 All right, so we're going to record this session for our participants online. There's been technical difficulties with sharing it for everybody in person. Welcome to Dr. Scott Furlan from Seattle Children's. I'll be presenting on bringing single cell genomics closer to the clinic for patients with leukemia. Everybody, give a big hand to Dr. Furlan, please. So, thanks for the organizers of the meeting today, Tim, for inviting me to talk. Sean and Mark Carlson also for their support. It's a real pleasure to be here. I'm a physician, scientist, an MD. I use genomic technologies to better understand relapse disease, relapse of leukemia in children. And I am a bone marrow transplant, so I attend at the Seattle Children's Hospital and perform bone marrow transplants on our patients who need them. Today, I want to share with you really a project in the lab that's just been building over the last year or so, and was really motivated by one patient's sample. And I'll show you most of the data from that sample today. But I do want to just take the time right now and acknowledge his sacrifice and willingness to participate in research. He really opened our eyes to the possibilities of using single cell RNA-seq for residual disease measurements in AML. And unfortunately, last month, he died after a long battle with AML. And he was a great kid and had a love of science and will really, really miss him. I devote this talk to him and his family, and thank you for listening. So overview, I want to introduce the concept of measurable residual disease, as I said, from the lens of one of our patients. Provide a rationale for the use of single cell genomic technologies to potentially improve measurable residual disease assessment after transplant. For some of our preliminary data, highlight some novel molecular and computational approaches we're taking in the lab to enhance MRD, enhance the confidence of MRD assessments. And I won't touch on it today, but by explaining our work in MRD, I hope you'll kind of see the broader applicability of single cell genomics to understanding biology of AML, acute leukemias, and mechanisms of relapse. So really, we're just kind of building that toolbox. And hopefully in the next few years, we'll be able to get even more samples and have a greater understanding of relapse leukemia. So I have no relevant conflicts of interest or disclosures to make. And so, again, going back to our patient, he's a 12-year-old boy. And our history begins with us when he was diagnosed with myelodysplastic syndrome in 2018. Myelodysplastic syndrome is a disorder of blood-producing stem cells that produces dysmorphic or malformed or strangely shaped red blood cells, white blood cells, as evidenced by the picture here shown in this diagram. And unfortunately, he went on and evolved to what we call acute myeloid leukemia, which is not uncommon for patients with MDS. He received a standard of care, which is a match-unrelated bone marrow transplant in April of 2019. And two years later, unfortunately, he developed low blood counts again on routine monitoring and was found to have a relapse of his disease. After relapse, he underwent re-induction chemotherapy and was being considered for a second transplant, this time using a different donor source of cord blood donor. And there was a conundrum, and we were struggling to figure out what to do next. We were measuring his disease in the marrow using flow cytometry. And by doing so, we were achieving measurements in the 1% to 2% range. So 1% to 2% of the cells in his marrow were leukemic. But when we were testing for mutational burden using bulk NGS testing, the variant allele frequencies of those data seemed to suggest that he actually had a much higher level of disease, around 10%. And so we really were struggling with this conundrum, and it was very frustrating for both the treating team and the family, as you can imagine. And you might say, OK, well, he's got relapsed disease, we'll just treat him, move on. And there's an important reason why we can't necessarily do that or why we need to be very careful here. So let me explain the sort of nuances in clinical care for patients who have low levels of disease after transplant. So if you have levels of disease that are less than 1%, you can consider maneuvers to try to boost that graft versus leukemia effect. And those maneuvers include altering immunosuppression, giving more donor lymphocytes, potentially treating with chromatin modifying agents to try to boost antigen production by the leukemia in hopes that the graft will destroy the leukemic cells. But we generally don't consider those maneuvers unless the disease level is low. So greater than 1%, they won't work. It's just not enough. And when we do this, we can alter the trajectory of relapse for a large number of patients, and we can cure them with just these small maneuvers. On the other hand, if the burden of disease is high, so if we have greater than 5% of disease, it is generally not considered a good idea to go to a second transplant. The outcomes for patients who have levels of disease