 Welcome everyone to the afternoon session of the second day of our summer school which starts with an outstanding keynote speaker. We have Christoph Bock here. Christoph is a star in the field of bioinformatics and machine learning. He's currently full professor for biomedical informatics at the Medical University of Vienna and also at the research center of the Austrian Academy of Sciences for molecular medicine. As I said, he has done outstanding work. He has won the Tohan Medal for his PhD studies at the Max Planck Institute in Zabrücken. Then he was a postdoc at the Broad Institute, and then he became a principal investigator at the Center for Molecular Medicine in Vienna. He won the Overton Prize for Early Career Accomplishments in Computational Biology and two ESC starting grants, just to name a few of the honors that Christoph has received throughout his career. He's also a fellow of Ellis, this European lab for learning and intelligence systems. He's the expert, one of the star experts for epigenetics and how to analyze it with single cell technology. We are very happy to have you here Christoph and excited to learn about your work and the future plans and your vision for the future. Thank you so much for inviting me. Summer schools are a particular pleasure. I have spent kind of doing my PhD, the first two years I think each summer I was at a summer school that has very much shaped what I'm doing, and I learned a lot of the things like how to approach a problem from summer schools. Obviously we're missing out on the in person character this year, hopefully the next year. That's also part of the game and summers over at least in Vienna. So perhaps we're not missing out on the outdoor part too much, but certainly the discussions and I'm very happy that you've been organizing it such that we will have a chat with the students after the presentation. So as customers already alluded to, I'm a computational scientist by training and with the focus on machine learning for the longest time, but I realized early on that machine learning is strongest when it's combined with really relevant data and also with the ability to validate your, your, your predictions and your findings so the lab I'm leading here in Vienna is combines computational methods development, applications of computational technology, as well as web lab and kind of high support biology to kind of understand in order to dissect the role of cell state. I add here kind of an epigenetic because I think the epigenome is really what give us a perspective into the developmental past and future potential of cells. So today I will really focus on kind of you should describe how we use all these technologies to approach biological problem with with real medical relevance, and I will take the epigenetic perspective to it. And the epigenome being this intermediate level or layer of control on top of the genomic DNA sequence that switches on and off or keeps genes in a closed or open state, but I like to think, and sometimes the epigenome has been referred to kind of a second of the of the genome. I find that to read a reductionistic and I would rather like to think about it, especially in the context of today's talk as the state space the landscape in which cells live and develop and die. This was first kind of very intuitively visualized by Conrad Waddington already in the 1950s as a model of development with the idea that a stem cell essentially the fertilized egg would start in a state of pluripotency a state of being able to develop into pretty much anything. And then as it differentiates along these different trajectories. It would actually become ever more restricted to certain patterns. As you could think of kind of the first trajectory is here between the kind of hematopoietic and non hematopoietic lineages and then the myeloid linus lymphoid linus or further breaking up into B cells T cells etc. Now this is the power of this this paradigm is that it provides kind of and it's no longer sees cells as a snapshot in a single point in time and sometimes when we look at on a seek data, we are simplifying essentially kind of seeing cells as backs of RNA molecules as was recently suggested which is useful in some cases, but it, and there's really kind of a history and the future to each cell and that oftentimes can be captured very nicely with the I will show a few example of this epigenetic memory concept. We can also capture what happens in diseases for example in cancer a common situation is this block of differentiation that a leukemic cell just doesn't reach terminal differentiation anymore, or it goes totally off the chart and starts behaving and unexpected and and disease associated patterns. Importantly for me the epigenome is not something that should be kind of pitched against gene regulation like if you don't like the term epigenome you can think of this as a chromatase team based memory based form of gene regulation that we're talking about, but to me the best way of seeing is, is, is the epigenome and transcription regulation as essentially two sides of the same coin so if you look on top of the epigenetic landscape, you see this kind of development factor is, and here already in the original study, or in the original book that proposed these ideas was a drawing of what is kind of underneath this epigenetic landscape, essentially regulators, you can think of transcription factors, keeping this landscape in shape, and then you can already very visually get an idea what will happen if