Part of our Transcriptions series. Protocol Labs has hundreds of recorded talks that have lived only as video. We are going through the archive and publishing the ones that still matter as text: a short introduction, the speaker’s own words lightly edited, and a note on what has happened since.

Ed Boyden is the Y. Eva Tan Professor in Neurotechnology at MIT and director of the Synthetic Neurobiology Group. He co-invented optogenetics — the use of light-sensitive proteins to switch neurons on and off — with Karl Deisseroth, and his lab invented expansion microscopy, which physically enlarges preserved tissue so its nanoscale wiring can be imaged on an ordinary microscope. His lab's tools are used by thousands of researchers worldwide.

The stakes he sets out: brain diseases such as Alzheimer's, Parkinson's, and epilepsy affect roughly a billion people, and for most of them treatments exist but cures do not. New treatments routinely take more than a decade and several billion dollars to reach patients. Expansion microscopy lets researchers see the brain's molecular wiring in fine detail. Optogenetics lets them switch specific circuits on and off with light. In mice, it revealed the 40Hz brain rhythm now being tested, non-invasively, as an Alzheimer's therapy. It's also being tested directly in patients, to restore partial vision in retinitis pigmentosa, an inherited form of blindness. A newer set of genetically encoded voltage sensors lets them watch brain activity unfold in real time in animals.

Boyden gave this talk at LabWeek Field Building 2024 — Protocol Labs' gathering of frontier-tech builders — held at Edge Esmeralda in Healdsburg, California, in June 2024. It closes on the collaboration he was then forming with Konrad Kording at the University of Pennsylvania to combine all three tools into data sets detailed enough to build predictive brain simulations, starting with the roundworm C. elegans and the larval zebrafish and scaling toward the mouse and eventually the human brain. An update at the end of this post covers what has happened since, including Boyden and Kording's October 2025 essay in The Transmitter, which sets out the full program (link).

This talk lays out Boyden's plan to build ground-truth simulations of the brain — computer models assembled from a brain's measured molecules, wiring, and activity, rather than from assumptions about them. His argument is that brain disease is hard to treat because neuroscience lacks what physics had by the 1930s, a practical description of the system's actual parts, and that three tools his lab helped build now make that description reachable for small brains.

Thanks all for coming. It's really exciting to get to talk today about some directions that are just starting to converge, and it's interesting both because a lot of work over the last couple of decades that might at first glance look quite separate and independent is all kind of colliding in an interesting way, but also hopefully you can see there's some method to all the madness, and hopefully we'll achieve some kinds of deep understanding not too far from now.

I direct a group at MIT that works on technologies for imaging and controlling brain circuitry, and I also advise lots of organizations — we've spun out several companies from our group. We've also got a lot of alumni running interesting organizations, and lots of interesting connections and collaborations, some of which I'll get to mention throughout. Our hope is that we can really, as a community, say something really interesting toward understanding the brain.

For me the core interests are philosophical. I've always wondered, ever since I was a kid, about human existence and what is the nature of it — why do we think what we think and feel what we feel. I've been lucky to work on a lot of different projects over the years that are on the border of something philosophical and something very practical, but neuroscience really has been my focus for the last really 25 years, I guess, and it's been on the technology side, in part because the brain is really, really, really complicated.

Why understanding the brain is so hard

I know many people around this room are experts on the brain, but not everybody is, so we'll just try to have a couple of slides to get everyone to the same page. The brain is incredibly spatially complicated — in a cubic millimeter of your brain you might have 100,000 brain cells, and they have very elaborate shapes, and those brain cells are connected through connections called synapses, and those synapses are also very dense — a billion synapses per cubic millimeter. Of course these brain cells are not full of empty space; they're full of biomolecules. We have something like 30,000 genes in our genome, and they encode for who knows how many kinds of biomolecules — probably hundreds of thousands to millions of biomolecules — and those are all very small, so brain cells are gigantic by comparison. We have some that are a meter in length going down our spinal cord, but the biomolecules that make up the brain, and all these things we just talked about, are nanoscale, and it presents a huge tension: how often do you have to understand a large macroscopic object in terms of its nanoscale components? Maybe the brain is the only example where you kind of have to do those two things at the same time.

And then of course you have the time dynamics. As biological systems go, the brain maybe presents more differing scales of time dynamics than any other system we have to deal with. If you care about brain aging, or the learning of a language, or the formation of a memory, these things take months, years, sometimes decades. But the quantal events of the brain — millisecond-timescale electrical events inside brain cells, and millisecond-timescale chemical exchanges between brain cells — these are extremely brief. Maybe not compared to silicon electronics, but definitely extremely short compared to almost anything else in biology. So, sometimes half-jokingly, I say the group has really only two goals: conquering space and time.

The scale of the problem: disease burden, drug failure rates, and cost

The reasons, of course, go beyond philosophy. I'm personally very motivated by understanding why we are what we are, but there's a lot of really pressing and practical issues that are very daunting. If you look at some of these names — Alzheimer's disease, epilepsy, Parkinson's disease — pretty much nothing on this slide can be cured. And the brain of course interacts with everything in our body, so there's a lot of concern now that metabolic and immune and other conditions also relate to the brain in some way. On the order of a billion people are affected directly by something that could be considered a condition of the brain, and I think all of us, if we live long enough, have to deal with this.

These are some numbers from 2010 — I think the numbers are actually far worse by now — but in Science magazine in 2010, the editors wrote a little editorial where they were lamenting that treatments for brain conditions take a long time to generate. I think the actual number is far more than a decade at this point. The failure rate is pretty depressing — I think now it's probably like 95% failure to get regulatory approval. The cost is enormous; I think Tufts Medical Center had a recent survey putting it at more like three to four billion dollars per drug. And then, even if you make it all the way through those hurdles, there's a good chance that your therapy doesn't work as well as most patients would have hoped. So what can we do about that?

