 Thank you all for being here today. I'm very delighted to introduce Dr. Lars Fritsche from the University of Michigan. And as you know, age-related macular degeneration is a terrible blinding disease, and you're going to hear this morning about the genetics of this from the largest case control study on the genetics of this disease. And Dr. Lars Fritsche, who earned his Ph.D. from the University of Regensburg in human genetics, and also did a postdoctoral fellowship under the direction of Dr. Bernard Weber, also at the University of Regensburg. We were lucky enough to get him here in the United States under the direction of Gonzalo Abacases. He did a second postdoctoral fellowship in biostatistics, and he has shouldered much of the burden in the first authorship for the papers that have come out of the AMD consortium. And you're going to hear about that today. And those papers have been published in Nature Genetics, and Lars has contributed vastly to the field in methodology and studying these large cohorts of over 70,000 individuals from all over the world and of different ethnicities. So please join me in a warm welcome for Lars here. Thank you. Thank you so much, ma'am. So good morning, everybody. It's a real pleasure being here and being invited to have this opportunity to present in front of you. I know you are the phenotype expert, so I'm not going to discuss that so much, but more under the genetics, because that's the field I'm more familiar with and where I'm most comfortable. So I'm going to talk about the large international AMD genomics consortium, and the talk will be first about the known genetics of H-related macular degeneration, that I'm going to talk about the work of the consortium, talk a little bit about the design, what we are planning to do or what we plan to do in this experiment, and then talk about single variant test results and gene-based test results. I'm going into details into this later on. And at the very end, I'm trying to do some kind of accounting of AMD genetics to give you an idea of where we are and where we might be going in the future. So H-related macular degeneration, as you all know, is one of the leading causes of blindness in the elderly. Americans also, the age of onset usually about 50 years of age, and the prevalence increases with age. It's a progressive disease, and estimates right now that there are about 10 million affected individuals with any form of AMD in the U.S. And it's a complex disease. So it has numerous risk factors, like smoking and diet. And one of the strongest underlying causes are the genetics, and the estimates for the heritability range between about 50 to 70 percent. And phenotypes, I'm going to talk about these are large drusen, so early forms of AMD, yellowish deposits underneath the RPE, as you all know. And then we have geographic atrophy in the area of the macular, or the second late-stage form, which is called idonovascularization. So we have an ingrow of small vessels underneath the RPE, and it can lead to bleeding and detachment of the RPE, and death of photoreceptors as well. And in my talk, I'm going to talk about those, mainly on those late-stage forms, because they severely affect the vision, and also we know that the genetics of the late-stage forms are stronger compared to the early forms, and they give us more power, more chances to discover a genetic risk variance, that's why we're focusing on them. And so it's a complex disease. We have genetics environmental and demographic risk factors, and these factors, they influence expression of genes or protein function, so there's a pathogenesis ongoing that lead then to the phenotype itself, and so understanding each of those categories will help to understand the other categories. So understanding the genetics will help to understand what's going on in the genes, are the expression differences that can be explained by genetic variance, and also how can we understand the phenotype that is developing, and better understanding the phenotype will again help us to better understand the genetics, so the purer our phenotypes are, the better we can define the genetics of those phenotypes. And there was a lot of progress and success stories in the field of genetics, and actually the first genome-wide association study that was published in 2005 was focusing on AMD, and these are the results as shown here. So there are 100,000 SNPs on the X-axis, so these were genome-wide distributed on the Y-axis, you can see the association signal, so it's the log 10 of the P-value, and this line indicates the level and the threshold for significance, and they identified on chromosome 1, which is here, variance in the complement factor H that are strongly associated with AMD, and this was the first genome-wide association study ever to identify such a strong signal. And since then, there was a lot of progress, because the complement factor H indicated the alternative pathway of the complement cascade, and people were doing candidate gene analyses, and they identified several other variants and other complement genes that are also associated with the disease, and so there are at least four other complement genes associated with AMD, and also people not only did candidate gene analyses, but also did larger genome-wide association studies, more variants that were analyzed, and larger case control studies, so they have more power to find such variants, and in the last, in the latest analysis, which is a meta-analysis of different genome-wide association studies, we increased the number by seven, and right now we're looking at about 20 risk loads that are known to carry risk variants that influence the development