Episode 4 | Transcript
Episode 4: “Optical Methods for Molecular Connectomics”
Published Wed, 3 Oct 2018 | Transcript
Allen Sulzen: Hi there, and welcome back to the carbon copies podcast. My name is Allen Sulzen and this is the fourth episode on our workshop series. Today we’ll hear a presentation from Dr. Adam Marblestone, on optical methods for molecular connectomics. Before we begin, if you’ve been enjoying this podcast, take the time to share it with your friends. If you have our vision as well and are interested in volunteering for carboncopies.org, please visit our contact page to join. Dr. Randal Koene will introduce the next speaker.
Randal Koene: Our next speaker is Dr Adam Marblestone. Adam spent a good number of years driving research at the MIT media lab of Ed Boyden, with work on optical connectomics and expansion microscopy and many other things. He’s also known as the main author of a number of strategic papers such as the ones that Tony already mentioned, exploring ways to accomplish high resolution brain mapping. Adam is the chief strategy officer of the neurotechnology startup colonel. I think that Adam’s talk today is going to be an excellent compliment to the one that Tony just gave. So Adam, please feel free to begin sharing your screen.
A. Marblestone: Yeah. So it’s great to go right after Tony, because it’s going to make it a lot easier to explain what I’ll be talking about, which is mostly additional layers of chemistry and and some extra explanation that will hopefully clarify some of the things that he said as well, but very much in the same spirit. The purpose of what we’ll be talking about is this notion of a Rosetta brain, a very comprehensive mapping of as large a volume of a brain as possible; ideally an entire brain. This is describing work that’s done in the Boyden, Church and also Tony’s Zador’s lab over the past few years. So, as I think some other speakers talked about, kind of the gold standard way to, to look in fine detail at neurocircuitry today is the electron microscope, which involves doing a very fine slicing and imaging essentially of the morphology of neurons.
So, the cell membrane is typically the main thing that’s labeled. You’ll see a little bit of extra electron density for certain molecular complexes like the Post synaptic density that tells you something about synapses or synaptic vesicles in some cases, but mostly you’re looking at the cell membrane; and that’s about it, or at least the membranes of the cell and some organelles. In terms of doing a Rosetta brain type concept of being truly comprehensive, this is sort of going to go over similar ideas as Tony presented, but just in a slightly different angle on them. There are a few other things that we would like to be able to see. It is still an outstanding challenge for electron microscopy to be able to give us this additional information.
So, one of them is basically molecular. So on the left here, this is a slide from Eve Martyr, who studies the Crab Stomatogastric Ganglion, which is a region in the gut of a crab that basically has about 30 or so neurons, and it actually has more neuromodulatory molecules that affect it, influencing how the synaptic transmission works, or influencing how the ion channels and the cells, respond to time or things like that. It has more modulators than it has neurons in this circuit. It’s interesting to think about whether that will also be true for other circuits, or at least the complexity might not have gone down much, as you go up into something like cortex of a mammal. The other is a sort of very fine spatial resolution with electron microscopy has but, we would also like to combine with these barcode aspects and also to combine with a molecule.
So, on the right there is a diagram from a region of the Thalamus, called the lateral geniculate nucleus, that relays information from the retina into the visual cortex. What you’re seeing is basically the incoming synapses onto one of these thalamic relay cell dendrites. What you can see is that not only are these synapses arranged in a complex pattern, in space with different branching axon terminals and so on, but also each type of synapse is using slightly different combinations of transmitter and receptor molecules that are going to influence what the actual meaning of these signals is and how they affect this Thalamic relay. Likewise, it’s quite plausible that this kind of complexity also extends into other circuits throughout the Mammalian brain.
So, these are things that we would like to have. Electron microscopy has a bit of that in terms of the shape in the spatial layout, but doesn’t have the molecules right now. Of course we want to have what Tony talked about, I think, which is basically what makes it dramatically easier to scale this to large volumes and long distances for any kind of structural mapping, because, often times what we would want to look at, in neuroscience is not just a local circuit (a few hundred microns or cubic millimeter sized circuit), but you might want to look at something like the connections between the Cortex, the Thalamus and the Basal Ganglia and how those are all mutually interacting, to have a kind of complex architecture. You really want to know the inputs and outputs to an area in addition to just what that area is doing internally.
