WEBVTT 00:00:00.000 --> 00:00:03.150 align:middle line:84% JOEL HIRSCHHORN: So the topic is basically comple-- the title is 00:00:03.150 --> 00:00:04.830 align:middle line:84% Complex Traits, What to Believe, which 00:00:04.830 --> 00:00:09.810 align:middle line:84% I think tries to get the gist of one of the main messages 00:00:09.810 --> 00:00:12.930 align:middle line:84% here and fit into the number of characters allotted 00:00:12.930 --> 00:00:15.150 align:middle line:90% for a title of a lecture. 00:00:15.150 --> 00:00:20.550 align:middle line:84% And it basically has to do with current efforts 00:00:20.550 --> 00:00:24.180 align:middle line:84% in human genetics to understand the role that 00:00:24.180 --> 00:00:27.570 align:middle line:84% inherited DNA sequence variation plays in regulating 00:00:27.570 --> 00:00:30.210 align:middle line:90% what are called complex traits. 00:00:30.210 --> 00:00:32.220 align:middle line:84% And let me just go through all that. 00:00:32.220 --> 00:00:36.807 align:middle line:84% So this is basically the plan for the talk, 00:00:36.807 --> 00:00:38.640 align:middle line:84% at least as it stands now, and just give you 00:00:38.640 --> 00:00:40.640 align:middle line:84% a little bit of introduction into complex traits 00:00:40.640 --> 00:00:44.950 align:middle line:84% and common disease and how they relate to each other. 00:00:44.950 --> 00:00:46.410 align:middle line:90% So here's a list of-- 00:00:46.410 --> 00:00:48.870 align:middle line:84% a somewhat random list of diseases 00:00:48.870 --> 00:00:53.730 align:middle line:84% that you might see in a hospital or in a doctor's 00:00:53.730 --> 00:00:57.840 align:middle line:84% office and account for a lot of morbidity and mortality. 00:00:57.840 --> 00:01:01.290 align:middle line:84% And it turns out that most of these diseases cluster 00:01:01.290 --> 00:01:04.800 align:middle line:84% in families in ways that tell us that genetics plays 00:01:04.800 --> 00:01:08.010 align:middle line:84% a role in regulating who will and will not 00:01:08.010 --> 00:01:09.420 align:middle line:90% come down with these diseases. 00:01:09.420 --> 00:01:10.868 align:middle line:84% Same thing is true for things that 00:01:10.868 --> 00:01:12.660 align:middle line:84% aren't diseases but are quantitative traits 00:01:12.660 --> 00:01:13.830 align:middle line:90% that you can measure. 00:01:13.830 --> 00:01:17.012 align:middle line:84% Some of these, however, are very significant risk factors. 00:01:17.012 --> 00:01:19.470 align:middle line:84% I list here just things that are relevant to endocrinology, 00:01:19.470 --> 00:01:20.678 align:middle line:90% which is one of my interests. 00:01:20.678 --> 00:01:22.470 align:middle line:84% But things like blood pressure, obviously, 00:01:22.470 --> 00:01:23.610 align:middle line:84% if you have high blood pressure, that's 00:01:23.610 --> 00:01:25.902 align:middle line:84% a huge risk factor for heart disease, stroke, et cetera 00:01:25.902 --> 00:01:26.820 align:middle line:90% et cetera. 00:01:26.820 --> 00:01:29.820 align:middle line:84% Increased obesity is a huge risk factor for death and diabetes 00:01:29.820 --> 00:01:32.280 align:middle line:84% and heart disease, and many of these other things 00:01:32.280 --> 00:01:34.290 align:middle line:84% also have to do with future diseases. 00:01:34.290 --> 00:01:35.790 align:middle line:84% And these also cluster in families 00:01:35.790 --> 00:01:39.610 align:middle line:84% in ways that tell us that genetics plays a role. 00:01:39.610 --> 00:01:41.100 align:middle line:84% But for all of these, whether it's 00:01:41.100 --> 00:01:43.600 align:middle line:84% disease susceptibility or the quantitative trait illustrated 00:01:43.600 --> 00:01:47.130 align:middle line:84% here, adult height, we know also from the way 00:01:47.130 --> 00:01:49.980 align:middle line:84% that they cluster in families that multiple genes are 00:01:49.980 --> 00:01:53.490 align:middle line:84% involved and furthermore, that environmental and other 00:01:53.490 --> 00:01:57.400 align:middle line:84% non-genetic influences must play a role as well. 00:01:57.400 --> 00:02:00.480 align:middle line:84% So this is not quite so simple a picture as, for example, 00:02:00.480 --> 00:02:01.980 align:middle line:84% trying to tract down the gene that's 00:02:01.980 --> 00:02:06.920 align:middle line:84% responsible for cystic fibrosis or for Tay-Sachs disease. 00:02:06.920 --> 00:02:10.263 align:middle line:84% And so the goal that a lot of people in genetics have 00:02:10.263 --> 00:02:11.680 align:middle line:84% is to try to make this connection. 00:02:11.680 --> 00:02:15.670 align:middle line:84% There's variation in the human genome, which I'll 00:02:15.670 --> 00:02:17.570 align:middle line:90% talk quite extensively about. 00:02:17.570 --> 00:02:20.604 align:middle line:84% And there's variation that we observe out in the world. 00:02:20.604 --> 00:02:22.333 align:middle line:84% And we know there's a connection here, 00:02:22.333 --> 00:02:24.250 align:middle line:84% and it's a question of making that connection. 00:02:24.250 --> 00:02:26.860 align:middle line:90% 00:02:26.860 --> 00:02:29.490 align:middle line:84% And if you compare genomes as, again, I'll 00:02:29.490 --> 00:02:33.270 align:middle line:84% get to in a little bit, find that people are mostly 00:02:33.270 --> 00:02:33.780 align:middle line:90% the same. 00:02:33.780 --> 00:02:37.680 align:middle line:84% But there are differences between every person. 00:02:37.680 --> 00:02:39.180 align:middle line:84% And again, most of these differences 00:02:39.180 --> 00:02:42.210 align:middle line:84% are random spelling differences in the genetic code 00:02:42.210 --> 00:02:43.380 align:middle line:90% that have no consequence. 00:02:43.380 --> 00:02:45.690 align:middle line:84% But some of them do have biological consequences. 00:02:45.690 --> 00:02:47.232 align:middle line:84% Again, the challenge is to figure out 00:02:47.232 --> 00:02:49.140 align:middle line:84% which one of these differences matter. 00:02:49.140 --> 00:02:51.450 align:middle line:84% So I said that we know that genetics plays a role, 00:02:51.450 --> 00:02:52.950 align:middle line:84% and I just want to briefly get into, 00:02:52.950 --> 00:02:54.750 align:middle line:90% how do we know that, actually? 00:02:54.750 --> 00:03:00.330 align:middle line:84% So one way is from something called twin studies. 00:03:00.330 --> 00:03:01.338 align:middle line:90% I forgot the little-- 00:03:01.338 --> 00:03:02.880 align:middle line:84% this is from some country music group 00:03:02.880 --> 00:03:04.290 align:middle line:84% that I have no affiliation with and have never 00:03:04.290 --> 00:03:05.665 align:middle line:84% heard called the Taylor Brothers, 00:03:05.665 --> 00:03:08.940 align:middle line:84% but notable for part of their promotion 00:03:08.940 --> 00:03:11.280 align:middle line:90% is that they're identical twins. 00:03:11.280 --> 00:03:14.650 align:middle line:84% And you can see things like they look the same, 00:03:14.650 --> 00:03:15.900 align:middle line:90% they're about the same height. 00:03:15.900 --> 00:03:17.070 align:middle line:84% I'm not sure whether you can tell 00:03:17.070 --> 00:03:19.020 align:middle line:84% whether they have the same preference in clothing or not 00:03:19.020 --> 00:03:19.520 align:middle line:90% from this. 00:03:19.520 --> 00:03:22.100 align:middle line:84% But they may have some of the same aptitude. 00:03:22.100 --> 00:03:24.780 align:middle line:84% So and what you can do is can do twin studies 00:03:24.780 --> 00:03:27.510 align:middle line:84% where you compare identical twins to twins 00:03:27.510 --> 00:03:31.170 align:middle line:84% who are what are called dizygotic twins, basically 00:03:31.170 --> 00:03:33.245 align:middle line:84% where also commonly called fraternal twins. 00:03:33.245 --> 00:03:34.620 align:middle line:84% If you compare same sex fraternal 00:03:34.620 --> 00:03:37.980 align:middle line:84% twins and identical twins, they're, in theory, 00:03:37.980 --> 00:03:40.600 align:middle line:84% raised in about the same amount of shared environment. 00:03:40.600 --> 00:03:42.060 align:middle line:84% So they're both raised in the same household, 00:03:42.060 --> 00:03:42.935 align:middle line:90% and that's the thing. 00:03:42.935 --> 00:03:44.790 align:middle line:84% But identical twins, or monozygotic twins, 00:03:44.790 --> 00:03:48.240 align:middle line:84% share 100% of their DNA, whereas dizygotic twins share only 50% 00:03:48.240 --> 00:03:49.710 align:middle line:90% of their DNA. 00:03:49.710 --> 00:03:51.900 align:middle line:90% They share the same as siblings. 00:03:51.900 --> 00:03:56.010 align:middle line:84% And so if there are things that are more concordant or more 00:03:56.010 --> 00:03:59.040 align:middle line:84% similar in monozygotic twins than in dizygotic twins, 00:03:59.040 --> 00:04:01.650 align:middle line:84% the usual conclusion is that genetics plays 00:04:01.650 --> 00:04:03.990 align:middle line:90% a role in that similarity. 00:04:03.990 --> 00:04:06.210 align:middle line:84% So for example, if you look at type 2 diabetes, 00:04:06.210 --> 00:04:08.820 align:middle line:84% if you look at monozygotic twins and follow them 00:04:08.820 --> 00:04:10.800 align:middle line:84% over their lifetime, there's at least an 80%-- 00:04:10.800 --> 00:04:13.140 align:middle line:84% and some people thinks it's actually close to 100%-- 00:04:13.140 --> 00:04:14.520 align:middle line:90% concordance rate. 00:04:14.520 --> 00:04:16.440 align:middle line:84% What that means is actually sort of striking, 00:04:16.440 --> 00:04:22.320 align:middle line:84% which means that there is a particular set of genotypes, 00:04:22.320 --> 00:04:25.650 align:middle line:84% which is incredibly predictive of getting diabetes. 00:04:25.650 --> 00:04:26.430 align:middle line:90% Yes, thank you. 00:04:26.430 --> 00:04:26.930 align:middle line:90% Sorry. 00:04:26.930 --> 00:04:28.245 align:middle line:90% I found a-- I found a-- 00:04:28.245 --> 00:04:29.370 align:middle line:90% I found a little cable. 00:04:29.370 --> 00:04:32.070 align:middle line:84% So one right one and one wrong one. 00:04:32.070 --> 00:04:33.390 align:middle line:90% So I'm all right. 00:04:33.390 --> 00:04:36.450 align:middle line:90% 00:04:36.450 --> 00:04:38.122 align:middle line:84% Whereas if you look at dizygotic twins, 00:04:38.122 --> 00:04:39.330 align:middle line:90% there's a lot of concordance. 00:04:39.330 --> 00:04:40.590 align:middle line:84% And part of that is because they share 00:04:40.590 --> 00:04:41.840 align:middle line:90% 50% of their genetic material. 00:04:41.840 --> 00:04:43.423 align:middle line:84% And part of it's probably because they 00:04:43.423 --> 00:04:45.150 align:middle line:84% share the environment that they shared. 00:04:45.150 --> 00:04:47.130 align:middle line:84% But the concordance rate is much, much higher 00:04:47.130 --> 00:04:48.280 align:middle line:90% for monozygotic twins. 00:04:48.280 --> 00:04:52.650 align:middle line:84% So that says that there's a strong genetic influence. 00:04:52.650 --> 00:04:54.690 align:middle line:84% And you can do other studies that 00:04:54.690 --> 00:04:55.980 align:middle line:90% don't depend on twin studies. 00:04:55.980 --> 00:04:57.240 align:middle line:84% And you can also do twin studies where 00:04:57.240 --> 00:04:58.948 align:middle line:84% there are monozygotic twins reared apart, 00:04:58.948 --> 00:05:00.460 align:middle line:90% but those are much less common. 00:05:00.460 --> 00:05:04.260 align:middle line:84% But basically family studies where if somebody has diabetes, 00:05:04.260 --> 00:05:07.530 align:middle line:84% what's the risk to their sibling for getting diabetes 00:05:07.530 --> 00:05:09.810 align:middle line:84% as opposed to the risk of somebody next door, 00:05:09.810 --> 00:05:10.350 align:middle line:90% for example? 00:05:10.350 --> 00:05:13.350 align:middle line:84% And again, it's much greater for siblings 00:05:13.350 --> 00:05:16.560 align:middle line:84% than it is for the general population. 00:05:16.560 --> 00:05:18.960 align:middle line:84% That ratio is something called lambda S. 00:05:18.960 --> 00:05:21.600 align:middle line:84% And you can estimate also something called heritability, 00:05:21.600 --> 00:05:26.220 align:middle line:84% which purports to be the percentage of variation that's 00:05:26.220 --> 00:05:29.010 align:middle line:84% attributable to genetic factors and actually 00:05:29.010 --> 00:05:31.870 align:middle line:84% certain types of genetic factors. 00:05:31.870 --> 00:05:33.370 align:middle line:84% So that's basically an introduction. 00:05:33.370 --> 00:05:37.200 align:middle line:84% So most common diseases are these complex genetic traits 00:05:37.200 --> 00:05:38.680 align:middle line:90% where we know genes play a role. 00:05:38.680 --> 00:05:40.513 align:middle line:84% And we also know, from those family studies, 00:05:40.513 --> 00:05:42.990 align:middle line:84% they don't just segregate in a simple pattern 00:05:42.990 --> 00:05:44.340 align:middle line:90% where if the father-- 00:05:44.340 --> 00:05:46.740 align:middle line:84% if the father has it, half of the children will have it, 00:05:46.740 --> 00:05:49.073 align:middle line:84% and then half of their children will have it, et cetera, 00:05:49.073 --> 00:05:50.760 align:middle line:84% and/or in a simple recessive pattern. 00:05:50.760 --> 00:05:54.180 align:middle line:90% But it's clearly multifactorial. 00:05:54.180 --> 00:05:56.492 align:middle line:84% So the next section is basically why 00:05:56.492 --> 00:05:58.200 align:middle line:84% we think that there might be a connection 00:05:58.200 --> 00:06:03.010 align:middle line:84% between common genetic variation and these common diseases. 00:06:03.010 --> 00:06:06.360 align:middle line:84% So this is a stretch of a genome. 00:06:06.360 --> 00:06:09.630 align:middle line:84% I think I borrowed this from David Altshuler or Eric Lander. 00:06:09.630 --> 00:06:11.100 align:middle line:90% I'm not sure. 00:06:11.100 --> 00:06:14.370 align:middle line:84% And this is what genetic variation looks like. 00:06:14.370 --> 00:06:16.200 align:middle line:90% So you have a long stretch here. 00:06:16.200 --> 00:06:19.950 align:middle line:84% And if you compare two stretches taken from different people, 00:06:19.950 --> 00:06:22.410 align:middle line:84% you'll find that there's one place here and one place 00:06:22.410 --> 00:06:25.470 align:middle line:90% here where maybe they differ. 