How fly neurons compute the direction of visual motion
March 22, 2022
March 22, 2022
Alexander Borst, Max-Planck-Institute of Neurobiology, Martinsried, Germany
All Captioned Videos Brains, Minds and Machines Seminar Series
Detecting the direction of image motion is important for visual navigation, predator avoidance and prey capture, and thus essential for the survival of all animals that have eyes. However, the direction of motion is not explicitly represented at the level of the photoreceptors: it rather needs to be computed by subsequent neural circuits, involving a comparison of the signals from neighboring photoreceptors over time. The exact nature of this process represents a classic example of neural computation and has been a longstanding question in the field. Only recently, much progress has been made in the fruit fly Drosophila by genetically targeting individual neuron types to block, activate or record from them. Our results obtained this way demonstrate that the local direction of motion is computed in two parallel ON and OFF pathways. Within each pathway, a retinotopic array of four direction-selective T4 (ON) and T5 (OFF) cells represents the four Cartesian components of local motion vectors (leftward, rightward, upward, downward). Since none of the presynaptic neurons is directionally selective, direction selectivity first emerges within T4 and T5 cells. Our present research focuses on the cellular and biophysical mechanisms by which the direction of image motion is computed in these neurons.
TOMASO POGGIO: So welcome to everybody to this CBMM talk in our weekly series. As several of you know, I spent 10 years at the Max Planck Institute working with Werner Reichardt on the visual system of the fly. Among others, Martin Heisenberg was there. Martin was the son of Werner Heisenberg.
At the time, one of the problems that really fascinated me, like today, is how the brain computes. And I thought at the time that limitless eye was a very good example of a linear computation edge enhancement. But it was linear filtering, lateral inhibition.
And the first, I thought, non-trivial example of an interesting algorithm and nonlinear one, was the computation of the direction of motion. And there was a kind of phenomenological theory at the time by Reichardt and Hassenstein based on directional selectivity computed by certain algorithm model, the key operation of which was multiplication.
There was an attempt at that time by Baron Levick to explain in terms of neural mechanisms, invoking the threshold mechanism of the spike. But we found there was a more interesting way to do it, which was on dendritic trees.
And this was a model. And this takes understanding how neurons and synapses do computational. Basic operation is important, not only because we want to understand the nervous system at various levels, including the detailed ones of the circuits, and the elements, but also because the mechanisms themselves are likely to influence the algorithms and how the organism behaves.
And so the talk today will really answer this question about the basis of their actual selectivity, the cellular and synaptic basis of it. Axel Borst has focused his career on solving the problem of how the fly's neuron compute motion. And I would say, he has solved the problem. And he will explain today how that whole thing works.
Axel was born in Germany, is 10 years younger than me. He is director of the Max Planck Institute of Neurobiology Martin Heisenberg was his advisor. And after that, he worked in the same institute I was at, the Max Planck Institute for Biological Cybernetics, and in the Friedrich Miescher lab, and is now leading his own group, director of the Institute.
And it's very nice to have you here, Axel. And I hope very much that we can repeat this in person. But in the meantime, [INAUDIBLE] virtually on Zoom.
ALEXANDER BORST: So, yeah, thank you very much, Tommy, for the nice invitation and the introduction. And, yeah, I promised to really come by MIT and see you there next time I'm in the US. And then meet each other in person again.
So what Tommy told you about his interest is really not that different from my interest. In fact, I am a computational neuroscientist. And I want to understand how neurons make an interesting computation with the signals they receive from other cells, and compute an interesting output quantity from that.
And I want to understand that at the level of neural circuit components, which cells provide the input which participate in the circuit. But, eventually, I also want to understand that, at the biophysical level, what transmitters are used, and what is the electrical equivalent circuit of that, so really very much into the nitty gritty details of computations performed by single neurons.
And as a practical example for neural computation, I found the question of how neurons compute the direction of motion when I came to Tubingen. And this question never left me. And I spent almost all my career with this question now.
