Scientist Stories: Elaine Mardis, Genomics and Medicine

So welcome, everyone, for coming here. We are extremely pleased to have today as our speaker for the GBM Seminar Series Dr. Elaine Mardis, who is the Robert and Louise Dunn Distinguished Professor of Medicine, Genetics, and Molecular Microbiology at the Washington University School of Medicine, where she is also the co-director of the Genome Institute. So Elaine is a graduate for both her undergraduate degree in zoology as well as her PhD degree in chemistry and biochemistry from the University of Oklahoma. Go Sooners. And since then, she spent most of her career at the Washington University School of Medicine, where she's really been a leader, a cutting-edge leader in technology development, a major leader in the Human Genome Project. Her major research interest is very appropriate for this seminar series. Where she's mixing high-throughput cancer genome sequencing with really applying it to therapeutic decision-making in the clinical setting. She's a recipient of a number of awards, too numerous to list, as well as a supervisory board of a number of institutions. So it is our great pleasure to have Elaine speak to us today. Elaine.

Well, thank you so much, and thanks especially to Eric Schott for the kind invitation to present here. It's been a while since I've been at Mount Sinai. I'll tell you about that in a little bit, but it's really a pleasure. I was going out to Long Island anyway to be at a Cold Spring Harbor meeting, so I thought I'd just stop in the city for a few hours and give this lecture while I was at it. And what I'm going to do today is really feature some recent work from our group, but also with a bit of a historical context so that I sort of take you on a journey from the early discoveries we made in cancer genomics to the real clinical translation that's up front now. And I'm going to do that across a number of different tissue types, if you will, ranging from breast cancer to AML to the most recent work that we've published in melanoma in immunotherapy. And so I think I'll just go ahead and get things started. Focusing on some work that we've been doing and is not yet published, but we're preparing for publication in the genomics of late-relapse ER-positive breast cancer. So, as many of you know, breast cancer is a very prevalent disease, and ER-positive disease is about 60% of all breast cancers that are diagnosed. Some work that we've been doing for some time looking at ER-positive disease recently revealed, as did other work, that there are gain-of-function mutations and translocations that occur into the estrogen receptor itself. And this is typically in late relapse patients that have been on estrogen suppression for very prolonged periods. So effectively, the new cancer cells have found a way to invent or evolve around that estrogen suppression by these modifications. We wanted to delve into this further. Further with a study that I'll tell you about now, really sort of taking a retrospective look at very old samples for which there had been a consistent type of treatment applied and known outcomes. And it— from a technology sort of side, the reason that this is incredibly challenging is that in breast cancer, typically these patients are not relapsing until 15 or 20 years out. So that means you would have had to have the foresight to keep all of those clinical trial blocks together, to follow all of those patients over time, and to really understand their outcomes. That's not the technology challenge. The technology challenge is actually now taking fully 25-year-old blocks in formalin and extracting nucleic acids that are going to be suitable for sequencing to give you the answer that you want. In the context of that clinical outcome for that patient. So, just to kind of set the stage in terms of why this is hard, I'll focus on that a little bit more. So, where we started was really by saying we've got this clinical— set of clinical samples from a trial that Torsten Nielsen and colleagues in British Columbia did many years ago, and this has been reported on. You can see it's a large number of patients, 5 years of adjuvant tamoxifen monotherapy, and a 10-year follow-up. On average for these patients. And so the idea was to take the primary tumor, sequence it, and then correlate it to outcomes for these patients. And I'll show you data for that in just a moment. Where we started was essentially taking the big data that had emerged from sequencing thousands of breast cancer patients in the primary disease setting and really sort of putting together a list of genes that are frequently mutated, with a special emphasis, of course, on ER-positive of disease. So this is just sort of a brief vignette on how we did that. It was much harder than these 4 bullet points capture, but essentially we came up with a list of genes that were selected from 8 large published genomics discovery, including TCGA and some ICGC projects and others. The genes were simply selected if they were mutated in 2% or more of cases, or were considered to be druggable by having more than 2 drug-gene interactions indicated by our drug-gene interaction database, which is cited here. And then we essentially identified this novel reagent called Ultramers. These are available from a company, IDT, which is based in Iowa, and these are 125— 120-mer biotinylated probes, and we just tiled these across each axon of every gene, and the final gene set is about 83 genes and only a small proportion of the genome, about 400 kb. So we set about then preparing libraries from all of these 600-plus samples and sequencing them through a capture step using the Ultramer probes that I just described. And this sort of gives you some idea of the metrics around these. So average 25 years since diagnosis for these old age blocks. The known relapse status, of course, was critical. Here's a challenge: no match normal DNA, because who would have thought, you know, 25-plus years ago that it would really be more facile to do a tumor-to-normal comparison? So we actually had to come up with a number of inventive steps, which I'll show you in the next slide, about how to sort of subtract out suspected germline mutations so we got a true picture of the somatic landscape for these patients. So we were able to do automated library construction from essentially fairly low amounts of input, 50 nanograms is really sort of pressing the limit, put together indexed or barcoded libraries so that we could pool 96 together at a time, do the capture, and then do the sequencing. And remarkably, for these old age samples, we passed about 99% of them, which is defined by 80% of the targeted bases being covered at 20-fold or higher coverage. And you can see some of the metrics maybe here along the bottom, although the resolution isn't great. But essentially, very low duplication rate. I'll just point out we worry about duplication rate when we have very limited quantities of sample going into the capture. Because you tend to capture less and it tends to amplify more, jackpotting through the PCR steps that are made to amplify the library prior to sequencing. So this just kind of gives the overview of the study design that I've just described, but in particular I'll focus you here on the variant filtering that we were able to do for these patient samples, where essentially we removed false positive and germline variants and step the numbers down pretty dramatically here by doing comparisons out to a variety of different databases, such as the TCGA breast cancer blood normals, about 900 samples, also from whole genome sequencing in TCGA, and then a variant detection step that we use that we haven't published yet, which is the database of curated mutations that essentially reexamines the sequencing data for known mutated sites to ensure that we haven't just missed those out because of some aspect of coverage or some other reason. And so we're able to salvage things, actually, at this step. And then finally, an intensive manual review of all of the passing variants so that we could really identify and filter out any remaining false positives. So one way that you compare what you've got when you're all said and done is you compare back to other large datasets. So this is just a representative mutation waterfall plot of all of the— of all of our 600 samples compared to TCGA. And essentially we recapitulate the predominantly mutated genes, especially in ER-positive breast cancer shown here along the— along this axis, and you can see that we've been able to detect a variety of mutation types. In the data that were generated. So this actually looks quite promising, and if you can't read the slide, PI3 kinase, PIK3CA, here is at the top of the list. Okay, so what does this mean in terms of outcome? We were actually able to get, as shown here, a couple of really exciting and basically unexpected results here. So keep in mind, these are primary tumors that are being sequenced and then compared to the outcome for the patients. So two of the surprises that I show you here are DDR1 compared to breast cancer-specific survival. You see a statistically worse survivability here for the mutated versus the wild-type. Numbers are still a little bit on the small side, but the p-value looks good. NF1 was also a complete surprise. Again, a worse outcome, again, but still a small number of patients with that mutation. In it and versus the wild-type NF1. More on that just in a second, which is that if we then really filter for only the most deleterious mutations in NF1, we identify again the same statistical— or the same relative Kaplan-Meier curve for breast cancer-specific survival percentage. This is a gene that's been sort of implicated, if you will, in the context of metastasis of breast cancer. FOXC1, so we wanted to take a look at it. And you can also see that the mutated FOXC1 portends a worse outcome in the disease than the wild-type mutation— or sorry, the wild-type gene. And what else do we have here? Metastasis-free survival in particular, we also see quite a difference here in the FOXC1 metastasis-free survival. I'm not sure you can see that green line very well. But this is— the green line is for distant metastasis-free survival versus locoregional, just in terms of where the recurrent tumor was at for the patient. So, overwhelmingly, the distant metastasis-free survival is the worst outcome on this curve. So, here's just another sort of interesting vignette. This is ARID1B, and here we see sort of subtype-specific outcomes, so there are two subtypes in ER-positive disease by what are called intrinsic subtyping, luminal A and luminal B, and you can see hopefully that the green line here is mutated luminal B tumors for this ARID1B mutation class have a statistically worse outcome, whereas luminal A essentially is completely untouched by this particular type of mutation. Mutation. So these are just a few examples of what we were able to do combining this very large dataset with outcomes. We're actually poised to repeat this on another set of samples coming from the ICR, from Mitch Dowsett's lab, similar banking and clinical protocol, all ER-positive disease, and hope to increase these numbers so that we not only sort of reinforce or refute the statistical relationship of these different mutations to outcome, but also can begin to hopefully start looking at co-occurrence. And we actually have some interesting data that I haven't shown because it's not yet reached statistical significance but needs more samples in the context of two frequently mutated genes, namely PIK3CA and MAP kinase— blanking on the gene name, sorry, I got up too early this morning to get here— but MAP3K1. And they actually have different co-occurrence versus mutual exclusive patterns in terms of the Kaplan-Meier curves that are predictive of better outcome versus worse outcome depending upon the mutation status. So we hope to not only reinforce those statistically, but also maybe come up with new rules about co-occurrence and mutual exclusivity for other genes that have been