Scientist Stories: Luke Gilbert, Repurposing CRISPR as an RNA guided DNA binding platform

[Speaker:JOSHUA ELKINGTON] Thank you all for coming to the second day of the 2017 workshop. This will be part 2 of my lecture. So yesterday was sort of an elaborate introduction and then a segue into our early work using dCas9 to control transcription in bacteria by by simply sterically blocking RNA polymerase elongation. Today we'll expand into higher eukaryotes, including yeast and human cells, and with a focus on appending additional functionalities to dCas9 to control transcription, edit the epigenome, or image the genome, or mutate the genome. So I think, Today's a really sort of nicely matched set of talks, I think. Some of what I'm talking about will be a segue into some of the things Mike's talking about. And I'm gonna talk about sort of single gene approaches, and then you can imagine scaling a lot of these technologies up to ask additional questions about how the genome controls function of various cell types. So this is the same outline you saw yesterday. We have a lot of slides, so I'm gonna, Skip through the intro. So initially, as I said, we used dCas9 in bacteria. Just by targeting this protein to specific sites along the genome, we can inhibit RNA polymerase elongation. This unfortunately does not work so well in human cells, and so we've taken to appending additional domains to more robustly control transcription. And so our goal early on, once we realized that dCas9 alone didn't, perturb transcription by much in human cells was to ask whether we could append domains to either silence transcription or activate transcription to dCas9, so that in essence we have the platform shown on this slide, where we have an RNA-guided DNA binding platform with which we can fuse a variety of effector domains, and then easily target these effector domains to arbitrary regions of the genome to enact some function. And so, we sort of, one of the reasons I talked extensively yesterday about the, historical perspective of synthetic biology and controlling transcription was we moved very quickly early on because we could take a playbook from what had been done and just repurpose it over to CRISPR. And so I showed you ways that you can silence or activate transcription yesterday using various domains that have been defined over the last 20 or so years. And so we, not being very creative, took a handful of those domains and appended them to dCas9 and asked whether we could control transcription. And so this is one of the earliest experiments that we did, where we took this well-characterized KRAB domain that I had shown you yesterday. This is a scaffold that recruits additional enzymes to change local chromatin marks. And so we appended this domain to dCas9 and transfected or infected reporter cells that expressed a simple GFP reporter with a guide targeting the reporter and this KRAB-dCas9 fusion domain, and then just asked a very simple question a few days later, so that we could turn GFP expression on. Sort of exactly the same experiment as I showed you in E. coli, but here we're using human cells and we've got some additional effector domains. And so this data is shown on the right, where there's a fax plot looking at human cells that express either this GFP reporter, and then we can compare those cells to cells that express no GFP, or if we target the GFP reporter with an sgRNA, that targets this dCas9-KRAB fusion protein to the reporter, we can see that we can pretty robustly turn transcription off. So these two histograms are largely overlapping, suggesting that there's almost no GFP remaining expressed from this reporter. So this was sort of the first experiment that worked to the point where we were excited about using dCas9 to control transcription in human cells. But our ability to semi-robustly control transcription doesn't really mean that much if it's not a general property. And so we went about testing quite a few guides targeting this GFP reporter just to try to derive rules sort of similarly as we had done in bacteria, as to where we needed to position these complexes along a gene to most robustly control transcription. And so you can see, you know, you can test, there's 8 guides up there, some of them work, not all of them work, we started to learn preliminary rules. Some of those I'll go into in additional detail. But the point of this slide, I think, is to show you that we are actually controlling transcription with these domains. So this is a slide looking at repression of this GFP reporter. If you just target the GFP reporter with dCas9 alone, and here you can see, some of these guides are somewhat active, but you don't really see more than about 2-ish-fold repression. And if you add this domain, now you start to see more robust repression. We've improved on this over time, but this is our first hint that we actually are targeting these domains functionally to a gene to control transcription, and that this is happening through local changes to the epigenome. So we're not directly editing the epigenome with an enzyme, but we're using a scaffold to recruit endogenous enzymes that edit the epigenome to turn off transcription locally. So what I showed you previously was Reporter, we also early on showed that we could repress endogenous genes, and that we could, similarly to bacteria, repress at least two endogenous genes. So this system could be multiplexed. And this was early results, maybe not, you know, we're not completely turning transcription off, as I'll show you we can later, but these were encouraging that we began to suspect that we could control expression of a variety of endogenous genes in artificial reporters, either as single genes or as sets of genes, using these artificial transcription factors in human cells. Sort of similar to what I had shown in bacteria, we wanted to know whether we could also use this technology to target promoters to silence transcription, not by targeting a gene body, but instead by targeting promoter, and so we were able to show using this well-characterized SV40 promoter driving expression of GFP, that we could turn off transcription of the GFP reporter by targeting various sites along the promoter. We also were interested in a more subtle question of whether we could use dCas9 alone, just the protein, in a eukaryotic cell to compete with an endogenous transcription factor for a binding site at a promoter. So you can imagine this This is sort of a different question where we were sort of very, in a simple way, probing whether we could map regulatory regions across the genome by competing two proteins for a specific binding site. And you can imagine this would have a variety of applications in mapping enhancers or promoters. But the first experiment we did to test this was we didn't use a completely endogenous system, but we had a reporter that had TET binding sites that drive Venus expression, and we constitutively expressed TET activator protein that induces Venus expression. This is in yeast, actually, so you can see that when you express this RTTA, or this TETR activator protein, you induce Venus expression. And then if you co-express dCas9 targeting these TET sites, we see that we can turn gene expression off. So this is illustrated here. We don't go all the way back to baseline, clearly illustrates that we can use dCas9 in a eukaryote to compete for a transcription factor binding site and control gene expression, sort of functionally mapping sites in an artificial way, functionally mapping sites along a promoter which are required for controlling gene expression. And so this is, I think, sort of an underexplored use of dCas9, but it has a lot of potential for mapping regulatory regions sort of in a very fine way. So our ability to control transcription is only useful if it's specific. So if you have, doesn't matter how robust your ability to control transcription is, if you're deregulating many genes across the genome, that's not gonna be a very useful tool for understanding how biological systems work. And so I'm not presenting these experiments in temporal order, but after we saw that we could use these dCas9-CRAB fusions edit the epigenome and control transcription, we immediately went and did an RNA sequencing experiment to see whether our ability to control transcription of these reporters was specific. And so this, I like this experiment for a couple reasons. One is it gave us the answer we want. So that data is shown here, this is the transcriptomes of two populations of cells, one of which expresses a negative control, and the other expresses sgRNA targeting this GFP reporter. So when we express the sgRNA targeting GFP, we guide the dCas9-KRAB fusion protein to the GFP reporter, the reporter and turn transcription off. And you see that GFP expression goes down. Very satisfyingly, though, we saw that really the expression of no other genes changed. So this implies that we can both robustly and very specifically control transcription. When we saw these results, this was really where we got excited about using this as a means of controlling transcription in human cells. The reason this experiment works sort of well, and the reason we did it was that you could imagine it can be hard to test specificity of, your specificity for modulating transcription for endogenous genes. If you overexpress or knock down an endogenous gene, that's gonna have primary effects on the gene you're targeting, but could also have a variety of indirect effects. You can imagine if you're knocking out an enzyme that's required for some metabolic function, you're gonna knock that enzyme down, but then over time you're gonna accumulate secondary transcriptional changes. One of the ways that we like to measure specificity is repression or activation of proteins that have no functional consequence on the cell. And that allows you to really measure your direct effect on transcription and not convolute this with other effects. So these are additional control RNA experiments that show that expression of a guide alone or the KRAB fusion alone don't really have much of an effect on cells. So these aren't grossly toxic to a cell. And this is an important point for thinking about, you know, applying these to a variety of models, you wouldn't wanna have some gross underlying toxicity if you're trying to study some stress response or developmental response. Similar to what we showed in bacteria, we showed that our ability to turn off transcription in human cells is controllable, so we can create inducible systems where we can induce the expression of our CRISPRi constructs to turn off transcription acutely, and then we can withdraw these inducers, in this case, using doxycycline, the promoters go off and the artificial transcription factor goes away and genes go back on. So that's shown more or less here where we've done an experiment where we knocked down these two genes, we then wash out doxycycline, so dCas9, the CRISPRi expression goes off, and you can see transcription comes back on, sort of with delayed kinetics relative to what I showed you yesterday for bacteria. But nonetheless suggests that these, our ability to control transcription is inducible and reversible, and the changes that we make to editing the epigenome are not inducing permanent silencing of a gene, but rather are locally changing chromatin in a way that is fully reversible. [Speaker] So you're controlling transcription of dCas9 here. Have you tried it with any of the inhibitory proteins you talked about yesterday? Which ones? [Speaker:AUDIENCE] It's like the small molecule, like small protein inhibitors. [INAUDIBLE] [Speaker:JOSHUA Elkington] Yes. No. So you're talking about the anti-CRISPRs? [Speaker:AUDIENCE] Yeah. [Speaker:JOSHUA Elkington] Yeah. No, so we have used more sophisticated schemes to control activity where we have small molecule dimerization systems where you can, what was shown, one of the things that was shown yesterday was these split Cas9 systems where you're not controlling transcription, you're controlling protein Dimerization using a small molecule. We have done that. We have not used an anti-CRISPR to control our dCas9 constructs. You definitely should be able to, but we're doing those experiments, but they haven't been, we've been a little shorthanded recently, and so we haven't gotten around to them. The anti-CRISPRs are relatively new, so these really have only been applied in human cells over the last like 6 months or so successfully. And we just haven't had a strong biological motivation for testing that. I mean, I think it was shown to work with Cas9, and so once that worked, it's very likely to work with dCas9. But definitely an additional level of control that would be exciting.

