 Thank you very much. All right, let's want to thank both speakers. It was excellent presentations. And I'd like to open it up to a broader discussion around ways to functionalize these variants of unknown significance or credential them, or in some way, get a ranking of which ones we should care about and which ones we should care less about. So who wants to start it off? So I'm going to come back to the modeling that you've done of trying to use what you're learning to immediately model sort of going forward. So pressing you, like, so how, how different, I mean, it looked like it was a lot better than sort of existing algorithms for predicting functionality. What is it that you're able to leverage? What's the new data, the new information that's making your models now so much better than sort of the generic ones were before? Well, I don't, I mean, with these experiments, we explore so much more sequence space than I think evolution has been able to explore. So BrSA1, for example, is really not that old in comparison to a lot of other genes. So, you know, at each position, we're testing the effect of like every possible amino acid. So we get a lot more information out of that, I think, than the evolutionary conservation. So take a different example. When you look at newborn screening and when people started adding new disorders, there was a workgroup and a panel and now it's become official as part of the President's Secretary Committee on. And they came up with specific criteria and it seems like you want something it's going to be fairly common. So all the individual intellectual disability genes while they're important, very, very. So BrSA1, where there's some sort of treatment or something you could do again and where there is a reasonable chance of developing the assay. And so I think to try to come up with some list of criteria and then there gonna be all the nuances and then some group or to plug it into that and use that as a framework to me might make the most sense. Beyond those three, I'm not sure what other criteria are, but I'm sure everybody could come up with some common treatable doable. Can I just ask what is not doable? Are there particular things that are very hard? Well, of course the bigger the protein there it is with anything that has to deal with sequencing or gene construction and stuff like that. So the bigger the protein, it'll kind of scale the size. So for example, I'm terrified of BrCA2 which is 350 amino acids. Yeah, we've been developing things for membrane proteins, again just trying to see do they get to the membrane, things like that. So those are questions that we can start answering. But the things that typically make genome sequence based analysis difficult like a lot of repetitive regions and stuff like that, those are not a problem for you, is that right? They may be, for example, when we have to make these repair templates, sometimes making the homology arms to those, if it's in a really repetitive region we have trouble with that. But once we get it cloned once, we can usually get at it pretty well. And the splice site issue that you raised, is that something that you can attack at all or is it just not within the realm? One of the things I think we're planning on moving forward with is going with the CRISPR-Cas9 based editing, saturation genome editing and that gets both things, right? So if we can tie protein function also and we'll get splicing and protein all at the same time. We may not know actually what the effect is whether the depletion of that variant is due to RNA or protein, but it doesn't really matter. I mean, if it's gone, it's gone. A couple of simple questions actually for Leah. So with the CRISPR editing, are you mostly making homozygous changes or will they mostly be heterozygous? In this case, because we were looking at RNA it didn't matter, but going forward we're actually going to be knocking out one copy and then targeting the second. So you'll delete one copy entirely and then go, yeah, right, that makes perfect sense. And regarding the splicing changes, so maybe I misunderstood, but it seems as though your assay would allow you to pick up exonic splicing changes. Can you also pick up intronic? So, potentially, I mean, one class of variants I'm very interested in is deeper intronic splice gain variants, we've seen a number of these now from RNA-Seq in muscle disease patients, but they're incredibly difficult to predict in advance. We have not targeted them yet. Again, with the repair templates we could take our variable region and just spread it kind of 50 bases on both sides of the exon and that would be something that we could definitely do. Although the only problem is that when you're trying to sequence the RNA, it's not there, but it would drop out from the genome. If you're selecting for function of the protein, those variants would be depleted from the genomic DNA too, but not being able to see them in the first place would be kind of an absence of evidence kind of situation. Yeah, that's what I figured. It seems like, at least with exonic variants, you can see it directly in the CDNA, so that makes life a lot easier. Yeah, so that might be a harder thing to do. Okay, cool, thanks. All right, Scott, and then Mike. This is really not a comment on the last two talks, but a more comment on functional variation generally and what's needed in the context of translating this into the clinic. One of the biggest problems, I think, is that there's a lack of standardization of the nomenclature around molecular genetics, molecular biology, and the functional assessment of variants, and so some people will approach this from using nomenclature that's more related to the protein and some approach it from the molecular genetic side, but