implicated so far, and perhaps new ones as well, to sort of add to the outlook. Work in terms of late relapse disease. So just to conclude for this part, I think what I'm trying to demonstrate here, just one example, is that this large bolus of discovery data can now really permit these subtype-focused questions to be asked. And I think, you know, this is just one example, certainly, but you have to have the right samples and the right technology approach to get the data out and to really begin to understand it and to have lots of samples available to get some replication as as well. And I think we're going to need really large cohorts of patients with similar treatment and long-term follow-up, as I've showed you. And also, just the right bioinformatics approach can, I think, take you a long way. Okay, so I'm going to switch gears now to the hemalignancies, specifically AML. And for those of you who know the work from our group, this is really the disease in which I think we've had— as I'll try to demonstrate— a pretty major impact in terms of understanding the genomics When we set about this in 2002, we didn't have next-gen sequencing. We had two things: PCR and capillary sequencing, a powerful combination. So, back in the day, what we wanted to approach using PCR and capillary sequencing was this fundamental question about AML, namely that most of the patients actually fall into an intermediate risk category. They have cytogenetically normal chromosomes, unlike patients with favorable risk who typically have known translocations such as PML-RR alpha, or adverse risk patients that have very complex cytogenetics. These patients really had no indicator from cytogenetic examination of how they would outcome in the disease. And we also just sort of wanted to begin testing through these samples on a series of sort of candidate genes to see if we could come up with some new ideas about genes that were contributing to the development and progression of AML. So that was really the idea, so we started with our list of genes, our PCR primers, our sequencers, and off we went. And at the end of this sort of first year, we, with an NCI-funded program project grant, we had paced through a lot of genes and a lot of samples but didn't really have a lot of new hits. So along came next-generation sequencing, and as we were developing what was then the new Solexis sequencer in our laboratory, we started to get kind of an idea about what we might be able to do with this that would get us away from this sort of look under the lamppost and actually allow us to do unbiased looks at the genome. And so we started to ask ourselves, well, gee, why wouldn't we want to just use this method, this technology to sequence whole genomes and compare tumor to normal and let the genome tell us what was changing in the tumor cells themselves. So that sounded like a good idea, but of course there were lots of reasons not to do it. One reason is that it would be too expensive, so at the time we were sequencing just one end of every fragment at 32 base pairs. This would take on the order of about 100 machine runs just to get 30x coverage on the tumor. Then you'd need another 100 to do 30x coverage on the normal. So the dollar count started to mount quite a bit. This did not deter us from suggesting it during our renewal for the program project grant for our project, and we were, of course, soundly kicked to the curb by the reviewers who said, no, it was too early and it was too expensive. So we went to a private donor in St. Louis and asked for the million dollars that we needed to do this project, and he gave it to us. And I've presented this data at Mount Sinai in 2008, the exact day that the paper came out, and it made the front page of the New York Times below the fold. I guess that's important for some reason. So on my way back to the air— at the airport at LaGuardia, I picked up a couple of paper copies of the paper and then noticed that the chancellor of our university was actually standing in the first-class line waiting to get onto the same plane as I was. Back to St. Louis. I said hello to him. I don't think he knew who I was, but I pointed out to him on the paper that we actually had the front page of the New York Times, and I think he clicked in then. He kind of knows who I am now, so it's— things have gotten a little bit better. But this was truly the Million Dollar Genome, as I've already said. Not only did we need that amount of money to generate the data, but the bioinformatics of comparing the tumor to the normal you know, had basically never been done before. So, fast forward, we also sequenced the second tumor-normal genome with next-gen methods and reported this in the New England Journal, 2009. This was a remarkable N-of-1 patient because not only did we discover some genes that we knew about in AML, NRAS and MPM, nucleophosmin, but we also picked up on this novel gene that, you know, nobody would have ever, let's face it, put on their list of suspect genes, which is isocitrate dehydrogenase. Glycolytic enzyme. And this was important because as we showed in the paper, there was a statistical association with worse outcome if you're mutated in IDH1. And following on from that, several groups— we were not the first to describe IDH1, it was Bert Vogelstein's group in secondary gliomas, we described it first in AMLs— and then several other groups including Craig Thompson, who I've heard runs some sort of a hospital around here. And his group actually identified that this mutated protein was not just mutated but actually had a completely novel function in the cell now. So you guys all remember glycolysis, right? Converting isocitrate to alpha-ketoglutarate is the job of isocitrate dehydrogenase. But of course in the mutated form what happens is that alpha-ketoglutarate is converted to this novel metabolite, 2-hydroxyglutarate. Interestingly, you can actually detect this in the bloods of patients with IDH mutations, and this is a novel biomarker, et cetera, et cetera, but more importantly, in a fairly short period of time, we reported this mutation in 2009, Steve Vogelstein, a group at Agios Pharmaceuticals now has a drug in clinical trials of AML patients, primary refractory AMLs. That, for all intents and purposes, is looking pretty good, that essentially takes the brakes off of hematopoietic development that are imposed by IDH1 mutations and appears to allow these patients to go forward to normal hematopoiesis. So that's actually, I think, a reasonably remarkably short time to go from genomics to therapeutics. In the course of all this happening, we also started to ask questions about evolution. And so this beautiful figure came from a Nature paper in 2012, as you can see, where what we did was to compare the whole genomes of patients in the primary diagnostic setting to their relapse specimen and to begin to characterize the differences that crept in at relapse. One of these differences is that you get a lot more mutations in the relapse disease, and what we were able to show in this paper, only studying 8 patients, is that these all have a common signature. And it's essentially the signature of damage from chemotherapy that's meant to put these patients into remission comes back in their relapsed disease as a predominant type of mutation, studying the transition transversion rates. What we also showed in this study was that while AML had historically been thought to be largely a homogeneous presentation, in fact, patients, even to our limited ability to detect down to low levels at this at this early point in time and using next-gen, had often multiple subclones represented in their primary disease. And interestingly, in every case, one of the subclones from the primary disease survived chemotherapy, brought on additional mutations as I've just described, and emerges as the sole component of the relapse disease in these patients' marrows. And so this was a fundamentally, you know, new understanding of primary to relapse transition. So more on that in just a second. Now I want to sort of talk about where we're going with all of this, having been through lots and lots of discovery. So we did a few patients here and there when sequencing was pretty expensive, and then we scaled up as part of TCGA. This was again a New England Journal paper just a couple of years ago, and all 200 samples for this study actually were banked individuals from Washington University. And you can see the array of different types of omics that were used to study these. Whenever you do one of these big projects like this, you have to, have to produce a figure that looks like this, completely incomprehensible, to be sure, but just sort of describes the landscape of the disease, if you will, across different genes that are shown here that sort out into different pathways. You can see co- and mutual exclusivity patterns come out. Et cetera, et cetera. So I'm not going to talk about this or even try to interpret it, I'm not sure that I can, but what I want to show you is it's a big dataset and now we have to try and make sense out of it in the context of how we treat patients daily in the clinic. So where do we start with that? Well, first, the other thing you have to produce when you sequence lots and lots of tumors is a list of significantly mutated genes. It's a requirement. You have to have one of these charts, and in AML which is a disease that typically has on the order of about 10 to 12 average mutations in genes in the, you know, in every patient, the list is not terribly long, 23. You can see it here, and the so-called long tail happens pretty quickly, as it always does. So what we wanted to do, since 200 is a big number but it's not a huge number of patients to get this information from, is we actually extended out to just recurrently mutated genes, and the number there is about 260. 60, and we defined that just as any gene that we'd seen mutated twice at a minimum in our 200 patients. So that's how we defined that. There's no math around it in terms of statistical significance, but it just allows us to broaden the look a little bit more in the clinical setting, as I'll describe for you in just a moment. We designed probes to these 259 recurrently mutated genes. Just similar to the ones that I described earlier for the breast cancer, and we did some experiments that showed that if we actually dope these into an exome reagent, into an exome capture, we could return very high coverage on the genes that we were most interested in. And so the idea was, going forward in the clinic, let's not restrict ourselves in terms of more discovery. We'll do exome sequencing, but we'll also enhance our coverage at these known frequently mutated genes. And we'll begin to better understand this across patients in real time in terms of whether we can raise some of those recurrently mutated genes up to statistical significance or not, and so on and so forth. So this was sort of the logic behind what I'm going to tell you about next. Before I transition to that, though, I also want to point out that RNA is important. Okay, so in AML, we know that a lot of fusion proteins, for example, PML-RR alpha, which I mentioned earlier, are known drivers in the disease. Disease. This is the necessary and sufficient fusion that drives patients with acute promyelocytic leukemia, a highly treatable form of the disease with a high cure rate. And even in our study at TCGA, we found about 80 of 179 patients with fusions, some of which we'd seen before, like PML-RA-RA alpha, some of which were completely novel. Furthermore, and I don't know if this is sort of an obvious thing to say, but I'll just say it. By studying DNA only, we miss a lot. We don't actually see the impact, for example, of changes in chromatin binding, changes in methylation, or other signals such as mutations in promoter regions that actually may cause a gene to be over or underexpressed relative to expectations, and we won't see that unless we go and look at RNA-Seq as well. From a clinical standpoint, this is a pretty scary prospect. Most people