Scientist Stories: Gerry Rubin, The Connectome of the Fly Brain

So the first science talk is by Jerry Reuman. He's a senior group leader at Janelia, a member of the National Academy of Sciences, and perhaps best known as Janelia's founding director from 2003 to 2020. He had a bachelor's degree from MIT, PhD from Cambridge, uh, did his postdoc at Stanford and had faculty positions at Harvard, the Carnegie Institute at Berkeley, before moving to become Vice President of Howard Hughes Medical Institute in 2000. Jerry has been a leader in genetics, genomics, and developmental biology of Drosophila, um, and, uh, he, as I said, founded Janelia, uh, developed the design, construction, direction and Janelia opened in 2003. And today, Jerry will share a big project that came out of Janelia. His science talk is called "The Connectome of the Fly Brain: An Example of Big Science." So with that, we'd like to move now to Jerry Rubin. Welcome, Jerry, and we look forward to your talk.

Thank you, Ron. I am going to talk today about the connectome. And when I was a PhD student at the LMB in Cambridge in the early 1970s, work on the worm connectome was just beginning. It took about 15 years to reconstruct the worm's nervous system. The connectome of the worm's 302 neurons was published in 1984 in this 340-page paper. Determining connectomes and sequencing genomes are the two big science biology projects of my generation. I was fortunate to observe and participate in both, and I have been struck by the many features they share. Neither was constrained by the need to have an hypothesis. Neither had broad support at the time they started, as you can see from these quotes about them from key leaders. A key reason behind these objections was the high cost of these projects relative to their expected return of insight. And the fear that they would divert funds better used for other research. Both required extensive collaboration. As such, they only appealed to a minority of scientists, and academic environments were unsuited to support them. Both projects aimed to provide information to accelerate biological research, but depended almost exclusively on the very same non-biological disciplines to achieve the scale-up and cost reduction that led to their success. Instrumentation and applied physics, computer science and software engineering were by far the main contributors, with much smaller contributors from chemistry and chemical engineering and biological disciplines. What eventually made the Genome Project popular was an extraordinary increase in cost efficiency. I'm confident that the same will be true for the Connectomics. I would like to give you a very personal example of what I mean by extraordinary increase in efficiency. I sequenced 156-base RNA as part of my thesis. This was nearly 50 years ago, before sequencing DNA was even possible. This was my first scientific paper. I'm very proud of it, largely because it was my own work, the sequence itself was correct, and the paper has been cited over 300 times. But it had no biology, as you can see from its abstract. When we were sequencing the fly genome 20 years ago, I realized that I could have done my PhD in 100 milliseconds. Progress did not stop, but I stopped keeping track when it got down to nanoseconds. This is a 10 to the 12-fold increase, in case you were calculating. In my talk today, I will describe how the fly connectome is being generated and used. Our goal is to understand at the level of general principles how the brain produces behavior. Admittedly, an ambitious goal that will take the entire field many years to achieve. I believe that such general principles can be discovered in any animal that has a rich enough repertoire of behaviors. It follows directly from that belief that to discover them efficiently, it would be best to use the most experimentally tractable animal that displays suitably complex behaviors. For me, that animal is the fruit fly. Flies display complex motor behavior, as illustrated by this fly crossing a gap. Flies are able to judge their body size and judge the width of the gap so as not to attempt to cross gaps that are too large for them. Flies also display a range of social behaviors illustrated by this example of male-male aggression. A fly's wins and losses affect its future behaviors, and flies can establish dominance hierarchies. Flies can use landmarks to remember the position of a safe spot, a cool spot in an otherwise uncomfortably warm floor. In this video, you can watch as flies learn to associate the visual landscape with the position of the cool spot. The position of the cool spot is shown by the dotted box in the arena. The landmarks rotate when the cool spot moves, so the relationship between the landmarks and the cool spot is always the same. The flies just have to use those landmarks to quickly find the cool spot. The performance for each run will be plotted here. So here you see a bunch of flies running around. They're all tracked, and you can see where the cool spot is, and you can see when they get there, they sort of hang out. Then when they're all happily made it, the cool spot is going to be switched and they all have to find it again, which they do more rapidly the second time around, or at least nearly all the flies do it more rapidly. Then the cool spot moves again. And you can see if you look in the performance window, I'm going to see how fast they're getting there, how their performance improves. This means that the flies are actually learning to associate the landmark with the position or place of the cool spot. The fly can accomplish these tasks with a relatively small brain. This shows a fly brain sitting inside a mouse brain at exactly the same scale. And the mouse brain is shown here with a single neuron that was reconstructed by the MouseLight project team at Janelia shown. This gives you an impression of just how small the fly brain is. The fly has only 100,000 neurons compared to the mouse's 70 million neurons. So small. The small size gives us the advantage of completeness. We can look at the entire brain. We can look at individual cell types and combine different types of data, taking full advantage of the genetic tools. That the fly affords us. So again, we want to understand