 Okay, so now we have all the suspects in one place. Aravinda. I'm going to bite. Aravinda is going to lead us off. By the way, this has really been a great day and I thought I could do without alcohol. And I think the last two talks made me do it. It's great. So tell me, I mean if you're going to do this, this is too... Will, if you can do all of this in the future, why do we need a catalog of encode elements? Meaning you can do it at a base pair level or at a few... So you're saying if you do the screening that I was suggesting, yeah. So that would be done probably in a cell line, right? That's a pretty... That's a context that isn't necessarily that relevant to lots of different tissues. I'm sorry, I didn't want quite a formal answer in that way. I recognize why we do it. I guess what I'm saying is, could you do it at a base pair resolution? I understand the value of encode. Oh, base pair. Yeah. So I'm not a Cas9 guy. But usually when you target these things, they make small indels. So if you could tile across... I mean they also have a PAM sequence that sort of... Is there some constraints to where exactly you can put it? People are figuring out ways around these constraints left and right. So I would imagine that you will be pretty much unlimited in where you apply these things relatively soon. And then, yeah, I think that you could at least make indels. It's hard to make changes, I think, at specific locations because then you have to do directed repair. Yeah, the non-homology. Right. So that's harder. But I think that there's a lot to do with it now. Wasn't the catalog a prerequisite for designing the CRISPR step? Right? It's some kind of catalog to start, wouldn't you? Well, I think that there are plenty of putative enhancers out there now. Is that right? I guess my argument would be like, there's plenty... If you look at encode, there's plenty of ways to infer whether or not something is predicted to be an enhancer. And I think capture C and others would increase that level of probability. And so you take, you know, 50 of these things that might be working and then you use those as your knock-outs. I agree, it's not necessarily totally hypothesis-less, but it is very limited in its constraint with the hypotheses, which I think NHGRI would love. Let me try to channel Aravinda, because you might have been still coming down from your seat high there when he asked this. And at least this is what I heard Aravinda ask, and I'd love to know the answer. This is fantastic, what you're proposing to do. So, you know, could you really hit every base pair? I mean, could we just forget about all of the stuff we're doing now and you're going to discover, or these technologies will discover what we need. Well, start a prep. So there's a lot of problems with what I proposed. It was more of a jumping off point than anything. So the idea would be that you want to test, say, the couple hundred genes that you know to be important or rate limiting for growth, let's say. So you come up with a system where they're all important for growth, so you have sensitivity, and then you could do this CRISPR knockout or up-regulate, testing, recruiting specific factors to specific locations and figuring out how that affects gene expression, I would argue. But again, this is like a gleam in someone's eye at this point, but I think it's a jumping off place for trying to understand how we might go about, because I think a lot of the depth of the ENCODE lines are in cell, like the Tier 1 lines, which are extremely well characterized and really set the stage for these kinds of in vitro in vivo. I mean, cell line growth assays. Yeah, I think. Ravid, did you have? I wanted to make a couple of comments on the CRISPR side, but completely irrespectively of the technical ability to go and edit at will, which is true for frame shifts and more challenging when you need homologous recombinations for editing of bases, you also have to remember that when you want to run a screen like this, you're actually going to be limited by cell numbers. So the issue with a full combinatorial higher order and not just pairs, the world does not end in pairs, is actually that, that the biggest directors in the world in sacrificing all the mice out there wouldn't solve the problem, because the cell numbers grow to numbers that are unmanageable. So that is why a more model-driven, a notation-driven and so on approach is needed. It's not really the hypothesis, it's just the limitation of the source space. So we do need a catalog. And so, yes. Okay, good. How are you? I think for the CRISPR screen, I mean, right now what have been done is really gene-based, right? Like you screen for very dramatic survival phenotype, level of death, and then you see enrichment. So I think natural extension would be to target those enhancers or regulatory element based on the assumption that single one will cause dramatic phenotype. But I agree that the combination is really the problem, because you're going to multiply it by so many fold, you will not cover the space. And also when you talk about in the same cell to up-regulate and down-regulate simultaneously, you also need this orthologous system to allow you to do that. But really, publication allows you to do that. I agree. But the key is you just, the combination of events, I don't know how you're going to cover other space. So I totally agree. My goal wasn't to, I went real fast. My goal wasn't to do all possible enhancers against all enhancers. I definitely want to use knowledge that exists to say, okay, here are 20 enhancers, putative enhancers for this gene that are implicated by both epigenetic information and also high C information and say, okay, we're going to screen those just to see if they do anything to this gene. And let's say five do, or let's say 10 do or something. And then you can imagine those 10 squared different assays, right, for maybe a couple hundred genes. So that's, I think, a limit. I didn't do the math, but I think that's a more appropriate scale, hundreds of thousands, let's say, of SGR and A-synthesis, which I think is eminently doable. Mike. Mike. But also going all the