or hypotheses can interact. And I'm going to go through these two ways.
Some of you may already know them. But if not, or if you disagree with me, or think of another way, don't hesitate to interrupt. The first this is something called deduction. I think we all know that word deduction. The deductive approach is that you have some kind of theory about the way something works and then you check it by seeing if data, real data in the natural world, is consistent with it.
And in this sense you're testing your hypothesis by looking at data. Often when people use the deductive approach the data they obtain is explicitly acquired as a test of a theory. And therefore, they don't just get any old data. They often get data that manipulates a variable that is related to their theory with an expectation that if they manipulate a variable one way or another they will change a result.
And they can manipulate a variable like using a pharmacological agent that might block a neurotransmitter, or using a line of animals where a particular gene is knocked out, or a particular type of cell is made inactive, or they use it as something to manipulate activity. Optogenetics is a very popular tool for testing hypotheses. Or they put the animal in an unusual behavioral paradigm which based on their expectation of the theory they think the animal will behave in a particular way.
And in this sense the theorist, the person who is doing the thinking about this, can dictate exactly what experimental data is required to test the hypothesis. Now sometimes theorists have a theory already, but data becomes available to them that they didn't generate. Genomics data is a good example and connectomic data-- you know wiring diagrams is.
And then when they have their theory already they use the available data to constrain, they sometimes say, or to validate, or sometimes to invalidate theories that already are in existence. If you know anything about deductive science you know it's very hard to validate a theory. But it's easy to find data consistent with a theory.
But it's not so hard to invalidate a theory, that is find data that is in disagreement with your own hypothesis. So in principle the deductive approach can rule out hypotheses that you've been harboring a long time by showing you some piece of data is incompatible with it. In practice that rarely if ever happens because those are considered negative results.
And so many modern scientists use deductive work to constrain their hypothesis, to modify their hypothesis, but not necessarily to abandon it. Because they're in love with their hypothesis as you come to be if you have been living with a particular idea long time. You begin to believe it's true and all evidence to the contrary, you're still going to believe it's true or you'll modify your hypothesis but you won't get rid of it.
So you can already tell that I have some issues with deductive approach. But I hope everybody's clear on what I've said so far. If you disagree or you want to argue already, please go right ahead. If not I'm going to move ahead to the alternative, which is the inductive approach.
Which is you don't have a theory yet, but the theory comes based on data that already exists. That is you use data to generate hypotheses rather than to test them. Often the data that you might use was not obtained with any particular theory in mind. And it's the theory that you come up with must be configured to be consistent with the data.
Hence this approach requires really close scrutiny of the data itself, the data that was not obtained by testing a particular theory per se. And thus, you have to look at the data. And that may not be particularly in the comfort zone of a theorist. For example, a theorist interested in neural circuits if they're going to use connectomic data might actually have to look at the EM data, which is not something they really want to do.
But that's kind of what you have to do in the inductive approach. However, once you get a theory based on the inductive approach, once you've formulated it, then you go back to number one and you find ways to test and validate or invalidate. Try to validate or invalidate the theory. So let me go a little further here.
And although the vast majority of science is deductive, especially in, biology I believe that many great discoveries-- I might almost say all great discoveries in neuroscience-- have come from observations that generated hypotheses. That is that the inductive approach despite the fact it seems to be aimless, you're completely dependent on the data and not your own thinking.
It is the way most of the things we know about the nervous system that have stood the test of time. That are not theories anymore but have become facts. The idea that there are two types of processes coming off nerve cells, axons that send information and dendrites that receive them, that wasn't that someone cooked up the idea and then looked in the brain and found, I'm right. It's that people, starting with people like [INAUDIBLE] saw these different, distinctly different processes sticking out of cells. And after some work could figure out that one were receptive and the other was sending information away.
Similarly it wasn't that someone thought up synapses. People saw synapses and that then generated a theory of what they were doing. Action potentials were not inferred from theory, they were seen and then theory explained them, if you will. More modern things like place and grid cells were not coming out of the mind of great scientists and then they checked whether they actually existed. They recorded from brains of animals and saw these things. And now theories have to be created to allow us to explain why the hippocampus has place cells and why the nearby entorhinal cortex has grid cells.
Chemical organization of olfaction, the use of g coupled protein receptors as a means of organizing odor information was not inferred from a theory. It was shown. The fact that retinal ganglion cells in the eye use center surround receptive fields. This was not a theory. It was a fact and then theories grew up around it. And I could go on and on.
None of these things started with a theory. They all started with an observation and then in the hands of very talented people they used this iterative first you have an observation, then you generate a theory, then you do the experiments that might lead to other observations, and go back and forth. And finally, I think in almost every one of these cases Nobel prizes came out the other end.
So it's a pretty powerful approach. And I think for those of you who are thinking of the interface between data and sort of theory, I think theorists should figure out therefore how to leverage data to create theories rather than depend on being really, really smart and coming up with theories in the first place. This is somewhat controversial but that is my view here.
And I would say that the field I'm in now, where it's not so much theory for me but data, connectomics data will soon be ripe for this kind of approach. Where people will, with the right turn of mind, generate theories based on the data they see. And I think they will be more successful than people who already have a theory and hope to shoehorn in their theory into the data that they want. That's my point of view.
Is that clear? I can't really see very well. But if I was in the room I saw at least one nod. Did someone raise their hand? Was that an agreement, or you want to argue with me at this point?
AUDIENCE: I just wanted to ask a clarification question. The way you've put it here it seems like the theory itself is not really of any value until the second step is done, which is like the deductive process is repeated. Because I can just make up any theory to explain the data. But unless that actually can be validated or tested or at least tried to be falsified and be failed in the next step, it seems like it doesn't really get value until then.
Other than that, any reliably measured data will fall in the category of a great discovery.
AUDIENCE: Is that the right way to think about it?
PRESENTER: Yeah, well I think you're on to something there. My view is that what we're ultimately after is a kind of complete description of the natural world, including brains. We want to know exactly what they do and how they do it.
Interesting that the way I've just said that, that is really a description of the natural world but at very high resolution. Not necessarily spatial resolution, but temporal resolution, or causal resolution. That you know, what are the causes the proximate causes of every single thing that happens. None of that, that I just said actually requires you to do an experiment. Which is an experiment in the sense a manipulation to test your ideas.
It just requires having tools to see things very, very clearly whether they're in the temporal domain, the spatial domain, or in the causal analytical domain. It doesn't really matter. And I'll give you an example. There's a lot of great progress in astronomy.
That's a field that doesn't do experiments. And it works pretty well. And I would argue paleontology is another field that has made great progress but it doesn't do experiments. It just observes and then builds a knowledge that is internally consistent. And when new data appears, the theories may be modified. But it's not because of an experiment that a person thought up.
