Scientist Stories: Howard Chang, Genome Regulation by Long Noncoding RNAs

Good afternoon, everyone. The first day of June and a very warm Wednesday, and welcome to all of you to Mazur Auditorium and to those who are watching remotely on the web. We have a treat today in terms of our Wednesday afternoon lecture series, that being Dr. Howard Chang, who's going to talk to us about genome regulation by long non-coding RNAs. Howard got his undergraduate degree at Harvard, his MD-PhD at Harvard and MIT, with his PhD component serving under the mentorship of no less than Professor David Baltimore. After that, he moved to Stanford, where he's been in the Department of Dermatology since 2004 and currently serves as professor as well as director of the Center for Personal Dynamic Regulomes. Would you like to be the center of the Personal Dynamic Regulomes Director? That's a pretty cool title. He has been honored in many ways by recognitions including the Paul Marks Prize for cancer research last year. And his work, which I'm sure he's going to tell you about, relates to a fascinating and somewhat unanticipated feature of the human genome, the existence of large numbers of long non-coding RNAs whose existence was not entirely anticipated, at least not in the numbers that they are now clearly represented in the human genome, and whose function has been the source of many people's interest over the course of the last decade. And I think no one has done more than Howard to begin to elucidate what the functions are of these particularly important nucleic acids. So please join me in welcoming to the podium Dr. Howard Chang.

