Scientist Stories: Sami Farhi, In situ Omics & Profiling Tissue Spatial Organization

I tried to gear this to be as generally accessible as possible. I made the slides purposefully very information-rich so that if something goes by, you can use them as a resource. But please interrupt me at any point with any questions. Today I'll be talking about this neologism, which is in situ omics, which are general methods for profiling tissue spatial organization. Very briefly, I think it's important to define your terminology. So omics is very popular at the Broad. These are high-content, preferably unbiased characterizations of a collection of biological molecules. We know these very well. We've done genomics for many years. We're now doing transcriptomics and proteomics all the time. And we're doing metabolomics in many groups throughout the institute. I think we have a platform. At the Broad for every one of these four modalities. What we're adding on now is the in situ aspect, which means without destroying the sample and retaining the spatial information. So this requires really rethinking how you do almost every part of these measurements. And I'll tell you, this has been a very active area of research over the last 6 to 7 years. And I'll tell you about some of the major highlights throughout. I'll show you a few of the relevant papers and now the biological results that are just starting to come out over the last couple of years. But overall, why would you even want to do something like this? One is that we think about tissues a lot from dissociated measurements. So single-cell sequencing has become a transformative technique in almost every field of biology, but we tend to ignore the fact is that there are compositional biases to those measurements. So some cells are less well receptive to being dissociated and measured by single-cell sequencing. If you do the measurements in situ, the idea is you're less biased and you can get a better sense for how the tissue actually looks. By retaining spatial information, you can also look at cell-to-cell interactions based on proximity and making computational inferences from the data. Of course, you can also look at spatial variation of biomolecules across tissues. You can look for concentration gradients of morphogens or anything like that over a developing sample. And for me personally, the most exciting thing you can do is you can start to combine many of these assays into the same measurement to look at multimodal pictures of the same tissue. You can look at how morphology and transcriptomics are linked in the same cells. So I'll go over this timeline. I'll spend almost all of the time talking about the in situ omics methods that are available right now. I see I didn't get rid of every snarky thing for YouTube, but I'll talk about some of the methodological tech dev opportunities that are still remaining, very briefly talk talk about the untapped biological areas, and then for you, possibly more interesting, which method might be right for your particular system, and how you might go about making that decision. And then finally, the capabilities that we're building up both at the Broad and specifically at the platform I direct on spatial methodologies. So the— You're familiar with a lot of in situ measurements for just one species. We've had— if you were talking about genomics, the analog would be DNA-FISH. You can make a fluorescent oligomer which can hybridize to a sequence of DNA that you've designed. And you can do this in situ without taking the genome out of the cell and get a sense for where a particular sequence is within that cell. Very similar, you can do the same thing for mRNA and get a sense for how many copies of a particular mRNA you have inside of a cell. And you can do this depending on how many fluorescence channels. You can do maybe 2, 3, or 4 species at the same time or different regions of the genome. Similarly, we've had immunofluorescence for many decades at this point, which allows you to detect the concentration of a protein using an antibody. And then convert that signal into fluorescence with a fluorescently tagged secondary antibody. And then finally, I wasn't sure what to put in for the chemical sensors, so I took something that I did in my undergrad. Obviously, this is not a representative section, but you can look at chemical modalities here, copper(I), which we're detecting with a fluorescence sensor, but you can imagine many such molecules developed for a variety of different biomolecules, and really I'm just trying to kiss up a little bit to the therapeutics crowd here. I made this. But the question is now, how do we increase the bandwidth from one to many? One of the ideas is to overcome the spectral crowding of fluorophores by doing sequential rounds of imaging. Another is to assign spatial information somehow to a barcode that you can then read out together in the data acquisition stage. And then the third common modality is to leverage mass spectrometry, which has a much larger detector bandwidth since you can detect a variety of different elements, a variety of different mass-to-weight ratios. And having the four different modalities versus these three different ways to multiplex things has resulted in an awful matrix, which we will go through for the rest of the, this talk, which I'm going to call methodological menagerie. So here you have the model, the thing you're trying to detect in a high-plex fashion versus the various different modalities to look at it. I'm going to point out one particular separation, which is over here. Cyclic fluorescence, I think, has become the most popular of these methods at this point. And I'm going to subdivide that based on hybridization-based techniques, which rely on sequential rounds of FISH, if you will. Sequencing-based techniques, which do in situ sequencing leveraging something like sequencing by synthesis, sequencing by ligation, or just commercial Illumina chemistry. And then there are methods for that directly detect the signals without going through DNA oligo intermediate. I'll talk about only these 5 boxes today. If you want to know about any of the rest of them, we can talk offline or afterwards in the questions. But these are the most popular, methods so far and the most widely used. And we'll just get right into it, starting with the hybridization-based technologies for