High Throughput Microscopy

00:00:13.24 Hello my name is Jan Ellenberg
00:00:16.10 and I head the cell biology and
00:00:17.13 biophysics unit at the
00:00:19.04 European molecular biology
00:00:20.17 laboratory in Heidelberg.
00:00:23.10 Today we want to talk about
00:00:25.12 high-throughput microscopy
00:00:27.01 for systems biology.
00:00:30.03 Now the motivation for such
00:00:31.18 studies is that cells carry out many
00:00:34.01 essential functions in life.
00:00:36.22 This is illustrated by this panel of
00:00:38.24 different cell types on this image and
00:00:41.05 each of them actually has multiple
00:00:43.07 essential function that it's carrying out
00:00:45.21 If we want to understand molecularly
00:00:48.00 how these functions work, the first
00:00:50.16 task we have is to identify the genes
00:00:52.18 responsible for these functions.
00:00:55.17 So even for very basic functions of life
00:00:57.19 such as cell division which is depicted here
00:01:00.06 going from interphase to prophase
00:01:03.12 metaphase, anaphase
00:01:05.01 and cytokinesis, the main goal
00:01:07.04 of which is genome segregation
00:01:09.04 we don't know all the human genes and
00:01:11.16 what they do for cell division.
00:01:14.10 Now that that's the case was
00:01:15.22 the very sobering recognition
00:01:17.10 after the sequencing of the
00:01:18.23 human genome a few years ago,
00:01:21.06 where we knew that we had
00:01:22.14 23,000 protein-coding genes
00:01:25.13 in our genomes. But only for about
00:01:27.16 a hundred at the time had a function in
00:01:30.07 such a basic event as cell division
00:01:32.21 been demonstrated.
00:01:34.16 Now we knew from model
00:01:35.17 organisms, such as the
00:01:37.01 roundworm C. elegans, that over 300
00:01:40.09 genes are required for division.
00:01:41.23 So, clearly, we were lacking a lot
00:01:43.11 of the knowledge in human
00:01:44.19 to understand this process comprehensively.
00:01:48.05 So, how do you go about
00:01:49.06 identifying genes for essential
00:01:51.01 functions. The first thing you need is
00:01:53.22 you need a way to inhibit the function
00:01:56.09 of the gene. The most popular way of
00:01:59.05 doing that, is to use RNA interference,
00:02:02.05 libraries of small interfering RNAs that
00:02:04.07 either cover the whole genome or certain
00:02:06.12 subsets of it, to inhibit the expression
00:02:09.12 of a gene one by one and record the
00:02:11.19 phenotypic consequence of that.
00:02:14.23 Now, an alternative is to use
00:02:16.02 overexpression putting genes into cDNA
00:02:19.03 constructs to artificially increase the
00:02:24.05 gene dosage and thereby perturb the
00:02:26.00 function. And more recently very popular,
00:02:28.17 is also to use specific chemical
00:02:30.21 inhibitors in drug libraries or also
00:02:34.04 unknown target libraries to
00:02:36.06 systematically inhibit gene functions.
00:02:39.03 Today for simplicity I would give
00:02:40.21 examples using siRNA libraries to
00:02:44.05 silence the expression of genes.
00:02:46.17 Now, if you have such a
00:02:47.12 genome-wide siRNA library
00:02:49.03 for the human genome, the next
00:02:51.02 thing you need is an assay that can
00:02:53.24 report to you whether the cellular
00:02:55.16 function of interest is affected if
00:02:57.07 a gene is silenced or not.
00:02:59.10 So, I will use the example of
00:03:01.09 cell division for simplicity.
00:03:04.02 It's a quite typical cellular function,
00:03:06.03 it's essential, it’s dynamic in space and
00:03:08.07 time and we can report on it easily on a
00:03:11.12 microscope here using the core histone 2B
00:03:15.12 fused to green fluorescent protein.
00:03:18.11 Now, in this image you can see
00:03:20.10 the bright green nucleus which is where
00:03:22.09 the chromosomes are, and the chromosomes
00:03:24.23 consist of DNA wrapped around histone
00:03:26.22 octamers, and one of these eight proteins
00:03:29.15 in the histone octamer is histone 2B
00:03:31.21 tagged with GFP.
00:03:33.13 So, in every single cell in the field
00:03:35.06 we can see where the genome
00:03:36.10 lies inside the cell. And we can also
00:03:39.05 immediately see the function we're
00:03:40.20 interested in cell division. It's rarely
00:03:42.22 detected because the cell cycle takes
00:03:45.21 about a day, and cell division, in this
00:03:48.11 cell here it's visible, only takes about an
00:03:50.08 hour. But if we record these cells alive
00:03:54.00 on the microscope for about a day then
00:03:56.05 every single one of them that was
00:03:57.23 resting at the beginning of the movie
00:03:59.24 will go through the normal vision process
00:04:02.15 and tell us if genome segregation
00:04:05.00 occurred accurately or not.
00:04:07.04 This is a control situation
00:04:08.11 where no gene is silenced
00:04:09.20 and we can see the cells
00:04:11.00 dividing normally on the microscope.
00:04:13.22 Now, we can compare that to a situation
00:04:15.16 where we silence a particular gene.
00:04:17.21 The positive control gene Plk1 kinase
00:04:20.12 silenced by a specific siRNA.
00:04:23.14 Already the first image shows us
00:04:25.03 these cells are in a different state
00:04:26.22 than the controls.
00:04:28.02 They arrest at the prometaphase stage
00:04:30.19 of mitosis with condensed mitotic
00:04:32.17 chromosomes. And if you run the movie
00:04:35.00 we can see they arrest like that for
00:04:36.16 several hours before they realize that
00:04:39.19 genome segregation is impossible without
00:04:42.12 the activity of this enzyme
00:04:44.04 and they die by controlled cell death,
00:04:46.14 fragmenting the surface and their DNA.
00:04:49.15 So, that's the kind of phenotype we're
00:04:51.00 interested in, if we want to find genes
00:04:53.16 required for cell division. Now. how do you
00:04:56.11 do this across the entire genome with
00:04:58.22 23,000 protein-coding genes.
00:05:01.16 For this we need to develop
00:05:02.23 technology and the first
00:05:04.07 technology is to deliver silencing
00:05:06.10 reagents in high-throughput to cells on
00:05:08.24 a microscope. So, for this, we put together
00:05:11.21 a microarray platform which is assembled
00:05:14.18 inside ELISA imaging chambers and
00:05:17.05 each spot on these microarrays now
00:05:19.00 contains a different siRNA, silencing
00:05:21.18 a different gene. The entire human
00:05:24.11 genome with about 51,000 siRNAs
00:05:27.13 can be spotted on only a hundred fifty
00:05:29.09 microscope slides and therefore in a few
00:05:32.03 experiments you can go through the