of AMD, and so the meta-analysis basically is an approach where we combine multiple genome-wide association studies and analyze them together to have more power to identify the signals, and the two factors that influence the chances to discover such risk factors that are mainly a little frequency, so we can have very rare variants or very common variants, and the second factor, these are the effect sizes, and the effect size in this case, these are odds-rages, so estimators for relative risk, so we have the effect of 1.1, meaning that the risk is increased by 10%, or odds-race of three would mean a three-fold increased risk, or even the effect size to 50, so 50-fold increased risk, and these dotted line, these indicate the area that can be identified in genetic association studies, so we know of rare variants that are very small effects, but we don't have enough power to find them, on the other hand we have, it's quite unlikely to have very strong genetic risk factors that are common in the population, but in between there are factors, common variants with modest risk effects, these are the common variants that are presented before they were identified in the Genoma Association Studies of AMD, they are quite common, and also we have indications that there are maybe some rare variants also associated with the disease, and we are trying to fill this gap and to identify more risk variants to better understand the genetics and the functions that cause the disease, and for example these are the results of the genome-wide, of the meta-analysis that we performed, and again you can see the chromosomes on the x-axis, colored in different shading, and in blue these were the known loci at that time, and in green these were the novel loci we identified in this approach, and we tested 2.4 million variants in a sample that consisted of over 7,000 AMD cases and over 50,000 controls, quite large study, and so we could identify those signals, and these signals are common, so the allele frequency in the populations is about 10%, which can be seen here, and the risk effects that we saw, especially of those novel loci, they were quite moderate, so the odds ratio range between 1.1 to 1.3, 10% or 30% increased risk, indicating that the earlier studies, they were too small, could not identify those modest risk signals, but now gaining this additional power we could identify them. Another thing you can see is that the two major risk factors are the CFH gene, that's the one I mentioned before, and there's a locus on chromosome 10, arms 2, HTRA1, which has a similar large risk effect, and those are the two major susceptibility loci that can explain already a large proportion of the disease, but we are also interested to identify those variants and those genes, because every gene will help us resolve this conundrum of the pathways, involved pathways, and might help us to guide future functional essays and maybe also therapies, so these were common variants, but in addition we know of low frequencies and rare variants that have quite large risk effects, and they were all published, each one of them at least, in one nature genetics paper, and the frequencies are quite rare, and one variant in the complement factor H is extremely rare in controls and quite enriched in AMD cases, and the odds ratio here was estimated to be about 20, this is based on our current study, so we have a very strong signal that almost behaves like a Mendelian disease, so we have kind of spectrum that reaches from rare variants to common variants, but we are also interested in the spectrum in between, so we know they are rare variants, we know they are common variants, so we expect there might be also some variants in between, but they are not so strong effects, so we want you to fill this gap with the variants with low frequencies but intermediate effect sizes, and for that we followed up with, we formed a new consortium and now brought together 26 groups from over 10 countries or distributed all over the world, and this star represents Salt Lake City, so glad to have you on board, and the good thing about this analysis is now, it's not a meta-analysis but it's a mega-analysis, so the genotyping itself was funded by the National Eye Institute, meaning that we could genotype all our centers in a single genotyping facility which helped us to have a centralized quality control, a harmonized data set which made the analysis much more easier and cleaner also and helped us to develop some analogies that I'm going to show later on. And the sample sizes, so I mentioned there were over 50,000 individuals that were genotyped, but I'm only focusing on the unrelated individuals with recent European ancestry, so having, focusing on one ancestry group will help to minimize population stratification effects that will influence the discovery of association signals, so that's why we're focusing on Europeans only, and we were mainly comparing advanced AMD cases, so we had about 60,000 advanced AMD cases, so some had a CND only, some had GA only, and some had both forms of late-stage AMD, and we had about 18,000 controls. Also we had some data of early forms of AMD, like 4,000 cases with large drusen, or over 2,000 cases of intermediate AMD, but as I mentioned before, we obtained the best power by focusing on advanced AMD and comparing it with controls. And this is a quite large case control study, and it's quite balanced when you look at the ratio between cases and controls, also preventing some bias in the analysis. So this is the, so we had a large sample, but now what about the genotyping? And for the genotyping, we were lucky to be the first to test a new genotyping platform