Good afternoon. Can you all hear me? Yes. Okay. Well, thank you, Dr. Collins, for that kind introduction. Thank you to Dr. Tyson for bringing me here. And I want to tell you today about long non-coding RNAs, and my talk will be in 3 parts. First, I'll give you an overview of the field. Second, I'll tell you about some new progress looking at the long non-coding RNAs involved in sex. And third, I'll talk to you about the evolution of long non-coding RNAs. Okay. So the new century started with a biological mystery. As the sequences were rolling off our sequencers with the completion of the human genome, we had these vast array of transcripts. And that quickly showed that there were two flavors. One set, which were the classic protein coding genes, behaved very well. You can put them into the computer and quickly realize what kind of enzyme domains, what kind of functions they may be, and off you are to a fairly intelligent experiment. But on the other hand, there were these large numbers of transcripts which ultimately got named long non-coding RNAs, which I'll pronounce the abbreviation as lncRNAs. So these are transcripts that are defined to be greater than 200 nucleotides. They're made by RNA polymerase II. They're capped. They're spliced, sometimes polyadenylated. And they appear not to have conserved open reading frames. And so their function, if any, was a large mystery. And so the active questions in the field were, do they have any functions? Could they be involved in human disease? And finally, the long-term goal, how can we systematically decode the information content so we can actually understand function from the primary sequence alone? So as we go through this journey, I think that the challenge we face at a time may be analogous to another scientific mystery, one maybe some of you are familiar with. So what I'm showing you here are Egyptian hieroglyphics, the meaning of which was lost for many years until Napoleon's soldiers found this tablet, the Rosetta Stone, which had the same text written 3 times. Once in hieroglyphics, once in Egyptian, and once in Greek. And by systematic correspondence between these 3 texts, the archaeologists were able to work out this code of transformation. And so breaking this code then allowed this text, not only this text to be read, but also all the ancient texts written in hieroglyphics. So this analogy is, I think, very apt for our task because RNA, like these hieroglyphics, fold into complicated shapes, and really our job is to systematically associate form with function, and that's what I'm going to be telling you about. Okay, so in my lab, we've been pursuing a potential function of lincRNAs by discovering novel lncRNAs by their patterns of expression during cancer or disease states, understanding their functions at a molecular level or actually better in vivo, and finally, making technologies, tools to define key features in the lncRNA and their associated partners, nucleic acids or proteins, that define their function. Okay. And so, our foray into this field started from my clinical interest, and that was understanding how cells in different positions of the body acquire their positional identity. And I was interested in that question in the skin. So if you look in the mirror or look around, you'll see that skin, even though it's the largest organ that covers your entire body, is not the same everywhere. Some of us have long hairs on our scalps, but I hope none of you have long hairs on your palms and soles, even though it is one organ. So clearly these cells, know their location in the body, and know how to behave. And so we're interested in the mechanism, the positional memory, and genes that will be expressed in a very position-specific manner. Now, in our analysis, in the search for such genes, we use a tool, novel at the time, called a tiling array, and this led to the initial discovery of a copious amount of these lncRNAs that had a position-specific expression. And also, our initial analysis showed that these lncRNAs did more than what people suspected. So, in the past, a few lincRNAs were known, and we're going to talk to you much more about this, such as Xist, involved in dosage compensation of the X chromosome, lincRNAs involved in imprinting, and those always acted on the neighboring genes, that is in cis. And they would recruit chromatin remodeling factors, as shown here. And the RNA that we discovered, called hotAIR, which expressed only in certain positions in the body, actually turned out to act at a distance. Hot Air will bind to a chromatin modeling complex called polycomb and target polycomb to hundreds of sites throughout the genome, so action at a distance. The function of Hot Air in the animal was revealed when we knocked out Hot Air, and which showed then this kind of a birth defect, a homeotic transformation. On the left, I'm showing you basically CT scans of the spine showing the morphologies from the L1 spine all the way to L6. This L6 spine has these two lateral processes. This is the sacral S1 spine. The iliac crest is joined here. In the HOTAIR knockout, there's a change in positional information so that this last segment, the L6, has been turned to S1. And so you still see the two spines here, but now the iliac crest is joined one level above. And so this transformation of identity is due to a deep repression of HOXD activity. These are a family of transcription factors whose expression determine segment identity. And so in the hot air knockout, we see expansion, for example, of the HOXD10 expression domain and increased expression of HOXD11. And hence, one of the normal functions, we believe, of hot air in the body is to maintain the positional identity, maintain the positional information. So it turns out that this process, this mechanism, could be subverted in disease, and this was very clearly shown in the case of cancer. A few years after discovery of HOT-AIR, my lab found that in early human breast cancers, approximately a quarter of the patients have very high levels of HOT-AIR, and high meaning 250-fold or more above normal. Those patients then had a much higher risk to develop metastasis and eventual death from their disease, as shown in this Kaplan-Meier survival curve here. So this RNA, lincRNA, is a very powerful biomarker, very powerful predictor of patient prognosis. So it turns out that HOT-AIR is not only a biomarker, it's actually a driver of metastasis. If we take human cancer cells, program them with the level of expression of HOT-AIR seen in patients, and introduce them into the animal model, as you can see here, these cells are now much more able to metastasize to the lung and cause disease progression. So the overall idea is that this is a system that normally that indicates positional identity, and cancer cells basically subvert the system to convince themselves, to reprogram themselves to think they belong somewhere else in the body. And by this process, they can access the genetic mechanism to spread. So since that time, high levels of HOT-AIR has been shown to be a poor prognosis factor in 26 different human tumor types, and that there are now actually FDA-approved diagnostics based on lincRNA for cancer prediction diagnosis, first in 2013. So it's a very active area of investigation. Now, the way we come to know how lincRNAs would act was by actually mapping where they would actually interact with the genome. And so here we were inspired by chromatin immunoprecipitation, ChIP assays, which were used to define protein-DNA interactions. And so the analogous technology we developed was named ChIRP, where the R stands for RNA. This is a technology where we can take living cells, cross-link the chromatin and the RNA together, retrieve the RNA of interest with all of the nucleotides that are complementary to that RNA, and then retrieve the associated DNA or proteins, and which can be subject to deep sequencing or mass spectrometry. So, this technology showed us that hot air indeed co-localizes and guides polycomb occupancy, and in fact, the RNA goes first, and that the polycomb occupancy depends on the RNA. There is a specific sequence motif that's enriched at HOT AIR target sites. And so, there's an interdependence between the RNA and chromatin remodeling factors. And it's important to point out that many of these chromatin remodeling enzymes actually lack intrinsic DNA binding specificity. And hence, these RNA can serve as one of the guides to take them to specific locations in the genome. So, extending beyond that idea, We later found that the RNA, a lincRNA, can do actually more than one job. Early embryonic development, histone modification— histones would actually have both two kinds of modifications, one indicating activation, such as histone lysine 4, K4 methylation, and also one indicating silencing called K27 methylation. So, the presence of both markers is called the bivalent domain. It's a poised state. We found that HOT-AIR actually acts as a scaffold by interacting with two different enzyme complexes. The 5' end of HOT-AIR interacts with polycomb, shown here, and the 3' end interacted with a lysine demethylase, the LSD1 complex, which then removes the K4 methylation. So, these two actions together, one removing K4 and one placing the K27 mark, would then enforce a univalent, very strong signal for silencing, and also then shows the logic of how a modular lncRNA organization can drive enforced proximity and targeting. So summarizing many sort of studies now in the literature, we now know that lncRNAs don't have a single mode of action. They actually have several lncRNA mechanisms. But they can be understood by pairwise combinations of lncRNA protein and nucleic acid interactions. Some lncRNAs act as guides, as I just showed you, taking histone modification complex to specific locations. The part of the RNA dictates the targeting, and they actually— the cargo dictates the effect, either activation or silencing. So, this modular design is also now frequently used in designing, for example, CRISPR-based guide RNA, synthetic guide RNAs. Other lncRNAs can act as a scaffold, and that is bring— multiple enzymes together for basically proximity to each other or