00:06:25.470 --> 00:06:28.228 align:middle line:84% And most of those variants change a single nucleotide 00:06:28.228 --> 00:06:30.270 align:middle line:84% or single DNA letter, and it's called, therefore, 00:06:30.270 --> 00:06:33.420 align:middle line:90% Single Nucleotide Polymorphism. 00:06:33.420 --> 00:06:36.000 align:middle line:84% And means many forms, or in this case, 00:06:36.000 --> 00:06:39.510 align:middle line:84% would be just two forms, or a SNP for short. 00:06:39.510 --> 00:06:41.760 align:middle line:84% And what you might find is there's this stretch of DNA 00:06:41.760 --> 00:06:42.600 align:middle line:90% where everybody is the same. 00:06:42.600 --> 00:06:44.642 align:middle line:84% And then at this nucleotide, some people have a C 00:06:44.642 --> 00:06:47.520 align:middle line:84% and some people have a T. And what we want to know 00:06:47.520 --> 00:06:49.490 align:middle line:84% is, of all these different genetic variants, 00:06:49.490 --> 00:06:51.990 align:middle line:84% which ones are going to be the ones that are going to affect 00:06:51.990 --> 00:06:53.110 align:middle line:90% things that we care about? 00:06:53.110 --> 00:06:54.693 align:middle line:84% So here's something that we should all 00:06:54.693 --> 00:06:57.390 align:middle line:84% be passionate about in the next few months, 00:06:57.390 --> 00:07:00.210 align:middle line:84% so whether you're happy or sad that A-Rod 00:07:00.210 --> 00:07:02.800 align:middle line:84% went to the Yankees or something like that. 00:07:02.800 --> 00:07:07.020 align:middle line:84% But no, in seriousness, if this were healthy and diseased, 00:07:07.020 --> 00:07:10.470 align:middle line:84% not to say which is which here, then we 00:07:10.470 --> 00:07:13.078 align:middle line:84% would be interested to see which variants track along 00:07:13.078 --> 00:07:13.620 align:middle line:90% with disease. 00:07:13.620 --> 00:07:15.780 align:middle line:84% And that's going to be the general gist of what 00:07:15.780 --> 00:07:18.070 align:middle line:90% people are trying to do. 00:07:18.070 --> 00:07:19.650 align:middle line:84% So I said that there's a connection 00:07:19.650 --> 00:07:21.580 align:middle line:84% between common variants and disease. 00:07:21.580 --> 00:07:24.120 align:middle line:84% And so one question is, why might we be 00:07:24.120 --> 00:07:26.000 align:middle line:90% interested in common variants? 00:07:26.000 --> 00:07:28.760 align:middle line:84% And if you think about it, before I begin, 00:07:28.760 --> 00:07:31.018 align:middle line:84% what's known so far about genetic disease? 00:07:31.018 --> 00:07:32.810 align:middle line:84% You might say, well, that's the wrong place 00:07:32.810 --> 00:07:34.977 align:middle line:84% to look because if you look at things like Tay-Sachs 00:07:34.977 --> 00:07:38.360 align:middle line:84% or cystic fibrosis, those are mostly 00:07:38.360 --> 00:07:41.060 align:middle line:84% caused by genetic variants that are 00:07:41.060 --> 00:07:42.410 align:middle line:90% quite rare in the population. 00:07:42.410 --> 00:07:47.180 align:middle line:84% They're severe deleterious mutations that 00:07:47.180 --> 00:07:48.870 align:middle line:90% cause these severe diseases. 00:07:48.870 --> 00:07:51.930 align:middle line:84% So if those genetic diseases are caused by rare variants, 00:07:51.930 --> 00:07:53.465 align:middle line:84% why wouldn't things like diabetes 00:07:53.465 --> 00:07:54.590 align:middle line:90% be caused by rare variants? 00:07:54.590 --> 00:07:56.250 align:middle line:84% And it's an open question, and I'll get into that. 00:07:56.250 --> 00:07:58.167 align:middle line:84% But let me just take you through at least some 00:07:58.167 --> 00:07:59.480 align:middle line:90% of the background. 00:07:59.480 --> 00:08:02.000 align:middle line:84% So I showed you earlier that if you sequence a stretch 00:08:02.000 --> 00:08:04.677 align:middle line:84% of genome and imagine this is about 1,000 bases from two 00:08:04.677 --> 00:08:07.010 align:middle line:84% different chromosomes-- it could be two chromosomes from 00:08:07.010 --> 00:08:09.920 align:middle line:84% the same person, or it could be one chromosome from somebody 00:08:09.920 --> 00:08:13.250 align:middle line:84% here and one chromosome from somebody around the world-- 00:08:13.250 --> 00:08:14.880 align:middle line:84% you'd get the same answer more or less, 00:08:14.880 --> 00:08:16.713 align:middle line:84% which is that, on average, you'd see about 1 00:08:16.713 --> 00:08:19.770 align:middle line:84% every 1,200 bases would be different. 00:08:19.770 --> 00:08:21.650 align:middle line:84% You can then ask the question, well, 00:08:21.650 --> 00:08:24.875 align:middle line:84% having identified these two alleles, the C allele and the A 00:08:24.875 --> 00:08:28.970 align:middle line:84% allele, if I then look in the same 1,000 bases of lots 00:08:28.970 --> 00:08:32.944 align:middle line:84% of other chromosomes, is it that this A allele or maybe this C 00:08:32.944 --> 00:08:35.630 align:middle line:84% allele is just a really unique event that's private to one 00:08:35.630 --> 00:08:36.620 align:middle line:90% of these chromosomes? 00:08:36.620 --> 00:08:41.190 align:middle line:84% Or do I see both alleles around the whole world? 00:08:41.190 --> 00:08:44.450 align:middle line:84% And the answer, most of the time, about 90% of the time, 00:08:44.450 --> 00:08:47.370 align:middle line:84% is that you see both alleles around the world. 00:08:47.370 --> 00:08:49.070 align:middle line:84% And so what that means is that most 00:08:49.070 --> 00:08:52.850 align:middle line:84% of the differences that you can find between two chromosomes 00:08:52.850 --> 00:08:56.850 align:middle line:84% are explainable or attributable to the variation, 00:08:56.850 --> 00:08:58.685 align:middle line:84% those variants that are common, that 00:08:58.685 --> 00:09:01.310 align:middle line:84% are shared throughout the world that are not like the Tay-Sachs 00:09:01.310 --> 00:09:02.090 align:middle line:90% mutations. 00:09:02.090 --> 00:09:04.070 align:middle line:84% But they're actually old variants 00:09:04.070 --> 00:09:06.050 align:middle line:84% that happened a long time ago and are 00:09:06.050 --> 00:09:10.640 align:middle line:84% present in multiple populations around the world. 00:09:10.640 --> 00:09:11.770 align:middle line:90% So right. 00:09:11.770 --> 00:09:14.110 align:middle line:84% And that's basically what that says. 00:09:14.110 --> 00:09:16.100 align:middle line:84% And so this has led to a hypothesis. 00:09:16.100 --> 00:09:18.782 align:middle line:84% So the assumptions are as follows-- 00:09:18.782 --> 00:09:20.740 align:middle line:84% first of all, that most variation in the genome 00:09:20.740 --> 00:09:22.540 align:middle line:90% is evolutionarily neutral. 00:09:22.540 --> 00:09:26.680 align:middle line:84% That actually may or may not be true, but actually, 00:09:26.680 --> 00:09:28.750 align:middle line:90% whether that-- 00:09:28.750 --> 00:09:31.330 align:middle line:84% as long as it's not, most variation 00:09:31.330 --> 00:09:32.840 align:middle line:90% is evolutionarily beneficial. 00:09:32.840 --> 00:09:33.340 align:middle line:90% We're OK. 00:09:33.340 --> 00:09:38.180 align:middle line:84% But let's assume there's some background for the genome. 00:09:38.180 --> 00:09:42.460 align:middle line:84% So then you have a snip that's 22 kb away from the nearest 00:09:42.460 --> 00:09:44.590 align:middle line:84% gene and probably doesn't affect the function 00:09:44.590 --> 00:09:45.445 align:middle line:90% of that gene at all. 00:09:45.445 --> 00:09:47.320 align:middle line:84% So assume that evolution has not particularly 00:09:47.320 --> 00:09:50.350 align:middle line:90% cared about that base change. 00:09:50.350 --> 00:09:52.240 align:middle line:84% We know that most of this variation, most 00:09:52.240 --> 00:09:53.560 align:middle line:84% of the background variation is probably 00:09:53.560 --> 00:09:54.880 align:middle line:84% what you would call this, rather than neutral, 00:09:54.880 --> 00:09:56.525 align:middle line:90% is due to these common variants. 00:09:56.525 --> 00:09:58.900 align:middle line:84% That's what I just showed you, that 90% of that variation 00:09:58.900 --> 00:10:00.492 align:middle line:90% is due to common variants. 00:10:00.492 --> 00:10:02.950 align:middle line:84% We know also, what I've been telling you with the Tay-Sachs 00:10:02.950 --> 00:10:05.200 align:middle line:84% disease, that traits under negative selection 00:10:05.200 --> 00:10:07.580 align:middle line:84% will, in fact, be largely due to rare variants. 00:10:07.580 --> 00:10:09.130 align:middle line:84% And there's a nice description of why 00:10:09.130 --> 00:10:14.410 align:middle line:84% this should be true in an article by Jonathan Pritchard. 00:10:14.410 --> 00:10:15.940 align:middle line:84% But there's a hypothesis that traits 00:10:15.940 --> 00:10:18.148 align:middle line:84% that are not under negative selection or at least not 00:10:18.148 --> 00:10:20.380 align:middle line:84% under severe negative selection will actually 00:10:20.380 --> 00:10:24.280 align:middle line:84% be due to variants that look more like the background, 00:10:24.280 --> 00:10:26.227 align:middle line:90% so that are common variants. 00:10:26.227 --> 00:10:28.810 align:middle line:84% And so this is what's called the common disease-common variant 00:10:28.810 --> 00:10:29.770 align:middle line:90% hypothesis. 00:10:29.770 --> 00:10:33.550 align:middle line:84% And one of the more forceful, I would say, descriptions of it 00:10:33.550 --> 00:10:37.360 align:middle line:84% is a paper by Reich and Lander in Trends in Genetics, 2002. 00:10:37.360 --> 00:10:39.880 align:middle line:84% And the idea for why common diseases might 00:10:39.880 --> 00:10:42.610 align:middle line:84% be like-- might be due to alleles 00:10:42.610 --> 00:10:44.500 align:middle line:84% that are not under severe negative selection 00:10:44.500 --> 00:10:45.680 align:middle line:90% goes like this. 00:10:45.680 --> 00:10:49.730 align:middle line:84% First of all, they're actually common. 00:10:49.730 --> 00:10:52.300 align:middle line:84% So in some sense, it sounds tautological, 00:10:52.300 --> 00:10:53.680 align:middle line:90% but it's actually not. 00:10:53.680 --> 00:10:56.230 align:middle line:84% They can't be that evolutionarily deleterious 00:10:56.230 --> 00:10:57.850 align:middle line:90% if they're extremely-- 00:10:57.850 --> 00:10:59.450 align:middle line:90% if they're extremely common. 00:10:59.450 --> 00:11:00.460 align:middle line:90% So that's one argument. 00:11:00.460 --> 00:11:02.050 align:middle line:84% Two is that they're multifactorial. 00:11:02.050 --> 00:11:05.620 align:middle line:84% So each gene that contributes to these common diseases actually 00:11:05.620 --> 00:11:07.630 align:middle line:84% only contributes a little bit to susceptibility. 00:11:07.630 --> 00:11:08.620 align:middle line:84% And that's going to be a theme that'll 00:11:08.620 --> 00:11:10.270 align:middle line:90% run through this whole talk. 00:11:10.270 --> 00:11:13.210 align:middle line:84% So even if it were a-- even if they were somewhat 00:11:13.210 --> 00:11:15.490 align:middle line:84% evolutionarily deleterious, each individual allele 00:11:15.490 --> 00:11:18.280 align:middle line:84% would only be contributing a small fraction of that. 00:11:18.280 --> 00:11:20.290 align:middle line:84% Third is they often tend to be late onset 00:11:20.290 --> 00:11:23.270 align:middle line:90% or after reproductive age. 00:11:23.270 --> 00:11:25.790 align:middle line:84% So again, evolution cares less about stuff like that. 00:11:25.790 --> 00:11:26.790 align:middle line:90% Although, we don't know. 00:11:26.790 --> 00:11:29.170 align:middle line:84% May be good to have had your grandfather around in Africa 00:11:29.170 --> 00:11:30.100 align:middle line:90% 50,000 years ago. 00:11:30.100 --> 00:11:32.620 align:middle line:84% So it could be the late onset diseases were bad. 00:11:32.620 --> 00:11:34.780 align:middle line:84% Another thing is that it's been hypothesized 00:11:34.780 --> 00:11:37.660 align:middle line:84% that some of these diseases like hypertension or diabetes 00:11:37.660 --> 00:11:40.420 align:middle line:84% actually have what's called balancing selection, which 00:11:40.420 --> 00:11:44.170 align:middle line:84% is that yeah, they give you diabetes late in life, 00:11:44.170 --> 00:11:46.060 align:middle line:84% but the reason that they give you diabetes 00:11:46.060 --> 00:11:51.980 align:middle line:84% is because they help make you hold on to calories better. 00:11:51.980 --> 00:11:55.180 align:middle line:84% So that gives you obesity now when we have McDonald's around. 00:11:55.180 --> 00:11:58.045 align:middle line:84% But 100,000 years ago when food was scarce, 00:11:58.045 --> 00:11:59.170 align:middle line:90% maybe that was really good. 00:11:59.170 --> 00:12:01.430 align:middle line:84% There's something called the thrifty gene hypothesis. 00:12:01.430 --> 00:12:02.805 align:middle line:84% So for hypertension, maybe it was 00:12:02.805 --> 00:12:04.540 align:middle line:84% good to hold onto salt if you lived 00:12:04.540 --> 00:12:07.880 align:middle line:90% in a desert, that sort of thing. 00:12:07.880 --> 00:12:12.850 align:middle line:84% So if we're looking for a common variation 00:12:12.850 --> 00:12:14.950 align:middle line:84% or thinking about that this may be a good place 00:12:14.950 --> 00:12:17.950 align:middle line:84% to look for the genetic variants that underlie common disease-- 00:12:17.950 --> 00:12:19.570 align:middle line:84% and again, this hypothesis does not 00:12:19.570 --> 00:12:21.010 align:middle line:84% say that rare variation won't also 00:12:21.010 --> 00:12:22.600 align:middle line:90% contribute to common disease. 00:12:22.600 --> 00:12:24.373 align:middle line:84% Just says a common variation will 00:12:24.373 --> 00:12:25.540 align:middle line:90% be part of the contributors. 00:12:25.540 --> 00:12:27.082 align:middle line:84% So if we're looking for it, it's been 00:12:27.082 --> 00:12:30.730 align:middle line:84% estimated that there are about 10 million common SNPs where 00:12:30.730 --> 00:12:33.280 align:middle line:84% both alleles are found at more than 1% 00:12:33.280 --> 00:12:34.960 align:middle line:90% in the general population. 00:12:34.960 --> 00:12:37.970 align:middle line:84% About 6 million of those are actually already in databases. 