So having these unique identifiers for cells, these barcodes that Tony talked about is also something that I view as essential for doing a comprehensive mapping at a kind of whole brain level. As Tony mentioned, a way that we can try to put all these things together is by doing the sequencing of barcodes and as I’ll mention also potentially of other molecules, that are present in the sample in addition to the barcodes. We want to be doing the sequencing of those as an imaging approach. As Tony talks about, basically this mirrors the way in which if you get your genome sequence today, what you’re essentially doing is, your genome gates get chopped into lots of small fragments of DNA which land at different positions on a glass microscope slide.
Then essentially the sequencing machine is actually a fluorescent microscope that takes a series of images of each of those fragments, reading out as a sequence of colors, one letter at a time, the identity of each strand of DNA. Then computationally, what people do is they take all those short fragments and they try to re-stitch that back into your genome. This is a technique that was developed in the Church Lab. Now, as part of this collaboration with Ed and Tony, they further developed this notion of fluorescent in situ sequencing and it is being applied to the cellular barcodes. So, I think that the question that we get to, if we want to combine all these elements though, is how do we get a sufficient spatial resolution of doing this in situ sequencing, to read these barcoded molecules or other endogenous molecules? We can illustrate this question of spatial resolution, by considering just how much spatial resolution do we actually need to see every individual synapse; because the synapses are packed closely in space, with about one or a few per cubic micron of the brain tissue, at least in the cortex.
So one of the issues that you run into, if you’re trying to use an imaging based approach, to read whether it be barcodes or morphology or other things like other molecules, is that the synapses are packed densely in space and you don’t want them to overlap, or else you won’t be able to resolve, what you’re looking at. So what various people have concluded, starting with Uray Machenko, back in 2011 who studied a related problem of optical detection of synapses, is that in order to actually uniquely see each synapse, you don’t necessarily need quite the level of resolution of electron microscopy, the kind of resolution that Ken Hayworth is talking about where you would maybe have five or 10 nanometers, at XYZ resolution. However, you do need probably resolution of more on the order of 100 nanometers.
Whereas, a conventional optical microscope is more like three or four hundred nanometers in the XY direction, and slightly worse than that, for most microscopes in the z direction, that the light is actually propagating along. So, we felt that we needed to introduce some additional elements into this picture to try to boost the spatial resolution, initially to try to resolve synapses better, and as I’ll talk about, also to potentially resolve other aspects of the cell better. So, back in 2014 or so when we were starting to think about it, this seemed like a rather difficult problem. There were certain existing super resolution microscopy approaches that could be applied. Some of them involved in looking at the stochastic blinking of single molecules, kind of sparsifying the sample.
So, most of the molecules are off, and you can turn a subset of them on and try to find the centers and then stitch the image back together by localizing the centroids of certain dots. Another idea that came up early in the work around in situ sequencing was maybe there are other ways to sparsify the image so you wouldn’t be looking at everything at once. Everything would not have to overlap. You could look at it in a series of stages where only some of the molecules are being looked at. There were other groups that were doing ultra thin sectioning, kind of mirroring some of the work that’s done in electron microscopy, but here being used with an optical microscope. Then there are some other ideas around computational approaches to resolve overlapping objects.
But, basically, I was fortunate to be in the loop at that time with Ed Boyden’s lab as well, which had started developing a totally different approach to increasing the spatial resolution of any kind of optical observation of hunks of intact 3d brain tissue. The new approach is called expansion microscopy and it basically means physically making the brain itself larger, rather than trying to essentially make the light smaller. You take the same microscope, but you physically swell the brain. The way they do that is by infusing the brain tissue, with this swellable polymer, sodium polyacrylate, which when it is exposed to water with low salt content, will start to create chains that repel each other and will swell and hydrates, as an expanding hydro Gel and will actually blow up physically the sample initially, in what they demonstrated by about four or five times along an axis.
So, that gets us kind of exactly into the range of resolution. Starting with a regular optical microscope, of several hundred nanometers to a micron resolution down into the sort of sub 100 nanometer resolution range, that seems to me to be kind of the right range for doing a Rosetta brain mapping effort. In order to actually allow the brain tissue to swell, one of the things that the folks in the Boyden lab had to do was come up with a trick, which is how do you allow the tissue to swell uniformly, right? If you just infused this polymer, different parts of the tissue are going to have different mechanical stiffness and so on; and what you expect to happen now is basically that the tissue will tear, rather than swelling uniformly or it will swell to a certain point and then not be able to swell any further without tearing.