So for the non-fly person, let me give you an introduction to the fly's nervous system first. So here you see the fly's nervous system in blue. And if we zoom into the brain, you see these colored areas there. And the optic lobe is shown in red, which takes a large part of the central brain.
And this is-- the optic lobe in red is like the equivalent of the mammalian retina, and is concerned with image processing. And it's built from several layers of neural tube that is shown here. Here you have the retina, the lamina, the medulla, the lobula, and the lobula plate.
And what you notice is the columnar structure of each of these neural [INAUDIBLE] which is retinotopic so that neighboring columns, neurons in neighboring columns process the signals that they receive from neighboring facets. And at the level of the lobular plate, you see large tangential cells which with the large dendrites receive the signals from many hundreds of columnar neurons.
And here's the central phenomenon that I want to understand. If you take a bar and move it to the right and to the left again, and you record from the photoreceptors, then you see a depolarization each time. So clearly, the signal is non-directional. From just watching the signal, you cannot tell whether the bar move to the left or to the right.
Now, you take this electrode and move it just a few synapses downstream into one of the large tangential cells. And here you see a clearly direction selective signal. The cell depolarizes in response to rightward motion. That's called the preferred direction. And it hyperpolarize in response to leftward motion. That's called the cells null direction.
So within a few synopsis, you grow from non-direction to directional signal. So how come? And that's exactly what the model explains that hassenstein and Reichardt proposed in the '50s. This paper that is cited here came out in '56.
This is one year before I appeared. And the model that they derived from behavioral experiments on stationary walking beetles on this spherical treadmill here is shown here from the original paper with all the bells and whistles. And here is the essence of the model.
It consists of two subunits which share the same two photoreceptor input. And within each subunit, the signal from one photoreceptor is delayed by a low-pass filter and multiplied with the instantaneous signal it receives from the neighboring photoreceptor. That's done twice in mirror symmetrical subunits.
And when you then subtract those subunits from each other, you get a nice, directional selective signal, as we've seen before. And this model not just explains qualitatively how direction selectivity arises, it also describes very much in a quantitative way of what we've seen in the tangential sense of the fly.
And I spent much of my career together with my colleague Martin [INAUDIBLE], thinking about any kind of stimuli, doing computer simulations, running it through the Hassenstein Reichardt detector, and then doing experiments on the fly lobular plate tangential cells. And each time, it confirmed the model prediction pretty well.
So that, of course, led to the question of which cells then, presynaptic to these tangential cells, perform this computation. So let's take a look at what cells we find in the fly optic lobe. So here you see again the schematic of the fly optic lobe. And we take a horizontal cross section through the optic lobe.
And then we see, first of all, that photoreceptors are 1 to 6. They terminate in the lamina. And in the lamina, these signals are picked up by lamina monopolar cells, which then bring the signal into the medulla. And within the medulla, we find different types of cell classes, so-called transmedulla neurons, TM, which take the signal from the medulla to the lobula.
There are intrinsic medulla neurons, MI cells, which connect different strata within the medulla. And there are the so-called T4 and T5 cells. And they bring the signal-- T4 cells bring the signal from the medulla to the lobular plate. And T5 cells have the dendrite in the lobula, and then have the axon terminal here in the different layers of the lobula plate.
So in the following, I will really concentrate on these T4 and T5 neurons. Well, here you see an account of all the different cell types we find in the optic lobes and this paper shows the Golgi stainings that Fischbach and Dittrich published in '89.
And this is a real jungle of about 100 different cell types, so in terms of complexity, comparable to the mammalian retina. But the problem was that for a very long time, these cells were way too small to record from them. And we could didn't have any calcium indicators at that time.
So for about two decades, this was like a silent collection of all different cells. That was very unsatisfying. Because on the other hand, we had the Reichardt detector, which describes so well the input, output transformation, and what we then see in the lobula plate tangential cells.