that was developed in cooperation with Illumina. It's called Illumina Human Core Exumship, which is a genotyping array that allows the genotyping of genome-wide tagging snips, so similar to the analysis I've shown before, so genome-wide distributed common variants that might give us ideas about a low side that we're not aware of, but just randomly picking signals among the genome we might get lucky and find an associated variant. In addition to those genome-wide tagging snips, we had non-synonymous coding snips on the array, about 45%. And these are variants that likely influence the protein function of the splicing gene, and there's a strong negative selection because the effects of such variants are quite large, and we were interested to see if some of those variants are associated or not. And these variants were selected in large sequencing studies, and they helped to design this array. In addition, we had some custom content, so we already knew about 20 AMD risk-low side, and we were interested to get a better picture about the regions, and so we added extra variants in those regions to have a better coverage and also added some additional rare variants that we know are associated. And the good thing about those genome-wide tagging snips is that we now can do an so-called imputation. So there are maybe only 250,000 variants included, but using an imputation, you can increase this number tenfold, or even more than that by using a new approach, and you're probably aware of this. So we can use a project, so for example, the 1000 Genome Project Reference Panel, which is a project where they sequenced over thousands of individuals. And we now know the haplotypes of those individuals, the chromosomes and the different alleles on those chromosomes. And if you compare those haplotypes with our case control data, you can see that our data is much sparse also. We have some gaps in between the question marks. We don't know what's going on there, but because we can compare those genotypes and those haplotypes with the reference haplotypes, we can fill in those gaps and now increase this number of variants that we can test for an association tenfold. And in our case, we can impute over 11 million variants using this approach. So it didn't cost as much money besides computational power, and we got much more variants that we could test. When we look at the content of the array, again, focusing on the genotype variant, so this is the chip content itself, you can see that we found 160,000 of those non-synonymous coding variants that were included on the design to be polymorphic in our genotype platform. So some of them were monomorphic. We could not observe in any of our patients, participants, but 160 of them, 160,000 carried those variants. In addition, we had those genome-wide common SNPs, and they were present in our samples, and so we had 270,000 of them. When we look at the frequency spectrum of those variants, you can see what I indicated before, there's a strong selection against such non-synonymous coding variants, making them rare in our populations, and most of them have frequency below 0.5%, so there's just one heterozygous in 100 genotype individuals. In contrast, those tagging SNPs that are genome-wide distributed, they are mainly common, because that's how they were designed. So the special thing about this platform is now that we can look at those rare non-synonymous variants, and this was not possible before, and as I said, we can now also impute our data using the 1000 genome reference panel, and we could increase the number of variants that we could test by 11.5 million SNPs, and most of those SNPs are not non-synonymous coding, so they are intronic, intergenic, so we might get a nice idea about regions that were not indicated before by using those variants that are distributed genome-wide, and we could test them now in 16,000 advanced ND cases and compare the results with 18,000 controls, and this is what we get. Again, this is the so-called Manhattan blood, the chromosomes, and there's those oscillation signals, and you can see there's a gap, and so there's a different scale on top, so we obtained extremely small P-values, 10 to the power of minus 600 or 800, this is pretty crazy low, I was always told, so people are questioning this approach, but so just indicating these are very strong signals, and there's no doubt these are strong risk factors for AMD, and we could confirm most of the known AMD loci shown in blue that reach genome-wide significance that is indicating with this red line. In addition, there are six, maybe hard to read, 16 novel AMD loci that we could identify, so we almost doubled the number of AMD loci using this approach, and just a quick summary, we could confirm 18 of the 20 known AMD loci with P-values below 5 times 10 to the power of minus 8, this is the genome-wide significance threshold. Three of the previous reported AMD loci we could not confirm, so one we found is not independent of another locus, so there were two loci that were pretty close together, but when correcting for the one stronger signal we found that the other one was not significant anymore, so this was kind of a shadow effect that we identified. In addition, there was a gene on chromosome 6 that was much smaller, the signal was much weaker than reported before, and we called those kind of those effects as the winners curse, so the initial discovery study had maybe some random more power to find this signal, and now we, with more samples, we could not quite replicate this finding, but still there was some kind of peak over there, and the other locus that we could not replicate was previously described in Asian