Scientist Stories: Howard Chang, Personal Regulome Navigation of Cancer

Nice to see you, Dr. Chang. Professor Chang is Director of the Center for Personal Dynamic Regulomes and the Virginia and DK Ludwig Professor of Cancer Research at Stanford, also an HHMI investigator, Professor of Dermatology and Genetics at the School of Medicine.

Um, so how do you go from dermatology to, um, what you're doing now? Yeah, dermatology is really an entire field that has all kinds of diseases, including cancer, autoimmune disease, many other kind of problems. So it's actually a great field to kind of keep an open mind and think about all the different potential applications. Cool.

Well, I'll turn it over to you now, and I'm really interested in what you've got to tell us.

Great. Well, thank you, Casper, for the kind introduction. Let me share my screen to make sure that works.

We can see it. Yeah.

Okay, perfect. Yeah. So what I thought I'd do today is kind of share a little bit of our journey of how we started working on trying to understand the gene regulation landscape, what we call the regulome, and how that led us ultimately to study cancer and specific kind of problem cancer called extrachromosomal DNA. My lab has been fascinated by the non-coding genome for a long time. When we think about the genome, we often think about genes, maybe disease-associated genes. But this cartoon helps me— tells me to remind you that each gene is connected to a series of regulatory elements, really switches that decide when and where genes turn on and off. These regulatory elements are really the binding sites of transacting factors like transcription factor proteins, regulatory RNAs that together make these decisions. The picture in the human genome actually looks like this. Just 2% is protein coding. The other 98% is really dedicated to this non-coding space. We also know that actually the vast majority inherited human genetic variants associated with disease map to the non-coding genome. So one approach that we really like and pioneered really takes the idea of the concept of DNA accessibility. In every human cell, 2 meters of DNA is packed into a 10-micron nucleus. Therefore, most of your DNA is highly compacted, not accessible, except for the active elements that are being read by the cell's machinery. Therefore, simply finding the accessible bits of DNA tells you really about the software program the cell's currently running. So several years ago, my colleague Will Greenleaf and I at Stanford developed this method called ATAC-seq, or assay of transposase-accessible chromatin. And this makes use of an enzyme called transposase that copies and pastes DNA. We load up this enzyme with DNA adapters that go to your sequencing machine. So when we expose this prokaryotic enzyme to eukaryotic chromatin, it can only copy and paste into the open chromatin bits, but not into the closed elements. So in a single step, you selectively and covalently tag the regions of interest that allows you to basically then amplify and sequence these areas. So this methodology led to ultimately a million-fold improvement in the sensitivity and 100-fold improving the speed of mapping regulatory elements in a genome-wide scale. So let me show you what this kind of data would look like. So here we're looking at genomic tracks on the x-axis. These are the locations of genes. On the y-axis, each peak, the height indicates the level of accessibility. So the first row is the prior gold standard, DNA hypersensitivity. It needed 10 million cells. The first version of ATAC-seq used 50 million cells. 50,000 cells, and ultimately single-cell ATAC-seq. Basically, you can see that these patterns look virtually identical. But now when you zoom in, this is the summation of several hundred single cells. Now when you zoom in, every row is a single cell. There are 254 single cells here. Then at every position, you have either 0, 1, or 2 reads because this is a diploid genome. Okay? So in this way, this kind of analog signal has turned to digital data. Armed with this very powerful way of mapping the regulatory landscape, we then turn to think about other kinds of maps that might be useful, and we're inspired by the electronic maps that we all perhaps use in our daily lives. So these electronic maps represent the real world in multiple layers, for example, basically information about land usage, elevation, different businesses, the streets, and maybe your friends or customers. And then each of these layers makes the map more and more useful. And by analogy, we think that if we have a snapshot of gene regulation, a personal regulome, we can then infer, extract from this kind of map information about cell types, cell states, tissue microenvironment, lineage, drug response, connect them together with advanced computation. And, indeed, so many studies over the last several years really showed that this is now possible. I like to really directly learn from patients. And so I want to show you kind of some of the fruits of that kind of effort, really principally in the space of cancer. And Ryan Korsis led this work when he was a postdoc in my lab, now a faculty member at the Gladstone and UCSF. And so we went and used this methodology, ATAC-seq, to look at the chromatin landscape in primary human cancers. This is using a collection called the Tumor Genome