Single-Cell Imaging: Imaging and Reconstructing Mouse Development at the Single-Cell Level

Top

00:00:12.06 Hi.
00:00:13.06 My name is Philipp Keller, and I'm a group leader at the Janelia research campus.
00:00:17.00 And a key focus of the research in my lab is the development of new microscopes
00:00:23.09 and computational methods for image analysis.
00:00:27.02 And our efforts in technology development to this end are also motivated by our desire
00:00:34.19 to investigate and understand a fundamental question in biology.
00:00:39.15 And this question is, how does an entire functioning embryo form from a single cell?
00:00:47.06 Now, an embryo is a fairly complex biological system
00:00:50.20 that comprises numerous tissues and organ systems,
00:00:56.06 which all have fairly intricate three-dimensional architectures, tissue architectures,
00:01:02.02 and also remarkable functional capabilities.
00:01:06.07 So, how does something like this form from scratch?
00:01:11.02 How do the cells in the embryo divide, move, interact with each other, change their shape,
00:01:18.06 differentiate, to gradually give rise to this, you know, complex embryo?
00:01:24.06 And given that there are so many cells forming during early embryogenesis,
00:01:29.11 how do these myriad cells coordinate their dynamic behaviors?
00:01:34.05 Now, there are many aspects of this process that we don't really understand yet.
00:01:39.01 We do know that it's certainly a highly dynamic process at many different scales.
00:01:43.18 And so, in order to work towards a comprehensive understanding, it is an important prerequisite
00:01:51.14 that we find a way to visualize and quantitatively analyze this dynamic behavior,
00:01:57.22 for all of these cells, as an entire embryo develops.
00:02:01.28 And so, we've always been curious to find a way to study embryonic development
00:02:06.25 at the single-cell level in a mammalian embryo so that we can gain insights, make observations,
00:02:15.06 in a biological model system that is relatively close to human.
00:02:20.07 And so this goal led us to the mouse embryo.
00:02:23.16 Because in the mouse embryo, if you want to follow the formation and development of
00:02:28.02 the early tissues and organs, from this initially pluripotent ball of precursor cells,
00:02:35.16 there is a 48-hour time window -- shown down here -- from embryonic day 6.5 to 8.5,
00:02:44.05 that is of critical importance.
00:02:46.04 During this period of time, the germ layers are forming.
00:02:49.22 So, we have the mesoderm and the endoderm and the ectoderm emerging.
00:02:54.18 And then these germ layers, in turn, give rise to the early tissues and organs.
00:02:59.05 So, by the end of this 48-hour time window, we have an embryo that has, for example,
00:03:06.03 a neural tube, it has somites, it has an early stage gut, and it even has a beating heart.
00:03:14.02 Now, the tricky part is to find a way to actually measure what all of the individual cells
00:03:21.01 are doing in the embryo during this process.
00:03:22.27 And in principle, if you want to do live observations and live measurements of cellular dynamics
00:03:31.10 at a high spatial and temporal resolution, light microscopy is a very powerful technology
00:03:35.28 for that purpose.
00:03:37.10 However, so far it's not impossible to image cellular dynamics
00:03:42.24 across an entire developing embryo at the single-cell level.
00:03:47.21 And so there are a number of challenges and difficulties associated with that goal.
00:03:53.17 In general, it is... and it's not trivial to image mouse embryonic development,
00:03:59.25 because the mouse embryo is very sensitive to light, which of course we have to rely on
00:04:04.22 in the light microscope to do the measurement.
00:04:06.22 It also needs to be cultured... and it's a sophisticated culturing protocol, you know,
00:04:13.26 to be kept alive and healthy in the microscope.
00:04:16.19 Now, if we want to measure cellular dynamics and do live imaging of mouse development
00:04:22.24 at the single-cell level over this entire 48-hour time window, we have added challenges.
00:04:28.28 Because, now, the embryo also grows by a factor of about 200 in volume over this period of time,
00:04:34.17 and it has fairly complex optical properties.
00:04:39.12 So, for these reasons, we set ourselves several goals in this study.
00:04:45.06 The first goal is to develop a microscope that would enable us to do these types of measurements,
00:04:51.11 to image mouse embryonic development at the single-cell level...
00:04:54.19 through gastrulation, early organogenesis, at the single-cell level.
00:04:59.01 And then secondly, to use the resulting image data in order to analyze how cells
00:05:04.00 divide, move, and essentially, assemble in these early tissues.
00:05:09.18 And then, considering that there are a number of resources that would come out of such an investigation,
00:05:15.25 and also technology that needs to be developed, we also want to make sure
00:05:19.02 that all of these resources and methods would be publicly and freely available
00:05:23.27 to the entire scientific community.
00:05:25.04 That would be the third goal.
00:05:28.14 So, starting with the first goal, we needed to find a way to develop a microscope
00:05:34.02 that allowed us to do long-term, high-resolution imaging of mouse embryonic development.
00:05:40.10 And so a key idea... a key concept in the microscope that I'm about to show you,
00:05:46.10 and in fact in pretty much all of the microscopes we've developed in the lab so far,
00:05:50.21 is the idea of using a light sheet for live imaging.