the problem is that when clinical geneticists are evaluating a VUS, they go to the literature and they have to read every article. And there's not an ability to bioinformatically mind the literature for what these functional effects are in any sort of organized way. So it would really be, you know, along the lines of what Callum was saying about phenotype, it would really help if we had standardization around functional nomenclature because that would allow us to much more rapidly evaluate the literature and help to translate this into clinical practice. It's a huge issue, I think, because what you find is people sometimes are calling the same gene or variant by completely different names, and people are confused and don't even know what they're talking about. So that's clearly something that would be of value. Mike was next, but since your hand rose directly from, is this directly from Mike? Yeah, so I'm gonna talk a little bit later about this, but we're actually, when you look across differences in variant calling, it's always the functional evidence that causes differences in those calls, right? So we've been working as part of an NCI contract to build exactly what you just suggested, which is different computational models for how you represent functional evidence so that we can start building those sorts of representations that can live outside the papers so that we can aggregate them and compute over them. And I'd love to talk to you more about requirements for that. Well, most of this will be directed and then for sure, Mike, after that. One thing I wanted to say about standardizing functional evidence, there is a resource called Vario, I think it's variational ontology, but really it's functional terms for standard terms to describe, it's a decrease in the protein activity or something. It's a fairly limited set of terms right now, and I think there's a lot of groups who would like to use it, but I think they are short on funding. So that's one thing to consider. The other thing that was mentioned are standards for genes and variants. I think genes, there is a standard HGNC, I think we just have to encourage use of that. For variants, that is a big problem in standardization of variants, but there is work going on NCBI to develop an allele registry, which would be a database where you can submit your description of your variant. We can normalize it, so no matter how you describe it, left-shifted, right-shifted, HDBS or chromosome coordinates, we can let you know you're all talking about the same variants, so this could help with literature mining and making sure that everybody knows when they're talking about the same thing. So an RSID for function or something like that? Yeah, so this would be more specific than an RSID before the allele. Right. Yeah. So I was thinking about our two talks together this morning in a question that Dan Rodin was asking, and it occurs to me that there should be a class of variants out there that's going to be especially hard to track down. If you look at receptor tyrosine kinases, sometimes when they're expressed at low levels, when they're activated, they trigger differentiation. Sometimes when they're expressed at high levels, same intrinsic activity of the enzyme, yet now they trigger cell proliferation. So expression differences can lead to very different phenotypes without necessarily changing the intrinsic function of the protein. I don't think this is a problem in these kinds of studies. You'll still find a lot of useful stuff if you miss certain rare things. Any advance is a good advance and that's helpful. But in terms of interpreting negative results, one might sometimes say, oh, well, this shouldn't have mattered according to this one line of evidence. Yet it does have, there's an effect, so maybe the connection is wrong, but in fact, this emergent property could be important. So it seems to me, we're going back to the focus of the meeting, which is to think about the issue of buses and functionalizing buses, that the two talks we just heard and maybe the talks yesterday are really kind of in the basic science trenches. It's really about developing fundamental methodologies to scale our ability to apply some sort of rational, functional significance to the variants that we're discovering. I think one of the things that's missing on that treacherous bridge between basic science and clinical translation is having more opportunities for people developing these technologies to be in the presence and discuss with people who are more on the clinical side about maybe how to direct the actions and the next steps in developing the technologies. And I think, and I'm gonna put that on the spot, because we had a quick meeting in the lobby and he was actually saying this meeting was maybe changing a little bit about how he thought about focusing just because he wasn't necessarily aware of it for some of the clinical parameters about the significance of the work he's doing. So I think even though what we are hearing today is really basic science, it's to me, if we're gonna make this relevant, we have to have more opportunities for the communication with the basic scientists again and the clinicians. And I'm gonna ask Doug to say just a few words if you would. I guess you should be careful about lobby conversations. Come back to bite you. No, I mean, I think it's true, right? We very much want to contribute to this problem in a real way, right? And nobody can know everything. And clearly on the basic side, like what we don't know is exactly what kind of information is gonna be most useful. And also where to point ourselves, where to point the ship. And we have an inkling, but the devil is in the details. So just hearing how you all are thinking about