how behavior is generated by circuits made up of cells whose properties are determined by molecules. How do we propose to do all that? We will start with the connectome, the main subject of today's talk. But we will also need many other types of data. We will need to use our genetic tools to manipulate individual cell types. We will then study the effects of those manipulations on behavior, and we will make physiological measurements. And all this will take place in a tight loop between theory and experiment. A connectome is often confused with an image stack where you can trace individual neurons. Such an image stack is not a connectome, but rather a necessary intermediate step. A connectome is a complete graph showing all the connections between between all the neurons. I offer you an operational definition. You know you have a connectome when you can sit in a coffee shop with your laptop and in the time it takes you to drink a cup of coffee get answers to questions like these. And there should also be seamless ways to go from the graph both to the original grayscale EM data as well as the reconstructed 3D morphologies of the neurons. No, this is not a distant fantasy, at least for the fly. The fly team at Janelia— the fly EM team at Janelia released such connectome for roughly one half of the adult fly brain in January. I will first tell you how the connectome was generated, and then I will share some examples of what we are learning from it. The first task was collecting the image data. Harold Hess and Chang Su, adapted the FIB-SEM microscope, normally a tool for metallurgy, to biology, which is a tour de force of applied physics and instrumentation. Sample preparation was continuously improved over several years by Zhiwang. FIB-SEM allows you to image with isotropic voxels, that is, voxels that have the same dimension in each of the 3 dimensions. This is a great help with image processing. Here's a movie going through a Z-series of a volume collected this way from a project we completed 4 years ago. To give you a feeling of what the data actually looks like. Here's one slice through that image volume. Our goal here was not to trace individual neurons, but to reconstruct all the neurons simultaneously to generate an image like this, where all the processes have been assigned to an identified cell type. We have been steadily improving the rate at which we can generate connectomes. But more progress was needed. Between 2017 and 2019, we achieved an additional 50-fold improvement, putting us in a position to tackle the full fly connectome. A large part of this was achieved through a collaboration with Google. It's important to keep in mind that this was a large collaborative project, big science, involving many different areas of expertise. Steve Plaza was a project manager who had to coordinate many activities, starting with sample prep to imaging to aligning the dataset to segmentation to algorithms to synapse prediction tools, proofreading biologists for doing the proofreading and analyzing the results, supported by a lot of infrastructure. It required patient funding, and it was not inexpensive, even without putting the dollar figure on Google's contribution. So this was big science, both in the amount of time, the amount of people, and the budget. For imaging, the brain needed to be sliced like a loaf of bread into 20-micron sections, which could then be imaged in parallel using a method developed by Ken Hayworth. Our first large-scale target was the red portion of the adult fly brain, which we call the hemi-brain because it's about half the total brain. The image slabs that were imaged were then computationally aligned and assembled into a single volume, which you can see here in this movie. You can see the lines between the original slabs, and you can see how they're seamlessly joined together, nearly. And you can see the quality of the data. Of this image set as we zoom in. And we'll end up with looking at a single synapse crossing the border that will be in the middle of your screen right there. Okay. So now we have the images. We have to define the neurons. And this required image segmentation. And improvements in image segmentation were key for us. We have been using machine vision to segment images, that is, to identify the borders of individual cells. For some time. The computational results are not perfect, and the segmented images require human proofreading, something you're going to hear about from Ruchi. Extrapolating from the small volume we had completed led, unfortunately, to estimates that it would take 5,000 person-years to do the entire fly brain, to go from the segmented image volume to the full connectome. So we're very fortunate to establish a collaboration with Google on image segmentation. The resulting algorithmic improvements and computing power that they provided allowed us to reduce our estimate from 5,000 to 200 person-years for proofreading the full brain. A manageable number. The result of segmentation looks like this in 2D, with individual neurons having different colors from their neighbors. Synapses are also identified by machine vision. In a separate process. There are nearly 9 million presynaptic terminals in the— in the hemi-brain, so machine vision is essential for predicting the synapses as well. Segmentation plus synapse prediction does not generate a finished connectome. Remember those 200 person-years of proofreading we estimated for the full brain? That would correspond to 100 person-years for the hemi-brain, which is about what it took. A team of 50 proofreaders worked on this project for the better part of 2 years under the direction of Ruchi and Pat, and you will soon hear from Ruchi what it takes to recruit, train, and manage a team of that size. The Connectome of the hemi-brain was released a few months ago with no restrictions on its use, along with analysis tools to make the data accessible. It is being heavily used. Analysis tools are necessary if the Connectome is to be broadly accessible. Producing these tools required a significant software engineering effort. I am going to show you such a tool, NuPrint, that was released in January. As illustrated by the video I'm going to show you, which is but one of many instructional videos.