So for a very long time, this relationship was there. It was like a black box. And we couldn't open the black box until we really started to work with Drosophila. Because the sophisticated genetics that we have in Drosophila together with calcium indicators, et cetera, then allowed us to open this black box.
And just within the last, I would say, 12 years, we made a series of discoveries that allowed us to identify the elements of the Reichardt detector within the [INAUDIBLE] of the fly. The first thing we found was that we're not dealing with one elementary motion detector, but actually with two.
Because the fly visual system computes the direction of motion in two parallel on and off channels, just like the mammalian retina. Second thing was we found that the first primary motion sensing neurons are the T4 and the T5 cells, T4 in the on pathway, T5 cells in the off pathway.
And, finally, we found lobula plate interneurons, which implements the subtraction stage of the Hassenstein detector. So let me spend the next few minutes introducing the T4 and T5 cells and these lobular plated interneurons in a bit more detail to you.
The T4, T5 cells were actually, like every other neuron on Earth, seen by and drawn by Ramon y Cajal. And here, from his paper that appeared in 1915, you recognize that these are T5 cells and these are T4 cells. And it's very funny to read his commentary on these cells.
And he calls these cells [SPEAKING SPANISH], which means those strange things giving rise to bi-tufted sticks. To him, it looked like a stick with two bushes at each end. And he didn't know what was a dendrite and what was an axon terminal. So he could not draw the arrows like he did with pyramidal cells and all that. So that was something that he had never seen before.
But meanwhile, we know that there are four subtypes of each of these cells. And each subtype ramify in one of the four layers of the lobular plate. So in total, we have eight cells, anatomical subtypes, four T4 cells, four T5 cells. And they ramify in the four layers of the lobula plate.
And here you see a single cell flip out, again, with the dendrite of a T4 cell here. The dendrite in the medulla and the axon terminal in the lobular plate. And here you see an enlarged top view of the dendrite of a single T4 cell with the different columns in magenta. And you see that the dendrite covers 7 to 8 different columns. So this is where it collects the input signals from.
That's a very important feature of the cells. And since the fly has 750 columns in the visual system and we have eight of these cells per column, then we should come up with about 6,000 cells in total on each side of the brain. And this shows you a cell-body staining of hemi brain. And I really counted all of them and made sure that the number matches the expectation.
So what are the visual response properties of these cells? This is something we wanted to know for a very long time. So about 10 years ago, we got a driver line from Gerald Rubin's lab at Janelia Farm, which stains the T4 and the T5 cells. And here you see a confocal image of these driver line.
This is the dendrite of T4 cells, dendrite of T5 cells. And this is the inner chiasm. Here you see beautiful four layers of axon terminals formed by both T4 and T5 cells. And now two people in the lab, Matt and Jeurgen Haag, they took this driver line, crossed it with [INAUDIBLE], and took them under the two photon microscope.
And the area that they imaged is shown here. And already in the two-photon microscope, just the structure, you see 1, 2, 3, 4 different layers. And now they stimulated the fly with rightward motion. And they saw a very strong activity largely in layer 1.
When you move the pattern to the left, the activity is shown in layer 2. When you move the pattern upward, you see activity in layer 3. When you move the pattern downward, you see activity in layer 4. And now if you take this image and you color code each pixel according to which direction it responded the strongest, rightward, leftward, upward, downward, you get these four functional layers, which really correspond to the four cardinal directions of motion.
So that was an extremely satisfying result, very, very nice. It rarely happens to you in your career to get such a clean result that the four subtypes of the T4 and T5 cells really correspond to the four directions of motion. So the directionally tuned, and they project according to their preferred direction to one of the four layers.
Keep in mind that this is not going in a circle. But we have opposing directions in adjacent layers. So this is rightward. This is next to leftward, and upward is next to downward. That will come up in a few minutes again. But after getting this result, the next question we wanted to answer is, what is the functional difference between T4 and T5 cells?
And for that, we then had, of course, the idea that maybe they correspond to the on and off channels of motion vision. So we got separate driver lines T4 cells and T5 cells. And we repeated this experiment, but this time with eight different stimuli.