Atlas, TCGA, Cancer Genome Atlas, which at the time had many kinds of information. Really, this is a flagship project that's been building for many years. By the time they had whole genome sequencing, RNA information, other mutation data, but didn't have any chromatin information. We were able to use this very sensitive method because by the time we got to it, they really had depleted almost all the samples. And so we were able to map the landscape of chromatin axes, really the landscape of regulatory DNA in 23 human cancer types. They're illustrated here on the left. You can see that they span from glioblastoma to lung cancer, kidney cancer, bladder cancer, and so forth, really some of the most common, also aggressive cancers. We mapped 410 tumors, identified more than half a million active DNA regulatory elements across these cancer types. Now, the important thing to know is that actually nearly half of these elements were not seen in other maps of normal tissues, for example, in ENCODE or other kinds of epigenomic roadmap. They're only exercised during the pathology of cancer. And these kinds of maps taught us a lot about the landscape and also the— the biochemical causes are both inherited and somatic variations that contribute to cancer. For example, for many years, geneticists have been mapping families, taking care of families that may have elevated risk of certain cancers. Most of these risk elements map again to the non-coding genome. There's no explanation for how they may work. On the right-hand side, I'm showing you some, again, primary data where on the x-axis, these are genomic location, the y-axis, these are accessibility, and we're looking at the MYC oncogene locus right at the bottom here, right in the center. All these cancers basically activate MYC. But on the top 5 samples colored in orange, these are colon cancers. You can see that these cancers principally use regulatory elements to the 5' end to the left of the MYC locus, whereas in kidney cancer shown on the bottom 5 in purple, these cancers use elements to the right-hand side of MYC, on the 3' end. So same oncogene, different oncogene switches. The second point is that, in fact, these risk SNPs, variants for cancer, they map to cancer-specific enhancers. So for example, this colon cancer risk actually maps to a colon cancer-specific enhancer here, and the kidney cancer-specific enhancer maps to, again, a kidney cancer enhancer in the middle of this PVT1 locus lincRNA gene. By this way, we can actually provide, really, a biochemical explanation for a lot of genetic association with cancer. Now, I want to spend some time talking about this locus, PVT1, because long non-coding RNA genes, there's really one of the major focus in my lab. And this particular gene, PVT1, was actually perhaps the first long non-coding RNA gene nominated by Cancer Genetics back in the '80s. This location, PVT1, is actually a breakpoint translocation in Burkitt's lymphoma. It led to the cloning of MYC and recognition of MYC as an oncogene. So that's how far back it goes into history. And interestingly enough, there are a lot of frequent mutations and deletions within this locus. And of course, MYC is a very powerful oncogene. It's overexpressed in over half of human cancers. And it can really drive— it's a transcription factor that can really drive all the hallmarks of cancer. And so we, through a series of experiments, we were able to understand that really this PVT1 gene, even though it's basically transcribing a long non-coding RNA, its function is better understood as a series of DNA regulatory elements. What's going on here is that, in fact, there are 4 intragenic enhancers that activate the PVT1 gene promoter. But if the PVT1 gene promoter were not firing, those enhancers are still active and they would activate MYC instead. PVT1 promoter is really like a boundary element that prevents these enhancers from spilling over. We came to that conclusion because we did a genome-wide CRISPRi gene silencing screen identified that PVT1 promoter was the number 1 lncRNA locus in the entire genome, that when silenced help cancer cells grow. We further show that, in fact, that there are frequent deletions, mutations that inactivate the promoter, and that specifically this PVT1 captured the 3-dimensional looping and prevented enhancers from spilling over. And so this will basically pioneer the concept of a lncRNA promoter as a DNA tumor suppressor DNA boundary element. You're thinking about lincRNA genes really as a series of DNA control switches. Now, this particular study of this locus also led us to the next topic, which is potentially chromatin clues to a very unusual DNA structure. Now, I've shown you a number of cases of these tracks, basically genomic location on the x-axis and nice peaks like this. But in this particular case, you see this very unusual pattern where we have this almost 100 kilobase-wide contiguous high accessibility, again, spanning our friend MYC and the PVT1 gene, the 5' end. What is really going on here? In fact, these are extra-chromosomal DNA amplifications. It's not just because there are lots of copies. At the bottom, I'm showing you a metaphase spread where these are cells arrested in mitosis. These big structures in the middle, these are chromosomes. Stained with DAPI. These little dots are also DNA. These are the extrachromosomal