in the clinical setting don't want to think about sequencing RNA. It's very complicated. It's very labile, but I think it's very, very important. And so our plans, as I'll describe in a moment, are to also pursue RNA-seq in the patients that we're studying in clinic. Okay, so now I want to layer a little bit of the clinical context of AML on top of all of this before I get down to the nitty-gritty details. What you can see in front of you is a sort of assortment of patients' outcomes in this disease. So this is a highly fatal disease. Overall, but some patients actually do pretty well. They have long first remissions, like those patients that we studied in that Nature paper that I showed you from 2012. Those are all patients that were more than a year out in remission and had not yet recurred. There are a handful of patients that actually have a prolonged first remission of greater than 2 years, but very, very few. Then there are these sort of intermediate patients. But really, the confounders in all of this are over at the left. So, firstly, primary refractory patients. These are patients that essentially never respond to chemotherapy at all and don't diminish their disease burden by going through even sometimes one or two very aggressive chemotherapies. We need to know about those patients earlier than we are learning about them now. I'll describe that in a second. And then the other conundrum are early refractory patients, which are patients who look like, for all intents and purposes, after this 30-day course of aggressive chemotherapy to rid them of their disease and get them into remission, they actually come back very quickly with recurrent disease, and we don't really understand why that is. But rather than sort of scratching our heads about why, what we thought might be a good idea would be to actually monitor patients from the point at which they present with the disease, so-called day 0,, and at the point where they complete their first round of induction chemotherapy, and then to do sort of a compare and contrast between the genomics of their first tumor and the genomics of whatever residual tumor might remain behind. And so I'll tell you how we're going to do that in just a second. I just want to point out that we do have some other predictors, and most importantly, that this is not meant to be sort of a standalone diagnostic. But an idea that we want to pursue in the clinics to see if we can actually add information to help figure out patients that are on the left or patients that are on the right. So that's what I'll describe for you next. So this is just going to be the clinical workflow that I talked you through just a moment ago. We have our diagnostic marrow from which we'll extract DNA. A skin normal is typical, although we can use a mouthwash because sometimes skin is problematic, and then a day 30 marrow, which is commonly taken as part of the endpoint for induction chemotherapy, so this is obtained as part of the standard of care, and we'll extract DNA from all of those, and then as I've already sort of walked you through, sequence those by using an intermediate step of exome plus recurrently mutated gene hybrid capture, and we'll do this for the tumors and the normal, and then we'll do high sensitivity variant detection and essentially say, what are the mutations in day 0? What are the mutations, if any, that we can detect at day 30? And what are the variant allele fractions? So that means the number of sequencing reads in this sample versus this sample, how much has it gone down, if at all? This is really part of the sensitivity that you can gain from using next-generation sequencing. And as I already mentioned, we'll study the RNAs for these two, compare RNA fusions as well as expression levels in the day 0 versus day 30 to see what's persisting there. So that's the— that's the paradigm that we plan to pursue, and the idea is that for each newly diagnosed de novo AML patient, we'll do all of the sequencing, we'll compare data from day 0 and day 30. And what we're really going to focus on are these recurrently mutated gene sites. Specifically, as I've already told you, the variant allele fractions and how they change. We also want to look at the exome for any novel mutations that might crop up so we can add those to our catalog. This also gives us a better ability to estimate heterogeneity in the primary disease in particular, but also in the recurrent relapse disease. And then the transcriptome I've already walked you through. The bottom line is, is that by looking at variant allele fractions, especially at day 30, We hope to compare evidence from next-gen sequencing versus the standard of pathology, which is typically flow cytometry of these marrows to identify any residual tumor cells. So, whenever you propose doing something like this, obviously, you now want to go back and look at a bunch of retrospective samples for which you know outcomes, similar to the breast cancer story that I told you at the beginning. And the idea there is sort of, will it work? If we actually know whether these patients are primary refractory, you know, if they had a very initial return to disease progression, et cetera, we want to know— we need to know that, and we need to now examine through these samples to see if even our hypothesis looks like it can predict patients that will recur in the disease. And how quickly. So these are just kind of two vignettes, two different patients that we applied this to. We've now done this to about 50 patients, so I'll show you the composite data in just a second. But this is sort of the ideal case, right? So this is a patient with complete clearance at day 39 when we've evaluated the marrow versus day 30. You can see that there's clear evidence of mutational burden here. I mentioned MPMC earlier. Here's an NRAS mutation that looks like it's in a minor subclone. But in essence, the entirety of the subclone population goes to zero and stays there even out to day 384, which was the most recent marrow sampling that we had for this patient. So, that's technically what you would love to see for patients because this is a patient who's having a long first remission. By contrast, here's a patient with fairly incomplete clearance. So, this is day zero, this is day 31. You can see diminution, but not down to the baseline for all of these for most of these mutations except for this minor subclone here with a FLT3 internal tandem duplication, but then after a short while, this is day 90, this patient actually is a very rapid relapse patient and comes back with essentially the full complement of mutations that were present in the first place, some at differing variant allele fractions, but a patient with with fulminant disease, no doubt. So then here's just the composite outcomes for event-free survival less than 12 months. So these are patients that get— come back in relapsed disease very quickly. You can see that it's not all the same diminution to baseline at day 30. There are still patients with active disease, and these are probably patients that are, to a first approximation, fairly treatment refractory. And then here are the patients with event-free survival greater than 12 months, and they look for all intents and purposes just like that individual patient that I showed you earlier, with some patients perhaps emerging towards relapse disease, but it's too soon to tell, which is a small number of lines that are going up above the baseline over here. So, the bottom line is, is that this retrospective study tells us that it's time to really move this idea forward in the CLIA lab, studying patients that are being treated in our clinics,. And we're going to focus on de novo, or first presentation, AMLs. Just to be clear, there are a large increasing class of secondary AMLs, patients that come back with AML who have typically previously been diagnosed with a primary solid tissue malignancy, often women with breast cancer. And this is, again, due largely to the DNA damaging effects of some of the chemo and radiotherapy that we're using for patients. They emerge with AML. Okay, so that's kind of where that's at right now. I hope that we'll be moving forward into our CLIA lab with that assay very soon. We essentially have all of our paperwork in place, and that's where we plan to go with this all next. So I'm just going to finish with something completely different now, which is work that's come out of a couple of avenues that I've been pursuing at WashU. With our immunology folks, and we have a very strong immunology department, I found out one day. And the way that I found out was I was standing in for my colleague Rick Wilson during the sort of practice talks that we give to prepare for our renewal for the cancer center grant that gets renewed, or hopefully gets renewed, every 4 years from NCI. And the guy who was going before me in terms of the speakers was this guy Bob Schreiber. Who's a very well-known immunologist. I didn't know that at the time, but I know it now. And a couple of weeks later, after we sort of had a chance to chat during the rehearsal, he called me up and he said, you know, can we get together and have a cup of coffee because I've got some ideas that I'd really like to talk to you about? And I said, sure, I'm always up for a free cup of coffee. So we met, and what I'll tell you about next, just in one brief slide because it's already published, was really what emerged, but has now led to the second study, which just came out in Science, and I'll finish with that. So the ideas that Bob and I sort of talked about over coffee and eventually came forward in a Nature paper in 2012, and then this most recent paper that was just out at the end of last year, 2014, was could we springboard immunology ahead in terms of identifying those neoantigens, those tumor-specific peptides that were most immune stimulatory to the immune system. This had historically been done through a series of, I guess, what are called cloning steps, but was incredibly arduous. And Bob had this idea, and it was his idea alone, I was just the help here, that we might actually be able to do this using genomics to sort of point out through a variety of steps that I'll walk you through now which of these novel mutated peptides coming out of a cancer genome were actually the most likely to have strong interactions with that immune system. So Bob's work is all so far in mice using a mouse sarcoma model that emerges from treatment with a chemical carcinogen, methylcholanthrine. And methylcholanthrine, like most chemical carcinogens, UV light, you name it, causes loads and loads of mutations. So this is a very sort of different set of cells than normal cells. In the mice that are treated with this chemical. So the pipeline that we developed is just a synopsis here, and you'll see the similarities next to the human disease pipeline; they're quite similar. But basically what we did was, in Bob's paradigm, we started with RNA converted to cDNA from the tumor. We compared it to normal DNA by exome sequencing, and we identified, using the fact that we're starting from RNA, expressed mutations, okay? So whenever you are working with a tumor type like melanoma or lung cancer in a smoker that has a very high mutational load, I can guarantee you that 40 to 50% of those mutations are not actually expressed at the level of RNA, and we need RNA to tell us that. So hurrah for RNA, yet another reason why we should sequence it. So this identified expressed mutations. And then what we did was we actually used a variety of predictive algorithms to take a look at each one of these novel peptides, so a new amino acid, so we're just talking about non-synonymous mutations here, and predict which of those looked the most different to that mouse patient's immune system. And so they all had a common immune HLA MHC class I type. So through epitope prediction, we then applied filters to sort of filter out things that we knew probably were not going to be real, such as hypothetical proteins. We synthesized peptides, and in this paper in Nature from