Except that the anti-CRISPRs freeze the bound complex, and so they might not be effective in preventing the transcriptional regulation.

That's true. We don't actually know. You're right. Yeah, yeah. Although for the one— the anti-CRISPRs, like the first three characterized, all have different mechanisms for inhibiting cascade. So even if the one that was shown yesterday doesn't work for dCas9, there's some rationale to think that we could find one that would work. But we'll see. We haven't tried.

So I see crabs on the middle terminus. Is that the best end, or does it matter?

We've appended it to both. I think, no, I actually like the C-terminal fusions better. I think they're a bit more stable. The N-terminal fusions, you can see the protein is quite unstable. So when you wash out doxycycline here at day 4, by day 5, the protein by Western blot is almost completely gone. The half-life is quite short in human cells. That's a good thing for an inducible system, 'cause it makes it less leaky, but it means to get, in our hands, you need a reasonably strong promoter to get full activity of these N-terminal fusions, where you can use a weaker promoter for the C-terminal fusions, just 'cause the protein's more stable. So the bottom plots also show that you can use this to study essential genes. One of the things that is a little bit tricky to study using wild-type Cas9 is an essential gene. If you, you know, make an edit to the genome to knock out an essential gene, that's a bit difficult to recover a dead cell. With CRISPRi, we can use these inducible versions of CRISPRi to control transcription, and in that way, sort of, we can set up experiments and then knock down an essential gene when we want, and then carry out CRISPRi analysis. So here below, this is the same inducible system where before, we're looking at knocking down these essential genes over the course of several days. If we don't add doxycycline to turn on expression of the CRISPRi construct, we can see that these sgRNAs targeting essential genes have no effect on cell proliferation, so the cells proliferate like wild-type cells. If you then knock down these variety of essential genes, cells where you've knocked down the essential gene deplete very robustly relative to their wild-type counterparts. And then lastly, we've shown that you can use this to control expression of both protein coding genes and non-coding genes. This is a little bit later work than what I showed earlier, where we're knocking down a variety of endogenous genes. And you can see that our ability to control transcription has gotten more robust over time as we've derived algorithms that are more highly enriched for highly active guides. So this is qPCR data on the top and bottom, and it's a log 10 plot. So you can see we're really, you know, really robustly controlling transcription for many of these genes, both protein coding and non-coding, where we're getting like 99.9% repression. And if you look by, for example, FISH down here for Xist RNA, which coats the inactivated X chromosome, you can see that there is, like really no signal left at all for these cells. So we're completely turning off transcription for this endogenous non-coding RNA. So that was, all that work was done in cancer cells, a chronic myeloid leukemia cell line called K562s. Well, not all of it, but most of it. Some was done in HEK293s and HeLa, but really like standard sort of very robust and easy-to-use cancer cell lines that may or may not model many types of biology, and so we were curious whether we could expand our ability to control transcription out of cancer cells into more developmental models or other models. And so, together with Bruce Conklin's lab, who's in the Gladstone at UCSF, we decided to ask whether we could control transcription in human iPS cells. What we did was, rather than, all the previous work I showed was either transient transfection or lentiviral delivery. In this case, we've gone to knocking our constructs into a safe harbor locus within the cell so we can control their expression very stably over time in iPS cells. And then we used the same inducible system that I showed on the previous slides to ask how efficient is our ability to control transcription, is this tunable, is it reversible, can we multiplex this, we continue to ask many of the same questions. And so this is a series of pictures comparing the ability to, the gene's not on here, they're targeting, sorry, what is this? So this is actually just a schematic to show that the system works, that you can knock these constructs into the genome of an iPS cell, and that we can control expression of either Cas9 or dCas9 regulatable system. So in both cases, we've targeted either our CRISPRi constructs or a Cas9 nuclease to the AAVS1 locus. We can use a TALEN to make a double-strand break, which increases homologous recombination, and then we use a long double-stranded linear donor to knock in all of this business into these AAVS1 loci. And so what we've knocked in is both a constitutive promoter driving the RTTA, that in the presence of doxycycline will activate expression of a CRISPRi construct or a Cas9 nuclease. And you can see that in the absence of doxycycline, the CRISPRi construct or Cas9 are not expressed. If you add doxycycline, you get homogenous expression across the population of cells. So then the question is, is this functional? And so Bruce's lab targeted a variety of genes, but shown here is targeting NANOG with either CRISPRi or CRISPRn,, and we're able to show that here in iPS cells we can very robustly control expression of this gene NANOG. So NANOG here would be stained in green, and you can see there's no green there, so that's pretty good control of gene transcription. If you target NANOG with CRISPRn, you see more heterogeneity. So some cells will be completely knocked out for NANOG, other cells will be heterozygous and still express some NANOG. Both tools are great, they just give you different outputs. Here you're not making nulls, you're turning transcription down. Here you make true nulls, but you also have heterogeneity in protein expression across the population. And Bruce's lab showed that this system in general, as in cancer models, works quite robustly to control transcription of a variety of types of endogenous genes. This is just qPCR data. But overall, it was a nice proof of principle set of experiments showing that we could turn off transcription in iPS cells in an inducible way. And so you can imagine this is especially important for studying differentiation, where you might wanna grow your iPS cells up, infect in a guide, induce some differentiation protocol, for example, that guides you down a neuronal lineage or a cardiac lineage, and then at a specific time point during that differentiation process, turn off a gene and ask how does that gene How is that gene required for neural differentiation or cardiac differentiation? And so that's all possible with either of these systems, really, in most ways. But definitely we're able to control transcription in iPS cells. One of the things we noticed really early on that I like to reemphasize in this workshop is that in many settings in human cells, Expression of an sgRNA is the limiting factor for a lot of types of experiments, whether it's imaging or, in many cases, editing, if you're not using RNPs, or for controlling transcription. So one of the ways we saw this was that we were knocking down these GFP reporters, and we saw that the level of knockdown, so this is expression of a GFP reporter, was to some degree correlated with indirect measurements of guide expression. So the more sgRNA you express, the more you repress this GFP reporter. And so that was a problem for us early on, 'cause we had an eye towards doing functional genomics, where we perturb expression of all genes encoded in the genome, but to do that, you need to build a platform that relies on guide expression from a lentivirus that's integrated at single copy into the genome. And so if we had, you know, a dose-limiting expression of our sgRNA, that wasn't a very satisfying, position to start from for building a functional genomics platform. And so together with Bo Huang's lab and Stanley Qi, we tested a series of changes to the original sgRNA design. So this is the original published design from Jennifer that was published early on. It's actually a modified form of the natural pyogenes hairpin that's a bit truncated, so it's truncated by 5 bases. And so we tested a number of things, but we tested just, what happens if we just extend the hairpin to the wild-type length? And that seemed to be more active, as these were the active changes that we made. And then the other thing we noticed is that there's a Pol III terminator in the stem of the pyogenes tracker RNA. So if we remove that stretch of 4 Ts and just break it up with an A, you increase expression of the guide RNA by about 5-fold, which is beneficial in many cases. And then if we combine both this change that stabilizes the sgRNA structure along with the change that increases transcription, what we saw was that we got a lot better activity when we were trying to control transcription with single-copy lentiviral integrations of the guide. So this isn't always affected if you're doing transient transfection experiments. Often, you know, expression off a plasmid is really robust and your experiments won't be limited by either expression of the Cas9 or dCas9 proteins or the guide, they're