DNA. In the middle panel is a DNA in situ hybridization or FISH experiment against the MYC gene. You can see that the ecDNA, extrachromosomal DNA, contains the oncogene MYC, not on the chromosomes. The third picture here is a version of ATACs that basically you can read out with fluorescence. Which we call ATAC-C. You can see that this little dot is as bright as this entire chromosome, which is quantified on the right. This shows that on a per-molecule basis, the ecDNA has that more basically open accessible chromatin structure leading to altered gene output. In fact, we've now learned that extrachromosomal DNA represents a major challenge for cancer patients. This is a very prevalent kind of a problem present in half of human cancer types, up to a third of human cancer cases over time. They are unfortunately associated with much worse survival for patients and outcome. And these are actually double-stranded, circular, large pieces of DNA from 100 kilobase to a megabase that always can— nearly always contain oncogenes and associated DNA regulatory elements. And I'll describe to you then the transcription of the energy in a moment. These extra-chromosomal DNAs, they don't have any centromeres. They don't have any telomeres. So they're randomly— basically, they're randomly segregated into daughter cells. So they can lead to very rapid change in copy number. It's really a moving target, causing accelerated evolution. And importantly, these kinds of clonal ecDNAs are unique to cancer. They're not found in normal cells. Our studies together with my colleague Paul Michel led to a realization that there's something very different that happens when ecDNAs basically pop out and tether themselves from the rest of the genome. That is that every molecule puts out basically about 4 times more messenger RNA than the same sequence on the chromosome. Previously, ecDNA was thought of as a copy number problem, and our work really revised this concept to say that this is actually an epigenetic problem. Once the sequence comes out into this circular form, it is much less compact. It transcribes much more. And therefore, we have a situation where we have the same sequence, DNA sequence, two different chromatin states, really a new kind of epigenetic phenomenon. And so a lot of our ongoing work now is to understand how this has happened. I've shown you a number of pictures of these basically metaphase spreads. But one surprise turned up when we looked at cells in interphase. This is an image, again, of DNA in situ hybridization. In the green color is in situ hybridization for MYC. In red is against a chromosomal locus adjacent to MYC. If you have a chromosomal copy, you'll see basically paired green and red dots. There you can see there are multiple copies. If it's ecDNA by itself, it should just be green. Now these cells have actually about 10 copies of ecDNAs per nucleus. You can see that all 10 copies congregate together in one single spot, and this is called extrachromosomal DNA hub, ecDNA hub. This is a micron-sized collection of multiple ecDNA molecules. You can see 3 cells in the field, they're all doing exactly the same thing. Notice that there's plenty of space, the chromosomes are nicely spread out, but the ecDNAs all want to get together. Okay. We're really intrigued by these kinds of images. So we went on and studied what might be the— where— what made the oncogene RNAs, where are they coming from, right? Are they coming out of these hubs, or are they coming from singleton molecules? So first of all, I just want to show you that when we look across many cancer types, we basically always see the ecDNA hubs. For example, in prostate cancer, colon cancer, GBM, glioblastoma, gastric cancer. And we can also see this even directly in primary cancer samples. For example, in neuroblastoma, you can see this here. Most cells in this field have these giant ecDNA hubs. Okay? And secondly, we can actually combine RNA-FISH with DNA-FISH. So we can look at the nascent RNA transcription and see that most of the nascent RNA, for example, in the MYC locus, is coming from the ecDNA, not from singletons outside of the hub. Okay? So we were able then to go on and compute the per-molecule probability of transcription for every DNA molecule. So you can see that the ecDNA has a very high probability of transcription, whereas a chromosomal locus, much less. On the right-hand side, we can further quantify how likely each ecDNA molecule is is to fire, is to transcribe, based on its relationship with other ecDNA molecules. We discover that the more they're clustered together, that is, the more neighbors are nearby, the more likely each ecDNA is likely to fire. In other words, this is the cooperative transcription. They care about other molecules. In contrast to your chromosomes, we have two separate alleles, they each do their own thing. In this case, the ecDNA transcription is somehow cooperative. So a major question then is, okay, so these ecDNA hubs, there are all these molecules together. What is holding them together? And we discovered that there's, in fact, a very important protein core that holds them together. And this— and our realization came from the fact that I just told you that ecDNAs transcribe like crazy, and that they must be somehow using a lot of the enhancer-promoter contact and transcriptional mechanism. One