Instead of gratings, which have on and off edges, we used pure edges, off or on, going into the four cardinal directions. What we found was that T4 cells very selectively respond to on edges, and very little to off edges.
And the opposite is true for T5 cells. They responded very strongly to off edges and very little to on edges. So from this, we conclude that T4 cells are motion sensing neurons in on pathway, and T5 since are the motion sensing neurons in the off pathway.
So now when we look at the signals of the tangential cells, then we-- I will focus on two groups of tangential cells, namely HS and VS cells. This is the horizontal system neurons and these are the vertical system neurons. And HS cells depolarize to rightward motion in the right hemisphere, and they hyperpolarize to leftward motion.
VS cells depolarize to downward motion, and they hyperpolarized to upward motion. So now if we take a top view of the lobula plate and compare it with the layers that we've seen before from the T4 and T5 cells, we see that the preferred direction of the tangential cells match to the T4, T5 cells that terminate in exactly the layer where these tangential cells have their dendrites.
HS cells have the dendrites in the layer 1, where the T4, T5 cells ramify that respond to rightward motion. And VS cells sensitive to downward motion have their dendrites in layer 4, where T4, T5 cells terminate that are sensitive to downward motion. There are also cells in between that we did not reconstruct here in this study.
So the question was, how come that these cells then hyperpolarized to the opposite direction of motion? Are we having a separate population of inhibitory T4, T5 cells, or what's going on there? So this is where Alex Mauss comes into play.
And when he joined my lab as a postdoc, already like 10 years ago, he brought optogenetics to the lab. And he expressed [INAUDIBLE] rhodopsin in T4 and T5 cells while recording from the tangential cells, in this case from VS cells. And this is the real experiment. And this is the schematic of the experiment.
And when he optogenetics is stimulated to T4 and T5 cells, he saw the following. T4, T5 cells, when activate T4, T5 cells here at-- for just like 1 millisecond or 2 millisecond, you're seeing the VS very fast depolarization, followed by a delayed inhibition.
So this suggests that T4, T5 cells provide the fast, excitatory signal to these cells, and also activate some interneurons, which then provide the delayed inhibition. So one is synaptic, the other signal is bisynaptic, [INAUDIBLE].
And the question was, which type of interneurons is doing that? And we came up with a search profile for these cells. And, clearly, these signals, these cells should be stratified. They should be postsynaptic in one layer. And they should be presynaptic in the adjacent layer. And, actually, we should have four types of these neurons, one from layer 1 to 2, 2 to 1, and 3 to 4, and 4 to 3.
And so with this search profile, we asked our colleagues at Janelia Farm to look whether in their cell collection they can find such neurons. And [INAUDIBLE] came back with this beautiful stack of neurons, which is shown here. And these neurons were not known before. And we termed them lobula plate interneurons.
And if we look at one of these cells from the side, then compare it with the layers that is formed by the T4 and T5 cells, we see that this lobula plate interneuron 3, 4, really covers layers 3 and 4. Now, in order to see which of these branches is presynaptic, we stained for synaptic [INAUDIBLE].
And then we saw that this neuron was really presynaptic in layer 4. So we presumed that it was postsynaptic in layer 3. If that is the case, then it should respond to upward motion. And so we did calcium imaging. And we found, indeed, that this type of lobula plate interneuron is sensitive to upward motion, which was great.
So the next question was, are these cells inhibitory? So now Alex expressed [INAUDIBLE] rhodopsin in this lobula plate interneurons and recorded from the VS cells. And when he stimulated them up optogenetically, he saw a fast monosynaptic hyperpolarization. So these lobula plate interneurons are, indeed, inhibitory.
And the last question, of course, is are they responsible now for the null direction hyperpolarization of the tangential cells? And for that, he first recorded from the control flies to see nice depolarization for preferred direction and nice hyperpolarization for null-direction motion with everything intact.