populations, and because we were focusing mainly on European samples, we could not find the signal, so this seems to be specific for Japanese population, and maybe also Chinese populations, and as I said, there are 16 novel AMD loci that we could identify, and I'm not going to talk too much about the function of those 16 novel loci because there's a lot of research still ongoing, so we don't understand those loci at this point, so we don't know which gene exactly is associated, but this is something that's still ongoing right now. When we look at all those signals, there's 34 independent association signals. We can again look at the two factors that influence the discovery of those risk variants, so the risk allele frequencies, from 0.01% to 50% to 99.99%, and this again is the effect size, the odds ratios, and you can see this is the 80% power we had, so the chances to find this, such variance is 80%, and you can see that most of our signals we identified are as expected in this area where we had more than 80% power, but another thing you can see is that we could not identify additional signals of rare variants, so we had 160,000 non-sonomers coding variants, but we could not find new signals in any other locus than the known AMD loci or the novel loci, so on a genome-wide level, we had maybe not enough power to find those variants, but you can see we need a lot of strong effects, we need strong effects to be able to identify them, but maybe we did not have enough power even when analyzing 16,000 cases and 18,000 controls. Of course, we were not interested to further characterize the signals that we identified, and so we were interested to see if in those 34 loci there are additional signals, so that we have strong signals, we're interested are there other signals underneath this strong primary signal, and we already know of such variants from previous studies, so we know multiple variants in CfH, CfI, CQ, CfB and C3, and now we could, because we now have individual level data at hand, we could do a so-called sequential conditional analysis, so we could analyze common and variants combined, we had all the data at hand and we could stepwise condition on the top signal, and I'm going to show you some examples to maybe identify secondary signals underneath this top signal. There's one negative example, so this is the locus on chromosome 2, one of the two major susceptibility loci, arms 2, HTR1, and this is the top signal of the p-value of smaller than 10 to the power of minus 700, so we were interested to maybe get an additional variant in this region that might help us understand this region, and maybe if we find a risk variant in this gene, might help us to better dissect this region, but when we conditioned on this top signal, so we removed the signal, we could see basically a flat line, so there was just one single signal, so even we had 10 to the power of minus 700 when conditioning on this signal, there was nothing left in there, so there's one single signal that can explain the whole association peak in that region, so no help in differentiating those two loci here, which was kind of disappointed, but now people can focus on those variants and maybe identify the most interesting and putatively functional variants. The other example is the complement factor I, which is shown here, again there's a strong association signal here, and you can see that there are other signals, and the color code indicates the correlation to this top variant, so these variants seem to be correlated, those variants not so much correlated, and this signal here is quite strong, and it's dark blue, meaning that the correlation is pretty low, and now by doing a conditional analysis, we could see that this is indeed an additional independent signal, in this case this was the rare variant of the CIF gene that was also identified before, but we can clearly see when removing the other signal that this is very strong, and actually by doing this conditional analysis we gained some power, so the association signal was stronger after doing the conditioning than before, and we did this for all our 34 loci and we found nine loci where we could identify 18 secondary signals, for example CIF H we could do this eight times, so we found the top snip, conditional on the top snip, and repeated this seven times and still could find signals with genome-wide significance, and for C2 CFB we could do this four times, and so on so, and some of those signals were very rare, and we could find rare signals in CIF H, C3, and CFI, and these were the rare variants that were identified before, so we kind of confirmed those signals, but we could not identify new, interesting, missense variants with low frequencies, and those multiple signals can represent independently associated snips, but also it could be an indication for haplotypes effects, so we're still always assuming that the genotype, the true, maybe causal variant, and those, but we could have actually missed these variants, and how to impute maybe, or not present on a genotyping array, so now we can follow up on those regions doing haplotype analysis to maybe identify haplotypes that explain those multiple signals, so it's an excellent dataset to do follow-up haplotype analysis, and the three variants that I mentioned before, that we could identify to be independently associated from the common known risk variants against CIF H, CFI, C3, and the signals are in our data quite strong, 10 to the power of minus 10, up to 10 to the power of minus 28, and so we had enough power to find those risk variants with genome-wide significance, and this was also not possible before, but now it was possible, but these were the