It's all data driven. So I'm very much leery I think of the rush to do experiments to manipulate variables to see if you're on the right track. That is the standard approach, I can't deny it. It's just not one that I like very much. And I don't know if that satisfies your question. But that's--
AUDIENCE: Yeah, I was kind of wondering if looking through a telescope an experiment. Or like how do you define experiment because any observation should be a result of an experiment. And it just depends on what you define experiments to be. So I would say that paleontology has plenty of experiments and astronomy has plenty of experiments. It's just not of the nature of like biological manipulation.
PRESENTER: Yeah. I think biological manipulation are the kinds of-- when I use the word experiment, that is what I'm talking about. It's people who knock out genes or use optogenetics, or whatever it is, to modify something to see if they're on the right track. And then inevitably the papers that they publish say they are on the right track. Inevitably.
I mean, that's basically the nature of the game. I'm just not sure they are, even though they might find data consistent with a theory, that doesn't prove the theory true. And so whereas finding something in the natural world that is incompatible with your theories, if your eyes are open and you have an open mind, can change the way you think about the world.
And I would say all the things I've listed there, and I could go on and on and list many other things that are just like that. They change the way you think. Not because they're an experiment in the sense of a manipulation, but because they're inductive data that you've gathered and then you generate a theory.
AUDIENCE: So how do we-- I mean even to collect some data we need to decide where to start looking. And choose where to start looking it says there must be some reason. Right? So is that reason not a theory or a hypothesis?
PRESENTER: That's another really good point. I think there's this idea called the prepared mind. That if a person knows enough, when they see something that is odd, that is not compatible with their world view, that if they're in control of the parameters that they're looking at they know that that's where they should spend some time.
A good example of this is Bernard Katz, whose name will come up a little later again. He was, you probably know, the discoverer of the way synaptic transmission works. He figured out the quantile hypothesis, a role of calcium called the calcium hypothesis. And basically the vesicle hypothesis that the quanta are little packets of neurotransmitter when the EM pictures of synapses first appeared, they were basically showing exactly what he said you might find. Even though that wasn't an experiment, it was all very consistent.
Anyway, when he was studying neuromuscular junctions he put an electrode near a neuromuscular junction inside a muscle fiber, which records the membrane potential of the muscle fiber. And some of these little blips, which we now call spontaneous miniature input potentials, but back then they were just little blips. They had a sharp rise and slow fall.
And because he understood the tools he was using, he didn't discount them as just some kind of electronic noise. He said there's something interesting about these. And then he showed that if he took his electrode out of the muscle fiber near the end plate, which is the site where the neurotransmitter is released by the nerve, and put the electrode far away the blips went away.
So he knew this was coming out of the-- this was somehow related to the nerve terminal. And then one thing led to another and then he realized that the synaptic potential when all the neurotransmitter released is just a big version of one of these miniatures. And therefore, he said synapses must be made up of many miniatures whether not spontaneously released but released to the electrical activity. And you have a quantile content, the number of vesicles released tells you how big the synapse would be.
Anyway, this was really an Earth shaking idea because it explained something about why synapses worked. And it stood the test of time. It's a great discovery and part of the reason he got the Nobel Prize but it was based on an observation that was just accidental. And he was just smart enough, and he definitely is a genius, but he was smart enough to realize that was telling him something.
It wasn't that he intentionally said I'm going to find minis. He didn't know anything about them. He just observed something and then believed enough in the regularity of what he was seeing that he could then fashion a theory that ultimately stood the test of time. So I think that's the gift of figuring out what it is interesting to look at. Some people are just very good at that.
But there's nothing-- this didn't require differential equations or despite the fact that he was very smart this was just common sense basically. And that would be what I'm arguing that virtually anybody can do this if they're interested in looking at things and studying them. They will see things that are counter to their expectation, and that's what they should go after. Not things that perpetuate an idea but things that challenge accepted dogma.
What I'm going to do is just tell you three vignettes today. These are all little pieces of very large studies. And it's more to give you a sense of exactly this observations forcing you to think about theory. And I am, I think I'm reasonably good as an observational biologist who does inductive science. I'm not sure I'm great as a theorist.
So if you have good ideas about this data, I'm all ears. In fact, none of these things I'm going to show you are published. So if you have a really good idea I will take advantage of your ideas. Maybe it will help me understand what to say about them.
So the first vignette is about a study that we've been working on in my lab for about 10 years. That's a very long time for a scientific project. And it's because it's a hard project that we are now finally writing it up. And I'm not going to tell you the whole story. That would take more than hours to do. I'm just going to tell you one little bit of it.
And it is related to the change in the wiring diagram of the axons that innervate muscle fibers, something we were just talking about, over development in mice. So this is sort of a developmental connectomic study of the neuromuscular system, the connections between nerve and muscle, in mice.
Now just three days ago I was part of a team that published a paper in Nature that was a developmental connectomic study of C elegans. And that's published and it's filled with interesting observations and theories to explain them. And I think if you're interested in how connectomics can be leveraged by having more than one time point, developmental time point, that's a paper you might look at. The first author is Daniel [INAUDIBLE] and it just came out in Nature this week.
So that would be a place to look for something that's finished. But what I'm going to show you is something that is-- we're finishing up but it is it's not finished. And it is a comparison of the wiring diagram to the same muscle at various stages from birth to adulthood. And it's a muscle called the interscutularis muscle. And we use three different techniques to look at the wiring diagram at three different ages for technical reasons.
So this image here is three panels showing the interscutularis muscle in almost reverse size order. The big panel at the top is a serial reconstruction using electron microscopy, very high resolution electron microscopy, to reconstruct every axon from the brainstem that goes to this little muscle that pulls the ear back in a mouse. It's called the interscutularis muscle.
So you see these big bundles here. Each of these bundles has many, many, many, 30, 40, 50, 80 different branches of axons in it. And you can't at this resolution see any of that. All you see is this sort of gross pattern of connections. And then at the end of the connections there are these little places where these branches come off and make synapses on muscle fibers. And these long gray things here are the muscle fibers.
This interscutularis muscle is at birth. And at birth, you probably don't know this, but the neuromuscular junctions which are sometimes called end plates, and I'm going to use that term a lot. So I know that's not a term you're familiar with, probably. The end plates tend to be contacted by multiple axons at birth. This is quite different from the same muscle in adults, where the branching pattern of the whole nerve looks the same. But now when you look at the number of axons in these bundles, there are many fewer and they terminate at single end plates. And there's only one axon going to each end plate.
So you start out with the same number of axons in this muscle as you have in this muscle. But in this muscle they're branching like crazy, going to lots of endplates. And here they're branching much less and going to fewer endplates. And the colors here-- this is connectomics, so the colors represent the different axons.
So the pink axon is going to the junctions. The neuromuscular junctions are end plates where there's a red pretzel shaped thing which is a nerve terminal releasing neurotransmitter on the muscle there. And the yellow ones are a different axon. And the blue ones are a different one. And this is a muscle that has around 19 to 20 axons.