of these proteins that's involved in this job is called BRD4. This is a protein that contains a bromodomain, which is a reader module for basically histone H3 lysine 27 acetylation, which is an enhancer-associated histone mark. The additional domains, the low-complexity domains, they also form phase separation. And so this image on the right just shows that both the BRMOT domain and separately the IDR is capable of basically forming these condensates, both in vitro and also in vivo with very high concentrations, and they can basically group molecules together without membranes. Okay. So in the— now in this image, I'm showing you that, in fact, we believe that BRD4 is tethering these EZ DNA hubs together. On the left is a live cell imaging strategy so we can actually visualize the ecDNA shown in green. In the middle, we have now knocked in an epitope, a Halo tag, into the endogenous BRD4 locus. So that's shown in purple. You can see that there's actually a BRD4 throughout the nucleus, but hopefully you can see that every place where you have an enrichment of BRD4, you also see an enrichment of ecDNA. Okay? So finally is the key experiment. So on the left here, now you're showing— I'm showing you again a 3D movie of these ecDNA hubs. So you see again these giant dynamic sort of condensates of ecDNA. On the right, if we treat with a molecule called JQ1 that blocks BRD4 function, within 20 minutes, all these ecDNAs fly apart. Okay, you can hopefully see lots and lots of individual molecules now. Separated. Okay. And this actually has a profound effect on ecDNA transcription. And so on the left, I showed you first ecDNA has a very high rate of transcription of the oncogene per molecule. If we break up the hubs, now that transcription advantage goes away. There's also a selective impact on cells with extra chromosomal DNA shown in blue compared to isogenic cells, that still have elevated Myc copy number, but now on the chromosomal locus, so-called HSR cells shown in green. Much less of an effect here. Indeed, JQ1 allows selective killing of the cancer cells with extra chromosomal DNA shown in blue. Now, we're really curious about this phenomenon because now we have these proteins that hold these DNA together, but we wonder how do different DNA molecules know to join that hub? What is their membership card to basically say, "Let me into the club," okay? And so our old friend PVT1 sort of comes to the fore again, because our prior work on promoter-enhancer competition showed us that PVT1 was able to basically take transcription away from the Myc promoter because it was a very powerful binding site for Brd4 and K27 acetylation. A number of experiments, including this one, really showed that PVT1 promoter was a DNA element that allows you to join the hub. The key experiment said if you just put PVT1 promoter onto a heterologous episode, a plasmid, like a reporter gene, that is enough for this new plasmid to join the ecDNA hub and now transcribe like crazy. This effect is, of course, now blocked by JQ1, and which, of course, then disperses hubs, which is shown on the right, that indeed, basically, if we add a PVT1 promoter to this reporter gene, selective activation in cells with ecDNA hubs, they're blocked by JQ1, and not so much— much less so in cells that have, again, elevated copy numbers of the same locus, isogenic cells, but now on a chromosomal locus. Okay. Now, some of you listening to me may have realized that I just described for you something very unusual, and that is that normally gene activation between enhancers and regulatory and promoters, as I showed in my first intro slide, happens on the same chromosome, right? They have to be on the same DNA molecule for these long-range contacts. But this experiment I just described for you implies an intermolecular gene activation. This reporter gene plasmid has no enhancers, is getting activated by enhancers in the ecDNA hub. Is this really something that can happen? So, in fact, indeed, we turn to an endogenous situation where we have the gastric cancer cell line, which has two different flavors of ecDNAs, one coming from MYC, the second coming from the FGFR2 gene originally from two different chromosomes. And indeed, these imaging experiments show us that these ecDNA hubs, they intermix. On the right-hand side, molecular experiments with chromosome conformation capture shows, again, extensive intermolecular contacts between FGFR2 and MYC extrachromosomal DNA. Okay. So King Hung, a very talented student, then went on to form a genetic screen where we use CRISPR interference, which is a way of using dead Cas9 as a programmable DNA-binding platform connected to a KRAB domain to silence elements that we tile across basically ecDNAs and ask, do these elements affect expression of the oncogene on the same ecDNA or the oncogene on a different ecDNA? And indeed, we found a number of elements, for example, in the FGFR2 ecDNA that's required for FGFR2 activation. For example, the F— its own transcriptional start site, and also similarly for MYC. But really, to our delight, we found there are a number of elements that act in trans, that is, in intermolecular fashion. A number of elements on FGFR2, actually, I believe 5 of them, that are needed to activate MYC, and conversely, an element that's needed from MYC to activate FGFR2. Okay. So what I've