only genome-wide significant single variants that we could identify, so but that's not the end, because we can increase our chance to find such rare signals by doing gene-based burden tests, so not only we can compare one single variant and the frequency between cases and controls, but we can collapse variants of a gene and compare the burden, the presence of such variants between cases and controls, and for example, we did this pooling the non-sononymous coding variants of genes and compared the frequencies or the enrichment of such variants, and we did this for 17,000 genes with a test that's called the variable threshold test, and we could indeed find genes that are genome-wide significant associated, and some of them were mentioned before and are known before CIF H, CFI, CFI, but also we found Tim 3, and there were other signals here indicated in gray, and we found that those signals are shadow effects of common variants, so after correcting for the stronger common variants, those signals disappeared, but CIF H, CFI, and Tim 3, they remained genome-wide significant, and the reason is, and we found that the underlying driver of those associations are very rare variants, so these are extremely rare. We did a variable threshold test, so we did not start with 1% allele frequency cutoff, but we let our software detect the optimal threshold to do this test, and the optimal threshold was here below 0.1%, so indicating these are very rare variants, and when we look at the burden between, when we compare the burden between cases and controls, you can see that there's enrichment in cases of those variants, and the odds ratio of those pooled variants are about 2 to 3, or in case of Tim 3, the odds races were about 30, so almost fully penetrant risk variants only found one, so we only found one of those Tim 3 variants in controls, but 29 times in cases, and the reason why we find this Tim 3 as finding quite interesting is that we enriched our genotyping platform for a special kind of mutations, and so we added known source piece of fundus dystrophy variants to our genotyping platform. We could predict also maybe other such similar variants with similar consequences, and source piece fundus dystrophy is a monogenetic macular degeneration proven down to you, and so we could find among our AMD patients some that carry such extremely highly penetrant variants that were not found in controls. So just to summarize the findings, we almost doubled the number of AMD loci on our analysis during the single variant analysis. We did a sequential conditional analysis, and we could in total identify 52 statistically significant independent variants, and in addition, using those gene-based burden tests, we found very rare variants in CFH, CFI, and Tim 3 that influenced the risk of developing AMD, and these are excellent starting points for follow-up experiments, so those variants are interesting to follow up, and as I said before, the novel loci that we identified, they are currently also followed up using expression data and doing additional fine-mapping, additional imputations to better understand what's going on there. So, for example, when pooling those, the loci that we identified, we can perform a pathway analysis to see are there certain pathways that are enriched in those AMD loci that are associated with AMD, and indeed, we can identify several pathways that show quite an enrichment. So, not surprisingly, we found that the pathway, the regulation of complement cascades, is associated. Lipid protein metabolism is associated, and also collagen structures of fibrill, the assembly of collagen fibrils and other multimeric structures seem to be associated, highly enriched in our signals, and in addition, there's a rather weaker signal, but also degradation of extracellular matrix seems to play a role in AMD. So, this is just an example of how we can follow up on the signals that we identified. So, the question is, where are we now? So, how much can we explain of the heritability of the disease? It's always an interesting question to know. Did you already explain all the data? So, is there still work to do, work left to do for us to do association studies? And one helpful estimate is the chip heritability. So, estimating the variance explained by the genetics, by the genetics. So, we could use our genotype data to estimate the chip heritability, and this estimate depends on the prevalence, we assume. And so, we assume three different prevalences of one, five, and ten percent. And so, the heritability could be between thirty and sixty percent. And we can compare this to the signals we identified. And what you can see here, these are the independently associated variance we identified. And so, there's still a large gap that we cannot explain with the signals we identified at this point. But another thing you can see is that the fifty percent of the chip heritability seems to be explainable by four AMD loci, a CFH, arms to HTR1, C2, CFBN, and C3. And those novel loci have much weaker effects and maybe are less frequent. They only contribute a little to this overall chip heritability. But as I said before, they help us identify interesting pathways that we need to explore and maybe find targets for therapy. Another interesting thing is to use this data now to estimate the risk of individuals. So, we can, using all the variants, we can come up with a generic risk score and we can classify maybe cases and controls and what we found when doing standard area under the curve analysis, a receive operator curve. We could find that our variants can distinguish cases and controls quite well. So, the area under curve is 0.8. This is much higher