Each muscle is slightly different and has about 200 muscle fibers. So there are about 200 different muscle fibers here. And many of these muscle fibers just have one site on their surface contacted or innervated by one axon.
This was done in the adult using a transgenic line that labels all the axons the same color of fluorescence. But because things are so big we can just with confocal microscopy trace out all the branches of all the axons. And we've been doing that in this muscle for a number of years. But we've started doing it again recently for this study I'm about to tell you about.
So we have a P0 muscle where there's a lot of extra branching. I don't, haven't shown you evidence of that yet but you will see it in a second. And the adult where everything has been trimmed away. And now what remains are strong super threshold connections between a single axon in a single muscle fiber for every single muscle fiber in this muscle.
In between at post-natal day seven, so the adult at about postnatal day 14, pretty much it's all over in mice. Everything has gotten to the adult state. This might be older, this is more like probably postnatal day 30. But postnatal day seven you're about halfway between the way the muscle is at P0 and the way the muscle will be in the adult.
And at that age things are so close together we can't use a transgenic line of mice where all the axons are the same color and then colorize just by tracing. We have to put color in the axons itself. And so we use a technique that I helped develop a number of years ago called brainbow where each axon has its own unique color.
And some of these end plates now, you can see have more than one axon there. There's a green one and a blue one at this particular neuromuscular junction. There's a red one and a little piece of a green one at this neuromuscular junction. Whereas in the adult, it's 100% one or the other.
So there is still some sharing of single endplates by more than one axon at this age. But it is becoming rarer and rarer. Although, if you look carefully, you see many, many examples here. At birth they're all multiply innervated. Every neuromuscular junction has lots of axons. At this age about half of them are multiply innervated, that is more than one axon is at the junction. And in the adult there are no multiply innervated junctions.
And this is not due to the fact that some axons die. It's just that axons start out making a lot of branches, and then they prune their branches away until each muscle fiber is single there. Is that kind of clear, so far?
OK. So here is a sort of diagram, a connectomic diagram, of the adult connectivity. And I'll explain what you're looking at here. These little black dots-- there's a big black dot and then the dots get smaller, and smaller, and smaller, until we get to the very smallest black dot.
And those black dots are the muscle fibers. The 200 or so muscle fibers in this muscle. And each of these circles is a different neuron. And each of these lines is one of the branches of this neuron to one of these muscle fibers. And as I said, in the adult each muscle fiber is only contacted by one axon. So if you look at this little black dot, I hope you can see my cursor, my green cursor there, there's only a red line going there. And that's because there's only one axon innervating it.
And you can see that some axons innervate a lot of muscle fibers. And they're arranged by me in this order here. Actually Jerome [INAUDIBLE] did the arrangement. This is work that he has been doing and I've been watching as he does it. But this is something he did.
From the biggest motor unit, the axon that innervates the most muscle fibers to the axon that innervates the least. And in this particular muscle this tiny neuron is only contacting three muscle fibers. So that's the adult connectivity, or at least part of the adult connectivity. I'll show you a little more of it in a second. You'll see what I mean in a second.
And now if we do the same exact wiring diagram at first there's no question it's a lot different. And connectomics allows us now, even though this one was done with optical microscopy and this was done with electron microscopy, I have complete confidence that this is accurate. At birth, this number of muscle fibers is the same. But now this axon, instead of branching to a large number, branches to well more than half of the muscle fibers.
And even the smallest, weakest axon, instead of only contacting 3 is contacting in this case 1, 2, 3, 4, 5, 6, 7. And you just see as you go up along here there's just a lot more branches in this muscle. Number of muscle fibers is the same but there's a lot more branches. And this is very complicated. And in every single animal, in every single muscle of every mammal, every fish, every amphibian, every avian, every muscle undergoes a transition like this after birth.
At the time animals are beginning to experience the world a profound amount of synaptic rearrangement in which synapse loss, synapse elimination, is a major player. Now one of the things we discovered in doing the connectomic birth that was completely unexpected was that what I'm showing you here are muscle fibers that have one neuromuscular junction on their surface.
Which is what we assumed was the rule. But not only is there an adult one axon innervating one neuromuscular junction. But in animals, in babies, there is one neuromuscular junction per muscle fiber. But it turned out that these we now call singly end plated or single end plate fibers, because these are muscle fibers that only have one end plate. And in the baby we were surprised to find there's another type of muscle fiber that we didn't know existed that have more than one end plate.
Each end plate is contacted by more than one axon. But there are also, and you can see there are not as many of these muscle fibers. There are fewer black dots there than here. But there are many muscle fibers that not only have a lot of axons converging on a neuromuscular junction but they have more than one neuromuscular junction. More than one end plate on the same fiber.
And it's a lot. It's about 40% of the muscle fibers-- of the axons, sorry. 40% of the axons participate in innervating these multi end plate fibers. And that made us think, well, maybe we should just check and make sure that there are no multi end plate fibers in the adult. Because we had just assumed from the textbooks and from my own papers that there is no such thing as a multi end plate fiber in adult muscle fibers, muscles.
And to our surprise, in the adult connectivity they were rare, but there were certain number of muscle fibers. Here's one, for example, that's receiving innervation at one end plate. Singly innervated at one end plate by this blue axon and another end plate on the same muscle fiber by this green axon.
Every one of these black dots here is a muscle fiber in the adult that's innervated by more than one axon. And this is actually just a subset of them. Because in some cases, as you'll see, two axons from the same neuron branch to innervate both end plates of the same muscle fiber.
But there's no question that the number of multi end plated fibers is dropping in adults, meaning that in addition to losing multiple innervation event plates adult muscle fibers are losing end plates themselves. There's two kinds of elimination-- each neuromuscular junction becomes singly innervated, and most of the multi end plate muscle fibers become single end plate muscle fibers. Most of them are single but there still are a few left that are multi end plates and fibers, meaning that there's synapse elimination within junctions and also there's complete elimination of one junction or the other.
So that every adult muscle fiber ends up with one end plate except for a small number with two. Now this was just-- this was a finding that we were not looking for. It was in a way embarrassing. It was not something we should have known, that this would have happened. It doesn't have any known significance functionally.
But there was one thing that was really interesting. That in addition to these muscle fibers that shared input from two different axons, there was way too many multi and plate fibers like this. Here's a muscle fiber. It's more than a millimeter long. And there is one neuromuscular junction site where there's [INAUDIBLE] receptors.
And there's the other neuromuscular junction site very far away on the other end of the muscle fiber. There should only be one. Here is a muscle fiber with two. And when you trace the axon it bifurcates you and one of the branches takes a very arbitrary route through the muscle, and ends up here. And the other branch takes a completely different route through the muscle.
But you know there's muscle fiber, after muscle fiber, after muscle fiber here. And it ends up here. This should not happen by chance very often. And yet in adults the remaining multi innervated muscle fibers that are far apart are innervated by the same axon more often than expected by chance.