greater than 5% are dismal, and therefore we don't even consider really taking those patients to second transplant. And so the paradigm or the clinical practice is to reduce the burden of disease by chemotherapy, and then once we achieve less than 5%, taking the patient. So you can imagine, we have two tests. One is saying go to transplant, and the other one is saying don't. And this was a real struggle. And so this really opened my eyes into really how the clinical tests that we have for measurable residual disease are not good enough, and we need to do better. And so this motivated us to try and use single cell RNA sequencing to try to potentially understand what the real level of disease is. Now, I'd like to start back up and just give you a little bit of background about how we test for measurable residual disease. Generally the tests fall into four categories, PCR-based tests, NGS-based tests, so the bulk NGS testing like I was explaining before. And then multi-parameter flow cytometry is the gold standard for pediatric AML. And then chimerism testing is another group of tests that we use. So let me just go through the strengths and deficiencies of each one. So PCR tests are really aimed at capturing fusion. So RNA fusion, chromosomal translocations, leukemias are often caused by chromosomal translocations, so we can PCR and capture those transcripts. They have very high sensitivity, but they're limited in their applicability, and you need to have a priori knowledge about what you're going after with the PCR test. You can't just fish for every fusion, and here's why. So we like to say that pediatric AML and ALL2 have what's called a long tail. And so these are data that Tim alluded to earlier in our collaboration with Tim and Sohail Mishinshi, my mentor. We have 1,210 patients who were profiled on the last cooperative children's oncology group trial, study AML-1031. And we just looked at all the different fusions that we see. You can see the major problem is that a large number of patients don't have a fusion at all. So there's nothing to actually PCR. And then the long tail problem is the other problem. You can't possibly go after all of these. I mean, you could, but it would be very challenging and hasn't been implemented clinically. So that's a problem for long tails, a problem for molecular assays. And similarly, we don't always for mutations. We don't also have, we have to have a priori knowledge. And there's a similar long tail of mutations that are associated with leukemia. So bulk NGS testings are even less sensitive, and so they don't work very well either. In this case, they did show that discrepancy. So that's interesting. So the standard of care really in pediatric AML is to use multi-parameter flow psychometry. It's more broadly applicable, it applies to many different kinds of leukemia. But it's limited in sensitivity. And I think what, what our experience with this patient has shown us is that it probably is, it underestimates the burden of disease. And that's probably for a lot of different reasons. But namely because I think there's clonal heterogeneity in AML in particular, ALL less so. And I think that confounds the ability to detect AML in clinical samples. So additional problems include the fact that it's really hard to run across multiple centers, right? You've got, everybody's using different antibodies, different instruments. Maybe someone didn't bleach it properly, who knows, right? And so it's a challenge. And chimerism testing is not currently sensitive enough to be used as an assay for MRD right now. It's plus or minus 5% on either side. So 10%, that's not good enough. So we, as I alluded to earlier, or I might have just said it, single cell RNA sequencing we think can meet this need. So because single cell RNA sequencing, and we're 10x genomics here, three prime and five prime measurements. And because of the massive parameter space in genes, a number of genes that we're able to detect in single cells, we think that this will give us a better ability to differentiate leukemia versus not. And so we'll compete with multi-parameter flow cytometers, the single cell method, and it can measure many, many, many new parameters. And what I'd like to share with you today is our experience using natural variation and mining the data for SNPs and single nucleotide variation to further differentiate recipient versus host, I'm sorry, recipient versus donor in samples. And I think this is a really important advance and will be a clinically useful tool for single cell chimerism. We have ongoing work using next generation detecting mutations using long read sequencing, and I think Tim alluded to that a little bit earlier. And we are working on trying to detect fusion transcripts in 10x data as well. So we think this is an exciting area for development. We're excited to talk about our experience with the molecular approaches and computational approaches, but we need help. We need help from computational people, from molecular people. So if there's anything you'd like to reach out to me, you want to get involved, love