00:12:37.970 --> 00:12:39.190 align:middle line:90% So it's actually most-- 00:12:39.190 --> 00:12:41.170 align:middle line:84% a good fraction that have been found. 00:12:41.170 --> 00:12:44.470 align:middle line:84% And as you'll see, there's a lot of redundancy between SNPs. 00:12:44.470 --> 00:12:47.513 align:middle line:84% So probably a very high fraction of the variation, 00:12:47.513 --> 00:12:49.180 align:middle line:84% common variation in the genome, actually 00:12:49.180 --> 00:12:51.370 align:middle line:84% has been cataloged, which is a remarkable thing. 00:12:51.370 --> 00:12:54.850 align:middle line:84% When I started in this field not very long ago, 00:12:54.850 --> 00:12:56.980 align:middle line:84% I think six years ago, there was a paper 00:12:56.980 --> 00:13:00.220 align:middle line:84% that came out of the Lander lab, which 00:13:00.220 --> 00:13:03.550 align:middle line:84% is a massive tour de force, cataloging 00:13:03.550 --> 00:13:05.200 align:middle line:90% 2,000 SNPs across the genome. 00:13:05.200 --> 00:13:08.050 align:middle line:84% And this quintupled the number of SNPs that were known. 00:13:08.050 --> 00:13:09.070 align:middle line:90% And so this is now-- 00:13:09.070 --> 00:13:12.460 align:middle line:84% this 3,000 times as many now, just in the last six years. 00:13:12.460 --> 00:13:14.950 align:middle line:84% And you can find all of these in websites 00:13:14.950 --> 00:13:17.320 align:middle line:84% like dbSNP or the UCSC Genome Browser, 00:13:17.320 --> 00:13:19.030 align:middle line:90% and CBI has a genome browser. 00:13:19.030 --> 00:13:21.350 align:middle line:90% And they're all right there. 00:13:21.350 --> 00:13:24.070 align:middle line:84% So this is a screenshot from the UCSC Genome Browser. 00:13:24.070 --> 00:13:25.842 align:middle line:84% Each one of these little lines is a SNP. 00:13:25.842 --> 00:13:27.550 align:middle line:84% And this is probably already out of date, 00:13:27.550 --> 00:13:30.550 align:middle line:84% even though I took it not too long ago from there. 00:13:30.550 --> 00:13:33.340 align:middle line:84% And this is just in one gene, each of these are-- 00:13:33.340 --> 00:13:35.470 align:middle line:84% on each of these two columns is a SNP. 00:13:35.470 --> 00:13:38.850 align:middle line:90% 00:13:38.850 --> 00:13:43.500 align:middle line:84% So if we want to use this catalog of common variation 00:13:43.500 --> 00:13:47.218 align:middle line:84% to try to find the alleles that are responsible for disease, 00:13:47.218 --> 00:13:48.010 align:middle line:90% how can we do that? 00:13:48.010 --> 00:13:49.410 align:middle line:84% Well, one way would actually be to try 00:13:49.410 --> 00:13:51.690 align:middle line:84% to study every variant or at least every common variant. 00:13:51.690 --> 00:13:52.523 align:middle line:90% This would be great. 00:13:52.523 --> 00:13:55.230 align:middle line:84% It would be an unbiased search of the entire genome, basically 00:13:55.230 --> 00:13:57.540 align:middle line:84% the entire universe of common variation. 00:13:57.540 --> 00:14:00.180 align:middle line:84% And this would be great if we could do it, 00:14:00.180 --> 00:14:01.680 align:middle line:84% but it's not currently practical. 00:14:01.680 --> 00:14:04.530 align:middle line:84% 10 million people-- sorry, 10 million variants. 00:14:04.530 --> 00:14:07.140 align:middle line:84% And as I'll show you, you need many thousands of people 00:14:07.140 --> 00:14:07.720 align:middle line:90% to do this. 00:14:07.720 --> 00:14:09.540 align:middle line:84% So it actually works out to hundreds 00:14:09.540 --> 00:14:11.410 align:middle line:90% of billions of genotypes. 00:14:11.410 --> 00:14:13.890 align:middle line:84% So that's something like the Pentagon budget 00:14:13.890 --> 00:14:15.870 align:middle line:84% or something for a year to do one study. 00:14:15.870 --> 00:14:18.820 align:middle line:84% So NIH doesn't like grants like that. 00:14:18.820 --> 00:14:22.050 align:middle line:84% So basically, until this gets much cheaper 00:14:22.050 --> 00:14:23.730 align:middle line:84% and/or we apply some of the shortcuts 00:14:23.730 --> 00:14:26.820 align:middle line:84% that I'll tell you about, you need to pick which genes 00:14:26.820 --> 00:14:29.610 align:middle line:84% and variants we're going to study. 00:14:29.610 --> 00:14:32.010 align:middle line:84% So this brings into methods for finding traits. 00:14:32.010 --> 00:14:33.390 align:middle line:90% Any questions so far? 00:14:33.390 --> 00:14:36.430 align:middle line:90% 00:14:36.430 --> 00:14:39.970 align:middle line:84% So we're on the track of trying to find which 00:14:39.970 --> 00:14:42.100 align:middle line:84% of these common variants might be contributing 00:14:42.100 --> 00:14:45.050 align:middle line:90% to common diseases. 00:14:45.050 --> 00:14:47.348 align:middle line:84% So really, three approaches, two of which 00:14:47.348 --> 00:14:49.390 align:middle line:84% have to do with common variants, and one actually 00:14:49.390 --> 00:14:51.307 align:middle line:84% which explicitly has to do with rare variants. 00:14:51.307 --> 00:14:54.430 align:middle line:90% And I'll get to that at the end. 00:14:54.430 --> 00:14:58.218 align:middle line:84% So the traditional approach to finding genes for disease 00:14:58.218 --> 00:15:00.010 align:middle line:84% has actually been something called linkage. 00:15:00.010 --> 00:15:00.552 align:middle line:90% I don't know. 00:15:00.552 --> 00:15:02.170 align:middle line:90% Have guys done linkage? 00:15:02.170 --> 00:15:05.480 align:middle line:84% Is that something that's been done here? 00:15:05.480 --> 00:15:05.980 align:middle line:90% Yes? 00:15:05.980 --> 00:15:07.120 align:middle line:90% No? 00:15:07.120 --> 00:15:09.610 align:middle line:84% So I'll briefly go through linkage. 00:15:09.610 --> 00:15:12.760 align:middle line:84% So linkage requires families where 00:15:12.760 --> 00:15:15.850 align:middle line:84% you have, in the case of disease, 00:15:15.850 --> 00:15:19.340 align:middle line:84% multiple affected relatives in the same family. 00:15:19.340 --> 00:15:20.860 align:middle line:84% So those could be brother/sister, 00:15:20.860 --> 00:15:24.070 align:middle line:84% or it could be cousins or an uncle 00:15:24.070 --> 00:15:26.720 align:middle line:84% and a niece or something like that. 00:15:26.720 --> 00:15:31.690 align:middle line:84% And the idea is that there is a single disease-causing allele 00:15:31.690 --> 00:15:33.130 align:middle line:90% that's segregating that family. 00:15:33.130 --> 00:15:35.005 align:middle line:84% So it's being passed on to some of the people 00:15:35.005 --> 00:15:36.970 align:middle line:90% in the family and not others. 00:15:36.970 --> 00:15:40.828 align:middle line:84% And the other assumption is that the disease will more or less-- 00:15:40.828 --> 00:15:43.120 align:middle line:84% doesn't have to be perfect, but has to be pretty good-- 00:15:43.120 --> 00:15:47.098 align:middle line:84% will more or less track with that allele. 00:15:47.098 --> 00:15:48.640 align:middle line:84% Now, for linkage, you're not actually 00:15:48.640 --> 00:15:49.883 align:middle line:90% typing the allele itself. 00:15:49.883 --> 00:15:51.550 align:middle line:84% What you're doing, you're typing markers 00:15:51.550 --> 00:15:52.700 align:middle line:90% that surround the alleles. 00:15:52.700 --> 00:15:56.080 align:middle line:84% So they may be a couple million bases away, actually. 00:15:56.080 --> 00:15:58.660 align:middle line:84% And that's a typical thing is about markers 00:15:58.660 --> 00:15:59.950 align:middle line:90% spaced every 10 million-- 00:15:59.950 --> 00:16:02.950 align:middle line:90% every few to 10 million bases. 00:16:02.950 --> 00:16:05.590 align:middle line:84% And the idea is that since you're just 00:16:05.590 --> 00:16:09.070 align:middle line:84% looking in a family, it turns out 00:16:09.070 --> 00:16:11.825 align:middle line:84% that a million bases is what's called one centimorgan, which 00:16:11.825 --> 00:16:13.450 align:middle line:84% means that recombination events happens 00:16:13.450 --> 00:16:15.730 align:middle line:90% once every hundred meiosis. 00:16:15.730 --> 00:16:18.100 align:middle line:84% so within a few centimorgans, there's 00:16:18.100 --> 00:16:19.990 align:middle line:84% generally not going to be any recombination 00:16:19.990 --> 00:16:21.370 align:middle line:90% events within a family. 00:16:21.370 --> 00:16:23.830 align:middle line:84% So the marker that's a few million bases away 00:16:23.830 --> 00:16:26.710 align:middle line:84% will travel with the disease allele in the family. 00:16:26.710 --> 00:16:29.930 align:middle line:84% So you use that marker as a surrogate for the disease 00:16:29.930 --> 00:16:34.210 align:middle line:84% allele, and you follow the marker through the family. 00:16:34.210 --> 00:16:35.677 align:middle line:84% And you say which of these markers. 00:16:35.677 --> 00:16:37.510 align:middle line:84% So if this marker is traveling with disease, 00:16:37.510 --> 00:16:39.843 align:middle line:84% you say maybe that's pretty close to the disease allele. 00:16:39.843 --> 00:16:41.800 align:middle line:84% But a marker on some other chromosome 00:16:41.800 --> 00:16:43.880 align:middle line:84% won't be segregating with the disease. 00:16:43.880 --> 00:16:46.090 align:middle line:84% So you'll say that region isn't linked. 00:16:46.090 --> 00:16:47.350 align:middle line:90% So that's linkage, basically. 00:16:47.350 --> 00:16:48.965 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:16:48.965 --> 00:16:50.590 align:middle line:84% JOEL HIRSCHHORN: So traditionally, it's 00:16:50.590 --> 00:16:53.560 align:middle line:84% been microsatellites, which are markers 00:16:53.560 --> 00:16:57.940 align:middle line:84% that are things like repeats of the sequence TATATATATA, 00:16:57.940 --> 00:16:59.020 align:middle line:90% and some people have-- 00:16:59.020 --> 00:17:01.280 align:middle line:84% DNA polymerase doesn't like to copy those very well. 00:17:01.280 --> 00:17:05.170 align:middle line:84% So some people have 50 copies, some people have 44, 00:17:05.170 --> 00:17:07.390 align:middle line:84% some people have 53, that sort of thing. 00:17:07.390 --> 00:17:09.535 align:middle line:90% And you can type those. 00:17:09.535 --> 00:17:12.460 align:middle line:84% Those are a pain in the ass to work with, actually. 00:17:12.460 --> 00:17:15.460 align:middle line:84% And so people are shifting over, starting to, 00:17:15.460 --> 00:17:17.805 align:middle line:84% to use SNP-based linkage analysis. 00:17:17.805 --> 00:17:19.180 align:middle line:84% Because all you need is basically 00:17:19.180 --> 00:17:21.970 align:middle line:84% a marker that lets you distinguish 00:17:21.970 --> 00:17:25.119 align:middle line:84% each of the chromosomes that are segregating in the family. 00:17:25.119 --> 00:17:28.900 align:middle line:84% So probably, microsatellites are-- 00:17:28.900 --> 00:17:31.120 align:middle line:84% it's going to-- there's going to be a shift. 00:17:31.120 --> 00:17:33.640 align:middle line:84% But it's been wildly successful for single gene disorders. 00:17:33.640 --> 00:17:35.223 align:middle line:84% And the reason it's been successful is 00:17:35.223 --> 00:17:37.540 align:middle line:84% because there's always just a single copy in a family 00:17:37.540 --> 00:17:41.050 align:middle line:84% of the disease allele, and the segregation with disease 00:17:41.050 --> 00:17:44.350 align:middle line:84% and that marker is really very good because if you have 00:17:44.350 --> 00:17:47.140 align:middle line:84% the disease-causing allele, you get disease, 00:17:47.140 --> 00:17:48.700 align:middle line:90% and if you don't, you don't. 00:17:48.700 --> 00:17:53.060 align:middle line:84% So for common diseases, none of those things are true. 00:17:53.060 --> 00:17:55.885 align:middle line:84% First of all, because multiple genes are involved, 00:17:55.885 --> 00:17:57.760 align:middle line:84% you might not have the disease-causing allele 00:17:57.760 --> 00:18:00.010 align:middle line:84% but get the disease anyway and vice versa. 00:18:00.010 --> 00:18:02.620 align:middle line:84% You might have the disease susceptibility allele, 00:18:02.620 --> 00:18:05.460 align:middle line:84% really it should be called, and not get the disease. 00:18:05.460 --> 00:18:07.210 align:middle line:84% And furthermore, there might be multiple-- 00:18:07.210 --> 00:18:09.010 align:middle line:84% because it might be a common variant, 00:18:09.010 --> 00:18:10.660 align:middle line:84% there actually might be multiple copies 00:18:10.660 --> 00:18:13.160 align:middle line:84% of the susceptibility allele traveling in the same family. 00:18:13.160 --> 00:18:16.510 align:middle line:84% So the markers that are a couple million bases away now 00:18:16.510 --> 00:18:20.650 align:middle line:84% no longer perfectly mark the disease susceptibility allele 00:18:20.650 --> 00:18:23.740 align:middle line:84% because maybe on this side of the family, it's marked by-- 00:18:23.740 --> 00:18:25.890 align:middle line:84% it travels with this marker, this allele. 00:18:25.890 --> 00:18:28.390 align:middle line:84% But then on this side of the family, where it's segregating, 00:18:28.390 --> 00:18:29.950 align:middle line:84% it travels with a different marker. 00:18:29.950 --> 00:18:32.590 align:middle line:84% So all of these things conspire against you 00:18:32.590 --> 00:18:33.910 align:middle line:90% for common diseases. 00:18:33.910 --> 00:18:36.910 align:middle line:84% Nonetheless, people have found a couple of things 00:18:36.910 --> 00:18:39.880 align:middle line:84% through linkage of common diseases and tracked it down. 00:18:39.880 --> 00:18:41.650 align:middle line:84% DeCODE genetics in Iceland is probably 00:18:41.650 --> 00:18:45.200 align:middle line:84% the most prolific example of that. 00:18:45.200 --> 00:18:47.860 align:middle line:84% So if linkage isn't working, it brings us 00:18:47.860 --> 00:18:49.520 align:middle line:84% to the other main approach that's used, 00:18:49.520 --> 00:18:51.790 align:middle line:84% which I'll spend a lot of the talk talking about 00:18:51.790 --> 00:18:53.290 align:middle line:90% because it is the main method. 00:18:53.290 --> 00:18:54.975 align:middle line:90% It's called association studies. 