We can use "Find Neurons" to pull up neurons of the type mBon19 and FB06a, and with this information, we can look for the shortest pathways between a single mBon and a single FB neuron. Here, we can see that although there are no direct connections between our two neurons, There are 7 one-hop interneuron pathways that connect the two. We can take a closer look at one of these interneurons with Find Neurons. We can check the morphology of this interneuron. And we can pull up the synapses that connect this neuron to another neuron. Here, we can see the inputs from our mBond-19 in green and the outputs to our FB neuron in purple. We can also view EM data using Neuroglancer. The segmentation of the selected neuron is shown in purple.

I would like to use the remainder of my talk to give you an example of how we're using the connectome. I will discuss a part of the fly's brain called, because of its shape, the mushroom body. The mushroom body, or MB as I will refer to it, is the main site of learning and memory in the fly's brain. Here's an example of how we assess memory. You may be wondering, how do you know if the fly learned anything?. In this experiment, flies are first exposed to an odor, Odor A, with an aversive stimulus, so they hopefully will learn that Odor A is not a good thing for them. Then they're exposed to Odor B, a neutral odor, and then the flies are given the chance to choose between those two odors in an arena with two quadrants each with with the odors. And this is— you can see this in this movie. The flies are running around in this quadrant where the two odors are present, and there's two odor A quadrants and two odor B quadrants. And what you're going to see is that the flies leave the odor A quadrant. You watch this purpley track fly here, and it thinks about going back into the odor A quadrant again. It doesn't like that. It avoids it. And here you're going to see some flies that are in the B, the safe odor, and they stick their nose into the aversive odor quadrant, they turn around. So you can see the flies have learned that odor A is bad and odor A is good from this movie. They've learned the association and can remember it, at least for a few minutes or hours actually in this case. This slide shows a cartoon of the olfactory learning circuit in Drosophila. Odors are sensed by olfactory receptor neurons in the fly's equivalent of a nose. Neurons with the same receptors converge on common downstream neurons in the fly's brain, just as they do in your own brain in your olfactory bulb. From there, projection neurons go from these convergent areas to the dendrites of the input area of the 2,000 Kenyon cells. Unusual cells that have long axons. And their long axons make up the mushroom body lobes. This part of the fly brain plays a similar role to your piriform cortex. An odor generally activates a few percentage of the Kenyon cells. So odor identity is encoded by which Kenyon cells are activated. In the prevailing model of olfactory learning, the lobes of the mushroom body are divided into zones.— that are each innervated by a dopaminergic neuron and an output neuron. During training, like I showed you in the movie, synaptic plasticity occurs at synapses where there's a coincidence of Kenyon cell firing, representing the odor, and dopamine release, representing, in this case, a punishment. This change in strength of the synapses between the Kenyon cells that represent a particular odor, and the output neuron is the physical manifestation of the memory. It's the change in the brain that records the memory. This in turn decreases signaling from the output neuron from that