joint targeting. Yet other lincRNAs literally link three-dimensional chromosome conformation, taking basically enhancers to their target sites. And finally, when the regulation is done, there are other lincRNAs that act as decoys. They can actually be expressed, they bind to a factor of interest, and then move it away from the genome, basically terminating gene regulation. So, these kinds of molecular mechanisms were very exciting, and they were quickly then followed by, I think, very nice examples from actually genetics demonstrating the functions of lincRNAs. So, what I'm showing you here is a very nice collaboration between my lab and the lab of Carla Kirkgaard at Stanford, where a lincRNA was found to be responsible for a genetic locus involving infection control, we went from a purely forward genetic approach. So to keep the story kind of straightforward, there are two strains of mice initially characterized at the Pasteur Institute in Paris. One strain would get infected by certain organisms and they would all die. One strain would get infected and they would all live. So it's a very simple phenotype. And so people are trying to— would then perform forward genetic crosses trying to map out the locus of interest. So this project went on for about 10 years and was a really hard project because when people kept mapping down the locus, there was actually not a protein coding gene there. Okay, so this was very discouraging for the people involved until finally, I think, the concept of lincRNAs came around. So the locus in question was here in mouse chromosome 10, shown here, and in humans in chromosome 12 right next to interferon gamma and a number of cytokine genes and also variation— genetic variants in this location were actually a main— is one of the major hotspots of basically proclivities for a number of human diseases shown on the graph. And this actually affects intron— the last intron of human NEST. This is the name of the lincRNA that we eventually gave this locus. So, it turns out that this is a defect in a lincRNA called NEST. NEST is needed basically to activate interferon gamma through actually activation of K4 methylation. And the formal genetic proof came from a transgene encoding NEST was able to rescue basically the deficiency phenotype, showing that this is in fact the gene responsible. So I think this example nicely demonstrated that natural variation in lincRNA activity can confer disease susceptibility. So what are other ways in which basically our sequence or information in lincRNAs can be altered to affect biology? What we learned is that there are actually another class of variants which change RNA structure. And so, these have been named ribosnitches. Ribosnitches, ribo because it's RNA, and S is supposed to be a SNP, but it's a snitch because it changed the structure. So, this came from our analysis of the heritability of RNA structure, looking at, for example, this family trios. And we can then see that whenever one— when this allele was in this pattern with an A, This is the shape, and if this position was a C, then this is an alternative shape. So, such ribosnitches that change RNA structure are a minority, of course, of all the SNPs, but actually, in fact, we discovered that many of the disease-linked non-coding variants were actually ribosnitches, and that provided a potential mechanism for these orphaned variants that previously lacked a biochemical explanation. More recently, we found that another way to change a link— a non-coding RNA meaning without actually changing the primary sequence is through RNA chemical modification. What I'm showing you here is basically the most prevalent RNA modification called N6-methyladenosine or m6A. What we discovered is that whenever this modification was placed, it caused a double-stranded region to actually spring open into a single-stranded region. The biophysical basis of this is that every loc— place where this m6A is placed causes a steric clash if you force it into a double-stranded conformation. This has the consequence of destabilizing RNAs if it's at a certain position and allowing their timely turnover. So embryonic stem cells, what we found when we knocked out the m6A writer enzyme was that the m— these cells would basically get stuck as embryonic stem cells. They could never exit the stem cell state. The reason is because that they cannot forget the past. They cannot turn over the transcripts encoding stem cell fate, and therefore they can never actually embrace the future of the new phase in differentiation. So this concept then led to the idea that these modifications would be a kind of anti-epigenetic mechanism, a necessary process for cells to forget. This mechanism, we believe, would also have implications in cancer. When we basically create— took these cells and injected them into animal models, we see that these cells are incapable of forgetting the past. These metal-3 knockouts form these huge tumors that was even bigger than the entire slide, and they're very primitive cells reflecting their stem cell origin. So, extending along this concept of changes in structure changing meaning, our latest set of findings have to do with the idea that RNA helicases could actually act as transcription factors. This sounds very strange, but this reflects really the intimate connection between regulatory RNAs on chromatin. And hence, if you want to change information on chromatin, you can actually use an RNA helicase that would occupy chromatin just the same as transcription factors, but will actually change the meaning of the RNA and therefore change the outcome in gene expression. And so, there are two examples that we came across. First example, all in the last year. First example is an RNA helicase called DDX21 acting on RNA called 7SK and affecting the transcription of ribosome protein genes. So, this is actually the binding pattern of DDX21. Looks very much like a transcription factor. The second example is another helicase called DDX5 working on a lincRNA called rmRP, which is critical for making a kind of inflammatory T cells called Th17 and actually very necessary for for basically disease progression in certain models, for example, of colitis. Okay, so this is a very interesting sort of flow of information where we start from a protein to an RNA controlling DNA, essentially reversing the sort of the canonical dogma of the flow of information, another surprise from the lincRNA world. Okay, so I would like to now change gears a little bit and focus my talk on a, you know, I think really a classical Lingardnet, and I hope I'm allowed to use the government network to show this, but I'm showing, of course, two famous Americans, two icons, and you can see that in this, and so they're both actually showing a dance move called brush off the shoulder, okay? But you can see that this high-impact journal happens to be really focusing more on her and not so much on our Commander-in-Chief, because he's kind of out of the picture here. But perhaps this is all appropriate because only Beyoncé has a very powerful epigenetic system, and that is the X chromosome inactivation. Okay, so of course, all of you know that this is a very special process. So let me just remind you, of course, that males have an X and a Y chromosome, and females have two X chromosomes. And to make the gene expression dosage equal in mammals, one of the two X chromosomes in females is epigenetically silenced. This is achieved by a long non-coding RNA called Xist that's transcribed only from the inactive X chromosome. It somehow spreads across the entire inactive X and basically causes permanent epigenetic silencing for that cell and all of its progenies. This is a very classic example of somatic cell epigenetics. Now, it's important to know that some genes escape X inactivation. And so therefore, that causes a fundamental dosage difference between males and females. So one of the ways that we've now started to study this problem is using a new technology for access, basically, the regulatory landscape in the genome. And this technology is called ATAC-seq, or assay of transposase-accessible chromatin. So I'll briefly explain. This is a technology that my lab and the lab of Will Greenleaf at Stanford invented a couple years ago, and essentially you can think of it as spray painting the genome with an enzyme. This is a transposase, which is an enzyme that copies and pastes DNA. And we've loaded this enzyme with adapters, sequencing adapters I use for sequencing. And the idea is that in humans, in mammalian cells, in every human cell, 2 meters of DNA is packed into a 10-micron nucleus. So most of the DNA is packed away and not accessible. Only the active regulatory elements are accessible and hence can be spray-painted or tagged by our transposase. So, in a single step, you biochemically covalently inserted the tag into the desired locations, and then you can just amplify and sequence. So, this technology allowed a million-fold improvement in the sensitivity and a hundred-fold improvement in the speed of conducting this type of epigenomic studies. Let me show you some primary data. So, this is looking at a particular location in the genome. The x-axis is basically different location of different genes. The y-axis indicates each— these peaks indicates an accessible element. Okay? So, this is the classic technology, DNase hypersensitivity, done with 10 million cells. Each of these peaks is an active element. This is the first version of ATAC-seq, 50,000 cells. You can see the pattern looks identical. And last year, we think this technology went to its ultimate resolution, single-cell ATAC-seq. So this is now the summation of several hundred single cells, again, reproducing this beautiful pattern. What is nice now is that if you zoom in on this data, you can actually now see the variation at the single-cell level. Every row here at the bottom is an individual cell— is a single cell. And so for every position in genome, we see there are 0, 1, or 2 reads because it's a diploid genome. So this analog kind of information has turned into digital data. From this kind of analysis. So using the kind of technology, we recently surveyed the