And given that there are so few multiply end plated fibers in the adults the pairs of distant endplates innervated by different axons are more often leading to the loss of one or the other. But when both of the sites are innervated by the same axon they tend to persist. That's the idea.
Now that's interesting and there's many potential reasons why that could be. It could be there's some specificity that this axon likes the chemical nature of this muscle fiber. But we didn't know anything about this. But we did know that survival of multiple end plates on the same muscle fiber is an exception in the adult.
Given how many there are in babies, the vast, vast majority of them lose one end plate or the other. And yet here is an adult one where they both persist into adulthood and coincidentally they're both innervated by the same axon. And I'm interested in this change from multiple to single innervation injunction. So this seemed like another thing we could study.
And indeed if we make a graph of the percentage of muscle fibers that have multi junctions at birth, we're dealing with 40, 42% of the muscle fibers have more than one neuromuscular junction. And by the time we get to postnatal day 60 it's down around 9%. So you know, 3/4 of the multi end plated muscle fibers lose one of those end plates with age.
But the interesting thing is the ones that survived had a different property than the ones that got eliminated. And that was the distance between the end plates. And here at P0 at birth, here is one end plate in black at the bottom. And then as a percent of the total length of the muscle, which we say is let's say one, the second end plate at birth, when 40% to 45% of the muscle fibers have more than one end plate, they're distributed over the whole length of the muscle.
If you look in the adult, all of the multi innervated fibers that remain with more than one end plate, the other end plate is very far away from the first one. So the end plates got eliminated are close to the first end plate and the ones that stayed are far away.
That's the first thing that's interesting is that end plates that are far away have more of a likelihood of surviving than end plates that are close. But even the ones that are far away, there are many fewer than there were at birth. And these are normalized distances. At birth the muscle is smaller. So this is relative to the length of the muscle being one in all cases.
So the last remaining nearby end plates at P 10 and P 12 are these. And after P 12 they become further away. And after P 30 they're pretty far away. But the last remnants of multiple end plate muscle fibers are these end plates down here. So we looked at these.
For just curiosity, because we thought maybe we could understand how end plate elimination takes place by looking at these. When we found that all of these had exactly the same property. And I'll show you a picture of them.
The endplates are close together that's why-- and we know they're not going to last. And in every case, 100% of the cases, the two end plates are innervated by the same axon. Each of these cases. The axon bifurcates and goes to both of them.
And these end plates are the last remnants of the close-by end plates that survive. Now we know that when they're far apart and they're innervated by the same axon they survive into adulthood, but these endplates don't survive. So it means being innervated by the same axon is not sufficient to prevent you from being eliminated.
But somehow when the axons are far apart if it's from the same axon the branches they will survive. And if it's from different branches they will often be eliminated. So this is all the data I'm going to show you about this. And now I want to just talk about theory. About how to think about this.
And I don't really know exactly whether I'm on the right track. Maybe before trying to explain how this could be, I want to make sure that you understand what I said. I didn't, I'm not sure I explained it very well about multiple end plates that are far apart persist if they're innervated by the same axon. And multiple endplates that are close together disappear always. But the last ones to disappear are the ones that are innervated by the same axon at both sides.
AUDIENCE: Which fibers are we talking about? You have a progression in time. And is this something that happens at any junction in any part of the body?
PRESENTER: I mean we're only looking in one muscle. There are a number of muscles in the head and neck we now have found where multiple end plates exist on the same muscle fibers in development. And a much smaller percentage persist in adulthood. I wouldn't go so far as to say all muscles have multiple end plates on adult muscle fibers. But it is, it was certainly once we were looking for it then suddenly I felt like a fool that we'd never noticed this before.
It's not easy to find because in order to do this, you have to actually stain the muscle fiber. Most cases people stain synapses but they don't stain the postsynaptic cell. So if you see two synapses you don't know it's on the same fiber unless you're doing something as we're doing here, serially adding.
These stripes you are seeing this way is because we're cutting the muscle in this direction. And we just reconstruct and see these two end plates are on the same muscle.
AUDIENCE: Yeah I just have a question. When I see this, what is this is a function of any kind of experience. And as you said we're talking about the neck muscles. Would you expect to see the same in muscle from a different part of the body? Like I guess the question is, over the time course that you showed, is there something that's going on that affects differently the neck than any other part of the body that would tell us that we might see this in the neck but not other parts?
PRESENTER: In all parts of the body, from internal muscles like the diaphragm to limb muscles, during the first several postnatal weeks, the number of axons converging on a single neuromuscular junction goes from multiple to single, close to one. It is a new phenomenon, I think, to talk about end plate elimination. That's the idea that there are two endplates and then later on there's only going to be one.
And I know that in this case, none of these are both going to survive because if we just go back here, after P 12 there are no nearby end plates on any muscle fiber. Then this in four muscles. They're just, they're gone. These are temporary. And oddly, every one of these that are temporary, of these there were many that were different axons at the same end plate. But here that are nearby, but the ones that persist at least to postnatal 12 are always the same axon.
I don't think this is a neck muscle issue. I think the muscle we're using, which is actually pulling the ear back which is on the back of the head, I don't think has a lot to do with this phenomenon. Although, truth be told, we've seen it in a bunch of head and neck muscles. But the very long muscles in the limb, I haven't actually done. It's hard to do. You'd have to trace an axon, sorry, a muscle fiber, over a very long length to convince yourself two end plates are on the same muscle fiber.
And that's why I think people have not noticed this before because it's just hard to trace things that far. Connectomics makes is sort of the way you get this data for free if you stain the target cell as well as the axons. But I don't think there's something specially-- I'm not trying to make a case that the neck has some special property that requires this to occur.
I think this is a developmental trend that I'm showing you here that probably occurs in many muscles, not only throughout the mouse but through the vertebrate animal kingdom, I would say. So these guys are far enough apart that if one staves and the other leaves there has to be some action at a distance.
And that's even more striking for the end plates that are more than a millimeter away that are at the opposite ends of the same muscle. Many of these multiply end plated muscle fibers also disappear, even though the other end plate is down here where these black dots are. This is, there has to be some way that the muscle fiber knows that it's got another end plate because there's no muscle fiber we've ever seen that doesn't have an end plate.
So if there's one surviving the other one is probably going to disappear in most cases. And that requires some kind of action at a distance. How can a muscle fiber know to get rid of an end plate if the other end plate is a millimeter away? This is our thinking about this.
And this is not because we started wanting to think about this, it's just data that comes out of looking. Loss of end plates, which is quite ubiquitous, we saw many fewer in adults than in babies, requires distant synapses on the same muscle fiber to know that they're both there because no fiber ends up with no end plates.
If both end plates might spontaneously disappear, we should end up with a substantial number of uninnervated muscle fibers. We never see that. So that makes us think that one of these end plates is doing something to destabilize the other end plate.
Can anyone think of a signal that could pass from one end plate all the way to the other end of a muscle? Does anybody have an idea of what that might be? What kind of signal would allow one end plate to make its presence known to another end plate that's very far away?