to hear from you, please do so. So what's the deliverable here? I think that we can deliver a more confident assessment of what that level of measurable residual disease is. And I think that's the direction that we need to move. Okay, so let me tell you about the experiment that we did on our patient that I described to you earlier. So we took a bone marrow sample. We really didn't know what the best method for processing the sample would be. There's lots of papers out there on single cell RNA-Seq. You usually use FICOL or gradient centrifugation, but we wanted to go take a step back and say, okay, what's best? So we did red blood cell lysis step on the marrow. We did gradient centrifugation. We also did CD34 enrichment. So we ran the cells through an immunomagnetic column enriching for CD34 with the idea that if flow was right, and we're only at 1%, hey, let's boost this up a little bit so that we can actually see some enough disease to be able to look at the genomic profile of it. And then we performed single cell capture using both 3' and 5' 10x, V3, and sequenced those six samples. So three different processing, and then two different library preps to get the data. So I want to talk about a little, I think as we begin to think about scaling these things clinically, I think we also have to think about, is this feasible? I won't get to cost, but is it feasible to actually do? We collected the sample in the afternoon on a Thursday, captured the cells later that afternoon. Library preparation we did in one day on a Friday. We lost the weekend because the core wasn't open, and we now have our own sequencer, but we didn't then. And so we started sequencing at 8am on Monday. Sequencing ran for 24 hours. There's a little delay there, as partly my fault, but the demultiplexing started in the afternoon on Tuesday, and CellRanger was done in a few hours, and I was able to look at the data on Thursday evening. So is this feasible for turnaround? Yes, I think that it is. Flow is a one to two-day turnaround time. Those molecular tests, much longer. I think this is feasible to do. And so I don't want to spend a lot of time talking about droplet partitioning, single cell RNA sequencing. I think a lot of you are familiar with it. Suffice to say that the microfluidic device is a workhorse that brings cells and specialized beads together that barcode the messenger RNA from those cells uniquely, and then we perform library preparation and were able to recover the transcripts that are associated with single cells because of those barcodes that were tagged onto the messenger RNA molecule early on. I want to just point out a fantastic team of people I have in the lab doing these experiments and who've been leading our molecular development work in this space. Sammy and Truti are great to have in the lab and they're doing a wonderful job. All right. So I'm just going to skip all over the QC. I know you've probably seen tons of single cell RNA sequencing QC. I'm not going to bore you with that. We know how to do QC. I'm assuming you do too. Let's just look at the data. So we know that this patient, this is the flow, this flow cytometry. The blasts here are high for CD34 and low for CD38, which is classic for AML. So what's the first thing we're going to look for? We're going to look for a CD34 transcript. So here's a dimension UMAP embedding of the data after filtering, cleaning the data up, removing doublets, et cetera, et cetera. And what we see is a large hook of cells over here. Now this is all of the cells, all of the conditions combined. So CD34 enriched, FICOL, and RBC lysis. And then I'm just throwing the three prime data. I'm just going to fast forward and tell you that the five prime data don't look much different than the three prime data. So we'll just get rid of that right now. So three prime data, big, ugly cluster of cells that look concerning. Now this includes the CD34 enrichment. So could we just be enriching for the leukemia? And the answer to that question appears to be yes, if these are leukemia, I'll get to that. But this hook, we see in the CD34 enriched sample, we see many cells in the gradient centrifugation sample and RBC lysis sample. And these are about 10%. So in more agreement with those NGS tests than flow cytometric measurements. But now we have a lot of cells to look at, which is good. Okay, so the next thing we really wanted to do was annotate these cell types. We know that the, we suspected leukemia shouldn't really cluster with healthy blood cells in the marrow. It should be different. It should be expressing a lot of different other genes. And so the method that we first used to look at this was Surat's RPCA integration. And what that allows us to do is, you know, the reference dataset that we're using here, which is from Greenleaf Labs publication in 2019, is a great reference dataset. I highly recommend it for marrow. We've seen really good performance with that as a reference dataset, but it used 10xv2 and we're using v3. And so we had to, you know, obviously do some integration because there's a big difference