a transcription factor here is mutated in cancer. These ropes essentially regulatory mechanisms disappear, the landscape changes and perhaps accounts for itself gets stuck here and keeps proliferating but never reach terminal differentiation. Now today we are in a computational meeting so myself being a computational person is, it's hard to look at these things without wanting to overcome the metaphorical character and actually make it something that we can calculate infer and and essentially transform mathematical operations on top of this. And this has been proposed so the idea is that this epigenetic landscape is indeed a some form of quasi potential landscape in which we can see cells develop and perhaps eventually do real predictive science on this in the sense that when a cell differentiates towards a certain trajectory and it's being hit here by a an oncogene to predict whether what kind of cell states these cells enter into. This was kind of more of a pipe dream for for many years, including 2011 when this kind of attempts were an attempt was made by three while to formalize this concept, but with the recent boost of single cell technologies in particular single cell thermodynamics profiling. This is really getting in reach and to reach and I think this is a really exciting area and today, I will present some initial evidence or initial applications where we, we've combined computational methods with experimental methods to try and exploit this this concept, but it's certainly just becoming feasible as we speak and much more work needs to be done to really be able to compute on the cell state landscape and or epigenetic cell states landscape. The first part, I will talk about single cell analysis of epigenetic cell state started working on this as part of the International Human Epigenome Consortium and focusing on the hematopoietic lineage essentially on the human blood. The idea here is that the hematopoietic lineage the human blood is relatively well understood system. So we kind of knew what we what we were expecting, which is kind of makes it useful for for take computational technology development. At that time we had in the context of the blueprint project we had developed the methods single cell hygiene and visor fight sequencing that allowed us to get the enamel relation profiles from low input samples but also from single cells so essentially what we did together with colleagues in Cambridge we sorted various hematopoietic stem and progenitor cells from the blood of of healthy individuals we perform DNA methylation and when then we used a prediction framework to essentially classify each of these cells are samples into their cell identity with them. To protect stem cells the progenitor cells, the myeloid lineage and the lymphoid lineage, and essentially from the misclassification rates we constructed a network that look surprisingly similar what we were expecting from prior knowledge here. This was a proof of concept that machine learning methods train fuel can purely trained on epigenetic data in this case DNA methylation can provide insights into development for trajectories essentially into the memory of cells as they, they, they develop. Now obviously in that first study we just validated what we already knew from from prior research so we were then really excited to take this those methods to applications where we have no clue what might be happening. And one of the really kind of kind of most unresolved. applications and biology are rare genetic diseases where we often know the mutation, but a whole kind of new essentially you're having human knockouts genetic diseases where individual genes are knocked out and then we see kind of a comprehensive process of biological phenotypes that is often much more than we were expecting. In one such case here was long hand cell histocytosis, particularly interesting because it is clearly driven by a cancer associated mutation. In the RAFV 600 E that's a very well established cancer driver mutation. At the same time, you see here these lesions, these lesions are for the most part just accumulating a myeloid immune cells they never get into this trajectory of metastasis and clonal evolution, etc. So many great advice this is also now to immune disease. And so it is not quite a cancer not quite enough to be immune disease. So we really wanted to take a look at these lesions and see what's happening here. To that and we did single cell army seek and and a taxi on sorted cell populations and essentially using the concept of entropy the idea that stem cells tend to have kind of more broadly expressed genes, and kind of more open chromatin, and then essentially as the balls roll down this Whattington landscape that I showed earlier, the cells become ever more restricted in the right entity. So we applied these type of methods to our data set. And we've indeed found a structure here, we couldn't feel in further there are two types of progenitor cells and then a bifurcation of material ring and kind of more destructive disease associated cells that seem to be driving this this rear childhood disease. Now in both cases, we are essentially capturing developmental trajectory so in a way, in a way, during developmental history what I call the past or developmental past of cells. Now we can also try to use the epigenome to look into the future of cells. And for that kind of the model that I'm thinking for today's talk is really on immune regulation and new regulation is particularly useful for thinking