Episode 5 | Transcript

Episode 5 | Transcript

Episode 5: “High-Throughput EM and X-Ray Nano-CT”

Published Fri, 26 Oct 2018 | Transcript

Allen Sulzen: Hi there and welcome back to the Carboncopies podcast. I'm Allen Sulzen, your host, and I'm excited to be presenting another episode on our workshop series. This time around, we'll be hearing Dr Shawn Mikula speak on his work, taking serial section Electron Microscopy further than its historical applications. He has successfully imaged the whole mouse brain using these techniques. Their lab hopes to revolutionize cellular-resolution connectomics. If you're interested in our work and would like to contribute as a volunteer, please get in touch. You can find us at carboncopies.org.

Randal Koene: So Shawn has just recently joined the National Institute for Physiological Sciences there and previously worked at the Max Planck Institute. He's been involved in connectomics for a long time and electron microscopy, and so this connects back a bit to what Ken was talking about.

Dr. Mikula: I'm excited to be here and I wanted to thank you, Randal, for putting this together, and the other people involved - it's a great opportunity I think, and I've learned quite a bit even though I wasn't able to catch everything (I had to get some sleep). So I'm going to be talking about high throughput EM, and x-ray nano-CT, of whole mammalian brains for circuit reconstruction, cellular resolution connectomics. And at first the title might seem slightly odd, why I would put together both EM and x-rays in the same talk, right? As will become clear shortly. I think there's very good reason to do that. And, so, to begin my talk, I want to go back in time a little bit. If we go back five years, and we look at the state of the field for volume electron microscopy, there were several landmark papers that were published, a exploiting all of the spectrum of different methods - EM-based methods - for reconstructing neural circuits. So we had serial section EM methods, we had FIBSEM (the focused ion beam scanning electron microscopy, which Ken talked about earlier), and we had serial blockface electron microscopy techniques. And what all of these studies and all these results had in common was that they were using conventional sample preparation for their electron microscopy. And this type of sample preparation always uses this reduced osmium type of stain, which is shown here. So the idea is that you have a mouse brain, you extract a small area of the brain and you put it into reduced osmium solution. And then there were some variations on this of what you did afterwards, and what I show here is what Ellisman and Deerinck published as the ROTO method (or what's commonly known as the ROTO method), where after reduced osmium, you would then amplify the "osmication," if you will, by mixing in TCH (thiocarbohydrazide), and then having a second osmium step and then some additional staining before embedding in plastic and cutting and imaging.

Now, why use reduced osmium? And there's good reason for this. It's because it gives you very strong membrane contrast, which is essential if your interest is in reconstructing circuits, which is what these studies were interested in. Now the problem was, when I tried to apply this method to something bigger, like a whole mouse brain, it quickly became evident that this would not work. That the Reduced Osmium had penetration problems, that it would only penetrate a few hundred microns into a sample. And so you could not use it practically for anything a millimeter in size or bigger. And so this presented a dilemma to me because of course I was interested at the time in applying volume EM techniques to mammalian brains, specifically the mouse brain. And so I had to spend a couple of years figuring out how to modify the existing protocol to get it working in larger samples. And I'll just present the result here.

This is called BROPA, or the Brain-wide Reduced Osmium staining, with Pyrogallol-mediated Amplification. The basic idea was to just first examine the reduced osmium step and just explore what I would call an "augmented" or "expanded" parameter space. And this parameter space included different things that you would add to the solution in order to expect enhanced penetration. So this would include detergents, solvents, various different ions, different things were looked at in buffers. And after the screen I found out that formamide was the missing ingredient. And what's interesting is formamide is a very small molecule and if you look at variations on the formamide molecule, like acetamide, which has the addition of just a single carbon, it does not enhance penetration at all. And so there is something very peculiar about formamide. It seems like it solubilizes things that would otherwise precipitate in solution, and build up to form this diffusion barrier to subsequent, to further diffusion of the stain.

The second thing I did was just modify or replace the tch that was in the standard protocol because this was liberating nitrogen gas and causing ultrastructural damage in large samples. And so this was pretty straightforward. I just did a large screen of all small molecular osmiophilic molecules that were commercially available. And the best one turned out to be a black and white photographic developer called pyrogallol. And structure is shown there. And then the rest of it was pretty much the same as the ROTO stain. And so what was interesting about BROPA is that not only did it work in the mouse brain, but it worked in a variety of small mammalian brains that I looked at. So it worked in the pygmy shrew and also up to the rat. And I didn't explore much beyond the rat, so I can't say how well it scales at this time to larger brains. And on the right hand side I'm just showing corresponding EM, with a high membrane contrast that is typical of reduced osmium staining.