in situ transcriptomics. Now, I mentioned RNA-FISH, which is the basis for all of these methods. But we'll do it again one more time so everybody's on the same page. The idea is you have a collection of different RNA molecules inside of a cell, here Different transcript types are represented by different colored thick gray lines. And you can come in with complementary probes and detect these. You can do one gene per color channel, so up to 4 or 5 genes per experiment. Now, this has been done for many years. People have gotten very good at it. And the same idea but with brighter signals is commercially available. You can buy it from RNA-Scope. From ACD Bio in the form of RNA-Scope. You can buy in situ HCR probes from Molecular Technologies, and then progressively better and better amplified signals, including most recently, SABER-FISH from Punggyun's lab. But what if you want to do more? The easiest thing you can imagine is just doing this a bunch of times. So you can take those same probes and strip them off your sample using something like formamide or high heat or a variety of different methods. And then restain with another set of probes and do another round of imaging against a different set of mRNAs. And you can do this as many times as your sample and either you or your postdoc will tolerate, which usually means less than 50. But it works very well, and it's been done many times, most recently in this paper from the Lindner's lab, I want to say, which I'll highlight now to give you a sense for what this kind of data can generate. So they actually did a very interesting thing. They published— this is the entire paper. These are both figures in it. They did 37— or I'm sorry, 15 rounds of imaging for 37 genes. Across this portion of the mammalian somatosensory cortex. This is a mouse section. And they chose canonical genes, which are known to define particular cell types. When they were able to identify, as far as we know, every single cell, and they found very good correspondence between their spatially defined cell types and the cell types that they knew existed there by single-cell RNA sequencing. So what this gives you, which is new, is that now you know where these cell types are actually positioned inside this piece of tissue. And you can use this to perform computational clustering to define anatomical regions and potentially to infer cell-to-cell connectivity. I mean, connectivity here does not mean axon to dendrite. It means some sort of chemical proximity-based interaction. So for example, you can think about what type of astrocyte is closest to what type of excitatory or inhibitory neuron. And this is sort of the default experiment. Every single in situ RNA method has done a version of this, usually in basically the same tissue. And in fact, there is now a comparison project going on to figure out which of these tools does the best in terms of classifying different cell types within this exact system.. But what if you want to go to more than 30 or 50 genes? Your opportunity here is to go to a combinatorial detection-based approach. So here you have to take a slightly different avenue. Rather than directly detecting the mRNAs, you go through an encoding probe. These are these little angular oligos that I've drawn on here. So each one of these corresponds to a particular mRNA species. And is assigned a barcode, which you will be able to read out later. The reason you do this is that it's been found that this increase or drastically decreases the amount of time you have to spend hybridizing and rehybridizing. So instead of doing overnight destains, you're doing 20-minute destains and restains. So the idea is now you come in with a fluorescent oligo against your secondary or your against your encoding probe, strip it off, and restain, and perform this a smaller number of times than you would have to do with linear encoding single molecule FISH. I'm going to pause here in case none of that made any sense, and see if anybody's brave enough to ask a question. Yeah? [Speaker:AUDIENCE] You said the cycle time is like 20 minutes for this? [Speaker:JOSHUA Elkington] Yeah. I think you're about to get a mic thrown in your face. But the question was— Did I say that the cycling time was about 25 minutes? And yes, that's what I'm saying. This stain is much faster than staining directly onto the mRNA. I actually don't know why, but it's been shown by many labs that this is the case. The assumption is that this is something about the crowding of the mRNA. And the restain is 5 minutes. So there's been two versions of this protocol that are basically identical to this. You've probably heard of one or two of them at least. They're called MERFISH and seqFISH. MERFISH is from Xiaowei Zhuang's lab at Harvard. The first paper was in 2015 in Science. seqFISH is from Long Cai's lab, who's now at UPenn, which was started up in 2014. And they've gone through about 5 iterations each, and they've kept the nomenclature consistent. So when you think about MERFISH or SeqFish, you really should think about what year of MERFISH or SeqFish you're talking about. The only real difference between the two protocols is the degree of error checking. So I'll show you what I mean here. This is the figure from the first seqFISH paper. And it gives you a better sense of how the method works as well. Here you have mRNAs that are detected in every round with a particular color. And the number of mRNAs you can detect is the number of barcodes to the number of cycles you have available. I think— So the basic idea is that with 5 barcodes and 3 cycles of imaging, you can do 125 genes. The difference between this and MERFISH is that in MERFISH, they assign a much larger codebook space. So they're not— they're using many more rounds of hybridization and staining, but they're much more robust to errors in hybridization, spot calling, or stripping than the original SeqFish protocol. Now, both protocols have borrowed from each other over the years, so SeqFish now has a degree of error correction in it. MERFISH has borrowed some of the amplification strategies that SeqFish used in order to get bright signals. And then I actually made you guys a movie so you can see what the data actually looks like. This is one cell from MERFISH, which is going through 16 rounds of imaging. And there is one particular dot which is lighting up in only 4 out of