00:05:33.17 entire human genome,
00:05:35.07 silencing in the cells cultured on these
00:05:37.14 microarrays each single gene one by one.
00:05:40.17 Now, the next piece of technology you
00:05:42.03 need, is the focus of this lecture, is a
00:05:44.18 high-throughput microscope. So, we need a
00:05:47.00 microscope, which is fully incubated so
00:05:49.04 the cells are dividing as happiest in
00:05:51.23 the incubator, and the state of which can
00:05:54.19 accommodate in this case up to four of
00:05:57.18 these microarrays, allowing us to image
00:06:00.02 about 1,500 such gene silencing
00:06:02.11 experiments in parallel. Now we're
00:06:04.19 recording movies for two days with the
00:06:07.04 time resolution of 30 minutes so the one
00:06:09.07 hour event of cell division is never
00:06:10.17 missed, resulting across the entire
00:06:13.09 genome in a data set of 260,000 movies
00:06:16.17 or 34 terabytes of digital data.
00:06:20.18 So, this really is the technology
00:06:22.04 challenge. One of high-throughput
00:06:24.03 microscopy record such large amounts of
00:06:26.07 data, which reproduce the quality, in an
00:06:28.16 automatic fashion. But immediately after,
00:06:31.13 we have a computational challenge
00:06:33.09 with the data.
00:06:35.02 Before we come to that, we take
00:06:36.20 a step into the lab and actually look at
00:06:38.22 the operation of a wide field fluorescence
00:06:42.01 time-lapse microscope in screening
00:06:44.09 operation.
00:06:46.03 We see here is fully automated
00:06:48.02 right-field time-lapse fluorescence
00:06:50.02 microscope in a temperature
00:06:51.23 control box imaging each siRNAs
00:06:54.20 spot of the 4 microarrays with a
00:06:57.11 time resolution of 30 minutes for two
00:06:59.22 days. Cells stably expressing
00:07:05.12 GFP-tagged H2B highlighting the
00:07:08.02 cell’s DNA throughout the cell cycle.
00:07:12.12 The data that such a microscope
00:07:13.24 produces looks like this:
00:07:15.23 in the background, we have the entire field
00:07:17.21 of view of a 10x objective. The dashed
00:07:20.05 line indicates the spot where the siRNA
00:07:22.18 was deposited, in this case
00:07:25.02 silencing the gene Kinesin-5, which
00:07:28.11 is again a known mitotic component for
00:07:30.09 reference. And in this rectangle, we have
00:07:33.21 zoomed up this area of the experiment
00:07:36.12 for simplicity, where we can now see
00:07:38.10 cells dividing for two days, or rather
00:07:41.04 not dividing because you can see, similar
00:07:43.14 to Plk1 kinase, these cells arrest in
00:07:45.21 mitosis at the beginning of the movie
00:07:47.20 and then undergo apoptotic death by
00:07:50.04 fragmenting their DNA. So, this again is
00:07:52.17 another example of a mitotic phenotype
00:07:55.04 that we would like to find automatically,
00:07:58.08 but we have to find these genes that
00:08:00.06 have such phenotypes in the 260,000
00:08:03.06 movies that we recorded. So that requires
00:08:06.20 automatic and unbiased computational
00:08:08.20 analysis. How do you go about doing that?
00:08:12.00 You take a stepwise approach. The first
00:08:14.13 thing the computational analysis does is
00:08:17.08 to segment the image, finding the
00:08:19.22 individual cells that can be of very
00:08:21.15 different brightness depending on
00:08:22.22 whether they divide or not, and defining
00:08:25.18 the outline of each of the nuclei of the
00:08:28.00 cells. Now after that each cell
00:08:30.15 computation now is one object in the
00:08:32.10 compound image, and human expert can
00:08:35.05 annotate what cell cycle stages the
00:08:38.04 different cells were in when the movie
00:08:40.22 was recorded. For example, interphase,
00:08:43.15 mitosis, and apoptosis. Now the computer
00:08:47.07 then extracts from each of the cells
00:08:49.11 about 200 features that describe the
00:08:51.23 shape and the labeling and the texture
00:08:54.14 of each of the cells. And by comparing
00:08:57.02 the features between the different
00:08:59.13 annotated classes, it can make
00:09:01.15 quantitative distinctions. This is
00:09:03.18 illustrated in this feature plot
00:09:05.22 for two such features. Here
00:09:07.24 we look at how round a cell is and at
00:09:10.24 the standard deviation of the gray
00:09:13.00 values inside the cell, which means how
00:09:15.14 evenly labeled the cell is. You can
00:09:18.13 immediately see from the scatterplot
00:09:20.13 apoptotic cells in red are very different
00:09:23.15 from interphase and mitotic cells because
00:09:27.10 they are no longer round and they are
00:09:29.00 no longer evenly labeled for their genome,
00:09:31.10 which is fragmented. So, if you imagine
00:09:33.22 200 such parameters, you have a very
00:09:36.05 powerful automatic classification
00:09:38.22 algorithm that can in fact in the entire
00:09:41.20 genome-wide data set differentiate 16
00:09:44.15 different morphological classes. So, this
00:09:47.07 allowed computational analysis in a week
00:09:50.07 of two billion images of individual
00:09:52.24 nuclei or 20 million movies of cell
00:09:56.07 division events. Now the classes that are
00:09:58.20 particularly interesting for cell division
00:10:01.07 are either the stage of cell division
00:10:03.04 itself highlighted here in green, or the
00:10:05.23 consequences of an incorrect division,
00:10:08.18 which are highlighted here in blue.
00:10:11.09 So how do we find those classes
00:10:13.04 automatically in the movies and also
00:10:15.04 score with statistical significance if
00:10:17.13 they are over-represented compared to
00:10:19.23 a control situation. Now our automatic
00:10:23.16 recognition of morphologies has
00:10:25.19 transformed the movies now into
00:10:28.13 quantitative data. Each cell has an
00:10:30.17 outline and a color code depending on
00:10:33.06 the morphology recognized. Here we will
00:10:35.14 focus on individual nuclei in interphase,
00:10:39.12 cells with two nuclei after failed
00:10:41.24 cytokinesis event, and cells with more
00:10:44.13 than two nuclei after a second failed
00:10:46.15 cytokinesis event. If I run the movie, you
00:10:49.03 can see at the beginning many green
00:10:51.13 cells then taken over by many
00:10:53.09 bi-nucleated dark blue cells and at the
00:10:55.22 end of the movie many light blue cells
00:10:58.11 with more than two nuclei. So, this can now
00:11:01.09 be plotted quantitatively, where we plot
00:11:03.20 the percentage of cells in a particular
00:11:05.18 morphological class over the two days of
00:11:08.23 the movie. And you can see exactly as in
