GF2045(2013) Proceedings: Dr. Ed Boyden
Lincoln Center, New York City, June 2013
Global Future 2045: Towards a New Strategy for Human Evolution, Congress Proceedings
Tools for Analyzing and Engineering the Brain
Dr. Ed Boyden
Associate Professor of Biological Engineering and Brain and Cognitive Sciences at MIT and Co-Director, MIT Center for Neurobiological Engineering
Abstract: The brain is a complex, densely wired circuit made out of heterogeneous cells, which vary in their shapes, molecular composition, and patterns of connectivity. In order to help discover how neural circuits implement brain functions, and how these computations go awry in brain disorders, we invent technologies to enable the scalable, systematic observation and control of biological structures and processes in the living brain. We have developed genetically-encoded reagents that, when expressed in specific neuron types in the nervous system, enable their electrical activities to be precisely driven or silenced in response to millisecond timescale pulses of light. I will give an overview of these “optogenetic” tools, adapted from natural photosensory and photosynthetic proteins, and discuss new tools we are developing, including molecules with novel color sensitivities and other unique capabilities. Often working in interdisciplinary collaborations, we have developed microfabricated hardware to enable complex and distributed neural circuits to be controlled and observed in a fully 3-D fashion, as well as robots that can automatically record neurons intracellularly and integratively in live brain. These tools are in widespread use to enable systematic analysis of neural circuit functions, are also opening up new frontiers on the understanding and treatment of brain disorders, and may serve as components of new platforms for diagnosing and repairing the brain.
Transcript
So, I direct a neuroengineering group at the MIT media lab, and our goal is to try to reveal new inventions that allow us to make progress in understanding and engineering and repairing the circuits of the brain. Now, the last century of neuroscience has taught us that the brain is an incredibly complex circuit, made out of hundreds of billions of cells called neurons, and this is evidenced through all sorts of studies - animal model studies; studies of human patients. These neurons are incredible computing machines. Each of these cells is connected to some tens of thousands of others, perhaps, and they compute using millisecond-timescale electrical impulses. But it’s also the case that these neurons have all sorts of modalities of communication and computation that we’re just learning about. For example, neurons are known to release gaseous messages that allow communication across volumes, for example, not just through their connections.
So we’re kind of in this interesting era: We’re trying to build technologies through fundamental principles of brain operation, while we’re also trying to figure out how to use these principles in order to repair brain circuitry, and even - as I’ll show you at the end - to build brains from scratch in a dish. Now, one thing we know about the brain is that it’s incredibly dense and heterogeneous. So, within a cubic millimeter of brain tissue, you have perhaps a hundred thousand cells, connected by something like a billion connections. Each of those connections can also be active at the millisecond timescale, so if you think about it, that means that a cubic millimeter of brain tissue could, in theory, be capable of a trillion different distinct events per second (although it probably has even more subtle things going on within the cells).
But, even not knowing how many kinds of cells there are in the brain - this is sort of how little we know about what’s going on inside of our heads - we know that there are some cell classes that are distinct. For example, these small star-shaped cells, known as basket cells, are amongst the cells that are compromised in patients with diseases such as schizophrenia. We also know a lot about long-range excitatory projection neurons that will send information even throughout large fractions of the brain, and these are amongst the cells that might be overactive in patients who have epilepsy, for example. So, we need to figure out how to build technologies that allow us to answer fundamental questions, like: How many kinds of cells there are? How are they connected? What molecules are at those connections? And also, how do we repair these brain circuits - both at the anatomical and the molecular level? Can we actually form connections and even build parts of the brain if we need to? But - also at the dynamical level - can we enter information into the brain in order to re-sculpt the computations that are happening?
So, what we’re trying to do today is to talk about technologies that we’re building in order to enable this kind of approach to happen. We build tools that we then disseminate to the community, and over a thousand groups around the world are using these tools from our group, in order to analyze, engineer, and construct brain circuits. But our main goal is to try and figure out how to ‘push’ the technology in new directions. Now, this is important not only for philosophical reasons - we want to understand how the brain computes - but also, as you’ve heard, to address fundamental medical needs about the brain.