the 16 rounds. So you can assign that to a particular gene identity. The key point to think about, though, is that you need single-molecule resolution to perform this imaging barcoding. If you zoom out so you can't resolve single molecules, none of this will work. You can't call the bases. And I want to point out a very common misconception that very intense lasers are required. That's not true anymore. That was true in the original Merfish paper, which used photobleaching as the way to remove the fluorophore. Now that they've switched to chemical means, they go both much faster, and if your laser is a little more dim, you just take a little bit longer to take your images. Of course, brighter light is always helpful because it takes less time to do the imaging. So now that these tools have been around for a while, they actually both have real neuroscience studies that have come out with them. On the MERFAR side, they performed a similar cell type clustering to what was done with OSMFISH, also called awesomeFISH., which is to perform 130-plex profiling of the preoptic region of the hypothalamus. And one of the things they found is that we typically draw— or by we, I guess I mean the Allen Institute typically draws the anatomical definitions of the nuclei of this region based on subtle variations of neuron density. But it turns out that if you map the cell type clustering onto this same space, they don't really match up at all. And so one of the key outputs of this study, in addition to a very nice comparison between the performance of MERFISH and single-cell RNA sequencing, was a more refined understanding of the anatomical structure of this region of the hypothalamus. Similarly, SeqFISH has really pushed the boat out in terms of how many different transcripts they've been able to detect. The latest paper last year was able to detect basically the full transcriptome, doing 10,000 genes. They did this in several different tissues. Here I'm showing the same example they highlight in the olfactory bulb, where based on the the statistically significant increase in the proximity of two different cell types next to each other, and their knowledge of the full transcriptome of the cell, they can back out, or at least propose, receptor-ligand interactions that may be mediating these kinds of interactions, which is a very cool thing that I don't know how you would do with single-cell RNA sequencing. So those are the hybridization-based methods. On the other hand, there's several methods that just take a different approach to the readout, and I'm going to highlight them as a separate category. They're asking the same kind of information, but they're better suited, and people have started applying them more in distinct ways from the hybridization-based approaches. The first approach came out of Mats Nilsson's lab in 2013. And I'm going to step you through this flowchart. But basically, you have an mRNA, you put on a random hexamer primer, and you use reverse transcriptase to get a cDNA. You can then come in with what's called a padlock probe, which is a piece of DNA which can hybridize to two regions on the cDNA.. And you can make these in two different versions, either a gap-filling or a barcoding padlock. The gap-filling padlock has a little space over here, which you can then fill up with a polymerase and a ligase, while the barcode padlock doesn't have that. It just has a nick, which you can seal with a ligase. Regardless of what you do, there is a site here that you can use to initiate rolling circle amplification., and get many, many copies of this small locked padlock. These then typically carry a short barcode, which can be read out by in situ sequencing, either Illumina, solid, or a variety of chemistries that have come out since then. And this is it in action. This is the darkest picture I've ever seen in a paper, but the point is this is a cell, and there are 5 bases in here, which you can see over several cycles can be assigned to sequence the same way that you would with Illumina chemistry. So I'm going to briefly highlight 4 derivatives of this technique. There's the original in situ sequencing and the protocol all the way through. There was a follow-up the year later from George this year is wrong. That should be 2015, from George Church's lab, which instead of using padlocks, tried to circularize the entire mRNA cDNA, allowing you to read out the full sequence. Now, that's been very difficult. A lot of people haven't been able to follow it up. But in the first paper, it looked very nice. There's been a subsequent follow-up from the first author of that paper, who now has his own lab, I think, at Cold Spring Harbor, which appears to be doing a little bit better by having more sophisticated sequencing chemistry. And then finally, our own Xiao Wang, in her postdoc project, developed a derivative of this, sort of a hybrid between in situ sequencing and hybridization-based sequencing. Approaches, which uses hybridization of a padlock and the primer simultaneously on the mRNA to avoid the initial reverse transcriptase step. This is very important because reverse transcriptase is the lowest efficiency step in the whole process, and she's able to capture much more lowly expressed genes this way. The rest of the techniques really work best for higher expressors. And she's also developed her own in situ sequencing chemistry, which is more robust. So now highlight a couple of different approaches where this is particularly appealing. This is for the gap-filling approach. Tony Zayder's lab at Cold Spring Harbor has been developing synthis viruses, which are expressed— various barcodes, which can be injected wherever you like in the mammalian brain, and then will traffic out to the axonal projections. What they've used in situ sequencing for is to figure out which cells are expressing which barcode, while then taking out chunks of brain that they think might contain these axonal projections, and putting it through single-cell sequencing. I think the reason that they do that is that it's much easier to do imaging when you know where you should be imaging. But regardless, this approach works well. They're able to see the individual cells' barcodes and link them up to projection maps, allowing them to generate these kind of connectomics datasets where you can classify the cells you've infected on the basis of their