In a lot of scientific fields, especially biomedical sciences, there's a tension between curiosity and application — the tension between basic, curiosity-driven research and translational research. Both are really important. Big revolutions like CRISPR and fluorescent proteins came from curiosity-driven research, and then you also have translational research, which is really important if you want to get something into the clinic or to patients. But both are tricky conversations to have — curiosity can go on forever, and how do you know what direction you're headed in; and translational efforts, as this slide alludes to, are mostly failures. So what can we do?

Learning from physics: a search for a ground truth theory of biology

When I started our group at MIT in 2006, our question was whether you could take a third point of view. I really like to study the history of science, and when I switched out of engineering and physics into biology, I spent a lot of time reading biographies and scientific stories about people who made that switch — Max Delbrück, who switched from physics into molecular biology, for example; Seymour Benzer, who also went from physics to molecular biology, many generations ago. In recent years, of course, large numbers of physicists have been entering biology, and physics is inspiring not just because it's so powerful, but also because of the intellectual journey that physicists took.

You can kind of divide physics into many chapters, but a very interesting standpoint is around the 1920s and 1930s, where quantum mechanics became a discipline and many people came with extremely practical ways of dealing with a quantum analysis of systems — Heisenberg and Schrödinger, and the list goes on — and came up with equations and formalisms that let you deal with real complexity. And that really led to a big revolution: we have the transistor and the microchip, lasers, the internet, modern communications, and the list goes on and on. So what happened here? We're now in an era where people are able to drop out of college and start Facebook, Apple, and Microsoft — all three were started by college dropouts. I would argue that's all based upon the fact that the microchip makes physics programmable by anybody. So this slide is metaphorical, but maybe also meant to be taken a little bit literally.

What happened here: in quantum mechanics you have a list of the parts — the fundamental particles of existence, at least on Earth's surface (who knows what's happening in deep space or near a black hole) — but on Earth's surface, quantum mechanics is pretty darn good, and also there's a practical recipe of what to do with it. With semiconductor electronics, microchips, and so forth, it makes the fundamental aspects of physics a good substrate for you to interact with. And so it's fun to think about whether there are analogies between these things and biology. Of course, the problem with biology, as I've already pointed out, is that we're dealing with a very large number of building blocks — many kinds of molecules, cells, and so forth — and also lots of ways that they interact. In quantum mechanics you have a small number of things that interact through a small number of forces — really a handful of each; you can write down all the laws of physics on a single sheet of paper. But in biology you have tens of thousands, maybe millions of things, and they interact through who knows how many kinds of chemical means.

What ground truth means for the brain: mapping molecules and building an AI interface

So what is ground truth for the brain — or you could argue any biological system beyond a certain level of complexity? I would argue it's probably the same: you want to know the building blocks and how they interact, and you need some easily accessible way to deal with that complexity in a practical fashion. So the data probably means mapping biological systems down to the molecules and their interactions, and then, because that's a colossally huge number of things to deal with, it probably means some kind of computer-scientific or AI interface, because you're not going to be able to hold all that in your head.

So let's get the numbers down, just to have them all in one place. The brain has enormous complexity in terms of its wiring — these numbers are astronomically huge, and that's just for a cubic millimeter. But also the molecules — lots of genes, gene products, biomolecules, molecules that we absorb from our environment through our diet or through the kinds of interactions we have with the everyday world. And then finally the time dynamics, which cannot really be easily analyzed, much less interpreted, at this point, because we know, for example, from wonderful work from groups like Yang Dan at Berkeley or Michael Brecht in Berlin, that if you stimulate just one brain cell in, for example, a rat brain — which has something like half a billion brain cells — you can change the entire dynamics of the entire brain. So a lot of these ideas, like the brain is modular, or we can think of the parts of the brain just in terms of cell types, are becoming more and more clearly approximations, and the mechanisms of the brain could go on for a very long list. We know that brain cells can make cannabinoids — things like the active ingredient in marijuana. We know that brain cells can make gases like nitric oxide. This goes on and on in terms of things we know the brain can do, but what is the function in thinking or feeling, or the dysfunction in a condition like Alzheimer's disease — the answer for almost any such question is we don't really understand it, and there might be mechanisms we don't even know about; people keep finding new molecules that have interesting functions we didn't anticipate.

So over the last 20 years or so, what we and our collaborators and colleagues have been trying to do is build technologies to confront these things head on. In the first part of this talk I'll briefly try to describe these different tools; in the second part I'll talk about how they're being applied to different problems that have clinical relevance; and then the third part we'll talk about where we could integrate these into, hopefully, data sets that are appropriate for ground-truth-oriented simulations of the brain in a computer. They're not really going to be three parts in strict sequence — they're actually interdigitated, because that makes it more fun — but that's how I think about it: three different topics.

The resolution gap: from brain scans to microscopes

All right, let's start with mapping the brain. Again, huge spatial tension between centimeter-to-meter scale and the nanoscale. I think we've all seen pictures like brain scans — these are often taking pictures of basically blood flow of the brain, or something related to brain metabolism. Each of these voxels that light up, those little yellow blobs, could contain millions, perhaps billions, of brain cells, and that's tricky, because we know that individual brain cells could have extreme power. We really would like to be able to map the brain with great fidelity, and of course the advantage here is it's not invasive — that's one of the reasons why we love this kind of brain scan — but even those of us who do a lot of this have to admit the resolution is just orders of magnitude away from where we want to be.

At the other extreme we have microscopes. Microscopes magnify images of things, and that's great, but whatever energy form is being emitted or transmitted through the specimen will have a size, or wavelength, and you're not going to see things much smaller than that. The wavelength of light is in the hundreds of nanometers, which is too big to see the fine wiring of the brain or to detect biomolecules in isolation from each other.