00:11:11.03 the movie, the green interphase cells
00:11:13.02 disappear, the dark blue bi-nucleated
00:11:15.17 cells appear after one day, and the
00:11:18.03 light blue, more than two nuclei cells, come
00:11:20.23 up after two days at the end of the
00:11:22.19 movie. Now to measure again in 260,000
00:11:26.14 experiments which gene was actually now
00:11:28.15 specifically affecting morphologies
00:11:31.16 representing cell division defects, we need
00:11:34.00 simpler measures than a plot like this.
00:11:37.03 So, what we score is the maximum
00:11:39.12 penetrance of A mitosis relevant phenotype
00:11:42.10 at any timepoint in the movie. Here, for
00:11:45.03 bi-nucleated cells after about a day of
00:11:48.03 the movie representing many cells that
00:11:50.11 looked like this or like that in the
00:11:52.14 movie. We compare then this difference to
00:11:56.04 all the genes that we have silenced in
00:11:58.04 the entire genome, scoring their
00:12:00.18 penetrance in the genome wide
00:12:02.00 distribution, and if the siRNA causes
00:12:05.16 a deviation more than two standard
00:12:08.02 deviations away from the whole data set
00:12:10.06 we say it is a real hit, it significantly
00:12:13.04 perturbed this morphology in the movie.
00:12:16.12 Now we then add to the binuclear
00:12:18.06 morphology the other morphologies relevant
00:12:20.20 to mitosis, mitotic arrest, and delay,
00:12:23.23 polylobed nuclei, or grape shaped nuclei.
00:12:26.17 And if we sum all these siRNAs together
00:12:29.00 and map them to the genes
00:12:30.18 in the human genome, we
00:12:32.03 come to about 1,200 hits from the
00:12:34.21 primary screen. We then validate whether
00:12:38.04 these hits are reproducible by having at
00:12:41.12 least a second siRNA that gives the
00:12:43.19 same phenotype in the movies, coming to a
00:12:46.07 set of about 572 validated genes that
00:12:50.03 reproducibly, with two independent
00:12:51.19 reagents, perturb mitosis. So that's the
00:12:54.22 set of genes close to 600 rather than
00:12:57.16 23,000 that we should work on if we're
00:13:00.23 interested in cell division. Now the
00:13:03.13 movies that we recorded contain a lot
00:13:05.15 more information about the genes than its
00:13:08.09 simple score is a hit or not. And we
00:13:10.12 would like to use that information to
00:13:12.14 predict which genes have common
00:13:14.17 functions in mitosis, and which genes
00:13:16.20 have different functions in mitosis. But
00:13:19.14 in order to do that, we need to measure
00:13:21.05 how different the movies are from each
00:13:22.20 other and that's not so simple as
00:13:24.10 scoring if an siRNA has a significant
00:13:26.18 phenotype. So, for this we have
00:13:28.18 to transform our data.
00:13:30.13 Here is the plot that
00:13:31.22 you already saw before. We have the
00:13:33.16 percentage of cells over time and for
00:13:36.15 simplicity we plot only two morphologies:
00:13:39.08 binucleated cells and polylobed cells.
00:13:42.18 In this movie first, you have binucleated
00:13:44.13 cells coming up and later polylobed. In
00:13:47.09 this movie also both of these classes
00:13:49.09 come up, but the polylobed are first,
00:13:51.18 followed by the binucleated. So, these are
00:13:54.00 biologically two different results.
00:13:57.02 However, most computers would say these
00:13:58.20 are very similar because both classes
00:14:01.17 are over represented with relatively
00:14:03.23 similar kinetics over time. Now to see
00:14:07.02 the order of events, we need to transform
00:14:09.06 this data into a trajectory
00:14:11.09 representation where we plot the polylobed
00:14:14.02 class directly against the binuclear class
00:14:17.02 and we use time as the implicit
00:14:18.21 parameter to link two data points
00:14:20.19 together. Now in these two trajectory
00:14:23.06 plots we can see immediately that these
00:14:25.11 two experiments are very different. And
00:14:27.10 to measure the difference between them
00:14:29.12 what we need to do is we need to
00:14:30.23 transform the data into two vectors that
00:14:34.01 connect the two different points in time.
00:14:37.03 These are two because we're scoring two
00:14:38.19 cell cycles over two days. And then between
00:14:41.17 these vector pairs we can mathematically
00:14:43.18 measure the distance here with two
00:14:47.08 morphologies, but in a very similar
00:14:50.05 way with 16 morphologies, which are all
00:14:52.22 the morphologies we measured. Now these
00:14:55.11 distances can then be used to transform
00:14:57.15 the data into a heatmap. This heatmap
00:15:00.00 measures now functional similarity
00:15:02.07 between the genes. So, the dendrogram here
00:15:06.18 shows the relationship between the genes
00:15:08.23 based on the distance between the
00:15:10.06 vector pairs and the heatmap
00:15:12.01 visualizes that distance for the human
00:15:14.00 eye.
00:15:14.24 Each column shows how over-represented a
00:15:17.17 particular morphology is. Let's take cell
00:15:19.17 death as an example: the darker blue the
00:15:22.12 more cells exhibited cell death, the
00:15:25.12 lighter to what's yellow the less cells
00:15:27.07 did that. Each column has a timeline
00:15:30.17 for the two days that we recorded so we
00:15:33.08 also can see when a particular phenotype
00:15:36.00 appeared. You can see this clustering
00:15:38.23 really groups genes together that have
00:15:41.02 similar phenotypic signatures in the
00:15:43.05 movie and that means these
00:15:45.04 clusters should predict common function
00:15:47.04 of genes in a particular process in
00:15:49.15 mitosis. So, let's zoom in on one of these
00:15:52.04 classes: the first box here in red
00:15:55.00 where at the beginning of the movie we have
00:15:56.17 a mitotic arrest phenotype followed by a
00:16:00.00 polylobed phenotype indicative of
00:16:03.06 aberrant chromosome segregation. And at
00:16:05.22 the end of the movie the cells are dying
00:16:08.04 because of cell death. This cluster
00:16:10.15 contain genes that were already known to
00:16:12.24 play a role in mitosis in the assembly
00:16:15.18 of the mitotic spindle the central
00:16:17.13 structure made up of microtubules that
00:16:19.17 segregates the genome. So, this prediction
00:16:22.16 for the new genes we found needs to be
00:16:24.07 validated. To do that we performed a
00:16:27.09 secondary screen where we visualize the
00:16:29.16 microtubules directly on the confocal
00:16:31.21 microscope. So, in red you can see again
00:16:34.02 the chromosomes which look like in the