about the epigenome, because cells need to be prepared cells are confronted with a lot of things, relatively suddenly in particular infections, and they need to cope. And our hypothesis here was that non immune cells essentially the structural cells of the body fibroblast epithelial cells endothelial cells those cells that line our organs that line our blood vessels. And that kind of as fibroblast can keep our tissue together that those cells are really kind of an underappreciated first line of defense and alarm for the immune system. So we sorted all of the cell, the cell types from 12 different organs of the mouse and we performed transcriptome and epigenome sequencing on those samples and we found indeed that there is more kind of activity in the of immune genes than you would normally expect from an immune a non immune cell. But it was particularly interesting when we started correlating what we are seeing on the epigenetic level, the chromatin accessibility with what we are seeing on the gene expression level. So we found that oftentimes in these structural cells immune genes were carrying a highly open promoter, but we're only moderately expressed so normally there is kind of a good global correlation between promoter chromatin accessibility essentially how open how active a promoter is and how highly the genius transcribed, whereas in this case, the cells tend apparently afforded kind of invested energy into keeping the chromatin of these immune gene promoters open, although they were not heavily expressed. So essentially our interpretation here kind of systems biological systems in a lot of logical perspective was that these cells keep important immune genes in an unrealized potential. So essentially, although these are commonly differentiated cells in the epigenetic landscape they seem to be keeping a certain level of altitude, and essentially to have the potential to very quickly respond and one or the other way as, for example, an infection might might come about. So our model really that needs validation of course is that structural cells use the epigenome to program themselves for future challenges. Now to test this model we infected the mice with a virus that attacks essentially all organs of the mouse and then we looked at which genes were up-regulated or down-regulated. And here in a type of enrichment curve that's kind of similar to rock curve in many ways, we saw that there was indeed high sensitivity and specificity of these genes to be out that we found kind of epigenetically prepared for a quick reaction to indeed go up in response to an infection. So the model that emerges from this is that the structural cells keep an epigenetic state that allows them to rapidly respond to all kinds of future challenges such as viral infection which we tested here but likely also wound healing physical problems such as cell inflammation, metabolic challenges, etc. So kind of to sum up this part we use the epigenome to reconstruct and model what cells have gone through in their developmental history and at the same time we can use the epigenome or cells are using the epigenome to program themselves for those things they need to be prepared for in the future. Now if this is quite a fundamental thing that cells are doing then they should likely be useful for both diagnostics and therapies and kind of the next part of the presentation I would like to connect some of these topics that we've discussed so far with kind of relevance and precision medicine. So here epigenetic biomarkers have been an interesting concept for quite a while already in particular because they fill a very useful gap between the stability of DNA genome based biomarkers and the fluctuating nature of RNA based biomarkers. So DNA methylation seems to be set primarily during early development when tissue identity is first formed and then keeps a long term memory of that. Today, DNA methylation is one of the best markers of tissue, tissue type for example is even used in forensic applications, if you find DNA at a crime scene, and the police wants to know whether this is blood from a fight, or it is just kind of hear our skin tissue from someone who walked by. The DNA sequence cannot tell us this information but the DNA methylation profile that can be read from many DNA samples that gives this information quite precisely. And also DNA methylation has emerged as well as perhaps the best biomarker of chronological age that we have at the moment. So based on the DNA methylation profile in the blood, you can predict with relatively high accuracy, plus or minus a few years, how old a person is. And indeed deviations between the biological age of a person based on the DNA methylation and the actual age appear to be predictable, I have been shown to be predictive for disease and death. If someone who is epigenetically younger than their chronological age will live longer to the degree that this has been licensed by an American insurance agency life insurance agency and this is actually being used on for essentially adjusting premiums insurance premiums. You can also see that these topics raises all kinds of ethical questions and topics that we am happy to discuss in a follow up. But let's stay on the scientific side for the time being, and how we got into this question of epigenetic biomarkers was through a collaboration with Stella, who challenged us to find a computational method that could identify cell of origin in metastatic cancers. So there are about 10% of all metastatic cancers or