00:05:56.11 Now, light-sheet microscopy is a... it's an idea that's been around for a long time.
00:06:02.07 The concept of light-sheet microscopy dates back more than 100 years.
00:06:06.07 But over the last 15 years or so only has it really found its way into live imaging in biology.
00:06:13.28 And now, in these modern forms of light-sheet microscopy, what we would do is to take the sample,
00:06:21.13 which is shown here as a cylinder in this conceptual illustration,
00:06:26.08 and illuminate it from the side with a thin sheet of light, of laser light.
00:06:31.06 And so this light sheet is only a few micrometers thick.
00:06:33.09 So, it's a thin optical section that we've formed with the illuminative light.
00:06:38.08 And then we would detect the fluorescence that is emitted by molecules located
00:06:43.18 within that illuminated plane, and form an image of that... of that fluorescent light.
00:06:48.23 So, for that purpose, we use a detection arm with a detection objective that is
00:06:53.12 facing that illuminated plane.
00:06:54.16 And we have a camera, further downstream in that arm, that takes the image.
00:06:59.26 And so, with a camera, that means immediately we can take... you know, we can...
00:07:03.17 we can acquire information for millions of pixels instantaneously.
00:07:06.16 We also have some easy means to move the light sheet relatively quickly in relation to the sample,
00:07:16.07 and thereby acquire an image sequence that represents the entire volume of the specimen.
00:07:21.11 And so that means we can do volumetric imaging at the rate of
00:07:25.02 hundreds of millions of pixels per second, if we use a state-of-the-art camera.
00:07:29.08 So, high speed... high imaging speed is one of the key strengths of light-sheet microscopy.
00:07:36.02 But there's another key advantage that's arguably even more important.
00:07:41.17 And so, light-sheet microscopy are also very gentle methods... imaging methods.
00:07:46.11 They minimize the amount of light that the sample is exposed to in order to get the image data.
00:07:53.07 Now, let me show an example to explain that.
00:07:57.10 In a conventional confocal microscope, if the image... a plane, you would use a single objective that...
00:08:06.22 it would first focus a light beam into the sample,
00:08:09.12 and so the focus would be formed somewhere in that plane of interest.
00:08:13.07 And so, the fluorescence light that is then emitted by the sample would be detected through...
00:08:18.08 it would be collected by the same objective, and then would form an image.
00:08:23.15 And so, that arrangement and that approach means that even just to record this one single plane,
00:08:30.02 we need to send light through a large fraction of the... of the volume of the sample.
00:08:35.06 Basically, the light beam travels through regions above the focal plane,
00:08:39.00 and then it diverges again, and so, you know, moves through regions below the focal plane.
00:08:44.02 So, that introduces the risk of light exposure, and certainly... of light exposure for sure,
00:08:50.08 but also potentially damage that is mediated by the light in this large fraction of the volume.
00:08:57.01 And so this is fundamentally different, now, in the light-sheet microscope,
00:09:00.08 where the light is really confined to that plane of interest, because we don't illuminate any out-of-focus structures.
00:09:06.09 And therefore, also, we can't really damage them.
00:09:09.00 And generally it just reduces the amount of light inside the specimen.
00:09:13.11 And so that's a striking difference, now.
00:09:15.12 And particularly in volumetric imaging, because you can see from this comparison that
00:09:21.08 with the conventional approach, if you look at the cumulative light load,
00:09:25.21 we would have a linear increase in the energy exposure in any local region in the sample,
00:09:31.19 a linear increase with the size of the volume and with the number of image planes in the volume.
00:09:36.16 Whereas in the light-sheet microscope, we have a constant light exposure... local...
00:09:42.11 a local light ex... light load that is independent of the number of planes in the volume
00:09:46.19 that we are recording.
00:09:48.18 So, a light-sheet microscope... microscopy is particularly useful for live imaging
00:09:54.08 of relatively large samples for that reason.
00:09:57.16 And that includes, certainly, large developing embryos, such as a mouse embryo.
00:10:03.00 And so based on this idea of light-sheet imaging, we developed a multi-view light-sheet microscope.
00:10:09.16 And so the idea in the multi-view light-sheet microscope is that we illuminate the sample
00:10:13.23 with two light sheets that come... that enter the sample from opposite sides.
00:10:20.13 And then we also have two detection arms to record the images.
00:10:23.18 We can acquire opposing views of the sample simultaneously.
00:10:26.22 And so this is useful when imaging a large sample that may not be completely transparent,
00:10:32.28 because then the light basically has to travel through the sample and, you know,
00:10:38.06 encounter some amount of... subject to some amount of light scattering, light absorption, and aberrations.
00:10:43.17 And so, with the multi-view imaging, acquiring these multiple complementary views
00:10:47.20 from multiple different viewpoints,
00:10:50.12 we can simply increase the fraction of the sample that is covered in high resolution,
00:10:56.05 in high image quality, in the experiment.