But let's focus a little bit more on the mouse brain and just see whether BROPA is, in fact, suitable for something like a brain-wide circuit reconstruction. So if we look at the SEM through a full coronal slice through the mouse brain, we would see something like in Part A, and we can zoom in on each of the regions, starting from the peripheral part of the brain and just work our way inwards, from E to E. And so we see different structures, the cerebral cortex, white matter, the external capsule, striatum, and also some additional areas that aren't shown in Part A, such as the cerebellum and reticular formation. And all of these showed high membrane contrast and appeared suitable for reconstructing circuits. So this is just showing a fly-through, which I'm sure (based on what I've seen earlier today) will appear choppy to the viewers. But that's okay. It's just showing a fly-through, through the striatum from one of these BROPA brains. And what you can see is a large number of neurites and synapses and quite a few solid bodies, presumably mostly medium spiny neurons - the dominant population in the striatum. And so, of course I wanted to make this a bit more quantitative about the traceability in synapse detection. And so I just went through and randomly selected neurites from different regions of the brain. So one of these neurites that is shown in Part A has been fully segmented along with all of the incident boutons. And I assigned, starting at seed locations, I assigned this neurite to a different people to trace. So this is shown in F, actually. So 20 people redundantly traced it, and so, based on this redundant tracing, you can estimate error-free path lines at creator than a millimeter in the two great regions examined in greater than 70 millimeters in the region of white matter that was examined. And Part E is just showing how consistently people were to able identify synapses, which is greater than 90% for threefold identification. And so all of this looked good in seeing what we wanted, for then taking the next step of imaging the entire brain.

But before I get to that, I want to just talk about looking at larger pieces of how you can actually trace processes through very long stretches, of the mouse brain (in this case). So this is a cortical striatal stack. It's one and a half millimeters high, and a few hundred microns in width. And so basically you could trace neurites through almost the full extent of the stack. Now the stack was also taken at coarser resolution so that we can see synapses. It was basically just limited to tracing neurites and identifying cell bodies. So with regard to now, now that we have a brain, we want to be able to cut it up and image it using high throughput EM. There are different ways we can go. I've mentioned three previously; FIBSEM is ruled out because the limited depth that you can actually mill away before problems occur, which is limited to, I think, 100 microns or less. And so that left the serial block face approach and the serial sectioning approach.

Now with the serial block face approach, I was working for quite some time with others trying to get that to work. Currently, it is not working. However, about one and a half years ago I started to focus my attention on the serial sectioning approach. It seemed like it might be better suited for, cutting through large samples. And so I purchased one of these ATUMtome machines. So this ATUMtome, it's, it's this box on the left hand side. This was based on a design that Ken Hayworth developed some years back in Jeff Lichtman's group at Harvard, and they've been developing this technique extensively. And what it is, is basically a conveyor belt for picking up sections that are produced automatically from commercial ultra microtome, which is shown on the right hand side. Now, when I first attempted this, I explained to the people selling me these machines and other ultramicrotomists what I was intending to do, what I was intending to use their machines for, and the old school ultramicrotomists were just giving me warnings that something terrible would happen if I tried to actually cut through a centimeter-sized sample with these ultramicrotomes. And so, this didn't discourage me at all, but it actually got me quite curious to find out what would happen if I seriously attempted this.

So of course I went ahead and did this. And so what I show here, the sample that's in the sample holder is, is one of these whole mouse brain samples. And if you go through and just start cutting - so this is through the olfactory bulbs of a mouse brain sample - actually it looks to be just a slab and not the whole brain in this case - so this is not the full size of the mouse brain. This is about maybe four millimeters by two millimeters, area that it's cutting through. But you should be able to see both olfactory bulbs in these sections that are coming off. The sections are 50 nanometers in this case. Now the much more interesting thing is what happens when we try cutting through an entire brain. And this is shown here. So this is a full mouse brain at the midsagittal plane. So this is about 15 millimeters high by about a seven, seven and a half millimeters wide, using an eight millimeter diamond knife to do the cutting. And you can see that the cutting is also a much slower here. And you can also see the sections on the conveyor belt on the ATUM. So they have a slightly purpleish color and you can see that they're full sagittal mouse brain sections that are being cut in this case at 80 nanometer thickness. So this was the first time that ultra thin sections had been cut through a full mouse brain. And what I found most remarkable is how easy it was to get this working. And so this made me optimistic that whatever remaining problems there are with regard to section thickness or section quality or number of sections, these are things that would be much easier to solve than trying to get alternative approaches to work. And so I focused my attention, almost exclusively, on the serial sectioning approach.