told you then is that we discover a kind of emerging behavior of the extrachromosomal DNA. Like these schools of fish swimming together, these ecDNAs will actually congregate, so the self-organizing principle, into this basically transcriptional hub. And the important concept then here is that even though we might be tempted to call these things chromosomes, this is really a different kind of concept. Okay. A chromosome, as we know, is now a very long piece of DNA. Proteins on top like nucleosomes and histones. And that is kind of the modern understanding of a chromosome. In contrast, the ecDNA hub has an inverse kind of topology. There's a protein core, for example, BRD4, and then DNA molecules decorating on top of it. Okay. And the basis for this interaction is actually this bromodomain interaction with a histone modification. And that is the basis of these hubs holding together. We believe that this concept of ecDNA hub, I think, will be important for this field because we think that the ecDNA hub is really the unit of the oncogene function, not individual DNA molecules. Because inside these ecDNA hubs, we actually have intramolecular gene activation. So enhancer-promoter contact can occur across different molecules. This allows cooperative transcription, so basically more spatial proximity more gene activation, and really also a combinatorial and very promiscuous sampling of the regulatory input, something that you can never achieve when you're tethered to chromosomes. Conversely, we think that there are some really interesting consequences, and that is that this cooperative lifestyle might drive ecDNA evolution, oncogene diversification, But we also think that this is a therapeutic target because, as I showed you, if you break up these ecDNA hubs, this transcriptional advantage stops, and these oncogenes are basically rendered much less powerful. Normal cells have chromosomes, right? So these ecDNA-specific mechanisms are only unique to cancer cells. It's really a truly unique and differentiated target. Okay. So I want to spend kind of the— the remainder of time talking about some of our new efforts on trying to read and write extrachromosomal DNA. And one of the important challenges is that once we realize these DNAs are so important in cancer, we really like to basically determine the primary sequence of extrachromosomal DNAs. Okay? And that actually may not be as easy as you might think because every time we sequence a sample, we're sequencing all the DNA in the cell, extrachromosomal DNA and the chromosomes together. This basically makes a good mesh and then it's hard to tease it apart. Recently, King Hung, the very talented student, we developed a method for targeted profiling extrachromosomal DNAs, a method called CRISPR-Cas. CRISPR-Cas was a method previously developed to purify bacterial DNA segments. We have now adapted it to purify extrachromosomal DNA from human cancer samples and cells. This method starts in the beginning by embedding your cell or DNA sample in agarose, which prevents shearing. And then we— you go on to prepare the sample for a method called pulse gel electrophoresis, which is a classic method for separating very large pieces of DNA. Okay? So I've mentioned that pulse field gel was invented by Gil Chu and Ron Davis at Stanford. And the key insight is that large circular pieces of DNA will not go through the gel. They're trapped in the well. But if you cut that DNA just once, circle turns into a line, and now it will run according to its size into the gel. The chromosome may actually get cut, but the chromosome segments are so big, hundreds of megabases, that they're still stuck in the well, so they don't go anywhere. Conversely, if you want to purify the chromosomal DNA, you have to make two cuts outside of the amplicon, and then that will also release a segment that I can now run into the gel and basically go on to purify and sequence. Okay, so that's the concept. Let me show you what this actually looks like. So on the left, this is again a metaphase image, and that kind of shows you the challenge. This is a, a glioblastoma sample near sphere, The red dots are basically ecDNAs that contain the EGFR, epidermal growth factor receptor oncogene. You can see there are lots of copies, about 100 copies. The blue blobs are chromosomes. We want to just somehow pull out just the red dots, not the blue stuff. Then conversely, even the same sequence is actually present on the chromosomes. There are about 7 copies, maybe 5 copies here. And so again, we want to be able to pull them out selectively one over the other. Okay. So at the bottom, that is how we design our CRISPR guides is using the Cas9 for very simple cutting function. And based on where we cut, we think we can isolate either ecDNA or the chromosomal DNA. On the right-hand side, I'm showing you now a pulse field electrophoresis gel. And you can actually see that this works beautifully. You can actually release the ecDNA just by itself here. And we can actually— now, this is about 1.3 megabase. And so the first time we can actually directly verify the ecDNA size on the absolute scale. And of course, secondly, after we sequence, we now get a massive enrichment in ecDNA content, a 30-fold enrichment. Rather than sequencing everything together, we just sequence the stuff we want. So this is like exome capture but