newly diagnosed metastatic cancers, which are called cancers of unknown primary site. So in these cancers, you are the patient presents with for example a liver metastasis, but it's not clear whether this was originally breast cancer or colon cancer or something completely different so, and these are then very difficult to treat because you don't know which protocol you need to follow. So essentially the hypothesis was here was that these metastases that can be surgically removed from the liver that this might keep an epigenetic memory of their cell of origin, from which this tumor was originally derived. And so we collected a large data set we trained and evaluated elastic net classifiers and we were indeed able to assign these these tumors based on their DNA manipulation profiles to the self types or organs they were originally derived from. And this was then taken forward by a lot of collaborators into a large scale retrospective and a smaller scale prospective validation study, where it was shown that you can this really really works well. And in those cases where you can then assign a specific molecular targeted therapy in kind of with the knowledge that this is for example metastatic breast cancer. There was even an improvement in overall survival scene although this is really preliminary data that would require further validation in a randomized controlled clinical trial. So with this idea that the cancer cells keep an epigenetic memory of the cells from which they were originally derived. We teamed up with colleagues at the terms cancer research Institute in Vienna, the team of Eleni Tomaso, and we wanted to find out in the cancer Ewing's a coma for which very little was really known about how this cancer and develops and where it takes its heterogeneity from. We wanted to use DNA manipulation essentially as a window back into time of where this cancer originally developed. The interesting thing here is we did not see what is often seen at what what colleagues in Heidelberg for example I've seen demonstrated very nicely in brain tumors we did not see kind of many nice clusters that were separated but we see what we saw based on the DNA manipulation landscape was a spectrum trajectory and so kind of no clear binary separation but but a quantitative spectrum. So what seems to be happening here and is conserved in the DNA methylation profiles is that a stem cell develops into a mesenchymal cell and during this phase that is a gradual process. It is susceptible to transformation by the WS fly one fusion protein that is really driving this, this disease. So what seems to be happening here is that the transformation into cancer cell freezes a cell in its epigenetic state of where it was in the developmental phase when it first became a tumor. So in a way with the right type of inferences methods we can use tumors as the kind of evolutionary windows into our kind of developmental past as humans. And we can also take this further into cell based bio kind of biomarkers based on liquid biopsy based on blood. So in a follow up study we essentially exploited these unique patterns as seen in in these tumors to reconstruct or to follow the disease life as unfolds in in these these patients. The concept here is when we know it's very well established from the field of liquid biopsy that tumor cells break apart and give DNA into the bloodstream. And what has been realized relatively recently is that this DNA fragments in the bloodstream from broken cancer cells that they are very much driven in their size distribution based on the chromatin structure essentially epigenome of the cells they were derived on. And we can use this was a dedicated software that is called licorice to using machine learning method methodology infer the cell of origin of these these tumors such as we can, over time, follow not only which part of the body the tumor has been derived from, but also kind of how it develops over time how it responds to chemotherapy. What other side effects of the chemotherapy has for example, harming liver cells, essentially providing us which is particularly important in children, because you cannot, you cannot really biopsy these these children very frequently because it's such an invasive essay was based on the blood sample we can really follow how the therapy develops over time. So we are always committed to making our software available. So many times it's really the application of the collaboration with clinicians clinicians that drives our research and then an interesting ideas and merge of what could be useful tools. So one of these tools we then make available for the community as kind of production quality software tools. I think you will find if you're working with DNA methylation data you will find aren't beats most useful. The law law tool is the most interesting if you have any set of region, genomic region data and you want to do a biological interpretation. So essentially you can think of Lola as a form of gene set enrichment analysis, but for genomic regions so other tools such as great use essentially met genomic regions to genes and then do genes with enrichment analysis on genes, but we found it often, at least complimentary sometimes more useful to do a gene set or like a genomic region set enrichment analysis and for that we've developed the Lola software and curated a data set that can be used as a reference method reference data set. Now I said that we are going to do some precision medicine here so and that is is a challenge because