And if he now blocks these interneurons, then the preferred direction motion is really untouched. There is no change at all. But the null-direction motion response is completely gone. So from this, we conclude that these lobula plate interneurons really implement the subtraction stage of the [INAUDIBLE] detector, and lead to the motional potency of the tangential cells.
So now having talked about the key players in the circuit, the elements of the neural [INAUDIBLE] in the fly that correspond to the Hassenstein-Reichardt detector, I want to talk about mechanism. And the question is, how do these T4 and T5 cells become directionally selective?
And there were two proposals in the literature. And Tommy has already mentioned them in the introduction. The first proposal was done by Hassenstein and Reichardt in the '50s. And they proposed that a preferred direction enhancement would account for the directional selectivity.
So you take the signal here from one side. You delay it, and then you multiply it with the non-delayed neighboring signal. And as an alternative mechanism, Barlow and Levick proposed that you take the delayed signal from the null side, and you divide the signal by this one. And so this leads then to null direction suppression. And division has been proposed to formally account for that.
So in order to discriminate between these two mechanisms, we did apparent motion experiments. In apparent motion, you stimulated at one place, record the response, you stimulate at the other place, you record the response, and then you give the sequence.
And from the response to the sequence, you subtract the sum of the individual responses. If you get more-- if you get a positive result, then you have a preferred direction enhancement. If you get less, you have null direction suppression.
And so Juergen Haag in the lab used a telescope to stimulate the flies, and placed the stimuli exactly on the facets of the eye. And first measure the receptive field of the cell that he's recording from, and then use these stimuli at these locations here. And he used the driver line which just stains for T4, T5 cells for upward motion.
And then what he saw was the following. When he played this game, apparent motion, between the two bottom columns, then he saw only preferred direction enhancement, and no direction suppression. If he played the game between the two central columns, he saw both null direction suppression and preferred direction enhancement.
And between the two top columns he saw no preferred direction enhancement, only no direction suppression. So depending on the location within the receptive field of the T4, T5 cells, we find both preferred direction enhancement and null direction suppression.
So this led us to propose a new model that combines preferred-direction enhancement and null-direction suppression. And it's shown here. And that was also confirmed by studies from Clendenin's lab at Stanford.
So what is the advantage of having such a dual mechanism? Well, for that, I simply did simulations where you take this signal here, multiply it with this one, and then divide it by this. This is A times B divided by C. And then when you probe such a motion detector with grating motion along all 360 degrees, then plotted in this polar plot here, you see a very tight directional tuning profile.
If you now consider only preferred-direction enhancement, you still see directional tuning, but it's much broader. And when you only have null-direction suppression, you have still a more narrow directional tuning, but significant responses to null direction, which is not seen here.
So the idea is that the combination of both of these mechanisms, underlying direction selectivity, leads to an ideal tight directional tuning of the cells. So what cells are providing these input signals? And thanks to the connectome efforts that Janelia Farm has undertaken in the last 10 years, we know every input neuron to T4 and T5 cells here.
And in the following, I would concentrate on the major inputs, Mi1, Mi4, and Mi9. For the James Bond fans amongst you, and MI6 is not on the list. So we took these driver lines for these cells under the two photon microscope and did stochastic noise imaging, or simulation with them use the calcium indicator, and then applied reverse correlation to calculate the linear receptive fields to that fitted linear filter responses to the results. And this is what is shown here.
So how does the receptive fields of these cells look like in space? Well, they are very small, about six degrees. That corresponds to one facet in the fly's eye. And some of the cells have a antagonistic surround. Others have less of an antagonistic surround.
And as expected, Mi1 and Mi4 are on cells. Because they provide input to on pathway T4 cells. But to our great surprise, we found that Mi9 is an off cell. That was really surprising. Now, how about the temporal properties of these cells?
Well, Mi1 turned out to have band-pass properties. And Mi4 and Mi9, these have step responses, turned out to have low pass properties. So that was very interesting with respect to the motion detecting mechanism.