compared to other complex diseases, but it's maybe not high enough to do prism thematic testing. And one thing here is that, of course, the known AMD loci with the strong risk effects with the more common variants, these are the strongest contributors here. And this analysis that I'm showing here is just based on genetic risk factors. By adding additional risk factors like smoking or age, we can even increase the predictive power of such a test. But what does it mean for a population? And there's another interesting analysis we could do so we can kind of analyze the impact of the associated SNPs. And what we did, we split a simulated population. So, we simulated a population using our controls data and simulated population. The AMD prevalence of 5% into 10 equal parts. So, for example, if there are 1,000 individuals, 50 of them will develop AMD and we found that the distribution of those individuals increases with the higher risk score. But we found only 3 of the 10 deciles have a risk that is higher than the average, which is in this case 5%. Meaning that the risk itself is as expected in... So, the proportion of affected individuals is higher in the individuals that have a higher risk score. But when you look at the top risk category, you can see that only 20.7% of them will be affected by AMD. So, even being in this high risk group means not necessarily that you will develop AMD. In this case, we estimated that the chances to develop AMD will be 23%. This kind of indicates that there are limitations in using those risk scores to predict the disease. Of course, we can ask what if you look at the percentiles at the top 1%, so the 1% of the population that have the highest risk and still in this category, less than 50% will develop AMD according to this estimation. So, there are limitations in using those genetic risk scores to predict the development of AMD. So, in conclusion, I hope I showed you some interesting results from the largest centralized unotyping effort on AMD. Again, we doubled the number of AMD low side. We could identify 52 statistically independent variants that now we can use for risk analysis and also represent starting points for follow-up analysis. We think that the major common genetic risk variants have been identified, at least the common that have strong risk effects or moderate risk effects, but still there could be common variants with weaker risk effects. But based on the common variants, we might have identified most of them. But there are many more rare variants and low frequency variants that still are undiscovered and we're working on that to fill this gap, as I mentioned before. So, now having this better picture of AMD genetics now can do expression analysis, protein function essays to decipher the multiple signals we identified and hopefully we can maybe predict even the late stage forms. And one thing I did not mention is that we found one signal that seems to be specific for one form of AMD, but not the other. So, in this case, it was a gene involved only, it was only associated with wet AMD, but not with geographic atrophy. So, there might be some risk factors that can differentiate between the late stage forms and this might help us to understand how those two late stage forms develop and what are their courses. There were a lot of people involved and I'm just mentioning here the centers of the International AMD Genomics Consortium. It's probably not enough time to read all of them. So, there were 26 centers and we are now drafted a first paper and it's currently under review or I think it was submitting it and so there are a lot of authors involved and so this was quite challenging I think to come up with this author list and get to the affiliations right and there was, I would say, there was one core team that helped to do this kind of analysis and with this phenotype data, genotype data and everything was analyzed and so I'm very grateful to work in such an excellent team and I'm also thankful to the people of my colleagues in Michigan that helped me analyze the data. So, we were on a retreat, we did some cowboy games and so actually the struggle in our department is that we're all struggling to get most of our computational resources so that's what we're doing in Michigan. So, thank you very much for your attention also and I'm happy to take any questions you might have. Yeah, that's an excellent question and this is actually something, okay so the question was we analyzed advanced AMD cases together so we lumped GA and CNV and at the very end then tried to maybe make sense to differentiate between the two forms. The question is, is there a developed approach maybe or why are we doing this? So, yeah, so one thing we already knew before doing this analysis is that the genetic risk factors that are known are shared between the two late stage forms. So CFH, arms 2 and the other component genes that are associated with wet AMD as well as with dry AMD. There are some slight differences in the effect sizes but overall the risk factors we knew at that time were associated with both late stage forms so you can imagine by pooling them together we have more power to identify them because they are associated with both late stage forms and what we did, we also tried to understand if they're to estimate the overlap between genetic risk factors between those late stage forms and we estimated this overlap to be about 80% so kind of confirming our approach that it is helpful to analyze them together. But also we were able to analyze them separately. So I presented the whole genome and the givers of the advanced AMD forms but also we did the same for CNV only compared to GA only and that's