last year sort of showed through a variety of analyses, both in terms of mass spectrometry and CyTOF, as well as in the mouse patients, that you could actually identify the proteins that were predicted to be immunoepitopes, see which immune molecules were actually carrying those that validated our predictions, and then look at in the mouse model whether if you immunize with these peptides you could actually get a mouse to either fend off cancer, so prophylactic vaccine, or most importantly defeat a tumor that was growing already in its body by vaccinating with a cocktail of these peptides. And so that was that 2014 paper. So many people, you know, will say, well, that's really great, you've cured the mice of the cancer, what about people? So I won't talk about cure today, but I will talk about a phase 1 study that we're just finishing up that was reported in Science just in— what is it, May? April 2nd in Science. And so this is work with Beatriz Carino and Jerry Lynette in melanoma. So what we did here was we actually took multiple metastatic lesions from several melanoma patients. These were all a specific HLA type, HLA-A02, and we sequenced tumor and germline, identified somatic mutations, and used RNA capture to, again, verify the expressed mutations. Of course, UV-associated melanoma has a very high mutational load, so we have to do this RNA step out here as well. And we also do this to look at the expression level of the RNA. It's not perfect. We'd ideally be going after protein by mass spectrometry, but we use that RNA expression level as a way of only going after the most immunogenic peptides that are expressed at high levels for our vaccine development. So this also gets looked at by an algorithm, as I just told you, and filtered out to identify the best immunoepitopes of the expressed variants in that patient's tumor genome. Of course, we would never sort of stop at this and say, "Right, that's the list." We actually go on with some apheresis samples from each patient to do some in vitro assays and really tell us which of these antigens, maybe there's some small amount of existing immunity, you'll see evidence of that, or whether they have the ability to mount an immune response to that particular peptide. So this is a culling mechanism that helps us to sort of reorder our list. And the reason why is because we want to be as precise as possible, as correct as possible, but also these peptides are quite expensive to synthesize, and so we really want to winnow the list down as much as possible. The other question that comes up when you think about doing something like this is, well, there's heterogeneity in solid tissue malignancies, there's heterogeneity in heme malignancies, so how do you really design a vaccine that's going to be as applicable as possible across all of the metastatic lesions in that patient's body? Body. And so to really get at the point of this question, we actually did, as you can see, for 2 of the 3 patients that were reported from this Phase 1 study, look at multiple different lesions, often taken at different points in time, to really get a feel for missense mutation load that could be predicted, those immunogenic peptides from the HLA-A*02, and then culling down using the RNA-Seq data those expressed antigens. And now you can see here the overlap where the blue-labeled peptides are the ones that were novel response, and these red-labeled peptides were ones where there was actually some hint of existing immunity from the patient. We applied all of this in a dendritic cell vaccine platform, which is shown here. There are many ways to use vaccines, so I've already mentioned the peptide cocktail that's just injected, which is what we did with the mice. Dendritic cells, I don't really know much about the immune system, but I know one thing, and that's that dendritic cells have a, have a role in the immune system of reminding T cells that they have immunity to a specific peptide. And so the idea here is that we mature peptides from the patients outside of their body, obviously, with our final list of peptides that have gone through a GMP synthesis, and because they're going to become part of the vaccine. So the dendritic cells will mature with these and several other components that we know are important and essentially carry those peptides back into the patient's immune system in the form of the vaccine, which is infused in the patients. So as you can see here, we've already given the vaccine to 4 patients. Patient 5 actually is now finished the vaccine transfusions, and we just have room for one more patient in this small FDA IND trial. The paper in Science that I've been referring to refers to our first patients, 1 through 3, because they were the furthest along and had been tracked for the longest period of time in terms of their response to the vaccine. And again, to emphasize, this is a Phase I study, so we're trying to accomplish two things here, not one of which is therapeutic benefit. First question is, are there any adverse events for these patients? So far, although numbers are small, no adverse events have been recorded, and this stands a bit in stark contrast to some other types of immunotherapies, which have been, you know, reported as having very extreme adverse events. The other question, of course, is if we think we can use dendritic cell vaccines to stimulate the immune system against specific peptides, can we? And can we actually monitor before and after the vaccination to identify those peptides that we chose that work the best? And so that's sort of what this little flow diagram shows here for the 3 patients, MEL21, 38, and 218. And you can see sort of before vaccine and after vaccine where you can see in the blue and red peptides that we actually did achieve some T cell response. Response from CD8-positive T cells. And of course it's not 100%, so actually each patient got 7 different peptides. We're not going after shared antigens here, although I will tell you that we had 2 shared antigens in GP100 that were included in every patient's vaccine as a control. They also did show T cell memory responses. So what does this tell us? Well, one thing, we're not as good as we need to be about predicting peptides that are going to result in a, in in a T-cell memory event. We're pretty good. We're 3 for 7, basically, in all 3 of the patients, and we're still in the process of evaluating patient 4 to see if we've done any better. But I think this is an area that needs continual improvement, as I'm sure you would not argue. However, based on those 2 criteria for the Phase 1 clinical trial, we felt that these results were worth reporting. And so just to conclude for this, we've shown safety, as I've mentioned. We've shown partial response in eliciting CD8+ T cell memory for the tumor unique antigens in 3 patients. And we have 2 patients that are still in process. One that's been identified, the other one is not. Oh, sorry, this is patients 4 and 5, and then we have patient 6 to still identify. And we do need to improve our neoantigen prediction approach, as I've just demonstrated. We're trying to evaluate the idea of using machine learning approaches, because we will ultimately know this endpoint for these patients, which comes from the T-cell monitoring. We may be able to use that feedback loop to essentially improve how well we predict neoantigens through the pipeline. And also, you know, we've only scratched the surface here, obviously, so we're just looking at non-synonymous mutations. How much more immunostimulatory, albeit more rare, would, for example, an insertion mutation be where you've actually gotten an insertion event that adds an amino acid, or indeed a translocation event that gives you a completely novel protein. So these are going to be rare events, but certainly we've seen them, and we do want to test them out, and moving forward, I think we'll be able to do that. I should also say that we're not going to stop at melanoma. We actually have a proposal for funding in on melanoma to continue this study where we're now going to combine the precision vaccine approach that I just described for you in combination with the checkpoint blockade therapy that approaches PD-1 and see if the combination of immunotherapies, checkpoint blockade, and vaccine is actually even more powerful. I will point out, in case you're wondering, all 3 of the patients that we described in the science paper are still alive, all still have stable disease or no disease.. And the first vaccinated patient was from January of 2013. So that just gives you kind of an idea of the timeframe that we're working in here. In addition, we want to move on and look at two other tumor types that we think we have the right expertise and the right patient population to go after. So those are early-stage resectable lung adenocarcinoma, where these patients would essentially, at stage 1, just be having the lesion removed from the lung. And would not go on to additional therapy. So we have a window of opportunity there where we can actually develop a vaccine, administer the vaccine, and then continue to monitor them. And of course, in adenocarcinomas, typically the relapse rate is fairly short, so 18 months to 2 years if the patient is going to recur with their disease. And then for triple negative breast cancer, we have our FDA paperwork in now for this disease type with some funding from Komen. This is an obvious target as well. There's really one standard of care for these patients, which is aggressive chemotherapy. Many patients survive this disease, but almost equal numbers recur with metastatic disease. So, again, we'll be going for a window of opportunity trial here for patients who have their disease resected, go through chemo and often radiation therapy, and this gives us sort of a 4- 4- to 6-month window where we can design the vaccine, have the final evaluations done and ready to give that patient at the end of their chemo/rads and once we've established that their immune system has recovered from that assault as well and is ready to respond if possible. So those are sort of the plans for the future. I'm willing to answer lots of questions about that because I will tell you that there are many more questions than answers at this point in time for the for the immunotherapy approaches, but we're pretty excited about what we've been able to do so far. So I'll just finish by saying that I hope I've given a few examples of how really the clinical translation of cancer genomics is underway. You know, we've spent the last 5, 6 years, along with lots of other folks, some of whom I'm sure are in the room, really trying to lay the groundwork for what goes wrong in the somatic genome, and I think it's now time, and many other people are doing this as well, to really understand how we're going to be able to change the lives and treatment of cancer patients by applying what we know in myriad and different ways. So I've kind of tried to illustrate that a bit with the work in breast cancer, which is pretty early on, but I think looks quite promising, with the work in AML, which we're just poised to put into clinical practice, and with the use of these individual cancer vaccines, which are already being used in the clinical care of patients. So, I'll finish and just acknowledge that there are a lot of people who never step on an airplane nearly as often as I do, and good for them. So, this is my team at the Genome Institute. In particular, these folks that are identified here in bolder print are the ones that really made substantial contributions to what I talked about today. Rick Wilson is the director of our center, and we absolutely couldn't do what we do without these incredible MDs and MD-PhDs, investigators that we work with both at WashU and some expats. Matt Ellis moved to Baylor recently. Torsten's also always been at Vancouver. And especially want to thank the patients and families that contribute their samples to these studies. Thanks for your attention. And now we're supposed to move to the Chairs, so we can answer questions.