just more limited by other factors. But we and others have seen that this guide design in our hands works a lot better, and just so you don't think I'm, I'm not only talking about our work. This is one example of another lab from Texas, I think, that showed the same thing. And at this point, there are 4 or 5 other papers that have shown that this modified guide design is generally better than what most people are using. So it's a bit, I'll get off my soapbox, but basically it's a bit surprising that people haven't switched toward this, 'cause there's no consequence, as far as we can tell, to having, well, unless you want an inducible system like Dana's using. But in many cases, you want to maximize your guide expression, and these changes help with that. So there are a variety of tools for silencing transcription or doing loss-of-function experiments, including, you know, RNAi and CRISPR, now CRISPRi, in human cells. So we were excited about additional functions that we could specify using dCas9, and so we tested whether we could activate transcription as well. You can imagine this would have a variety applications for gain-of-function experiments that could be useful in studying cancer biology or development. And so, again, we took a page from the artificial transcription factor field and used the sort of who's who of domains to try to activate transcription of reporters. We fused this VP16 domain in 4 copies to dCas9, and then targeted a GFP reporter and asked with, and this reporter importantly has 3 binding sites for the sgRNA we're expressing, and asked whether we could turn GFP expression on. And so if you look a couple days later by flow cytometry, we can see that specifically when we co-express the dCas9 activator protein together with a guide targeting this promoter, we turn on GFP expression. So this is proof at least that we can fuse these domains to dCas9 and use them to target the genome and activate transcription of reporters. Disappointingly, when we started to try to test our ability to activate transcription of endogenous genes, we saw very quickly that this scheme doesn't really work very robustly to activate endogenous gene expression. And this is not entirely surprising. Yesterday I showed you a slide from Keith Jung's lab, where they were using tail VP64 fusions, and I didn't really walk you through the slide, but essentially what the paper The point of the paper was that one tail VP64 fusion is usually not enough to activate transcription of an endogenous gene. And they saw that with tails you had to have, you know, 4 adjacent, or 3 or 4 adjacent binding events to get activation of most endogenous genes. And so that's essentially the exact same thing we're seeing here, where if you have 3 adjacent binding events, you get nice activation of a gene, where if you have a single binding site, you're only recruiting one copy of this VP64 domain, you don't get much activation. And so at that point, I think the field in general, like, started to engage with this idea that a simple artificial transcription factor that activates transcription with a single binding event was not gonna be enough, or at least not enough with a single activation domain. And so we and others started to search for additional ways to recruit more activation domains with the expression of a single sgRNA. And the way we solved this problem was we fused a highly optimized epitope tag to dCas9. So this is derived from a yeast GCN4 protein. It's a fairly linear epitope that can be connected by a short flexible linker. So we fused 10 copies of this epitope to dCas9. And so now in theory, with expression of one sgRNA, if we co-express a single-chain variable fragment fused to this same activation domain, we can recruit up to 10 copies of the activation domain, sort of using a single sgRNA with a version 2 system to amplify our signal. And we showed in human cells for a variety of endogenous genes that this really worked much better than the version 1 system. So here we're looking at just a test reporter gene called CXCR4. The gene's not really important, although for the aficionados, this is an HIV co-receptor. And you see with our version 1 system, when we target CXCR4, we really don't activate transcription at all. If you use our version 2 system, where we're now recruiting 10 copies of this activation domain, we see we turn on gene expression by about 40-fold. This gene is pretty much off in these cell types, and so it's difficult to normalize a fold change. If you subtract background fluorescence, you can get artificially large fold changes, and so the field struggled a little bit with how to display these kinds of data. What I like a lot better is just to display the primary flow cytometry data. K562 cells that are unstained, so they just, this is autofluorescence. If you stain them for CXCR4 expression, and in cells that express a negative control guide RNA, you see, you know, they look the same as the unstained cells, indicating there's really almost no CXCR4 expressed on this cell type. If you then activate CXCR4 expression by expressing a guide in the presence of this version 2 CRISPRa system, we see that most of the cells in the population activate expression and it's fairly homogenous across the population. And so this, our ability to, this is a little bit of a subtle point, but our, sorry.

[Speaker] I've just got a question about the dynamics of the system. So do you have an idea of whether or not the Cas9 is actually actively docked onto the site improvements of the additional factors, or do you think of like leaves, and just because it has a strong binding affinity for this general area, that's where you're getting activation of the genes? And then a follow-up question, do you see activation of other genes like in this locus or this area in general?

[Speaker] So, I'll answer the second question first. No, we don't, for the most part, we don't see positional effects where you activate neighbor genes. If you have two anti-parallel promoters that are very close together, you'll activate both of them. But the window in which we activate transcription is quite narrow, it's only about 1 to 2 KB, so the transcription start sites for endogenous genes are generally more spaced out. There are caveats to that answer, there are a lot of described non-coding RNAs are anti-parallel to coding genes and are very, very close with their divergent transcription start sites. And so we can't really decouple the function of an adjacent non-coding RNA, although that's a bit of a tricky thing to say, 'cause often the expression of those anti-parallel non-coding RNAs is intimately also involved in regulation of the coding gene. So you might need additional methods to dissect those. For the first question, that's sort of hard to measure The affinity between the epitope that we're using and the SCFE is extraordinarily high. I would be surprised if you didn't have the SCFE bound to dCas9 constitutively, and that as it, we haven't measured it. If you do, like we've done some FRAP experiments where we're just imaging telomeres using this same setup, And I don't know, that doesn't really tell you the answer to your question, I don't know. So like I said, the field was searching for a series of second generation CRISPRa constructs. There have been quite a few published at this point, from us and Ron Vale, from George Church's lab, from Feng Zhang's lab, and Jaenisch's lab, and who else is on here, Charles Gersbach, uh, maybe Stanley. Uh, anyways, there are at least 8 now published. Um, and George Church did the field a service by comparing them all in a Nature Methods paper, um, described here, uh, last year. And I have a few issues with the way he did this, but, uh, we'll just take him at his word, and he says that, uh, ours, uh, his, this VIPER system, and the SAM system from Feng are the best. All three sort of convergently arrived at the same idea that you need to use multiple activation domains to robustly control transcription of, or robustly activate transcription of endogenous genes. And so you can read about these, they're all published, but they all converge on the same idea, essentially. So I wanna introduce an idea that's been really, really useful over the years in synthetic biology. So far today I've talked predominantly about controlling transcription or doing a variety of things by fusing proteins together artificially. Another way that you can control protein localization is to use RNA aptamers that have high affinity for proteins. So a lot of these depicted here are a bunch of RNA stem loops, most of these are derived from phage viruses, and these have been used, or they've been characterized since the '70s or '80s as having very, very high affinity for specific phage proteins. So they're involved in phage genome and viral assembly, and so these are often coat proteins that coat the viral phage genome and are bound very tightly. And so there's quite a number of them, actually. There are whatever, 6 listed here. This top one, MS2, is perhaps the most famous one, but what we can see is that this MS2 coat protein derived from a phage binds the MS2 stem loop with a very, very high affinity. So this is the KD listed here. And so this has been used in the past in synthetic biology to control recruitment of proteins using an RNA rather than using a protein. So we're talking about RNA-protein interactions, not protein-protein interactions. And so one of the ways in which a second-generation