[pause] So yeah, why do we want a brain on tape in the first place? And I've tried to explain the main advantages here. So, in the images, before I get to the advantages, in the image on the left hand side is just a custom tape collecting machine that I built using aluminum foil as the tape substrate. And you can see a whole brain in the sample holder. And the upper middle image is just a collection of these brain sections. I ended up cutting about 6,000 consecutive sections at 80 nanometers, which is about five percent of the mouse brain and I collected all of those on tape. And the upper right is just explaining that a single mouse brain on tape would be about one kilometer. It, it would actually be closer to two kilometers if you're cutting at, like, 40 to 50 nanometer thickness, which is actually the range that I'm aiming for. And I have 500 meters of aluminum foil tape, which is something like 12 microns thick, shown up there. So you would need maybe anywhere from two to four of those rolls to have the whole mouse brain on it for imaging.

Now, why? Why do we want the brain on tape? So I list three advantages, advantages relative to other volume EM techniques such as the serial block face technique. So for instance, it's a nondestructive technique that allows for random access volume EM. So nondestructive. I mean, once the brain is on tape, it's always on tape. You can go back and re-image it at higher resolution if need be, if there's some ambiguity you're trying to resolve it's not a problem, whereas with FIB and SEM, once you a cut or otherwise vaporize or ablate your sample, it's gone forever. You can never go back to it. If you're going through the data analyzing it and you see something that you're unsure about and you want to go back to reimage, you're stuck. You can't do it. So the fact that the brain is on tape and it's always on tape and allows for repeated imaging, is really a big selling point. And the second point is the random access. You don't - I mean, if you want to go from section one, for whatever reason, jump to section 1001, there's not a problem accessing that section. Whereas in the destructive techniques, you would have to actually cut through all the intervening sections (and presumably image them) before getting to section 1001. The random access means that you don't need to image the entire, you don't need to image everything, you can just go to the regions that you're interested in, assuming the sections are properly indexed on the tape, which is pretty straightforward to do. The second advantage is that there are two modes of imaging the brain on tape. One is just do automated bulk imaging of everything on the tape using high throughput EM. You can do that by either using multi-SEM, use multiPMS EMs, or you can just cut the tape up in a bunch of strips and send each one off to a bunch of single beam or multi-SEMs if you want to increase the throughput. But the other mode is do sparse imaging of the sections for efficient reconstructions of just certain sub circuits that you're interested in, which is compatible with the lower throughput single beam EM. You don't have to image the entire tape to get at brainwide circuits. And the final thing is this nanoscale molecular interrogation. So I'm thinking specifically of immuno-EM labeling, not so much of Mass Spec because this mainly tells you elemental analysis and there're other problems. Now with immuno-EM, there is loss of immuno reactivity, so this is something that would have to be troubleshot; I mean the preparation for EM causes lots of problems.

So this is just showing some of the results for the mouse whole-brain serial section EM that I've attained so far. So this in Part A is the micro CT with the slice at about the region that I was collecting the sections from. And Part B is just showing three consecutive sections. You can see section quality needs, still, considerable work. There were wrinkles in that first section. There's chatter in the second one. Chatter is one of the big issues that has to be solved, but I have some tricks up my sleeve for that one. And then in C and D it's just showing higher resolution EM. And so with this being said - one final thing is that, of course, the brain on tape, and all these brains, are compatible with the multi-SEM. So this is actually showing blockface imaging with a 91v multi-SEM. The contrast is actually much higher with thin sections, like as you would have with a brain on tape.

Now I've sort of gotten through the EM; I don't have that much time, so I want to just quickly say something about the nano-CT. So I promised to explain why there's the relation between the two. If you have a sample that is suitable for EM, x-rays will also interact with those metals and you can use such a sample for x-rays without any problem. And in fact, EM samples are very good samples for x-ray nano-CT. So I'm going to show you some results from my collaboration with Bobby Kasthuri, Rafael Vescovi, and others at the Advanced Photon Source at Argonne National Labs. And this is showing pygmy shrew, just a single slice, midway through the brain. I'm not sure what resolution you guys can see on your end, but what I see is a bunch of little dots which are cell bodies. Now if you zoom in on that yellow box, and just do a fly through of about a thousand sections, you would see this. So the resolution here is sub-micron. It really is nanoscale, it's closer to half a micron. And so this is showing cortex and the upper part. And then at the very bottom right, you see the external capsule, the white matter coming into view. Now that's fine. But what's interesting is - so this is showing, if you use a lookup table and just color code pixels and do a projection-type image. So this is just identifying the cell bodies in blue and the vasculature in red. And again, this is not using any sort of machine learning or anything - there's nothing here except using a lookup table to color code things according to the gray scale intensity level. And so this is promising, this is interesting, but the most interesting part of it is shown here. The fact that if you look carefully at this data in the white matter, if you rotate that stack a little bit and look at it so the external capsule is at you, is facing, you, and you adjust your lookup table correctly, you see something like this; and now you see the telltale morphology of blood vessels in the upper part. But if you look at the lower part, you can see bundles of tube-like structures, which I think can only be one thing. And those are large-caliber axons, the vast majority of which would be myelinated axons.