The brain is not only important for our lifespan, and our ability to move, and other things like that, but it is highly important in generating our identity, our feelings, and our consciousness. And so, that’s why these disorders of the brain (pretty much none of which can be completely cured, and very few of which can be treated) are so devastating. Now, of course, there’s been some success - you’ve heard a little bit about neuroprosthetics and the use of electricity, for example, to control brain circuits, and pharmaceuticals have had an impact over the last hundred years, in treating brain disorders. But one could argue that even a well-trodden field, like brain pharmaceuticals, is an area in need of great innovation.
These are some numbers from Science magazine a couple of years ago, where they point out that brain drugs take about nine years to be developed and approved; they fail, about 92% of the time, to get clinical approval; and the cost per drug is enormous - about $850 M each. So, if we think about it, a drug that just bathes the brain in a chemical might affect circuits in the brain that you want to fix, as well as circuits in the brain that you’d rather leave untouched - and that, if the drug does perturb, will result in side effects. And so, part of the goal, I think, is not only to think of neuroprosthetics in ways of controlling brain circuits, but even (when we’re thinking about pharmaceuticals) - are there ways for us to figure out how to make them target circuits in the brain?
I’ll tell you about a couple of projects we’re working on. The first one is a project that we’re working on with Craig Forest’s group at Georgia Tech, in which we’re trying to find a way to analyze cells in the brain. As I mentioned, we don’t even have a complete list of the cells types in the brain - a ‘parts list,’ if you will, of the basic circuit building blocks. Now, the brain doesn’t feel pain, so we can actually - in the living brain - introduce robots in the brain that can help us figure out what are the integrative properties of cells in the brain. If we can record the electrical pulses in a cell, and we can also observe the molecular contents, and look at the morphology, that would tell us something important about how that cell works.
So what you see here is a robot that we built, where we introduce a very small glass micro-needle into the brain, and what we’re trying to do is to detect a cell and then stop when we get to that cell, so we can do automated neurorecordings; automated surveillance of the molecular contents of cells; and also injection of dyes into the cells, so that we can see the shape. The way we do this is, we deliver little electric pulses into the micro-needle, so that we can do a sort of impedance survey of the environment. The closer we get to the cell, the less current will flow through the micro-pipette - you can sort of think of it as getting ‘blockaded’ by its environment. By doing a time series analysis, of the electrical impedance of this needle tip over time, we can actually predict when we’re going to hit a cell, with high accuracy - perhaps 80% nowadays. Then, once we’ve detected a cell, we can apply a bit of suction - this is called the patch clamp method, which we did not invent, but which we‘ve been able to automate for use in the living brain - and when we apply suction, we can have an intimate interface between our micro-needle and the cell. That’s exactly what we want to do if we want to get intricate and integrated surveillance of what’s going on inside the cell. So, if we apply a little bit more suction, we can get access to the inside, and now we have an electrical connection to the inside of the cell, so that, for example, if that cell fires an action potential - one of its electrical impulses - we can observe it. We have an amplifier that’s connected to this micro-needle. Similarly, we can inject a dye into the cell to get its shape, and we can harvest the contents by sucking it out.
Now we’re looking at a variety of other strategies. Can we scale up these robotic arrays, to be able to do the analysis of many cells in a circuit? Can we do this in the awake brain, for example - so this could be used as a neurosurgical tool, to give synapse - level information about what’s happening in cells in the brain, to help people who are undergoing neurosurgery for various conditions? By being integrative, we can connect the dots. We can associate electrical, circuit, and molecular properties, all in the same cell, and this is something that we’re working on, as we’re thinking about how to make nanorobotic strategies for doing this as well, so this can scale up to significant fractions of the brain.