00:16:36.04 primary screen but now we have a second
00:16:38.03 fluorescent channel in green where we can see
00:16:40.14 the microtubules directly in these cells
00:16:43.04 after the same genes are silenced. And
00:16:46.01 wherever and arrow is pointing we can
00:16:48.00 see defects in the process of spindle
00:16:50.24 assembly. The spindles are not forming
00:16:53.01 normally or not forming at all as in
00:16:55.15 this case cell division failed as a
00:16:58.03 consequence of that. So, the secondary
00:17:00.19 screen that visualizes directly the
00:17:03.03 predicted function can confirm that
00:17:05.05 these genes truly belong together
00:17:06.24 functionally but it doesn't yet tell us
00:17:09.14 what the molecular mechanism was that
00:17:11.17 underlined this particular common
00:17:13.23 function. So, in order to do that we now
00:17:16.10 have a better hypothesis that
00:17:18.00 microtubules in the mitotic spindle are
00:17:20.01 somehow affected and we can again use
00:17:22.10 high-throughput microscopy to understand
00:17:25.05 mechanistically what this phenotype is
00:17:27.16 due to.
00:17:28.21 So here we make use of another reporter
00:17:30.20 protein called EB3 again tagged with GFP
00:17:34.07 and EB3 has the property to bind to the
00:17:36.18 growing tips of microtubules. So, in a
00:17:39.14 short 30 second movie of EB3 on a
00:17:42.20 confocal microscope we can see the
00:17:45.07 growth trajectories of the microtubules
00:17:47.11 within the cell. If we track
00:17:50.15 these data automatically in the computer,
00:17:52.19 which is shown here by projecting the
00:17:54.19 computer tracks on the time projection
00:17:56.14 of the movie, we can extract
00:17:58.11 quantitative parameters about microtubule
00:18:00.18 dynamics from these data. These parameters
00:18:03.19 are for example the average growth speed,
00:18:06.15 the length of the microtubules, and their
00:18:08.24 lifetime.
00:18:10.12 This can now be done again on automatic
00:18:12.12 high-throughput confocal microscopes and
00:18:14.12 we can do it for all the genes where we
00:18:16.03 predicted and validated in the secondary
00:18:18.10 screen that they affect the mitotic
00:18:20.05 spindle. So, putting all that data
00:18:22.15 together then again can be done in the
00:18:24.16 form of a heatmap, where blue means one
00:18:28.03 microtubule parameter was reduced, yellow
00:18:31.08 means it was increased, and white means
00:18:33.21 no phenotype compared to control. And we
00:18:36.10 can see for this set of spindle assembly
00:18:39.04 factor genes that most of them perturb
00:18:42.04 microtubule dynamic parameters, and again
00:18:44.15 we can cluster them by the similarity in
00:18:47.16 which way they affect microtubule
00:18:49.16 parameters, now giving us very detailed
00:18:52.09 hypothesis to really understand
00:18:54.19 molecular function. Now what I've showed you
00:18:57.21 so far, I've let you through the workflow of
00:19:00.04 high-throughput microscopy applied first
00:19:02.16 to a genome-wide RNAi screen for
00:19:04.19 basic cellular functions such as cell
00:19:06.09 division, then a secondary screen to
00:19:09.10 validate for one particular cluster of
00:19:11.15 the discovered genes that it truly is
00:19:14.06 involved in this function, and a tertiary
00:19:16.09 screen that then shows you molecular
00:19:18.10 detail in the phenotype. But all of these
00:19:21.13 examples are linked to mitosis and I
00:19:23.19 want to make clear that the technology
00:19:25.13 platforms that I used to do this are
00:19:28.06 completely generic. And so, this is
00:19:30.24 another example where we can use the
00:19:32.24 same method of genome-wide screening,
00:19:35.09 secondary screening and tertiary
00:19:37.03 screening to validate predictions for
00:19:39.07 a completely different cellular process
00:19:41.07 which is protein secretion.
00:19:44.02 The whole process and workflow of
00:19:45.21 the genome-wide siRNA library,
00:19:48.02 computational analysis of the data
00:19:49.23 and automatic data acquisition,
00:19:52.14 secondary screening, clustering network
00:19:55.21 construction, and validation is generic.
00:19:58.08 It's the same for any cellular function.
00:20:00.14 What you need to do is you have to
00:20:02.09 change the assay on the microscope
00:20:04.14 because now the function that you want
00:20:06.14 to visualize and assay is very different:
00:20:10.02 we want to study how proteins move
00:20:12.06 from the endoplasmic reticulum to the cell
00:20:14.14 surface. So, Rainer Pepperkok's group here
00:20:17.04 at the EMBL who carried out the screen
00:20:19.05 use an antibody that can visualize if a
00:20:22.01 protein is delivered to the cell surface if a
00:20:24.19 particular gene is silence. When this
00:20:26.18 delivery is impaired, we can see that by
00:20:28.19 the absence of the signal on the cell surface
00:20:31.07 and can again quantify this in the
00:20:33.07 microscopy images. So, this genome-wide
00:20:35.20 study has led them to understand for
00:20:38.24 about a hundred fifty genes in detail
00:20:40.17 what they do in the secretory pathway, a
00:20:43.12 completely different cellular function.
00:20:45.14 So, these methods really can be used to
00:20:48.04 understand any function of cells that
00:20:49.22 can be visualized on a microscope.
00:20:53.16 Nevertheless, also for this study we stay
00:20:56.04 at the genetic level of information. We
00:20:58.21 understand the genes required and we
00:21:00.17 understand in detail their function or
00:21:03.12 phenotypic profile but eventually we
00:21:05.22 need to move beyond that and study
00:21:07.19 directly the proteins. Before I tell you
00:21:10.22 how we do that we move again to the lab
00:21:13.11 and look at an automatic microscope for
00:21:16.18 immunofluorescence-based assays, which
00:21:18.20 doesn't just automatically image but
00:21:20.14 also, automatically dispense reagents
00:21:23.10 using computerized dispensing system.
00:21:27.18 This automated confocal fluorescence
00:21:29.20 microscope is equipped with an on-stage
00:21:32.19 liquid dispenser.
00:21:43.11 The chemically fixed cells to be stained
00:21:45.16 are first washed with buffer.
00:21:53.09 Thereafter, the first solution
00:21:55.19 containing fluorescently
00:21:57.02 labeled antibodies is applied to the
00:21:59.13 cells for staining.
00:22:04.21 After staining cells are washed again.
00:22:22.21 Following the autofocus and auto
00:22:24.22 exposure adjustments, cells are
00:22:26.24 automatically imaged in 3D.