00:18:54.975 --> 00:18:56.350 align:middle line:84% And basically, what this means is 00:18:56.350 --> 00:18:58.540 align:middle line:84% that rather than searching through the whole genome, 00:18:58.540 --> 00:18:59.770 align:middle line:90% you pick candidate genes. 00:18:59.770 --> 00:19:01.450 align:middle line:84% And Rich and [INAUDIBLE] have shown 00:19:01.450 --> 00:19:05.890 align:middle line:84% that this is much better suited than linkage to this scenario 00:19:05.890 --> 00:19:07.810 align:middle line:84% where you have common alleles that 00:19:07.810 --> 00:19:10.060 align:middle line:84% have only modest effects on disease, what's called 00:19:10.060 --> 00:19:12.340 align:middle line:90% modest penetrance. 00:19:12.340 --> 00:19:13.930 align:middle line:84% And then there's another approach 00:19:13.930 --> 00:19:15.550 align:middle line:84% that people are taking, which is they 00:19:15.550 --> 00:19:17.500 align:middle line:84% say, OK, I understand linkage might not 00:19:17.500 --> 00:19:18.820 align:middle line:90% be the best way to go. 00:19:18.820 --> 00:19:21.160 align:middle line:84% But I don't believe this common variants stuff. 00:19:21.160 --> 00:19:23.690 align:middle line:84% I want to look for rare variants that contribute to disease. 00:19:23.690 --> 00:19:25.160 align:middle line:84% And so there, what you are doing is, 00:19:25.160 --> 00:19:26.952 align:middle line:84% again, you're picking the genes to look at. 00:19:26.952 --> 00:19:29.530 align:middle line:84% But instead of just taking SNPs off the public map 00:19:29.530 --> 00:19:31.150 align:middle line:84% and looking for common variation, 00:19:31.150 --> 00:19:32.650 align:middle line:84% they're actually resequencing to try 00:19:32.650 --> 00:19:35.390 align:middle line:84% to find those rare variants that are present only in disease. 00:19:35.390 --> 00:19:37.398 align:middle line:84% And the best examples of this are 00:19:37.398 --> 00:19:39.940 align:middle line:84% BRCA1 for breast cancer, which actually originally identified 00:19:39.940 --> 00:19:43.630 align:middle line:84% by linkage, but that's a rare disease, and also 00:19:43.630 --> 00:19:44.995 align:middle line:90% MC4R mutations. 00:19:44.995 --> 00:19:47.860 align:middle line:84% That's a melanocortin-4-receptor gene, 00:19:47.860 --> 00:19:49.510 align:middle line:84% where there's a lot of rare variants 00:19:49.510 --> 00:19:53.800 align:middle line:84% that contribute to early onset, particularly, severe obesity. 00:19:53.800 --> 00:19:55.810 align:middle line:84% The difficulty with this approach right 00:19:55.810 --> 00:19:57.350 align:middle line:90% now is it's very expensive. 00:19:57.350 --> 00:19:59.740 align:middle line:84% It's much more expensive to resequence 00:19:59.740 --> 00:20:02.830 align:middle line:84% than it is to do genotyping of polymorphisms. 00:20:02.830 --> 00:20:06.370 align:middle line:84% And if the expense were flipped, people 00:20:06.370 --> 00:20:07.870 align:middle line:84% would probably do this resequencing 00:20:07.870 --> 00:20:10.010 align:middle line:84% because this gets you the common variants as well. 00:20:10.010 --> 00:20:12.640 align:middle line:84% And it's really just-- it's largely an expense issue 00:20:12.640 --> 00:20:13.690 align:middle line:90% that we can't do this. 00:20:13.690 --> 00:20:16.370 align:middle line:84% So I talked about the study that would take just the Pentagon 00:20:16.370 --> 00:20:16.870 align:middle line:90% budget. 00:20:16.870 --> 00:20:18.335 align:middle line:84% If you were going to do this, this 00:20:18.335 --> 00:20:20.710 align:middle line:84% would take probably the GDP for the history of the United 00:20:20.710 --> 00:20:21.960 align:middle line:90% States or something like that. 00:20:21.960 --> 00:20:25.210 align:middle line:84% So we're still a ways away from this. 00:20:25.210 --> 00:20:28.625 align:middle line:84% AUDIENCE: That flies in the face of the whole hypothesis, 00:20:28.625 --> 00:20:30.455 align:middle line:90% though, [INAUDIBLE] 00:20:30.455 --> 00:20:31.330 align:middle line:90% JOEL HIRSCHHORN: Yes. 00:20:31.330 --> 00:20:33.490 align:middle line:84% So I'm just saying that if you choose to-- 00:20:33.490 --> 00:20:36.820 align:middle line:84% I don't mean to say that the hypothesis has been proven. 00:20:36.820 --> 00:20:39.850 align:middle line:84% I believe that common variants contribute to common disease. 00:20:39.850 --> 00:20:42.220 align:middle line:84% There are still people who believe that that's not true 00:20:42.220 --> 00:20:44.428 align:middle line:84% at all and that really, this is the only thing that's 00:20:44.428 --> 00:20:48.490 align:middle line:84% ever going to yield anything is looking for rare variants. 00:20:48.490 --> 00:20:51.190 align:middle line:84% So and there are definitely a couple 00:20:51.190 --> 00:20:53.140 align:middle line:84% of cases where this has been true. 00:20:53.140 --> 00:20:55.628 align:middle line:84% And this really hasn't been explored to any degree 00:20:55.628 --> 00:20:56.920 align:middle line:90% just because it's so expensive. 00:20:56.920 --> 00:20:59.170 align:middle line:84% And I put it there just because if all 00:20:59.170 --> 00:21:01.510 align:middle line:84% of a sudden, to sequence somebody's genome for $100 00:21:01.510 --> 00:21:04.450 align:middle line:84% became feasible, you would take 1,000 people with diabetes 00:21:04.450 --> 00:21:05.590 align:middle line:90% and sequence their genomes. 00:21:05.590 --> 00:21:08.330 align:middle line:84% And that would be the way you do the experiment. 00:21:08.330 --> 00:21:12.940 align:middle line:90% So that's why it's on there. 00:21:12.940 --> 00:21:15.490 align:middle line:84% So this is a cartoon version of what I just told you. 00:21:15.490 --> 00:21:17.600 align:middle line:84% Linkage analysis lets you search the whole genome 00:21:17.600 --> 00:21:19.600 align:middle line:84% and points you to a region, although it does not 00:21:19.600 --> 00:21:20.517 align:middle line:90% point you to the gene. 00:21:20.517 --> 00:21:23.620 align:middle line:84% So remember, those markers are 10 million bases apart. 00:21:23.620 --> 00:21:26.470 align:middle line:84% And in fact, because of some of the vagaries of complex trait 00:21:26.470 --> 00:21:30.280 align:middle line:84% genetics and the reasons that linkage isn't that well-suited, 00:21:30.280 --> 00:21:32.230 align:middle line:84% usually you get about a 20 centimorgan region 00:21:32.230 --> 00:21:33.940 align:middle line:84% or about 20 million bases, which can have 00:21:33.940 --> 00:21:36.910 align:middle line:90% something like 200 genes in it. 00:21:36.910 --> 00:21:39.550 align:middle line:84% And it'll say somewhere in those 200 genes, 00:21:39.550 --> 00:21:43.090 align:middle line:84% there's something that's contributing to disease. 00:21:43.090 --> 00:21:46.390 align:middle line:84% Whereas association, you're looking one gene at a time, 00:21:46.390 --> 00:21:47.230 align:middle line:90% so you're not-- 00:21:47.230 --> 00:21:49.670 align:middle line:84% but you have to guess correctly where to go. 00:21:49.670 --> 00:21:53.020 align:middle line:84% And so you have to-- so there's this haystack here of 30,000 00:21:53.020 --> 00:21:55.210 align:middle line:90% genes or 30,000 straws. 00:21:55.210 --> 00:21:56.950 align:middle line:84% And you can pull one out and say, 00:21:56.950 --> 00:21:58.162 align:middle line:90% is that a straw or a needle? 00:21:58.162 --> 00:21:58.870 align:middle line:90% No, it's a straw. 00:21:58.870 --> 00:21:59.410 align:middle line:90% Next one. 00:21:59.410 --> 00:22:00.310 align:middle line:90% And look. 00:22:00.310 --> 00:22:02.170 align:middle line:84% And obviously, the success depends on you 00:22:02.170 --> 00:22:04.150 align:middle line:90% being able to guess correctly. 00:22:04.150 --> 00:22:06.700 align:middle line:84% But we actually-- at least for the first time, 00:22:06.700 --> 00:22:08.680 align:middle line:84% we know what the 30,000 straws are. 00:22:08.680 --> 00:22:11.840 align:middle line:84% So there's a website now with the entire genome on it. 00:22:11.840 --> 00:22:13.840 align:middle line:84% And we're getting pretty good, as I'll tell you, 00:22:13.840 --> 00:22:19.060 align:middle line:84% at telling whether things are real contributors to disease 00:22:19.060 --> 00:22:19.720 align:middle line:90% or not. 00:22:19.720 --> 00:22:21.730 align:middle line:84% But it's this interpretation of whether this 00:22:21.730 --> 00:22:25.900 align:middle line:84% is a straw or a needle that is going 00:22:25.900 --> 00:22:27.835 align:middle line:84% to be the focus of a lot of the talk 00:22:27.835 --> 00:22:29.710 align:middle line:84% because that's the study that a lot of people 00:22:29.710 --> 00:22:31.452 align:middle line:84% do is they pick a gene and they test it, 00:22:31.452 --> 00:22:33.160 align:middle line:84% and they claim it either does or does not 00:22:33.160 --> 00:22:34.118 align:middle line:90% have a role in disease. 00:22:34.118 --> 00:22:37.480 align:middle line:84% And I just want to give you hopefully 00:22:37.480 --> 00:22:40.330 align:middle line:84% some sophistication in reading those sorts of studies 00:22:40.330 --> 00:22:44.298 align:middle line:84% to try to say should you believe everything that you read? 00:22:44.298 --> 00:22:45.340 align:middle line:90% The obvious answer is no. 00:22:45.340 --> 00:22:48.520 align:middle line:84% But so this is what an association study is. 00:22:48.520 --> 00:22:52.120 align:middle line:84% You can imagine that you had healthy individuals and people 00:22:52.120 --> 00:22:54.220 align:middle line:90% with Alzheimer's disease. 00:22:54.220 --> 00:22:59.530 align:middle line:84% And there's a common variant in the APOE gene called APOE4. 00:22:59.530 --> 00:23:01.482 align:middle line:90% It encodes apolipoprotein E. 00:23:01.482 --> 00:23:03.190 align:middle line:84% And it is a common variant because if you 00:23:03.190 --> 00:23:07.750 align:middle line:84% look in the control population, you'll see it's there about 10% 00:23:07.750 --> 00:23:09.375 align:middle line:84% of the time, something like that. 00:23:09.375 --> 00:23:11.500 align:middle line:84% But if you look in people with Alzheimer's disease, 00:23:11.500 --> 00:23:15.590 align:middle line:84% you'll see those frequencies increased about threefold. 00:23:15.590 --> 00:23:18.670 align:middle line:84% And so this is the-- this is the paradigm for associations 00:23:18.670 --> 00:23:20.838 align:middle line:84% of common disease and common variants, 00:23:20.838 --> 00:23:22.630 align:middle line:84% and this has been seen over and over again. 00:23:22.630 --> 00:23:25.600 align:middle line:84% You look at just about every study that's been done, 00:23:25.600 --> 00:23:29.170 align:middle line:90% you see this increase in risk. 00:23:29.170 --> 00:23:29.670 align:middle line:90% Yeah. 00:23:29.670 --> 00:23:30.210 align:middle line:90% AUDIENCE: Sorry. 00:23:30.210 --> 00:23:30.510 align:middle line:90% JOEL HIRSCHHORN: No, no. 00:23:30.510 --> 00:23:31.010 align:middle line:90% Sure. 00:23:31.010 --> 00:23:35.803 align:middle line:84% AUDIENCE: Should I-- You also looked, obviously, [INAUDIBLE] 00:23:35.803 --> 00:23:36.720 align:middle line:90% JOEL HIRSCHHORN: Yeah. 00:23:36.720 --> 00:23:37.470 align:middle line:90% So actually, yeah. 00:23:37.470 --> 00:23:39.210 align:middle line:84% So there's a little bit of a subtlety 00:23:39.210 --> 00:23:40.500 align:middle line:90% to even this association. 00:23:40.500 --> 00:23:42.990 align:middle line:84% So it has been seen in other ethnicities. 00:23:42.990 --> 00:23:46.378 align:middle line:84% It looks-- although, the last time I looked at this, 00:23:46.378 --> 00:23:48.420 align:middle line:84% the data were not strong enough to say that there 00:23:48.420 --> 00:23:49.710 align:middle line:90% was definitely a difference. 00:23:49.710 --> 00:23:51.840 align:middle line:84% But it looks like the association with Alzheimer's 00:23:51.840 --> 00:23:53.820 align:middle line:84% disease might be a little bit weaker 00:23:53.820 --> 00:23:56.310 align:middle line:90% in African-derived populations. 00:23:56.310 --> 00:23:58.950 align:middle line:84% There could be a wide variety of reasons why that's true, 00:23:58.950 --> 00:24:01.800 align:middle line:84% and I'll go through a bunch of them just in general. 00:24:01.800 --> 00:24:07.285 align:middle line:84% AUDIENCE: Because there's less [INAUDIBLE] 00:24:07.285 --> 00:24:10.153 align:middle line:84% like if you just made that for all individuals-- if you made 00:24:10.153 --> 00:24:19.317 align:middle line:90% that [INAUDIBLE] 00:24:19.317 --> 00:24:20.150 align:middle line:90% JOEL HIRSCHHORN: No. 00:24:20.150 --> 00:24:24.890 align:middle line:84% So the rate, the frequency of APOE4, in this case, 00:24:24.890 --> 00:24:27.050 align:middle line:84% doesn't actually vary that much across populations. 00:24:27.050 --> 00:24:28.670 align:middle line:90% It's an old allele. 00:24:28.670 --> 00:24:30.830 align:middle line:84% Some alleles do, and you're getting into something 00:24:30.830 --> 00:24:32.747 align:middle line:84% called ethnic admixture, which I'll definitely 00:24:32.747 --> 00:24:33.860 align:middle line:90% get to in a little bit. 00:24:33.860 --> 00:24:36.580 align:middle line:90% 00:24:36.580 --> 00:24:40.440 align:middle line:84% So as I said, you have to guess right. 00:24:40.440 --> 00:24:42.652 align:middle line:84% So there's an issue of which genes do you guess? 00:24:42.652 --> 00:24:44.860 align:middle line:84% And ideally, again, one would like to do all of them. 00:24:44.860 --> 00:24:47.600 align:middle line:84% And it's getting closer to be able to do that, but-- yeah? 00:24:47.600 --> 00:24:50.800 align:middle line:84% AUDIENCE: Individuals with the BCRA1 example, 00:24:50.800 --> 00:24:54.163 align:middle line:84% when you take the linkage study and find that region that 00:24:54.163 --> 00:24:55.080 align:middle line:90% appears to be linked-- 00:24:55.080 --> 00:24:56.622 align:middle line:84% JOEL HIRSCHHORN: So that's definitely 00:24:56.622 --> 00:24:57.580 align:middle line:90% one thing you would do. 00:24:57.580 --> 00:24:58.980 align:middle line:84% So it's actually been hard enough 00:24:58.980 --> 00:25:06.250 align:middle line:84% until basically this year to do that. 00:25:06.250 --> 00:25:07.980 align:middle line:90% So you have 200 genes. 00:25:07.980 --> 00:25:11.250 align:middle line:84% And to thoroughly go through a gene 00:25:11.250 --> 00:25:14.712 align:middle line:84% and basically look at all 200 of them 00:25:14.712 --> 00:25:17.170 align:middle line:84% has been something that's just basically been out of reach. 