Now, these procedures are destructive to the cell. If we harvest the contents, the cell no longer has its recipe to make it do what it does. So, we also have been working on extracellular methods, to tile the brain with electrodes, so we can record cells in circuits distributed throughout the entire brain. What you see here are some microfabricated devices we’re building with the Fonstad group at MIT, and at the top you can see a three-dimensional array, that has prongs that can span a significant fraction of a mammalian brain, such as a mouse model that’s used commonly in neuroscience. If you zoom in, as you see at the bottom here, we’re packing electrodes in incredibly densely, in the sense that we’re sort of tiling these electrodes. In some ways, this is almost like an image of the brain - we’re imaging the electrical activity by picking it up, in a 3D pattern, throughout the entire brain circuit. That’s a very powerful way to collect information off the brain, because if you can pick up neural activity in regions that are mediating sensations, emotions, decisions, and actions, you might be able to connect the dots, to explain how these different cells are working together, to integrate information towards goals, or decisions, or complex behaviors.
Now, we also want to be able to fix the brain. And to do that, we have to take a step back and think about how neurons compute. As we’ve talked about earlier, there are all sorts of messengers, even gaseous ones, and different ions, and kinases, and all sorts of things that mediate neural computations. But a very useful abstraction has been to think about neurons computing using electrical pulses. So, if neurons are computing using electrical pulses - and if we are able to find what are essentially the molecular equivalent of solar panels, that convert light into electricity, so that we can install them into neurons - then shining light into those cells should allow us to turn those neurons on or off, depending upon the way that those electrical charges move. And as I mentioned, the brain doesn’t feel pain, so we can implant prosthetics in the brain, including light-emitting prosthetics, like an optical fiber that you can put into the brain, and now you can deliver light to a circuit to turn it on or off.
Over the last forty years, many groups have been able to find molecules that will convert light into electrical energy, and a couple of years ago, the Hegemann-Nagel-Bamberg groups, working together in Germany, were actually able to pinpoint a molecule that’s used in unicellular green algae, to help it swim towards light. This molecule stems from a little eyespot of the green algae that help the flagella turn in an optimal fashion. If you zoom into the eyespot, you can see these little proteins respond to light, because if you were to shine light on them and record the electrical currents, you can actually observe electrical currents moving through these little pores. If you think about it, that’s exactly what we want: a molecule that, when we shine light on it, will respond by changing the electrical gradient across the membrane. Now we’re lucky - it turns out that this molecule is a protein that’s encoded for by a single gene; and that gene is small enough that you can actually put it into other cells, using techniques from the field of gene therapy.
So, we take this snippet of DNA, from this algae, and put it into a vector, and deliver it to neurons. This is a collaboration that Nagel, Bamberg, Diesseroth and I did in 2004. Now, this is where serendipity kicked in. In the algae, this molecule is expressed in a small eyespot, inside the cell. But when neurons receive this gene, interestingly enough, they were able to manufacture the protein, and install the proteins on the boundary - on the outer membrane - of the cell. So that was one piece of luck. A second piece of luck was that the currents that were mediated by this protein were just the right timescale and amplitude that when we shine light on a neuron, we can turn it on - we would be able to mediate those fast electrical impulses. So I always tell students when they start out: the most important thing is to be very lucky, when you’re doing a project. So, this is great, because we can aim light at a cell, and unlike electricity, which will go in all directions, activating not just neurons you care about, but all sorts of other cells (that might be just the kinds of cells we don’t want to affect), we can now manipulate neural activity very precisely.
So, how does this work in practice? Let’s take that basket cell - the one that’s compromised in patients with schizophrenia, that we talked about earlier. It turns out, we can take a little piece of regulatory DNA, put it in front of your light-activated protein or opsin, as they’re known, and then we deliver it to the brain - only these cells have the factors that turn on this piece of regulatory DNA and make the molecule. So you can actually target just these cells and not their neighbors, even though they’re all densely packed together, as only these neurons are light-activated. Now, when you bathe this circuit in light, these cells will be active, while their neighbors will not be directly affected by light. So imagine that we can take cells that atrophy, and if we can augment their performance in some way, maybe we can repair computations directly. Or, for many cells in the brain, we don’t know what those cell types necessarily do, or if we have hypotheses, it’s hard to prove that those cells do something specific.