So people can come to the microphones and ask their— actually, people go to the microphones to ask questions.

And the mics are because it's being broadcast, I think, to other places, so people can hear the question and I don't have to repeat it. Ross. Thank you. My question is about the peptides, which obviously is very promising. Do you have ways to figure out which antigens are likely to be shared across all the subclones, which would be an issue? And does chemo help you actually get a richer set of these peptides?— you know, that would allow you to have more choices and so on and so on. So those are great questions. The answer to the first is that we can, by doing that sort of Venn diagram-based approach, look at the prevalence. You know, sequencing through massively parallel means is very digital, and so we can get how predominant those mutations are in different disease sites if we have the ability to do that. So melanoma is obviously a special case, right, because there are lots of cutaneous lesions, they're fairly easy to punch biopsy, etc. That's not always the case for every tumor, so I think the challenge there is going to be getting a biopsy to begin with in many cases, but then studying it as well as we can and hoping that it's representative. So that may be more of a specialized case than for other solid tissue types. I think the question about chemotherapy giving you a richer landscape to choose from is a really fundamentally important question. And we're actually pursuing that. I told you about the work in AML, and AML is a great example where this might be brought to bear because AML de novo with 12 genic mutations is a really bad place to start, right? There are hardly any good choices there because probably 2 to 3 of those genes are actually drivers of the disease. And maybe not likely so much to be neoantigens. Where we're really pursuing this is right now today is in pediatric brain cancer. So pediatric brain cancer, again, is highly fatal. About 95% of kids die from that disease. And they have sometimes one or multiple recurrences, often with an intermediate of high-dose radiation and/or chemotherapy. And often the relapsed disease is removed by neurosurgery. Registry, so we have it available to study. Even most recently, our neurosurgeons have put together a whole brain legacy registry. So these kids that are going to die from brain cancer and their parents actually consent to donate the entire brain for regional spatial studies to look at heterogeneity. But the bottom line is we think that after high-dose radiation and temozolomide treatment, we're actually going to see a much more enriched landscape of mutations in these kids' recurrent disease and hopefully make them a candidate for a vaccine-based approach. But are you really— it's okay to go after brain cancer with this approach? Yeah. And not have massive brain inflammation and things like that? Yeah, we think it will be fine.