Well muscle fibers propagate action potentials. So activity of end plate could propagate as an action potential to another end plate and maybe destabilize a site at the other end of the muscle. That's a hypothesis, but I don't know of any other signal that could go that far.
And it's interesting to know that action potentials conduct in muscle fibers very, very slowly. Only about 1 meter per second. That may seem like why, why does it matter how slow they go. But I think you'll see in a second. Why the fact that muscle fiber action potential conduction is slow has an interesting effect.
If we think of both distant end plates are innervated by the same axon, they're both going to generate action potentials, and those action potentials are going to head towards each other. But if you know anything about electrophysiology you might know that action potentials can't pass each other because after the rising phase is a refractory period where sodium channels are all inactivated.
So the action potentials collide and they never pass each other. So if you have the same axon innervating two end plates that are apart they don't know they're on the same muscle fiber because they never sensed the activity of the other end plate because their own action potential collides with it.
So you have synchronous activity, if you will, that doesn't potentiate, it doesn't do anything because it's like two planets on the opposite side of the sun. They never know they're there because there's always the sun in between. In this case, there's always a collision in between. So that rate and I'll just say, we thought about that. Well if that is why maybe two endplates can persist that are very far apart because their action potentials are colliding, then why can't two endplates that are close together, innervated by the same axon, coexist on the same muscle fiber?
We went back and read some Bernard Katz papers on the delay between the release of neurotransmitter, the invasion of an action potential in a nerve terminal and the release of a neurotransmitter, and we calculated that the likelihood for collision-- if endpoints are nearby, collision is very unlikely to occur even if it's the same axon innervating both sites because of the stochastic time between action potential invasion and the release of neurotransmitter, which according to Bernard Katz is about plus or minus 1.5 milliseconds.
So the two sites, even though they're activated by the same axon, they're not releasing their transmitter at exactly the same moment. One action potential get started slightly before the other. And because they're so close they can get to the other endplate before the other endplate can fire its action potential.
So you get collision when the two endplates are far apart. And when the two endplates are close together, you don't get collision even though it's the same axon. So this suggests a really weird form of spike timing dependent plasticity, where synchronous inputs that are far apart are oblivious to each other because of collision and are incapable of eliminating each other because they're not actually ever at the same junction at the same time. But if they are close because of the absence of collision, even though they're synchronous they're not synchronous enough and one action potential gets to the endplate, activates the receptor, activates the action potential machinery when the muscle fibers receptors are silent and the muscles-- that endplate is destabilized because it's not participating in the action potentials in the muscle.
So this is a very, I know, detailed analysis of a very peculiar phenomenon. For us it's much more important than I think it is for any of you. But I would say that we were not looking for this particular phenomenon. But we have an interesting way in which synapses that are far apart on the same target cell interact, not by potentiating each other but by either colliding or destabilizing depending on the timing of when they're activated. That's one vignette.
I'll give you want one more neuromuscular vignette, which is related to brainbow. We go a little later in development of postnatal day 6. Now we have muscles where most of the fibers are singly innervated, but some of the muscle fibers are still multiply innervated. And if I just zoom up on part of this picture, you'll see here, that these are neuromuscular junctions. And some of them, here's one that's purple and green, where two axons are co-innervating the same muscle fiber at the same site.
Now we're only going to be looking at single neuromuscular junctions not at multiple junctions on the same muscle. So this is a multiple innervated muscle fiber. This is a singly innervated muscle fiber by the orange axon. There's another multiply innervated muscle fiber with green and purple. There's another multiply integrated muscle fiber with purple and a little bit of green.
There's another neuromuscular junction with purple and a little green. And here's another one with purple and a little green. And there's several others as well. There's another one that's purple and green. Seems like these two axons seem to be sharing the same neuromuscular junction more often than expected by chance.
Because we had every axon in every connection with connectomics we could ask, given the number of branches each axon made are the branches randomly distributed at postnatal day 6 or 7 when synapse elimination has already started, or are there some places where multiple innervation is more likely to persist?
And if we look at, for example, in this very muscle at all the muscle fibers that the red axon-- with brainbow every axon has its own unique color-- co-innervates with other colored axons, we find that there are lots of muscle fibers where the red axon and the blue axon co-innervate. 65 muscle fibers.
Even though given the number of red branches and blue branches we expect to see 20 co-innervated muscle fibers, which is way too many. We see about the number as expected between red and pink, but there are too few muscle fibers sharing red and lavender. Even fewer than expected by chance of red and green. And red and purple, there's only one muscle fiber.
So red has a strong tendency to co-innervate with blue and a very, very low tendency to innervate with purple. And this is one part of a vast scheme of preferences. Every single axon over innervates muscle fibers with some axons and under represents itself with others. So for example, I showed you in that picture. There were a lot of purple and green neuromuscular junctions.
There only should have been, let's say 15 or 16 in the muscle but there was close to 60. Whereas they were about the right number of purple and lavender, but very few purple and pink, purple and red, purple and blue. And every axon seems to have some weird desire to co-innervate muscle fibers with a particular other axon and not with others.
One reason this could occur is if the red and blue axons were, all their terminals were in one part of the muscle. So they contact each other a lot because they're in the same place. But when we look in the muscle, let's say at the red axon and where it synapses were, they were distributed all over the muscle.
This is the topographic map of where the red axon terminals were. And the 66 red and blue terminals are distributed randomly everywhere. So it's not anatomy that has generated this specificity. So now I'm going to show you the full specificity of this muscle and see if you can explain this for me because I need to have this explained by somebody.
Here are all the different axons and here are the number of muscle fibers they share. So the red and the purple share 66 muscle fibers. That's a large number of fibers they share. The light green and light pink share 61 muscle fibers. That's the second most common sharing in the muscle. The light green and that middle shade purple share 60 muscle fibers.
Light green and the tan color, there are 54 muscle fibers shared. Whereas tan and this purple will share 50 muscle fibers. And I've drawn this as a sort of hexagon because there are a total of six neuromuscular junctions in this particular neck muscle. It's a different, it's not the interscutularis, it's the [INAUDIBLE].
So you might expect there should be a lot of sharing between these two. There's only one muscle fiber they share. So whatever is going on here, this muscle, this neuromuscular pit, this axon and this axon have very little in common in terms of sharing. Or at this stage, all their shared junctions have been eliminated. But a lot have persisted between this and this.
Now what about between red and pink? Fewer shared, but more than red and this dark pink, 29. The next most common sharing also skips one in this order, 22 shared between these two, 21 shared between these two. In this order now we're skipping one. So they are close together. These are one further apart in this order, whatever this order is. And every one of these ones that I'm adding now are shifted by one apart.
This one 18 junctions shared between those two. And now if I skip two, there were 16 neuromuscular junctions shared by these two axons. And there were seven neuromuscular junctions shared by these two. Again, skipping two in this particular order. And there were six between those two, skipping two, in that order.