individuality and the variation in the regulums of healthy donors in CD4-positive T cells. So we found that in these healthy donors— and this is a longitudinal study, so we had these same people come back again and again, and we're looking for sources of variation. And you can see that basically a vast majority of elements— I would say 90% plus, you know, maybe 95% or so— were actually quite stable. Only a minority were varying across individuals. But among these, we can then map to different demographic factors that we can associate. And to our surprise, or perhaps not, we found that the most significant factor was actually the sex of the donors. And this pervaded through all the sex chromosomes, but also the autosomes. So it was initially a surprise, but as we read more, we realized that these are actually very strong literature and epidemiological data connecting sex and differences in the immune system, particularly autoimmunity. This is a graph showing the sex preponderance of different diseases. You can see there's a very— here females is marked blue. You can see there's a very strong trend here. You know that 80% of all patients with autoimmune disease are women, and some diseases like lupus, the ratio is more like 9 to 1. For Sjögren's, 19 to 1. And so sex chromosome changes are long thought to be postulated, but the actual elements involved, right, and the mechanism were actually still under investigation. So let me show you some data of a sex-specific regulome. So if we say, for example, look at an autosome, comparing all male ATAC-seq data from human and all female ATAC-seq data, we see actually a very nice 45-degree diagonal, meaning that They're essentially very similar. Okay? But if we do this on the X chromosome, we see very interesting three different patterns. In the gray spots are all the elements that undergo X inactivation. So they're dosage compensated. So, again, female here, male here. So they're now basically on a slope of 1 even though females have two copies of the X chromosome. We see a second regime of elements that have a slope of 2 shown here. And these are all the elements that escape X inactivation. So we basically rediscover the ones, the elements, the genes that are known to escape, as well as identify new elements that escape, which we subsequently validated. Thirdly, we have to have a set of elements that have an infinite slope, and they're only present in the inactive X chromosome and not present in males. So that includes Xist and several other elements. So this was a nice start, okay, but in these human cells, we can only infer the escape by the dosage of 2 copies versus 1, it definitely does show that the ATAC-seq data is very sensitive, precise to tell that apart. We want to directly visualize the 2 copies in the same nucleus doing something different. And so to that end, we went on to a mouse experiment where we can actually follow allele-specific information. So we're looking now at an F1 mouse coming from a cross from two highly divergent strains shown there, 129 and castaneus, where we have a presence of a single nucleotide variant approximately every 100 base pairs. So when we do ATAC-seq, not only do we have the ends where the transposase is accessible, when you sequence in, you simultaneously genotype the allele that you've inserted. And so therefore, you can have allele-specific information and you can track paternal versus maternal regulome in any cell type. Okay. So this is, of course, a technology that's very suitable to study also human variation, disease-associated SNPs or cancer variants and so forth. Okay. So now, X inactivation co-occurs with differentiation. So embryonic stem cells— this is the entire X chromosome. In embryonic stem cells, we have two active X chromosomes. Here's the Cas9 allele. Here's the 129 allele, but when they differentiate, now one chromosome stays active and the other chromosome gets shut down and it loses globally accessibility, except for these peaks, which are all the escapees. So this high-resolution map gave us now really the detailed information of how these elements escape, and I'm gonna combine that with another technology called chromosome conformation capture, which maps basically the spatial proximity of DNA elements that might be far apart in a linear genome, but when they come close together in space, can be captured and sequenced. So, the way to look at this map is that this is basically the coordinates of the chromosome again from left to right, and then again, basically from top to bottom. And so, the diagonal is the element interacting with itself, and then these off-diagonal squares are long-range contacts. And I'm showing you three different viewpoints, the whole chromosome. You can see the whole— basically, the active X chromosome here across the whole chromosome, and then basically a 100-megabase scale. So you see these— each of these blocks approximately a megabase. These are the topologically associated domains, long-range contacts. And in a further zoom, you can actually see individual enhancer-promoter contact points. That's on the active X chromosome. But on the inactive X chromosome,, we see something completely different. Now, the entire chromosome has been segregated into two large mega-domains, okay? So there's now extensive long-range contact across the entire chromosome. The TADs, the topological associated domains, have been removed, so they no longer exist, and instead, it's replaced by this kind of a bilobe-shaped pattern. What we further discovered is that this boundary element occurs at this repeat called the DXZ4 repeat. And that this long-range contact has a logic. When genes escape X chromosome inactivation, it's not the entire gene that escapes. Only the promoter escapes. The enhancers do not. And instead, the escapee promoters start having long-range contact with each other and backing each other up. So we think that might be one of the reasons for this kind of long-range mega-domain structure, a very unusual structure. Okay, now this is a collaboration, I should mention, with the labs of Job Dekker and Edith Heard. And so, of course, we want to know what's the function of this boundary, so we went on to perform allele-specific deletion of that boundary and repeated all of the analyses. So this is indeed a boundary, because if we remove it, now these two mega-domains no longer form. Instead, they fuse into one giant long-range contact structure, into one giant domain. The TADs do not come back. But instead, what is influenced is the rate of facultative escape, and so that is that— in different cells, some elements always escape, that's all constitutive. Some elements only escape in some cells. And so we see a correlation between the frequency of facultative escape and accessibility. And that is that, for example, in the wild-type situation, we see that this is the boundary. These are elements that escape X-inactivation, and they're also transcribed. In this deletion, we see that there's a lack of accessibility and also a lack of expression indicating that there's now an influence perhaps on escape. These elements that get affected are most frequently proximal CTCF sites near promoters. So, again, we think there's a role in these proximal accessible sites in mediating this long-range contact. Okay. So, what's happening then on this chromosome is that this active X chromosome with these lots of topological associated domains are under the influence of a long non-coding RNA exists turn to this bipartite megadomain structure with a boundary. The boundary is needed to maintain the structure. And we think there's an influence of the boundary and the structure somehow on the frequency of facultative escape. We don't know if there's actual physical contact between boundary or not, but then the only parts that escape are the promoters, the escapees, which then have this unique sort of structure for their continued expression. Okay. So, I've told you quite a bit about the— in detail about how the DNA template reacts to this kind of epigenetic reprogramming. So, now I'm going to turn my focus to the RNA and its associated factors for how this lincRNA exists, does its job. So, the approach we used was actually very similar to what I mentioned to you before, was that the factors that were associated with— exist, again, a long— 17 kb long RNA has been under active been under investigation for many years, but it was a difficult problem. And so, we went at it with an RNA-centric strategy using ChIRP again to pull down the RNA from— against the endogenous RNA complex and subject the complex to mass spectrometry. So, as we reported last year, we discovered then that this exist particle is comprised of 81 proteins, and that's assembled in a modular way. And I'll just show you some highlights of how we figure out what is the most important factor. So this RNA has been studied extensively in the past, so it's known that even though it's a very long RNA, there's a small region, approximately 700 base pairs in the 5' end called the A repeat, that's absolutely necessary for silencing. And so, in this A repeat mutant, the RNA can still be transcribed, still codes the entire X chromosome, but the genes don't get silenced. So there's got to be something very important on this A repeat. So, we performed ChIRP-MS on wild-type Xist and also the A repeat mutant. This is the scatter plot showing you, again, the wild-type versus the mutant. You can see that everything is on the 45-degree diagonal except for 3 proteins whose peptide count falls to 0. Okay, so these 3 proteins absolutely need the A repeat for binding, and hence must be something important for silencing. So, we've narrowed from 81 proteins down to 3, much more manageable. Now we knocked down each of the three proteins, and we discovered that this factor named SpEN was the key factor for mediating A repeat-dependent exist silencing. This is a factor that had never been previously implicated in X inactivation, but it's actually a very interesting gene. It's a gene first discovered in Drosophila, has the name split ends as a developmental mutant, and was later then rediscovered in human cells from biochemical approaches