Expansion microscopy: making the brain physically bigger instead of the microscope stronger

So about a decade ago — our group likes to practice different methods for creative problem solving; I'm a big believer that creativity can be learned, taught, and practiced, and so one of the things we often do is try to be the most opposite of everybody possible — two grad students in my group, Fei Chen and Paul Tillberg, were trying to do nanoimaging with classical nanoimaging means: electron microscopy and super-resolution microscopy. And they were getting a bit frustrated, because these are very slow techniques that are difficult to apply to large 3D objects. So we started thinking, what's the most opposite we could be from zooming in on an image? And we started really thinking about this idea: what if we could make the object itself bigger?

One idea we honed in on was: can you basically take a preserved biological specimen — not a living thing — and fill it, infuse it, with the swellable material you find in baby diapers? Baby diapers have swellable polymers — sodium polyacrylate, for example — add water, and it'll swell; an experiment that millions of babies do every day. So this is what it looks like — I'll tell you how we do it in a second, but just to show you that it works: we can take a piece of brain, a couple of millimeters on a side, and chemically synthesize this baby-diaper polymer throughout the brain specimen, so it's not just outside the cells, it's inside the cells too — it weaves in between and around the biomolecules. Then, about a day and a half later, we add water; the baby diaper material swells, but because it's permeating the brain, we can make the brain itself bigger. The polymer starts out super dense, and then it becomes swollen. Here's a little time-lapse movie, about a half hour sped up to half a minute, just to show it in action — it's a piece of the mouse brain we polymerized earlier, and we just took it out of the water, and I hope you can see it swelling now.

So the original protocol — the whole thing is a day and a half; now we have it down to a few hours, and I wouldn't be surprised if we could get it down to minutes from start to finish. So I think the whole thing can be done really quickly.

How expansion microscopy works: anchoring, polymerization, softening, expansion

Okay, so how does this really look at the nanoscale? So far I've shown you big chunks of the brain at the nanoscale. What we hope for is to take a piece of the brain — let's say one brain cell, like this golden brain cell on the left — and make it look like, on the right, a constellation of biomolecules hovering in space, but pulled apart from each other: two biomolecules that were touching are now some minimum distance apart, and two biomolecules that were some larger distance apart have been scaled up by a linear factor. Of course this is not compatible with a living state — this is only going to give you molecular anatomy. I always get these questions, so let's just get them out of the way: the specimen is not alive; it's preserved; by this point we're pulling the molecules apart from each other, so they're no longer interacting. People often ask, "Have you disrupted the specimen?" Yes, it's completely disrupted — let's just get that off the table. Does it change interactions between molecules? Yes, they're all gone. That's okay, because what we want to know is the organization of the molecules — we have lots of other tools for looking at the living state that I'll tell you about later. Our goal here is to get molecular anatomy.

So what does the process look like at the nanoscale? We made some cartoons to help people visualize this, because it is a very weird way of doing imaging. The first key step is anchoring: we want to give little handles to all the biomolecules — that's how we pull them apart. Biomolecules are shown in brown, and we have handles shown in purple, and our goal is to put handles on all the biomolecules. Next step is polymerization: we can't just dump the baby-diaper material on top of the brain — we have to get it inside, and ideally very evenly. So we use a process called, not surprisingly, polymerization: you add monomers, shown as little white spheres, and they self-assemble on the polymers — these long chains — and whenever one of these chains of polymer encounters a handle, it forms a covalent bond. So now we have this densely permeating polymer mesh, and it will swell if we add water; we have handles that convey force to the biomolecules. We're almost there, but there's a third step that needs to be done, which is softening: we have to loosen up everything from each other, so we can use heat and detergent to soften up, or get molecules to not adhere to each other — sometimes we chop molecules into pieces — and then we add water, and the baby diaper swells. As it swells, because of that softening, and because of those handles, we have an expansion of the molecules away from each other. So the molecules are being pulled apart.

Validating expansion: checking distortion and resolution against known ground truth

We designed the process to be very even, but this is biology, so we should show that it is even. Especially in the early years, we spent a lot of time doing pre-expansion imaging with a classical nanoimaging method, like a super-resolution microscope, and then post-expansion imaging using a regular, old, boring conventional microscope. In light microscopy we use fluorophores to decorate certain proteins — for example, microtubules being seen inside a cultured cell; we add antibodies that bind to the tubulin proteins, and they decorate these microtubules with fluorophores. Pink is pre-expansion images taken on a super-resolution microscope, and green is post-expansion images taken on a regular microscope. We shrank down the post-expansion image and overlaid it, and what you can see is that's pretty good — the error is not zero, but it's very, very good; it's like a few percent over a typical microscope field of view. That's distortion. What about resolution?

We can compare what we see to the known ground truth — we know a lot about nanoscale things thanks to decades of other people's work using electron microscopy or super-resolution microscopy, so we can look at microtubules that we know exactly what they should look like, and figure out how much error we've added. With a four-and-a-half-fold expansion — that's how much our original expansion protocol did — we get about 60 nanometers of resolution; if you take the resolution of a regular microscope and divide by four and a half, you get about 60 nanometers, if you have a good microscope. So it raises the question: if you expand more, can you get better resolution? The answer is yes, but even with four-and-a-half-fold expansion, there's still a lot of things you can do that are really hard or even impossible to do with earlier means.