00:22:40.23 Finally, the fluorescent labeling is
00:22:42.23 bleached away completely and the whole
00:22:45.12 procedure is repeated for another
00:22:47.12 antibody. Automatic immunofluorescence
00:22:50.18 microscopy allows labelling and imaging of
00:22:53.20 up to several tens of organelle markers
00:22:56.15 in the same cells in parallel.
00:23:04.22 Now back to the lecture. We would like to
00:23:07.09 move from the genetic to the protein
00:23:09.24 level to understand how cellular
00:23:12.11 functions are carried out and again we
00:23:14.04 can use high-throughput microscopy in
00:23:16.12 order to characterize the proteins
00:23:18.13 encoded by the genes we identified.
00:23:20.20 So far, I told you about RNAi screening,
00:23:23.09 how it can identify genes, and how it can
00:23:26.08 provide a lot of information about their
00:23:28.04 phenotypes. But once we know the genes,
00:23:30.19 we need eventually to look at the proteins
00:23:33.10 and for many cellular functions that are
00:23:35.15 regulated in space and time very
00:23:37.21 intimately,
00:23:39.01 this is most useful to do in the context
00:23:41.01 of the intact cell. So again, we turn to
00:23:43.09 imaging and by physical measures to
00:23:46.00 understand for example where do these
00:23:47.24 proteins localize and when do they
00:23:50.18 interact during the biological function
00:23:53.02 we are interested in.
00:23:54.08 And having such data, we
00:23:55.16 hope, will provide enough information to
00:23:57.24 propose models about the molecular
00:24:00.11 mechanism of how these proteins are
00:24:03.01 carrying out the essential cellular
00:24:04.21 functions. So, I want to briefly give you
00:24:07.15 a glimpse of what the technologies are
00:24:10.01 that can be done in high-throughput
00:24:11.07 under microscope to characterize protein
00:24:13.21 systematically. Now the first information
00:24:16.23 we would like to have about cellular
00:24:19.05 protein networks are systematic
00:24:21.00 localization information. Because where
00:24:23.14 a cell has its protein localized is often
00:24:26.13 informative about the function of the
00:24:27.24 protein. Now to do this let's go back to
00:24:31.04 the example of cell division, which I
00:24:33.02 used at the beginning of the talk where
00:24:35.02 we had about 600 genes. So, what we
00:24:37.08 now need is we need to GFT-tag the
00:24:39.20 proteins encoded by these six hundred
00:24:41.15 genes, put them into cell lines and
00:24:43.23 multiplex the analysis in multiwell
00:24:46.03 plates. Now the assay for localization
00:24:48.22 then would be high throughput,
00:24:50.08 high-resolution confocal imaging in
00:24:52.06 three-dimensions overtime,
00:24:54.11 record many such movies of the different
00:24:57.03 proteins, and then derive in the computer
00:25:00.11 the co-localization and construct the
00:25:02.23 localization networks. So, in order to do
00:25:05.17 that, we need again microscopy technology.
00:25:08.15 But before we put cells on the
00:25:10.02 microscope we have to express the
00:25:12.04 fluorescently-tagged version of the
00:25:14.21 protein we are interested in. And to make
00:25:17.04 sure that fluorescently-tagged version is
00:25:18.22 also working, we first need to
00:25:21.10 demonstrate that it is working by
00:25:23.10 rescuing the RNAi phenotype. So, to tag
00:25:26.17 the protein what we use is the
00:25:28.22 autologous gene from mouse, which is
00:25:31.05 automatically resistant to the human
00:25:32.24 siRNA. We tag it at the last exon using
00:25:36.02 GFP and then we introduce it back into
00:25:38.16 the human cell line that we used for the
00:25:41.08 primary screen. So, in the primary screen, we
00:25:43.22 had many mitotic defects either during
00:25:46.12 mitosis or after mitosis. Now the rescue
00:25:49.11 cell lines are the same human cell line
00:25:52.01 that have the mouse RNAi-resistant
00:25:54.05 gene to complement the phenotype and you
00:25:56.06 can see all the post mitotic nuclei are now
00:25:58.23 normal. So that means these genes are
00:26:01.02 working and it's useful to look at the
00:26:03.21 localization of the encoded proteins. Now
00:26:07.04 we want to do that not just throughout
00:26:09.11 the entire cell cycle because we already
00:26:11.07 know the function is critical during
00:26:13.05 mitosis so we want to focus our analysis
00:26:15.19 on cell division only. In order to do that,
00:26:18.21 we have to make the microscope more
00:26:20.05 intelligent than the microscope for the
00:26:22.06 genome-wide screen. And to do that we
00:26:24.18 have developed a software platform
00:26:26.11 called Micropilot, which is open source
00:26:29.09 available, which allows the microscope to
00:26:31.17 automatically focus the cells, take a
00:26:34.08 low-resolution prescan image, make an
00:26:36.19 online real-time prediction of which
00:26:38.10 cells are about to divide and then go
00:26:41.20 into high-resolution mode in order to
00:26:43.22 record just the dividing cells. So, this
00:26:47.01 works in practice. Here is such an overview
00:26:49.11 image of a field of cells, one cell that
00:26:52.22 the computer predicted to divide. The
00:26:55.16 microscope then automatically zooms in
00:26:57.14 on top of that cell and checks whether
00:27:00.03 the protein of interest is actually
00:27:02.03 expressed in the GFP channel. And if
00:27:04.05 that's the case, it launches its
00:27:06.10 3D time-lapse movie
00:27:07.20 of the dividing cell. So, with this
00:27:10.21 intelligence on the microscope we can
00:27:12.24 record such difficult experiments as a
00:27:15.17 3D movie of cell division in
00:27:17.14 high-throughput automatically. And so, we
00:27:21.12 want to go back to the lab again for a
00:27:23.10 moment to show you one of these
00:27:25.02 automatic high-throughput systems that
00:27:27.01 are intelligently controlled in
00:27:28.23 operation.
00:27:30.13 The high-throughput confocal
00:27:31.21 microscope scans the sample first to
00:27:34.17 rapidly acquire low-resolution images of
00:27:37.15 the nuclear Cherry-H2B labelling.
00:27:48.15 Acquired prescan images are then
00:27:50.14 analyzed by the Micropilot software in
00:27:53.06 order to identify cells that are going
00:27:55.20 into mitosis.
00:28:23.06 After the cell of interest is found, the
00:28:25.14 microscope settings are changed to
00:28:27.18 zoom into this cell and to acquire
00:28:30.23 high-resolution 3D multicolor time-lapse
00:28:33.18 image series.
00:28:41.21 The data that such a microscope
00:28:43.11 produces then looks like this:
00:28:45.09 here we have a panel of cells that we
00:28:47.04 all caught automatically at the
00:28:49.00 beginning of the mitotic process. You can
00:28:51.16 see prophase condensed chromosomes here
00:28:54.16 in our red reference channel of histone 2B
00:28:58.01 tagged with mCherry, and in green
00:29:00.06 we can see the localization of all the
00:29:02.03 different proteins of interest that we
00:29:04.04 know have mitotic functions. So where
00:29:06.13 they are during mitosis should be
00:29:08.03 informative for what they do and you can
00:29:10.04 see the patterns are very different. Some
00:29:12.04 proteins localized to prominent mitotic
00:29:14.17 structures like the central spindle or
00:29:17.21 the kinetochore. Other proteins are in
00:29:21.10 places where the function is not
00:29:22.19 immediately obvious like the cell
00:29:24.00 surface, or simply just the cytoplasm. But
00:29:27.09 we can again use the dynamic localization
00:29:29.14 information to predict which proteins
00:29:32.22 are most similar to each other and if
00:29:35.00 they are in the same place over several
00:29:37.08 stages of mitosis probably also interact
00:29:40.06 with each other. But how do we do this if
00:29:43.02 we have this data for 600 proteins? Again
00:29:45.02 we need to use computational tools and
00:29:47.11 what I can do today is not yet present
00:29:49.10 you the end result of that, but I can
00:29:51.05 present you what the strategy is to
00:29:53.10 integrate dynamic 3D localization data.
00:29:56.18 The input data we have is like this: we
00:29:58.08 have 3D data over time of dividing cells.
00:30:01.01 We perform these experiments in a cell
00:30:03.01 line that again gives us landmarks.
00:30:06.01 Now here's our old friend landmark
00:30:08.08 chromosomes labeled by histone-GFP.
00:30:11.07 But in addition to that we need to know
00:30:12.23 where this cell is ending so we have a
00:30:14.16 landmark on the surface of the cell. And
00:30:16.21 we want to know the position of the
00:30:18.16 mitotic apparatus, so we have a landmark
00:30:21.07 on the spindle pole as well. All these
00:30:23.24 three landmark proteins are tagged in
00:30:25.14 the same color, cerulean or cyan
00:30:28.01 fluorescent protein and
00:30:30.00 are therefore taking up just one channel
00:30:32.05 on the fluorescence microscope.
00:30:34.03 Now our protein of interest that we want
00:30:36.03 to probe its localization of is tagged in
00:30:38.12 mCherry, in this case localizes to the
00:30:41.03 mitotic spindle, and now needs to be
00:30:43.11 described quantitatively in order to
00:30:45.22 compare it to other proteins in
00:30:48.13 reference to the same landmarks. So, what
00:30:50.23 we do is for each cell and time point to
00:30:53.24 extract again quantitative features that
00:30:56.10 describe the localization at that time
00:30:59.01 point of the protein of interest, that
00:31:01.14 have to do with its special location, its
00:31:04.01 distribution, the morphology of the
00:31:06.06 labelled structures, as well as the
00:31:07.24 texture inside the cell of the fluorescence
00:31:10.18 channel of interest. So, these extracted
00:31:13.12 features then allow us to cluster the
00:31:16.11 data again so we can compare different
00:31:19.17 cells at different time points that are now
00:31:22.08 single data plots in such a feature plot.
00:31:26.00 And the differences between different
00:31:27.19 groups can be again trained by
00:31:31.04 supervised classification of proteins
00:31:33.16 that have known localizations. So that
00:31:36.10 then allows us to go for an individual
00:31:38.23 protein in an individual dividing cell,
00:31:41.05 look at the feature parameters and map
00:31:43.10 it to a cluster of proteins that have
00:31:46.04 similar dynamic localization properties.
00:31:49.04 Now the prediction would be that all the
00:31:50.17 proteins that share these properties are
00:31:52.21 indeed interacting so the output of this
00:31:55.09 prediction is a prediction about
00:31:57.12 protein-protein interactions. Now this is
00:32:00.11 only a prediction because the confocal
00:32:02.09 microscope has a resolution of about 200
00:32:04.16 nanometers that is not sufficient to
00:32:06.22 demonstrate direct interactions. So that
00:32:09.08 means if we want to validate such
00:32:11.13 protein interactions we need to turn to
00:32:13.22 methods that can provide this
00:32:16.07 information directly. And so, these
00:32:18.16 methods then are really biophysical
00:32:21.08 methods that look at the properties of
00:32:23.09 individual molecules. So, what we need to
00:32:26.18 do to validate interaction
00:32:28.17 systematically is record single-molecule
00:32:30.23 fluctuations in high-throughput in order
00:32:33.21 to extract the network information from
00:32:36.14 this data.
00:32:37.20 What does this technically require? It
00:32:39.13 requires a lot of cell culture to make
00:32:41.23 now not single but double knock-in cell
00:32:43.10 lines that have the predicted protein
00:32:46.14 pair tagged in GFP and mCherry. These
00:32:51.14 can again be multiplexed on multi-well slides
00:32:54.08 and then we need to perform automatic
00:32:56.20 high-throughput fluorescence correlation
00:32:58.20 spectroscopy - that I will explain in a
00:33:01.09 second - record many thousands of spectra
00:33:03.16 and from that data extract in the
00:33:05.22 computer the interactions during cell
00:33:08.14 division systematically to finally
00:33:10.08 assemble an interaction network during
00:33:12.19 mitosis. Again, here we don't yet have the
00:33:15.17 end result but I can show you how this
00:33:17.08 technology works because the automation
00:33:19.12 has already been achieved.
00:33:21.22 So, first of all, the principle of
00:33:23.23 fluorescence correlation spectroscopy.
00:33:26.19 An FCS microscope is in essence a normal
00:33:29.00 confocal microscope that has two
00:33:30.16 different features: one is the confocal
00:33:33.01 laser beam. It's not scanning the cell to
00:33:35.07 generate an image but is rather
00:33:37.04 stationary at one position in the cell.
00:33:40.02 The other feature is that the detectors
00:33:41.19 used are very sensitive and have
00:33:44.04 single molecule sensitivity in terms of
00:33:47.15 the photon count. Now the beam of the
00:33:50.23 laser can be focused to a very small
00:33:52.19 volume into which only a few
00:33:54.17 fluorescently labeled proteins fit. The
00:33:57.03 data that you record on an FCS then is the
00:33:59.12 fluctuation of these molecules through