00:11:01.01 And so, if you look at a photo of this microscope, now, as it actually exists in the lab,
00:11:07.09 you can see the incubation system, here, in the center, and then the sample chamber
00:11:13.01 would be in the inside.
00:11:15.04 So, this is how we then can control the temperature and the atmosphere.
00:11:19.13 That's critical for keeping the embryo healthy and alive.
00:11:23.17 Inside the sample chamber, you would find this glass cylinder that we use for sample preparation.
00:11:29.01 So, that's a cylinder that has about a three-millimeter diameter
00:11:33.27 and is filled with an extracellular matrix called Matrigel.
00:11:37.26 And so we would anchor the embryo, at the location of its ectoplacental cone, in that matrix.
00:11:43.04 So, that provides mechanical support.
00:11:46.04 But at the same time, because the cone is an extraembryonic structure
00:11:49.27 that we don't really need to image to see what the embryo itself... what's happening in the embryo,
00:11:54.04 we have no optical constraints on the embryo.
00:11:58.04 We have a clear view of the embryo itself.
00:12:00.20 And there's a Teflon tube, a transparent tube, surrounding the embryo to provide
00:12:06.09 a small compartment for the culturing medium to keep the embryo healthy and alive.
00:12:11.10 And so, the bottom line is that, with this arrangement, we have mechanical support.
00:12:15.02 The embryo is not really floating around in space.
00:12:17.21 It's moving just a tiny little bit.
00:12:20.17 And there's no mechanical constraint on its growth.
00:12:23.04 It's important for this 48-hour time window that it can grow so much.
00:12:26.20 And at the same time, we have multi-view axes.
00:12:29.08 We can really image it from all sides without blocking the view of the sample.
00:12:34.06 So, with what I've shown you so far, we basically have, now, a microscope.
00:12:39.27 We can culture the embryo in the microscope, keep it alive to do the multi-view imaging.
00:12:43.26 But there's one more important problem that we need to solve.
00:12:47.27 And so, one challenge that I, you know, haven't told you about yet is that because the embryo
00:12:52.17 is relatively large and, you know, that concerns, really, most, you know,
00:12:57.19 if not all large multicellular organisms... certainly the mouse embryo...
00:13:02.01 it is large enough and complex enough in terms of its optical properties
00:13:06.05 that it fundamentally perturbs the light that we have to rely on
00:13:09.10 in order to acquire images in the light microscope.
00:13:13.06 So, you know, the way we can think about this is that inside the embryo we have, you know,
00:13:21.05 some amount of varia... variability of optical properties.
00:13:25.07 So, you know, as we look at different parts of the... of the embrdyo, different local...
00:13:31.04 you know, local optical constellations ... different refractive indices, for example,
00:13:35.26 in terms vary with the local concentration, the local density of cells,
00:13:39.17 or changes in the biochemical composition in local domains...
00:13:44.03 and that all constitute some amount of optical heterogeneity in the specimen.
00:13:48.23 Likewise, the optical properties of the sample can vary not only, you know, inside the sample,
00:13:56.14 but also if you compare the sample to the surrounding culture medium.
00:14:01.02 We might have a mismatch there, with some different optical properties as well.
00:14:05.04 And so all of these factors contribute to aberrations in the imaging process.
00:14:10.23 And that means, for example, for the light sheets, that it will change their paths
00:14:18.18 depending on the geometry of the surface of the embryo, and depending on exactly how they...
00:14:25.07 how they enter the sample and, you know, what the relative angle is... relative to the surface.
00:14:29.09 They can change their path, and they can change the path even further as they then travel
00:14:32.26 through the embryo.
00:14:35.04 And likewise, for the detection focal planes, we'll see an increase in curvature as we
00:14:39.01 go deeper and deeper inside the sample.
00:14:41.00 And so that means that the illumination planes and the detection planes are spatially mismatched.
00:14:46.05 So, we effectively acquire images out of focus.
00:14:51.06 Now, what's even worse is that the situation is also changing as a function of time, because,
00:14:56.28 you know, the embryo is not a static structure, but as I've mentioned before,
00:15:00.17 it grows in size quite a bit and changes its shape.
00:15:03.26 So, to address this problem, we designed the microscope such that it will adapt itself
00:15:10.11 to the sample.
00:15:12.04 And so that involves the following concept.
00:15:14.13 So, first of all, we introduced a number of degrees of freedom that allow us to...
00:15:22.00 so, these are digitally controlled, hardware-based degrees of freedom with... you know,
00:15:27.10 we have optical modules in a microscope to introduce them... which allow us to change the geometry
00:15:32.27 and the relative three-dimensional orientation of light sheets and detection focal planes.
00:15:37.27 And then we have a control software that is fully automated.
00:15:40.27 It takes advantage of these degrees of freedom and uses them to map out the optical properties
00:15:45.21 of the sample.
00:15:46.21 So, it will basically use software modules that are automated and... you know,
00:15:56.18 that try to determine and then also apply the optimal optical settings, across the imaging volume,
00:16:03.28 that are required to get image data of the optimal... of the highest quality,
00:16:10.01 of the highest resolution.