Of course, the first thing we wanted to do is multiscale imaging: take a piece of the mouse brain, a couple of millimeters on a side, and each white box is zoomed in, with a frame below. We're using fluorescent antibodies to decorate certain biomolecules, and the left is pre-expansion, so very quickly you get fuzzy images — green is the wiring of the brain, we've labeled one brain cell with a fluorescent protein, and pink and blue are the synapses, the neural connections, although as you can see they're fuzzy. Post-expansion, though, you can see a very crisp image of this branch of a neural dendrite, and the fuzzy things become clear — we can see pink and blue protein densities, which are the pre- and post-synaptic proteins at a synapse. If we look at the distance between these, we get exactly the same answer that Catherine Dulac and Xiaowei Zhuang measured many years ago with STORM super-resolution, invented by Zhuang — but now you can do this measurement with a regular, conventional, boring microscope.

Multiscale imaging: whole fly brains, mouse brain wiring, and scaling color-based labeling

Now, if you want to use a regular microscope for nano-imaging, you can, but when you expand something a hundred times in water, it's mostly water, so it becomes almost perfectly transparent too, and what that means is you can use clever tricks like light-sheet imaging, where you shine light from one angle and take a picture from an orthogonal angle, and you can go at blazingly fast speeds. Working with Eric Betzig — another super-resolution pioneer — we showed that by expanding and then using a lattice light-sheet microscope, one of these light-sheet microscopes developed by his group, we can go on the order of a thousand times faster than the nearest competing, equal-resolution super-resolution method. Here we're looking at mitochondrial and lysosomal proteins; at the bottom is myelination work done by Ruixuan Gao, Shoh Asano, and Gokul Upadhyayula. This work really illustrates the fundamental multiscale nature of this kind of work — we're zoomed in, with the same kind of color code as before, so we have fluorescent protein in yellow, and pink and blue are synaptic proteins; we zoomed in way in there, and now we're pulling back, and you can see how the nanoscale context can be visualized within a macroscale environment. It's an intrinsically multiscale method of imaging.

This is sort of a hint of where I want to end the talk, but already in 2019, working with Betzig and others, it was pretty clear this was going to be possible — this is the whole fruit fly brain, and all the dopamine neurons have been labeled (dopamine is often called the pleasure center of the brain, but of course we all know it's more complicated than that). Here is a fruit fly where every dopamine neuron has been labeled, and now we're going to fly through, so to speak, the fly brain — here's the whole fly brain, and now we're going to zoom in right into the middle, and now we're going to turn 90 degrees and go to the edge. Not shown here — they also labeled all the synapses, but they're not shown in this movie, because otherwise you'd see nothing else — and now you can zoom back out. These lattice light-sheet microscopes, with these kinds of expanded fly brains, optimized and done in high throughput, you can imagine scanning brains in a matter of hours to dozens of hours each, depending on how many colors you want to image.

That's fluorescent protein labeling. You can also label neurons with combinations of fluorescent proteins — some people call this method "Brainbow," after the first paper that did this. Here, different mouse hippocampal neurons are expressing different fluorescent-protein combinations delivered by viruses, and different neurons express different combinations. If you zoom in to a small piece of the mouse brain with a regular microscope, you quickly hit the resolution limit, because the wiring of the brain is tiny — you might see this green banana shape here, but what is it after expansion, same specimen, same microscope? Now you can see the green banana shape has four wires within it. We're working on trying to scale this up and create more and more barcodes — E11 Bio, which I think you heard from earlier in the week, is working on getting even more barcodes going for the whole mouse brain; I'm a science advisor to them, and we're very excited about the work we can do together. This is just for fun, because we did a fly-through for the fly brain — I thought we should do a fly-through of a piece of the mouse brain too, and it's just really beautiful how you can see what would have been a blur resolve into clean, non-overlapping colors — green wires, blue wires, red wires, and so forth. It's very clear this is going to work very well, and it's a very exciting time ahead of us to see if we can scale up to very, very large brain circuits.

Beyond proteins: expanding DNA, RNA, and other molecules — and where expansion microscopy is headed

So far, everything I've shown you is protein visualization — there are lots of other biomolecules, and I don't want to belabor the point, but suffice to say: if you can anchor it and you can stain it, you can see it. For example, Fei Chen, when he was in our group, and later Shahar Alon and Asmamaw Wassie, worked on anchoring and expanding proteins, DNA, and RNA all at the same time, and now we can also look at lipids and sugars as well. This is an example of expansion where we've taken a piece of the mouse brain, and you can see a fluorescent protein in green, and each pink dot is post-expansion labeling of a messenger RNA — so this is really cool, you can now visualize lots of different molecule types at the same time.

So where is this all headed? Expansion is very easy to do — we share all the protocols on our website, expansionmicroscopy.org. The point is that the number of experimental discoveries — actual experimental preprints or papers — I think is about to hit 700 any day now. It's being applied to the human kidney, to early cancer detection, to worms, to plant roots, and everything you can think of. It's very easy to modify — you want a bigger expansion factor, no big deal, you can change the recipe; we and others have made bigger expansion factors, and you can probably at this point expand as far as you want. You want to apply it to bone? Okay, you can add a softening step that softens the bone — each of the steps, anchoring, polymerization, softening, and expansion, can be modified to suit your purposes, and that's one reason I think this is taking off: if you have a nano-imaging microscope, most biologists would find it very hard to modify it for their purposes, but if it's a fairly simple chemical strategy, and you want to expand something like a bacterium with a hard cell wall, okay, fine, add an extra softening step that softens that cell wall — a group at Stanford, for example, showed a couple of years ago that they could add a cell-wall-softening step and expand microbiome specimens. We've posted step-by-step photographic tutorials on our website, so anybody can learn how to do it.

The future is getting really exciting — the resolution is a lot better than we originally thought; there are some hints that maybe you can even look at parts of a molecule (some preprints have been put out on that topic), and you can expand and then follow expansion with lots of other interrogation — people expanding and then doing mass-spec imaging, or expanding and doing x-ray imaging, and the list goes on and on.