Okay.

I'd like to encourage folks to come to the microphones and ask ask questions. So one question I have in thinking about just the wonderful object of taking this to the clinic and getting CLIA certification is a workforce question. So you're dealing with AML, hematology oncologist folks are intelligent. How do you train the clinicians who are not necessarily doing research in this area to address questions or to deal with the clinical information that they're getting back, and how do you feed information back to them?

So I think it's a partnership in the way that we practice it. Physicians are like anybody else. That's good to know, right? There are those who are more adventurous and those who are a little more reticent to jump into the fray. So we tend to choose the former rather than the latter, or they choose us, right? And these are the collaborators that I mentioned at the end of the talk. So you do see, and this is really led by Tim Ley, you know, who years ago would come to us and say, "Right, you guys are sequencing the human genome. I ought to just be able to do PCR of all of the genes in the genome, right, and figure out what's going wrong in my patients." Yeah, we could do that, but now we really can do that. So I think it is a little bit of a self-selection where you're at a, you know, an academic medical school. It's not going to be every physician, but there are going to be those physicians often that have had research experience who really want to know more about what genomics can do for them. I think probably the best example of that, although I don't want to single him out, is Will Glanders, who is my breast cancer collaborator, who is the PI for this Komen grant where we are developing the personalized vaccines. I mean, I am actually proud to say that I have taught Will everything he knows about genomics because really as a surgeon when he came to us he sort of said, I want to take advantage of this, but I actually don't know what you guys do, so, you know, can you teach me? And he's, you know, come along and mastered the topic very, very nicely. Now, probably me doing surgery on somebody is not such a good idea by contrast, but I actually haven't been training for that, so I think that's probably okay. Yeah, but I do think it's just a specialized subset. It does raise an issue though that I think we all have to face up to, and I'd love to hear how you I know you guys are doing it here, but, you know, major medical schools I think have to really start, even in first-year medical genetics, start training, you know, physicians about genomics and what genomics is used for in medicine, not just in cancer care but writ large. There's certainly a lot going on. And then I think beyond that, in terms of the specialty residency training programs, you know, I can honestly say that I went to Vancouver earlier this year and spoke to their pathology residents who are learning about genomics as pathology residents, and I spoke about, you know, cancer sequencing and how it's done. But I've never been asked to give that lecture at my institution to pathology residents, and I'm not just trying to single our pathology residents out, but I think it's a larger goal, really, to introduce the concepts to more and more people at those basic levels. And then we have this real challenge on our hands, which is to take practicing physicians and really, if they're interested, of course, and want to use it in their practice, and really get them to start thinking about what is genomics, how is it being brought to bear in my particular field of expertise, and so on.

That's a perfect segue in terms of teaching genomics to medical students, so Andrew, maybe a first note about what we're doing here in Sana'a along those lines, and then your question.

Yeah, very well. So actually, I, I don't think Michael Linderman's here. Michael is the course director for our Introduction to Your Personal Genome course, where students actually have the opportunity to sequence their own genome. Okay. And that's been a fairly useful course for us because, first of all, it's, it's got a mix of students, so medical genetics residents, genetic counseling graduate master's master's students, some people with bioinformatics backgrounds, some people with laboratory background, which sort of gives you the full complement of people— actually, the molecular pathology residents are usually in there, too. You know, it gives you the full complement of people that would actually normally be involved in analyzing a clinical genome, so they help sort of teach each other aspects of that process so the instructors don't take the burden. So that's been, I think, pretty useful for us. I think we're in our third year on that now. The challenge, of course, is that's hitting people fairly late in their education; it's not, you know, not introductory. To medical genetics, and so we've been trying to explore ways to get that more integrated with the medical curriculum. And, of course, the challenge there is everyone's got to pass the board, so, you know, until that's part of the boards, it's very, very hard to get it through the, you know, through the boot camp, essentially. So we can talk more about that later if you're interested, but the question I had actually surrounds the clearance of tumor cells that you showed in AML. So, I mean, obviously everyone would love to see a picture where, you know, you treat and, you know, the nasty things are gone. You'd like to see that in infectious disease, you'd like to see that in cancer. Obviously sometimes you see it, and so the questions I've got for you are two. One is, how much of that do you think is a limit of detection sensitivity phenomenon? The other one is, how much of it do you think that is because that there is a refuge in some patients where, you know, the drug just doesn't penetrate for whatever reason?— or some other factors such as, you know, metabolism, or, you know, some admin effects on the drugs, or something else. I mean, you must have engaged with your clinical colleagues on those questions.

I'm curious to hear. [Speaker:DR. LISA BROOKS] Yeah, no doubt we have, and thanks for bringing it up, because it does give me a chance to say that we don't have infinite sensitivity, right? So even though we're covering those RMG genes at much higher levels, maybe 300 to 500-fold, right? That still gets you down in the range of maybe 1% of the tumor cell population that you're able to inquire and get a solid answer. People would argue, and they would be right, that actually multicolor flow cytometry, if you had the expertise and the instrumentation to do it, could get you down to much lower levels of sensitivity. And so I think that's really the bottom line is that we don't have infinite sensitivity to detect residual disease. We're sort of at the point where we're doing the best we can. We could probably go a little bit lower, but I'm not sure like where the— where things would bottom out at with that particular assay that we're using. And you could even argue that maybe something a bit more sensitive might come out of, for example, multiplex PCR as opposed to capture, which is what we're— which is what we're doing. So we probably need to explore to explore that as well. In terms of a place where cancer cells can go to hide, I think in some way, shape, or form this probably invokes the stem cell hypothesis, right? And I don't think I really know enough about stem cells to comment on that intelligently, but certainly is a formal possibility that this is where the sort of reservoir exists that then comes back in terms of disease. I guess the question would then be why do some patients, you know, sort of go into remission and not recur, but that's a fairly tiny proportion of patients, so. Yeah.