it's really important for precision medicine not only to have a single point in time, but really follow the disease as it unfolds. And the Ewing sarcoma that I presented with a focus on liquor pyroxies already provides the first approach how this could work. The other very useful, I'm almost tempted to say model disease as chronic lymphocytic leukemia. That is not a huge unmet medical need because some people, most people live actually pretty well with it and there are good drugs, including a Bruggenep which inhibits B cell receptor signaling that really allow us to handle this disease quite well. But it allows us to ask fundamental questions about this role of genetic cell states connecting the developmental past of cells with the future potential of the tumor to respond to chemotherapy and to all kinds of kinds of treatments. And this is really the angle that we've been following here and we wanted to understand what happens when patients with this type of leukemia are going on a targeted therapy where we are essentially using a drug taking out the driver pathway of the disease. So thanks to a collaboration with clinicians in Hungary, and we had a very nice data set, very few patients, seven patients, but sample systematically day after treatment, three days, seven days, 14 days, etc. This is a very systematic time series analysis. We did single cell RNA-seq and depigene profiling, and then we used Gaussian processes to really try and sort out this time series character of this data set. And kind of an interesting thing that emerged here is that although all patients followed a different trajectory or like proceeded in a different speed, all of them followed the same trajectory. So we saw a consistent biological program that unfolded in different speeds in these different patients and essentially the Gaussian processes here provided also a form of alignment between the time series trajectories of these individual patients and the way of calibrating how far individuals had already advanced. From a biological side, what we see is that NFKB binding was rapidly shutting down as a result of the inhibition of the BTK kinase, and then the lineage defining transcription factors went down. Essentially, those kind of I showed this bias early that keep the epigenetic landscape in place, they were essentially cut through the signaling defect that we are pharmacologically inducing here. And then these cells eroded their leukemiques cell identity. In this epigenetic landscape it presented earlier essentially they're going off the chart and they are no longer in the kind of normal realm, and they seem to be still kind of able enough to sense that something is wrong, and then they go into a quiescence state. So these leukemic cells that we essentially blow out of their epigenetic landscape into kind of a no man's land off of cell identity. These then go into a quiescence state they do not go into apoptosis, but they go into a quiescence state and kind of hang around the bloodstream for for quite a while longer. This also means that, while the drug is very powerful the drug is not a cure that the moment we remove the drug these quiescence cells rapidly reactivate, and essentially lead to a full blown relapse of the disease. This is a problem because some people just don't cannot tolerate this drug forever and some and some other people that this patient, the leukemic cells become resistant and resistant and eventually overcome their therapy. So this kind of on top of this initial study we combined this with a pharmacological screening technology that was developed in the laboratory of Julius Pettifogart, our institute, and we are we essentially looked before and after the start of this targeted therapy. We looked at the epigenetic landscape as well as on the drug sensitivity using a high throughput imaging workflow. And then we got to kind of initial candidates that we that might synergize with the the initial drug given to the patient so essentially what we are doing for here we want to exploit induced weaknesses in these cells that go into quiescence and into kind of a broken epigenetic state to perhaps with a second drug wipe them out entirely such that it becomes safe to stop the treatment. In the early days and before you can bring those things into a kind of a clinical study, obviously much more work needs to be done, but it motivated me to, to go back to a concept we've formulated was 10 years ago, inspired by work in HIV. In the laboratory of Thomas Langauer, where I was a PhD student. Essentially the idea that cancer therapy in the elderly is not necessarily about wiping out the cancer, but it is about modulating the evolutionary landscape of the cancer in such a way that the disease can be kept chronic for long enough. If a person dies from some other disease at the age of 80, we are not necessarily needing to achieve a cure anymore. If we can live, have patients have a good life for 20 or more years with the disease, managing the disease that might be as good, or in better than the cure achieved at the cost of the very high toxicity of a classical chemotherapy. Essentially inspired by work on HIV and targeted therapy, they are adapted drug combinations adapted based on computational methods and machine learning. We are thinking of leukemia also as some fitness landscape of the disease in which we essentially played chess with the cancer cells so we put forward a move. We try to already calculate how the tumor might respond how the tumor will eventually go get