And so next we try to arrange them accordingly. And, again, thanks to connectomics, we know that a T4 cell which is tuned to rightward motion gets input from Mi9 in the left column, gets input from Mi1 in the central column, and gets input from Mi4 in the rightward column.
And that's very nice, because Mi9 and Mi4 have low-pass properties. And this one has band-pass properties. And just using again this a times b divided by c, this multiplication and the division, we can very nicely reproduce directional tuning, also when we take these cellular properties into account.
But what I really wanted to get at is the biophysical mechanism underlying these operations of multiplication and division. So I want to replace this by an electrical equivalent circuit, and then put the membrane equation in there. It has a first step towards that.
We and others figured out what type of neurotransmitters Mi9, Mi1, and Mi4 have. And what we found was that Mi9 is glutamatergic, Mi1 is cholinergic, and Mi4 GABAergic. After knowing that, we wanted to look at the receptors at the level of T4 cells.
And for that, thanks to rapidly evolving RNA-sequencing technology, we, again other labs, did RNA sequencing in T4 and T5 cells, and we found the following. First of all, T4 cells express nicotinic acetylcholine receptors with different subunits.
Second, they express different GABA A receptor subtypes, namely RDL and LCCH 3. But what was really interesting that the dominant glutamate receptor was GluCl alpha.
And GluCl alpha is a chloride conductance that is created by glutamate, and thus is inhibitory in nature. And it would reverse the sign of the signal delivered by Mi9. And since Mi9 is an off cell, it made really sort of click.
And so next question was, where on the dendrites do we find these receptors? And, again, an electron microscopy study from Janelia Farm published three years ago showed that the different subtypes of T4 cells and T5 cells have the dendrites oriented against their preferred direction.
So a T4a cell, which is responsive to a rightward motion, has the dendritic tips oriented to the left. A T4b cell has the tips oriented to the right, and it responds strongest to leftward motion. And T4c has the tips pointing downward. It's sensitive for upward motion. And the opposite is true for the T4d cell. And at the same thing is true for the T5 cells.
So now two brave people in the lab, Rene and Sandra, they developed a flip tack technique that allows to visualize the different subunits and their dendritic location. And here is the picture that they got for the T4a cells as an example.
And what you see is that the GluCl alpha receptor is really found only at the tips of the T4a cells. And the nicotinic acetylcholine receptor subunits are found in the center. And the GABA A receptors are found towards the base of the dendrite.
And they did this analysis for all the different subtypes, and then align the subtypes according to the intrinsic coordinate system so that it's always from the tip towards the base, and then evaluated it quantitatively.
And they found that all four subtypes have this arrangement, that the GluCl alpha is found primarily at the tip of the dendrite. The cholinergic receptors are found in the center, and the gaba receptor largely at the base. That's a common theme.
So that nicely fits the distribution that Janelia people found in the [INAUDIBLE] for Mi9, Mi1, and Mi4. So now we can come up with the biophysical model, which then incorporates conductances and everything. And this is the model that I came up with a couple of years ago. And this is how it works.
If we now consider a single pulse delivered to the center of this motion detector, of this T4 cell, then we see the following. The input signals are as such that Mi9 in the dark has a high activity. Mi1 is stimulated. If we look at the input resistance, it has a brief dip here when the cholinergic receptors is open, and the membrane potential depolarizes to some degree.
Now, if we move an edge from the left to the right, we see the following. Then the Mi9 signal all of a sudden drops because it's inhibited by this bright edge. And this leads to an input resistance increase. And due to this increase, because the channel closes, the membrane depolarization increases. So we get an amplification, like a multiplicative effect here.
And the opposite is true for the null direction. In this case, Mi4 is activated first. And that leads to a further drop of input resistance. So it's a decrease. And, accordingly, the output signal will decrease in amplitude. So this is how it could work.
And when I came up with that, I of course realized that I wasn't the first one to propose such a mechanism. But let's first look at the directional tuning of such a biophysical model. And here you see again the very tight, nice, tuning of such a dual mechanism, this time a conductance-based model when you have only preferred-direction enhancement. Again, it's broader. And the same is true for only null-direction suppression.