for ecDNAs. Okay. So now we can see the data. Notice that basically the coverage now is basically like up to 500-fold, very enriched. We have ecDNA, and these different cuts basically correspond exactly to the size that we wanted. And compare the coverage, for example, for the whole genome sequencing, 15x, ecDNA, 500x. Okay? So just that this has a number of implications for what we can learn. For example, in this particular case of the glioblastoma, we were able to identify all these single nucleotide variants and also subclonal SNPs and basically assign them to the two parental chromosomes. This is called phasing. And then this has a very important consequence because we learn that, in fact, in one allele is the origin of ecDNA. We see the ecDNA is all from one haplotype. And then there's a deletion at the original position, including a junctional scar read that spans the original locus. Okay? The other allele is intact. Second point is that you might notice the little white box here, that there's a deletion in the EGFR on the ecDNA. And this is actually an activating mutation, makes a variant called EGFR V3 that makes the receptor active always, independent of the presence of ligand. This is a clear oncogenic driver. 100% of the ecDNA molecules have this EGFR V3 point mutation deletion. Interestingly, the chromosomal version of EGFR is entirely wild-type. You can actually in fact now learn for the first time that this oncogene activation is purely an ecDNA problem. The chromosomal genome is actually pristine. Second interesting point is that there are actually models about how ecDNA is generated. One model posits that ecDNA happens because basically there's an extra round of copying during S phase. And that model predicts that you would basically have the intact locus in the chromosome plus an extra copy of ecDNA. A second model, alternative model, posits that this is a deletion mechanism. You basically make two cuts, you cut out the ecDNA like this, and there's a deletion of the original locus and extra chromosomal DNA. So our data clearly supports the second deletion model. And then basically, I think, makes an important point about— and then we believe that subsequently, one copy of ecDNA gained the EGFR V3 mutation, which then swept through the entire population through positive selection because it's an oncogene driver. Okay? Okay. If— we're also really happy to find that we can actually learn more just beyond the mutational landscape, but also learn about the epigenetic sort of mutational mechanism. And that is that we can take the purified ecDNA molecule— remember, no amplification— and pass it directly to single-molecule nanopore sequencing. And this allows us to basically measure both mutations and also DNA methylation status. And we discovered that, in fact, there are unique DNA demethylation events only on the ecDNA compared to the chromosome. So for example, right over the EGFR promoter, you can see there's a selective demethylation. And every row here is a different single molecule. You can actually read out the patterns in this kind of ticker tape. Okay? And so this, we believe, is a second reason for the selective gene activation on the extrachromosomal DNA because they're losing silencing DNA methylation works. And we are— you may know that I'm a physician scientist. I'm a dermatologist by training. So I'm always sort of really motivated to try to take the technology we develop in the lab and apply it to clinical material. And this is— and I'm really delighted to share with you that this has come to pass. Some of you may know that targeted therapy for cancer has had remarkable advances in the last decade, especially in the field of melanoma. But unfortunately, a lot of our targeted therapies such as BRAF inhibitors and MEK inhibitors, MAP kinase inhibitors, they are met with initial response, but unfortunately, and then therapeutic resistance, and the mechanisms are not always understood. And here is such a case where a patient was initially diagnosed with a cutaneous melanoma, which then has a BRAF mutation, was treated with a BRAF inhibitor and also subsequently a MEK inhibitor immunotherapy, but unfortunately, On therapy developed a resistant lesion, a metastatic lesion. Genomic surveillance showed that only at this resistant lesion did an ecDNA basically pop out. We're able to apply our method CRISPR-Cas and characterize this ecDNA. This turns out to be an 890-kilobase ecDNA spanning the NRAS oncogene. Let me first show you the genomic locus. Basically, this interval here, the ecDNA, actually spans basically two topologically associated domains. These are regions that like to fold up, the genome likes to fold up by themselves. The right-hand side of this topologically associated domain, which is this triangle, contains a number of enhancers that are now brought to the proximity of the NRAS gene on the left-hand side of this interval through circularization. We're pretty sure that NRAS is the oncogene because the NRAS gene has a point mutation. There's a G12R point mutation that locks the NRAS in the active GTP-bound form. Okay? 100% of the eCDNA contains this NRAS point mutation, again, showing us that this is very likely the driver event. And so this is a situation where you have a basic parallel pathway driving the MEK signaling, MAP kinase signaling, bypassing drug