projection patterns to different brain regions. Similar tool for very different application came out of Paul Blaney's lab last year, where they realized that this gap-filling approach could be used to sequence guide RNA, which can be introduced into any cell system that is transfectable. So they borrowed the CropSeq vector, which was published as part of a PerturbSeq approach, which that puts the guide RNA sequence into an mRNA. And they're able to sequence these directly. I think this will play. This is from their paper. But these are HEK cells, each expressing a single copy of a guide RNA, which you can then use to figure out what genomic perturbation you applied to this cell type. Now, the reason you wanna do something like that is because it allows you to perform pooled optical screening. What that means is that this is just sequencing, but you can also think about parameters such as morphology, protein tracking, calcium imaging. In their case— in their case, they looked at NF-kappaB translocation to the nucleus to look for regulators which are implicated in the trafficking of this protein. And they were able to— much as you see with most CRISPR studies— they were able to pick out almost every every known element of the pathway in a single screen, notably this time requiring much less resources because doing a 40 million cell screen is just a single 6-well plate, which can be done with the right automation, which we'll get to down the line. It can be done in a couple of days without doing any kind of library extraction or anything like that. And then finally, this just came out. I thought it was interesting. This is from Mats Nielsen's lab again, so the originator of the technique. And it's using a ligase, which only ligates if you have complete complementarity to a region of interest, rather than— and not ligating if you have a mismatch. So what they were trying to do here is to figure out the degree of RNA editing that different genes in the brain experience over the course of development. They chose a few different mutations and a few different genes that they were interested in. And of course, they found differences, or I guess it wouldn't have been published, showing that in the adult, you have much higher editing levels in a particular set of genes, while other genes tend to not be affected. OK. This brings us close to the end. I want to point out a couple of technical takeaways that I think I mentioned along the way. First, FISH-based methods are intrinsically more sensitive than the in situ sequencing methods. That's just because there's no RT step. They generally allow you to sparsify your signal, which basically means that some transcripts can be dark in some or many rounds, which allows you to get higher numbers of genes profiled and higher numbers of transcripts. But the signals with FISH-based approaches still tend to be quite a bit dimmer, because you don't have the rolling circle amplification step. And the low detection efficiency of in situ sequencing can sort of perversely be an advantage when you're looking at something that's extremely highly expressing, which would just crowd out your combinatorial FISH. Completely. And of course, in situ sequencing, when properly configured, allows you to read out the native RNA sequence, which can be potentially useful. This is a good stopping point before we move on to the next step. So there's another branch of tools which tries to assign a barcode to a particular RNA species based on its spatial information. It's really the answer to, what if I really hate to use microscopes? You don't have to. These approaches in general tend to do many more genes, and they can be untargeted. There's actually some commercial products out right now. 10x recently bought a company called Spatial Transcriptomics out of Sweden and renamed it Visium and made the protocol a little more accessible. That's something that you can just buy now. But one of the downsides is that these methods overall still tend to have quite a bit lower capture efficiencies. You're looking at transcript detection levels that are more similar to single-cell RNA sequencing rather than to FISH. An order of magnitude or more in terms of dropout. I'll only highlight one approach out of these, partially because they're extensively published, and partially because the idea is very similar. And that's the SLY-Seq approach that was developed in Fei Chen and Evan McCosko's lab at the Stanley Center. The basic idea is that you can take effectively drop-seq beads, and deposit them on a coverslip or a slide in this kind of circular array, which you can then perform in situ sequencing with, with any of the tools that we discussed in the previous section. So that doesn't have to be done by the user. That can be done by anybody. And these slides can then be shipped to an end user, who simply has to cut their tissue and smack it down directly onto the coverslip, which captures the RNA, performs an RT and tissue digestion step, and allows you to amplify a library which now contains the spatial barcodes associated with every single bead. They've done beautiful work demonstrating that these kinds of approaches can be used in brain, in cerebellum, olfactory bulb, kidney, liver, and the number of tissues is growing almost weekly. The interesting thing, of course, is that you're getting full transcriptome data for free. So you can perform the traditional single-cell RNA sequencing clustering analysis, and then map those clusters back to the spatially mapped— or to the spatial locations given by the DropSeq bead identities. So this, here, they've found that the top— I guess we're calling this dorsal since we're real neuroscientists— the dorsal edge of the hippocampus is identified as a single cluster in the single-cell RNA sequencing data. They're not color coding this based on any kind of spatial information. They're just merely mapping this cluster back into the original data space. And the same stands true for the habenula here. And because this is an unbiased approach, you can do interesting things. The example they chose in their paper is to reveal the transcriptomic response to a traumatic brain injury. So they take a mouse brain, and they stab it with a needle. And they can look for a variety of different programs which are upregulated over the course of this injury healing. We're about to leave RNA