this problem is hugely enlightening. It's exciting for us to talk about what we can do and to think about what Howard asks questions about cost and scale and how do we actually do that and we have good ideas, but ultimately we have to scale to like the right thing. Otherwise it's not gonna be as useful. So, and the right time to do that is before we invest all the time and effort, right? So, I mean, I think it's great. And we had talked about maybe trying to develop a sort of a specific list and we've heard more about what that list would be in terms of common, treatable and doable. That was a good one. But that's absolutely critical and not, you know. So, I anyway appreciate the opportunity to be here and listen, yep. So as the functional assessments are developed, I'm struck by your comment that it's hard when you have a clinical variant to search databases, could clungen develop an arm that would be for functional genomic data relative to specific genes? I don't know if that's something clungen has thought about adding, but I think it would be very important to put this data into some context that is searchable when you need it clinically. And also by doing that, you could bring in functional genomicists and drive maybe what they're doing by what is most clinically relevant or important or at least advise in some way. Since Queen clungen is not in her seat, her throne, I'll mention this one thing briefly and then Mark can also comment. Certainly within the working groups that I'm involved in clungen, the functional data is being annotated in a semi-quantitative way and working through some wording to try to reflect that. And that's pharmacogenomics where there's extensive metabolism versus normal versus poor or in the somatic area where there's some level of function. So it's in there, but it's not being done in a way that's uniform across the working groups that I'm aware of. But maybe Mark would have a better view. Yeah, as you know, Deborah, the clungen is having all the working groups utilize the ACMG approach to annotation or to at least explicitly define the annotation approaches that they're using, which should include functional data. Now at the present time, the actual ability to access the functional studies wouldn't be available through the site, but I think as Howard pointed out, that information is certainly being considered. The references for those that are published to be posted on the site, but there'd be no reason not to consider creating some type of link through to, if there were a data repository of functional information that could be pointed to through the resource, I think that that would be something that would be quite valuable. Well, my only thought is that that's helpful if clungen is working on what mutation you happen to have found or what gene you think may be causing your patient's disease process or symptoms, but if there were a repository somehow of functional data across the genome of whatever investigator is doing, whatever work that's functional genomics, you could access it as a clinical search even if clungen isn't targeting that specific gene yet. Yeah, I think that would be something that would have to be thought about within the scope of clungen, which has been primarily focused on being a clinically relevant variant resource. And I think one of the discussion points as well how it's not to say that this isn't clinically relevant, but it's probably not clinically relevant in terms of the target audience that we're thinking of for clungen. So then the question comes up is we have the ability within clungen to aggregate other resources and we've already done that on the website. So if a resource like this existed, would it be of value to have that represented on our resource page that people could then go through clungen to actually access it? And then for those variants where the functional information is directly relevant, then you could link that. Mike, you want to jump in on this? Yeah, I don't know what this would be considered to be clinically relevant functional genomics or not, but there are two fully available resources, RegulomDB and Haploreg. Both of those are connected to ENCODE funded by ENCODE and have ENCODE resource information as well as data from other projects like common fund epigenomics, GTX and so forth. And they could either be linked to by clungen or searched by anybody that has internet access. There's no password or sign in or anything. Thanks. Cal? Just to make the obvious statement that you could potentially connect the type of basic science assays with new clinical assays if you picked universal enough assays. Obviously you need, in terms of the proximate assays that we were talking about yesterday, you need things at multiple levels of scale. But one of the levels of scale that would be potentially powerful would be cell biology actually in the clinic and have those cellular assays be represented universally in some of the basic science projects would immediately make the connections that we're talking about. Well, there's different levels of use. You know, for trying to publish in cell that you want a certain quality of data, if you're trying to choose between two therapeutic options for a patient that will be seen on Thursday, you just need a feather that will tip the scale. And that's where some of the data that, you know, I think Doug, you were kind of coming about your old data and it's not quite, you know, it's not quite this, not quite that, but we're in a situation, clinically, it's in the cancer side where we just want some direction and we don't assume that we will be directed to perfect or maybe even to grade. And, you know, I'm not even sure if it's good, but it's something. Because the goal often in the patients