identifying this protein as part of the— basically a co-repressor complex along with histone deacetylases, part of the NERD complex. And so this protein has a very interesting structure. On the N-terminal side, 4 professional RNA-binding domains called RRM repeats. On the C-terminal side, a Spock domain that actually recruits it into the NERD complex. We later showed that the RRM domains directly physically interact with the A-repeat, And so, this is a very nice biochemical link from the lincRNA to then basically a repressive chromatin modification. To summarize what we learned about the exist particle, I just want to tell you that this is a modularly assembled and developmentally programmed complex. It's a very long RNA, but the proteins on it assemble in two steps. In embryonic stem cells, before X-inactivation happens, Polycomb repressor group 1, PRC1 complex proteins already exist, as are other proteins, but parts of the complex is missing. Only with differentiation does SPEN and other factors basically get on to exist, and that completes the structure. And there's actually a logic to this bipartite structure, and that is that this right-hand part of the structure basically is involved in maintaining the status quo. If a locus is active, it will stay active. If it's silent, it will keep it silent. What is needed to basically recreate an active lincRNA complex was actually this SpEN complex, which actually deactivates active loci. Once you deactivate the active loci, you can then put it into the status quo module and keep it in that silent state in perpetuity. Okay? So, there's a logic to this modularly assembled developmentally programmed lincRNA complex. So we said a lot about the proteins, but I'm drawing this RNA like a long line. Okay, is this really how it looks? Of course you know that that's not true. So I want to tell you more about really an effort to understand lincRNAs really at the level I think of its functional meaning, and that is at the level of the RNA structure. So I think all of you know that RNA being a single-stranded polymer can actually fold by base pairing with itself and with other nucleic acids into very interesting and complicated structures. And so the RNA structure, the structure around the entire set of RNA structures is a long-term goal for many investigators. But what we know so far, at least to date, has been largely what can be done in vitro, often short RNAs and naked RNAs in isolation. And one of the long-term goals in my lab has been to get RNA structural information in living cells that are in the full-length context and with our— together with RNA-binding proteins. I would further say that a number of the methods in the recent past are characterized as one-dimensional methods, and that is that you can tell for every base whether it's single or double-stranded, but the double-stranded regions, how exactly they pair up, was actually based on prediction. It's inferred. It's actually not based on experimental evidence. 'Cause that was a sort of a shortcoming that we had hoped to solve with the next sort of thing, advance I'm gonna tell you. So I think I was very fortunate that I was trained as a dermatologist and now I'm working on RNA biochemistry because I knew a lot about this next chemical called psoralen. Psoralen is a very interesting chemical because it cross-links nucleic acids in a reversible manner, but also used in, actually in humans as a medicine. And so I'm showing you the structure of psoralen, and the chemistry is as follows. Psoralen is a base intercalator, so it will intercalate into nucleic acids such as RNA or DNA, and with the right wavelength, it will cross-link the two strands together like this, okay? And then with a different wavelength, you can actually reverse the entire process. And the structure, the preferred substrate looks like this, staggered uridines, basically which frequently occurs, of course, in nucleic acids. And so the recent method that we reported takes advantage of this technology or this idea, and it's called Sorrel-Lynn Analysis of RNA Interaction and Structure, or PARIS. It's a way to get an RNA duplex map in living cells. So here's how this works. So you start again with living cells, and I'm going to focus your attention on these two, this red and green, RNA duplex, which you're going to track through the rest of the slide. We introduce psoralen into living cells and we cross-link with UV, okay? So then, once we trap this duplex, we can break the cell open, degrade the rest of the RNA, get rid of it with RNase. Then we can enrich for this duplex with an old-school biochemical method called 2D gel. So in the one dimension is denaturing, and the second dimension is native, okay? So the combination of this is that any uncross-linked RNA or single-stranded RNAs will basically fall apart, but the cross-linked RNA, doubly cross-linked RNAs, will run like a spider and in an off-diagonal pattern, like shown like this. So this species is something like less than 1% of the entire nucleic acid content, so this biochemical enrichment step is quite necessary to enrich for the molecules you care about. Okay, so once we have this cross-linked duplex, we can basically connect the two strands together, uncross-link them, and then go ahead and sequence. So in every sequence read, we have a single molecule evidence that these two strands of RNA were previously in a duplex together. Okay, so it's actually single molecule type evidence that we're going to be getting here. So let me show you what the data looks like. And so, indeed, we have a map now of the RNA structure and interaction in a living cell. This is looking at the well-known and abundant U4 and U6 RNAs, which have both intramolecular loops and intermolecular loops, which I'm showing with these arcs. Each of these arcs is supported by the primary sequence reads shown here. So each of these reads is supporting basically this structure with this structure, and this is another one shown here. If we zoom in on that, we can see that basically this arm corresponds to this arm, this arm corresponds to this arm, and this tells us that these two arms were in a Watson-Crick base pair in the living cell. And we have multiple of these reads giving us confidence that this is a real interaction. So using this technology now on exist, which is our topic of interest, we can see that this long 17-kb-long RNA has a very— actually, a very interesting domain structure. There are four modular domains that basically have basically folds within folds, okay, shown here, and that in addition, each of these, I'm showing the loops in two ways. On the bottom is the arc plot that I've shown you already. On the top, these duplex spans, each of these dots indicates that this and this arm are in a duplex. So the higher the dots are, that means the longer range the interaction is. So indeed, there are actually very long-range interactions, some as far away as 7 kb long, between the duplexes. Many of you know that the softwares that people use to predict RNA structure have a window size of 200 base pairs, okay, so they force proximal interactions. These are actually, I think, computationally predicted, they're convenient, not necessarily correct, and I think we can now do a much better job using experimentally guided structural data. And finally, if this map looks extremely complex, it's because it is, and there are actually alternative structures at one base involved in several different interactions. They cannot co-occur, so there's actually a family of structures. These are alternative structures. Okay, so let's zoom in on the business end of exist. So exist is comprised of a number of sequence repeats, okay? And this is now a situation where this kind of Paris map is really necessary, because, for example, the A repeat gets its name because in human there are 8.5 copies of the same sequence over and over again. So if you have a one-dimensional method, each of these repeats will be computationally equivalent for your prediction as to who's going to touch whom. But in the Paris map, we can actually have unique information. So on this track, I'm showing you here the repeats. Here's the sequence conservation. IC-Shape is a strategy that maps single-stranded regions, and Paris maps the double-stranded regions. And the duplex pattern actually looks like this, so we actually know then exactly how each each repeat folds with the next repeat. And I'm going to summarize this complex data in the following way. In brief, these repeats fold up into a kind of a module, which is repeated. The unit looks like this. It's a staggered repeat where basically the 5' end basically do base pairs with the next one. So this is the one copy starts from here, Goes over here, this is the end, this is the second copy. So the first half duplexes, the second half is single-stranded. This is a single repeat. This is the conservation, this part is highly conserved, this is the single-stranded region. And we learned that SPEN binds actually right here. This is the peak of SPEN interaction, okay? So this unique structure, it's a platform for the recognition of the RNA binding protein. That mediates silencing. This is the primary data indicating— confirming what I just said. This is a UV cross-linking experiment showing that SPEN binds exactly at these repeats, exactly at the very beginning, a little bit off to the 5' end in the single-stranded regions before the duplex regions. Okay, so I'm going to show this to you now in cartoon form. So, these 8+ copy repeats fold up in a specific way. It's really like RNA origami. So repeat 1 folds with repeat 2, okay, like this. It's a stagger repeat, and then 3 actually with 5, whereas 4 loops over with 8, and this mandates a fold within this kind of a higher order structure, and 6 with 7. We have 4 copies of this kind of a stagger repeat structure, and each span cross-links right here. There are 2 of these cross-linking sites for where the interaction is between the RNA and the protein, and there are 4 of these in this unit. Okay? So this unique fold mandated by the sequence to get into the structure is the binding platform for this long RNA silencing the X chromosome.