00:25:17.170 --> 00:25:19.710 align:middle line:84% So to give you an example, I just got, I think, 00:25:19.710 --> 00:25:22.830 align:middle line:84% funded to do exactly that for an obesity linkage peak. 00:25:22.830 --> 00:25:25.200 align:middle line:84% And we couldn't have really written 00:25:25.200 --> 00:25:28.740 align:middle line:84% to do that as part of a regular R01-type grant. 00:25:28.740 --> 00:25:31.220 align:middle line:84% We could not have written an R01 to do that a year ago. 00:25:31.220 --> 00:25:33.498 align:middle line:84% It was just too expensive and too hard. 00:25:33.498 --> 00:25:35.790 align:middle line:84% It's just the information about which variants to type, 00:25:35.790 --> 00:25:38.040 align:middle line:84% and I'll go through, again, using haplotypes and stuff 00:25:38.040 --> 00:25:39.213 align:middle line:90% like that to do it. 00:25:39.213 --> 00:25:40.630 align:middle line:84% The information just wasn't there, 00:25:40.630 --> 00:25:42.090 align:middle line:90% and the expense was too high. 00:25:42.090 --> 00:25:42.840 align:middle line:90% But yes, I agree. 00:25:42.840 --> 00:25:44.340 align:middle line:84% Those would be the first candidates. 00:25:44.340 --> 00:25:46.620 align:middle line:84% So you would use information from linkage studies. 00:25:46.620 --> 00:25:48.690 align:middle line:84% And you might just say, I don't care about anything else 00:25:48.690 --> 00:25:49.440 align:middle line:90% that anybody says. 00:25:49.440 --> 00:25:51.255 align:middle line:90% I'm going to take linkage. 00:25:51.255 --> 00:25:53.520 align:middle line:84% If your complex trait or your common disease 00:25:53.520 --> 00:25:56.260 align:middle line:84% is fortunate enough to have a linkage peak that you believe, 00:25:56.260 --> 00:25:58.135 align:middle line:84% which is not true for most of them, actually, 00:25:58.135 --> 00:26:00.270 align:middle line:84% but some of them it is, you would say, 00:26:00.270 --> 00:26:01.500 align:middle line:90% those are the 200 best genes. 00:26:01.500 --> 00:26:02.700 align:middle line:90% I'm going to go for those. 00:26:02.700 --> 00:26:05.400 align:middle line:90% That's definitely an approach. 00:26:05.400 --> 00:26:09.180 align:middle line:84% You might say, well, linkage is not 00:26:09.180 --> 00:26:10.920 align:middle line:84% very well-powered for common disease. 00:26:10.920 --> 00:26:13.530 align:middle line:84% Probably most of the genes are not under linkage peaks. 00:26:13.530 --> 00:26:15.450 align:middle line:84% And most of the genes under the linkage peak 00:26:15.450 --> 00:26:17.110 align:middle line:90% are not the right one. 00:26:17.110 --> 00:26:18.545 align:middle line:90% So I'm going to ignore that. 00:26:18.545 --> 00:26:19.920 align:middle line:84% I'm going to go for things that-- 00:26:19.920 --> 00:26:23.195 align:middle line:84% I'm going to say 100 years of research has meant something. 00:26:23.195 --> 00:26:24.570 align:middle line:84% And if I'm studying diabetes, I'm 00:26:24.570 --> 00:26:26.940 align:middle line:84% going to do the insulin signaling pathway and beta cell 00:26:26.940 --> 00:26:28.905 align:middle line:90% function and obesi-- 00:26:28.905 --> 00:26:31.030 align:middle line:84% fat cell differentiation genes and stuff like that, 00:26:31.030 --> 00:26:32.863 align:middle line:84% and those are the genes I'm going to go for. 00:26:32.863 --> 00:26:35.798 align:middle line:84% Or you might say, well, this is just lamp post 00:26:35.798 --> 00:26:38.340 align:middle line:84% science because it's the biology that people have been doing. 00:26:38.340 --> 00:26:40.882 align:middle line:84% And this is not very powered, but this is a genome-wide tool. 00:26:40.882 --> 00:26:44.130 align:middle line:84% We can use expression analysis, so survey all the genes, 00:26:44.130 --> 00:26:47.430 align:middle line:84% find things that are expressed at different levels in disease 00:26:47.430 --> 00:26:51.330 align:middle line:84% versus healthy people and say that some 00:26:51.330 --> 00:26:53.070 align:middle line:84% of those or many of those changes 00:26:53.070 --> 00:26:55.590 align:middle line:84% might be secondary to the disease state. 00:26:55.590 --> 00:26:58.830 align:middle line:84% Maybe some of those represent a primary regulatory variant 00:26:58.830 --> 00:26:59.890 align:middle line:90% that affects disease. 00:26:59.890 --> 00:27:01.348 align:middle line:84% And so I'm going to look for those. 00:27:01.348 --> 00:27:03.930 align:middle line:84% And what we're trying to do and other people are trying to do 00:27:03.930 --> 00:27:05.040 align:middle line:84% is actually the most interesting genes 00:27:05.040 --> 00:27:07.123 align:middle line:84% are probably the ones that happen to fall in here, 00:27:07.123 --> 00:27:10.048 align:middle line:84% although there's no saying that they have 00:27:10.048 --> 00:27:11.340 align:middle line:90% to fall into this intersection. 00:27:11.340 --> 00:27:13.890 align:middle line:84% Maybe they fall in here or there or there or even 00:27:13.890 --> 00:27:18.720 align:middle line:84% just one of these or in the universe outside altogether. 00:27:18.720 --> 00:27:21.000 align:middle line:84% But I'll just give you an example for type 2 diabetes 00:27:21.000 --> 00:27:23.340 align:middle line:84% just because we are actually doing this sort of thing. 00:27:23.340 --> 00:27:26.370 align:middle line:84% So there is hundreds of years of biology 00:27:26.370 --> 00:27:28.980 align:middle line:84% or 100 years of biology probably on diabetes, 00:27:28.980 --> 00:27:31.650 align:middle line:84% a lot of mouse models, single gene disorders that 00:27:31.650 --> 00:27:32.740 align:middle line:90% cause the disease. 00:27:32.740 --> 00:27:33.960 align:middle line:90% So that's another thing. 00:27:33.960 --> 00:27:37.140 align:middle line:84% If severe variation causes a severe form of the disease, 00:27:37.140 --> 00:27:40.770 align:middle line:84% maybe mild variation causes late onset form. 00:27:40.770 --> 00:27:43.800 align:middle line:84% So you can-- there's a list of genes there. 00:27:43.800 --> 00:27:46.590 align:middle line:84% So for diabetes, this is one of those things where there really 00:27:46.590 --> 00:27:48.990 align:middle line:84% isn't a perfect linkage peak to go under. 00:27:48.990 --> 00:27:53.340 align:middle line:84% Graeme Bell's group identified the Calpain-10 gene 00:27:53.340 --> 00:27:55.830 align:middle line:84% by going under a peak where they had significant linkage. 00:27:55.830 --> 00:27:58.110 align:middle line:84% Although, if you actually put all the linkage evidence 00:27:58.110 --> 00:28:00.008 align:middle line:84% together from everybody, you would not have 00:28:00.008 --> 00:28:01.800 align:middle line:84% said that there was any significant linkage 00:28:01.800 --> 00:28:03.210 align:middle line:90% to diabetes there. 00:28:03.210 --> 00:28:10.560 align:middle line:84% That association is probably true, but not-- 00:28:10.560 --> 00:28:14.890 align:middle line:84% so you can do this, but it's not the end all. 00:28:14.890 --> 00:28:17.520 align:middle line:84% And then there's been expression studies in diabetes 00:28:17.520 --> 00:28:22.230 align:middle line:84% that's pointed to a pathway of oxidative phosphorylation. 00:28:22.230 --> 00:28:25.958 align:middle line:84% So some colleagues of mine, [INAUDIBLE],, David Altshuler, 00:28:25.958 --> 00:28:28.125 align:middle line:84% [INAUDIBLE],, Cecilia Lindgren, and others, and also 00:28:28.125 --> 00:28:30.060 align:middle line:90% a paper from Nick Joselyn. 00:28:30.060 --> 00:28:32.350 align:middle line:84% Mary Elizabeth Patty's pointed to that pathway. 00:28:32.350 --> 00:28:36.285 align:middle line:84% And there are genes that are actually in this red area, 00:28:36.285 --> 00:28:37.080 align:middle line:90% a few of them. 00:28:37.080 --> 00:28:41.260 align:middle line:84% So we're going after those genes, for example. 00:28:41.260 --> 00:28:43.500 align:middle line:84% So that's picking genes, and it's more of a-- 00:28:43.500 --> 00:28:46.470 align:middle line:84% at this point, more of a art form 00:28:46.470 --> 00:28:49.680 align:middle line:84% than any computerized algorithm to do it. 00:28:49.680 --> 00:28:52.800 align:middle line:84% Although, just even getting all the information into one 00:28:52.800 --> 00:28:56.060 align:middle line:84% place where you could even do the queries of what genes are 00:28:56.060 --> 00:28:58.560 align:middle line:84% under linkage peak, what genes are differentially expressed, 00:28:58.560 --> 00:29:00.210 align:middle line:84% what genes are in known pathways, 00:29:00.210 --> 00:29:05.820 align:middle line:84% that's actually tools that we're just trying to develop. 00:29:05.820 --> 00:29:07.650 align:middle line:84% So which genes, and then which variants? 00:29:07.650 --> 00:29:10.320 align:middle line:84% So ideally, you have the disease-causing allele 00:29:10.320 --> 00:29:12.060 align:middle line:84% in your hand and you're genotyping it, 00:29:12.060 --> 00:29:14.820 align:middle line:84% and you test it because they're the marker that you're testing, 00:29:14.820 --> 00:29:16.572 align:middle line:84% as opposed to being a distance away, 00:29:16.572 --> 00:29:18.280 align:middle line:84% is perfectly correlate with the variants. 00:29:18.280 --> 00:29:19.930 align:middle line:90% That's the ideal situation. 00:29:19.930 --> 00:29:22.470 align:middle line:84% And so that implies that maybe we should actually really 00:29:22.470 --> 00:29:24.330 align:middle line:84% concentrate our efforts on finding the variants that 00:29:24.330 --> 00:29:25.800 align:middle line:84% are most likely to be functional, so 00:29:25.800 --> 00:29:28.258 align:middle line:84% similar to the way you might try to find the genes that are 00:29:28.258 --> 00:29:30.200 align:middle line:90% most likely to be responsible. 00:29:30.200 --> 00:29:33.030 align:middle line:84% So the most obvious things are missense variants. 00:29:33.030 --> 00:29:34.750 align:middle line:90% And these are easy to recognize. 00:29:34.750 --> 00:29:36.460 align:middle line:84% It's been shown, if you look at their-- 00:29:36.460 --> 00:29:38.085 align:middle line:84% they're much rarer than they should be, 00:29:38.085 --> 00:29:40.640 align:middle line:84% which means evolution has cared about them to suppress them. 00:29:40.640 --> 00:29:43.030 align:middle line:84% So many of them are actually mildly deleterious. 00:29:43.030 --> 00:29:45.400 align:middle line:84% Maybe they're not severe deleterious, 00:29:45.400 --> 00:29:47.500 align:middle line:84% but maybe this is just the right balance of things 00:29:47.500 --> 00:29:49.330 align:middle line:90% to cause common disease. 00:29:49.330 --> 00:29:51.070 align:middle line:84% And even if you're doing rare variants, 00:29:51.070 --> 00:29:52.987 align:middle line:84% you can recognize them and group them together 00:29:52.987 --> 00:29:54.660 align:middle line:84% and study them as a class, if you 00:29:54.660 --> 00:29:57.160 align:middle line:84% want to say-- if you have the model that rare variants cause 00:29:57.160 --> 00:29:58.610 align:middle line:90% disease. 00:29:58.610 --> 00:30:01.930 align:middle line:84% So but we know, and there are plenty of cases 00:30:01.930 --> 00:30:04.270 align:middle line:84% now, where the variant that's been finally identified 00:30:04.270 --> 00:30:06.120 align:middle line:84% is actually for a complex trait has not 00:30:06.120 --> 00:30:07.120 align:middle line:90% been a missense variant. 00:30:07.120 --> 00:30:08.578 align:middle line:84% It's been a regulatory variant that 00:30:08.578 --> 00:30:11.320 align:middle line:84% affects the level of the gene but not the structure 00:30:11.320 --> 00:30:13.150 align:middle line:90% of the protein it encodes. 00:30:13.150 --> 00:30:15.370 align:middle line:84% There's a great example from type 1 diabetes 00:30:15.370 --> 00:30:17.345 align:middle line:84% from John Todd's group and others 00:30:17.345 --> 00:30:18.970 align:middle line:84% where there was a missense variant that 00:30:18.970 --> 00:30:21.040 align:middle line:84% was thought to be the right variant, 00:30:21.040 --> 00:30:22.960 align:middle line:84% but then they did a huge study and showed 00:30:22.960 --> 00:30:25.050 align:middle line:84% that in fact, all of the association that they 00:30:25.050 --> 00:30:26.800 align:middle line:84% saw with the missense variant was actually 00:30:26.800 --> 00:30:28.690 align:middle line:84% explained by something in the 3 prime UTR, 00:30:28.690 --> 00:30:30.920 align:middle line:84% and it affected the level of the gene. 00:30:30.920 --> 00:30:33.382 align:middle line:84% So the question is, well, these are hard to recognize. 00:30:33.382 --> 00:30:34.840 align:middle line:84% There's lots of variation out there 00:30:34.840 --> 00:30:35.890 align:middle line:90% we have no idea how to reach. 00:30:35.890 --> 00:30:37.890 align:middle line:84% So the reason we can recognize missense variants 00:30:37.890 --> 00:30:39.490 align:middle line:90% is we know the genetic code. 00:30:39.490 --> 00:30:43.090 align:middle line:84% They cracked that soon after DNA. 00:30:43.090 --> 00:30:45.760 align:middle line:84% But we don't know the regulatory code. 00:30:45.760 --> 00:30:48.970 align:middle line:84% So there's some effort now to try to crack, essentially, 00:30:48.970 --> 00:30:52.720 align:middle line:84% the regulatory code by using multiple species 00:30:52.720 --> 00:30:54.110 align:middle line:90% and doing comparisons. 00:30:54.110 --> 00:30:56.290 align:middle line:84% So this is from Eddie Rubin's group. 00:30:56.290 --> 00:30:57.790 align:middle line:90% This is a great website. 00:30:57.790 --> 00:30:59.440 align:middle line:90% It's called the ECR browser. 00:30:59.440 --> 00:31:02.140 align:middle line:84% And you can see, what you see here 00:31:02.140 --> 00:31:04.600 align:middle line:84% is regions that are actually evolutionarily conserved 00:31:04.600 --> 00:31:07.000 align:middle line:84% across species and therefore, may be important. 00:31:07.000 --> 00:31:09.650 align:middle line:84% Evolutionists has cared about them again. 00:31:09.650 --> 00:31:11.860 align:middle line:84% So they may be important regulatory regions. 00:31:11.860 --> 00:31:15.488 align:middle line:84% And there's experimental backup for that hypothesis. 