And here's the paper I want to relate my finding to that came out in '78 by Tommy and Vincent Torre. And they proposed a very similar mechanism. They called it an AND/NOT gate to underline null-direction suppression in general.
And that is the mirror image of what I'm proposing here, saying that instead of having a shunting inhibition, to account for null-direction suppression, having the release of from shunting inhibition, to account for preferred-direction enhancement, that is a multiplication.
And so I had a very nice exchange with Tommy when this paper came out. It seems like really an amazing prediction also made-- I don't know-- almost like 40 something years ago.
Well, but these were proposals that are very stimulating for experiments. And so we still needed to find out about how it actually does work. And there is one major obstacle. And this major obstacle was that all the results we had so far were from calcium imaging. But calcium imaging is only a proxy for the membrane voltage.
It is slow. We don't see responses to fast changes in the membrane potential. And we see no hyperpolarization. And we cannot measure any currents. We cannot measure reversal potentials. We cannot measure input resistance. And this is very important for testing this model, as you can imagine.
And so we were very excited to see a few years ago a paper from [INAUDIBLE] group, [INAUDIBLE]. And [INAUDIBLE] published the first recording from the T4 and T5 cells where they patched from the soma of these cells. And that was an incredible achievement, because you have to imagine that the soma of these cells is 2.6 micrometers. It's really, really at the limit of what you can do.
And I was fortunate that at this time a very brave post doc came to the lab, Lukas Groschner. And Lukas is the guy with the golden hands. And he very soon also managed to record from T4 cells, and now does that on a daily basis. And here you see beautiful [INAUDIBLE] of a T4 cells that Lukas did.
And so now we can see the membrane potential of T4 cells, those signals that we always wanted to see for preferred direction of a grating. You see [INAUDIBLE] nice, strong, depolarization, and very little depolarization for null-direction grating. And this is the same for edges moving along the preferred direction. And this is the remaining depolarization for edges moving along the null-direction [INAUDIBLE]. And this is at a different velocity.
Now, Lukas has also trained a very ambitious and talented PhD student Jonatan Malis. And Jonatan Malis said, OK, if Lukas is recording from the T4 cells, then I want to record from all the input neurons using the same stimuli.
And together, they really did it. And within three or more years, they collected an incredible dataset of T4 cell responses and all the different input neurons, using an identical set of stimuli. Of course, they did not simultaneously record from these cells.
And so here you see the signals of T4 cells to gratings moving at various velocities, here to on edges, here to off edges, in preferred null direction. And here you see the signals that Jonatan Malis recorded from all the different input neurons in different colors for the different sets.
And I don't want you to look at all of them. Just pick out and concentrate on the three input neurons I have been talking about so far. So, first of all, what is the receptive field of these cells if we measure them with edges and now in the voltage regime?
And we again find that Mi9 is an off cell, and Mi1 and Mi4 are on cells. And now if we look at the edge responses of these cells, we see the following. We have the Mi9 signal when the edge comes in is suppressed. And Mi4 where the edge comes in is excited. And Mi1 has also an excitation to an on edge.
Now, if you look, if you just think about Mi9 and Mi4 both being inhibitory, you see that the T4 cell is basically inhibited the time. So how can it respond to an on edge? Well, you have to keep in mind that these profiles are all aligned in time to the center of the receptive field.
But the T4 cell is not sampling three signals of these three cells from the same column. It collects the signals of Mi9 from the left column, Mi1 from the center, and Mi4 from the right column.
And now if you align the signals accordingly, you see that the inhibition from Mi9 and Mi4 really leave open a window of opportunity during which the excitation from Mi1 can activate the T4 cell for preferred direction in motion. And for null-direction in motion, it's almost all the time covered, so very little excitation can happen.