00:16:11.06 So, as shown here in a little bit more detail, we have the advanced optical implementation
00:16:15.19 to get the degrees of freedom, and then we have all of these software modules on the right side
00:16:19.11 that do the real-time processing... image processing and data analysis.
00:16:25.26 And so by using this kind of adaptive imaging approach and optimizing the microscope
00:16:31.18 relative to the properties of the sample, we have the following improvement, now, in image data quality.
00:16:37.10 As an example, for two different timepoints in development,
00:16:40.13 you can see the uncorrected state, here, on the right side,
00:16:42.24 the corrected state of the microscope on the left side.
00:16:46.01 And so where we have, basically, a blurry... a fairly blurry image in the uncorrected state,
00:16:50.27 we can now recover high spatial resolution, and essentially
00:16:55.02 -- and that's the important part --
00:16:57.08 the ability to reliably distinguish individual cells -- in particular, neighboring cells --
00:17:01.20 here in the... in the corrected state.
00:17:03.15 So, across the different germ layers in the embryo, we have high resolution again.
00:17:09.11 So, by combining all of these features into a single instrument, into a single microscope,
00:17:14.09 we basically, now, have the tool that we need in order to
00:17:19.00 perform long-term high-resolution imaging of the mouse embryo at the single-cell level.
00:17:24.27 And so, here's an example of an imaging experiment performed in this way, where the mouse embryo
00:17:29.28 expresses a fluorescent reporter in the cell nuclei.
00:17:33.04 So, each little blob here, in this video, is the nucleus of one cell.
00:17:37.27 And we can see a number of striking, dynamic processes in this recording,
00:17:43.27 and interesting structures that are forming.
00:17:47.04 So, early on in development, at the beginning of this 48-hour time window, you can see the
00:17:52.28 amnion inflation.
00:17:53.28 So, the ballooning of the embryo, so growing very quickly, popping out.
00:17:58.03 Then, there's the node forming over here.
00:18:01.20 It's moving over, further to the posterior side.
00:18:04.11 There's a big gap here on the left side, which is... there's a little hole that is...
00:18:09.24 that represents the anterior intestinal portal of the embryo.
00:18:13.01 And then, from the left over to the right, we have the midline,
00:18:17.20 the anterior-posterior axis, of the embryo.
00:18:20.20 So, along this line we would probably find a neural tube, then, at this time point.
00:18:25.20 And then, on the left and the right side in the embryo...
00:18:28.02 on the left and the right side of the midline, we have the somites forming.
00:18:32.13 And then, very late in this recording, particularly now, at this last timepoint, you can see
00:18:36.16 a dynamic striping pattern here, on the anterior side, which is produced by
00:18:42.12 the onset of the heart beating.
00:18:44.09 So, while we were watching this video, the embryo actually formed a heart.
00:18:50.08 And if you look deep inside the embryo, which is possible, for example,
00:18:54.23 by using state-of-the-art near-infrared fluorescent reporters,
00:18:59.17 we can see that many of these very deep structures
00:19:01.05 are still visible at the single-cell level.
00:19:03.11 This is now a view from the anterior side, with the camera facing the anterior end,
00:19:06.15 and we're imaging fairly deep, you know, several mic... hundred microns deep inside.
00:19:10.20 You can see, slice by slice, the image volume that shows the beating heart.
00:19:15.08 Going basically all the way through that heart, pretty much at the single-cell level.
00:19:19.28 And there's one more trick that we can use to improve image quality,
00:19:23.17 even beyond what I've shown you so far.
00:19:26.16 And so, the idea here is that, instead of just acquiring the opposing views of the sample,
00:19:33.11 which are provided by the two opposing detection arms, we can rotate the embryo by 90 degrees,
00:19:38.28 mechanically, after each volume acquisition, and so record a second set of opposing views
00:19:44.16 with the same detection arms, but now at an angle that is perpendicular to the set of
00:19:48.26 the initial opposite views.
00:19:52.14 So, that means we have recorded four views, in total, that are all orthogonal to each other.
00:19:58.14 So, a set of four views.
00:19:59.17 And the advantage of this approach is illustrated here.
00:20:03.27 In the upper panel, you can see, you know, a particular image stack, now shown with the two opposing views.
00:20:11.23 In a light microscope, we would typically have anisotropic spatial resolution.
00:20:17.00 And so that means, for this view, that lateral resolution is typically significantly higher
00:20:22.16 than axial resolution in the microscope.
00:20:25.00 And so by looking at the side view of the volume, it's clear that the image is a...
00:20:30.01 it's a bit more blurry along this... along this axis, versus the lateral view...
00:20:37.07 versus the lateral dimensions in the... in the XY view.
00:20:41.00 But if we record the four orthogonal views, we can now combine all of the image data
00:20:45.18 into a single volume, and basically, by reconstructing that image dataset that combines
00:20:52.26 information gathered from all views, we can now obtain high resolution along all three axes simultaneously.
00:20:57.25 So now, then, if you look at the same image volume... basically, with this four-view imaging strategy,
00:21:04.02 you can see it's much crisper, particularly the higher axial resolution.