So, I’ll tell you about one concrete example here, that’s a basic science point, but a useful one, which is to try to understand which brain cells drive the processes of the brain that mediate sensations such as reward, or pleasure, or addiction. A lot of us have heard of dopamine, and I think that in the media, it’s often colloquially referred to as a ‘pleasure center of the brain,’ but it’s an open question what exact patterns in these neurons will drive a sensation that will reinforce whatever the brain was just doing. So we collaborate with the Fiorillo group, at KAIST, in Korea, and we’re able to label, through these acronyms at the top, just this set of cells - the dopamine producing cells, deep in the brainstem - a tiny cluster, and we are able to put the molecule from algae, called channelrhodopsin-2, just into those neurons, so when we shine blue light on them, they will be activated. The experiment, then, was to do a very simple behavior: to take an animal model like a mouse, whose neurons (that make dopamine) have been made light-sensitive, and to shine light on those neurons whenever the mouse goes to a certain point in this box. So when the mouse goes to the blue dot, it will get a pulse of blue light directly down an optical fiber that goes into its brain, and there’s a ‘control’ point - that black dot, where if the animal goes there, nothing will happen. So this is what happens: the mouse is going into its little righthand portal, and every time it pokes its nose into a little sensor, it gets a pulse of blue light delivered into its brainstem, and as you can see, the mouse pokes its nose, gets a pulse of light, and what the pulse of light’s doing is activating those cells and making the brain do more of whatever the brain was just doing. And the mouse doesn’t go to the other site, which has no associated reward contingency. So this allows you to prove that one-fifths-of-a-second-duration light pulses - just one pulse at a time - are enough to make the brain do more of what it was just doing. And that’s very important if you want to understand the basis of sensations like reward and pleasure, but also things like compulsions, or situations like addiction, where these circuits might be hijacked.
Now, in the years since, we’ve been branching out, and looking all over the tree of life for molecules that can do this, and it turns out that pretty much every kingdom of life, except for animals, have some molecule that will convert light into direct electrical gradients. So, for example, we’re now collaborating with this Thousand Plank project that’s headed by G. Wong and involves folks like Michael M. and others. We’ve been able to look at, literally, thousands of new molecules that convert light into various kinds of signals. And so, one could argue that, analogous to synthetic biology, we’re sort of seeing a synthetic physiology era, where instead of just controlling genes, we can now control real-time dynamic processes, in living cells. And we devise all sorts of high throughout optical equipment, to screen for these molecules, because they’re sometimes very hard to find.
Let’s describe one of these molecules that is of great use and is in use by hundreds of groups around the world, which is sort of the opposite of the molecule we looked at before (the channelrhodopsin-2 molecule that I told you about earlier - when you shine light on it, it brings a positive charge into the cell). Several years ago, we found a molecule that brings a negative charge into the cell, that can also be put into neurons, and is successfully expressed. When you shine light on neurons containing this molecule, it will now allow those neurons to be shut down with light. So you can imagine doing all sorts of different things with this molecule - for example, could you turn off neurons that are involved with epilepsy, to shut down a seizure, right when it occurs?
Actually, many groups are now engaged with such research activities. Here’s an example from Akihiro Yamanaka’s group, where they again label a tiny cluster of cells, but this time they made them silence-able with light, not activate-able by light. And they labeled the light-sensitive cells in the hypothalamus, known as the hypocretin or orexin cells, and these are cells that are compromised in patients with narcolepsy. So, the question remains: if these cells are gone in a patient with narcolepsy, does that lead to a sleepy state, or is the loss of these cells going to compromise other brain circuits that are going to adapt over periods of months or years, and those losses are going to cause the chronic sleepiness and other changes that characterise narcolepsy? So what Akihiro’s group did was to make a transgenic animal, with just this small cluster of cells, deep in the brain, made sensitive to light. And then they put in a laser, connected to a fiber (this time a yellow laser), so that you can turn these neurons off. And what you can see here is, this is the probability of being awake over time. When they turn the yellow light on, the animals fall asleep within some tens of seconds, and when you turn the light off, they wake back up (although if you look carefully, you can see that they’re a little bit groggy still, toward the end). So this allows you to prove that these cells, once they’re deactivated, can result instantaneously in a sleep condition.
