 I thought we would talk a little bit on sort of what we've heard today in terms of current best practices and some perspectives. We heard from our first speaker, Wendy, that confusing nomenclature is a major challenge. Collaboration is key, and it can promote careers rather than destroy them. Mechanisms are needed for supporting young investigators, an area that we really do need to pursue. It's important also to be sure what models and assumptions we've used. And I'll ask folks, as I'm talking, to sort of jot down areas where we might be able to develop some tools or future research. Large projects stimulate technology development, and vice versa. We heard some very good examples of this in the HAPMAP, which stimulated more rapid genotyping, which stimulated more genome-wide association studies, et cetera. Data sharing is crucial. I think, as someone said, if you don't come away with that message from this workshop, we have failed miserably. And Debbie talked about the need for tissue repositories for expression, function, methylation testing, and lots of things on the horizon that we're not even aware of just yet. We also heard from Debbie that even though we're only capturing maybe a third of the variation in the genome, we're still finding things. So this works. And we can expect to find even more things, perhaps of smaller effect, but of more interesting metabolic pathways as we go forward. She also talked about the importance of translation, gave the example of V-core C1 and Warfarin, and how it's so very predictive of Warfarin dosing, and yet we still don't have good ways of getting that into practice and of validating that rapidly and being sure that it should be in practice. We heard a lot about the importance of examining observed versus expected distributions of 500,000 SNPs. David talked about this, as well as some others, and looking at QQ plots, quantile-quantile plots, and other ways of being sure that your data are not fooling you, at least in the first blush. Importance of handling the DNA of the cases and controls similarly so that you don't introduce biases. And even though those biases may only affect a few SNPs, remember you're only pulling out a few from 500,000 or more of them. This doesn't bear repeating. Oh, sure it does. Replication, replication, replication is also another underlying theme of the day. And we heard from David and Bob that the CGEMS Association data are available now if you're doing breast and prostate cancer studies for immediate replication of your findings, and those are freely available without any need for request processes or approval. You heard from Elizabeth that genotype data are not available. We talked about QC filters, and that's a genomicist term, but they're basically QC criteria. And there's a nice description of them in the replication paper that was referred to by Stephen Chanak as the first author in nature just a week or two ago. Inspecting cluster plots, we saw some very good examples of why one wants to look at cluster plots, and also the importance of knowing what sample type you're working from, whether it's buccal or a whole genome or amplified or cell line, and you need to be able to tell your lab that. We heard that CNVs, copy number variants, are very exciting. People are very interested in them. We need better detection software for them because they're not easy to pick out with current software. And we also heard that more SNPs may not be better than more people. You do come to a point where you've captured pretty much all of the information the genome has to give you, at least in terms of SNP variation, and yet you may want to have people with more variation in other exposures and variants. We heard from Laura about the importance of planning for analysis, working with trial data sets that might be similar to your platform and kind of getting geared up so that when these data come to overwhelm you, you're at least somewhat ready for them. Most of them are too large for the standard programs we've used in epidemiology such as SAS, but there are programs that are widely available, often free shareware such as Plink and others, and it's good to sort of make yourself familiar with those. It's also good to make yourself familiar with the use of genome browsers. This is something that I think we could really contribute a lot to in epidemiology by annotating, as it were, the associations that we're seeing with diseases and traits and, you know, things like height and BMI and other stuff that would be very informative down the road. We also learned that unfiltered data, that is those that the data that's kind of don't pass your QC filters because they look funny, I think Nancy called those fleas, can really give you some very interesting clues. They also could be just plain wrong, but we don't want to throw those away, and DBGAP is providing those data. Collaboration does seem to be getting easier. We heard that from both Dan and Andy, actually from several people, that as people are recognizing the importance of this and how it does not seem to destroy careers and, in fact, does help, that it does seem to be getting a little bit easier. And Bob talked about multiple models of collaborations, shared collaborations, complementary where people have different parts of the whole, and then a portion where you really kind of separate out various disease areas. We also heard that DBGAP is going to accept association data as accessions, which is