how we came up with this one locus that is specific for CNV but overall what we could see is that the signals are much weaker in those stratified analyses. So, yeah, it's a power issue and we have for geographic atrophy we had I think only so compared to CNV much smaller sample size less power to identify risk for GA. So there might be still GA specific risk factors but we might not have enough power to identify them. Okay, that's an excellent question and ABCF4 is also an excellent example because it's also a monogenetic macular degeneration and when we designed our error we actually took care of this gene and so we added custom content to analyze disease variance as a disease associated variance and so we analyzed them but the mode of inheritance as you know of Stargard, or Stargard disease is recessive or compound. So, no, so, source-based fundus dystopia is an autosomal dominant inheritance I guess and Stargard disease is autosomal recessive inheritance. Oh, yeah, exactly, that's what I meant. And so the approach that we used are focusing on single variance or in the gene-based test and on pooling those variance but we did not consider the homozygosity of variance or the compound homozygosity of those variance. So we might have used the wrong model to test for those associations but when we used the single variance or the gene-based test we could not find any signal in the ABCF1 gene and this could be because we require at least two of those variance to cause the disease. There are two situations that are hard to discover in such an analysis. So we have, on one hand we had hospital-based controls so they were analyzed and free of signs of any macular degeneration but also we had population-based controls so they were recruited in population without screening for signs of AMD but this delusion of using maybe there might be some AMD patients among them but we found that it is still beneficial to use those population controls. So the delusion by using, is it delusion? Using potential AMD cases is pretty small because we expect that the prevalence of AMD is small compared to the actual controls that are added. But you're right, it would make sense to have better phenotype controls to make sure that we don't have any signs of AMD or any signs of other eye diseases in our controls and to compare them with extremely well phenotype cases and I think this is one of the next steps to get better phenotype and better characterization to do better statistical analyses. It's also a very good question and it's hard to answer so for example if you don't have any signs of AMD using these variants for prism tomatic testing you might identify some individuals that have a higher risk but actually we show that maybe not all of them will develop AMD at the very end and so what can you do? What can you recommend? You probably, hey, you can live a healthy life, exercise but you don't need genetic testing to do this, right? And if you have early signs of AMD, large bruising a lot of clinicians, they will increase the frequency of visits by seeing early signs of AMD. You don't need a genetic test to examine this patient but the genetic risk itself might have an influence on the progression of the disease so it might help you to better predict how soon will this person develop late stage AMD but then again, what can you do? How can you help this patient? So it's a tough question I personally, I don't want to get tested so I rely to the clinicians to find early signs of AMD and then react accordingly. So those companies, they use published results and so they generate a test that incorporates most of those variants that were published at that date and they come up with a risk score and they also incorporate age and maybe smoking behavior and they might give you a percentage what's your risk of developing age in the next few years, I don't know, like that but as I might have shown you is that there are multiple variants that were not considered in those tests so you might have a risk profile you get a risk profile of those companies saying you have a very low risk but you might have one of those rare TIN3 variants and that will maybe guarantee that you will develop AMD that wasn't a test so on a population based level those tests might make sense but on the individual level there are many factors that can influence the actual personal risk so I don't know so those companies, they have their goals and they have promises and some of them are helpful might help you to plan the future but I don't know exactly how exactly, yeah no, we... okay, so the question is why couldn't we find any GA-specific signals in our data so on one hand this is the sample size so what we did, we compared the variants we compared C and V versus GA and we could identify some differences so there were differences in the effect sizes between GA and C and V but those effects were observable in both late-stage forms so they were associated in both late-stage forms but there was a difference in the effect size and I could imagine that there are some GA-specific variants in the genome that can be discovered but maybe not with this sample size maybe not with this platform so we have a selection of non-synonymous variants that were tested we imputed some variants 11 million, 12 million variants that were tested it seems to be a lot but there are many more variants to be tested so the next goal I think is whole exome sequencing to get all the variants in the exomes or whole genome sequencing data but it is still quite expensive to reach the large sample sizes that are required to test those variants but yeah, we are looking for those GA-specific variants as well Thanks a lot