Okay.

In the final sort of segment, I want to kind of extend on that concept and talk about how we can use our knowledge of individual examples to decode new lncRNAs that might be present in evolution and in genomes that we haven't quite understood. And so, to do that, I'm going to turn— show you another— turn to another classic lncRNA system, and that's actually in the dosage compensation in flies, in Drosophila. And so insect cells in Drosophila do things differently. So again, females have two X chromosomes, males have an X and a Y. And the strategy for dosage compensation is that the single X in the male is transcriptionally upregulated by twofold. Okay? This is done, again, by two long non-coding RNAs called ROX RNAs. These ROX RNAs, called ROX1 and ROX2, are both on the X chromosome. They again coat the X chromosome, recruit chromatin-modifying complexes, but this time increase transcription. So even though the two ROCKs RNAs are genetically redundant— okay, so if you knock out both ROCKs RNAs, all male flies are dead, all females are viable, very specific phenotype— the two ROCKs RNAs look nothing alike each other. If you take the sequence of one and run BLAST, the other one doesn't even come up in the top 100. That just shows you the difficulty, the challenge of trying to understand lincRNAs using a primary sequence-centric view. Okay? But we know that these RNAs actually are mapping to the X chromosome. This is the view by RNA in situ hybridization. They code the chromosome. Using ChIRP, we previously mapped very high resolution the binding pattern of the ROCK RNAs, and you already explained the technology. So, here is the track showing you the data of a ChIRP. So, this is again X chromosomes. These are the peaks of the binding of the— one of the chromatin remodeling factors. This is the ROCK RNA. You can see that each of these peaks is exactly at the summit of the high affinity sites where are genetically defined binding sites for the ROCK RNAs. The ROCK RNAs only coat the X chromosome, not the autosome, so it's a very clear pattern. We carried out a lot of work on these ROCK RNAs to find that, in fact, the business end of the ROCK RNAs is again a structural motif. It's a tandem structural motif that looks like this. I'll summarize the data, then I'll show you the underlying information. So it's two stem loops separated by approximately 22 bases. These two stem loops has to be a specific sequence. The 3' arm of the stem is the binding site for the key binding protein, and this spacer has to be exactly 22 nucleotides, but could be any sequence, okay? Here's the data supporting that structure, including Shape mapping the single-stranded region, PAR mapping the double-stranded region. This is the structure conservation. You can see that this is the part that's conserved, and again, the 3' arm is where the binding happens for the key protein. So, the cartoon looks like this.