Yeah. I mean, just an interesting question because, you know, you showed two patients had a very different course, right? Yeah. And, you know, something was different about patient one and patient two, and, you know, the mechanism is something which would be actionable if you could figure it out.

Right. Yeah. But that's not— And we've tried long and hard to sort of, to sort of get at the genomic component of primary refractory disease, right, those patients that never respond. And I would say been woefully unsuccessful at identifying what that is through whole genome sequencing methods. So basically, really nothing, and though nothing in the genome sequence seems to correlate with that phenomenon where you sort of take the ends of the bell curve, right? So you take patients that are primary refractory versus long first remission, and we just fundamentally don't see anything different between— at the genomic level. Now that's not as comprehensive as we could be, right?

It's all a matter of time and money. But it is pretty interesting, actually. Yeah.

Because I mean, it sounds like it's a pretty good— Yeah, we had high hopes that we were going to come up with, you know, the gene or whatever, but that really has not panned out so far. So— Thank you. Yeah. Hi. About the AML study, I was wondering what was the reason that you didn't— you chose not to look at methylation differences in the patients? And I have another question in that study. For patients that received chemo and their cancer came back, the ones that relapsed for the AML patients, do you have any studies that are going on or you're planning to for the treatment of these patients? For which patients? For the patients that have relapsed after treatment, after chemo. Yeah, so we basically want to do this initial look using the genomics, and then in some proportion of those patients, for example, there may be an actionable mutation. Most of those patients, when they recur in the disease, just go on to a second, more aggressive round of chemotherapy. And the problem with that is— well, besides the obvious problem— is that a lot of these patients that do recur, and AML patients in general, like most cancer patients, are elderly. And so their ability to really go through a second round of aggressive chemo is pretty limited. So we think those are then going to be the best patients for if they have an IDH1 mutation, for example, they could go on the AGIOS clinical trial or wherever the drug is at, at that point in time. So that's one example. There are new FLT3 inhibitors, JAK2 inhibitors, etc. So I think that portfolio is flushing out a little bit and that would be sort of the most obvious place to take those patients, especially if they had trouble with the first round of induction chemo drugs. Drugs, they're really going to have trouble usually with the second round because of the more aggressive nature. So that's sort of the answer to your second question. And the first question was about— Methylation. Methylation. So great question. We know that one of the primary drugs that's used to treat AML is actually a demethylating agent. So I can guarantee you we're going to see difference between the before and after. I think the challenge there is that our skills at developing whole genome methylation have been developing, so we're not quite there yet. So it's sort of more us being slow than anything else. And secondly, you really are going to have to be able to correlate to the RNA-seq, I think, for those changes to really make any sense. And we don't have tools for that yet other than saying, "Gee, there's a change here. What gene is it next to?" to, and as you can imagine, that's going to get pretty tedious pretty quickly, so we need to glom on to something that's a little more automated, and we'll do that. We just have been slowly developing methylation. Thank you. Yeah, you're welcome. Yes? Very exciting talk. I had two quick questions. One was actually related to integrating the datasets that you're developing using exome sequencing, but also how does that How does that fit in the logic of looking at copy number variations versus LOHs and things like that? And will that— do you predict that would tell you anything about those mutations, somatic mutations? And my second question was, I guess, with the evolution of sequence technology, particularly focused on, for example, single-cell sequencing, how do you see that evolving in terms of of using that to identify potential stem cells, as you mentioned, or something along those lines. Right. So the data integration piece, we are— we're pursuing. What particular— Well, I was just curious with the whole genome sequencing, copy number variations. Oh, right. Exactly. So whole genome sequencing obviously is the gold standard, and we've pursued that long and hard, but, you know, oftentimes the budgets aren't there. We can do a reasonable job, I think, of predicting copy number from exome, but it's not great because the exome is a pretty limited space in the genome. One of the things that I've been looking at recently is sort of coupling exome with these really high-resolution arrays that Illumina offers, not to advertise for them but they're the only ones that are out there. It's like 4.2 million elements. So we've been making comparisons between what we get out of those arrays to what we get out of whole genome sequencing, and, you know, it's a reasonable representation for large-scale events, but will never get you down to the focused AMP deletions that you'll get from whole genome. As everybody here probably knows, whole genome sequencing costs are are dramatically decreasing with, you know, as long as you have $10 million in your pocket to buy an X Ten system, or I guess now there's an X Five, but anyway, so I think that'll maybe be a barrier that falls over time and we'll just hopefully go for whole genome sequencing on everything, but it's still a reasonably expensive proposition with cancer because you always have to oversample the tumor to really get that sensitivity that you want. But in general, I mean, we know, like, from TCGA, for example, that a gene can be altered by mutation, amplification, deletion, translocation, you name it, and we see that panoply of possibilities play out the more samples we sequence by whole genome. So it's— I think it's for the sake of comprehensiveness the best way to go to get the full answer in terms of genes that are frequently mutated. Okay, so I should just have the questions. One by one, I guess, because I'm too tired to remember what the second one was. No, and then just the one topic of single-cell sequencing. Oh, right, yeah, one of my favorite topics. Actually, I got in an argument over this last week. I don't think I won, but— right, so we did a small study in— that was published, I believe, in Nature Genetics a couple years ago, where we took a number of MDS patients that we had previously sequenced, and we sorted single cells and actually sequenced the single cells at different points, the MDS to secondary AML transition, to try and identify whether we could learn anything more from single-cell sequencing than we had learned from bulk tumor sequencing. So admittedly, only 3 samples, differing levels of heterogeneity in the MDS, blah, blah. I think we basically on one sample, which was the most complicated, were able to sort of temporally separate two different points in time in the evolution from MDS to secondary AML, but that was about the most that we got. And don't get me wrong, I'm not a fan of single-cell sequencing for the reason that it's not a cool, like, technical hurdle to overcome, but as we also showed in that paper, whenever you go through whole genome amplification, which everybody has to do to sequence single cells, you do miss out portions of the genome, right? So you're not always going to get the same answer from every single cell, but does that mean it's a different subclone, or does it just mean that that part of the genome where you saw something somewhere else, you know, was— has gone missing? So, I mean, I don't know. I think it's a cool technical thing for people to pursue, but whole genome amplification either has has to get better, or we have to be able to sequence from a single set of chromosomes somehow without amplification, or I don't know. There's still a lot of noise, I think, that has to get sorted out in terms of the single-cell-based approach before I'm really going to say, yeah, this gives us a huge advantage over what we can do with the bulk cell isolate. So— Thank you. Yeah, you're welcome. And I think the other thing that hasn't been done well with single cells is what I was just describing, which is sort of the primary tumor metastatic comparison to really say definitively, right, at high numbers of single cells for either one, that we missed out something that was there in that primary tumor that maybe swayed resistance or something in the metastatic disease. You know, so lots of interesting questions still to address, but you got to have a great big checkbook.

So I was really— I've always been impressed by the potential of leveraging nucleic acid analysis from pathology specimens. It's just this wealth of information that exists in pathology banks all across the country of just unparalleled benefit. How scalable is it? Because you described a, you know, a study, a great study, which I'm sure was incredibly time-consuming, tiling across 83 genes.