space and go into protein space. OK. So this is going to be similar in feel, but a very different kind of data. So if you want to look at proteomics, you have a bunch of different proteins that are hanging out in your cell. Approach that all of the multiplexed protein— proteomics tools have taken is to detect them with a variety of antibodies specific against each particular species, which are then tagged in a variety of different ways. So the first thing we'll go over is the mass-tagged approaches that use the large sensor bandwidth of mass spectrometry by tagging every individual antibody with a particular heavy metal isotope. Since there's hopefully no heavy metals in your biological sample, these are very cleanly isolated from the native biological signal using mass spectrometry. The idea is that you take a sample, you stain it with your library of antibodies, and you can raster the sample with either a UV laser or an O2 plasma gun. That's the difference basically between secondary ion mass spectrometry or laser desorption mass spectrometry. And then one way or another, you ionize the sample down to its constituent atoms and you detect it using a time-of-flight mass spectrometer. Then you can take the resulting spectrum and separate it out according to mass bins to give you the picture of every individual antibody's labeling. And then of course, they go through data processing and cell segmentation pipeline in order to give you the multiplexed expression data for every cell. Now, this has proven to be pretty powerful in cancer immunology, where people typically have questions about cell type identification, which can't be answered with just 4 or 5 different markers. The example paper I've chosen here from 2018 looked at triple negative breast cancer and stained, I think, about 46 different markers simultaneously. And then looked at the variety of different samples they had and found some interesting things. For example, they were able to cluster the different types of patient samples into samples which had more mixed populations between immune cells and tumor cells, and samples which had stronger compartmentalization between the tumor and the immune environments. The important thing to point out is that if you were to do the same kind of analysis by single-cell sequencing, you would end up getting the same ratio of cells for both of these sample types, even though the spatial localization is very different. And this allows you to do things like map out important immune markers, important checkpoint markers along the tumor-immune boundary. And at the end of the day, it allows you to find morphological features which are predictive of the survival of a particular cancer patient. So in this particular study, it turned out that— patients with this compartmentalized phenotype actually survived a lot longer and responded better to therapy than patients which had this more mixed phenotype. So this is a real predictive tool. And then if we now take the same approach and go over into fluorescence space, the idea is very similar, except that now the antibodies are not tagged with metals, they are tagged with oligos the same way that we did the original FISH work. You can then come in with complementary oligos, stain, image, destain, reimage, and perform this as many times as you want for the number of antibodies you have. Now, the typical numbers here are lower than with RNA. Do combinatorial imaging because there's just too much protein in a cell to get single molecule resolution, although it would be very exciting to figure a way out around that limitation. And this has started to be used. The same technology has like 3 different names at this point. Here I'm showing PRISM, which Jeff Cottrell had a hand in developing, which looks at what is that, 15 different synaptic markers simultaneously in cultured neurons, and allows you to see how the intensity of each of these markers changes in response to stimulations of these cells. They did a 2-day treatment with TTX to see how activity profiling affected the expression of various different synaptic markers, and I think as we may expect, Knowing that these cells are homeostatic, they increased their levels of excitatory synaptic markers. Another approach, switching back into the cancer— or into the immunology domain, is that you can track cell-to-cell interactions. This is work done in spleen, where a 40-plex panel, 36 or 40, something like that, was used to map out all of these cell types within the mouse spleen to find cell-to-cell interaction rates based on physical proximity of these cell types. This is just another way of showing that same idea, that you can look at a cell's nearest neighbors and figure out the most likely cell type next to every particular cell. Perhaps most interestingly, they found that these interaction maps changed in a mouse model of lupus. Again, even though the overall percentages of the cells seemed similar, the spatial arrangements between them were different. And so if you're thinking about whether you should use fluorescent or mass tag-based approaches, there's a few different considerations. One of the key ones is throughput. Fluorescence is, I think, 10,000 times faster than the mass spec-based approaches. The integration time is lower on a pixel-by-pixel basis, and you're acquiring on a camera rather than rastering a laser or an ion beam across your sample. The signal-to-noise is also higher because it can be amplified dramatically, for example, with tools like ImmunoSABER, which put many copies of an oligo onto a fluorescent onto an antibody. And it's cheaper. So the hardware cost is much lower because you're not building a ray gun with a mass spec attached to it. The microscopes, of course, you can spend more money on them, but you don't have to. You can do this with a conventional wide-field microscope. The problem, though, is that there tends to be autofluorescence background in a lot of fluorescent samples. This is particularly bad in adult human brain. It's bad in lung. It's bad in skin. Sometimes you can get around this with amplification. Other times you may not be able to. And there's also an aspect of the wet work that, because you're doing iterated staining and destaining for— the fluorescence-based approaches, whereas with the mass-based approaches you're doing one round of wet work. There's a lot more opportunities