And now if we go around this way there were only four neuromuscular junctions shared by this axon, and this axon, and this axon, skipping 1, 2, 3. And then finally, that one now makes sense because they're the furthest apart if this was a linear order, where this is nearest in some property to this axon, and this axon was nearest to this axon on one side of it in this axon on another side of it.
This is highly non-random and we saw this in six muscles. The numbers were different. The number of axons were different. But in every muscle we saw an ordering of connectivity that was not binary. It wasn't that this axon only co-innervated muscle fibers at P6 with this axon. It's that there was an order that sort of made sense.
So, there was a non-random connectivity. If we just look at the red axon here, you see the red axon has lots of muscle fibers it shares with the purple axon. Fewer with this. Fewer still with this. Fewer still with this. And fewest with that. And I can draw that as a diagram like this. This axon has 66 co-innervated muscle fibers with this one, fewer, 29 with this one, only six with this one, only four with this one and only one synapse with that one.
And I can now fill in the rest of this graph with all the other connectivity. And in every case, it's the same. The nearby neighbors have very powerful connections. But one removed connections are less strong. Two removed are less strong still. Three removed are less strong still. And the weakest connection of all is between these two at the end of this linear order connectional matrix.
That's a fancy term that I just made up. I have no idea what this actually means. But it's reminiscent of what could generate a kind of ordering of innervation, especially given what I've just told you about activity. If axons are firing together at the same neuromuscular junction they may be persisting longer then if they're firing differently.
So is it possible that this firing, this action potentials pattern of activity, is most similar to this? Next most similar to this? Next most similar to this, and so on? And that this axon firing pattern is most similar to the one on either side of it.
There is something known as the size principal where axons are activated in a fixed order from one that activates the muscle first, and the next axon activates the muscle second. Could this be some kind of firing where one axon fires first, and then sequentially another axon fires, and another one, and another one, and each axon is coexisting on muscle fibers where its firing pattern is most similar to other axons on the same muscle fiber.
It's been very interesting to me, because I'm very interested in how neurons that are active are participating in connections that perhaps persist a long time, like associated associative memory. So this is reminiscent also of something something called a synfire chain.
Importantly, this is not a theory. It was not that I was looking for this. This is just connectomics. This is what the data shows. And now it requires a theory that makes sense of this. But this could be perhaps the recruitment order of motor neurons somewhere.
All right, I want to end with something that will be less, I hope, difficult to understand. Which is a analysis from a large study that is in biorxiv right now. It's a 1.4 petabyte data set of human cortex where we've done the complete wiring diagram, Thanks to a lot of help from Google. And it's a biorxiv paper where if you want to read the paper it's here.
And I'm not going to go through the paper. I'm only going to go focus on one particular point. But to just give you a sense of how this data was generated, what you're looking at here is a piece of cortex which is the-- this is layer one where there are very few neurons but lots of astrocytes. And this is the white matter down here. So this sort of wedge is a piece of cortex that goes from white matter to the uppermost layer of cortex from a human.
This is a human sample that came from a patient who has epilepsy. And it was removed so the doctors and neurosurgeons could get into the hippocampus where the patient's epileptic focus was. To remove the epileptic focus they have to carve out a little piece of the overlying anterior temporal lobe.
And we were waiting for that. And they just dropped it into a beaker and we fixed it immediately. And we fixed it at the time of removal then stained it with osmium, using a particular osmium stain called ROTO. And then we embedded it in the hard resin. And then we cut it into strips, into sections, 30 nanometer sections that we picked up on a belt of tape using something called an ATUM, an automatic tape-collecting ultra microtome that we built as a means of collecting lots of sections from the brain automatically.
And sections are 30 nanometers thick, so about 1,000th as thick as a human hair. Very, very thin sections. And then those sections are, the strips of tape are put on a silicon wafer. And then they're imaged in a very fancy electron microscope. It just goes really fast. The multi beam scanning electron microscope. That's what the M is for. And it goes fast because it images with 61 beams. And each beam images one of these little boxes all simultaneously.
So one generates a rather large piece of tissue all imaged at the same moment with a scanning electron microscope from a single section. And then if you zoom up on any one part of this you see, you can see things like synapses. So this is a vesicle field profile using the vesicles that we use in our transmitter, making a synapse on a dendrite.
And to generate a whole section you move the stage around to tile a whole section with these little hexagons. There's one of those hexagons. So a section is about 3 millimeters because that's how thick the human cortex is. And at the resolution we're doing, that's 750,000 pixels and it's 2 millimeters or 500,000 pixels in this direction. So this is pretty big. And there were 5,292 30 nanometer sections. And each section was imaged at 4 by 4 nanometers.
A synaptic vesicle is 40 nanometers, so this is very, very fine resolution. And each of these images is about 350 gigabytes. So if you just add up those images 5,292 of them you get to 1.8 petabytes. But there's a lot of overlap of these individual tiles. And there's some areas here that doesn't brain in it.
So when we trim off all of the edges that don't have brain, that were just [INAUDIBLE] just the resin, and get rid of the overlap we end up with a data set that after stitching an alignment to remove overlap and non brain edges of the resin we have a data set that's 3 millimeters by 2 millimeters by 0.17 millimeters in depth. It's a little more than a cubic millimeter and it's 1,400 terabytes, or 1.4 terabytes.
So it's a pretty big data set. This is what it looks like. It's a cubic millimeter but it's not cubic. It's 3 millimeters long. And the region up here is the layer 1 part. And this part down here is the white matter where the myelinated axons are. And these little white dots you see all over are neurons.
So I can't really show you too many movies here just because bandwidth isn't good enough. But what Google did, we gave them some training, is they just labeled every object in here with a different color and a different ID. So these are neurons and each of them is a different color and in fact a unique ID. And the stuff in between, if we zoom up here you start to see the blood vessels, which are not labeled.
If we zoom up a little more you can, I think, make out the fact that there are nuclei in the nucleus of each of these nerve cells. These are like apical dendrites of pyramidal neurons. We zoom up a little more. We start to see these little black outlined objects, maybe. These are myelinated axons.
Here's a neuron and two other neurons sitting right next to it. A little higher you see the myelin more clearly here. There's lots of these, what we call parasitic cells. People who study glia don't like us to call them that. But these cells, these are usually oligodendrocytes, oligodendrocyte precursor cells, microglia, that just glom onto the side of big neurons.
A little higher you start to see the mitochondria inside nerve terminal processes. So between these cells are all these little cross sections. This is just one section. Go up a little higher, you can see these things a little better. And now, maybe you can begin to make out these little double colored bars.
These are synapses. In addition to labeling each object in 3D with a separate identity and color, we've also given them training data so they could localize all the excitatory in the end of the [INAUDIBLE] synapses. So vesicle film profile. So, there's 137 million synapses in here. This could not be done by hand but it could be done automatically. I'll zoom up a little more.