compartment. Which Kenyon cells were active determines the identity of the stimulus, identity of the odor. Which compartment receives the dopamine input determines the valence, whether it's positive or negative, of the stability of the memory. It turns out that dopamine neurons innervating the different compartments promoting memory formation are very different, and they promote different learning rules. For example, memory can be short-term or long-term. It can be positive, appetitive, or negative, aversive. Dopamine neurons in vertebrates like yourselves are also similarly heterogeneous in their activities. The mushroom body circuit architecture, with a series of separate zones of synaptic plasticity along the kenyon cell axons, may at first seem very peculiar to you, not like the cartoon neurons you've seen in textbooks. But I remind you that cerebellar granule cells, which are estimated to make up 3/4 of the neurons in your brain, have a similar architecture. Granule cells sparsely encode information like Kenyon cells. Mossy fibers provide sensory input. Purkinje cells are the output neurons. Climbing neurons provide teaching signals. The climbing fibers and the Purkinje cells define zones along the granule cells. Analogous to the mushroom body compartments. The synapses between the mossy fibers and the granule cells have a distinct morphology with claw-like dendrites surrounding the input axon. A very similar morphology is seen in the olfactory inputs to the Kenyon cells, which I can show you in the movie, which instead of a cartoon is actually a 3D reconstructed neuron from the Henry Brain dataset. We're going to show you a single Kenyon cell.

A single gamma Kenyon cell with its dendrites in the main calyx and its output synapses in the gamma lobe is shown. Postsynaptic sites are shown as white dots and presynaptic sites as yellow dots. The axonal terminals of an olfactory projection neuron are shown in green with its presynaptic sites as yellow dots. Note how this claw-like dendrite of the Kenyon cell wraps around a presynaptic bouton of the olfactory projection neuron. Projection neurons have an average of 5 such boutons and make more than 20 synapses onto the downstream KC in each bouton. The insert shows the original EM data that has been colorized so you can identify the processes of the dendrites.

Now I'm going to show you the second video, which illustrates the compartments that are formed by dopaminergic and output neurons.

The dendrites of each of the two mBond-14s define the alpha-3 compartment, as each of their arbors densely fill the compartment. Similarly, each of the axonal arbors of the two PPL1-06 dopaminergic neurons fill the same compartment. Here, we show the arbors of the two mBonds and two DANs together. Note how they define the same space: the α3 compartment. Now we turn to the α1 compartment. This compartment is filled by each of the dendritic trees of the two mBON7s. There are 7 PAM11 dopaminergic neurons. Unlike the PPL dopaminergic neurons found in α1, the axonal arbors of the individual PAM11 neurons are occupy only a portion of the alpha-3 compartment. Collectively, however, they fill the compartment. Here, the 2 M bonds and 7 dopaminergic neurons are shown together.

To analyze the mushroom body circuit, we have brought together a collaborative team of experimentalists and theorists. Our goal is to gain insights in the circuit architecture and function. In the last few minutes of my talk, I will share some of the highlights of that work. We discovered a new class of mushroom body output neurons consisting of 13 cell types, one of which is shown here.