Now, we have exquisite control over cells - and you’ve already heard about some of the things that can be done with electrical stimulation - but now we can turn on or off specific cells embedded within the dense matrix of the brain, and one very exciting area is to think about clinical applications. And, you know, that might seem a bit far out in the beginning, but if you think about it, over a quarter of a million people have some kind of neural implant already - cochlear implants or deep brain stimulators for Parkinson’s disease. We have a molecule that has to be genetically expressed in neurons, but gene therapies are also seeing a resurgence of interest (thanks to new viruses like the adeno-associated virus), and in fact, a gene therapy was recently approved in Europe. And recently, we’ve collaborated with groups like Rob Desimone and Graybiel to do non-human primate testing of these molecules to see how they’re tolerated by the brain, because these molecules come from plants and bacteria and so on, and so we want to know if the immune system, or cells, will regard these as foreign. And so the initial data is a small dataset, but more groups are now replicating them, and so we’re very excited to see how this avenue of pre-clinical testing goes in the years to come.
Let’s consider one concrete clinical application area, which is in the field of blindness. The retina, at the back of the eye, contains photoreceptors that capture light and transform those light signals into signals that can, through a layer of neurons, be transmitted to the brain. Many patients have lost photoreceptors - perhaps millions around the world - and can no longer do that initial light capture step. Here, we are zooming into that retina at the back of the eye, and those patients who have lost their photoreceptors will be lacking an entire row of cells. But there are still spared cells in the eye, and furthermore, light can still get into the eye. So what if we can take the molecules that convert light into electrical energy, and install them in one or another layer of cells that are spared? Maybe we could convert the rest of the eye into a high resolution camera. This is something that we’ve been pursuing, in projects led by Alan Horsager, and also being supported by a startup that I’m involved with, called Eos Neuroscience. Here’s a blind mouse that’s trying to solve a maze. In this maze, there are six arms, and there’s some water in the maze because mice like to swim, but they like to sit on platforms more. So this mouse is swimming around and it’s trying to get to a platform that’s at the top of this image. Now this mouse is blind. It turns out that many of the mice you find, for example at pet stores, have mutations that cause their photoreceptors to be degenerate. And so, mice are smart, they can solve the maze by systematically searching each of these avenues until the target is found, but the mouse is not using vision to do the task. The acronyms are for the different mutations that more or less address different kinds of human blindness, that the field of ophthalmology has discovered, over the years. Now this mouse was blind, say, a month or two before, and received one dose of an adeno-associated viral vector that contains the channelrhodopsin-2 gene, and what you see now is that mice can very quickly navigate to the platform. In fact, they can navigate as fast as mice that have seen their entire lives. It doesn’t prove that they’re having conscious vision (seeing form and shape), but it does show that they can make use of cognitive information in a way that they were not able to before. In this study we also showed that the eye did not have responses that are indicative of an immune attack. And so it’s all in a mouse, but we are collaborating with many groups to try and see how this might apply to many different cells within the retina, or, potentially, in different animal models that might be more accurate than in the mouse - say in something like a primate model.
Now the eye can be accessed from the outside. But what about the brain itself? It’s very dense, it’s huge - how do you get light into arbitrary 3D patterns in the brain? So with Cliff’s group, where we worked on this 3D electrode recording device that we showed you earlier, we’ve also been working on 3D light delivery devices. The way these work is, they’re essentially microfabricated optical fibers that are all packed together. So what you see here: each blue line is essentially a tiny optical fiber, perhaps 10 microns wide, and they conduct light down this probe, until you get to the desired depth in the brain, where then a mirror will steer the light out at a 90-degree angle. Recently we also developed ways to scale this up into three-dimensional devices, so that now we can play back light in a 3D pattern. Now, there aren’t any light sources shown here. What we do is we hook it up to a projector, not unlike, perhaps, the kind that are at the back of this room, and the projector is essentially a megapixel light source. You can play back a 2D pattern, and our array will convert that 2D pattern into a 3D pattern. So we can play back arbitrary patterns to the brain. Or for example, spell out ‘MIT’ (I’m not sure what the brain thinks of that, but that’s something MIT people do).