in measurements there. So after integrating the data, we see cells, and I'm showing the patient cells here in gray, that cluster with T cells and NK cells, B cells. We see myeloid cells clustering with myeloid cells. We see a group of cells. We don't know their myeloid yet, but they cluster with them at least. It's hard to see on this plot, but behind the gray line there, there's some green cells. Those are red blood cells. So these are red blood cells. And then there's this kind of ugly blob again of cells that looks very suspicious. The other thing we did to annotate our cells, now where I'm coloring the reference cells gray, and I'm annotating the patient cells using UMAP Transform. So I'm taking the UMAP data from the reference and we rigorously went through their data and we're able to pretty precisely recover their exact embedding, I would say. This is our redo of their embedding. It looks really close. And then we were able to project our data, so the patient cells onto this UMAP using UMAP Transform, and then color the cells by which cells they were most intimate with after that UMAP transformation step. And so now we're actually able to annotate the cells using UMAP Transform, and that was really helpful. And what you see is a really disordered pattern of what appears, as best we can tell using this projection method, a pattern of cells that appear to have a disordered differentiation pathway within that putatively leukemic blob. And you see cells that are most resemble stem cells going to early erythroid cells and then growing straight to monocytes. That's not how red blood cells are supposed to develop. Now this is not development, right? This is just what the best fit is, but nevertheless, it looks disordered. So and again we see the labels that we got up for our cells using UMAP Transform were intimate with the cells that we obtained with the RPCA embedding. And so okay, that's great. It actually took a little while to do that. I'd say an afternoon, maybe more a day. And we want something faster. If we're going to turn this around, I'm not sure RPCA integration, while it's good and fast enough. We want something a little bit faster. So and we like to code. And so we wanted to also kind of create our own classifier. So we created a wish list of things that we wanted to do. And so now I'm going to just kind of take a break from the patient sample and talk to you about a little bit of a computational tool that we've developed in the lab to help us annotate cells. We wanted it to work in R. We wanted to be able to classify cells with one line of code. We wanted it to be fast. We wanted to be able to edit the models or the learning algorithms that the structure uses. And we wanted it to be UMI count-based. In other words, we didn't want to have to take external data sets and do dimensionality reduction and call clusters and see how well it aligned with their calls, et cetera, et cetera, et cetera. We just want to be able to quickly download a data set from GEO and see how our cells annotate relative to that reference and quickly. So we developed a... So I was looking around and if anyone has any ideas about why I should be using a Rayfire, please tell me now. But I was looking around and I was intrigued by a C++ library called a Rayfire, which is a machine learning library that works using... And I'm saying all this, I don't really know it. So I'll just say that out loud. It's supposed to work. It does work. We've tested it. It works really well on both CPUs and GPUs. Flips between the two. There's nothing you have to do. It's really nice. And fortunately, the RCPP teams have figured out a way to easily create our packages using Rayfire libraries. And so we can plug it into those C++ libraries. And it makes it really fast. And so this... To try to think of a name for this project or this classifier, I came up with the word view master. If you remember... I'm a child of the 80s. So if you remember that old thing, used to flip through the little wheel, the little wheel of pictures and quickly just change your view. And that's what we want to do with single cell RNA sequencing data. We want to take our data and quickly flip through a reel of references and see, okay, yeah, this looks pretty good. But let's look at a different reference. Let's look at a reference from someone who was sick. Let's look at a reference from a mouse. And just see how things are similar or different. And so it works really well. So let me just show you how we've implemented it and how it benchmarks a little bit. Okay, so fortunately it came pre-baked with some code. So I was able to quickly implement the softmax regression, which is otherwise known as multinomial regression, logistic regression. And I didn't have to do much other than edit the code a little bit to feed it a serrat object or a monocle object or single cell experiment object. And so the idea here is that we feed at a reference dataset. It quickly learns how to classify cells based on what we're using as the most variable genes, as most packages use, the most variable genes, about 2,500 