Optogenetics: the search for a way to control neurons with light

Okay, that's the first part. Now, obviously, the limitation for expansion is that it's not going to work on a living thing, so the next couple of short stories I want to tell you about are the imaging and control of living signals, and then we'll put it all together — how this all works in terms of building models.

So this next part, we'll talk about controlling the activity of the brain, or optogenetics — here we focus on controlling the electrical signals of the brain. This started when my co-inventor, Karl Deisseroth, and I were both students at Stanford, and really wanted to find a way to stimulate the brain in a very precise way. So we used another one of my favorite creativity strategies, which is just to make a list of every possible idea — there's only so many kinds of energy you can deliver to the brain: electricity, magnetism, radio waves, light. There are only a few things, right — you can write it down probably on an index card — and then you have to convert that form of energy into some kind of thing the brain can understand. Electricity is probably the easiest thing, so what if you could install solar panels, for example, and convert light into electricity? We honed in on light as maybe the best way of doing it, because light can be very precisely aimed, and it's very fast.

We thought of other methods, like magnetic fields, which would have the advantage of being less invasive, but magnetic fields are hard to focus down to individual cells. To bring light into the brain — the brain doesn't have pain receptors the way the rest of the body does; that's why deep brain stimulators are put into the human brain while patients are awake (they have local anesthesia at the surface, where the surgery is done, but the brain itself, of course, they're putting the electrodes into deeply while the person's awake). Then the question becomes, how do you make the brain equipped with the solar panel?

From algae to brain cells: how optogenetics actually works

So I was reading lots of papers and was really fascinated by a class of molecule known as microbial rhodopsins. This little animation is a single-cell green algae, and they photosynthesize, and they want to be right at the surface of a body of water, so to navigate they have tails, or flagella, and to control the flagella they have a little eye spot that is full of light sensors, and those light sensors convert solar energy into ion flux — these are seven-transmembrane proteins, opsins, that convert light into the opening of an ion channel, and ions will flow across it — positive ions, for example, like protons and sodium — and then, in this single-cell algae, they'll go on to have a signal cascade that turns the tails to wiggle, and they swim to navigate toward the surface of the pond, which is where they can photosynthesize.

So I was reading these papers and got really excited about these molecules, and started writing to people asking if they would send us the genes that encoded for these proteins, and we were very lucky that people were willing to share these pieces of DNA with us. We put the genes into transfectable or viral vectors and put them into brain cells, and it was sheer dumb luck that it worked — they could have been toxic, they could have been misfolded, who knows what, but they expressed in the brain cells properly, and when we shone light on the brain cells, we would get activity — these brain cells would respond to light with electrical pulses, much like the kind that are happening right now in your brains as I say these words. And, maybe a more specialized comment, but to make this molecule work you need a vitamin A variant called all-trans retinal, and for whatever reason we don't really know, it's found in the mammalian brain and body, so we're very fortunate that that's true — fruit flies and worms, for example, you need to feed it to them.

So how can you make this work — this next movie has flashing images, so if people are photosensitive I can tell you when the flashing movie is over — but it illustrates how we use it: the brain again has many kinds of cells, some are big or small, some are fast, some are slow, some excite, some inhibit. We can use different tricks from the field of gene therapy and make one set of cells light-activated, and their neighbors will not be — for example, we can take these small, star-shaped cells called basket cells, or parvalbumin cells, make them express our light-gated channel, and then when we shine light from an LED or a laser, maybe for a deep target conducted down an optical fiber, those cells will be excited by light, but not their neighbors. Okay, the flashing images are over now.

Optogenetics and Alzheimer's: the 40Hz gamma entrainment story

Now, why is this useful? I'll just give you one example, from my collaborator Li-Huei Tsai's lab — work led by Hannah Iaccarino, who was a student with Li-Huei, and Annabelle Singer, who was a postdoc with me at the time. Li-Huei's group works on Alzheimer's disease and also on brain rhythms, so they decided to try to drive those small star-shaped cells I showed in the animation (this is kind of a coincidence — I made the animation not knowing they were interested in this topic and working on it; it was all in parallel). They had this idea to drive these star-shaped cells at a specific rhythm, 40 times a second — 40 hertz, or 40 times a second, is a magical frequency in neuroscience; people have found it associated with attention, and linked to all sorts of interesting features — and so they wondered, what if you drive this frequency, could you do something about the brain?

Li-Huei's group, and Hannah in particular, had this real desire to try to drive this frequency in Alzheimer's model mice — mice engineered to get Alzheimer's-like conditions — and to everybody's surprise, there's no theory that led to this; it's really about intuition and serendipity, and the brilliant jumps people make in science. What they found was that if you drive the brain at this frequency, the Alzheimer's symptoms clear up — you get removal of amyloid plaques, and hyperphosphorylated tau, brain inflammation goes down, a whole bunch of things get better.

Now, a lot of people have Alzheimer's — my own grandmother has Alzheimer's disease. If you have to give everybody a gene therapy, I guess that's your goal, but if you could do it non-invasively, that would be even cooler. Emery Brown, an expert in brain rhythms and also a clinician in his own right, had this great suggestion, and what the teams did, under Li-Huei's guidance, was try to do it non-invasively by basically playing a movie — no more light-gated proteins, no optical fibers, nothing, just a movie — so mice would watch a flickering light and hear a clicking sound. Amazingly, and this is work that was continued by Anthony Martorell and others, again with Li-Huei leading our collaborative, what people found was that this flickering light and clicking sound was able to actually improve cognition and slow down brain degeneration and all sorts of stuff in Alzheimer's model mice. We now have three clinical studies — one led by Annabelle, now running her own group at Georgia Tech; one led by Diane Chan and Li-Huei, with help from me and others out of academia at Harvard and MIT; and then one from a company, Cognito Therapeutics, that Li-Huei and I co-founded, to try to do very large-scale clinical trials. These small-scale studies are intriguing because they all do suggest we can slow down brain atrophy in humans who are watching and hearing these movies. There's still a ways to go — nothing's been approved by any regulatory body — but we're very excited to keep going down this path and see if we can help a lot of people who have this otherwise pretty daunting and intractable condition.