And so this is what's currently being investigated now. The resolution is certainly there to permit identification of large axons. And if you can see and trace large axons throughout these entire brains, then while you might not have the complete synaptic level circuit map, what you would have are complete projectomes for various mammalian brains. And this is, I think, very exciting.

So I would like to conclude by saying that there are two methods that I presented that appear suitable for reconstructing brainwide circuits: serial section EM and x-ray nano-CT for the serial section EM. We've been developing the whole brain on tape for nondestructive random access EM and x-ray nano-CT as a promising alternative, but resolution needs to improve by 10 to 20 x before synapse resolution, before synapse-level circuit reconstructions are possible. Nonetheless, net-net, the x-ray nano-CT has the possibility, if you can get the 10 x improvement in resolution of really, I think, revolutionizing the whole cellular connectomics, or cellular-resolution connectomics, because it is a very fast method. Like for the whole brain on tape, if you cut it off, and you talk about imaging times using a multi-SEM, the optimistic numbers come out to one, one and a half years. X-ray nano-CT, the actual scanning itself with the x-rays, is a matter of minutes. And then what you have are a series of projections that you have to use some sort of algorithm to back-project. And this is something that you run on a cluster that might take a few weeks, and so we're talking about a few weeks for nanoscale whole brain data, versus over a year if you go the EM route. And so, if nano-CT can push the resolution a little bit, it's going to make cellular resolution connectomics commonplace. It's going to make it easy.

And so I want to just acknowledge several people involved, several groups that I've had the pleasure of working with and collaborating with located across Germany, Japan, and the US. I'm not going to go into all the details. I think what excites me is that all the pieces are in place to actually do a connectomic attack on a small mammalian brain. And so with proper resource allocation, this is something that could be done starting today. In five years time, starting with the cutting, the imaging, the annotation, and the complete circuit reconstruction all be done in five years.

So thank you very much for your attention and I'm happy to take any questions.

Randal Koene: Yeah. Thank you very much. Shawn. First of all, it's time for a little applause. Yeah, thanks. This was great. We're going to have time for maybe two or three questions because we have a rather full schedule coming right up. So Peter is asking, when you're doing whole mouse brains, why is the knife so much slower and if it is so much slower, why not just take multiple brains slabs?

Dr. Mikula: So the reason for the slowness is because cutting force is proportional to the edge of the sample that it's cutting through. These mouse brain samples have anywhere from 20 to 40 times the cutting forces required for them, versus the standard small samples that you can normally cut faster through. And so you have to slow down the cutting to compensate for that. And if you don't, what happens is chatter problems. So the slow cutting speed is because the sample is bigger, and you want to minimize the onset and the propagation of chatter throughout the sample. Now regarding why not cut it up into slabs, I think you mean like a hot knife approach or something. But if you do something like hot knife, then you're tied down to using an ion-beam milling approach, and this is something which, yes, it would be possible, but it still has to be demonstrated that it's possible. And with the ion-beam milling, it has not yet been demonstrated that you can actually do that through thick slabs through the entire mouse brain. If demonstrated, then yes, that would be interesting to do that, to just cut through slabs through the mouse brain at 10 microns and have each of those ion-beam imaged - not with FIP, but with one of these broad ion beams like Kend was talking about.

Randal Koene: Daniel is asking: how do you adhere the sections to the tape or did they just naturally adhere or is there some kind of glue?

Dr. Mikula: They naturally adhere. So there seems to be a preference for hydrophilic substrates for the adherence. So if you have a hydrophilic substrate, it tends to be more wrinkle-free. So, with pure hydrophobic substrates, I don't think I have tested those, but with every tape substrate I have tested the sections stick without a problem - and these include metal foils, various various PET and Kapton, and ones with and without different types of coatings on them. The adherence has never been a problem.

Randal Koene: I'm going to give you a question now that came in through the online forum. Leslie Seymour asks, is there any similar work for mapping the second in brackets, gut brain? A) A more manageable size? Like 100,000 neurons? B) A huge impact on the affective self? C) Ninety percent of DNA in the body is microbiome? D) Huge recent interest in commercialization and investments? E) Ease of impact, e. g. a diet change? So basically is anything happening with the gut rather than the brain, as such?

Dr. Mikula: With EM, not that I'm aware of. With x-rays, yes. There is some work there. But I don't know people doing volume EM that are looking at the gut.

Randal Koene: Thank you very much, Shawn.

Allen Sulzen: That's all for today. Thanks for listening. If you'd like to learn more, visit us at Carboncopies.org.