00:34:01.24 the focus laser beam. So, the data you get
00:34:04.20 is fluctuation data over time. You can
00:34:08.03 analyze that data in different ways that
00:34:10.17 I would not go into detail here because
00:34:12.14 we focus on high-throughput microscopy.
00:34:14.20 But from that data you can extract the
00:34:16.23 concentration of molecules, their
00:34:19.05 diffusion constant, and whether two
00:34:21.22 proteins tagged in different colors are
00:34:24.22 interacting with each other. So, in order
00:34:29.01 to run FCS automatically, a key
00:34:31.13 ingredient is automatic water immersion
00:34:33.22 over long times. And we will step out to
00:34:36.08 the lab again to look at how automatic
00:34:39.03 water immersion microscopes are
00:34:40.22 actually working.
00:34:43.10 The cap on top of the objective is
00:34:45.02 filled with water automatically.
00:34:53.08 Then the turret moves back to its
00:34:55.20 initial set position. The focus is
00:34:59.06 adjusted, and images and FTS data
00:35:02.10 are acquired automatically.
00:35:12.24 To keep the water level constant,
00:35:14.18 refilling occurs at regular intervals.
00:35:30.05 Now the data that this enables us to
00:35:32.00 generate automatically looks like this.
00:35:35.02 Again, we use Micropilot in order to
00:35:38.00 automate the process intelligently so
00:35:40.17 that we record fluorescence correlation
00:35:42.17 spectra from cells that are in the
00:35:45.08 desired cell cycle stage, for example G2,
00:35:48.17 and that have the co-expression of the
00:35:51.02 two markers we are interested in, a red
00:35:53.23 and a green fluorescently-tagged protein.
00:35:56.02 Also in the the right concentration to be
00:35:58.05 useful for FCS.
00:36:00.05 Micro pilot then automatically segments
00:36:02.13 the cells in both nuclear and
00:36:04.01 cytoplasmic compartments, automatically
00:36:06.08 positions the laser beam into the
00:36:08.11 different compartments, and then
00:36:10.07 automatically records the fluctuation
00:36:12.10 data here in the nucleus or in the
00:36:15.05 cytoplasm, and performs automatically the
00:36:19.08 autocorrelation analysis to extract
00:36:23.00 here the residence time of the molecules
00:36:25.18 in the laser indicative of their
00:36:27.18 diffusion mobility, and the number of
00:36:30.13 molecules indicative of their
00:36:31.22 concentration. So, measuring this for
00:36:34.07 hundreds of cells we can very quickly
00:36:36.24 compare the mobility of proteins in
00:36:39.12 either the cytoplasm or the nucleus to
00:36:42.00 each other with statistical significance
00:36:44.14 and also determine their concentrations
00:36:46.18 very accurately. This is one level of
00:36:49.04 proteomic information but the most
00:36:51.08 interesting actually is if the red
00:36:53.09 and the green molecule move together
00:36:55.09 through the laser beam, which would
00:36:56.23 indicate that they bind to each other. So
00:37:00.02 in order to do that we analyze the data
00:37:02.05 a little bit further. We record the red
00:37:04.17 and the green labelled proteins at the
00:37:06.05 same time on two different detectors in
00:37:08.04 terms of their fluctuation. Here an example
00:37:11.06 with an H2B-mCherry-tagged reference
00:37:14.10 and an H2A variant YFP-tagged protein of
00:37:17.11 interest. If we look in the cytoplasm, we
00:37:20.14 can see the fluctuation curves between
00:37:23.23 the red and the green proteins are not
00:37:25.24 correlated, so the black Cross
00:37:27.22 Correlation curve shows no significant
00:37:30.06 amplitude over background.
00:37:32.13 However, if we look in the nucleus,
00:37:34.21 where these two proteins are part of the
00:37:36.18 same protein complex, we can see their
00:37:38.24 strong correlation between the red and
00:37:41.20 the green protein, and the
00:37:44.00 black Cross Correlation curve now is
00:37:45.12 very significantly elevated above
00:37:47.07 background, indicating a physical
00:37:49.06 interaction between these two proteins
00:37:51.16 in the same moving particle. So, this kind
00:37:54.05 of technology, now applied systematically
00:37:56.13 to many green- and red-tagged protein
00:37:59.05 Pairs, allows us to measure interactions
00:38:02.05 of proteins in living cells at the time
00:38:04.19 point we're interested in and in the
00:38:06.20 compartment we are interested in by
00:38:08.18 intelligently controlling the FCS
00:38:10.19 microscope. So, we hope in the long run
00:38:14.13 that high throughput microscopy will
00:38:15.23 really allow us to measure many such
00:38:17.24 proteomic parameters, not only score the
00:38:20.24 phenotypes and the genetic requirements
00:38:23.00 but also characterize in detail,
00:38:25.04 biophysically and biochemically, the
00:38:27.06 proteins that carry out these functions.
00:38:29.16 By integrating all that data in the
00:38:31.07 computer, the vision is that we are going
00:38:33.14 to be able to construct a Google Cell,
00:38:36.11 which is not like Google Earth a static
00:38:38.05 map of the components of the server,
00:38:41.13 rather a dynamic map of how these
00:38:43.22 protein components work together to
00:38:46.00 carry out the essential functions of life.
00:38:49.11 Now before I close I would like to
00:38:51.18 acknowledge the people who contributed
00:38:53.23 to the work that I presented to you.
00:38:56.07 This has been a collaboration for many
00:38:58.17 years between my own group and Raina
00:39:01.14 Pepperkok's group, who heads the advanced
00:39:03.17 light microscopy facility here at the EMBL.
00:39:06.10 For technology development involves our
00:39:09.00 mechanical and electronic workshops and
00:39:12.22 also involves constant interaction with
00:39:15.10 our industrial partners who are willing
00:39:17.16 to make modifications on their
00:39:19.00 microscopes, most prominently Leica, Zeiss,
00:39:22.08 and Olympus. The names highlighted here
00:39:24.21 in green, from Rainer's and my own team,
00:39:27.06 have made major contributions to many of
00:39:29.24 the publications and many of the
00:39:31.21 unpublished data that I showed you
00:39:33.15 during this lecture. And last but not
00:39:36.01 least, I would like to acknowledge
00:39:38.05 funding from the German Research Counsel,
00:39:41.10 the Human Frontier Science Foundation, and
00:39:43.15 the European Commission in both the Sixth
00:39:45.11 and the Seventh Framework Programme. And
00:39:48.12 with that I'd like to close. I hope you
00:39:50.23 enjoyed the lecture on high-throughput
00:39:52.15 microscopy.