inhibition. So many laboratory experiments suggested that ecDNAs might contribute to drug resistance. I think this is one of the clearer cases in human patients that this is, in fact, the case. Okay. Now, I've shown you some examples where we have ecDNA discovery curves by CRISPR-Cas to cut a single band. And we can also— we have also worked together with computational colleagues to develop a method to automatically take these bands and intervals, reconstruct them based on the sort of the edges connecting a graph, and then we can actually directly match the molecular size to reconstruct end-to-end ecDNA primary sequence. This is actually useful because sometimes the situation is way more complicated than you may imagine. On the right-hand side, I'm showing you a PULX field gel of, again, of this gastric cancer sample that we thought had two flavors of ecDNA, MYC and FGFR2. But in fact, once you cut with a single reagent, you realize that you release, let's say, three different bands, six different bands. There are obviously multiple different species here. This is impossible to sort out without a method like this. In the past, if you do whole genome sequencing for this sample, you'll get something like this picture. Basically, now we're looking at genomic tracks. These are basically the copy number and these edges are trying to connect together the possible connections. And this, if you, if this looks like a mess to you, that's exactly the point, right? Because the computer is trying to stitch together all these things, and basically it doesn't know how many molecules are actually here. It's trying to stitch them all together. So this picture probably is just like a fantasy, right? Some of these molecules don't even exist. But if this picture becomes much simplified when you can cut out individual bands and then reconstruct the sequences together, as shown on the bottom. So each band has its own pattern, right? And we now know absolutely the right size and how they connect together. So let me show you what this means. So that very complicated picture turns into something like this on the left. Now I'm showing you the genomic contents of ecDNAs. These are the pieces of the chromosomes that are contained, including a piece from chromosome 8, chromosome 10, chromosome 11. These are the epigenomic profiles. Each row now is a different ecDNA species. We know exactly the copy number and kind of the interval that's included. But moreover, we actually have an end-to-end 360-degree reconstruction of extrachromosomal DNA, right? And so the outer ring is actually the primary sequence. The inner ring is a validation using another method called optical mapping that shows that these reconstructions are correct. And so this kind of analysis is showing us that there's actually remarkable extra-chromosomal DNA diversity. And so which I can summarize here is what we learned is that in these East DNA amplicons, sometimes we see a full locus amplification. So basically, just basically the oncogene and the regulatory elements are— there are more copies. And other times, we see enhancer rearrangement. Recombination events that allows the oncogene to hijack DNA regulatory elements so they can basically crank up their expression. But actually, I think further to our surprise, we see evidence of ecDNA specialization. We see oncogenes coding sequence that are amplified but without its regulatory elements. And conversely, we see extrachromosomal DNA that contain just enhancers, just regulatory elements, but no oncogene coding sequence. Okay? So in of themselves, these would be expected to be— you would call these defective oncogenes, right, because they can't transcribe. But they only make sense because of the evolutionary force driven by the ecDNA hub. They can come together, complement each other, and basically still give the cancer cell a selective advantage. And so I think that the occurrence of these specialized ECDNAs, especially enhancer-only ECDNAs, really speaks to a new concept in cancer genetics, right, really a purely regulatory oncogene. And I think speaks to the presence of these intermolecular gene activation mechanisms as a driving force for cancer evolution. Okay. Finally, I want to talk to you about the rules of ecdNA evolution. And so working together with my colleague Paul Michel and and Ben Werner and other colleagues, we've been really thinking about modeling ecDNA through computation and mathematical analysis. These are backed up by experiments that measure key properties of ecDNA expression, segregation, that allow us to make predictive rules for how ecDNAs may propagate themselves through different cell lineages. And then this then further allows us to get therapeutic insights about what would happen if we do different perturbations and what would have the biggest impact were we to make certain modifications. I'd like to show you some of the initial fruits of this labor. This comes back to the concept of ecDNA random segregation. I showed you this image already about the chromosome. This makes the point that chromosomes actually undergo remarkable dynamic rearrangement. So in interphase, chromosomes are very floppy, but in mitosis, the chromosome condense 10,000-fold to make these rod-like structures that you'll remember from textbooks. ecDNAs actually do something the opposite. Okay, this is