00:31:15.488 --> 00:31:17.530 align:middle line:84% So basically, that would suggest that we should-- 00:31:17.530 --> 00:31:20.050 align:middle line:84% and [INAUDIBLE] have suggested this approach, 00:31:20.050 --> 00:31:23.620 align:middle line:84% at least for missense variants, that we should resequence 00:31:23.620 --> 00:31:27.010 align:middle line:84% targeted regions, try to find every variant that's in there, 00:31:27.010 --> 00:31:29.690 align:middle line:84% and make sure we type all of them. 00:31:29.690 --> 00:31:34.750 align:middle line:84% And again, the issue is that resequencing is expensive. 00:31:34.750 --> 00:31:37.160 align:middle line:84% But we might not have the causal variant in hand 00:31:37.160 --> 00:31:38.710 align:middle line:84% if we don't go through this approach 00:31:38.710 --> 00:31:41.320 align:middle line:84% or if we're bad at guessing what the regulatory variants are. 00:31:41.320 --> 00:31:42.945 align:middle line:84% So we should try to understand also 00:31:42.945 --> 00:31:44.320 align:middle line:84% the correlation between variants. 00:31:44.320 --> 00:31:45.778 align:middle line:84% I alluded to the fact that the SNPs 00:31:45.778 --> 00:31:47.687 align:middle line:90% are redundant with each other. 00:31:47.687 --> 00:31:49.270 align:middle line:84% So if you have the causal variant that 00:31:49.270 --> 00:31:52.030 align:middle line:84% hasn't been genotyped, you have to infer its effect 00:31:52.030 --> 00:31:54.760 align:middle line:90% by genotyping neighboring SNPs. 00:31:54.760 --> 00:31:56.560 align:middle line:84% And so they have to be correlated 00:31:56.560 --> 00:31:58.810 align:middle line:84% or what's called in linkage disequilibrium 00:31:58.810 --> 00:32:00.820 align:middle line:90% with the causal variant. 00:32:00.820 --> 00:32:03.520 align:middle line:84% It turns out that a lot of the times, that's true. 00:32:03.520 --> 00:32:05.830 align:middle line:84% So this is actually what the genome 00:32:05.830 --> 00:32:10.150 align:middle line:84% looks like most of the time or most of the genome, where 00:32:10.150 --> 00:32:11.858 align:middle line:84% that slide that I showed you before, 00:32:11.858 --> 00:32:13.150 align:middle line:90% imagine this is now zoomed out. 00:32:13.150 --> 00:32:14.918 align:middle line:84% We're looking now 30,000 bases of a bunch 00:32:14.918 --> 00:32:15.960 align:middle line:90% of different chromosomes. 00:32:15.960 --> 00:32:18.695 align:middle line:84% What you see is if you have a C here and a C here, 00:32:18.695 --> 00:32:20.570 align:middle line:84% you can predict that you're going to have a G 00:32:20.570 --> 00:32:22.540 align:middle line:84% here, a C here, an A here, and a C there. 00:32:22.540 --> 00:32:25.660 align:middle line:84% And these patterns of alleles are called haplotypes. 00:32:25.660 --> 00:32:28.090 align:middle line:84% And in most of the genome, there are these blocks 00:32:28.090 --> 00:32:30.580 align:middle line:84% of correlation or blocks of linkage disequilibrium 00:32:30.580 --> 00:32:32.477 align:middle line:84% where there are only a few common haplotypes. 00:32:32.477 --> 00:32:34.810 align:middle line:84% In this case, there's the red, the blue, and the yellow. 00:32:34.810 --> 00:32:39.410 align:middle line:84% And occasionally, there's a shuffling by recombination. 00:32:39.410 --> 00:32:42.010 align:middle line:84% These were first identified by Mark Daly and colleagues 00:32:42.010 --> 00:32:44.500 align:middle line:84% in genetics, and then this shown to be 00:32:44.500 --> 00:32:47.870 align:middle line:84% true on a genome-wide scale with Stacey Gabriel, et al. 00:32:47.870 --> 00:32:49.870 align:middle line:84% And the reason this is useful is that if there's 00:32:49.870 --> 00:32:53.020 align:middle line:84% a causal polymorphism but not one that you've typed, 00:32:53.020 --> 00:32:55.458 align:middle line:84% you can predict it pretty well by the neighboring markers. 00:32:55.458 --> 00:32:57.250 align:middle line:84% So that if you have a C here and an A here, 00:32:57.250 --> 00:32:59.866 align:middle line:84% then you probably are carrying this causal polymorphism. 00:32:59.866 --> 00:33:04.430 align:middle line:90% 00:33:04.430 --> 00:33:05.940 align:middle line:90% I'll take you through briefly. 00:33:05.940 --> 00:33:07.940 align:middle line:84% But I'm going to blow through this a little bit, 00:33:07.940 --> 00:33:13.325 align:middle line:84% just the range of what these patterns look like. 00:33:13.325 --> 00:33:14.700 align:middle line:84% There's a measure of correlation, 00:33:14.700 --> 00:33:17.870 align:middle line:84% which is not important to understand, called D prime. 00:33:17.870 --> 00:33:20.570 align:middle line:84% Basically, red means that the SNPs are correlated 00:33:20.570 --> 00:33:22.070 align:middle line:84% with each other, and any other color 00:33:22.070 --> 00:33:23.695 align:middle line:84% means that you're not sure if they are. 00:33:23.695 --> 00:33:26.250 align:middle line:84% And I'll show you a little plot like this, 00:33:26.250 --> 00:33:29.480 align:middle line:84% where this is the data for markers 1 and 2, 00:33:29.480 --> 00:33:33.410 align:middle line:84% and this is the data for markers 3 and 5. 00:33:33.410 --> 00:33:35.338 align:middle line:84% And this is what blocks look like, 00:33:35.338 --> 00:33:37.130 align:middle line:84% where you see a whole bunch of red triangle 00:33:37.130 --> 00:33:38.690 align:middle line:84% here, which means that all the markers 1 to 6 00:33:38.690 --> 00:33:39.982 align:middle line:90% are correlated with each other. 00:33:39.982 --> 00:33:42.503 align:middle line:84% And then here, 8 to 13 are all correlated with each other. 00:33:42.503 --> 00:33:44.420 align:middle line:84% And that corresponds to there being two blocks 00:33:44.420 --> 00:33:47.420 align:middle line:90% of linkage disequilibrium. 00:33:47.420 --> 00:33:50.870 align:middle line:84% And again, from the Gabriel, et al. paper, if you look, 00:33:50.870 --> 00:33:53.490 align:middle line:84% just concentrate on the figure on the right, 00:33:53.490 --> 00:33:59.030 align:middle line:84% in, for example, a European population or Asian population, 00:33:59.030 --> 00:34:03.440 align:middle line:84% over almost 40% of the genome is found in blocks of over 50 kb, 00:34:03.440 --> 00:34:06.740 align:middle line:84% and the average block size is about 22 kb. 00:34:06.740 --> 00:34:09.199 align:middle line:84% So on average, if you look in the genome, 00:34:09.199 --> 00:34:11.780 align:middle line:84% there's this correlation that extends for tens of thousands 00:34:11.780 --> 00:34:14.120 align:middle line:84% of bases on the order of the size of a gene, 00:34:14.120 --> 00:34:17.193 align:middle line:84% although it doesn't necessarily correlate perfectly with genes. 00:34:17.193 --> 00:34:18.860 align:middle line:84% And within those blocks, as I showed you 00:34:18.860 --> 00:34:20.527 align:middle line:84% with that red, blue, and yellow pattern, 00:34:20.527 --> 00:34:24.029 align:middle line:84% there's only a few common haplotypes present, 00:34:24.029 --> 00:34:27.737 align:middle line:84% so in the order of three to four for non-African populations 00:34:27.737 --> 00:34:29.570 align:middle line:84% and one or two more for African populations. 00:34:29.570 --> 00:34:31.219 align:middle line:84% And they explain the vast majority 00:34:31.219 --> 00:34:34.670 align:middle line:84% of the chromosomes in that population. 00:34:34.670 --> 00:34:36.800 align:middle line:84% And furthermore, the haplotypes themselves 00:34:36.800 --> 00:34:38.750 align:middle line:84% are generally shared, and Africans 00:34:38.750 --> 00:34:39.750 align:middle line:90% have a little bit more. 00:34:39.750 --> 00:34:42.409 align:middle line:84% And this is because when humans emerged from Africa, 00:34:42.409 --> 00:34:45.560 align:middle line:84% only a subset of the geniculate diversity went with them. 00:34:45.560 --> 00:34:48.020 align:middle line:84% So Africa is actually a more diverse continent 00:34:48.020 --> 00:34:51.219 align:middle line:90% than the rest of the world is. 00:34:51.219 --> 00:34:53.110 align:middle line:84% And I won't go through the reasons 00:34:53.110 --> 00:34:54.630 align:middle line:84% that blocks exist, but basically, it 00:34:54.630 --> 00:34:57.430 align:middle line:84% has to do with the nature of recombination, the fact 00:34:57.430 --> 00:34:59.830 align:middle line:84% that there are hotspots of recombination 00:34:59.830 --> 00:35:02.350 align:middle line:90% and this emergence-- 00:35:02.350 --> 00:35:04.900 align:middle line:90% this out of Africa pattern. 00:35:04.900 --> 00:35:08.350 align:middle line:84% And so that's led to a proposal that basically, 00:35:08.350 --> 00:35:12.340 align:middle line:84% if you can recognize these red, blue, and yellow patterns, 00:35:12.340 --> 00:35:14.980 align:middle line:84% you can then identify a few tagged 00:35:14.980 --> 00:35:18.530 align:middle line:84% SNPs, in this case, two, which distinguish those patterns. 00:35:18.530 --> 00:35:20.670 align:middle line:84% So if have the red variant, you're-- 00:35:20.670 --> 00:35:23.198 align:middle line:84% if you have the variant for tag SNP 1, you're red, 00:35:23.198 --> 00:35:25.490 align:middle line:84% if you have the variant for tag SNP two, you're yellow, 00:35:25.490 --> 00:35:27.032 align:middle line:84% and if you don't have either variant, 00:35:27.032 --> 00:35:28.730 align:middle line:90% you must have been blue. 00:35:28.730 --> 00:35:32.800 align:middle line:84% And so this means that if you can type enough markers 00:35:32.800 --> 00:35:34.690 align:middle line:84% to distinguish these patterns, you can then 00:35:34.690 --> 00:35:35.800 align:middle line:90% identify these tag SNPs. 00:35:35.800 --> 00:35:37.390 align:middle line:84% And that greatly reduces the effort 00:35:37.390 --> 00:35:40.340 align:middle line:90% you need to cover a region. 00:35:40.340 --> 00:35:42.443 align:middle line:84% And so that's the goal of the human haplotype 00:35:42.443 --> 00:35:43.360 align:middle line:90% map, which is ongoing. 00:35:43.360 --> 00:35:46.000 align:middle line:84% It's a $100 million international project. 00:35:46.000 --> 00:35:48.940 align:middle line:84% It's basically to type about a million or more markers 00:35:48.940 --> 00:35:51.940 align:middle line:84% in three different reference populations-- 00:35:51.940 --> 00:35:55.030 align:middle line:84% European-American, West African, and then 00:35:55.030 --> 00:35:57.640 align:middle line:84% East Asian, which is Japanese and Chinese. 00:35:57.640 --> 00:35:59.710 align:middle line:84% And you can go to the hap map website 00:35:59.710 --> 00:36:02.170 align:middle line:90% or see this nature description. 00:36:02.170 --> 00:36:05.387 align:middle line:84% But its goal is to identify all these tagged SNPs. 00:36:05.387 --> 00:36:06.970 align:middle line:84% And so the approach would be basically 00:36:06.970 --> 00:36:11.110 align:middle line:84% take the SNPs from the database, genotype these SNPs 00:36:11.110 --> 00:36:14.883 align:middle line:84% in reference panels, measure the linkage disequilibrium, 00:36:14.883 --> 00:36:17.050 align:middle line:84% and then identify the tag SNPs, and then select them 00:36:17.050 --> 00:36:20.320 align:middle line:84% and then take those SNPs and type them in the population. 00:36:20.320 --> 00:36:22.840 align:middle line:90% 00:36:22.840 --> 00:36:25.810 align:middle line:84% So that's how one would set up, in 2004, 00:36:25.810 --> 00:36:29.365 align:middle line:84% to do an association study, where you pick genes. 00:36:29.365 --> 00:36:31.240 align:middle line:84% You would understand the haplotype structure. 00:36:31.240 --> 00:36:32.620 align:middle line:84% You might even do some resequencing 00:36:32.620 --> 00:36:34.180 align:middle line:84% to make sure you had all the missense variants. 00:36:34.180 --> 00:36:36.910 align:middle line:84% And then you would take some set, some subset of variants, 00:36:36.910 --> 00:36:41.358 align:middle line:84% and type them in populations and do your association study. 00:36:41.358 --> 00:36:43.900 align:middle line:84% The problem is that after you've done all that work to set up 00:36:43.900 --> 00:36:47.950 align:middle line:84% to do this, you do an association, you publish it, 00:36:47.950 --> 00:36:50.200 align:middle line:84% and you find that it's very difficult for anybody else 00:36:50.200 --> 00:36:51.310 align:middle line:90% to reproduce it. 00:36:51.310 --> 00:36:53.650 align:middle line:84% So that was a lot of work for nothing, right? 00:36:53.650 --> 00:36:55.703 align:middle line:84% So we wanted to understand, if we 00:36:55.703 --> 00:36:57.370 align:middle line:84% were going to try to use this as a tool, 00:36:57.370 --> 00:37:01.810 align:middle line:84% well, why are association studies not reproduced? 00:37:01.810 --> 00:37:06.170 align:middle line:84% And there's really three possible explanations. 00:37:06.170 --> 00:37:08.410 align:middle line:90% One is could be false positives. 00:37:08.410 --> 00:37:11.800 align:middle line:84% So when you did your association, it was just wrong. 00:37:11.800 --> 00:37:13.780 align:middle line:84% Maybe you've tested a lot of genes 00:37:13.780 --> 00:37:15.790 align:middle line:84% and tested a lot of variants and a lot of genes 00:37:15.790 --> 00:37:18.970 align:middle line:84% and by chance, you got a p value of 0.01. 00:37:18.970 --> 00:37:20.893 align:middle line:84% So you go ahead and you publish it. 00:37:20.893 --> 00:37:22.810 align:middle line:84% And it was just because you had flipped a coin 00:37:22.810 --> 00:37:24.643 align:middle line:84% a bunch of times, and you got a row of-- you 00:37:24.643 --> 00:37:27.160 align:middle line:90% got a run of heads. 00:37:27.160 --> 00:37:29.770 align:middle line:84% Doesn't mean the coin is two-headed. 00:37:29.770 --> 00:37:32.950 align:middle line:84% It could mean that, in fact, you were right, but then 00:37:32.950 --> 00:37:36.100 align:middle line:84% the people who came after you had little, small studies, 00:37:36.100 --> 00:37:38.750 align:middle line:84% and they were looking for a modest effect, 00:37:38.750 --> 00:37:41.080 align:middle line:84% and they just didn't have enough power. 00:37:41.080 --> 00:37:43.300 align:middle line:84% So it actually was-- the coin was unfair. 00:37:43.300 --> 00:37:45.610 align:middle line:84% So it does come up heads more often than tails, 00:37:45.610 --> 00:37:48.363 align:middle line:84% but they only flipped it five times, and it came up tails-- 00:37:48.363 --> 00:37:49.780 align:middle line:84% or they only flipped it six times, 00:37:49.780 --> 00:37:52.072 align:middle line:84% and it came up heads three times and tails three times. 