00:21:08.15 And so that means it becomes even easier to track, for example, cells in regions
00:21:15.03 that are densely packed with cells.
00:21:19.10 And so that brings me to the second goal.
00:21:22.10 Now that we have these image datasets, these time-lapse recordings of mouse embryonic development
00:21:29.17 at the single-cell level for this entire period of gastrulation and early organogenesis,
00:21:34.16 the next step is to really find a way to convert the raw image data
00:21:39.06 into a biologically meaningful representation,
00:21:42.02 to extract useful information about the dynamic cell behavior in this recording.
00:21:48.07 And so, what's challenging is of course that the dataset it's not only fairly complex
00:21:53.03 as far as, you know, the complexity of the embryo is concerned, but it's also fairly large in size.
00:21:59.27 So, we have about 10 terabytes of image data for every single embryo that is recorded.
00:22:04.14 And the embryo in that video grows from a size of about 1000 cells to...
00:22:08.19 all the way up to about 50,000 cells towards the end of that recording.
00:22:11.24 And we have hundreds of timepoints that capture the dynamic behavior of these thousands,
00:22:16.08 and then eventually, tens of thousands of cells.
00:22:18.24 So, now, to analyze this, and to extract it, we developed a computational pipeline that
00:22:25.10 essentially allows us to do everything from basic steps,
00:22:28.05 such as basic large-scale data management and multi-view image processing,
00:22:36.26 down to the crucial steps of image segmentation, tracking of cells, detection of cell divisions,
00:22:44.01 and also registration of multiple embryos in space and time.
00:22:48.27 And so, over the next few slides, I'd like to show you a couple of examples of
00:22:52.00 what these methods are capable of.
00:22:53.20 And then also some examples of biological analyses that are enabled by these tools,
00:22:58.10 and what we can learn from the data.
00:23:00.08 So, one of the methods that we developed is the cell tracking framework TGMM,
00:23:07.08 which stands for Tracking with Gaussian Mixture Models.
00:23:11.05 And an example of tracking results obtained from one embryo using TGMM is shown down here,
00:23:16.15 in the lower... in the lower row.
00:23:18.08 So, in this video, we use a color code that represents the movement speeds of the cells.
00:23:24.22 And now, because it's an automated method, it's very convenient and fast to use,
00:23:28.09 but it still makes some mistakes.
00:23:30.15 So, linkages might not always be correct in time, and you could also lose cells temporarily,
00:23:37.03 and therefore disconnect cell tracks in time, which makes it harder to analyze
00:23:41.04 the dynamic behavior of the cells.
00:23:42.08 So, therefore we developed a second method, a statistical method,
00:23:45.27 that is complementary to TG... to TGMM, and processes the output of TGMM.
00:23:51.12 And it takes that data, and then, in a statistical way, examines the neighborhood of each cell
00:23:59.00 in the developing embryo, and looks at the dynamic behavior, and uses this information in a...
00:24:04.24 in a statistical way that allows us to correct for many of the most common mistakes
00:24:09.11 that happen during automated tracking.
00:24:12.08 And in this course, also ensures that cell tracks are continuous in time,
00:24:17.00 from the beginning to the end of the recording.
00:24:19.04 That is much more useful, then, for the data analysis.
00:24:22.24 And so, if we combine TGMM and the... you know, the statistical method,
00:24:27.26 and produce a tracking database, we can, in the next step,
00:24:32.00 use this information to build a dynamic fate map of the entire embryo.
00:24:35.04 So, you can think of this as the developmental building plan of the mouse embryo.
00:24:40.26 The way this works technically is that we take the last time point of the dataset,
00:24:46.24 the image data at the last time point, where we know that the cells have already assembled
00:24:50.14 into tissues and organs
00:24:53.00 -- we can actually anatomically identify these structures just from looking at the last timepoint, fairly clearly --
00:24:57.28 and thereby annotate which tissues cells contribute to.
00:25:02.28 So, each cell... basically, we have that information about whether it's part of the neural tube,
00:25:08.04 the notochord, or the heart field, and so on.
00:25:10.22 And so, then, from annotating this one timepoint, we have tissue labels that we can
00:25:16.01 associate with the tracking database so that each track in the database has a label linked to it.
00:25:22.26 And with that information, we basically know immediately what the developmental origin of
00:25:27.09 each tissue is, where it came from... you know, where it started out.
00:25:31.25 But also, we really have the entire sort of timeline of the development of this tissue in detail,
00:25:36.02 and in high spatial and temporal resolution.
00:25:39.22 Now, there's a lot we can do in terms of analysis, now, with a dataset like this.
00:25:45.22 And you know, even just from looking at the... at the results by eye, looking at this video
00:25:49.26 of the... of the tracking database, there are some interesting things that stand out.
00:25:55.02 You can, for example, see that already fairly early on there's a remarkable level of separation
00:26:01.25 -- not very much mixing of cells of different cell types, of precursor cells for different tissues --
00:26:08.03 once these cells have emerged from the primitive streak.