Okay?

So, what we realized is that even though this is a very long RNA, only these little bits actually really matter. Okay? And so, one proof of that concept is that if you look at the RNA, we can actually tell what part of the RNA matters. So we can make a mini gene that would have most of the function of the RNA. So, for example, ROCK1 RNA is approximately 4 kb long, but we learned that this domain, D3, approximately 400 bases, has the key stem loop features. And indeed, if we actually rescue, genetically rescue, ROCK1 or ROCK2 knockout flies— remember, 100% of the males would otherwise die— This mini gene is able to rescue as well as the full-length RNA. So essentially we have now trimmed the function down to 1/10 of the original size, indicating that we actually do know the functional parts of the RNA. A very talented student in the lab, Jeff Quinn, said, let's take this one step further. Can we take this knowledge and search for rocks RNA orthologs even in species where people couldn't find them before? So the strategy is to take really, again, genome sequence, always starting with the genome sequence, looking in syntonic regions, then looking for this repeated microhomology of these repetitive elements that can form into the right RNA structure. If we see that, we score that as a candidate rocks RNA, then we bootstrap from that to the next species and so on and so forth, okay? So what's amazing is that even in these species where we have no sequence identity beyond background. Okay, we can actually discover these lncRNA orthologs, which in a moment I'll prove to you are true orthologs. So I think this really speaks to the debate about, you know, what's junk DNA, what's the information content. Just because we cannot find it by sequence alignment, the simple strategy, doesn't mean that there's no information. Okay, so Jeff used this strategy and basically was able to go back in 40 million years of Drosophilid evolution. This is about kind of as different as kind of brown bears is to raccoons, so quite a bit of evolutionary diversity. And in fact, it's so distant that even the chromosome arm that's used as the X chromosome has actually changed. So some species is Mueller element A, that's one of the chromosome arms, and the other species is actually a different Mueller element, okay? So in this analysis, Jeff was able to basically triple the number of ROCK orthologs in the literature by this analysis. All the ones shown in blue were basically not known in the past and discovered in this analysis. So in total, he discovered something like 43 ROCK RNAs, both ROCK1 and ROCK2. What we learned is that these ROCK RNAs are rapidly evolving. Some species have 3 ROCK RNAs, some of them only have 1, and the structure was the key to the function. And I'm going to show that to you by focusing on 4 species that span the entire breadth of the evolutionary history. Drosophila monologaster, D. mel, willistonii, D. will, D. virilis shown here. Okay. And also buskii, this is the last species on the left here. Okay. These two here. Okay. So one of the proofs that these are actually true RNA, lncRNAs involved in dosage compensation is that they should bind the X chromosome. So Jeff went on to use ChIRP to map in each endogenous species the binding pattern. You can see that indeed each of these— basically, the ROCKS2 always bound the X chromosome, the Miller element A, but not— this species, both A and D's X chromosomes, so it's binding both, but not the other autosomes. ROCKS1 was a very good ROCKS RNA in melanogaster, but not so much in the other species. Species, and that's quantified on the right— on the left here. So Drosophila melanogaster, basically both ROCKS1 and ROCKS2 are functional, and these other species correspondingly the ROCKS1 is becoming weaker. Now the interesting thing is that this degeneration of ROCKS1 was actually readily predicted from our sequence analysis of the structure. So if you look at the complement of these double-stem loop structures, ROCKS1 and D-MAL has have 3 of these, these 2 guys have intermediate levels, have only 2, and this one that's the worst now has only 1 left, okay? So the complemented dosage of the repeated structural motifs predicted the potency of the activity of the ROCK's RNA. So you might say, well, that's a very nice correlation, prove to me that those elements are sufficient to actually do the job. So Jeff went on to actually construct synthetic link RNAs and ask, can they suffice in dosage compensation in vivo? So I mentioned to you that in Drosophila melanogaster, if you knock out both ROCKs RNAs, all the males are dead. If we come back with the transgene from melanogaster, we can fully rescue that effect. If we come back with the ROCKs from D. virilis, a different species grafted into melanogaster, we can actually also rescue quite a bit. Now we predicted this one is not quite as good because it only has one of these stem loops, so Jeff constructed a transgene now with an extra one, and now it rescues even better. So, indeed, the dosage, just a single stem loop, was able to confer some level of function. So, let's take this one step further. So, we thought, okay, what if we took just a single stem loop and added it to a completely irrelevant RNA, let's say E. coli lacZ? And even that has a measurable rescue at 2%. Okay? So this is— now you might think, well, this is not very impressive. It's only 2%. What I should tell you is that this is an experiment that could only be done in Drosophila because the escape rate from this lethality phenotype is less than 1 in 10,000. We counted basically tens of thousands of flies. We never had an escapee. So this 2% rescue is actually something like 100-fold better than the control, right? So we would assert that if this allele were happening in the wild, this would have a huge advantage that would quickly spread over the population. So you can even see how the rocks RNAs could evolve from these simple structural repeats being appended into a transcribed sequence. Finally, what I want to tell you about is how did these rocks RNAs target the X chromosome. If you have these new RNAs evolving, how did they find the right sites in their genome? When we did the ChIRP experiments, we of course mapped the binding site and the enriched motifs that these RNAs would target. And we can indeed see that all these different ROCK orthologs target the same sequence, this GAGA repeat, very simple appearing repeat. But the binding sites themselves were not always at the same location. Okay? So you can see that in this gene HDAC4, in melanogaster it's here, willistonii is over here, moving around over the place. This is a minority situation where the peak is always in the same location. This top one is the frequent pattern. This is quantified here. We're looking now at a 4-way comparison of these 4 species, sites that are in all 4 species, 3-way, 2-way, and singletons. You can see that the singletons are by far the majority, okay? But the pattern of evolution is such that even though you have the binding site at exactly the same location, the ROCK's binding tend to be in the near proximity of another binding site in another species. So if there's one in, let's say, D. mel, there will be another one nearby, not the exact same location. And so we think of this more as more like cell phone towers, right? They don't have to be exactly at the same position. They just have to be in some proximity to cover the entire X chromosome. So this still leaves the question of, okay, so these sites are moving around. How do the lincRNAs target the right sites in the genome? So we have a great opportunity here because of all these singletons where we know it's a binding site in one species, we can ask the question, what is the identity of that motif, that element in the other species before it became— before it evolved into a ROCK's link RNA binding site? And so here, I think Jeff made a very interesting discovery. What he found was that there's actually a preexisting element, the polyperimidine track that's used for splicing, that's often co-opted to become a ROCK's binding site in another species. This is a sequence that has the basically a CU-rich motif going from 5' to 3'. And what we discover is that, of course, the reverse complement could easily turn into GAGAGA with a slight shuffle. And this is kind of the frequency of how often these palindromic tracks in X chromosome can turn into a ROCK binding site in another species. The diagnostic test for this kind of a dual function moonlighting idea is that there has to be a strand bias that basically when we see an intragenic ROCKS binding site, they obviously have this strand bias so that one strand can still serve the RNA function of C-U-C-U-C-U, whereas the other strand on DNA can be G-A-G-A-G-A. Okay? So what's very elegant about the strategy is that this cooption of a splicing signal for dosage compensation automatically then targets your compensation machinery to actively transcribe genes. Right, so we have a limited amount of resource. We don't want to target it to heterochromatin, to some random location. We want to target to transcribe genes that's going to matter for dosage compensation. And so, this system is very flexible and allows the evolution of these new binding sites, but with also very efficient targeting and parsimonious strategy. What I've told you then is that there's a rapid evolution that underlies even apparently conserved and essential lncRNA-genome interactions. Something as fundamental as sex, which is 100% lethal without it, is still rapid evolution. So for lncRNAs, it's a refinement by repetition and acquisition of these structural repeats. And on the DNA side, it's the cooption in proximity of these preexisting signals, of preexisting sequences that can be changed by just a slight reshuffling into a lincRNA binding site. And these together then allow lincRNAs to have this surface— this kind of plasticity to basically manage dosage compensation even in changing genome backgrounds such as different chromosome arms. I'm going to end by showing you this sketch by M.C. Escher, which I like very much. It shows that basically coming out of this blank slate, two hands drawing each other. And what I've told you today is that I believe that the relationship between RNA and chromatin is very— has really a lot of analogy, and that is that these two polymers template each other to reinforce fidelity in the transmission of information. And furthermore, as I show you that because the link RNA and genome can co-evolve in a rapid way, this serves as a mechanism for evolutionary plasticity and innovation creating new meaning out of the blank slate of genome information. I want to thank you for your attention. Many people in the lab were involved in this work. I want to particularly— so their names are shown here. I want to acknowledge my collaborators, Will Greenleaf, Edith Hurd, Job Dekker, and Asifa Akhtar. And I gratefully acknowledge my funding sources, especially NIH, especially NHGRI, NCI, and NIEHS. Thank you for your attention.