of those. And then learn using probabilities, the, or spitting out probabilities, learn what the appropriate cell type is for each one. And because we have a reference already, those labels already there, we can learn. And then we can apply that model, the model that we get from ViewMaster, to take ourselves here and quickly annotate them to look, to give them labels according to the reference. And it works really well. Let me just give you an overview of an experiment we did to benchmark it. So the reference dataset, we held out 80, we trained on 80%, hold out 20 test set, feed it into ViewMaster, develop a model, and then we check it on another bone marrow dataset, another separate dataset, where there's an actual cell type. And I say actual because, you know, it's just the label that the authors gave. But then we can compare the actual cell type to the predicted cell type that ViewMaster spits out. And so the two datasets that we used, again, are the green leaf reference of bone marrow and stem cells. And then you may be familiar with the SROT BM site seek dataset. I harmonized the labels so that they would work. And let me show you what it spit out. So using the green leaf data as a reference and the SROT dataset as a query, accuracy on training is high as you would expect. But the hold out dataset is pretty impressive. I don't think it's overfitted. 98% accuracy and 29 seconds to actually do this classification, which I thought was reasonably fast. And then comparing the actual cell type to the predicted cell type across the range of different cell types, I think we do pretty well. There's some confounding here of lymphoid progenitors with pre-B cells. Those are very similar cell types. But I think there's actually another reason why this is confounded. The SROT dataset has very few numbers of lymphoid progenitors and very few numbers of B cell progenitors. I don't think this dataset was CD34 enriched very much. This dataset here on the left is CD34 enriched. So it has CD34 enriched and unenriched. So I think that's hurt the performance a little bit, especially around classifying that more rare cell type. And then when we do it the opposite way, we do see really nice confusion matrix and similar speed, et cetera. So we really want to use this more. People would like to weigh in on what they think of it. We'd love to hear it. I'm sure we've got a lot of room for improvement. So that's ViewMaster. And this is being led by a very talented technician in my lab, Olivia Waltner. Okay. So I alluded to this earlier, but how can we also... So going back to that patient sample now, how can we use natural genetic variation to predict whether or to figure out whether we can find out whether a cell is donor versus recipient? So we're mining the data for SNPs, essentially. And what I want to point out is that it's really critical to pair this with cell type. So as a transplanter, when I perform a transplant on someone, I'm less worried if their T-cells don't become 100% donor. That's not uncommon for T-cell engraftment to take a really long time. And they may never achieve 100% donor T-cell engraftment. But it's when those myeloid progenitors are anything but 100% donor, that's when I'm worried that it's been a relapse of leukemia. And so that's why pairing single cell chimerism, or at least a genotype demultiplexing, with cell type identification, they think it's where the advance is and where we're going to be able to more confidently measure residual disease, if that makes sense. So there's a tool, really great tool that was written by Haynes Heaton. And he's at Auburn now and he's agreed to collaborate with us. So we're working with him and extending this package to more data sets. And so we're excited to work with Haynes. But the idea is that we measure SNPs in reads that align to the genome. And because these reads are tagged with the cell that they came from, it's easy to cluster the cells, not by gene expression, but by SNP genotype. And so this is just data from his paper where there was five samples that were mixed together and then perform PCA on the SNP genotype matrix, the probability matrix that supercell spits out. And you can see clearly five different clusters of genotypes. And he went on even further to call cells that kind of had an intermediate genotype between two clusters as doublets. And those are genotype doublets. Presumably. So we did this on our patient sample. And we can clearly identify two very different clusters of SNP genotypes. We missed some doublets, right? Because when we did our doublet discrimination using scrublet, which is what we usually use, it's looking for gene expression and modeling gene expression doublets doesn't know how to look for genotypic doublets, right? So we found a few extra. And we can color this graph now by cell type. And you can see that there's this sort of smeary cloud of cells that don't really cluster very well into a genotype here. And when we look at what cell types those are, not surprisingly, those are red blood cells. And red blood cells have RNA, but they don't have a diversity