So, what they did is they played a trick where they first deliver, to the molecules of interest that they want to look at, which in this case was proteins that were tagged with a chemical tool called an antibody which binds specifically to certain proteins and could deliver other molecules to them. So, they tag the proteins that they want to look at with an antibody and attached to the antibody, is a little strand of DNA, and the DNA has on one end and molecule which will link chemically into the expanding gel network. On the other end it has a fluorescent dye. So, what they do is they deliver these tags, anchor the fluorescent molecules into the expanding gel, and then they digest away with an enzyme, a lot of the endogenous structure. They get rid of a lot of the protein, a lot of lipids, and they’re only leaving behind a kind of a shadow mask, of what they originally wanted to see.
That shadow mask is this very dilute solution, that is basically able to be expanded uniformly; so this works very nicely. As they showed with brainbow labeling of neuron membranes here, this is the exact same image of the exact same location in a chunk of brainbow labeled tissue, on the left without any expansion and on the right with expansion, and you can see that you’re resolving finer details, after expansion against the same microscope. So what they’ve done here is they’ve anchored fluorescent tags indicating the different colored rainbow proteins and allowed those two to expand uniformly. So, we have been trying to combine this method of expansion microscopy with some of the things that Tony has been talking about and also with a whole suite of other molecular tagging and indexing and readout chemistries, to try to read out as much as we can from the same expanded sample and again, feel for each thing we want to see.
We have to find some way of tagging it into the expanding gel and allowing that to expand, along with the gel while everything else gets digested away. What we’re showing here, is actually a simulation based on electron microscopy data of what we would see, so that we can understand the parameters in kind of an ideal scenario. We have a simulation of what we will see looking at labeled synapses, which could, for example, be Barcode Tag Synapses, at different levels of expansion. You can see that at one x or two x, it’s sort of blurry and hard to resolve individual objects, but once you get up to about three x or four x expansion, you start to see the individual synapses as very well resolved objects. This is similarly borne out, in some of the papers on expansion microscopy where they’ve been labeling synapses.
So, essentially this approach for Rosetta brain mapping, that we were trying to do is to combine these three technologies. So expansion, microscopy, barcodes of the type that Tony described and in situ sequencing into a single approach. There has been some progress on this from Richie Comin, in collaboration with Tony’s lab and the rest of this group on doing that combination of all three technologies, that we’re continuing to work on. At the same time, we are asking the question of whether we can push this, to even higher resolution. So, the original expansion microscopy papers did about four fold expansion. Again, this is getting us down into the sort of 100 nanometers or so resolution range, which is quite a great boost for our in situ sequencing technologies, but is not at the level of electron microscopy.
What we’re showing here is an iteration on the approach, which is literally done by iteration. Actually what they do is they take the expanded hydro gel and then they do a second round of the same expansion microscopy process on top of that initially expanded one; so they’re doing it twice, from Jin Wu Chang, in Ed’s lab where, now with the 20 fold expansion, you are starting to get down into the kind of 20 to 30 nanometer resolution range, which is still not quite at the level of electron microscopy, but it’s starting to get close. Here, what you can see in these images is actually that we’re starting to become limited by the chemical tagging process, where we’re starting to see the discreteness of the chemical tags down at the molecular level, as a limiting factor in our resolution. There’s ongoing work to try to push beyond that. So here also, you can do this with synapses, showing now optical imaging of several different molecular tags on the synapses, and starting to get at the idea that when you’re looking at the connectome, it would be great to also characterize something about the molecular makeup of each synapse. So here on the upper left they’re showing three different antibody tags, against different components of the synapse.
Now, how do we fundamentally put all these things together? So, we’re also thinking about computational strategies that you could use, where the morphology information that you would get from a high resolution expansion microscopy picture of the cell membranes, kind of like an electron microscope image except probably at still somewhat worse, a spatial resolution than the best that one can do with electron microscopy (FIBSEM), that was talked about. If you imagine you have a kind of crude, optical approximation of the morphology, but still at a relatively comparable spatial resolution, and you also had barcodes indicating the cell identity, in a unique way. So from a single point, wherever you see a barcode, you could uniquely determine the identity of that point, but perhaps of nowhere else can you use that to synergistically, reconstruct the full geometry of the circuit over longer distances than you can with electron microscopy, or with purely the morphological, imaging.
Also, you kind of take a barcode that you find and fill out from there, and then, go and identify the nearby synapses and identify the detailed shape of the cells that are surrounding that barcode that you’ve identified, maybe comes from some other part of the brain. So, this is also something we’ve been trying to evaluate in computer simulations where we’re starting with data sets that are collected, in this case from Jeff Lichtman’s lab, in a published paper, showing the reconstructed geometry of a small volume of neurophil, using the electron microscope and then simulating what that would look like if we took that same volume and imaged it with expansion microscopy with this 20 x form as it currently stands roughly. In terms of the performance of labeling and resolution that it can get.