for things to go wrong with the fluorescence-based approaches until the systems are better stabilized. So if you have high autofluorescence or you are doubtful of your equipment, it's maybe better to do the mass spec-based approaches. Otherwise, the fluorescence ones will win out. So, again, this is a good breakpoint. Okay. So, there's a lot of things that are left over to do in terms of tech development, and it's at almost every stage of the process. One of the absolutely biggest problems is that it's difficult to translate these approaches across different systems. Because you're not taking the sample out of the native context, different samples have different levels of autofluorescence, they have different levels of RNase activity, they may require staining or expansion, they may have lower or higher expression levels. So it's proven to be challenging to move out of whatever context the assay was originally developed in. It doesn't mean it's impossible, it just means that you need to plan for a little bit of extra work to adapt it to your particular system of interest. Now fortunately, a lot of the people in this room are neuroscientists, and a lot of the method development is actually being done in brain, so it's— if you want to do mouse brain studies, you're probably in a good position. If you want to do, I don't know, human lung, you're in a bad position, and you have to do a little bit more work. There's also serious bandwidth limitations that have to be thought about. In situ proteomics still requires antibodies. There are not that many good antibodies, and there's even fewer that have been validated in each of these tools. I think the largest panel is still smaller than 100 for the same systems. So if you're using a system that requires one of those, Great. If you're not, you have a job in front of you to validate a whole set of antibodies. Now, the companies are getting better at this. The panels have doubled in size in the year that I've been working on this. So hopefully that will not be— that'll be exponential and not linear, but we'll see. Like I mentioned, there's no combinatorial coding approaches for proteins. So that means that even if we do get 1,000 different antibodies will be imaging for a really long time. And on the RNA side, the capture-based tools are very powerful, but they're also very expensive. You're doing Illumina sequencing there. So you're paying upfront in order to do the assay. If you're getting it from 10x, well, you know what 10x— their model is to charge you for upfront and then give you over to Illumina so they can charge you again. On the other hand, Imaging tools are very slow when they go to the full transcriptome. You have to do many rounds of imaging. I think the lowest is 23 different rounds, and sometimes you have to expand the tissue, which slows down your imaging even more in order to cover the same area. This is— now I'm going to start periodically plugging the work that we do in my lab. Oh, never mind. I'm going to do that the next slide. There's also an overwhelming amount of data that's can be generated by these technologies. This is specifically for the fluorescence-based methods, but right now the state-of-the-art labs are still using the OrcaFlash 4.0 camera. This is going to get really nerdy, but these cameras generate 8 megabytes of data per exposure, which is about 200 milliseconds. And they're still not operational. They're not up all the time because most of the labs are in a method development stage. Stage, but better cameras are available. They generate fold more data. They can go faster than what we're doing right now. And once you figure out the protocol and are just pushing samples through, the operational uptime can easily be 80% because these— and that's of the hours in a year, not of workday hours. So I expect that we will be able to generate 5 petabytes a year easily in our lab, which for reference is the total amount of genomic data generated by GP at this point. So we're working— You have a job on the side. I guess marketing. Promenade or something. I think my backup career plan is marketing for Intel and compressed sensing. And then, in case you didn't think those problems were enough, I think we should have more. The menagerie can always be bigger. I'm just highlighting that these fields are pretty densely populated, but the DNA field, it's pretty sparse still, and I think there's a lot more to be done here. Now, part of this is because it's a question of biological utility. Most cells have the same genome, so it doesn't make sense to be doing genome sequencing on a cell-by-cell basis. But similar approaches can be used to, for example, look at chromatin structure, to look at epigenomic modifications. So I think this will be a rich area of development in the future. And at the same time, the metabolite-based approaches really all rely on spatially resolved mass spec, which is slow. And it would be great if we could figure something out over here. But I confess I've been thinking about this problem for like 6 years, and I still have nothing. So if anybody would like to work on this and has an idea, I would like to work on it with you. Finally, there's many different areas that are untapped. Applications outside of neuroscience and cancer immunology are still basically up for grabs. Very few people have migrated the tools over. And of course, we can keep going with refining anatomical definitions. This is already happening as part of the Human Cell Atlas, Human Tumor Atlas network. As the tools become more and more high throughput in terms of the number of transcripts you can profile, we'll be better able to identify receptor-ligand interactions. And I'm hoping that we can eventually get to a place where we're having more informative and predictive clinical pathology diagnosis. And particularly excited by the possibility of merging these tools together to generate multimodal pictures of the same cell. So this is going to be an awful flowchart, but you can have the slides later and then walk yourself through it as you want. And this is updated as of last night at— 11:30. The question you have to ask yourself right now is what kind of information are you looking for? If you want proteins, you want antibody-based readouts. If you want RNA, obviously you want transcriptomic-based readouts. The only real exception, if you say