of RNA. And so therefore they wouldn't be expected to have a lot of SNPs in those RNA molecules. So it's reasonable to think that, okay, if this genetic demultiplexing gets it wrong, if it's in a red blood cell, I'm going to be a lot less worried about leukemia, right? We don't see mature red blood cells as leukemic. So we can remove them. We get really nice clusters of genotype. And now, and this was just an amazing like wow moment for us, was when we colored the cells now. So this is the cell type identification. When we color this using genotype, it's so clear that these cells are of recipient origin and that this is very likely to be leukemia. Now I'm just showing you what this looks like in the hook embedding, taking it out of integrated. So this is what the, these are the same cells that were in the hook. Okay, so we're also working to be able to detect mutations in single cell data. And we might have like, let's say we have this clinically relevant SNP, an X on one of the mutation. And during capture 10X, we use a poly T oligo binds poly A, and then we make, you know, cDNA using reverse transcriptase going this way. We might miss that G molecule if we're only sequencing 30, you know, sorry, 90 reads into the back of the molecule. So we're not going to capture that G, but if we take, you know, because we, and then we fragment the library, et cetera, et cetera. But if we take the full length cDNA that we get from 10X before fragmentation, we might be able to recover reads that have that mutation embedded in them. The other accident that we can sort of take advantage of when we look at this is the fact that poly T oligos will bind to internal A repeats within the molecule. And so you will actually get some coverage randomly in other parts of the molecule. And so we can take advantage of that. And there was a really nice algorithm called CB Sniffer that was written to do exactly this, is to mine 10X data specifically for mutations in molecules, just using the standard 10X and relying on this priming that happens internal to the molecule. So we did, we ran CB Sniffer. Again, here's the genotype calls on the left. And here I'm showing all of the calls for the TP53 mutation that we're able to identify. And you can very clearly see that the mutated TP53 cells are in the hook and wild type or not. So not too surprising, but still pretty clear, black and white. When we look at IDH1, what is really interesting is that there seems to be a lot more tumor heterogeneity with respect to IDH1 mutations. And so we see about 50% of the cells have wild type, even the leukemia have 50% have wild type IDH1. And so as I alluded to earlier, we went back to that unfragmented library. And with the help of Jason Underwood, who's a principal scientist at PACBio, we're working on using hybridization capture to pull down those molecules early on and then sequence them with long reads. And I only have a snippet of data to show you, I wish I had more. But here's our coverage of the IDH1 locus. And we see that nice 50% ratio of IDH1 mutant to IDH1 wild type. And so we're working on computational approaches to be able to pull the PACBio data and integrate it with the short read data. And so that's more to come on that front. But this looks really promising in our ability to actually recover molecules and increase coverage over clinically important areas of the genome. So how does this extend to another patient? So I've got another patient, another two-year-old. She had undifferentiated leukemia. So this is a really rare leukemia that doesn't fit under the ALL bucket or the AML bucket. And we were able, and her parents were kind enough to let us take extra marrow. And she had an unrelated cord blood transplant and looked like she had early evidence of relapse after transplant as well. So we used ViewMaster and we used Supercell. And we're clearly able to identify a large number of cells here that looked, I won't go step by step through this. It's pretty clean. These are recipient origin, and they all look like early erythroid to HSC-like cells. But in this experiment, we added SightSeq. And so for those of you who are maybe unfamiliar with SightSeq, SightSeq is using oligo-tagged antibodies to on 10x data or other platforms to measure cell surface proteome. And so what we're working on now is marrying clinical flow data to the SightSeq data that we get from the same sample. So here I have the clinical flow data on the left. We have high CD34, low CD38, et cetera. And then this is the, we've recapitulated the same looking plot using flow data, it's not flow data, sorry. Cell surface immunoproteome data. And this time we have 150 surface molecules that we've measured on these samples. So there's a lot of opportunity to discover new molecules that are associated with leukemia. And you can see very similar trends. So high for CD34, negative for CD38, high for CD34, intermediate for CD7, et cetera, 56 negative. And so this works. And I think this is another exciting area of development and how we can kind of marry clinical flow with molecular techniques in the