This Talk

Speaker: Jan Ellenberg

Audience:

Researcher

Recorded: May 2012

Talk Overview

Determining the genes involved in different cellular processes is essential for understanding how cells function. However, with 23,000 protein coding genes in the genome, how can these molecular players be identified? In this lecture about high throughput microscopy, Jan Ellenberg discusses the methodology behind high throughput content screening by describing a series of experiments aimed at identifying genes involved in cell division. Both high throughput microscope imaging and computational analysis of phenotypes are demonstrated and discussed.

Questions

When developing a high throughput assay to identify genes involved in a cellular process, which parameter(s)/readout(s) need to be kept in mind?
1. Secondary assays to validate hits
2. Method for systematically perturbing the function of each gene
3. Unbiased computational analysis of phenotypes
4. Readout assay to identify perturbed functions
5. A and B
6. B and D
7. All of the above
Which secondary assay(s) would be useful to validate genes involved in cell division using initial results from a high throughput content screen? (select all that apply)
1. Single color fluorescence correlation spectroscopy of identified proteins
2. Fluorescent readout of apoptosis in siRNA cells
3. GFP-localization of identified proteins
4. Confocal imaging of fluorescently labeled microtubules in siRNA cells
During the initial siRNA screen for genes involved in cell division, how were the first round of hits identified computationally?
After the initial high throughput screen for genes involved in cell division, secondary high throughput assays were used to validate these hits and determine gene functions. List four of these assays.
Fluorescence correlation spectroscopy can be used to measure what two parameters of a fluorescently labeled molecule?

Answers

Speaker Bio

Jan Ellenberg

Jan Ellenberg

Jan Ellenberg is Head of the Cell Biology and Biophysics Unit and Coordinator of the Center for Molecular and Cellular Imaging at the EMBL, Heidelberg. His lab uses advanced fluorescence microscopy techniques to understand the mechanisms underlying nuclear organization during mitosis and meiosis. Continue Reading