both right now, is a technology called NanoString DSP. And if you're interested in that, you can talk to either me or to the Regev lab, who have the technology in-house. On the protein side, the question is, how many antibodies are you really looking to profile? If it's less than 3, then you already know the answer. If it's less than 7, there's a variety of commercial methods which will help you out, OPAL or Ultiview. If your tool is highly autofluorescent, you probably want either an IonPath MIBI, which we have, or a Fluidigm Imaging Mass Cytometer. Which we don't have, but they're basically the same in terms of data quality. Or you want to start thinking about sequential immunofluorescence, which really starts to get to questions like, do you need fast staining? If no, you can use Sisyphe, which gives gorgeous signals. Do you need fast imaging? If no, you can use a commercial product called the Maxima, although it's very expensive, so maybe you don't want to. If you need bright signal— if you don't need bright signals, you can use CODEX. Which we have. And if you absolutely have said yes to everything so far, it looks like the answer is ImmunoSapor, but it's not commercial. We'll get to a solution to this situation. And again, here are the commercially available ones. On the RNA side, the first thing to ask yourself is, are you ready to do imaging? If the answer is no, then SlideSeq, spatial transcriptomics, or NanoString are the way to go. These have trade-offs between high counts and high resolution. It's actually a conserved amount of data. It's just how finely you're segmenting it, as far as I can tell. Somebody from Faye and Evan's lab can correct me if I got that wrong. If you're willing to image, then again, the question is how many different genes you want to do. For less than 5, standard FISH is totally fine. There is a commercial technology from ADC Bio, which is RNA-SCOPE HiPlex, which is basically 3 rounds of RNA-FISH. But it's commercial, and those probes work really well. If you want to do less than 50, you can entertain the idea of using AwesomeFISH. And if you need more than that, you have to do one of these combinatorial methods, which really just depends. Are you doing less than 1,000 genes? You can go for something like FISH-Seq or Instaseq. Seq. And if you're doing low copy number genes, you want to go for a MERFISH-based approach, or at least a hybridization-based approach. If you're doing high copy number genes, you can get away with one of the in situ sequencing-based approaches. And again, these are the commercially available products right now. And for the rest of them— OK, we have 30 minutes. This is a great break point as well., but if not, I will tell you about what we're doing in order to make these tools more broadly available. Okay, so the problem is that all of these things require sophisticated hardware one way or another, and usually that's not something you want to be dealing with every day, day in and day out, so I lead a little group called the Optical Profiling Platform, which is a team of imaging experts, engineers, and methods developers who are trying to bridge the gap between experimental biologists and clinicians and computational biologists in order to get at these kinds of more sophisticated imaging data moieties. Our goal is to implement the latest and greatest in terms of imaging technologies, working either with the inventing company or the inventing lab. We try to upscale them, especially if they're coming from an academic lab. We want to make the protocols more robust. And of course, on our end, we're trying to develop the next generation of tools. We're trying to leverage our ability to combine all of these tools in one place. If you want to work with us, the general idea is that if there's something totally new, we'll work together as a collaborator. We'll figure out, first, is this something that we can work together on? We'll work together on the sample prep. And then we'll have imaging facilities available that we can use and evaluate how well this protocol is performing. If this is something a little bit better established, and I'll show you some of the things that are at this point, it may be better if the user would prep a sample, hand it off to us, and then we handle the imaging and the low-level data analysis and deliver to the original scientist. Established every step of this pipeline so far, but we're getting there on some of the techniques. And then behind the scenes, this is all enabled by an automation engineer, an optics expert, computational people, and method developers all working together on a variety of these different approaches. I'll show you where we are, and if you want more, you can read the Stanley Center annual report. [LAUGHTER] These are the tools in green. These are the ones that we are actively working on. In bold, somebody at the Broad has worked on this at some point, and I can help point you to them. And then finally, the last thing I want to say is that we're extremely close. Our lab space just opened on January 10th. It's— the Stanley Center is going to be over here, if you remember Tom's presentation from last week. Please come by, say hi. If you have an imaging problem, we'd love to talk to you. If you want to use one of these tools, please talk to us. And then here is everybody who you might see roaming around the halls. I'm eternally grateful for all of these people for taking a chance to work for something that is just starting out. I'm also extremely grateful to these people for supporting us as we do this. And we have time for questions. Sammy, the MIBI and CODEX data that you showed, were those generated with off-the-shelf antibody panels that anyone can acquire and require relatively little optimization? We did no optimization. We should do some. At least you have to titrate your antibody. We didn't even do that. Most of them worked. And the ones that didn't, there's a— this is technical details, and you can actually read about this specifically in the paper from Jeff Control and Mark Baffert, but bad samples imaged under CODEX tend to show a lot of nonspecific staining of the nuclei. This is a well-known problem when you're dealing with fluorescence. Oligos in any way. And so the worse the sample is, the more that staining overwhelms your desired staining. But yeah, it's actually just been working. Yeah.