00:26:11.14 But I would also like to point out that this is now reconstruction from a single embryo.
00:26:16.06 And so we don't really know how variable development is, and to what extent, you know,
00:26:20.13 this particular reconstruction, this particular embryo, can be considered to be representative
00:26:24.12 of mouse development.
00:26:25.12 In fact, we do know that embryos of the same age actually vary quite a bit in terms of
00:26:32.08 their size, their shape, and their developmental stage.
00:26:34.19 You now can see a couple of images over here.
00:26:37.12 Just by visually looking at them, it's clear they're not geometrically the same.
00:26:41.05 If you look more closely, from an anat... anatomical perspective,
00:26:43.14 you can also see the stage is not... it's not really perfectly matched.
00:26:47.18 So, for that reason, we decided to reconstruct development in multiple individuals.
00:26:52.24 You can see four here, with that same dynamic fate map visualization from the previous slide.
00:27:00.16 And then use this information from the multiple individuals in order to reconstruct
00:27:04.11 a statistical representation of their development.
00:27:07.21 So, for that purpose, we developed a method called TARDIS that allows us to register,
00:27:13.15 in space and time, the development of multiple embryos using anatomical landmarks.
00:27:19.26 And then once they're registered in space and time, we basically can now analyze
00:27:24.28 the local dynamic behavior of cells in all sorts of different parts... all different parts of the em...
00:27:29.04 of the embryo, while looking at corresponding spatial regions and at corresponding times
00:27:35.17 in development across individuals.
00:27:40.02 And so, equipped with this tool, we made a transition from a dynamic fate map
00:27:46.15 from one animal to a statistical fate map constructed from multiple embryos.
00:27:51.12 And that's the result shown here.
00:27:53.08 So, we basically have, now, the information of the probability that cells in a local...
00:28:01.04 in some, you know, local domain in the embryo will give rise to a certain type of tissue.
00:28:06.19 And we have that information as a function of space and time.
00:28:10.06 And so that means the color code, now, here, is not just solid colors, as it was before
00:28:14.20 -- where we had each color as one tissue -- but we also have a little bit of shading,
00:28:18.08 which then represents this probability.
00:28:19.08 So, there can be high probabilities to contribute a certain tissue or low probabilities,
00:28:23.20 and probabilities to really contribute to multiple tissues at the same time as in the statistical analysis.
00:28:29.22 And so, then, if we go to these fairly low probabilities, you can see how the colors
00:28:34.03 all fade into gray.
00:28:34.26 And that means that, in the video over here, we can now see the boundaries between tissues
00:28:40.09 in this gray color, statistically speaking.
00:28:44.28 We also use the TARDIS method to build an average mouse embryo.
00:28:50.15 And so, this is now a virtual mouse embryo that represents... exhibits the average statistical
00:28:57.06 shape of mouse embryos: their average, you know, cell density, cell counts,
00:29:03.14 and even their average cell movement patterns, you know, locally throughout the embryo.
00:29:08.00 And so, by reconstructing this average embryo, we also immediately have access, then,
00:29:12.11 to the information on variability of developmental processes.
00:29:16.12 And we can also use this information to measure how and to what extent an embryo
00:29:21.18 with perturbed development, such as a genetic mutant,
00:29:26.28 differs from wild type... you know, deviates from wild type development.
00:29:31.08 In principle, all we need to do for such an analysis is to use the same imaging tools
00:29:36.13 to image the perturbed embryo -- the development of the perturbed embryo --
00:29:40.02 and then use TARDIS to register it onto this reference atlas of mouse development.
00:29:49.02 And so, complimentary to this information on cell movements and cell fate, we also reconstructed,
00:29:55.02 systematically, cell divisions across the entire embryo.
00:29:58.26 And so, for this purpose, we collaborated with the groups of
00:30:02.04 Srinivas Turaga and Kristin Branson at Janelia,
00:30:05.14 who used a method... they developed a method based on neural networks
00:30:10.21 that identifies the condensation of the chromatin that occurs in the course of cell divisions,
00:30:18.11 and it picks up the associated changes in brightness and texture of the fluorescent marker
00:30:25.05 that we used in these live imaging experiments, this nuclear marker.
00:30:29.15 And so, this is the result of using this technique at the scale of an entire embryo.
00:30:34.03 This is actually the same image dataset I've shown you before, the same visualization...
00:30:38.19 timecourse visualization of embryonic development, but now superimposed with a color code
00:30:43.15 that points out where all of the different divisions occur, now using this color code.
00:30:47.08 So, there are several tens of thousands of divisions happening in each of these embryos
00:30:52.09 during these 48 hours.
00:30:53.09 I mean, you end up with 50,000 cells at the end.
00:30:55.14 And there are interesting patterns in these divisions that are revealed, for example,
00:31:00.15 by looking at a map that condenses all of that information into these...
00:31:05.27 you know, into this map of two axes.
00:31:07.26 So, we have the time axis along the vertical dimension.
00:31:11.09 And then, along the horizontal dimension, you can see space
00:31:14.02 represented by the anterior-posterior axis.