00:37:52.072 --> 00:37:55.330 align:middle line:84% This is equivalent doing a study with not very many people. 00:37:55.330 --> 00:37:56.800 align:middle line:90% So that's just a lack of power. 00:37:56.800 --> 00:37:59.510 align:middle line:84% And then there could be true differences between populations 00:37:59.510 --> 00:38:01.390 align:middle line:84% so that the marker really is associated 00:38:01.390 --> 00:38:03.765 align:middle line:84% with disease in your population but not in the population 00:38:03.765 --> 00:38:06.080 align:middle line:90% that somebody else studied. 00:38:06.080 --> 00:38:10.100 align:middle line:84% So we set out to overview the association study literature 00:38:10.100 --> 00:38:13.250 align:middle line:84% and say, well, what explanations are actually going on here? 00:38:13.250 --> 00:38:17.280 align:middle line:90% 00:38:17.280 --> 00:38:21.770 align:middle line:84% And we identified, as of 2000 or 2001, 00:38:21.770 --> 00:38:24.450 align:middle line:84% just over 600 associations between common variants 00:38:24.450 --> 00:38:24.950 align:middle line:90% and disease. 00:38:24.950 --> 00:38:28.460 align:middle line:84% And there are way more than this now. 00:38:28.460 --> 00:38:31.250 align:middle line:84% There were 166 that have been studied at least three times, 00:38:31.250 --> 00:38:33.710 align:middle line:84% so we get to get a handle on their reproducibility. 00:38:33.710 --> 00:38:37.280 align:middle line:84% And the first thing we noticed, of those 166, only 6 00:38:37.280 --> 00:38:39.260 align:middle line:84% were actually highly consistently reproducible. 00:38:39.260 --> 00:38:42.090 align:middle line:84% So only 6 were seen 3/4 of the time. 00:38:42.090 --> 00:38:45.080 align:middle line:84% So you could say, well, this is just terrible. 00:38:45.080 --> 00:38:49.700 align:middle line:84% Almost all associations just are not really reproducible. 00:38:49.700 --> 00:38:52.010 align:middle line:84% Or you could say, well, actually, 00:38:52.010 --> 00:38:53.840 align:middle line:90% you've only tested 166 things. 00:38:53.840 --> 00:38:55.010 align:middle line:90% You found 6 real ones. 00:38:55.010 --> 00:38:56.920 align:middle line:90% That's pretty good. 00:38:56.920 --> 00:39:00.210 align:middle line:84% There are 10 million things to test out there. 00:39:00.210 --> 00:39:03.110 align:middle line:90% So these are the 6. 00:39:03.110 --> 00:39:05.570 align:middle line:84% And one of them is this paradigm that I showed you, APOE4. 00:39:05.570 --> 00:39:08.560 align:middle line:84% Then there's things like CCR5-delta32, the CTLA-4. 00:39:08.560 --> 00:39:10.310 align:middle line:84% Interesting, this is that missense variant 00:39:10.310 --> 00:39:11.510 align:middle line:84% that I told you about that actually now 00:39:11.510 --> 00:39:13.400 align:middle line:84% this should be corrected to the 3 prime UTR. 00:39:13.400 --> 00:39:15.483 align:middle line:84% So even though they didn't have the right variant, 00:39:15.483 --> 00:39:17.630 align:middle line:84% it was still pretty reproducible. 00:39:17.630 --> 00:39:20.240 align:middle line:84% Factor V in deep venous thrombosis and a couple 00:39:20.240 --> 00:39:20.810 align:middle line:90% of others. 00:39:20.810 --> 00:39:23.390 align:middle line:84% And there are more on this list now, although not many, 00:39:23.390 --> 00:39:26.930 align:middle line:84% that are that highly consistently reproducible. 00:39:26.930 --> 00:39:28.520 align:middle line:90% So what about the other 160? 00:39:28.520 --> 00:39:31.640 align:middle line:84% Are they just complete nonsense, or what? 00:39:31.640 --> 00:39:33.620 align:middle line:84% Why are they-- why are they not reproducible? 00:39:33.620 --> 00:39:37.040 align:middle line:84% Well, the next thing we noticed was that of the 160, 91 of them 00:39:37.040 --> 00:39:38.780 align:middle line:84% were actually seen multiple times, so 00:39:38.780 --> 00:39:40.488 align:middle line:84% not just by the person who reported them, 00:39:40.488 --> 00:39:42.300 align:middle line:90% but by somebody else as well. 00:39:42.300 --> 00:39:44.372 align:middle line:84% And so that suggested that maybe there 00:39:44.372 --> 00:39:45.830 align:middle line:84% was something real going on, but we 00:39:45.830 --> 00:39:48.570 align:middle line:84% wanted to look at that in a more formal way. 00:39:48.570 --> 00:39:54.170 align:middle line:84% And so to do that, we basically picked 25 of these 160 00:39:54.170 --> 00:39:57.740 align:middle line:84% associations for some diseases we cared about and some others 00:39:57.740 --> 00:39:59.190 align:middle line:90% at random. 00:39:59.190 --> 00:40:01.310 align:middle line:84% And we got rid of the first person to report it. 00:40:01.310 --> 00:40:02.810 align:middle line:84% And we said, of all the other people 00:40:02.810 --> 00:40:06.530 align:middle line:84% who tried to replicate it, how well did they do? 00:40:06.530 --> 00:40:09.860 align:middle line:84% So there were 301 studies for these 25 associations. 00:40:09.860 --> 00:40:13.168 align:middle line:84% And if all of those 25 associations were incorrect, 00:40:13.168 --> 00:40:14.960 align:middle line:84% we would expect that the follow-up studies, 00:40:14.960 --> 00:40:16.910 align:middle line:84% in theory or naively, should have-- 00:40:16.910 --> 00:40:20.330 align:middle line:84% 5% of them should have, again, been statistically significant 00:40:20.330 --> 00:40:21.664 align:middle line:90% with a p value of 0.05. 00:40:21.664 --> 00:40:25.370 align:middle line:84% 1% should have had a p value of 0.01, et cetera, et cetera. 00:40:25.370 --> 00:40:27.300 align:middle line:90% So what did we see? 00:40:27.300 --> 00:40:29.630 align:middle line:84% We actually found that about a fifth of them 00:40:29.630 --> 00:40:31.182 align:middle line:90% had P values of less than 0.05. 00:40:31.182 --> 00:40:33.140 align:middle line:84% So this is way, way more than you would expect. 00:40:33.140 --> 00:40:35.630 align:middle line:90% This is not just by chance. 00:40:35.630 --> 00:40:38.055 align:middle line:84% And encouragingly, most of the associations 00:40:38.055 --> 00:40:39.180 align:middle line:90% were in the same direction. 00:40:39.180 --> 00:40:41.090 align:middle line:84% So it was the same allele that was associated with disease 00:40:41.090 --> 00:40:42.440 align:middle line:90% as was in the original report. 00:40:42.440 --> 00:40:45.270 align:middle line:84% So it's actually-- that would only happen not 1 in 20 times 00:40:45.270 --> 00:40:47.360 align:middle line:90% but 1 in 40 times. 00:40:47.360 --> 00:40:50.420 align:middle line:84% And furthermore, it wasn't that each of the 25 00:40:50.420 --> 00:40:51.470 align:middle line:90% had a few replications. 00:40:51.470 --> 00:40:53.690 align:middle line:84% It was that there was a subfraction of them that 00:40:53.690 --> 00:40:55.857 align:middle line:84% had a lot of replications, and then the rest of them 00:40:55.857 --> 00:40:58.850 align:middle line:90% were never seen again. 00:40:58.850 --> 00:41:01.930 align:middle line:84% So you might say, could this be publication bias? 00:41:01.930 --> 00:41:05.210 align:middle line:84% So we obviously could only look at the reports 00:41:05.210 --> 00:41:06.060 align:middle line:90% that were published. 00:41:06.060 --> 00:41:07.220 align:middle line:84% So it might be that there were lots 00:41:07.220 --> 00:41:09.000 align:middle line:84% of people doing these studies out there, 00:41:09.000 --> 00:41:10.580 align:middle line:84% and when they got a P value of less than 0.05, 00:41:10.580 --> 00:41:12.350 align:middle line:84% they broke out the champagne, celebrated, 00:41:12.350 --> 00:41:15.740 align:middle line:84% and sent the journal-- sent the report off to the journal. 00:41:15.740 --> 00:41:19.190 align:middle line:84% But maybe if they did it and the p value wasn't over 0.05, 00:41:19.190 --> 00:41:20.540 align:middle line:90% it lingered in the desk drawer. 00:41:20.540 --> 00:41:22.665 align:middle line:84% And that's actually the formal name for this thing. 00:41:22.665 --> 00:41:24.530 align:middle line:90% It's the desk drawer phenomenon. 00:41:24.530 --> 00:41:26.510 align:middle line:84% And so what we asked was, could there 00:41:26.510 --> 00:41:28.910 align:middle line:84% be a universe of unpublished studies lingering 00:41:28.910 --> 00:41:30.950 align:middle line:84% in desk drawers around the world that 00:41:30.950 --> 00:41:34.340 align:middle line:84% would explain why we saw so many of these studies? 00:41:34.340 --> 00:41:36.530 align:middle line:84% How many unpublished studies would we 00:41:36.530 --> 00:41:38.780 align:middle line:84% have to hypothesize per association 00:41:38.780 --> 00:41:40.460 align:middle line:84% to explain the fraction that we saw? 00:41:40.460 --> 00:41:44.840 align:middle line:84% Was it one or two or three, or was it just a ludicrous number 00:41:44.840 --> 00:41:46.677 align:middle line:90% of studies out there? 00:41:46.677 --> 00:41:48.260 align:middle line:84% And it turned out, actually, you would 00:41:48.260 --> 00:41:51.110 align:middle line:84% have to postulate a ludicrous number of studies 00:41:51.110 --> 00:41:52.520 align:middle line:90% to explain the positives we saw. 00:41:52.520 --> 00:41:54.230 align:middle line:84% You would have to explain possibly 00:41:54.230 --> 00:41:57.325 align:middle line:84% about 40 to 80 unpublished negative studies lingering 00:41:57.325 --> 00:41:58.700 align:middle line:84% in desk drawers around the world. 00:41:58.700 --> 00:42:00.492 align:middle line:84% And there just aren't even that many people 00:42:00.492 --> 00:42:02.929 align:middle line:84% doing association studies for any of these diseases. 00:42:02.929 --> 00:42:05.262 align:middle line:84% So we think this is-- and we did some other work to show 00:42:05.262 --> 00:42:08.510 align:middle line:90% this is not publication bias. 00:42:08.510 --> 00:42:11.010 align:middle line:84% So you asked about ethnic admixture stratification. 00:42:11.010 --> 00:42:12.440 align:middle line:84% So I want to talk about that for a little bit. 00:42:12.440 --> 00:42:13.370 align:middle line:90% AUDIENCE: I have one question. 00:42:13.370 --> 00:42:13.790 align:middle line:90% JOEL HIRSCHHORN: Yeah, sure. 00:42:13.790 --> 00:42:14.450 align:middle line:90% Go right ahead. 00:42:14.450 --> 00:42:15.425 align:middle line:90% Yeah. 00:42:15.425 --> 00:42:21.063 align:middle line:84% AUDIENCE: [INAUDIBLE] a rare disease. 00:42:21.063 --> 00:42:21.980 align:middle line:90% JOEL HIRSCHHORN: Yeah. 00:42:21.980 --> 00:42:22.480 align:middle line:90% Yeah. 00:42:22.480 --> 00:42:24.318 align:middle line:84% So yeah, it's not a common disease. 00:42:24.318 --> 00:42:24.860 align:middle line:90% You're right. 00:42:24.860 --> 00:42:27.500 align:middle line:84% So it's a very environmental disease. 00:42:27.500 --> 00:42:31.220 align:middle line:84% So what it is is actually, it's probably a common disease, 00:42:31.220 --> 00:42:33.590 align:middle line:84% but just the environmental exposure that's required 00:42:33.590 --> 00:42:35.400 align:middle line:90% is rare. 00:42:35.400 --> 00:42:37.310 align:middle line:84% So it's like-- and actually, another one 00:42:37.310 --> 00:42:40.310 align:middle line:84% was HIV, which unfortunately is not a rare disease now 00:42:40.310 --> 00:42:42.350 align:middle line:84% because the environmental exposure is so common. 00:42:42.350 --> 00:42:45.830 align:middle line:84% But you could imagine, if the world were a different place, 00:42:45.830 --> 00:42:48.590 align:middle line:84% would be a rare disease and the same sort of thing. 00:42:48.590 --> 00:42:51.800 align:middle line:90% 00:42:51.800 --> 00:42:53.560 align:middle line:84% And it may be the same thing as diabetes. 00:42:53.560 --> 00:42:55.840 align:middle line:84% Type 2 diabetes may be a common disease in part 00:42:55.840 --> 00:42:57.340 align:middle line:84% because the environmental exposures. 00:42:57.340 --> 00:42:59.170 align:middle line:84% AUDIENCE: Like I said, I'm not sure how they can-- 00:42:59.170 --> 00:43:00.850 align:middle line:84% how can you get how you're looking-- 00:43:00.850 --> 00:43:03.040 align:middle line:84% JOEL HIRSCHHORN: Yeah, so it turns out it's actually 00:43:03.040 --> 00:43:04.340 align:middle line:90% a huge risk factor. 00:43:04.340 --> 00:43:06.132 align:middle line:90% It's a huge genetic risk factor. 00:43:06.132 --> 00:43:07.840 align:middle line:84% There's an interesting article in Science 00:43:07.840 --> 00:43:09.715 align:middle line:84% looking at this variant, showing that there's 00:43:09.715 --> 00:43:12.400 align:middle line:84% been waves of selection in favor of the resistance 00:43:12.400 --> 00:43:17.275 align:middle line:84% variant in tribes where prion diseases are endemic. 00:43:17.275 --> 00:43:19.150 align:middle line:84% AUDIENCE: It's just a pretty common sequence. 00:43:19.150 --> 00:43:20.067 align:middle line:90% JOEL HIRSCHHORN: Yeah. 00:43:20.067 --> 00:43:23.420 align:middle line:84% It's pretty common, and it turns out that it's a big effect. 00:43:23.420 --> 00:43:25.780 align:middle line:84% So that's how you get power so that even if-- 00:43:25.780 --> 00:43:28.810 align:middle line:84% and there are enough sporadic-- people collect a couple 00:43:28.810 --> 00:43:31.300 align:middle line:84% hundred sporadic cases here and there. 00:43:31.300 --> 00:43:33.250 align:middle line:84% And it's just a big effect, and that's 00:43:33.250 --> 00:43:34.850 align:middle line:90% why they were able to see it. 00:43:34.850 --> 00:43:36.910 align:middle line:84% So even rare diseases can have common variants 00:43:36.910 --> 00:43:39.520 align:middle line:90% that contribute to them. 00:43:39.520 --> 00:43:40.930 align:middle line:90% So ethnic admixture. 00:43:40.930 --> 00:43:43.310 align:middle line:90% What is that? 00:43:43.310 --> 00:43:47.870 align:middle line:84% So this is an ideal epidemiology study. 00:43:47.870 --> 00:43:50.770 align:middle line:84% So when people collect cases and controls, 00:43:50.770 --> 00:43:53.350 align:middle line:84% they often will ask about ethnicity. 00:43:53.350 --> 00:43:56.500 align:middle line:84% And they have what's called self-described ethnicity. 