Optogenetics in the clinic: restoring vision in blind mice and in patients

This is kind of a technical slide, but there's a lot of different molecule classes that we and others have found, and we can use them for different purposes. We can put in light-driven channels that let in positive charge and activate cells, and you can then drive patterns like Li-Huei and Hannah and others wanted to do for Alzheimer's. You can also find molecules that put negative charge into cells, and those can shut down brain cells — Brian Chow and Xue Han worked with me on the silencing; Amy Chuong later found red-light-drivable versions — red light is great because it goes deeper in the body than other colors of light. Karl Deisseroth and I, as I mentioned, put light-gated ion channels into neurons many years ago, and then Nathan Klapoetke and others worked on trying to perfect them by again making them red-shifted, and so forth. Something very interesting has been happening in the last couple of years: optogenetics — opto for light, genetics because these are genetically encoded molecules — is starting to be explored in actual human patients.

A lot of the impact of optogenetics, and this is probably going to be true for the foreseeable future, is by discovering patterns of the brain that could be helpful for healing the brain, like we're doing with Li-Huei and Cognito and so forth. But a brave team in Europe took one of our molecules, the red-light activator that Nathan found, and showed that it might work in the human eye. We were trying to play with this many years ago, but I think we were a little too early — this is work led by Alan Horsager; we actually tried to start a company, but I think we were a little ahead of our time. This is a blind mouse — there are different mutations in mice, many of them natural, that phenocopy aspects of human blindness. Millions of people are blind because they've lost the photoreceptors in their eyes, the light-sensing cells. This mouse is in a water maze — mice like to swim, but they like to sit and do nothing even more, so the water maze has six arms, and one of the arms has a platform that's lit, and if they find that platform they can climb out and don't have to swim anymore. If they swim for more than a minute we take them out, so we can have them do multiple trials. This mouse has the rd16 mutation, retinal degeneration 16, which phenocopies some of the mutations that cause, for example, retinitis pigmentosa, which affects about two to three million people around the world. These are different mutant mice that have photoreceptor loss, and all of them are smart — they eventually learn to find the platform through brute-force search, but not through vision. This mouse got one gene-therapy dose of a channelrhodopsin, a light-gated ion channel, and you can see it's going right for the platform — it takes a while for the mice to learn to use their eyes again, but over time they get really good, and they can eventually solve the water maze as fast as a normal mouse.

So, this was 2011 — two summers ago, I guess, sounds like three summers ago now — a team in Europe, Botond Roska and José-Alain Sahel, bravely picked up where we didn't get to that finish line. They took our molecule "Chrimson," this red-light activator, and in their company, GenSight, they were able to deliver the gene for Chrimson using an associated virus. They took people with retinitis pigmentosa, delivered the gene to one eye, left the other eye at least for the short term untreated so they could do a fair comparison, and gave the person a pair of goggles that took pictures of the world and translated it into the right color — it's a red-light-activated molecule, but they used orange light because there's apparently a bit more activation with orange light. They were able to get a very good but partial restoration of functional vision — it's not perfect, this person was not able to read text, had trouble recognizing faces, but was able to recognize household objects, was able to see doors along a hallway or lines of a crosswalk. It might be the best restoration of functional vision of any kind for this condition. Roska now has ideas for how to use the opsin to get better restoration of fine vision as well, so it's really exciting that this is starting to take shape.

So, optogenetics allows us to really interface with biology by controlling it with light — they're genetically encoded molecules, so that's really simple to use.

Reading brain activity: voltage imaging with engineered sensors

I just want to mention one very last short story, just a couple of slides before talking about how we tie it all together, which is: how do you see what's going on in real time in the brain? With optogenetics we were so lucky — these molecules just, out of the wild, work, and even that vitamin A relative I mentioned that you have to add — well, the human body apparently, and the mammalian body, has it, so we don't even have to introduce any small molecule into our brain, which, as some of you might know, can be challenging. With the inverse — imaging — we haven't been quite so lucky, but humans have been ingenious where luck didn't take them.

This is a screenshot from a database of engineered sensors that biologists have invented. How do these work? People take a fluorescent protein, like the green fluorescent protein from jellyfish, and glom it onto some kind of sensor — a protein that binds calcium, or a protein that binds cyclic AMP, or whatever — and by engineering this protein cleverly (previously through high-throughput screening and design, but in this post-AlphaFold world there are lots of more quick ways of doing it), you can engineer molecules that will light up when some signal is there.

Our group has been working in this area as well — for example, on fluorescent voltage indicators. This is a snapshot from a paper led by Kiryl Piatkevich in collaboration with Xue Han's group at Boston University, and what we showed was that we could take — actually it's an opsin molecule, and Adam Cohen's group at Harvard discovered that some opsin molecules actually are weakly fluorescent voltage sensors, really another brilliant piece of serendipity. We followed their lead and started mutating some of the molecules they were looking at, and made a molecule we call "Archon" — it's a mutated opsin molecule, and what this does is: when you shine red light on the brain and collect the infrared light, you see fluctuations in the fluorescence, and those fluctuations are proportional to the voltage of the cell. So these look like electrical traces, but they're not — they're being picked up on a camera, and because it's a camera you can see many cells at once. These are head-fixed mice, they're awake and behaving but not moving, and you can see many cells at once within a microscope field of view.