another reason to think about ecDNA hubs as a different unit. So in interphase, they exist in ecDNA hubs, but in mitosis, they actually break up and actually tether onto mitotic chromosomes and go into daughter cells. And from these movies we made, we can see that, like, okay, the top— this cell on the right actually have more ecDNA pixels than on the left. And so there's asymmetric segregation. The challenge we face is that how do you scale this up so you can look at lots and lots of cells all at the same time? So the key insight that Paul's group had is that you can actually stain mitotic cells with this protein of the midbody called Aurora B. And midbody is basically a cytological structure that bridges two daughter cells as they're dividing. So now when you look at a field of cells, you can see that these two cells are the two daughter cells coming out of a single division, these two cells coming out of a division, and so on and so forth. Okay, so now we can easily count up the copy number on the top versus the bottom and really quantify the asymmetric or random distribution. So now I'm showing you exactly that analysis. We're looking at across, again, multiple different kinds of cancer cells and multiple different ecDNAs. And we can basically find that ecDNA distribution to two daughter cells completely follow a random Poisson distribution, right? Exactly matches up the theory. Some of you may think that, well, maybe this is asymmetric segregation. But the asymmetry is definitely random because if you imagine, isolating prospectively one cell that's gotten a lot of ECG copy numbers. If they were programmed, then the next cell division, you would also have another cell that enriches furthermore. But if they're random, you'll basically regress back to the mean. The experiment showed that we see regression back to the mean, meaning that it's truly a random segregation. Our model, after a number of analysis, showed that these just two features of ecDNA can remarkably capture their copy number and dynamic presence in patient samples. And these two features are the random identity by descent, which I just showed you, the random segregation, and presence of positive selection, that is, daughter cells with more copies would actually grow better or survive better. So on the middle here, I'm showing you a graph looking at the range of ecDNA copy number versus the frequency of cells with that kind of copy number. So you can see this very wide dynamic range. The red dots are what actually happens in the patient tumor sample. The black dotted lines is what we— what comes out of a simulation. We start with the one cell of ecDNA, simulate to lots of cells with lots and lots of runs. But just by following these very two simple principles, we can actually, in fact, basically almost capture the dynamic range we see in patients. I use the word selection, and so we should really prove that. Jack Rose in the lab developed a way to write extra-chromosomal DNA. This is a method called CRISPR-C, circularization. Basically, what we do is we basically make two simultaneous cuts mimicking the easy DNA biogenesis. At some low frequency, that the two ends will circularize, creating an ecDNA junction. We can also separately follow the scar that's formed, and this creates a scar junction. We chose to do this in the haploid cell type so that we only have one copy of the chromosomes to deal with, and we can basically analyze very accurate digital droplet PCR. Now we did this, Jack did this with the dihydrofolate reductase gene, which you can select with methotrexate, Indeed, we see that after ecDNA generation, in the absence of selection, we basically have decay and then neutral drift. But if you impose positive selection with methotrexate, you can basically drive up ecDNA copy number. The range that we see basically correlates, it comports very well with our model, and that shows for the copy number we see, that the ecDNAs must give like a 30% or 20% or 30% advantage every cell generation to have achieved this kind of a positive selection. In summary, what I've told you then is that chromatin accessibility is a powerful way to annotate a non-coding genome in diseases, especially cancer. Secondly, that hyper-accessible ecDNAs will converge in ecDNA hubs, to mediate massive oncogene expression. And this occurs through the new mechanism of intermolecular gene activation. Thirdly, that the mutational records show that ecDNAs are driving oncogene activation, therapeutic resistance, and transcriptional rewiring, and that I think these new methods of reading and writing extrachromosomal DNA would help us discover therapies that target these important elements in cancer. I want to acknowledge the folks who are involved in this work, starting with Ryan Korsas, who led the work on TCGA, now a faculty member at UCSF. Ting Hung and Katie Yost led the work on ecDNA hubs. Ting also led the work on CRISPR-Cas, whereas Jack contributed to this ecDNA evolution story with CRISPR-C. I gratefully acknowledge my collaborators, especially Paul Michel, for all the work together on extrachromosomal DNA. And grateful to acknowledge my funding sources. Thank you for your attention. So thank you so much. And glad to have a chance to exchange ideas with everybody.

Thanks so much for speaking today, Dr. Cheng. See you later.

Take care. Bye-bye.