we're seeing are how can we keep them alive long enough to benefit from new science? It's not, how do we cure them? I mean, we'd like to do that, but how do we keep them alive in a quality state long enough so when something that will cure them comes, they're ready. And so in that case, just tipping the scale amongst these two options is enough. So the bar in some context is quite low and in others it's quite high. No, I completely agree. And for germline conditions, you can imagine exactly the same. You don't need something that's necessarily a line of sight mechanistically involved with the gene. If, with the way that we conceive of how the gene leads to the clinical phenotype, if you know that, you know, that cell state is perturbed in that individual ab initio, just simply because you have a massive repository of functional annotation for that particular state. You know, the one example that cellular electrophysiology is so powerful. One of the reasons the ECG is actually a potent tool is because, and it's representing what's actually happening at a cellular level. There are real linear correlations between a 12-liter electrocardiogram and what happens in a xenopus oocyte. If we could imagine doing that across even, let's say, 10 or 12 domains, suddenly you have exactly the ability to do what you're saying you need in somatic cells. Mike and then Peter. So sometimes at this meeting I've been hearing the tension between the importance of diagnosis, between the importance of treatment. And I'd like to bring out that functional genomics could have importance in terms of treatment. Imagine a patient in the clinic they have a skin disorder, you have a variant, you wanna know is that variant causal or not. But perhaps you could also learn from that variant, does that variant work in a lymphocyte suggesting autoimmunity or inflammatory? Or does it work in a keratinocyte or a melanocyte? And that could lead to treatment options. If you've got a exuberant lymph immune response, immunosuppressive agents could be a place to start. So that's another way that functional genomics could help, not just telling you which variants might be causal, but what the pathogenesis might be and that could suggest treatment mode. I was just gonna respond to Mike's point, which I think is excellent. Actually in one of the projects of the Undiagnosed Diseases Consortium, that's actually the focus, is because we find it not universally possible to get to causality with a single patient and a genome, can we get to therapeutic intervention based on pathway analysis? In all of this wave of optimism, I hope I'm not going to make a lot of enemies by saying this, but in my experience, the maximum path length between any variant and any treatment is four. Or by anybody at BEER who says, connect this and that with a path length of four and PubMed, I will find it. And so I think by providing that kind of a resource, that's an enormous responsibility. And I think we're pretty far away from providing reliable clinical clues in a lot of areas. And I think if one does go this way, then it's really important to develop software that will basically not just provide links in a way that say ingenuity software often does exuberantly, but also to provide the complete evidence, the provenance of statements, and also to provide contrasting ideas and because human intuition is great, but there's a high probability of misleading it in a certain direction based on some chance to link in a database with this type of software. Realism is welcome just in moderation, Mark. Yeah, I think that those points are well-taken from the perspective of going from a disease-causing variant and trying to move into therapy, but I think there's another approach that we haven't discussed much, which is to say what are the opportunities to recapitulate the PCSK9 story? In other words, identify our unusual individuals that have not a disease phenotype but have some type of a strong protective phenotype. In this case, very low LDL cholesterol levels that also seem to have low incident cardiovascular disease that is due to a genetic knockout that just occurred nationally that it doesn't appear to have any sort of other phenotypic impact negative for the patient that then immediately led to a therapeutic intervention which is an inhibitor of PCSK9 that is now available for treatment that actually took place over a relatively short period of time. Now, the strategy of identifying those resilient individuals and defining what those phenotypes would be would be problematic, but I think if we were able to define a handful of those and take advantage of large-scale sequence data combined with the ability to do rapid functional assays to say what exactly is going on that's resulting in the phenotype, that could actually result in a therapeutic approach in a relatively short order. So I want to echo what Mark said. I think that the idea of looking at people who, I'm not sure this is on topic, but you started this so I'm going to continue. Looking at people who should have disease but don't is a really appealing one. I think the problem, as you and I both know, is that it's easy to find people with diabetes. It's hard to find people who should have diabetes but don't. So the denominator has to be really, really big before you start to sort of get into finding that handful of patients with confidence. It's hard to find them and it's even harder to find them and be sure you found them. So those are the, and so another way to do it is to go at it through a genetic sort of approach and say here are alleles that we think might be protected. That was the PCSK9 approach, in fact, initially. And then go after those individuals in some kind of prospective way or in the way in