It's kind of fun to think about — right now there are these brave clinicians putting opsin molecules into the human body, and the voltage reporters I'm talking about are mutated optogenetic molecules, so if optogenetic controllers are safe in the human body, maybe optogenetic controllers that are mutated to read voltage are also safe. It's fun to think that maybe all optical input and output to the brain might not be science fiction anymore — it might be quite tractable, because all the safety data that applies to the optogenetic controllers increases confidence that optogenetic readout molecules might, if designed and screened carefully enough, also have good safety profiles.

Why simulate the brain: the case for a ground-truth model of small brains first

So that brings me to the final part, which is tying it all together. All these tools I talked about — fluorescent reporters, optogenetic controllers, expansion microscopy — we share freely; I mentioned that thousands of people are doing expansion, thousands of people are also using the optogenetic controllers, and it's been used to make so many discoveries. The voltage indicators are much newer, so we're just starting to get those out into the community, but I'm very excited about them too.

So where is this all going? Up till now we've focused on building tools and kicking them over the fence, where other people — Li-Huei for Alzheimer's, or Roska for blindness — will then brilliantly apply them to solve clinically relevant problems or basic-science problems. But I got into biology in the first place because I was interested in the philosophical angle — what does it mean to be human, why do we do what we do and feel what we feel — and it seemed like we really want to understand the brain. Like I was alluding to at the beginning, its immense complexity really means that a computer simulation in software might be one of the best ways to try to understand it, because then you could really look at the processes that occur and play with them too. Causality doesn't scale very well — we can image at scale, we can simulate at scale, but if you want to go and perturb every cell with optogenetics, that's going to take forever, and forget the combinations — if you have even a small brain, let's say a small fish with 100,000 neurons, and you want to try all the five-way combinations, good luck, you're not going to be able to do that.

So we started thinking: what if we could, from bottom to top, image the high-speed dynamics, perturb it, and then, at the end of the day, preserve that brain and make a map using the expansion methods? Working with Konrad Kording at UPenn, we've been trying to build a network of people, initially in academia, but at some point I think the scale might be very amenable to starting some kind of new, interesting hybrid structure, which could, for example, be housed in a nonprofit but could spin out for-profit entities that have different applications — much like many structures people have envisioned for different purposes, universities, or Broad Institute, or OpenAI, or so forth. You see lots of interesting structures now, where you can collaborate in different ways and scale in different ways.

Building the map: imaging living activity and mapping molecules in worm and zebrafish brains

So how does this work? The idea is to start with some small brains, but then, as you'll see, we'll collect different kinds of data and machine-learning principles that could be applied to any brain — I think of this more as an approach: we're going to pipeline the imaging and perturbation and mapping of a system, and you can then apply it to any system. We're going to test it in small brains first, like a worm with only 302 neurons, or a small fish which has only 100,000 neurons, but this approach, as you'll see hopefully by the end, could be applied to any system.

Konrad and I have been working with a large number of people. Basically there's two kinds of data and three kinds of models we want to build in the next, let's say, three to six years, for the initial phase of the project. The first idea is to look at the living state — we want to be able to image all the activity of a small brain, a fish or a worm, while we're imaging their behavior — so we're designing, for example, microfluidic devices and extremely fast microscopes that could image all the activity of an entire brain, ideally at the same speed that the brain really computes at, which is very challenging — and then use the optogenetic controllers, like the ones I told you about earlier, to do causal testing, to be able to see the importance of a given node of the network.

I'll just show you an example here for the fish, because it's the one we were both working on — a larval zebrafish with 100,000 neurons, as well as the worm, C. elegans, with 302 neurons. Sometimes the fish comes through faster, sometimes the worm, but we pick them because they're genetically tractable animals, they've been studied for decades, they are small and transparent, but they also, thanks to evolutionary conservation, might teach us things that apply directly to mouse or human, and also, by practicing, we'll know where the sticking points are for scale, and we can then scale up to larger brains, or work with partners — I mentioned E11 earlier, which is tackling mouse brain mapping — there might be different kinds of structures or institutions that can tackle different scales of the problem, and so it really is going to be a very interesting and vast community network.

Here's an example where Zeguan Wang, who just finished his PhD in my lab, built what might be the world's fastest light-sheet microscope — basically scanning a sheet of light, like I described earlier, and collecting those sheets of light onto a camera — but he pulled out all the stops here: for example, he vibrates a mirror in order to scan the light sheet — this is a very small, low-mass mirror, so you can vibrate it at blazingly fast speeds — and found a camera that goes faster than any camera I know of, something like 3,000 frames a second, and not low-resolution frames, full-resolution frames.

Here's the larval zebrafish, on the order of 100,000 neurons — we're scanning it at 7,000 frames a second, 200 volumes a second, and what you get are neurons across the entire brain. These traces look like electrode traces, but they're not — we're imaging neurons across the whole brain, and it's just exciting to see how fast things are going.

I showed you this slide earlier, which we published in 2019 — a paper in Nature that got a lot of attention at the time, which I think was the world record for most neurons imaged in a living animal with voltage: this was 13 neurons in an awake, behaving mouse. Five years later, we are now at, I think, at last count, 25,000 neurons at the same time, so we're going way faster than Moore's law here, and it's very exciting to think that this idea of whole-animal imaging of voltage is now within reach. We posted this on bioRxiv last December, and this is, to my knowledge, the first attempt to image the neuron voltage of neurons distributed throughout an entire brain — there's more work to be done, we didn't get every neuron, and 200 volumes a second is fast, but we'd like to go even faster.