CRISPRa system for activating transcription was derived was to take advantage of the fact that we can append these MS2 stem loops onto an sgRNA. So now, rather than dCas9 recruiting activator domains, the RNA is recruiting the activator domain by interacting with a second protein that's expressed. And so Feng appended in his CRISPRa system VP64 domain directly to dCas9, and then he also appended two MS2 stem loops internally on the sgRNA in a manner that was informed by the crystal structure. So this was like really, really nice piece of work. They waited until they had the crystal structure and then didn't just blindly stick these MS2 stems onto the sgRNA in sort of an unguided fashion like we did, but instead said, we're gonna use the structure to figure out where we can append these impact the function of the sgRNA, and stuck to there, then showed that this doesn't compromise the activity of sgRNA loading into Cas9 or dCas9. If you co-express an MCP coat protein that binds to these MS2 stem loops, fused to two activation domains called a P65 domain, this is derived from the endogenous NF-kappa-B pathway in human cells, along with HSF1 activation domain, which is also an endogenous activation domain, you're now with a single sgRNA recruiting 3 different activation domains. So this one's recruited with the protein, and these 2 are recruited by the RNA. And he showed in this paper that this works quite nicely to activate transcription of endogenous genes, and is another way to multiplex your ability to deliver different domains using dCas9 in a single sgRNA. We had a similar idea at UCSF. We didn't have the crystal structure at the time, so we just appended these stem loops to the C-terminus of an sgRNA. This is work from Jesse Zalatan, who has his own lab at University of Washington now, but the work was done in Wendell Lim's lab at UCSF. And what he was able to show is that, as I've shown you previously, we can use dCas9 to control transcription, either just by using dCas9 alone or by appending protein domains to dCas9. In this setting, the guide RNA that you express only encodes the locus that you're targeting. It doesn't do anything functionally. In contrast, if you start to append these RNA domains to an sgRNA, now the guide RNA encodes both the locus that you wanna target and the action that you wanna perform. So you can express, you can append different hairpin structures to an sgRNA that are targeted to different genomic locations, and with one dCas9,, and both activate and repress transcription at different loci. So that's shown here where you have this purple stem loop that recruits this purple protein fused to this green effector domain, and that's controlled by this gray dCas9, and so this can all be targeted to gene X. And then in the same cell with the same dCas9, you don't need an orthogonal dCas9, you can append a different sgRNA with a different stem loop that recruits a different protein that's fused to a different effector domain to do something different at a different locus. So this increases your ability to use one CRISPR system to control the expression of multiple genes in different ways, in a way that you can imagine could be really useful for a lot of applications. And so the question was, does this work? Jesse fused 3 different stem loops to the sgRNA and showed that each of these, when you co-express a protein that recognizes a stem loop fused to an effector domain, can activate transcription of a reporter gene in yeast so they're quantitatively a bit distinct, but you can see that if you coexpress all these components together, each of these will activate expression of this Venus reporter. This can be used to, he showed that there was no crosstalk between binding pairs, so these are, you can coexpress these in the same cell and they don't encode the wrong functions at the wrong loci, so there's, another way of saying that is they're completely orthogonal, You can use this same sgRNA scaffold general structure to recruit, to amplify your signal. So you can append multiple stem loops to one sgRNA to either recruit 1, 2, or 3 of these effector proteins. And that can have an additive effect on transcription. And then you can also append 2 different RNA stem loops to the same sgRNA. So here we're appending an MS2 stem loop and a PP7 to the same sgRNA,, and in this manner you can recruit either two of the same domain using an RNA or two different domains using one RNA to control transcription. This also works in human cells, so we used the MS2 system or this ComCrab system to show that in human cells with one dCas9, we can control expression of two genes and turn one off and turn the other on.. And so that's shown here, where here we're activating expression of CXCR4, or we're turning off expression of B4-galant1, or we can turn on one gene and turn off the other gene at the same time using one dCas9 scheme. So it just increases your flexibility for controlling transcription across the genome. This works for metabolic control, which you could imagine, like, the ability to control multiple genes dynamically could be really useful for metabolic imaging. And so in yeast, Jesse was able to show that he could tune the output of this violacin pathway. This is a fairly complex pathway that has, it's a 5-gene pathway, so violacin A through D, and depending on which gene is activated, you get production of these different metabolites that have different colors. So this one's pinkish, that one's purple, yellowish, and sort of a cyan. And so what Jesse was interested in is whether you could use a single dCas9 to turn off and on different steps in the pathway and tune bioproduction of these different metabolites. So this works, and he showed that in the absence of any sgRNAs, you get this, the pathway is off, more or less. If you start to activate the pathway, you produce this PV intermediate. If you activate the first two steps, you produce V. If you activate this first step and turn off D, you now produce PDV, which is shown on these HPLC curves as different peaks for these metabolic intermediates. And then if you activate the first two steps and turn off the last step, rather than producing this, you produce that. And so, sort of a toy experiment, but nice ability to show that we can control expression of more than one gene to specify the output of a pathway in an organism that is thought of widely as a fairly useful metabolic model. Organism for bioproduction. So I think some of the questions that remained at this point, and some of these will be answered by Mike, are can we manipulate pathways in the same way that I just showed you for yeast in human cells? Can we use CRISPR-iRNA to control expression of all of the genes encoded in the genome and in that way build a functional genomics platform? I sort of already showed you this, which is can we use CRISPR RNA to control the expression of non-coding RNAs and try to figure out the expression of the many non-coding RNAs that are annotated in mammalian genomes? It's hard to study a non-coding RNA with Cas9, 'cause non-coding RNAs don't have an open reading frame, so there's nothing really to disrupt in the same way that most people are doing genome editing. If you wanna disrupt a non-coding RNA with Cas9, you have to delete a large region of the genome. Both less efficient and also can have unintended consequences if you delete additional regulatory regions. And so we think that CRISPRi is like a very surgical way to just turn off transcription without disrupting large chunks of the genome, and it's useful for studying non-coding RNAs. And then we'd like to start to move towards higher-order perturbations where we're controlling the expression of sets of genes. So I'll end this section by just saying, work at UCSF, we've done a lot of, a lot of genome-scale screens, and this is sort of impinging on Mike, but it's for one specific purpose. Essentially, we've done, this is an outdated slide, we've done tons of screens. This gives us a lot of data which we can use to train more and more sophisticated machine learning algorithms. So this is depicted in a schematic here, what we've done, but essentially the output is more and more sophisticated algorithm predicts which sgRNAs are gonna be highly active for controlling transcription in mammalian cells. And so this is published, I put this slide up only as a resource to say if you're trying to do anything like this, the algorithm that predicts sgRNAs is really important, and we're a little biased, but we think that our current algorithm described by Max Horlbeck in this eLife paper is the best one out there. You can use this algorithm, the sgRNA predictions, with any of the CRISPRa systems that are described by any lab. So it's not restrictive to our CRISPRa system, you can use it with the Broads or with George Church's. But the protospacer predictions are very, very useful in our hands. And just as an illustration of this, when we tested this algorithm on non-coding RNAs, we can see that by qPCR, we're targeting non-coding RNAs that have no phenotype, so we're just purely testing our algorithm here, The vast majority are highly active, and most of them give you robust repression of your target gene. [Speaker:AUDIENCE] So can I just ask a question?

[Speaker:JOSHUA ELKINGTON] Yeah. [Speaker:AUDIENCE] Is that based on the— also, Mac's study of the location relative to the transcription start? So it's not just predicting a good sequence, but a good location.