00:31:16.23 And so, these are now all of the divisions from this one embryo,
00:31:20.06 tens of thousands of em... of cell divisions.
00:31:22.28 We can see, now, how they cluster at the anterior and the posterior ends, where the colors become red.
00:31:30.00 And so, since the embryo is in a constantly growing state during this period of time,
00:31:35.09 which means that the tissues are constantly elongating
00:31:37.00 -- pretty much all of the tissues are growing along the anterior-posterior axis --
00:31:40.05 that means that the cell divisions seem to cluster at the leading edge of these growing tissues,
00:31:44.19 which then in turn raises the question of whether... you know, this... this...
00:31:51.14 this particular location, this particular placement of cell divisions,
00:31:56.08 actually actively supports the elongation of these tissues, or whether
00:32:00.02 it's perhaps just much more easy... much easier to find space for cell divisions
00:32:07.26 at these leading edges, rather than in the middle of the embryo, where you might have a harder time
00:32:12.19 pushing your neighbors to the side.
00:32:15.22 Now, finally, what I'd like to show you is one last biological analysis, which now
00:32:24.00 ties together all of the different resources we've looked at so far.
00:32:28.26 Information on cell divisions, cell movements, and the image data, you know,
00:32:32.02 at a more macroscopic scale as well
00:32:35.16 basically allows us to link large-scale tissue morphogenesis, large-scale changes
00:32:39.27 at the tissue level, to the local cell dynamics.
00:32:44.17 And so, there are numerous interesting mechanical processes happening during this 48-hour time window.
00:32:51.25 One example concerns the neural tube, which is a tissue that we're particularly interested in.
00:32:57.04 And so, if you look at the image data from the inside of the embryo...
00:33:02.19 so, same image dataset as before, but now looking from the inside.
00:33:06.06 And it's a surface rendering, so you can see the inner surface, so that you can see
00:33:09.27 the deepest tissues most clearly, you can see the dramatic morphological changes
00:33:15.00 that give rise to the formation of the neural tube.
00:33:17.12 So, initially we have... again, the embryo itself is in this elongating state.
00:33:20.18 The tissues are elongating, so the neural ectoderm is expanding along the anterior-posterior axis.
00:33:26.12 And then, eventually the neural plate is folding in upon itself to give rise to the neural tube,
00:33:30.21 as shown here.
00:33:32.00 You can see you, now, the separating of neural tube.
00:33:36.03 And at the same time, this video also nicely shows the formation of the somites,
00:33:40.01 which are these little blocks of tissue, here, on the left and the right side of the...
00:33:43.10 of the neural tube.
00:33:45.27 So, we're curious to understanding what is happening at the single-cell level, locally,
00:33:52.18 in the embryo during these large-scale tissue rearrangements,
00:33:55.25 these large-scale morphological processes.
00:33:58.25 And we can investigate that, now, with our databases of cell divisions and cell movement patterns.
00:34:04.12 So, by systematically going through the data and, you know, thinking of the process as,
00:34:13.17 like, this two-step process, where we first have the extension of the body axis
00:34:18.03 and we first have the elongation of the neural ectoderm, and then we have the folding process,
00:34:22.10 we find the following thing.
00:34:23.19 If you look at the cell divisions, initially during the phase of tissue elongation,
00:34:30.05 the divisions are for the most part aligned with respect to their division axis,
00:34:35.17 with the rostral-caudal axis of the embryo.
00:34:38.26 So, in other words, during the elongation of the tissue, the cell divisions are aligned
00:34:44.22 along this axis of elongation.
00:34:46.26 And then, once the tissue goes into this mode of folding, the orientation of cell divisions
00:34:52.21 gradually changes, until eventually, as we reach peak folding, the cell divisions
00:34:58.14 are pretty much all oriented along the mediolateral axis.
00:35:01.15 You can see that in the rightmost panel.
00:35:06.00 So, it seems that, basically, at any point in time throughout this tissue, cell divisions
00:35:10.07 always are oriented along the axis of the dominant mechanical process in the embryo.
00:35:16.21 So, we were curious to see if this is a conserved feature across other tissues as well,
00:35:23.17 maybe a more universal phenomenon.
00:35:24.17 So, we looked at neighboring tissues, for example, in somitic mesoderm,
00:35:28.14 which sits right next to the neural tube.
00:35:30.15 And so, this tissue, the somitic mesoderm, does not go through a switch from elongation to folding.
00:35:37.09 And we actually also can't see, in the cell divisions, this kind of switch in the orientation
00:35:42.21 of cell division axes.
00:35:44.06 So, that's consistent.
00:35:46.25 In the notochord, which is a tissue that sits right on top of the neural tube...
00:35:53.02 cells are also, again, very close to the neural tube.
00:35:55.28 This tissue is... but this tissue is always in an elongating phase.
00:35:59.15 It's not really switching to folding at any point in time.
00:36:01.14 It's always elongating and, again, cell divisions are basically aligned with the rostral-caudal axis,
00:36:07.07 with the axis of the elongating tissue throughout the entire time window.