We simulate that we are distributing in these neurons, a kind of smattering of Tony’s barcodes in space and that we can do the in situ sequencing process and at the same resolution in the same sample. So, the approach that we’re taking, is basically shown in this series of panels here on this slide where there is a relatively conventional approach that’s used to reconstruct neuro morphology, if in the absence of any barcodes, which is to say that what you do is you first identify the cell membrane. So you have a first stage, and this is often done with a convolutional neural network, where what you do is you just label each pixel. Is it membrane or is it not membrane, or essentially, what’s the probability that any given pixel is on the membrane?
That gives you a probability map showing how likely each pixel is to be a boundary. Once you’ve identified those putative boundaries, you can do a segmentation operation called watershed, to basically try to find what are the contiguous regions, that are bounded by that boundary map. What that tends to do, is to over-segment and create lots of little mini pockets that really should be identified as part of the same neuron, but instead are identified as their own little contained boundaries. So, the question is, is there anything that we can do at that stage to assist the process of merging those pockets back together to identify which actually go together and are part of the same cell?
In this algorithm scheme, that’s exactly where the barcodes come in, because if you attempt to merge together two local units, that have different barcodes in them, then you know that you’ve made a mistake. We would like to also use the DNA that naturally happens in the expansion microscopy process, and actually use that as a barcode to tag different proteins that are present in the sample as well. So, basically using this and building on what Tony said, we really do think that we could eventually achieve a kind of Rosetta brain type comprehensive mapping, where we’d be looking at connectivity with barcodes as well as with morphology, a gene expression by looking at endogenous RNA; using the same sequencing process, looking at many proteins, also using DNA, barcoding and potentially other things like the cell lineage tree which can be encoded in DNA or ticker tapes, as some folks were asking questions. So with that I will close and acknowledge a large collaboration that has mostly done all of this, including Tony’s lab, Ed Boyden’s lab and George Church’s lab, and funding from IARPA. Thanks a lot.
Randal Koene: Thank you, Adam. Thanks very much for this fascinating talk. Let’s give you an applause here.
Okay. So the questions about this, I mean, these are kind of questions for both you and Tony in a way because it’s sort of one big thing. There are lots of different kinds of questions we can ask. Here’s one that happens to reach all the way back to, you know, the origins of the whole workshop as being about mind uploading. This comes from Alexander McLin, and he’s asking if, as brain preservation becomes common, is there a possibility that the preservation protocol could make it difficult to use expansion microscopy techniques, if it becomes an important tool for mind uploading?
A. Marblestone: Yeah, that’s an interesting question. So, I think one thing that I should mention is that the expansion microscopy protocol, that I described, a formaldehyde base chemical fixative, in its current embodiment. So, it’s actually can be done on heavily fixed and preserved tissue. So, for example, this aldehyde stabilized cryo preservation work from Robert McIntyre and colleagues; something that has a fixative, you know, something that can work with expansion microscopy. I think that it relates to a question that was asked to Tony, which is, how would you do the more genetic and barcode based approaches in a human to begin with? That’s obviously open as to whether that’s possible still. I think the more difficult part is that these barcoding technologies rely on viruses to actually replicate inside the neurons and make copies of these barcodes. So, it’s something where the procedure essentially starts while the animal’s alive, but the preservation as such, I don’t think that that necessarily has to inhibit expansion microscopy and we’re working anyway on trying to preserve tissue as well as possible to get the best morphological information out of it with expansion.
Randal Koene: We’re going to go on with a whole variety of different questions. Some may be far beyond what you just presented, but before I go into a question that I thought was a nice tangent, let me just check the audience right here in the room. So Daniel is asking, what does your research reveal about how expansion microscopy changes protein folding, etc.?
A. Marblestone: Yeah, that’s an interesting question. So, it hasn’t yet been done at a level that tries to fully preserve protein folding. So, in the original version that I described, it has proteins in a state where they are chemically fixed and then they are addressed with antibodies to tag them, but then the proteins themselves are digested away, so we know that it preserves because the fixation preserves enough of the protein, confirmations (they’re called epitopes) that the antibodies can still bind, that’s pretty good. Then there have been various follow-on works that are trying to preserve more protein information and be less disruptive to the proteins in the sense of not actually digesting them away. This has not been done yet down at the kind of sub five nanometer length scale. We’re still kind of on the level of size above where we’re looking at detailed protein confirmations. Not to say it’s impossible to look at detailed protein confirmations, but that might be another couple of expansions on top of the two iterations there.