And so every one of these objects is a three-dimensional object that can be traced over distance. Here's a myelinated axon, for example. And the stuff inside like mitochondria are not labeled a different color. That object, the algorithm understands how to look at objects here. So it uses the outside edge of the object. And then as you play through the data set, this object will change shape as you move from section to section.
So we take all the cells in the volume and just render them. There's 52,000 cell bodies. It looks like a big mess actually. There's not a lot you can say. But if you then just measure the cell body size, just its diameter, and colorize the image based on cell body size, the cortex becomes, as expected, and very layered structure. But now the layers are not colored by layer. They're colored simply by size.
But now these layers, layer one, two, three, four, five, six and white matter all very obvious just by the size of cells. It knows that the smallest cells are blue and the very largest cells in layer five mostly are red. For the interneurons, when we looked we didn't find any layering at all. But for the pyramidal neurons, when we just show a picture of all the pyramidal neurons.
Those colors are not colored after the fact. They're colored just by size. And using cell size and several other automatic things we generated the layers, not by drawing these by hand, but just letting the machine decide where is the appropriate place to put them. So again, this is layer one where there are no pyramidal cells, just the apical dendrites of pyramidal cells. And this is the white matter, where surprise to me, there's lots of neurons that send processes up into the higher layers. This layer six, layer five, layer four with smaller cells, layer three, layer two, and layer one.
So there's lots I could say about this. And if you're interested, you go to that biorxiv paper. And there's also a wonderful website that Google set up, a landing site, where you can just click on any object, any of those colored objects, and render it in 3D and rotate the 3D object and render another thing in 3D, render a dendrite in 3D and then render the axons that innervate the dendrite. It's really quite nice and fun to look at.
One of the things, because we had all the synapses, is we could look at single cells and label all the synapses on single cells. So this is an inhibitory neuron that doesn't have a spine. Here's an excitatory neuron that does. And interestingly in humans, because this is a human tissue, there's very few synapses on the cell body. There are some and they are inhibitory. This dark purple are inhibitory synapses.
And as you get on the out on the dendrites where the spines begin then you start seeing lots of excitatory synapses. The axon has lots of inhibitory synapses right in the initial segment. This is from of type of interneuron called the chandelier interneuron. And these algorithms to locate synapses and tell whether they're excitatory or inhibitory are remarkably good, 95 or greater percent accuracy.
There are very few false positives. There are few false negatives. There are few synapses, especially inhibitory synapses, that are hard for the algorithm to find. But almost always when you look at them you have to judgment call whether you want to call it a synapse or not. Inhibitory synapses are sometimes tricky.
Nonetheless, we found that excitatory neurons almost always have more excitatory synapses on them, innervating them, than inhibitory neurons where it's more equal. Because each of these synapses is associated with an axon you can also get all the innervation to a nerve cell by just reconstructing all the axons that are making synapses on all the spines.
And I'll show you this for excitatory neuron here. It's not pretty. I mean it's depressing I guess. I mean there's just, these are, all you're seeing here are the axons that are making synapses onto this particular pyramidal neuron. Again, not too many synapses on the cell body, but when you go out onto the dendrites where the spines are there's just huge numbers of axons making synapses.
Because of this we could generate actually a wiring diagram of connections in the cortex. That is the canonical circuit in the cortex must exist because of connections between excitatory and inhibitory cells in each layer. And so I did this by hand. I took every class of synapse we knew existed in terms of excitatory neurons and inhibitory neurons.
What did we see in the connectivity? And verify that all of the ones I'm showing you here actually exist. This is a wiring diagram of the cortical circuit based on what we found. In these yellow neurons are excitatory neurons. There are no excitatory neurons in layer one so these are mostly pyramidal cells.
And these blue cells are inhibitory cells. And these small blue neurons are what are called chandelier cells that only contact the initial segments of excitatory axons. What you see here is something that's very odd, to me at least unpleasant.
It's not, it's almost an all to all connectivity between neurons of a particular type, that is simply said excitatory or inhibitory neurons. So a neuron that is excitatory in layer 2 is making synapses on virtually every other cell type in the volume.
We don't, there's a technical reason when we don't know whether they're innervating with dendrites or chandelier cells. But that's not because they don't. It's just that when we look at a dendrite we can't tell whether it's a chandelier cell. So the absence of input to these cells should is irrelevant.
What should be relevant here is that each of these cell types is contacting many of their cell types. So I can't come up with a nice theory of how information flows through a circuit this is complicated. And this is not a theory. This is all documented in the same brain, every single one of these connections exists.
But that's not really what I want to talk about. I want to talk about one thing we weren't looking for that came out of this same analysis. If we go back to one of these cells and look at all the inputs to the cell, in many cases we notice that the majority of axons that innervate the dendrites send a branch right along, right across the dendrites and they touch let's say a dendritic spine and make one synapse.
And then they move on and just keep going. They go straight. They don't seem to have any special affinity for this. A very small number make two synapses. Even a smaller number make three. And one or so make four. But this out of thousands of synapses. So it's very rare for an axon to dedicate many synapses to the same cell.
But we made this histogram for every single neuron in the volume. And none of these neurons are complete because the volume is so narrow. So this is not the entire complement of synapses itself. But we began noticing, and I noticed this myself, I was just kind of struck by this. Neurons like this one here, where again there's lots of inputs that only make one synapse on it. Some with two, some with three, and a diminishingly smaller and smaller number down to one with nine. Then absolutely no synapses on this cell of axons that make more than that.
But there's one axon that makes 19 separate synapses onto this one cell. And because this is connectomics you can just go and say, well, what is that axon? And here it is. This is an inhibitory target cell. And this axon makes a bunch of synapses on this dendrite. Then it crosses over here and makes a bunch of synapses on this dendrite. And then it crosses over and makes a bunch of synapses on this dendrite.
This input by itself would synchronously release 19 synapse worth of data onto a cell. And I suspect the cell will hear that. That's compared to LTP and other kinds of potentiation where an axon might double its synaptic strength. This is an axon with almost 20-fold more neurotransmitter release and depolarization-- this is an excitatory axon-- than a single synapse, which is the vast majority.
So even though the cell is largely dominated by weak inputs, there's one very strong input. And once I saw that I said well let's go through and look at every single outlier case. And it turned out there were thousands of them. And many of them were of the following sort.
So they weren't all e to i excitatory synapses to inhibitor cells. Here, for example, is an excitatory input in green, an axon making a synapse on an excitatory spiny dendrite. What I'd like you to notice is that there's a place here where-- these little yellow dots are where there's a synapse-- where this axon runs across this dendrite and makes a synapse.
It's not making a branch. It's just sort of landing and making a synapse and then it moves on. This is sort of an en passant synapse. It's just incidentally running across here and making a mess. But as it goes further and further away then it drops a bunch of synapses all onto spines of this very same cell.