00:36:12.06 So, it seems that, basically, across all of these different tissues and time windows,
00:36:17.14 cell divisions are indeed aligned with their division axis to the axis of
00:36:22.19 the dominant mechanical process that... you know, that happens in that tissue at that time.
00:36:26.05 So, it raises the question of whether cell divisions are systematically placed and,
00:36:33.28 you know, oriented to support this mechanical process, and only support it.
00:36:39.14 Or whether it's... so, if the causality works the other way around, and because of the forces
00:36:45.09 that act on the tissue during these morphogenetic changes,
00:36:48.03 the cell divisions are forced into a particular orientation.
00:36:51.22 The cell division axes are forced into a particular orientation.
00:36:54.11 So, now that we have these observations and these, you know, different hypotheses
00:36:59.03 we can formulate as a direct result of this dynamic atlas of mouse development, we can,
00:37:04.05 in the next step, now design follow-up experiments, perturbation experiments... experiments that
00:37:08.09 investigate this causality, and basically allow us to look at the mechanisms in more detail.
00:37:13.10 And so, I'd like to end on the note that all of the tools we've developed in this study
00:37:21.07 -- the microscope, our computational tools, but also the data resources,
00:37:26.19 so, this dynamic atlas of mouse development --
00:37:29.01 they're all available as free public resources.
00:37:31.23 Typically, that's a bit difficult to achieve for a microscope.
00:37:35.10 I mean, how do you make it publicly and freely available, but we were fortunate enough to
00:37:40.04 be able to partner with the Advanced Imaging Center at Janelia.
00:37:45.13 And so, the Janelia AIC basically now has a copy of
00:37:48.01 our adaptive multi-view light-sheet microscope that is freely available.
00:37:52.11 You can contact them to, you know, arrange for time on this instrument.
00:37:59.14 And if you publish data obtained within this instrument, we are also not co-authors on these papers.
00:38:04.10 So, we really want to make it as easy as possible for people to use these new technologies,
00:38:09.20 long before they're available as commercial instruments.
00:38:11.15 So, here's the website of the Janelia AIC.
00:38:14.21 Please feel free to contact them if you're interested in this method and you think
00:38:18.15 it might help you with your research.
00:38:20.21 We also set up an atlas website, the Janelia Mouse Atlas, where you can download information and,
00:38:27.28 in fact, the resources coming out of this... out of this work:
00:38:31.18 the dynamic cellular atlas of mouse development, the open-source computational framework,
00:38:36.06 and technical blueprints of the microscope, if you want to build your own microscope.
00:38:40.13 And then finally, thanks to the support of the Open Microscopy Environment team,
00:38:47.10 and the team also behind the Image Data Resource.
00:38:50.03 You can even browse online and download the complete four-dimensional image data
00:38:57.04 for multiple embryos through the IDR website.
00:39:02.21 And so, with that, I'm at the end of the talk.
00:39:04.22 And I'd like to thank and acknowledge the people who did this work.
00:39:09.23 Pretty much all of the research I showed you on these slides is the work of
00:39:14.24 two outstanding postdocs in the lab.
00:39:20.08 Kate McDole, who is a biologist.
00:39:23.02 She developed the microscope, did all of the imaging experiments, and did a lot of the data analysis.
00:39:27.13 Leo Guignard, who is a computer scientist.
00:39:30.10 He also worked on the data analysis and developed a number of computational methods,
00:39:34.11 such as the TARDIS framework for space-time registration of embryos,
00:39:38.13 and also the statistical method for the tracking.
00:39:42.08 I'd also like to thank our alumni, Fernando and Loic, who contributed to this work.
00:39:48.01 And I'd like to thank our Janelia colleagues and our external collaborators for their support,
00:39:53.21 and the exciting and fantastic collaborations.
00:39:57.23 Thank you very much.

This Talk

Speaker: Philipp Keller

Audience:

Educators of H. School / Intro Undergrad
Student
Educators of Adv. Undergrad / Grad
Researcher
Educators

Recorded: December 2018

Talk Overview

To better understand how an entire embryo develops from a single cell, Dr. Philipp Keller and colleagues developed a technique to image and quantitatively reconstruct mouse embryogenesis from gastrulation through early organogenesis at the single-cell level. Keller’s lab developed an adaptive light-sheet microscope to follow the mouse embryo for 48 hours while it’s developing its germ layers (mesoderm, endoderm, and ectoderm), early tissues, and organs. By combining long-term high-resolution imaging, computational, and statistical analyses, they generated a dynamic fate map of the embryo. These open-access resources aid in the understanding of the dynamic cell behaviors that allow for proper growth and development of the embryo.

Speaker Bio

Philipp Keller

Dr. Philipp Keller earned his Master of Science in Physics at the University of Karlsruhe and at the University of Heidelberg in 2005. He continued his graduate studies at the European Molecular Biology Laboratory (EMBL). After a short postdoc at EMBL, he joined the faculty at HHMI’s Janelia Research Campus in 2010. His lab uses… Continue Reading