Randal Koene: Thank you. Okay, so a question that maybe wasn’t directly related to the technique you’re showing, but it’s sort of a question about all of this connectomics related stuff that we’ve been discussing right now. Online, one of our viewers has asked the question, how can neuroanatomy be correlated with activity or thought?
A. Marblestone: One of the ways (and Tony started to describe this as well) is that one can try to actually measure both in the same brain. Regardless of any analytical, predictive method for correlating that, you can actually just try to take the time to measure them both in the same organism, in the same circuit. This is something that one can do. It’s actually something that this IARPA project is requiring people to do for a certain volume around the size of a cubic millimeter to try to get a what’s called calcium imaging. So, basically you’re seeing the dynamic movie of the cells, having influxes of calcium as the neurons are firing and the calcium basically triggers greater florescence of a fluorescent dye. So, we’re seeing a movie of the cells turning on and off, and then we can take that and then go and do structural connectome type mapping on it from a kind of analytical, predictive perspective.
I think it’s still an open question. Essentially, what exactly do you need to do to predict based on anatomy, some aspect of the dynamics? You may not be able to predict everything about it, but if there are things like stable attractor states or so on in the circuit, you might be able to at least get enough to know something about how that circuit is going to function or what information it is going to encode even if you won’t in a single trial predict, every ion channel opening or so on.
Randal Koene: Thank you. So I wasn’t sure if Ken is still actively listening to us at this moment, but I couldn’t help being curious.
Dr. Hayworth: I’m still here.
Randal Koene: Okay, great. I couldn’t help being curious, given these two last talks showing the promise of barcoding and expansion microscopy and using optical microscopy. What would you see as sort of the limitations that still make EM an interesting approach? Or where do you see the transition from one technology to another? Maybe Adam can then chime into that, but first I’d like to hear from you Ken if that’s possible.
Dr. Hayworth: Well thanks. Yeah, I’m just here waiting to have them put me out of a job, that’s all. No, I think that these techniques are fantastic. I think in terms of understanding the nervous system, there is probably a place for all of these techniques to attack the exact same systems so that they can be cross checking each other, for example. I think what Adam said to one of these questions though is really my main take on it. Some of these techniques require that the animal is injected with viruses, while is alive so that these barcoding techniques, etc, can have time to express, and that is very powerful in a research context.
I think it would be very difficult to imagine a way for it, although they’ve done a great job explaining one way, for this to potentially work its way into the brain preservation protocol, if you will. So, the way I look at it is that the brain preservation protocol, tries to do the least amount of change because we don’t know what the future technology is going to be and what is going to be necessary. So preserve as much information as possible. If we were to say, go down the route of developing a set of barcoding virus tags that are injected just prior to the brain preservation procedure, which I think is what they were saying, that would mean that we would have to figure out what tags are necessary. What molecular tags, etc., I believe would be necessary today and that might be too much to ask for. So, in any case, to answer your question, I think that these are fantastic techniques and they could vastly accelerate the imaging of the connectome for at least research study, beyond the type of very slow electron microscopy that I was describing.
A. Marblestone: Yeah. So I guess I agree with what Ken said and I would further add that these optical and molecular techniques are still very much works In progress, right? So, we don’t have connectome of anything, using them yet. As with any other biological technique, there are always unexpected hiccups and challenges. So, I think that it’s kind of a proof is in the pudding situation, in terms of the ability to do connectomics with this still, although I think it’s making progress. The other aspect is that we really don’t know that we can get the same spatial resolution, and the same preservation of this super detailed morphology and in the absence of that exactly what you can do is still open. So, for example, if we’re trying to traffic the barcodes to synapses, are they going to traffic to every synapse? If we’re not trafficking the barcodes to synapses and we’re relying on some kind of morphological filling out, starting with a barcode and going out into the morphology, down deep where the synapses are, then we have to make sure that the morphological quality of that is good enough and none of those things are yet established.
So, I think that both because of intrinsic differences, and because of where the technology stands right now, it makes sense that electron microscopy, is still sort of the de facto way to do connectomics and that it should be pushed, even further like in the ways that Ken is doing to try to get a super precise mapping that can allow you to do larger scale reconstructions with that.
Allen Sulzen: That’s all for today. If you’ve enjoyed this podcast, please visit us at carboncopies.org to learn more