Thank you for a wonderful romp through some really interesting biology. If you have questions, there are microphones in the aisles. Please use those so that people watching by the video can also hear the question. While people are thinking about it, I'm puzzled by— your example with rocks where you're able to recreate, at least I think I understood that, the biology of a link RNA with just a single short repeat, and you would think then evolution wouldn't bother to make things that are long and complicated if it could accomplish the same goal with something short and simple, but I imagine long and complicated has some benefits.

Maybe you could explain. Right. So that was kind of really at the— all sort of the— so what we see in the naturally occurring link RNA is that these motifs, these structural motifs, are actually repeated multiple times. That's the case in exist, that's the case in rocks, and I think that the repeats make them work much better. So it's a real stat, right? So the single copy was a proof of concept saying, okay, yes, it could be done, but it was very weak, right? But the more copies we inserted, either in the naturally existing RNAs or the ones we create synthetically, they became more potent. So you can dial it up and down.

Yeah, over here. Hey, Howard, very nice talk. I have two questions. So, first one is for the ROCKs story. So, how the ROCKs will avoid autosomes because the polyphenylamine track? And the second one probably is more philosophical. So, I think you showed the structure is very important, but there's still some kind of a sequence specificity, at least some weak sequence specificity for these long non-coding RNA to bind the target regions. And can you elaborate maybe the loop regions or other regions may be important to confer this kind of sequence specificities?

[Speaker:JOSHUA_ELKINGTON] Sure. Those are really good questions. So the first question is, why does the ROCKs go to the X chromosome? Okay, so the ROCKs RNAs themselves live on the X chromosome. So they're transcribed from the X. So the simplest model is that they just never leave, right, because they're sinks of binding sites. So that's one possibility. It's also known that that motif that I showed you, that GAGA motif, is more enriched on the X chromosome. Than autosomes. So I think those two factors have some contribution. The third factor is that there's actually a relationship between the three-dimensional organization of the X chromosome and kind of where the ROCKs tends to go. And so I think those factors obviously being in cis then basically preferentially loads the ROCKs onto the X chromosome. I think your second question is whether lincRNAs have sequence specificity. I think that definitely Even if it's a structure-based model, we need specific sequence to fall into a specific structure, so sequence is definitely important, right, because the structure comes from the sequence. So it's just a matter of that, not the entire length of the lncRNA, the sequence matter, only the specific regions. I hope that answers your question.

Okay, this is a little bit simpler question. At the clinical level, you mentioned that autoimmunity would be more common or diseases of that autoimmunity would be more common in the female. And other people have put forth that that would be perhaps due to the male fetus in a female, and the intrauterine fetal maternal transfusion, and there she acquires a Y chromosome, and the autoimmunity, that might also be a mechanism. Do you think that has something to do with what you're talking about, or maybe a factor?

Yeah, so the question is, could it be a consequence of pregnancy, right? Especially bearing a male fetus. So that is definitely a model that's been proffered, but I think that the sort of the risk ratio for the autoimmunity exists even if you account for pregnancy. And in fact, there's a particular syndrome called Klinefelter syndrome, which the patients have a gene, have XXY. So they're phenotypically male, they're never gonna get pregnant, but they still have the elevated risk of autoimmunity because they have two X chromosomes.. And so I think that's a good answer to that. That must be an independent contribution separate from pregnancy.

Thank you.

Yeah, I just want to follow up on the question of why long non-coding RNAs are so long. So for XIST, it's really quite long, but the A repeat is only a small portion of that. So if maybe you could expand upon why you think it's so long.

Right. So it's a— think of it as like a— it's a platform, or like, think of it like a computer circuit board where all these sort of the circuits plug into it, right? So, it exists, it's long, but it's also loaded up with 81 proteins. So, the A-repeat structure only dealt with really sort of 3 proteins that bind in ATP-dependent manner. So, I think many of these other repeats could have additional, as I mentioned, this modular assembly of all these proteins. And so, it's a way to basically organize a large complex and deliver together to some location in the genome.

Yeah. Well, fascinating presentation and terrific discussion. We're going to have a reception in the library. Please join the speaker for coffee and cookies and continue the conversation. But let's thank Dr. Chang again.