future. So the other thing I think we can do with this is maybe we will learn new immunofenotypic markers that are associated with leukemia and potentially improve multi-parameter flow cytometry and really push that envelope as well. So I'm excited to see what happens there. This is being led by two scientists, Melinda Brionnacki, who's a medical oncologist. And Shuyen Chen is our resident, our collaborating hematopathologist. And her expertise in clinical flow has been really invaluable in this project. So where are we now? We've profiled seven samples, five I'm showing here. They show good correlation here, the number of blasts that we detect by single cell RNA sequencing on the x-axis, the number of blasts or leukemic cells we detect using flow cytometry. There's good correlation. I'm circling the samples that, where I think where flow is actually getting a little, is underestimating the burden of disease potentially. We have two patients who have PCR positivity but are flow negative. And we're working to analyze that. And we did intriguingly find one, in one of the samples, one cell of early myeloid lineage and recipient origin. So in summary, inconsistencies and clinical assays have been an important motivator for this project and I think have motivated me to improve diagnostics for leukemia. Integration of single cell data with genetic demultiplexing, and I think we'll ultimately provide a more confident assessment of relapse after transplant. I think we have promising preliminary data suggesting we can augment coverage using alternative molecular approaches on Tenex data. We have nice immunoproteomic data showing that we can recapitulate clinical flow. And we want to increase our sample numbers. We want to build our toolkit further, our computational toolkit, detect fusions, be able to mine those data, integrate it with UMAP embeddings and expand the toolkit even more. Ultimately, I haven't talked at all about this and I think, I hope that by talking about MRD, you can kind of see the potential for this to really explore mechanisms of relapse, HLA expression on the leukemia, antigen processing machinery, we can look at T cell exhaustion, myeloid suppressor cells. But we need money, so if you know anyone who has any, yeah, give it, please let me know. And we also need people. So we're hiring at SERI and I'll just put this up here. So Sean put Sean's email on there and Jay who I see sitting in the back, please contact them if you have any more information and are interested in working with us. We'd be delighted to hear from you. I'd like to acknowledge my lab, my clinical coordinators and clinical help mentors, of course, our patients and their families, funding sources, and with that I'll take a break and we can have questions. Yeah, very great presentation, really enjoyed. I was wondering it's, I mean seven samples, I guess it's still the beginning to draw patterns, but often you can have very fast algorithms, but if you need manual curation, that's the time consuming part. So if you have to spend a week to understand the, what maybe progenitor population you are dealing with, that's the pipeline. So I was wondering how much you think you can automate this process, how much you think Leukemia is such a terogeneous that you need a person for a week for each case that you have to analyze or... Yeah, it's a good question. I mean, right, we don't know because we haven't done enough samples, but it's pretty fast once we get it rolling. The ViewMaster has pretty confident assessment can differentiate a T cell and a red blood cell and a myeloid progenitor. That's really all we need. We're not asking a lot. So we see recipient chimerism in a T cell, we're going to be less worried. We see recipient chimerism in a red blood cell less worried. We see recipient chimerism in a myeloid progenitor for AML, at least. That's problematic. And so, yeah, we'll need to do more samples to see if the algorithm falls down in some cases, differentiating T cells from Leukemia. I don't think that's going to happen, but we'll only time until. I'm just going to squeeze in here. To be really accepted by this crowd, I think it needs to be ViewMastR. Okay, sure. Sure. Okay, yeah, I love it. We'll do that. Yeah, we'll change the name. Hi. Just a question about the mutation detection with IsoSeq. So like Pacfire saying that they've got much greater fidelity of detection of mutations, and there was also a whole recent Twitter thread saying how we expect to see a certain amount of missed calls within 10X just due to the actual artifacts within that. Are you seeing much difference between the mutation calls in 10X and the IsoSeq? And if so, which do you think has greater fidelity? Yeah, don't know enough to answer that question intelligently yet, but I will say that the TSO artifact that we think confounds the ability to sort of just PCR off of 10X libraries and put them on a short read sequence or we've done that experiment and put it on a Lumina, we see real problems with a TSO artifact. And so I didn't include that because it's just getting a little bit too