The atypical mushroom body output neuron, mBON24, receives strong input from Kenyon cells in the beta-2 compartment of the mushroom body lobes. It also receives about 25% of its synaptic input outside the mushroom body So what makes this mushroom—

this new type of mushroom body output neuron unusual is it has inputs both inside the mushroom body lobes and outside the mushroom body lobes, in comparison to the classically described mushroom body output neurons whose total input, dendritic input, is inside the mushroom body lobes. We also found that the amount of visual as opposed to olfactory sensory information to the mushroom body was much greater than previously realized. And perhaps more importantly, this information was segregated into two distinct types of Kenyon cells, one of which is shown here. All sensory information appears to be visual. And here is the second type with exclusively visual input. So these showed that nearly 10% of the input to the mushroom body was visual input. And these visual inputs were in distinct subsets of Kenyon cells. And what that allows you to do is form memories of visual information independently of olfactory memories or in combination. I told you that dopamine neurons provide different types of teaching signals. Dopaminergic neurons learn about the fly's internal state and its external environment from their inputs. So for the first time, the connectome allowed us to look comprehensively at the neurons that provide input to each of the 157 dopaminergic neurons that innervate the mushroom body. In this matrix, dopaminergic neurons are clustered by the similarity of the inputs they receive from upstream neurons. Then the matrix has been sorted to group neurons that provide input to each of the mushroom body compartments. Given the number of neurons and synapses, this could only be done computationally using the database provided by the Connectome. Let's zoom in and look at one part of this matrix in more detail. Notice how even within this single compartment there are two clear classes of dopaminergic neurons. The darker the color of the individual squares in the matrix, the more similar the inputs are to the two dopaminergic neurons that overlap in that square. I'll explain the colored squares in the diagonal in a moment. The dopamine neurons to that compartment share very few inputs to the dopamine neurons to this neighboring compartment, but they do share inputs to another compartment far away. The colored squares in the diagonal show which dopaminergic neurons get input from a mushroom body output neuron. In some cases, the input is from the same compartment. Forming a simple feedback loop where the output neuron is going back and informing the dopamine neuron, which is sending its signal. In other cases, such inputs provide a mechanism for cross-compartment communication as well. We also see cases where the input is exclusively from a different compartment or where there's no input from an M-Bond at all. The neurons that are upstream of the dopamine neurons can be grouped into morphological classes. 35 of which are shown on the columns of the matrix, and the rows of the matrix represent different morphological subclasses of dopaminergic neurons. Class 3 neurons, for example, upstream of many different dopaminergic neurons, as are those in Class 7. However, there are 5 classes that are heavily biased to almost exclusively input to the dopamine neurons going to the alpha-1 compartment that I showed you in the movie, such as Class 18 and Class 22. However, these classes, although dedicated to these same dopamine neurons, provide very different classes— types of information. Class 18 provides information from a dorsal brain area called the SMP. Class 22 neurons provide information from the SEZ and thus are likely to relay information about what the fly is tasting. Now I'd like to turn to my final point, which has to do with how the fly brain integrates information, often conflicting information, from the many types of mushroom body output neurons. A few minutes ago, I showed you a movie of flies learning an odor preference. How does a population of mushroom body neurons drive the learned behavior, make the flies go to the right place? Optogenetic activation of individual output neurons has been shown to be either attractive or repulsive depending on which MBOB type. In this model, in untrained flies, the activity levels of neurons for positive and negative valence are balanced, so they do not alter the odor preference. However, after forming the aversive memory, we've turned down one of these output neurons.. And that upsets the balance. And in this case, since we turned down a positively appetitive neuron, it shifts the balance to avoidance. This is what you saw actually happening in that movie, we believe. This model predicts that different M-Bond neurons must converge their outputs onto shared balancing neurons that can assess and drive the appropriate behavior. And we found this was indeed the case. The connectome allowed us to look at all the outputs of all the and mushroom body output neurons. These outputs are delivered over a large part of the dorsal brain, as you can see in this slide. Here's a single neuron. The neurons also span a large area. This is a single neuron. And it's the champion for convergence, as it receives input from 12 different types of output neurons. Some of them use excitatory transmitters, other inhibitory transmitters. Exactly the same kind of push-pull arrangement you would want to have to weigh the balance of outputs. And the connectome also allows us to look at great— in great detail, like where the synapses are. And we can see the synapses from different transmitters onto this neuron are clustered. And you can see this in these two different views. Okay. So I've only had time to give you a glimpse of the kind of information we're getting from the connectome. We're only beginning this exploration. And our map is incomplete, missing information about gap junctions, neuromodulators, other physiological properties of individual cells. This information will gradually get filled in, and our map will get better. And a good analogy is the early geographical maps. Here's one of my favorite early maps. It was made by Moll, one of the great mapmakers of his time, 300 years ago. California is drawn as an island based on limited information he had from 17th century sailors who guessed incorrectly that the Gulf of California and Puget Sound were connected bodies of water. Fortunately, there are more than 50 labs around the world that work on the flying mushroom body, so I'm optimistic that, aided by the connectome and an anticipated influx of theorists, that our understanding will rapidly improve. And I believe that the things that the scientists will be able to learn will demonstrate the power of connectomes to provide insight, just as the publication of the yeast genome sequence in 1996 demonstrated the power of genome sequencing.