which the PCSK9 story evolved. So starting with alleles that you think are protective and then ask them a question, are they protective? Or starting with patients who ought to have, the 95-year-old smoker who doesn't have lung cancer, that kind of stuff. But it's hard to find those people. But it gets back to what I think Les was talking about yesterday is that if we can find those, you know, the bowel markers that don't have, you know, 10 inferential steps, but have one or two inferential steps where you have a pretty robust association. And that was what made the PCSK9 work is that we knew that LDL cholesterol, you know, was directly relevant to incident cardiovascular disease. And so we could look at that and look for outliers that had exceedingly low LDL cholesterol levels. You could potentially in the lipid field look at individuals that have extremely high HDL cholesterol levels and see, you know, what that is. I mean, in our initial 50,000 that we looked at, we have a couple hundred individuals that have multiple HDL measurements of 250 or greater. You know, there's probably something interesting going on with those folks. Now, whether that would lead to a therapeutic intervention, given what we're now learning about HDL is a different question. But that would be the idea is that, you know, are there measurable myermarkers that are quantitative that we could search for that we are confident are associated with a disease phenotype so that you could really then say, well, this is an individual that is unlikely to develop this disease based on this particular marker. Teri and then Jose. So it sounds like we're talking about, you know, a disease-first approach or a lack of disease-first approach and a genotype-first approach. And Daniel described already, you know, many, I can't remember how many you can remind us, you know, human knockouts you've already found for genes that you're not sure what they do. Can you remind me of those hundreds or dozens? Yeah, so there's thousands. And there's work underway at the moment actually to identify more knockouts. If we're interested in homozygous knockouts, of course, the right strategy is not to go after outbreak Europeans, it's to look at consanguinous populations. And there was a paper published from a UK group which we were involved in about a month ago looking at a UK consanguinous population. They identified, from memory, just over a thousand knocked out genes in that population. We have another paper coming out, hopefully in the next couple of months, looking at another consanguinous population with similar numbers. So these, if we're interested in finding human knockouts, those are great populations to go after. And these, I think, are telling us a couple of things. The first is a lesson that we learned a long time ago, which is that you have to be extremely careful about how you annotate knockout, so-called knockout variants. And many of these, the genes find some way of escaping what looks like it should be a protein truncating variant. There's exon skipping or splicing or something else that happens. But even so, once you do get rid of the real artifacts, there clearly are a lot of genes that can tolerate a complete knockout in an individual who appears to be healthy, perhaps as some subclinical phenotypes, but definitely is present in the population. I think that strategy is gonna teach us a lot about gene function in a set of genes about which we currently have virtually no, or in some cases, absolutely no, phenotypical functional annotation. So it would seem if, you know, we talked yesterday a little bit about how do you prioritize all the genes on Doug's list or on anybody's list, you know, in terms of what's important. When we've been asking that question since the GWAS days or even before, but it seems like these would be a reasonable set. And one could use IPS. I mean, what you'd like to do, obviously, is go back to these people and phenotype the heck out of them. But in some cases, you're not able to do that. So would IPS or other kinds of, you know, basic techniques be possible? I would think they would. For sure. I mean, in both of these cohorts it's actually possible to recall individuals for genotype directed follow up. And we've said there's been a number of experiments like that that have already been done. Now obviously recall doesn't always work. Sometimes you lose the person or sometimes you can't follow up. But I think, again, these are great examples of cohorts where particularly the UK cohort was collected right from the very beginning as being a cohort of consanguineous individuals that was deliberately collected as a group that we could return to and re-phenotype when something interesting was found. IPS, though, I think is a good model. Obviously, there are certain types of experiments you can do in IPS cells that can't be done in humans. And it's nice to be able to look at exactly, particularly for validating whether or not a variant is actually loss of function. Often what you need to do is look at the expression of the transcript or the protein in whatever tissue that gene is supposed to be expressed in. And for many of those tissues, you can't buy up to your living human to get that. So having IPS cells that you can use to validate those findings is extremely useful. Jose. And then Mike. This was related to these resiliency that we are talking about. I was going to comment that about that this afternoon, but it's already on the table. And I agree with the difficulties because, for example, we're talking about lipids. It's a monogenic disorder, relatively monogenic disorder, like it's FH, right? Familiar hypercholesterolemia. We thought, I mean, we had collected about a few thousand people with FH and there were