[Speaker:JOSHUA ELKINGTON] Absolutely. The location is really, really critical. So there are features of the machine learning algorithm that are general. And that would be useful for predicting active guides for Cas9 or dCas9, and they relate largely to transcription of the guides in human cells, and folding and loading into dCas9 or Cas9, so those would be applicable for any experiment using CRISPR. But then there are also a lot of features to this algorithm that are specific for either CRISPRi or CRISPRa, where we're controlling transcription, and those largely relate as Dana pointed out, where you wanna target these complexes relative to the transcription start sites of genes. So our systems are fairly short-range activators and repressors, and so you need an accurate annotation of the transcription start site of the gene to be able to turn it on or off. Another feature that goes into this algorithm, I'm incredibly proud of Max, he's awesome. This is actually Horlbeck, eLife 2016B, There's an A that is really useful. It describes his finding that both Cas9 and dCas9 don't seem to bind either in vitro or in vivo at sites that have a nucleosome occupied. Sorry, that was a very backwards way of saying that. Cas9 can't bind DNA if there's a nucleosome present. 2013, Cas9 had never seen a eukaryotic genome, it had never seen DNA wrapped around a eukaryotic canonical nucleosome, and what Max, together with collaborators from Berkeley, showed both in vitro and in vivo is that if you have DNA wrapped around a nucleosome, Cas9 just doesn't bind at all. So if it doesn't bind, it can't cut, it can't control transcription, it can't do anything. Fortunately, within the human genome, nucleosome breathing or remodeling, and so the average site across the genome at some frequency can be bound by Cas9 and dCas9. But for controlling transcription, in all eukaryotes, the nucleosome positioning around the transcription start site is fairly well phased. So if you look at a well-annotated transcription start site, you can see phased information for nucleosomes downstream from the transcription start site. And so if we just avoid those phased nucleosome positions, we increase our enrichment for active guides that control transcription. Unfortunately for Cas9, nucleosomes aren't phased away from a transcription start site, so the coding exons for most genes are too far from a transcription start site to give us information on avoiding nucleosome positioning. But it is still something to consider for other, more directed experiments. Okay, so everything I've talked about so far today is about using dCas9 to turn transcription on or off using a variety of domains. So we think about this as sort of controlling the transcriptome. We are editing the epigenome, but we're not doing it enzymatically, and our focus up until this point had been more on the ability to control transcription for reasons I don't entirely have time to get into. But a lot of other labs have been interested in using dCas9 as an easy-to-use RNA-guided DNA binding platform to surgically control epigenetic marks. And so there's been a lot of really cool work from a number of labs over the last 2 years that I wanna just summarize over the next 15 or so minutes. So I showed a slide like this yesterday. Basically my only point here is that the eukaryotic nucleus is really complicated. You have DNA wrapped around histones, the histones have a histone tail, and these tails are highly modified with a variety of methylation marks and acetylation marks as well as more exotic derivatives like crotonylation and other things and phosphorylation. But the overall integration of all of these marks at each nucleosome dictates transcription of much of the human genome. Human genome, or much of any mammalian genome. And so this is a really active field of research. You see all kinds of papers saying, you know, this mark at this site does this, whatever. But unfortunately, for the most part, the field of epigenetics up until now has been stuck with either using toy loci, where we have investigated, you know, for years or decades one locus and how the chromatin marks at that locus dictate transcription, or we've used sort of retrospective techniques like ChIP-seq where you fix a cell and you pull things out and you have to correlate the presence of different marks with different transcription statuses in human cells. And so what we'd really like to be able to do is in a living human cell, edit a mark at a specific site and ask how that turns transcription on or off dynamically over time, like through differentiation or through some drug treatment treatment. And so labs have started to do this. So building off of sort of our initial work in yeast and human cells, people have used dCas9 and appended on additional domains. And so some of this work, this work is really all from 2016, where people have, a couple different groups have fused on DNA demethylases to remove methyl marks off of DNA. So I talked entirely I talked about histone marks, but DNA also has epigenetic changes, just methylation, sorry, not changes, that can be dynamically deposited or removed to control transcription. And so to remove or deposit DNA, people have created dCas9 demethylase fusions and dCas9 methyltransferase fusions. People have also fused on histone acetyltransferases and histone demethylases to edit the marks on histones. And so I'll just talk you through some of this very briefly. One of the first papers was from Renee Mayer's lab at UMass, published in 2015, where they fused— actually, this is Charles Gerschbach's lab, not Renee Mayer's lab, sorry. Backwards. They fused the P300 histone acetyltransferase domain to dCas9 and asked where they could deposit acetylation marks at an endogenous gene using, by targeting dCas9 fused to P300 domain to that locus. And so they showed this, that they could deposit acetylation marks by ChIP-seq, but also just asked whether deposition of increased acetylation would activate transcription. So that's shown here, and you can see They can quite robustly use these dCas9 P300 domains to deposit acetylation at these three endogenous genes, and that turns on gene expression. So rather than, like we had shown previously, use scaffolds, now they're directly editing the epigenome to activate transcription. One of the coolest things I think about moving beyond scaffolds towards actual epigenetic enzymes is that we can, can control the chromatin or DNA epigenetic marks at regions very far from genes, so distal enhancers, and this has dramatic effects on gene expression. So this is really the first example of how this might be done using CRISPR. So here we have a really classic locus in the human genome. This is the hemoglobin locus, encodes the different hemoglobin genes that are expressed at different times in your life. So we have fetal hemoglobin and adult hemoglobin and whatnot. They're all encoded at the same locus, so that's shown in these white boxes. There's this classic hemoglobin enhancer that's upstream of these genes, and what the Gerschbach lab was able to show is that if you target this P300 domain to the enhancer, you get, activation of gene expression very distally. So we're editing, we're depositing acetylation marks at an enhancer, which activates transcription of genes that are up to, you know, however many KB that is away, almost 100 KB away. So this is thought to be mostly interacted by the 3D structure of the enhancer interacting with these genes. But it was proof of principle that you could use these tools both to activate transcription alternatively to annotate functional enhancers across the genome by depositing or removing specific chromatin marks. And they did all of the right controls, if you're interested in this, this is published, but basically they showed that if you look at these enhancers or the proximal promoters for these different genes, with the dCas9 VP64 domain, you deposit H3K27me3 K27 acetylation where you target the dCas9 fusion, but not at distal sites. So that's shown by these ChIP experiments here. But if, in contrast, if you target this P300 domain to the enhancer, you see acetylation marks that are deposited very distantly from the genomic location in 2D that you're targeting the dCas9 P300 construct. So it's difficult to know exactly why the VP64 fusions, if they're looped in 3D space, don't induce activation of a gene that's very far from the enhancer that you're targeting, but the data is what it is, and it seems to act very similarly to the way the native systems are annotated to act. And so this has been suggested to be a useful strategy tool for depositing histone marks to regulate gene expression. Similarly, at about the same time, Rene Meyer's lab at UMass fused dCas9 to histone demethylase domains and asked whether you could turn on or off transcription with a histone demethylase. Methylation marks are pretty complicated, so usually we think of if you deposit methylation, you're turning transcription off, but in certain contexts at certain residues on histone tail, depositing methylation can actually turn transcription on. So it's very context-specific, and these are gonna be complicated tools to use. But they started, they created this system, and then they started to try to characterize it using the Oct4 locus in human ES cells, maybe mouse ES cells, I'm not sure which one actually. To be honest. But either way, they were able to show that these LSD1 fusions can activate gene expression if you're targeting them— sorry. Yeah. Basically, they were able to show that you can either activate or repress gene expression depending on where you're targeting and depending on whether methylation turns turns gene expression on or off at both enhancers and promoters, and that this has different, the DNA methylation at enhancers can have different consequences than using a KRAB domain as a scaffold to indirectly recruit enzymes. And it's more of a precise change rather than using a scaffold that brings in an undefined set of enzymes to control the epigenome. And so together, I think these are really valuable methods for depositing or removing acetylation and methylation across chromatin. But they're not the full story in terms of modifying, in terms of epigenetic control. And so, as I said in my intro slide, people have also fused DNA methyltransferases and DNA demethylases to dCas9 to control methylation of DNA. And so this was the