inside baseball, but we have a TSO depletion step that we're doing to try to minimize that 10X artifact. So hopefully we're addressing it, but I think we need more data to answer that question fully. It was a great talk and then it was really great to see that our research can be applicable in the real patients. I just wonder whether you are, I mean, you are kind of playing to utilize this information from single cell assays in maybe future patient treatment or anything. Yeah, I think so. I mean, that's the goal. I think what would be my long-term vision is, and I don't know how well this would work in de novo ML or regular vanilla leukemia, but I think after transplant, mining the data for genetic variation is a huge advance and gives me a lot of confidence as a treating physician that we're seeing a true relapse. And so I think that as we think about mobilizing this clinically and what are the next steps, I think transplant patients make the most sense to sort of roll this out. And then hopefully, I mean, we're, you know, a transplant is expensive. We're already spending a ton of money on these patients. What's, you know, three to $5,000 more for a 10X one. Scott, for clarification, just in case anybody here doesn't know how ridiculously expensive it is to treat this disease, what's the median cost to treat successfully or otherwise a pediatric AML patient? Yeah, so a transplant, you know, I don't know if I'm allowed to say this, but I'll say it. So we ask for patients who come in out of the country, don't have insurance within the country. We ask for usually between half a million and a million dollars up front to do a bone marrow transplant. So it's not cheap. And so, you know, a couple of 10X runs, it's like a drop in the bucket. So we got some questions online as well. I think Sean was going to take off. Yeah. So Scott, Jared Andrews, did we? Okay. Jared Andrews asks, why not use something like single R? And he has the capital R. I mean, we did, we did look at that. And there was a reason, but I don't remember what, single R is good. We've used it in other settings. We just kind of wanted to write something. Is that all right to say? And then Ryan Thompson asks, how does this method handle cells with ambiguous expression, i.e. expression that is intermediate between two or more cell types? And then you follow it up to say, to clarify, since several methods were discussed, I'm referring to the cell type annotation method. Yeah. So we're working on that. I mean, there's a max function at the end of the multinomial classifier that just takes the maximum. And so we're toying. This isn't really going to answer the question, but we've implemented a threshold. And so maybe we're thinking about using only the cells that have a probability greater than X and saying that those are more confident cell types. But yeah, we've more work to do on those sorts of questions. And so yeah, more to come there. More questions in person? There are more online if there's no in person. Yeah, jump jump in now before you get crowded out by this pesky virtual folks. Right. Yeah. Okay. All right. Eric Quavis asks, what data is used from SNPs to cluster cell types if present allele, position, LD? Yeah. So it's a it's a there's a variant calling step. You get a BCF file and you quantify the variants and get a matrix of reference and alt and then supercell takes the ref and alt to matrix and then does its magic and spits out? Nothing. The answer. Yeah. All right. Do we have time for over? Okay. May Woods asks, do you think there should be a standard set on the number of cells sequenced for single cell approaches? Since mutations in cells circulating at low frequency will be subject to sampling variants? I don't know. I mean, I don't think we would use and I can't envision a situation where we would say, you know, oh, one cell has one mutation and we're going to, you know, alter treatment based on that. So I think we I don't think we're in danger of that because I think what we need to do when we we implement single cell genomics clinically is we need to look at multiple parameters at the same time. And so I don't think we're going to be sort of hamstrung by, you know, one cell and an aberrant measurement in that way. I don't know if that really answers the question. And maybe you might, one possible answer might be is that as long reads start to potentially become more and more certain, we know that one cell can restart leukemia. So maybe at some point it does become feasible. But right now? Yeah. You said probably not. Right right now. Yeah, we got we got a slow slow roll into this. May follows up. She says, but it's not one cell if you sample 5000 cells. Yeah, I'm not I'm not sure I get the question, but sounds like email. Yeah, email me. Yeah, my work. Yeah, let's let's talk. Okay. Thanks everyone. I appreciate continuing the glorious streak of ending on time. Everybody enjoyed lunch. And please be sure and check out the workshop this afternoon. Thanks everybody. Another big hand for Scott. It'd be so kind. And our birds of a feather coming up this afternoon.