people there with very high LDL cholesterol and no disease at all. So it was a given. We were going to find with GWAS the genes that were responsible for that protection. We couldn't. Even with that clear phenotype. And then remember the paper that was published last week in Nature Biotechnology in which they took the same approach. But they did it with 600,000 people and they found 13 people that they had to have a disease, but didn't have it. The complexity comes in that the investigators were not able to recall these 13 people to investigate them further. So great finding with a very large population. Minimal finding, but maybe very informative, but we are with our hands tight because we cannot back to them. I mean, if nothing else, it gives Dan Rodin a number of how many people he needs to search. Yeah. It's 600,000, Dan. Yeah. Yeah, I think as we think about our priority list for this, another way to think about it is we're on the verge of having large numbers of people with sequence data through PMI and other efforts. And one effort will be to go into that sequence data and figure out who's at risk for diseases that we prioritize or condition. So there's three public health genomics conditions that usually get talked about. One of them is FH has been brought up, BRCA and Lynch. So that's nine genes out of the 56. I think they're the priority. One thing that was mentioned, I think Daniel mentioned it, he's, there's a number of genes that have KIA and a bunch of numbers or whatever. And we don't really have a clue what they do. And those genes, there seems to be a predilection for finding something really interesting in a gene like that. And maybe we just remember that sort of thing, but be interested in strategies for high throughput, high, at least maybe somewhat superficial strategies for trying to figure out what do they mean. And Nancy, maybe you can kick that off. Well, one of the things that I'll talk about this afternoon is this gene by medical phenome catalog that we're creating at Vanderbilt using the genetically predicted expression of genes. So you can think of a big knockdown experiment for every gene in the genome that we interrogate at least through GTECS and an upregulation experiment on every gene in the genome where we're looking at the readout as what of the medical phenome is associated with the altered expression of these genes. And you're absolutely right. Of the genes with unattractive names, the KIA and the C1 or F163, there are hundreds of genes that have the same characteristics with respect to the phenome that they attract on say reduced predicted expression of multiple congenital anomalies, intellectual disability and a lot of other bad phenotypes. There are Mendelian genes in waiting. I mean, we will find, there will be Mendelian genes, Mendelian diseases associated with these genes and we'll have a database to help people find those because we know what congenital anomalies will be associated with three knockouts of these genes because they're associated with just being two or three standard deviations below the mean for expression of these genes. So that's one high throughput way that we can get an insight into the medical phenome function of genes whose function we otherwise don't have any idea about. It's funny you called them unattractive or actually we actually had a discussion. Unattractively named. Named, yes, yes. Beautiful genes. We thought about giving them a name that had cell death somewhere in there because then everyone would work on it because so, Melissa, and then. Since this is today, I wasn't here yesterday, but today seems to be a little bit about cross species informatics. I just wanted to mention a resource that we work closely with, which is the BG resource, which is a cross species gene expression database. And as I've seen, as I'll show later, there's a lot of data in the models that we just don't have about human genes. So based on orthology, if we combine that orthologous gene expression patterns from that resource, and that's a fantastic resource with the kind of work that you're doing, we might actually get somewhere. Great, is this it? I just want to just remind everyone, I think everyone knows that the COMP project really has generated a huge resource. So one could always look up your favorite KIA gene and find the mouse ortholog. And there may be actually information about whether it's viable because a lot of the genes are now systematically getting knocked out. And then as I mentioned before, also with mutagenesis, there are a lot of alleles being generated from many genes. And in some cases, they're incidental, but if you know the nature of the screen, then you might know whether it's actually, if it's embryonic viable, because there are screens, for example, adult phenotypes. And incidental mutations come along, they may include your favorite gene. And so if there's a way you could search that, and I know that there's a large-scale screen for immune dysfunction genes, so these are largely adult viable. I think all those mutations are also curated in MGI. So if you type in the gene name, you actually can pull them out. And so that may be some way to resource some information that may already be in the databases. That's very helpful. Good, please. I think Melissa might be talking a little bit about that, but one of the things that we've done is to create a cross-species phenotype resource where, for instance, we use ontologies to map hypoplastic snout to small nose, and we've created software to actually evaluate an exome or a genome sequence by entering human phenotype terms and matching this against corresponding mouse and fish phenotype profiles. And I think we're basically just the beginning of this thanks to projects like