way back to things that Ross and Laurie, in particular, referred to in the area of enhancers, kind of two coupled questions. The first is, again, not my area, so I don't know what the site guide says, but how many of the putative enhancers people feel strongly are going to end up being actual enhancers? And related to that is, what is the definition of functional enhancers? So if we want to screen, right, or we want to assay even an individual one, what is the readout that we're looking for in order to know that something is functional in the enhancer space? The intention largely of my talk to say that we really don't know, we don't have a good working definition of one of enhancers, we have several assays that suggest functionality, but I think, again, you probably would have to take a reductionist approach to some extent and choose a set of regions that, by independent data sets, whether it be high C, 4C, activity correlations, what have you, we're connected to a specific set of genes and start there. And then just kind of march along those sequences to ask what the contribution of each nucleotide in that sequence is, but if you know the target gene, or you may actually be able to screen for a phenotype, you may take some intelligent design of what you'd want to look for. But you can also just do it by reporters, which wouldn't be great, but at least that would give you a stepping off point. No, I agree that there are reporter ways of doing this as well, that I think are very natural and also multiplexable, I guess. Well, I think we all got to... I want to take a shot at that answer. There was a phrase in the PI vision document about the golden age of molecular biology. I think we owed at least somebody over there on the front row. And the vocabulary coming from that, these are operational terms. A promoter is defined operationally. And I know the enhancer is defined operationally by certain characteristics it had in the reporter assays that we're doing with SV40 and so forth. And maybe we should keep that in mind and we need something, we need a short, clear Chris definitions, but we can have multiple terms. We can go with colors, we can go with all kinds of things, but that they're... that's where having lots of assays would be helpful and so we can be more precise in what we're talking about rather than just saying enhancers when somebody's a modulator or something is whatever. You know, Ross, just a quick note on that. So if you go back to Schaffner's original definition of the enhancer, the whole concept was based on the idea that all of the important logic that the brain was in the promoter and that these distal elements were just flipping the dial a little bit. And now that whole paradigm has been completely reversed. So I just want to throw that out there. So I want to just sort of follow up on, I think what both Ross and Lori pointed out that we say enhancer as sort of a catch-all definition for regulatory elements that seem to do good things to gene expression, but I think it's clear that not all enhancers are functioning the same way. And so from the point of view of trying to get at function and how things are actually working, do you have specific assays? Do you have specific ideas or readouts that you think would help us distinguish between an enhancer with a classical definition and a region with K27 acetylation that may be doing something completely different? Because I think we desperately need to get beyond histone modifications and start thinking about what these regions are actually doing. But I think there are going to be lots of points of view about how you assay that and what the appropriate readouts are. But I would love if when I leave here tomorrow I have a better clue for what this group thinks those would be, because then I think the non-coding genome is what we most need to get at and I don't see a consensus as to how we should do that. But we clearly need to go beyond just calling it a regulatory element. I think I heard a specific question. Does anybody up here know what to do to fill this in? And I'm... Maybe I can just comment. I mean, we talked a bit about this before. What you really need is a readout that's immediately downstream of the signal itself so you know it's not too far removed where all sorts of indirect effects come in. And so reporters or RNA measurements, they can often be very far downstream. So it's sort of the cause and effect problem. So we need to have those assays. So what are those assays? Well, whether it be their own positions or other histone modifications or factor assembly, things like that, if we can measure those as a cause or consequence of a specific upstream modification. But there has to be a lot of those tests done. And sometimes it's really hard to know which is immediately downstream versus two or three steps downstream. If I could just add one thing, I think it's important, I think as was mentioned before, not just to find what's downstream, but also the upstream causative transcription factor or event or something. And so the best assays are going to be ones that give us some insight into an ideal world both. And I don't think there have been that many knockouts of candidate regulatory regions. Of course, I'm pretty sure there's a large number that have been done that were never reported on account of they didn't do anything. But what we need to know, and that didn't do anything was within a, it's a context-dependent thing. And maybe there's more interest in really looking at the knockouts. We might start to see some consistent phenotypes. I still like gain-of-function assays, but there, I mean, you know, we're designing that assay to show you one thing. So there's a lot more that needs to be done. So we've been playing around a lot with this and we've tried, for example, I think factor depletion is going to be key in terms of looking at the immediate consequence of loss of a particular factor that you predict might be important for either establishing that enhancer, be it a pioneer factor or the rate-limiting factor that activates. I had done a lot of this initially with, for example, S-I-R-N-A's, convecting