Recently we started to implant these devices into the living brain - these are again using mouse models, as commonly done - and we can play back arbitrary patterns into the brain. We can now dial in sequences, or synchronous patterns, or try to engage one part of a circuit, and then another. My hope is that with these technologies, we can start to speak more naturalistic neural codes, with the right spatiotemporal resolution that the brain expects. So, for example, we can actually play back a pattern in one part of the brain, while we record neurons in another part of the brain, and try to figure out, ‘Is there an optimal pattern of control for one circuit, to make another part of the circuit do something?’ This can be very useful in terms of screening for neural codes that might be therapeutic. This has obvious applications for sensory prosthetics - in this case, we are actually implanting this into the sensory cortex. But there are also a lot of areas where you can make cognitive prosthetics of sorts. This is an area we started to think very hard about. Could you build, essentially, ‘brain co-processors’ - devices that can read information from the brain, using for example those very dense 3D arrays that I showed you earlier? Then, through custom computers that we’re designing, to compute exactly the information that’s needed in the brain, and then finally, to use three-dimensional high temporal precision, high spatial precision devices, to enter information back into the brain? This could be very useful as a general architecture for prosthetics, and one of the ideas is that in neurodegenerative conditions, where large fractions of the brain might be compromised - could you try to build replacement parts for the brain in the style of co-processors that could work with a computer chip?
Now, it’s very interesting to think about computers. We are using, currently, computers that are implemented the same way everyone else does it - in silicon. One of the things that we’re thinking about though is, if neurons have such incredible information-processing capacities, and also a genome that allows them to express all kinds of functions at different times, and all sorts of molecular signaling cascades that can integrate over a range of different time scales, what if we built computers out of neurons themselves? In the short term, this could be very useful, because now you could actually build neuron-based computers in a dish and use that to explore, in a constructive style, how to assemble neural circuits. What are the rules through which different cell types wire up, and how do you understand how these circuits, that you build from scratch, compute?
Richard Feynman once said, ‘What I cannot build, I cannot understand.’ One of the ideas is that neuroscience can benefit from constructive approaches, where you try to build brains, in all their glory, by trying to test our theories of how cells wire up and how they form connections - connectivity. In the long term, of course, should this line of research prove valuable, it might actually be useful for building replacement parts for the brain that have realistic cell types, connections, and so on, and have properties of the brain that are hard to mimic, including all the computational properties of neurons, but also the very low power usage of neurons, and the ability to intimately wire up with the existing neural circuitry.
So, we’ve been collaborating with the tissue engineering group of Gurkan and Demirci to explore this. What we’re doing is using photolithography, the same method used to make computer chips, to try to deposit cells of defined kinds, at defined points in 3D space, because, unlike other forms of tissue engineering, the brain demands an incredibly three-dimensional, high resolution method of fabrication. So, essentially, we wash in a monomer in a solution with some neural precursors. And then we shine light with a projector (you can do it through a mask as well), and aim light so that only some parts receive light and others won’t. Wherever light has hit it will cross-link those monomers into a solid and trap those neuronal precursors which were there at that point. Then we can wash away all the other monomers and cells and repeat this. So very rapidly, we may start to build up little neural circuits. For example, you can, as you see here, build a little cube of this polymer, and trap a cell at a fine point, and build little highway paths that are regions where the dendrites and axons can wire up. And one of the goals is to build this in a scalable fashion (can you do this at brain scale?), which will probably require us to also think about how to integrate blood vessels and vasculature into the mix as well.
I want to close on a note about how we’re trying to build tools that are powerful, but that we can rapidly deploy in the community. We try to give away tools however possible, to accelerate research, but also recently, along with Joost Bonsen, we started a new class on Neurotechnology Ventures, and we’re trying to figure out how to get people to get ideas out of the lab and into the world at scale as well. And part of my hope is that through these efforts we can build the neurotechnology discipline, but also help contribute to the neurotechonology industry as well. So this is a vast collaborative effort, and there’s not time to name everyone here, but I’ll close on this slide, and thank you everybody for your time.