And on the other side it sends a bunch of terminal branches up to innervate spines on the same cell. So a total of eight synapses. But not just willy nilly. There's two here, and then there's an en passant one, another one right next to it, and then a bunch over here as well.
This was a motif. We saw this a lot. There's another example of an excitatory neuron passing by an inhibitory neuron, making a synapse and then dropping a bunch of synapses here and sending up a bunch of synapses here. And if these synapses went off in any direction other than straight up they wouldn't have bumped into this particular dendrite.
Similarly, these that came down, I can imagine this was by accident basically because the axon is running past it in one direction or the other. But these others look to me guided onto the same target. So now sometimes the axon is not running past this way but the axon is running in this direction. And when it's running in the same direction we saw situations like this with an axon that makes 18 synapses. An inhibitory axon that makes 18 synapses on a dendrite of an exciting [INAUDIBLE].
And here's a case where an inhibitory axon makes 13 synapses on another inhibitory neuron. These don't seem to me to be accidents of chance. These seem to be situations where in among a vast number of weak connections there's a small number of very powerful connections. Powerful enough that they probably would dominate the activity when they're active relative to any other axon which would have to be active at the same time, 18 or 30 other exons are synchronously active.
These are all synchronous because they're coming from the same source. So I infer, this is theory again now, that sometime in development there's an incidental contact of which is the vast majority of connections in the adult brain. But every once in a while for reasons that I do not know but I could imagine are related to salience, or importance, or some kind of positive feedback that says this is a connection that should be more powerful, that incidental contact induces a directed set of contacts that now allow this one axon to have a profound effect on the target cell.
And I'm struck by this idea that when we learn things we go from a very cognitive effort to know how to do something-- like if you're learning to drive and you see a red light you've got to learn to lift your foot off the gas pedal and put it on the brake pedal. And it's cognitively difficult. You've got to practice a long time.
But after you've been driving for a while, for a year or so, then you can be talking to somebody or listening to the radio, the light turns red, you're not even aware of it. Your foot automatically goes off the gas pedal and onto the brake pedal. There's no more cognitive-- it's sort of post cognitive. It's become autonomous. It's like a reflex.
It's like you've learned that you have a reflex. But you couldn't have had it genetically because you're genetically not encoded to put your foot on a brake pedal with a red light because lights have only been red for 100 years or so. It's a brand new idea.
So, could this be the way you instantiate from a cognitive integrated type of method to one where you now have very powerful connections that just drive target sets. So these are places where I have theories. I have no idea. As I said, I'm a much better observer than I am with theory. So I don't want you to take these theories too seriously.
I just want to end by saying, in this data set there are millions of things, I think literally millions of things, that require theories that I don't actually have the slightest idea what they are. I'm just going to show you one of these that's quite striking.
I was looking at these axons running through the volume and I saw an axon that did something very strange. In the EM all these little cross sections are different axons except for this bunch that are all purple. It's all according to the algorithm that does the automatic segmentation. These are all parts of the same axon. And that seemed very odd.
So I rendered this in 3D, which just means I clicked on it. And then looked at it and sure enough the axon that was wrapping around, making this little whorl right on the cell body. And then leaving that whorl and going on over here.
This is a huge amount of length. So this is a big delay line where somebody in my lab said, maybe this is a magnetic inductive coil. I have no idea. But what was interesting is when we traced this axon further, this axon kept doing that over, and over, and over again. It made lots of whorls. So maybe this is just a really sick axon. But everything in between it is quite normal. It makes synapses all over the place. It's just got these massive delays in here by having all these whorls.
And once we did this then we asked Google to just come up with an algorithm to look. Are there other axons like this? And right now there are about 35 out of several hundred million axons in there. So they're not common, but it's not a one off. It's something, it just needs some good theorist to come up with an idea of what it is.
So I end by just saying some observations about these very large data sets, which is going to be in your future. Like it or not, they're going to be around. Lots of them. The big data can be leveraged to find rare but biologically informative results that would be much less likely to be found by sparse sampling.
For example, these uncommon powerful inputs to neurons, which are probably one in 1,000. You'd have to do a lot of single axon reporting to find one of those super powerful ones. With connectomics, they show up quite commonly. So out of every million axons there are 1,000 of them. So they're not that rare.
Big data provide many examples that allow you to start to break single types into small types. I didn't show you this but we discovered totally by accident a new type of layer six neuron that had a geometry exactly mirror symmetrical to another layer six cells whose dendrites point in exactly the opposite direction.
I didn't show it. Again I have no theory to explain why those cells exist. But big data allow you to see things by seeing so many examples of something you start to see well there are two types what used to be thought of as one type. Actually it can be divided into two categories, whether they point the dendrites in one particular direction or the diametrically opposite direction in this case.
Also, even though machine learning is imperfect you have so much data you can pick up and then post validate wiring diagrams to generate much of the canonical circuitry of the cortex, for example. It's quite amazing what you can get out of data sets like this without a great deal of effort. The effort was done by the machine learning algorithms.
These data sets are too big. You can't, they're really hard to deal with. You can't download a petabyte of data or 1.4 petabytes of data. Definitely not. And most academics don't have the wherewithal to even set this up. If it weren't for Google there'd be nothing I could do. This can only be done by big research institutes like the Allen Brain Institute or Max Planck Institute or Janelia farm.
And as I said, without Google there would be nothing I could do with data. This is a big problem for this field and for people who want to do mining of data like this. And then at the moment the tools to cull and analyze big data are in short supply.
And then at the moment skills and coding are pretty much required. And that's really too bad. This data is so interesting, but people who want to just-- you can look at it online thanks to Google. But if you want to do any sort of quantitative study you've got to be able to code. And that, it shouldn't be that way. But at the moment that's the only way you can do it because there are no linguistic engines for example, where you can just ask a question like you can ask Google what's the weather today or something.
You can't ask, at the moment, connectomic data questions like that. But certainly, and I think for those of you interested in the interface between brains and the theory of brains and maybe how we could make machines like brains, there's lots of weird and wonderful stuff yet to be discovered. And certainly new theories need to be made for such data.
And if you're interested, just please have at it if you're so inclined. Just go to that release that landing page and you can find all sorts of tools just to look at the data, even if you don't want to code. If you just want to click on things and follow them around it'll take you maybe 20 minutes to get used to looking. But it's really quite fun.
So I'm going to end there. Obviously this was a lot of work of a lot of people. We would not have gotten very far at all without Google's help, especially Viren Jain's group of really talented scientists.
If any of you are interested in generating connectomic data for some project you might be doing in the future, we can help. We have a grant from the government that allows us to help generate data sets for labs who want them. So just email me.
And I have to like the BRAIN Initiative and CONTE from the National Institutes of Mental Health for support of this data. It's expensive.
So I don't know. I hope some of this was interesting. I know I've realized as I was telling you these stories that I've been so engaged in about the neuromuscular junction, they probably seem very obscure. But I'm happy to answer questions if you have any about them. Thanks.