them into the encode cell types and asking what happens. The problem is, if it's an essential factor, what you end up doing is you're only selecting for those cells that actually will maintain sufficient levels. And so I think we have to have other systems. For example, I think maybe this is a little antiquated, but we can bring it back. Some of these degron systems that have actually become better, where you could actually have the genetic rescue right there with you as well. And so we're playing around with some of those options and we do see that if you can immediately deplete a factor that you predict has a functional consequence at that enhancer, we can see changes in chromatin states either before or after changes in gene expression. And I think things like that are going to be really important because then we can take away the chemical that actually causes thepletion of the factor and ask when that factor can bind then what happens. And so I think assays like that might start to get at those types of functions and you can even do that in combination as well. Okay, so in the interest of time, I want to make sure people have time. So Stan and then Dana and then... So one issue which kind of dominates all functional type assays is the concept of effect size. Okay, and we are dealing with biology here and biology has built into it a couple of properties like massive amplification, right? One cell to two to the 44th divisions and there you go, it has tremendous robustness buffering all kinds of other things. The most miniscule change, smaller than we can measure in expression, mapped out over that interval could be in fact biologically a huge effect size. So, and you would expect in fact that since the system itself is doing that that most of the important effects out there are in fact incredibly small effects that are amplified biologically. So how then are these systems going to be relevant for informing us about what the actual biology is when we have no idea about the sensitivity and we assume a priority that's going to be extremely low? I think that's a factor where human genetics can also be combined to inform us about some of these effects. And I think you can actually measure the consequences on chromatin structure sort of independent of a major effect on RNA, which is also suggestive that there's some change there that's worth, you know, thinking about an aggregate or what some of these sort of downstream effects would be. You know, I think that again, combining it with disparate data sources perhaps or even, for example, human genetics as I said could be really valuable in trying to, you know, combine those data sets to see what those effects might be. And certainly... Could I just make a comment about that? You can make point mutations, single base mutations and just the guts of a promoter and have a big effect on transcription. There's some... Alicia and I have various other diseases where those are really important. It's really hard to knock down an enhancer with a single base change because they're generally bigger, there are more proteins that bind and so the effects are subtle. So I think that's why long-distance elements probably have more of a complex genetics... you know, little contributions rather than big ones. That may be one reason why I sort of mentioned what John... The main point I was trying to make is that the negative predictive value is practically zero. I mean, and unless you have a well-established thing. I think what Lori was trying to get at is, you know, everything downstream starts with this cardinal event, right? And then even though we can't measure that, at least we can measure this event and assume that there's something else, you know, going on there, but I think in a lot of cases, this is a huge problem in pinning down functions. Great when you see it, but if you don't, what do you do? Maybe Paula has an answer. Well, I can say, I don't know if I have a great... I mean, I think you have an interesting hypothesis about the effect sizes which could be testable in an organism where you could look at the whole organism phenotype so just a short-generation time model would allow you to do that, right? Well, we actually... it's not a hypothesis because probably, you know, I think what we have are data from GWAS and other things that are telling us that there are a ton of variants out there that have extremely small biological effect sizes or correlations with a particular trait, which implies that they themselves have small effect sizes, okay? And also, you know, globally, by the time you get to the whole organism, so there's, I think there's a lot of indication... I thought you were arguing that there were effects that had, at the biochemical level, had minor effects that you would... you were arguing that you're going to have major effects at the organismal level. And that, I think, is something that I would... I disagree with, but... So you think that nothing... that nothing that has a small... Not nothing, but I just think it's... I mean, it's going to be that... in general, if it's going to have a really... a 2% effect on transcription, it's not going to do anything at the organismal level. The entire field of epidemiology is based on exposure times time, right? This is like all of human disease and biology is an exposure times time phenomenon. Yeah, but there are systems where you can start asking some of these, but... Sounds like a great bar discussion. So Dana, we'll get the last question and comment, and then to make sure people can get back to the hotel, please get back, continue the discussion. Dana? So I think the question should be taken a step back. Not how do we assay for functional elements, because it's clear that it's important and it's clear that we're not quite sure how to do it, but how would we make a concerted team effort to figure out how we could best assay functional elements? So taking it a step back, how would we figure out in a team way what is right? And I've been told to remind you that the bus tomorrow is at 7 a.m. So can we thank our speakers, our panelists? Thank you all. We'll see you tomorrow morning.