Can you say anything about computational methods? Is it pretty well understood how to analyze this data?

Are there major open questions? No, it's not understood at all. I mean, that's not fair. People are working on it is the right answer, but I think there's nobody who works in this field who would say that the problem is solved. It's not solved at almost every level. Base calling, which we don't even think about with Illuminix anymore, we just throw that data away at this point, but we don't know how to base call properly varying background levels and the different thicknesses of the tissue and the autofluorescence signals coming in. Once you base call, how do you segment? It depends on the tissue, and that's extremely difficult. Extremely low-level problems. How do you align your images? How do you correct for uneven illumination? I mean, yeah, there's methods out there, but you need something that works day in, day out on terabytes of data, and the fact is that most datasets that are generated and have taken advantage of the currently available technologies are bespoke. Almost everything that's been done that way, you know, you do a bunch of imaging and then you put a grad student or postdoc on that data for, you know, weeks, months, depending on how long it takes to get to the analyzed product, and they're not directly transferable to what we're doing here. So even once you move past all of these problems and you move past the data flux, issue. How do you— what more is lurking in the data is still an open question. The analysis that I've been showing you so far is like cell type A is close to cell type B. There's surely more than that. There's, at the very least, there's morphological information that we're not taking advantage of. These cells look different from each other. Not every dendritic cell looks the same. We know that they don't look the same. The same, but we're not analyzing that. So yeah, I mean, there's an incredible amount to do.

I'd love to talk to you about it. So I just started my first RNA scope this week, so I'm a little new at it. But I was wondering if you could explain what image-based recordings is.

You said it maybe 2 slides ago. What did I mean? Image-based recording.

Tell me when it starts to look familiar.

No. Next one? Wrong way. Not electro— Oh, sorry. Optopatch, yeah. So I didn't go into that at any length because this was supposed to be about in situ omics. Sorry. No, no, it's totally fine. This I can actually do a chalk talk on, but the problem with electrophysiology is that while the data it generates is absolutely beautiful, high signal-to-noise, wonderful, it's very laborious to acquire. So something I worked on in my PhD is a way to get the same kind of data while circumventing the need to stick an electrode into the cell. The way we did that is by combining a spectrally orthogonal voltage indicator with a channelrhodopsin actuator. What that means is that we have a sensor that shines red when you are when the neuron is firing, or at least when the voltage goes up, and we have a channelrhodopsin which we can hit with blue light, which doesn't affect the sensor, which stimulates activity. And what that allows you to do is do things like F-I curves, so you can apply increasing intensities of blue light as we're showing here, —and you can see how the neuron is firing progressively more action potentials as you increase the stimulation. The shape of this curve tells you something about the intrinsic excitability of the neuron. So it's a way to get after electrophysiological information in a much higher throughput way. In principle, you should be able to get to tens of thousands of cells per day this way. These are only 4 in this field of view, so we have a little bit of work left. And it's not invasive, right? And it's not invasive, but it's also— we're also working in cultured cells, so you have to be honest about that. Working in brain slices is much more difficult. We've done it, but it's difficult. So it has strengths and weaknesses, but if you're— especially if we're working in iPS-derived neurons where you're not putting— they are already in culture, it's a very appealing technology. Has anyone made transgenics? I mean, people make optogenetic transgenics all the time. Oh, yeah, we have optopatch mice. We haven't put it under a CAG promoter. That seems crazy. Almost certainly the mouse wouldn't survive that, I guess. We, not me personally, but a postdoc in Adam's lab did make an OptiPatch2 and OptiPatch3 mouse. There we have one is in a tiger locus, the other one is in a ROSA locus. Yeah, they work. It's not as good as I would love it to be. You need high expression levels in order to— to get these kind of voltage signals. The indicator is very dim and you're imaging very fast. So transgenics generally have lower expression than introducing with an AAV. But honestly, the place where I'm most excited about it, we've done this, I've never talked about this publicly, but we did a collaboration in my hometown. So basically I would stop by every Christmas when I was visiting my parents. The lab was interested in the zonagranulosa cells of the adrenal cortex, which release aldosterone, and they had found that these cells are electrically active and are just awful to patch. They have a capsid in them, they're mechanically sensitive, if you stick an electrode in them, it's bad. I think one postdoc patched for 4 years and got 6 cells, something like that. But they took this transgenic mouse, they crossed it to the relevant promoter— I forgot what they used— and I helped them set up the imaging one Christmas, and then the next Christmas they showed a movie which had 8 cells recorded simultaneously within one little rosette of the organ. I hope it's working for them still. I haven't heard back in a while. All right. Thank you, Sami. Yeah. Thank you all for having me.