resistant, and rather than keeping the therapies at the front line until the tumor has evolved far enough to be fully resistant. We also kind of rapidly adjust the therapeutic regimens based on kind of continuous monitoring and the computational prediction like scenario prediction of what might happen if we are changing therapy at a relatively running stage. Now obviously that's much more complex to achieve in cancer than it is in HIV because the human genome is so much bigger than and has higher complexity than it has in HIV. But this is clearly something that provides a perspective how computational methods can make a contribution to kind of increasingly making cancer a chronic rather than an acute disease. So in the last 10 minutes or so I want to go a bit beyond kind of the more descriptive parts in the first two parts of today's presentations. While single cell analysis really gives us the resolution to inspect epigenetic development, looking into the past and future of cells. We ultimately want to not just model but predict and validate and causally validate through the mechanistic interventions. And that's really a challenge for people like me with a computational background and a interest in systems biological approaches because doing the mechanistic validation experiments often means that you pick a single gene, you make the mouse and you essentially start working like five years on that particular look up mouse to find out precisely what's happening here. Now this is great, but it's a huge responsibility for the supervisor or to say okay, wouldn't be a good idea to work on that particular gene because you're essentially betting a postdocs career on making the right pick here. And the same so and as a result, if you're kind of a responsible supervisor you will probably pick something that has at least a 10% success chance chance of success or perhaps better 50% or 80% chance of success. So perhaps this is one of the reasons why a lot of biology biological research is always happening on the same set of genes that are not only clearly p53 and many of the very well researched genes they are very important genes no doubt about that, but there's also a certain kind of risk of versatility of kind of trying to figure out new mechanisms around what we already know. Well, so essentially what I think could be very exciting to find ways that would allow us to do mechanistic biology at a scale of at least hundreds, perhaps thousands of genes at the same time. And we would not necessarily need to be genome-wide we can do some prioritizations but we certainly don't want to kind of narrow it down to kind of the single gene type of research to earn. I think here a combination of computational methods, health and perturbations and organic technology can really help us achieve this question of causality at scale. The first step in this direction, it's, you need to somehow kind of reduce 22,000 human genes to like the 500 or 200 genes that you can realistically work on with the type of methods I will explain later. And here we found it very useful to use kind of the power of deep learning combined with the interpretability of regulatory gene regulatory networks. And this is based on work of Nicholas Faterni in my lab who is now an assistant professor in Salzburg. And he has, he essentially kind of put forward this hypothesis that a gene regulatory network and an artificial neural network they share this kind of depth of the, the multilayer character and perhaps we could force a deep learning algorithm to be trained on a gene regulatory network such that each not in the gene regulatory network and each edge in this gene regulatory network has a biological interpretation as a, as a regulator or as a kind of regulatory mechanism. So essentially he found a way to translate these gene regulatory networks into directed acyclic graphs that can keep much of the biological interpretability of the initial gene regulatory map network, and to train them effectively with classical algorithms of deep learning curve network fitting in such a way that we really exploit the ability of deep learning algorithms to assign meaningful effects to intermediate knots in a relatively many layer architecture. And we've been using this successfully to prioritize transcription factors, based on single cell sequencing data, in terms of their regulatory potential. And essentially this based on kind of massive scale single cell sequencing data and this type of interpretable machine learning gives us an idea of which factors might be playing a role, including not only transcription factors but also signaling proteins that are difficult to incorporate with existing methodology, because they are many layers away from the genes that they actually regulate. And then we can take those forward with into CRISPR single cell sequencing screens, such as the CropSeq technology that we've developed, where we are essentially performing genome editing at scale in in in cells and then perform a sequence sequencing of the guide RNAs that induces the perturbations and the transcription response to that perturbation. So with this we can measure the effect of a thousand knockouts, knockouts of a thousand different genes in on the gene regulatory landscape in the cells. This been very successful. There's been other technologies developed that serve very similar purposes, such as the CropSeq, CRISPRSeq, MosaicSeq and it's really the combination of these technologies, as well as the growing scale of single cell