Part two, the molecular atlas: this is where expansion and barcoding of the wiring, and mapping the molecules, is so key — again we do both worms and fish. Konrad worked on graphics with an artist — it's a bit more worm-centric, but the same idea for the fish: we expand it, and then, once you've expanded it, you can label all these different molecules, maybe over many rounds of staining and imaging, or even sequencing — I didn't have time to talk about in-situ sequencing, but one hope is that you could expand an object and just sequence all the DNA and RNA and so forth inside of it. So now you have a map of the wiring and the key molecules, and we've been working to adapt the expansion tools for the worm and the fish too — Limor Freifeld, when she was a postdoc in the lab, showed that we could expand zebrafish brain and see individual synapses; Jay Yu, when he was a grad student in the group, expanded the worm and showed that we can get very fine features of the worm.

Three kinds of models — and why molecular mapping might scale best to the human brain

Okay, so how do we put this together?

Konrad Kording and I have been working on some proposals to put together a large collaborative — initially again in academia, but one can imagine different institutions being formed and collaborating in different and non-traditional ways to make this really scale — initially with mouse and worm as test beds, but also working on scaling the tools to mouse and maybe even human. So we're going to collect lots of real-time causal and physiological data, and then we're going to map all these key molecules and wiring.

There are different kinds of model we want to build — three kinds. The first is to use deep learning to basically look at the activity and see if you can predict behavior directly from activity, and a lot of people are doing that kind of work already — it's very cool, because you can look at activity and try to predict what will happen next, and we're doing this as well. It's powerful because you're able to do very direct linkages between activity and behavior. The downside, of course, as with any sort of abstraction-layered deep-learning model, is that you're going to be missing fundamental mechanisms, so it might be very difficult to predict out-of-distribution behaviors — imagine you image worms and fish doing 15 different behaviors, but in a 16th behavior you haven't thought about, some neuropeptide is released that you haven't thought about, and it reconfigures the dynamism of the entire brain in a way you have never seen before; the deep-learning network might fail to predict that.

So that's why we also want to start thinking about, from these activity models, can we infer network physiology, neuron physiology — we know the genomes of these animals, we've known that for many years, and we could try to guess: this cell had these bursting patterns, or these slowly changing patterns, let's try to guess what the underlying physiology is, and that might give us more predictive power for out-of-distribution predictions. That's of course very interesting, because whether you think AI is not powerful enough or too powerful — I've heard both opinions — in both cases, knowing what is happening out of distribution could be very helpful; if you want more powerful predicting of things that lead to creativity, or human ethics and so forth, that could be very interesting, and if you think it's too powerful, knowing what it could do is also helpful. So we're very interested in inferring physiological models from the data, and we can then use that to improve our network models.

Now, the molecular data is very powerful, because if you can learn how molecules yield physiology, then we can test whether our inferences were correct — these are guesses as to what underlies physiology, but we know the wiring and the molecules, or at least we can get that with our techniques. We can say: I know exactly where dopamine or this other transmitter was, I know exactly that there were 17 connections between these cells, and here's exactly their geometry — so now we can go ahead and actually populate the neuron model bottom-up with molecular information, and each model can improve the other: knowing what the network does can help us figure out whether our physiology is relevant, and knowing what molecules are there could tell us whether the physiology we were inferring was correct. So all three levels of model — network level, neuronal level, and the connection to molecules level — are quite powerful.

Now, I'm also very interested in the molecular mapping, because that scales very well. Live imaging and live control — it's hard to see how you would control and record every neuron in the human brain; I don't think there's any technology on the table that would scale to every single cell in the human brain, and even if you did have such a technology, the need for medical ethics, which of course is so important, would mean you'd need a path to doing that, which could take a very long time. But mapping molecules in the human brain — we can obtain brain specimens from brain banks and so forth, that's not so hard, and we already, and others, have applied the expansion method to human brain tissue.

So suppose you got really good at predicting physiology and emergent network dynamics from molecules — maybe someday you could go directly from molecules to the emergent behavior, skip the middle steps. My hope is that we can validate the linkages between these levels of description through small brains, and maybe through other in-vitro experiments, like cell lines expressing different kinds of ion channel and receptor, and learn how to predict physiology from molecules, and then we can use molecular mapping to scale up to the human brain.

This is a list of the investigators who are part of a proposal we're circulating with people interested — again, initially maybe an academic phase, but also maybe with nonprofit and other sectors' involvement — Konrad and I have been very lucky to work with this great team. We've been taking little baby steps in this direction — this is some work led by Quilee Simeon when he was working in Robert Yang's group, collaborating with my group (Robert is leaving MIT to go start an AI company, so Quilee is switching to our group).

This is some data where Quilee took off-the-shelf data — worm data from other groups — and just showed, using deep learning, that you could take histories of calcium traces (this is not voltage traces, because voltage is still coming online) and train a neural network to predict future neural activity from past neural activity. It's very early days — there's not enough data in the world really to make these models work at the level I'd like to see — but it suggests there is structure in the activity of the brain, and this is not an unforeseeable path, so it's very exciting to see that maybe the possibilities of brain modeling are well-posed, potentially.

So I think I'll end there. I want to acknowledge all the people along the way who've led these different projects — a lot of other people listed on the slide have helped, and then we have an even longer list of people we collaborate with all over the world; this is just a list of the leaders of the groups we work with on these and many other projects I didn't have time to talk about.

If people are interested in using the different tools, our website synthneuro.org has a list of them, and my email, if people want to learn more, is eboyden@mit.edu. With that, I can take questions.

Update since the talk: what's happened since June 2024

Since this talk was given in June 2024, several of the threads Boyden describes have moved forward. Boyden and Konrad Kording published the argument for ground-truth brain simulation in October 2025 as "Whole-brain, bottom-up neuroscience: The time for it is now" in The Transmitter. E11 Bio, the mouse-brain-mapping effort Boyden advises, released its PRISM method in December 2024. Cognito Therapeutics, of which Boyden and Li-Huei Tsai are scientific co-founders, completed enrollment of its roughly 670-patient pivotal HOPE study of 40Hz stimulation in July 2025; results are pending as of this writing.