So kind of what's the long-term potential do you see for this going across diseases? Yeah, I think— I mean, I see a lot of potential, and actually I I've used that 83-gene panel for a lot of other work that I didn't describe today. For example, we're working with Tom Rohan from Einstein, who's an epidemiologist and has banked benign breast disease lesions for years and years and years in women who then either did or didn't go on to develop invasive cancer. And we've used that 83-gene panel, we're actually just finishing up the dataset this week on hundreds and hundreds of samples where we know the outcome and we can really begin to look at what the landscape is of benign breast disease that predisposes to invasive cancer versus not. So you can begin to think about using those old pathology specimens for sort of predictive tools, if you will, especially given the age on some of these samples. And they're not always going to be great. I mean, I always tell my collaborators, bring twice as many samples to me. You want to sequence because we're going to miss out about 50% of them because they just won't be adequate for whatever reason. But we'll— but I think there is a wealth of information as long as, you know, the appropriate clinical data are there, and that's really to me sort of the harbinger of whether you want to or can ask us, you know, an interesting clinical question or not. So I think there's lots more to be done in that realm, and especially now because we kind of have the common— I think everything sort of, you know, you can do it first, but when you do it best is when you've sort of waited to get all the culmination of all the tools that you need, and then you bring it to bear on a sample set. I was just meeting last week with a person about a project in breast cancer susceptibility, and now we're really at the gate— at and at the point in time, as I was just describing, where you could imagine sequencing through thousands of whole genomes, right, at 30-fold coverage, you know, $1,400 or $1,500 a sample, on sort of a large R01-type budget, and really beginning to look at the whole genome susceptibility to a disease as opposed to exome, which we know misses out a lot of the genome, as the other question was misses out a lot in terms of copy number and sort of focal amplifications and deletions. So it's all, I think, culminating towards the capability to use clinical specimens with a lot of the technological details that have developed along the way. Yeah, please. [Speaker:MALE_SPEAKER] Yeah, so I wonder what's your thoughts on circulating DNA sequencing, the plasma circulating DNA sequencing on cancer diagnosis? For surveillance for relapse. Yeah, circulating tumor DNA, circulating free DNA, exosomes, circulating tumor cells, I think is absolutely the wave of the future. I'm not doing any of it unfortunately right now. It really takes putting in place a clinical trial where those bloods are being banked routinely throughout the course of the patient's disease. And actually there's a trick with circulating free DNA to getting the plasma spun out very quickly, so within an hour. So it does take writing a clinical trials protocol that really requires, you know, those things to happen. But there have been people with the foresight to do that. I actually just heard a talk last week on squamous cell lung cancer where this is being done very, very nicely. And I think also in my frequent trips to Europe, I know that now the Europeans are actually introducing clinical trials where the combination of recurrent disease through imaging coupled with hotspot mutation analysis of the liquid biopsy for that patient is all that's required to get them onto a targeted therapy trial if they have the mutation or mutations that are being targeted. So no real tissue biopsy is now being required because it's often very hard to come by those tissue biopsies for recurrent disease. Disease. So I think it's moving forward very, very rapidly and it's again in the commercial space very contentious, lots of companies, you know, trying to apply technologies to it and I think it's really where things are going to move because of the exquisite sensitivity to detect when patients are developing acquired resistance or what have you. Yeah. So if you're looking for somewhere good to work that would be— it's getting crowded very quickly. I'm sure, but yes. Hi. So my background is in machine learning, so I have a question in one of your conclusion slides. Yeah. So what type of machine learning algorithm do you have in mind to kind of improve the neoantigen predictions? Yes, I know this much about machine learning, but I know that we're planning to use or compare a random forest model to a deep learning— this is a new term, right? Deep learning algorithm to see which one gives us sort of an advantage using these retrospective samples. And then we would do, as part of our clinical trials, then we would start using the winner of those two as part of our algorithm to really improve on our prediction, or at least that would be the hypothesis. Okay, I got the picture. So basically extracting the features from the proteins and then doing the classification stuff. Yeah. I wonder if you— what's the reason that in your talk there isn't kind of like too much network analysis and the mechanistic picture of the relapse and non-relapse because of the patient samples that you couldn't model that network? I think that's just a place we have haven't gone yet. The data are there. Okay. But I think that becomes perhaps most important, at least in my opinion, when you're trying to decide from a therapeutic angle, which is in a patient with maybe a couple of different therapeutic possibilities, where's the best place to apply the drug, right? And there you kind of have to have the overlay onto the pathway or the network, whichever you like. To call it, to really understand the best point to apply therapy. So we've done that, I just, I didn't talk about it today, but we've done it in considering from a patient who's had their genome and transcriptome sequenced for therapeutic indications, where's the best place to drug. And was the result promising to you? So far, yes, we've had a number of patients that where we've actually done that approach and they've gone gone on to receive the therapeutic and it's, you know, given some benefit, although of course as you probably know, the challenge in targeted therapy is that patients do acquire resistance often very quickly. So we've certainly seen that happen as well. This is why I think ultimately these immunotherapies or combined targeted immunotherapies are really going to be the combination that give this nice durable response for the disease-free maintenance. But that's being tested already by lots of people, so we're jumping into that fray.

Thank you. Yeah, you're welcome. So you talked a bit about how you're introducing genomics for RNA and DNA into the clinic sort of routinely now for AML. And so I'm curious your thoughts about how you want to apply that for, you know,— or what— how you would envision a fairly routine future, hopefully fairly effective AML diagnostic and therapeutic regimen based on the genomic data. So, just sort of drawing a little bit of inspiration from the HIV world where, you know, initially it was a real big problem, and you had one drug, people evolved resistance, and now, you know— so, if you got your HIV in the early days, you got a real problem probably with your HIV, but these days, if you newly acquire HIV, you know, you get hit with 4 drugs at once, or 3 drugs drugs at once, and, you know, it's a relatively manageable condition.

So, you know, what's your thought on how to get AML to being a similarly manageable condition? Well, I think there's still a lot of work to do, right? So I think, I think the one sort of obvious thing I've already said, which is coming out of, you know, remission into relapse disease, are there obvious sort of druggable points where we could get that patient on a clinical trial? You know, the nature— and this is was new information to me because we sort of think about AML as just a blood cancer, right? You've got mutations, you know, they cause the, you know, growth of the cancer cells over the normal cells, and all that is true, but really it's fundamentally different from the standpoint of the biology of what's going on, right? Which is why, you know, the Adgios drug that I cited is quite interesting because what it does in the case of IDH mutated AML, is essentially alleviates the blockade on further downstream differentiation in the hematopoietic cascade. So that's really what, you know, leukemia is sort of a stop point that's been applied in differentiation, and what's accumulating are those not yet fully differentiated cells. That's the leukemia. So it's a little bit of a different paradigm than solid tissue malignancies in that regard. And I guess we just have to be able to better figure out ways to remove the blockade.

It's interesting also, just because I mean, it's not— you don't necessarily want to kill the cells, you simply want them to let them grow up normally. Exactly, yeah.

That's actually a very different paradigm from a lot of things, including my infectious disease analogy. Right, including that. And so it's a different sort of mindset that you have to apply to the to the problem. Of course, unlike, again, the solid tissue malignancies and maybe infectious disease as well, the potential to graft in bone marrow from a donor, for example, means that if you can eradicate it in your own special way, you know, with high-dose chemotherapy or what have you, then you have the chance to give that patient a new hematopoietic stem cell component that can divide, you know, can develop out into a normal blood set of, you know, products, if you will. And so I think that's the other difference in the malignancies is that if you can actually get that patient by whatever means to really eliminate that cancer-based component, then you have the opportunity to provide them with a stem cell transplant. So it's, it's, it's almost a completely different way of thinking than the solid tissue malignancies.

Yeah, well, I'd like to thank the audience for their questions, and I'd like to thank Elaine for just a spectacular—

thank you, it was really nice.

Thanks.