sequencing that allows us to do a perturbation biology at scale and go beyond this purely descriptive inferential perspective and get actual hard data what happens when you take out the main kind of a key regulator. And this can be done in cell lines, obviously it can also be done in primary tissue, but I see a lot of value in doing this in organoids because organoids give you the particular ability of primary tissue with the same flexibility and access to materials that you would normally expect from cell lines. So along those lines we've proposed and I'm leading a pilot project within the human cell at last where we do the last scale single cell sequencing of human organoids and working with Oliver Stegler, the kind of set on these topics, we are trying to establish kind of a Rosetta Stone approach where people working in organoids can map what they are seeing to the human primary situation as well as what you're seeing the human primary situation you can map back to organoids where you can perturb and this going back and forth between in vitro and in vivo using high throughput perturbation methods such as crop seek as well as interpretable deep learning, I think this has a lot of power in terms of making mechanistic biology I throughput. Now I'm almost at the end of my presentation to the last few slides are essentially, it's a bit of a sideline it's, it's motivated by what I described earlier, we need to be able to really really scale on methods in terms of singles sequencing and what has been the big bottleneck for crop seek and crop seek and similar approaches has been the high cost of single cell sequencing. So along those lines we've kind of look very carefully at these micro fluid devices and we're disappointed to see how much very expensive reagents really goes to the bin, because you are stochastically loading these droplet devices to avoid duplicates. And we tried to figure out the way how we can get the maximum of what we pay for out of these devices so essentially what we are doing is we are heavily overloading these droplets generators not like factor two but like factor 10 or factor 100 such that in each droplet many many individual cells and how we then disentangle the whole thing is because we do a plate based pre indexing as in combinatorial indexing approaches, so that we can learn afterwards really tell the apart what came from the individual cells and we can get from a single channel of 10x, we can get in the order of 150,000 transcriptome profiles. So I don't have much time here and anyway I wanted to give a computational talk. So if you're interested in this there's a very detailed protocol available, and a lot of other labs have already tried and set this up in their own lab. And we've applied this also for kind of arid screens so if you have a lot of perturbations you can also use this first round pre indexing to essentially get back coding for free in in kind of large scale single cell army experiments. So essentially, that wraps up the story. I talked in particular in the first two parts about the epigenetic landscape on how we can look into the developmental past and flexible future of cells and it seems that the epigenome is a way by which cells program themselves for future action. And as a computational person, if cells can epigenically program themselves for future action, that really makes me excited about figuring out whether we could perhaps program cells for future action, whether we could program cells in the same way as we can program computers for doing useful things like gene therapy, all these type of things. So kind of one major motivation why we've gone into this perturbation screens and causal analysis is really that we want to engineer cells to do those things to write the epigenetic code of these cells and program them to become therapeutics to become useful weapons against cancer. So with that I would like to conclude and thank all the people who have contributed to our work as well as the funding agencies and of course you for inviting me and for for your attention. Thank you. Thank you very much, Christoph and we send you our virtual applause here through through zoom. And that was a very exciting talk. Thank you very much and now we have time for for questions. Who would like to start. Yeah, I see a raised hands to money we saw now one of the years asked now a network will ask the first question please. Well, first of all, thank you, I find this topic extremely fascinating. So I'm very interested in what you mentioned as a, basically as a takeaway for the first part about the capability of trying to program the epigenome of the cells. And I remember seeing recently some work on a group that was trying to create basically an epigenomic version of CRISPR. So instead of editing the genome editing the epigenome. And if that were a successful endeavor. What kind of experiments do you think would be most useful. Like what kind of experiments would you want to see in the first place. So would you want to try to find specific applications that I don't know for example oncogenes or for some other field. What is the application that you would be most excited for with such a net as well. Yes, these technologies for epigenicus were based epigenome editing they are actually quite good already so they say they do work. Some of them, or most of them do not provide stable enough information that it does not have a road over relatively short period of time, but also their progress is being made by making a lot of changes. And