first paper that did that. They fused dCas9 to DNMT3A. This is sort of probably the most well-characterized DNA methyltransferase. DNA methyltransferase in human cells. This DNMT3A recruits additional DNMT3L scaffold domains to induce methylation on DNA. What they were able to show by sulfide sequencing or other methods was that if you look at the amount of DNA methylation around the region where you target this fusion protein, you see that for these two loci, with the appropriate controls, you see a large increase in DNA methylation at these endogenous loci. And they were able to show that this can turn gene expression for these genes off, and is a nice example of how we can control DNA methylation to control the output of the genome. And this, I think, this has important implications for studying development, developmental transitions involve DNA methylation, so you've perhaps heard of imprinting, where you have dosage control from a maternal and a paternal allele. This requires DNA methylation for imprinting, and this will be a useful tool for studying that, as well as, you know, generally genome control. DNA methylation is often considered to be deregulated in cancer biology, but we've lacked a tool tools to study it, and so now we have a direct tool for depositing and removing methylation marks at specific loci and asking how that drives specific phenotypes. Similarly, people fused, so last slide was methyltransferases, now we have demethylases. The most well-characterized demethylases in the human genome are the Tet proteins. They catalyze oxidation of methylated cytosines to hydroxymethylated cytosines, which eventually leads to demethylation, and this group was able to show that you could use the same MS2 scheme that I talked about, where they fused MS2 stem loops to the sgRNA. If you then co-express MCP tet core domain fusion protein, you can use dCas9 to demethylate DNA, where dCas9 and the sgRNA now serve scaffold to recruit this TET-MCP fusion protein. And the data are shown here, but essentially you can remove methylation marks and this can activate gene expression. About the same time, a paper from Rudolf Jaenisch's lab came out that did both of these in one paper. So there's, you know, I illustrate this only that a lot of people converged on the same goal, goals at the same time, and it's nice to have different smart people trying to do the same thing, 'cause you end up with options. So similarly, Jaenisch fused DNMT3A and the TET proteins to dCas9, and then asked whether they could turn off or on reporters. And I'm not gonna go through all the data, just because the bottom line is that both, all of these four papers show that this works. Deposit and remove methylation sort of at will. And this can be used to turn on or off gene expression, and by controlling gene expression, drive differentiation in a variety of ES cell models or neuronal differentiation models. One of the more interesting things I thought that they did in this paper was they used this Tet, sorry, What was that? Hard to see. They used this DNMT3, it's hard to keep it straight which way is activating or inactivating. They used their DNMT3 dCas9 fusion to deposit methylation at a CTCF site within the genome. So CTCF is an insulator element that enforces transcriptional boundaries between different regions of the genome and can also influence in three-dimensional genome looping. And so by depositing methylation at these CTCF sites, they broke open a chromosomal loop. So the chromosome was looped together, and that was influencing gene expression, and they deposited DNA methylation, changing the three-dimensional structure of the nucleus. The nuclear 3D structure has been very, very hard to study, all the way most of the ways that people have studied up until now have been, again, retrospective, so you cross-link cells and you pull out regions of the genome and then sequence using Hi-C or 5C, or there's a bunch of different methods, and ask what's next to what. But that doesn't give you the ability to, in a live cell, control those marks or control those 3D loops and ask what does this really do functionally. It all just has to be static time points. So, I'm sort of a mouse cancer biologist by graduate training, and so one of the things I always think about is like, okay, well, this is great in cell culture, but can we actually use this in a more sophisticated model to ask something about how organs are regulated or disease processes arise in a living animal? And it's hard to see here, but what they did was they injected these constructs lentiviruses into the mouse of an— or the brain of an adult mouse, and then depending on what they injected where, they were able to show that you can engineer DNA methylation in the brain of a living mouse. So I think this could be really cool for studying memory formation or addiction or a variety of things that are modulated by DNA methylation, or thought to be modulated by DNA methylation in vivo. All right, so I'll stop on transcription there. Basically, if I have hopefully convinced you of anything, transcription is complicated, and we're more and more able to artificially influence transcription, either by using scaffolds or by using enzymes. And I sort of think of us as doing this now, where we're using single domains to sort of blindly bluntly control transcription, but we're assembling toolboxes, I think, where we can combine these in very sophisticated ways to control transcription or epigenetic marks in a way that is much more similar to the way these things are wired during development. And I think this will be really useful for studying disease and thinking about therapies, and in general, moving forward with moving away from the nucleus being this scary black hole black box that nobody wants to work on because it's too hard to control any of the processes.

[Speaker:AUDIENCE] Can I ask a question? How far from the tRNA binding site do these chemical modifications spread? It seems to me if you methylate one CpG, it's going to have a modest effect, if any. So how do you get some sort of regional modification?

Either epistopes or DNA. Yeah, I think that's a question that's still being answered. So far it looks like the modifications are fairly narrow, but there's a paper that I didn't discuss that describes combining KRAB domains with DNA methyl, DNMT3A fusions and DNMT3L fusions, and there they see that if you temporally order these things where you recruit KRAB and then you methylate, you can induce a window of silencing that both spreads in the same way that normal heterochromatin does, and is permanent. So rather than, everything I showed today, you have to keep expressing the dCas9 fusion proteins. If you remove the expression, the methylation or acetylation marks get reset and transcription goes back on or off. This one paper is the first suggestion that we may be able to permanently control epigenetic marks, and we don't even know whether it's right or not, but if it is, it'll be a— if we can artificially imprint an arbitrary locus and permanently turn gene expression off, or propagate gene silencing across a broad window of a chromosome, that would be really useful for modeling a lot of types of disease. The more straightforward answer to your question is, people only looked at a handful of loci, and mostly using reporters, and so I don't, like, reporters aren't always a faithful mimetic for what's gonna actually be true in the genome, and I don't think you really know what you're doing if you're just studying 4 or 5 loci, so we need some more high-throughput way to measure what, your question. Okay, so in the last, 20 minutes or so, I'm gonna talk about some additional applications that people have used dCas9 for. And so one of the things that we have been involved with at UCSF is using dCas9 to image the epigenome. So like I said, the nucleus is a very complicated place, and people have used, you know, hi-C methods and other reconstruction methods to try to map 3D nuclear architecture, nuclear architecture, but this is generally done in fixed cells. So it would be really nice to be able to have a live cell way of measuring where each piece of every chromosome is relative to each other across differentiation or across different processes. And so we thought, well, why don't we try to see whether we can use dCas9 to enrich a fluorescent signal at a region of the genome, and in this way we might be able to image a genomic locus locus in a live human cell. And this was a collaboration with Stanley Qi and Bo Huang. And the first thing we did was just very simply fused dCas9 on C-terminus, or fused GFP on the C-terminus of dCas9. We used an inducible system to keep our expression levels very, very low to reduce our background, and then we co-expressed guide RNAs off of a third lentiviral construct. To a couple different human cell types. We tend to use cell types that are very flat, these are easier to image. So we're using mostly HeLa cells or RPE cells, which are a transformed retinal pigment epithelial cell line for these experiments. The first question we asked was whether we could image a very repetitive region of the human genome. So what we saw was that one of the, what we initially asked was, okay, we were trying to think of like, okay, what's the most repetitive image of the region of the genome so we can greatly enrich our fluorescent signal? And so we were thinking about centromeres, and I'm like, wait, what about the telomere? The telomere, does it have a GG PAM? And can we design a guide that'll target the human telomere? And so fortunately for us, this is the human telomere repeat sequence described here, there are many hundreds of these repeats at the end of every chromosome. There's a GG PAM here, and so we designed a guide that would bind to the repeats, and thereby with the expression of one sgRNA, we can target many dCas9-GFP fusion proteins to the telomere. And so we did this, and we're hopeful we'd see local enrichment of fluorescent signal that was indicative of dCas9-GFP binding to the telomere, And our initial experiment looked like that, it was not great. So we saw a few spots, but predominantly our dCas9 GFP signal is enriched in the nucleolus, so you can't really see it in this picture, but in the absence, in this experiment at least, a lot of this was enriched in the nucleolus. So one of the reasons, I'm giving this talk a bit out of order, but one of the