So now we can use these signals, and then model the T4 cell using measured input signals. And now these are the data of T4 cells to an on edge moving at 15 degrees per second for preferred direction, for null direction, the way that Lukas measured it.
And when we adjust just the slope and the threshold of the transfer function, and use the input signals that Jonatan measured, we get this model response. It very nicely shows the same behavior. And the same is true for other stimulus conditions, like 30 degrees per second. Again, the model and the data really match very nicely.
So I think this is pretty much where we can get it. But now there's one critical thing to do, namely to test the model. So we think we have an account of multiplication found at the biophysical level in these T4 cell. But we have to test the model and answer a couple of questions.
The first question is, is glutamate really inhibitory in these T4 cells? And here both of these guys did this experiment. They puffed glutamate on the dendrite of T4 cell. And what they see is a hyperpolarization following this glutamate puff.
So glutamate is indeed inhibitory. And if you block the GluCl alpha receptor by RNA I, you don't see any response at all. Cells depolarize slightly, and respond no longer to glutamate.
The next question is, does the glutamate really control the input resistance to these cells? And here you see a series of current injections into a T4 cell under control conditions. And you see strong depolarization and hyperpolarization following the current injection.
And if you puff glutamate, then you see a very strong reduction of this voltage response, saying that glutamate really reduces the input resistance by almost a factor of 100%. So it is indeed controlling the input resistance of these cells.
Next question, when we move an edge, do we see an increase of the input resistance as predicted by this model? And yes, we do. So here you see the passage of an edge. And you see the voltage response of the T4 cell depolarization.
And when you see the input resistance, you see a 30% or 40% increase of input resistance from about 5 almost to 8 gigaohm. So, yes indeed, during visual motion, the input resistance changes dramatically.
And final question is, do we see an effect on the directional tuning of these cells if we take out the GluCl alpha receptor? And this is, again, the direction of tuning for T4 cells under control conditions. If we take it out, we see a very strong broadening of it.
So I would say we really have nailed it down. This is really strong evidence for biophysical account of a multiplication in a neuron. So let me summarize what I've told you.
First of all, as my-- the scientific problem that I picked for many decades is to find the neural elements of the Hassenstein-Reichardt detector in the fly optic lobe. And from this jungle of cells, we found the T4 and the T5 cells to be the primary motion sensing neurons.
And they come in four subtypes, each one tuned to one of the four cardinal directions of motion. And since the first element in the Hassenstein-Reichardt detector is the output of the multiplicator, here the T4, T5 cells should correspond to this multiplication interaction there.
And the lobula plate interneurons then account for the subtraction of opponent motion detectors. And about the mechanisms, we found that the fly T4, T5 cells really combine both mechanisms that have been proposed in the literature, preferred-direction enhancement and null-direction, to come up within one step with the very tight directional tuning.
Anbd for T4 cells, we really found the biophysical basis of this interaction, namely the GluCl alpha receptor underlying the multiplication, and the GABA receptor subunit underlying the inhibition. And using this data, we can now have a model, biophysical model, of the T4 cells, which accounts for all the response properties that we have seen so far.
So for T4 cells, I really would say we have reached the finish line. What remains to be explored is what is the biophysical mechanism underlying the directional selectivity in T5 cells. And if you're interested, I can tell you a little bit. We're working on it, but we haven't solved this problem yet.
With this I want to close, and thank the funding agency, namely-- first of all, the Max Planck Society for supporting this research for so many years, the [INAUDIBLE] and the SFB for additional grants. Then I want to thank my collaborators at Janelia Farm for providing all the tools that we've been using, driver lines which were the essence of our work, and also [INAUDIBLE] indicators from Loren Looger, which really changed the world for us.
And finally and most importantly, I want to thank my collaborators in the lab. It's really an incredible bunch of people with very diverse background from molecular biology, all the way to computer science, and physics, and everything. And it's a great pleasure to work with them. And they have this diverse background, but they all share with me the enthusiasm to figure out how fly neurons compute the direction of motion. Thank you very much for your attention.