Thank you. We're going to take questions, but in the interest of time, I am going to read just a couple of the top 2 questions. I was hoping to allow the audience to ask their questions, questions, but we're running a little bit late, which we will tweak for the future. But the first question is from Nicholas Perez. Have you tried to perform some of the connectome analysis after training the animals to compare how the circuits change after learning?

That's something that we'd obviously like to do. It's a bit laborious to do that by electron microscopy, but since we can target individual and label individual cell types, I think that's best done by light microscopy, and labs are starting to do that, like in the mushroom body circuit, look at the number of synapses and the size of synapses after learning.

Yeah, great, thank you. Also, a question from Royan is, do you think it would be important to compare the principles gleaned from a vertebrate connectome to those from the fly connectome to see if there are general neural principles versus fly or invertebrate-specific principles?

I think that would be great. I tried in my talk to point out how analogous the olfactory system was in its organization and the similarities between the mushroom body and the cerebellum and architecture and how they appear to work. So I, I do believe in evolution, so therefore I believe that— in fact, I would say if there are general principles, they'll be true of the fly and human. There'll be human-specific principles and fly-specific principles as well, but I believe there'll be general principles that are conserved. All the nuts and bolts, all the transmitters, neuropeptides, ion channels are all conserved, so I think the general principles almost by definition will be conserved. And I'm hoping, you know, there's a lot of skepticism about connectomics, and I'm hoping people see what we've been able to learn in the fly that we that we totally had no idea about before the connectome. I mean, the connectome of the fly clearly taught us new things that will make people want to do vertebrate connectomes. And in fact, there is an emerging effort to campaign to do the mouse connectome, something that's 10,000 times bigger than the fly connectome.

Exciting.

So one other question. To what extent can you identify the type of synapse, for example, a stimulatory, inhibitory, or the type of neurotransmitter from the EM data? And what technology do you think is needed to gather this information and then overlay it on top of the connectome?

So, you know, if you asked me that question 6 months ago, I would have said we can't tell that from the EM. But actually, there's a nice study done by one of my Janelia colleagues, Jan Foucault, recently posted in Bioarchive, where they're using machine learning, being able to tell the transmitters of neurons with apparently pretty high accuracy from EM data like this. So that's optimistic. And of course, the other approaches we can relate— we have great genetic tools that allow us to make drivers specific for cell types, and we can determine those transmitters by antibody staining or RNA profiling. And then we can leverage that data back on the connectome. The good thing about that is we only have to do it once on one connectome. Doesn't matter what brain, because No one thinks that in general the transmitters of a given cell type change or are different from individuals. Excellent.

Great.

Wiring may be different, but not transmitters. Well, Jerry, thank you so much for your talk and for addressing some questions. Now I think we are going to move on to our second speaker.