reasons that we started making changes to the guide RNA to try to increase transcription of the sgRNA or increase stability of the sgRNA, was that we couldn't control transcription with a single lentiviral infection, and we couldn't image the epigenome. And in both applications, what we saw was that our sgRNA expression was the limiting factor in being able to carry out these experiments. And so if you switch to the sgRNA, the modified sgRNA design that I showed like 10 or 15 slides ago, we see that we go from very little signal at the telomere to now nice discrete telomeric foci. And that's quantified here as the number of telomeres that you can image, or here as the percent of the whole cell GFP, so the fraction of GFP signal that's actually at the telomeres. And so in both, in all three measurements, we can see that with our improved or optimized sgRNA, we can image the telomere in a live cell. So what does this actually look like? Looks like this. The telomeres are not super dynamic, they sort of wiggle around, but we can image 'em. So this was a cool proof of principle. People had never really been able to, people had done this with other telomere-binding proteins, but nobody had done it with a programmable DNA-binding protein. And so one thing you can see here is that you get fluorescent bleaching over time, and so that is a, that can be a problem for these experiments. But if you quantify the relative movement of these telomeres, you can see that they, although it's subtle at this resolution, they have different dynamics. So some of them are very confined within the nucleus, others seem to wander around in a random fashion, and still others have a directed 3-dimensional movement that we don't understand, so it was proof of principle that this could work. [Speaker:AUDIENCE] Could you take them through mitosis? [Speaker:JOSHUA ELKINGTON] We did, yep. You can take them through mitosis. Yeah, we didn't live image the entire time. We took snapshots over— we synchronized cells and then followed them through and imaged at specific times across mitosis, but yep. I didn't know what would happen as you get chromosome condensation during mitosis, whether dCas9 would stay bound, because we know that that it has a problem with binding to DNA wrapped around nucleosomes, but at least for the telomere and a variety of other spots, you can see dCas9 stay bound across mitotic progression. So telomeres are pretty specialized, and we wanted to know whether you could image additional regions of the genome, so we started by imaging some other repetitive elements of the human genome, as you can see here, This is the MUC4 locus. This gene encodes mucins that are present in your lung, and this gene, for some whatever reason, has a lot of natural repetitive elements. And so we can target either an exon or an intron that have repetitive elements and show that you get 3 spots. We were really confused by this initially until we karyotyped our cells and we saw that they were triploid at this chromosome. So they were supposed to be diploid and they were not, but better to have that than have some unknown off-target binding site that we couldn't explain. So it was a good thing, not a bad thing. This, you can see, you can image two different regions on the same chromosome that are distant from each other. So you can see sort of two dim spots here. Importantly, and I don't display all the data, we did a lot of work to show that the CRISPR imaging method gives you the same information as with more standard methods like oligo-FISH on genomic locations. Synthetic biologists are obsessed with multiplexing and orthogonal control. So we keep bringing up this idea of multiplexing and orthogonal control. We wanted to ask whether we could use different dCas9 systems bacterial species that bind different sgRNA structures to image two different loci at the same time within one cell. And so we fused dCas9 to a Staphylococcus aureus, we fused Staphylococcus aureus dCas9 with GFP, and then asked whether we could do the same set of experiments imaging either the telomere or these mucin genes, and we see that they're quantitatively a little bit different, the pyogenes version works a bit better, but we can image the genome in living human cells with two different dCas9 fusions. And then if you use two fusions fused to different colors, so here we have a pyogenes-GFP fusion or a Staphylococcus aureus-mCherry fusion, you can see over here on the right that you can label a locus with two different colors. So can you guys see this? Can we turn the lights down a little bit? So that's better. So here, really here's the take-home message for this. If you image, if you guide these two different fluorescent fusions to two different locations along a chromosome, you can image a red and green spot. If you guide them to the same spot, you see yellow, which is as you'd expect, If you guide them to two different spots, you see red and green. So these are guided to two different spots that are 272 kb apart, and you can see that they're, the spots are nicely separated. If you guide them to two spots that are 7 kb apart, you can see some red and green signal, but they're also partly co-localized. So this relates to the fact that these aren't single molecule imaging experiments, these are targeting repetitive elements, and, but definitely we can see What this illustrates is that you can see, you can image two different locations on one chromosome, and in theory, this allows us to image in a live cell the act of chromosome looping. So you can imagine if you targeted a looped region, sort of similar to here, where you have distinct spots and then induced a loop, you'd see those colocalize, and we can measure this in living cells, but we didn't, get all the way there, but that's certainly something you could hope to do. Other groups have done similar things. We returned to the idea of using the RNA to recruit proteins. So another group in Nature Biotech published that you could use multiple different RNA aptamers fused to the sgRNA to recruit fluorescent proteins to image the genome in living human cells. So in this case, they fused MS2P GP7 or BOX-B aptamers to the sgRNA with 2 copies, so that each sgRNA recruits 2 fluorescent proteins, and then showed that you could image the telomere with BFP, GFP, or mCherry. So that on its own is not that useful, but what they were able to show is that you can mix and match these RNA aptamers to create additional colors. You can color blend with blues and greens to form cyans, or greens and reds to form yellows, or green, red, and blue to form white. And so together this gives you 7 colors that you can image up to 7 different loci within the human genome in a living human cell. And so they actually did this, I'm still amazed that they could get all of this expressed in one cell, but with a transient transfection they seemed to be able to. And so that data is shown here, They're not imaging 7, I think they're imaging 6 colors, but they can see if you image 6 different loci, they have different dynamics. So this one sort of doesn't do anything, it stays fairly isolated. This genomic location is wandering all over the place, and we don't really know what that means, but definitely there's a lot of signal here, and it should be useful for imaging nuclear dynamics in a variety of biological applications. So, another way to image, so in these imaging experiments, everything has been done with defined guides, where we express one or a set of guides that are cloned in low throughput. Rebecca Heald's lab here at Berkeley wondered whether you could image a whole chromosomal region. Derives this sort of clever scheme for building a large number of sgRNA expression constructs from a desired region of the genome. And that's described in this paper, but basically you can use PCR to amplify regions of the genome, and then you digest with this specialized restriction scheme such that you can ligate all the digested DNA into an sgRNA format, express thousands of sgRNAs that tile along a chromosomal region. And so here they're doing this in Xenopus frog eggs, and you can see that here's the nucleus, and then they're expressing, I forget the exact number, it's I think on the order of 1,000 sgRNAs targeting this region of chromosome, and you can see a very, very bright spot that chromosomal region is. So I showed you, one of the things that we struggle with in imaging experiments is always fluorescent bleaching. So when I showed you the telomere movie, you could see over time it bleaches. So the more signal you have, the more experiments you can do, or the more exposures you can take. And so this is a way of enriching signal, and is also a very, very cheap way to image a whole chromosomal location. So along the lines of worrying about bleaching, we wondered whether can we— yeah. [INAUDIBLE] —what proteins are there? So like if you did a pulldown for that locus and then used mass spec to measure protein occupancy? I guess, or just like any other methods. For the telomere, we mostly know what's there. Those are very well characterized. So in that sense, yes. For an arbitrary locus, it's been difficult to use dCas9 to pull down a locus and measure the proteins present. 