their way of saying here's a fact and we can number it and label it and find it in some other way. And, in fact, that will likely be required by some of the key journals in the near future before they will publish associations. And documentation and display of protocols and things in studies like ARIDS and Framingham that Jim showed you really seem to be revolutionary for the epithelial, the Framingham people kind of coming forward and saying, this is great. We can hardly wait to use it, you know, when they've had their data sets for 60 years and have managed them in a sort of a non-informatician way, is something that we really ought to take a lesson from as epidemiologists. You served in, okay. And genomics, in doing genomics and prospective population studies, we really do need to include participant involvement, dental some very telling stories about experience in Framingham, the need for multiple consent provisions and for tracking those provisions, which can be quite cumbersome and the need for ethical oversight. Data sharing should be encouraged and planned for in advance. And the participants really do seem to want this. They want their data to be used, that's why they donated them. And then being sure we sort of, if not future proof our studies, at least future plan our studies, recognizing that there are going to be new technologies coming down the pike, we need to get past the idea that every time you change a technology you need a new consent for it, because that probably is not the case. It certainly wasn't the case in echocardiography or in X-ray or other things and it's probably not the case in this field either. And we did hear concerns about the tenure system, really it applies to the promotion and awards system and lots of other things granting and funding that really do need to evolve along with us and we need to recognize that that does take some time and so we'll need to nurture and encourage that change. Just a note of what some warm-up genotyping data sets might be, Andy mentioned the NINDS open access repository and I would remind you that all of these slides will be available on the GEI website, so don't feel you have to scribble down various URLs. The HAPMAP samples that were typed using the two platforms being used in the Genetic Association Information Network, those data are freely available, you can just click on them and download them and that's the website there. And I might just mention that ways that population studies can be used in really exploring genome-wide association is to kind of think about what the geneticists call the genetic architecture of complex traits, so how do you describe a gene, the number of loci, the frequencies, the type of loci, et cetera. We could describe the epidemiologic architecture of genetic variants, so kind of turning it on its head and looking at it from a different perspective. So what's the population prevalence of a variant that might be important, the SNP in the AQ24 region that's strongly associated with prostate cancer? What's its prevalence in different groups of different ancestral origin? Relative risk of rigorously defined incident disease can be done in population-based prospective cohort studies better than anywhere else. Consistencies of associations across subgroups defined by different characteristics, potential modifiability of associated risks are these things that are modifiable or not? Do they change with environmental exposures, et cetera, with treatment? And correlations with other traits and exposures. And is there some way, then, to use our population-based studies to say, all right, tell me everything you can about the SLC-388 variant that is associated with diabetes. Wouldn't it be neat if you could look up in a database or call up a prospective study investigator and say, tell me what is the prevalence in your population and what are the correlates of it and what is the risk and what are the modifiers of that risk, et cetera? And we might even be able to identify potential clues to gene function. If we knew that the AQ24 variant was associated with hormone levels and other things, we might then give us some clues as to what exactly that gene does. So this is an area that really is right for epidemiology to be involved in. And I would just sort of harken back to that first scan that we saw from, actually I didn't show this, but the macular degeneration scan that showed a very strong association in complement factor H. And we heard that there hasn't been a lot of work following up sequencing that. There has been some work though in population-based studies done by David Hunter and others here in the nurses, I believe it was, David, looking at risk of developing macular degeneration by this variant and modifiable risk factors. And Marta alluded to this earlier, finding that that association really is modified by the two other major risk factors for macular degeneration, obesity and current smoking in which the risk is much higher if you have that exposure. So I think overall we've heard that these really are areas that are converging and epidemiology has a lot to contribute to this. I'm very fond of showing this Gary Larson cartoon. Well, what have I always said? Sheep and cattle just don't mix when considering genomicist and epidemiologists. And I think that the days of that are passed happily and now we'll sort of move to a panel discussion of what kinds of tools and approaches we need to facilitate that. Thank you.