wherever this coming and so we've some projects in the lab where we try to use combination of computational and screening technology to really get at this combinatorics of making several epigenetic modifications that wants to make them complimentary and then more stable over time where there is when I talk with people from the pharmaceutical industry. The biggest excitement in epigenome editing, it seems to be with altering risk in common complex diseases, because what rare genetic diseases tend to be driven by a single cause of mutation and gene therapy provides us with a good way of handling it, whereas for complex diseases such as cardiovascular diseases such as various issues of the brain. The major factor seems to lie in a non coding space where it's about dosing a gene in just the right way and if you dose it a bit too much or a bit too little, you might have increased risk for heart attacks or stroke or other other types of diseases and targeting gene regulatory regions with gene therapy seems complex, perhaps risky, tedious, etc. And you would have to make quite a lot of changes to have a real impact there. So essentially tuning expression of major risk genes through epigenome editing. I think this could be very powerful. This is like at least five years off in the future, more like 10 years until this is really hard to get in. But since we as human species have evolved in an environment that is quite different from our current environment, one would think that with the right type of drugs, we could perhaps recolibrate our kind of evolutionary material in a way that we would just make some tweaks here that they are not kind of with the, I'm not talking here in favor of kind of design a CRISPR design and babies or anything in this regard. This would perhaps already to be too harsh, but just like modulating a few enhances here and there in the body to make us cope a bit better with overeating and with sedentary lifestyle and become a bit less fat, stay more sportive and live a happier life. I think that's the ground vision here. It could also be a little bit less controversial than CRISPR because you cannot pass the modifications through the germline or yeah they are possibly reversible so I would be excited for those applications as well. Thank you. I have further questions for Christoph. I have one Chris of when you when you presented this fascinating work on the perturbational screening with single celled technology and CRISPR. And so you describe largely as a technology so like how, how far are we in this in this field so with these different technologies that exist now, can you really get a mechanistic model of the full gene regulatory network or how do I have to imagine that so this so the state of the art in applying these methods. I think it's clear that adding these perturbational screening data on top of like classical Bayesian network influence of gene regulatory networks. There's a huge amount of additional information, because it's not purely correlational, but you have some hard perturbations that allow you to validate whether certain pathways are work kind of irrelevant or not. Today, we are seeing this with this knowledge brand neural networks with this concept of deep learning on gene regulatory networks that this is very powerful but it still stays correlational and this is just inferred and but the moment that we are merging in drop sick data. We, we provide selective confidence and this elective confidence can be used with the right mathematical tools to validate the model as a whole so you wouldn't necessarily need to knock out or perturb every single not in your network, and because if you're perturbing just 10% of the notes or something in that order, and because everything is connected it just gives you also quite some validation and confidence for for the rest of the network. So right now, kind of whether this is already full mechanistic explanation probably not, because, ultimately, and we are only taking out things here, we would probably also want to do more gain of functional experiments. And I see a lot of power in CRISPR activation screens. Essentially, to get control dose response to our regulator transcription factor to downregulate transcription factor, and then get towards the quantitative network, because that's something we are notoriously bad, except for perhaps some form of metabolic where we have good quantitative data, but in gene regulatory networks, we are so far off in the equilibrium that, except for a few very focused kind of differential equation models are very specific in regulatory processes. We do not have a good grasp for like quantitative processes and I would think that's that is a major challenge going forward where these perturbational methods need to be developed further but also have with the right tweaks will have a lot of potential. Thank you. Indeed a very fascinating topic. I have further questions. If not, we also very closely end of this this hour it was a great pleasure listening to you, Christoph we I truly enjoyed that and I speak here for the whole network. And I thank you again for coming. And for now, taking another half an hour to meet our doctoral students and to talk to them about career and research I am sure you have a lot to share. Thanks a lot. That was great. And yeah, enjoy the upcoming half an hour. Thanks again for joining us and summer school as a whole reconvene said three p.m. Central European time be continue with Andrea Garnas keynote then. See you in a bit.