'Cause for every on, you have two on-target binding sites, but we're frequently expressing a lot more than two dCas9 proteins. And they're searching across the genome at putative, you know, low-affinity interaction sites. So anywhere where people have looked by ChIP-seq, anywhere where you have a PAM plus 7 base pairs of homology to the sgRNA, you can get enrichable signal by ChIP-seq, and so that results and thousands of binding sites that are measurable across the genome, and the aggregate signal from the off-target binding sites that are very transient is quantitatively more than the two on-target binding sites. And so it's hard to see the signal from noise. There may be other ways to do that, but people have struggled to do that. Okay, so we wondered whether we could get more sensitive imaging or single-molecule imaging by recruiting multiple fluorescent proteins with one sgRNA. So this is a crude diagram of the SunTag, the scheme I showed you before with our second-generation CRISPRa system where we can fuse to any protein a series of epitopes. If we co-express a single-chain variable fragment fused to another protein, we can enrich either an activation domain or a fluorescent protein at an arbitrary sequence. This works 'cause you can express a lot of repeats and the binding is very, very strong. Tight for the epitope. And if you do FRAP experiments, the binding's very stable. So these systems were being used by Ron Vale's lab to image kinesins. So this is a movie of kinesins moving in a living human cell, where you fuse these epitopes to a kinesin, and then you co-express the SFP-GFP fusion, and you recruit up to 24 GFPs to an individual kinesin. And so if you motion blur out the background, You can image, single molecule image, individual kinases moving within a cell. And so this is an illustration of how we can use a multiplexed enrichment for fluorescent signal to increase our sensitivity and do single molecule imaging. We haven't actually done exactly this with dCas9 for a couple of technical reasons, but the potential is there for doing single molecule imaging of an arbitrary locus of the human genome. Even if you don't have a repetitive sequence. So you can, these are just nice movies, so I included this, but basically these are, this is one, this is a blown-up view of this, and the arrows are tracking a single kinesin imaging, moving along a microtubule. So pretty cool, I was sort of new to the field of single-cell imaging, single-molecule imaging, but there's a lot of interesting cell biology you can do. This can be used with dCas9 as well to greatly enrich your sensitivity. So this is our version 1 CRISPR imaging system where we're imaging telomeres, you see spots. This is version 2 where we have, we're with each sgRNA recruiting 24 GFP molecules, and you can see that the signal's basically blown out. So we have way more sensitivity recruit many fluorescent proteins with a single sgRNA. So moving out of the nucleus, you might wonder, well, what else could we use these programmable binding proteins for, and could we use this to track RNAs as well as, and not just DNAs? And this is really a special interest of Jennifer's, actually, and so you can, there are reviews published about how RNA localization localization impacts a lot of different kinds of biology. And people for the last 15 or 20 years have been using these RNA aptamers like MS2 that I've talked about multiple times today to recruit fluorescent proteins to RNAs to either image them or to pull them down and study what's associated with them. And so these are summarized here, but we wanted to know, well, could we use a CRISPR system in a similar way bind to an unmodified RNA. So all these RNAs have to be modified where you append multiple artificial binding sequences. And so there's always this issue of, well, you changed the RNA, is this actually really what's happening, or do you have some artifact that's associated with these chimeric RNAs? And so Jennifer's lab around 2014 started to explore whether you could use Cas9 to bind to RNA molecules. And what they were able to show is that if you trick Cas9 by providing a, what they call a PAMmer in trans, so Cas9 normally binds double-stranded DNA, right? It's not supposed to bind single-strand DNA or single-strand RNA, but if you provide a PAMmer in trans, and that's detailed in this fairly complicated experiment, you can get Cas9 to bind to single-stranded RNA and cut it. And that's shown down here where you, if you apply a PAMmer in trans, it'll bind and cut. This is a similar experiment using dCas9 showing that you get binding of dCas9 to single-stranded RNA only in the presence of the PAMmer. So here's binding to double-stranded DNA, you can see it's very robust. Here's binding to single-stranded RNA, it's very, very weak. If you add the PAMmer to trick it, trick Cas9 into thinking it's binding double-strand DNA by forming a duplex between the PAMmer and the single-strand RNA, now Cas9 will bind and cut. And so this was used early on for IPs, so you can immunofinity purify cellular RNAs using a Cas9 plus a PAM, and this works well works well for high-abundance RNAs, and is, I think, in my understanding, is still a work in progress for very low-abundance RNAs and intermediates, or sort of in between. But it's a new tool for pulling down endogenous cellular RNAs to study what the set of proteins that are associated with them, or a variety of other things. We wanna do more than just IP RNAs, probably, and so in collaboration with Jennifer Jean Yeo's, Yeo, I don't know how to say his name, lab at UCSD showed that you could use a similar Cas9 scheme where you're supplying both Cas9 or dCas9 plus a PAMmer to image cytoplasmic mRNAs. So the data is shown here. Essentially what they show is that if you express a dCas9-GFP fusion, with a nuclear localization, you see that most of it goes to the nucleus. If you express an sgRNA with the dCas9-GFP, plus a PAMmer that targets this GAPDH mRNA, you can see that you get relocalization of this dCas9-GFP fusion out of the nucleus into the cytosol. So here you have very little GFP signal in the nucleus, and most of it's cytosolic. We did a lot of controls to show that the GFP signal in the cytosol is predominantly bound to the target GAPDH mRNA. And so you can do a little bit more sensitive microscopy to track actin localization, and that's shown here, where you can see actin localization in the cytosol. I think this remains a bit of a work in progress, in that for these technologies, you can definitely pull down and visualize abundant RNAs. Very low-abundance RNAs remain a challenge. But certainly promises to be an additional kind of application for what these could be used for. Personally, I think that Cas9 will probably not be the winner for imaging or pulling down RNAs. I tend to think that there's gonna be a different CRISPR protein that'll be easier to use for this application. And rather than having to apply this PAMmer at high levels in trans, why not just use a CRISPR protein that's hardwired to bind single-stranded RNA rather than double-stranded DNA. And so this is already starting to be done. This is work from Feng Zhang's lab and Jennifer Doudna's lab, where Feng, together with Eugene Koonin, characterized additional CRISPR proteins and then, through sort of a smart bioinformatics strategy, looked for proteins that might bind RNAs. And what they found was this protein called C2C2. And they were able to show in this paper that C2C2 binds RNA binds to single-stranded RNAs in an RNA-dependent fashion, so it's an RNA-guided RNA-binding protein. And when it binds to target RNAs, it activates a nonspecific RNase activity that chews up all of the RNAs that are locally around this complex. So this is thought to be another type of phage defense system where the bacteria can recognize invading RNA genomes in an adaptive way that's dependent on the sgRNA, and if it recognizes those, rather than cleaving, you know, rather than just cleaving in them, the bacteria in a sense commits suicide, where it just chews up all of its RNAs and the cell dies rather than allowing the phage to propagate. Jennifer and later Feng was able to show that sort of this process system can also be repurposed for biotechnology applications. So she showed that if you use an optimized fluorescent probe, you can detect specific cellular RNAs using C2C2 by a fluorescence assay. So what's sort of amazing is that this system can detect RNAs even down as low as like 10 picomolar in abundance., and this is borne out by this second paper as well. But I think this promises to be very useful for diagnostics in a lot of applications where you might be wanting to fluorescently visualize the presence of an infection within a biological sample, or the expression of specific RNAs associated with cancer progression, or, you know, you can imagine a lot of applications. Okay, so I'd like to end with one additional application of dCas9 that's been pretty popular over the last year or so. And some of this work was done by Mike, I'm not sure if you're gonna actually talk about this, so I, just a little bit, I'll just say two words about it and Mike can do the real justice to it. Essentially, there's been this desire to ask whether we can mutate the genome without inducing double-stranded breaks. Stranded breaks, so double-stranded breaks are a bit messy, they can induce translocations, we don't really have currently the ability to specify the DNA repair pathways that repair these double-stranded breaks, so the repair events, you know, you can tune them between HDR and NHEJ, but they're still largely beyond our control. And so people started to ask, well, can we just use dCas9 fused to additional domains to mutate the genome in a specific way? This is a process that happens normally during B-cell development and antibody maturation in all of us. And so what David Liu's lab and Mike Bassik's lab showed last year was that you could fuse dCas9 to these cytidine deaminases, either from rat or the human AID protein that's involved in antibody maturation, and then mutate the genome without forming forming double-stranded breaks. So these are targeted mutations that are, these are mutations that are induced in an RNA-guided way, and you predominantly, either depending on the fusion proteins you use, get C to T mutations, or in Mike's paper, they showed that you can get C to N mutations. And so this is, I think, an additional application of dCas9 people seem to be quite excited about, and there was sort of an explosion of papers in Nature Biotech a couple weeks ago where people used one of these two systems to implement DNA editing without inducing double-stranded breaks in a variety of types of organisms, and showed a lot of applications. David Liu's initial work showed that you can also fuse on additional domains to increase your rate of editing, so they fused on a uracil DNA glycosylase inhibitor onto dCas9, and this increases the conversion of mutated base pairs, and that's detailed in the paper, I won't go through it for time, but both David and Mike have shown that this can be used to study anti-cancer drugs, or to, in theory, in a very specific way, mutate disease alleles from their pathogenic sequence back to a wild-type sequence, which could have therapeutic applications. And so I'll end there and just say I hope that I've convinced you that as awesome as Cas9 is, there is a place for thinking even beyond Cas9's canonical activity and using repurposed forms of CRISPR enzymes to do additional things in bacteria, yeast, and human cells. And we're pretty excited about that. So thanks.