COMP and IMPC and improvements in our ontologies that are underway. Oh, so gene editing on the snout thing sounds pretty appealing. I wonder, we just have a couple of minutes left in this session, and not to put you in the spot, but I'd be really interested, Monty, if you had any, as you've been thinking about this, because you've been kind of living this interface for a while, anything that comes to mind as you've been taking this session in. I'm sorry for not giving you a heads-up for the question. I'm always happy to have questions, particularly ectopic questions. So yeah, I really enjoy the way this conversation is going right now, because I think we're coming back to that interface between the clinic and the basic research. And so in addition to the COMP, there's also the wonderful resource given to us by the Sanger Institute, where over half the genes have been knocked out in zebrafish, and we have that phenotype information as well. And so in addition to MGI, I just put a plug in for ZFIM, which also provides this phenotypic information, and the resource that Peter and Melissa was working on really closely integrates that information with human data. So these are resources, I think that clinicians in general really don't know too much about. Yeah, as the flood of acronyms continues, it strikes me that a very easy thing to do as an outcome of this meeting would be an aggregation of all of the resources that we think might be potentially relevant to facilitate work between the bench and the bedside. And so an actual name with the acronym plus a brief description of what we think that would do that could be distributed and potentially even posted somewhere in the NHG, I think would be very useful and would be a very tangible output of the meeting. I received a text last night from one of my bioinformatics colleagues who requested that I just have one email with all the different sites that he went, that I want him to look at because I was sending him an email every time one popped up yesterday, and he said, please, before I put you on my spam list, just print one, right? All right, two last comments before we finish up. So to take that idea a little further, Mark, a list of all the resources would be nice, but a long, a medium term goal or a longer term goal might be to figure out ways of actually having those lists talk to each other. I think Dan said this yesterday and Doug said this yesterday. I think that you'd look up a variant, you might start in XAC, but it would have a link to something, with a link to something else with a link to something. Or all in one place, just sort of saying, what is the totality of the evidence? Where is the evidence? So maybe there is evidence in Zephin or not, but if there were, that would be listed. And so it's some way to integrate them as opposed to making us look up 27 different data sites. Yeah, I noticed Wendy is absent today, but that seems like a definitely, oh, I'm sure, oh, you were here, you were hiding, you were hiding. Oh, I thought you were hiding. I thought you were hiding. I thought you were hiding. That sounds like an excellent job for NCBI, I think. My comment was on the same Venn, some list and then working with each of those resources, even like an API, a standard API, that would let MDU wants to query them, send it that query and have something returned using good mappings for the different orthologs from all those resources would be great. There's some work in that direction already, but it's not comprehensive and it's not totally aligned. I would just add, I mean, a lot of the model organism databases already do have human disease orthology pages or orthologies between different model organisms. So maybe some resource that collected that all in one place would be helpful. And maybe you're already doing it. Well, maybe you need to talk just about that, yeah. Just to say continued investment in that work, because that's a large task would be critical. Melissa, and last word from the list. So, well, this is exactly what we work on is trying to aggregate all of the model data with the human data. There are other people doing similar things, but one thing, the point I wanted to make too is that, you know, one of the issues that, amongst all the data integration issues, and there are many to choose from, one of the things about disease modeling is there are many different ways of modeling diseases and associating a model with a patient disease. So you can assert that something is a model of a disease. You can look at it based on orthology. You can look at it based on it being part of the same pathway. You can look at phenotype comparisons. You can look for gene enrichment. There's a whole slew of ways that you can say that a model might be related to a patient. And we need to be aggregating all those different ways together into one resource with all the different models in order to really fill in these data gaps that we see from not having enough information on the human side. The other thing that would make it extraordinarily useful and add a lot of value, especially considering the theme of this meeting, which is how do we use functional data to modify the pathogenicity of variants, especially those of unknown significance, is obviously we want to use these data to as a conditional probability. So what we need to know for each of these attributes is what's the likelihood of observing that piece of data? Were it to be pathogenic versus were it to be benign? Because that's the probability we need to do the formal Bayesian posterior based on that piece of data. Well, thank you very much. I want to thank the two speakers and then also the whole group for the active discussion. We'll take a break and start up 1015 with Carol and the chair.