00:43:56.500 --> 00:43:59.260 align:middle line:84% And the ideal thing is that you would-- clearly, 00:43:59.260 --> 00:44:01.600 align:middle line:84% everybody matches on what people put down. 00:44:01.600 --> 00:44:04.660 align:middle line:84% But people will self-describe differently. 00:44:04.660 --> 00:44:07.240 align:middle line:84% And there may not be an accurate self-description 00:44:07.240 --> 00:44:12.190 align:middle line:84% for what people's actual ethnic and genetic backgrounds are. 00:44:12.190 --> 00:44:15.375 align:middle line:84% So you hope that somehow that this self-described ethnicity 00:44:15.375 --> 00:44:17.500 align:middle line:84% is good enough that you actually have a good match. 00:44:17.500 --> 00:44:19.870 align:middle line:84% But you can imagine that within some ethnic group, 00:44:19.870 --> 00:44:21.655 align:middle line:84% there might be a couple of subgroups. 00:44:21.655 --> 00:44:23.280 align:middle line:84% And this is obviously the distinctions. 00:44:23.280 --> 00:44:26.382 align:middle line:84% I've drawn this as there's the blue subgroup 00:44:26.382 --> 00:44:27.340 align:middle line:90% and the green subgroup. 00:44:27.340 --> 00:44:29.920 align:middle line:84% But in reality, would probably some shading 00:44:29.920 --> 00:44:34.430 align:middle line:84% from blue to blue-green to green-blue to green. 00:44:34.430 --> 00:44:36.370 align:middle line:84% So you can imagine that there might 00:44:36.370 --> 00:44:38.317 align:middle line:84% be some different ethnic subgroups. 00:44:38.317 --> 00:44:40.150 align:middle line:84% And you could think about this, for example. 00:44:40.150 --> 00:44:42.525 align:middle line:84% European might be Northern European and Southern European 00:44:42.525 --> 00:44:44.600 align:middle line:90% or something like that. 00:44:44.600 --> 00:44:48.850 align:middle line:84% And there might be diseases where the disease actually 00:44:48.850 --> 00:44:51.850 align:middle line:84% is more prevalent, let's say, in northern Europe. 00:44:51.850 --> 00:44:54.160 align:middle line:84% So you might have more cases from 00:44:54.160 --> 00:44:56.590 align:middle line:84% the Northern European ancestry and more controls 00:44:56.590 --> 00:44:58.070 align:middle line:90% than Southern European ancestry. 00:44:58.070 --> 00:45:03.340 align:middle line:84% And for example, type 1 diabetes might fit this pattern. 00:45:03.340 --> 00:45:06.370 align:middle line:84% Much more common in Finland than in South of France, 00:45:06.370 --> 00:45:08.480 align:middle line:90% for example. 00:45:08.480 --> 00:45:14.020 align:middle line:84% So if you then had a marker that was tracked genetically 00:45:14.020 --> 00:45:18.220 align:middle line:84% with Northern European ancestry but not 00:45:18.220 --> 00:45:19.775 align:middle line:84% with Southern European ancestry-- 00:45:19.775 --> 00:45:21.650 align:middle line:84% and there are not a lot of markers like that, 00:45:21.650 --> 00:45:24.880 align:middle line:84% but there are at least one that I know of-- 00:45:24.880 --> 00:45:27.820 align:middle line:84% then you might find that that marker would be overrepresented 00:45:27.820 --> 00:45:29.950 align:middle line:84% in the cases relative to the controls 00:45:29.950 --> 00:45:33.820 align:middle line:84% just because that ethnic group was overrepresented. 00:45:33.820 --> 00:45:36.310 align:middle line:84% And this is thought to be more of a problem 00:45:36.310 --> 00:45:37.960 align:middle line:84% also in populations that have had 00:45:37.960 --> 00:45:40.330 align:middle line:84% recent admixture, like African-American populations 00:45:40.330 --> 00:45:45.970 align:middle line:84% or Latino populations, where instead of this being Northern 00:45:45.970 --> 00:45:50.900 align:middle line:84% European/Southern European, this might be degree of ancestry, 00:45:50.900 --> 00:45:53.740 align:middle line:84% so what fraction of your chromosomes 00:45:53.740 --> 00:45:56.260 align:middle line:84% have alleles that are more common in Europe versus alleles 00:45:56.260 --> 00:45:57.880 align:middle line:90% that are more common in Africa. 00:45:57.880 --> 00:46:00.070 align:middle line:84% And you can end up with this same setup. 00:46:00.070 --> 00:46:03.070 align:middle line:84% So you can get, in theory, false positive associations 00:46:03.070 --> 00:46:04.700 align:middle line:90% from that. 00:46:04.700 --> 00:46:07.610 align:middle line:90% So there are ways around it now. 00:46:07.610 --> 00:46:09.970 align:middle line:84% So the first way that was proposed by Rich Spielman 00:46:09.970 --> 00:46:12.345 align:middle line:84% and then a whole bunch of other people is to use what are 00:46:12.345 --> 00:46:14.140 align:middle line:90% called family test-based-- 00:46:14.140 --> 00:46:15.603 align:middle line:84% family-based tests of association. 00:46:15.603 --> 00:46:18.020 align:middle line:84% And I'm not going to go through the details of those other 00:46:18.020 --> 00:46:21.783 align:middle line:84% than to tell you that they are immune to this problem. 00:46:21.783 --> 00:46:23.200 align:middle line:84% And the other way is to basically, 00:46:23.200 --> 00:46:25.480 align:middle line:84% if you're doing-- if you can't do the family-based test, 00:46:25.480 --> 00:46:27.520 align:middle line:84% and there are a lot of reasons you might not be able to, 00:46:27.520 --> 00:46:29.770 align:middle line:84% you can use something called genomic control, where you type 00:46:29.770 --> 00:46:31.300 align:middle line:84% a lot of random markers, and you see, well, 00:46:31.300 --> 00:46:33.970 align:middle line:84% do those random markers give you these spurious associations? 00:46:33.970 --> 00:46:36.790 align:middle line:84% And you can use those to correct your association results 00:46:36.790 --> 00:46:39.370 align:middle line:84% or even to rematch your cases and controls to make sure 00:46:39.370 --> 00:46:41.025 align:middle line:90% that their genetic-- 00:46:41.025 --> 00:46:43.150 align:middle line:84% regardless of what the self-described ethnicity is, 00:46:43.150 --> 00:46:46.510 align:middle line:84% you can match them by their genetic background 00:46:46.510 --> 00:46:50.170 align:middle line:84% as assessed by these 100 markers. 00:46:50.170 --> 00:46:53.207 align:middle line:84% So we actually looked-- at least for our little study, 00:46:53.207 --> 00:46:54.790 align:middle line:84% most of the things that had replicated 00:46:54.790 --> 00:46:56.542 align:middle line:84% are actually seen in family-based studies 00:46:56.542 --> 00:46:58.000 align:middle line:84% or in multiple ethnic groups, which 00:46:58.000 --> 00:46:59.410 align:middle line:90% makes admixture less likely. 00:46:59.410 --> 00:47:02.350 align:middle line:84% But doesn't mean that this isn't going on in a rampant way. 00:47:02.350 --> 00:47:04.640 align:middle line:84% And for example, the ones that were never replicated, 00:47:04.640 --> 00:47:06.640 align:middle line:84% could well be that some of those false positives 00:47:06.640 --> 00:47:10.640 align:middle line:90% are due to admixture. 00:47:10.640 --> 00:47:14.960 align:middle line:84% Then there's true population differences 00:47:14.960 --> 00:47:18.283 align:middle line:90% that could be coming into play. 00:47:18.283 --> 00:47:19.700 align:middle line:84% There's a lot of different things. 00:47:19.700 --> 00:47:22.730 align:middle line:84% It could be different diets in different populations. 00:47:22.730 --> 00:47:24.950 align:middle line:84% It could be that there are other genes that 00:47:24.950 --> 00:47:27.112 align:middle line:84% are more common in one population than in another, 00:47:27.112 --> 00:47:28.820 align:middle line:84% and there's some interaction between what 00:47:28.820 --> 00:47:30.930 align:middle line:90% you're looking at and that. 00:47:30.930 --> 00:47:34.190 align:middle line:84% And I want to talk about one very specific type 00:47:34.190 --> 00:47:37.122 align:middle line:84% of difference that at least we could address, 00:47:37.122 --> 00:47:39.080 align:middle line:84% which is that there might be different patterns 00:47:39.080 --> 00:47:40.747 align:middle line:84% of correlation in different populations. 00:47:40.747 --> 00:47:43.460 align:middle line:84% And the marker that's being used might not be the causal marker. 00:47:43.460 --> 00:47:45.210 align:middle line:84% It might be something which is correlated. 00:47:45.210 --> 00:47:46.668 align:middle line:84% So I showed you this diagram, where 00:47:46.668 --> 00:47:48.710 align:middle line:90% you could use this CA haplot-- 00:47:48.710 --> 00:47:50.480 align:middle line:84% if you look at this two-marker haplotype, 00:47:50.480 --> 00:47:53.600 align:middle line:84% you use this CA haplotype to try to predict 00:47:53.600 --> 00:47:55.940 align:middle line:84% whether the causal polymorphism is there. 00:47:55.940 --> 00:47:57.740 align:middle line:84% And it does pretty well because there 00:47:57.740 --> 00:48:00.150 align:middle line:84% wasn't very much of this recombinant haplotype here. 00:48:00.150 --> 00:48:01.190 align:middle line:84% So here, actually, the prediction-- 00:48:01.190 --> 00:48:02.330 align:middle line:90% I haven't drawn the arrows. 00:48:02.330 --> 00:48:03.980 align:middle line:84% But the prediction would actually not 00:48:03.980 --> 00:48:05.390 align:middle line:84% be right for this one chromosome. 00:48:05.390 --> 00:48:07.860 align:middle line:84% But most of the time, the correlation is pretty good. 00:48:07.860 --> 00:48:10.277 align:middle line:84% But what if there was a lot of that recombinant chromosome 00:48:10.277 --> 00:48:11.360 align:middle line:90% in some other population? 00:48:11.360 --> 00:48:15.950 align:middle line:84% Now, all of a sudden, you have a common CA haplotype 00:48:15.950 --> 00:48:17.420 align:middle line:90% without the causal SNP. 00:48:17.420 --> 00:48:20.600 align:middle line:90% And so in addition to being-- 00:48:20.600 --> 00:48:23.070 align:middle line:84% the correlation is now no longer so good, 00:48:23.070 --> 00:48:24.780 align:middle line:84% so you might not see an association. 00:48:24.780 --> 00:48:27.802 align:middle line:84% So that's at least one possible thing to keep in mind 00:48:27.802 --> 00:48:29.510 align:middle line:84% is that the marker that you're looking at 00:48:29.510 --> 00:48:31.910 align:middle line:84% might not be the causal thing, so basically 00:48:31.910 --> 00:48:35.130 align:middle line:84% understanding the LD patterns around an association. 00:48:35.130 --> 00:48:36.770 align:middle line:84% So once an association has been found, 00:48:36.770 --> 00:48:39.260 align:middle line:84% typing lots of other markers in there and seeing, 00:48:39.260 --> 00:48:42.200 align:middle line:84% is there another marker that explains it better 00:48:42.200 --> 00:48:43.880 align:middle line:84% will become important, and almost nobody 00:48:43.880 --> 00:48:46.627 align:middle line:90% does that right now. 00:48:46.627 --> 00:48:48.710 align:middle line:84% And then finally, the last thing that we looked at 00:48:48.710 --> 00:48:50.780 align:middle line:84% was, could it be that some of the associations that we were 00:48:50.780 --> 00:48:52.940 align:middle line:84% seeing in our meta analysis are actually true, 00:48:52.940 --> 00:48:56.510 align:middle line:84% but the studies that said that there was no association 00:48:56.510 --> 00:48:59.100 align:middle line:90% were actually incorrect? 00:48:59.100 --> 00:49:01.490 align:middle line:84% And so to try to get a sense for this, what that requires 00:49:01.490 --> 00:49:03.560 align:middle line:84% is that the sample sizes were too 00:49:03.560 --> 00:49:05.540 align:middle line:84% small to detect the association, which 00:49:05.540 --> 00:49:07.520 align:middle line:84% means, in turn, that the association must 00:49:07.520 --> 00:49:11.160 align:middle line:84% have been a weak associations or modest effect. 00:49:11.160 --> 00:49:15.500 align:middle line:84% So we estimated the effects for all of these associations 00:49:15.500 --> 00:49:17.960 align:middle line:90% and by pooling all of the data. 00:49:17.960 --> 00:49:19.940 align:middle line:90% And so eight of them replicated. 00:49:19.940 --> 00:49:22.280 align:middle line:84% And the question was, what was the effect size? 00:49:22.280 --> 00:49:25.010 align:middle line:84% And basically all eight increased the risk of disease 00:49:25.010 --> 00:49:26.120 align:middle line:90% by less than twofold. 00:49:26.120 --> 00:49:28.490 align:middle line:84% And some of them was just a 10% increased risk 00:49:28.490 --> 00:49:31.580 align:middle line:84% of disease associated with having the allele. 00:49:31.580 --> 00:49:33.800 align:middle line:84% And for 10%, for example, you need thousands 00:49:33.800 --> 00:49:34.982 align:middle line:90% of cases and controls. 00:49:34.982 --> 00:49:37.190 align:middle line:84% And almost none of the studies we looked at had that. 00:49:37.190 --> 00:49:39.607 align:middle line:84% So almost all the studies were underpowered for almost all 00:49:39.607 --> 00:49:41.030 align:middle line:84% of the associations we looked at. 00:49:41.030 --> 00:49:44.060 align:middle line:84% So lack of power is rampant in this field. 00:49:44.060 --> 00:49:47.480 align:middle line:84% And a negative study by itself, unless it's 00:49:47.480 --> 00:49:50.000 align:middle line:84% got many thousands of people, really 00:49:50.000 --> 00:49:54.350 align:middle line:84% can't be taken to be worth much other than by itself. 00:49:54.350 --> 00:49:58.760 align:middle line:84% It can be worth a lot in context of other studies. 00:49:58.760 --> 00:50:01.462 align:middle line:84% One last thing that I want to just take you through again 00:50:01.462 --> 00:50:03.920 align:middle line:84% because it's an important issue thinking about interpreting 00:50:03.920 --> 00:50:07.190 align:middle line:84% the literature, when we did this estimation of what 00:50:07.190 --> 00:50:09.800 align:middle line:84% was the actual genetic effect, and then we compared those less 00:50:09.800 --> 00:50:13.670 align:middle line:84% than twofold risks to what the first report had claimed, 00:50:13.670 --> 00:50:15.800 align:middle line:84% almost always, the first positive report 00:50:15.800 --> 00:50:18.363 align:middle line:84% had overestimated the actual genetic effect. 00:50:18.363 --> 00:50:20.780 align:middle line:84% And we wanted to know if it's consistent with a phenomenon 00:50:20.780 --> 00:50:22.598 align:middle line:90% called the winner's curse. 00:50:22.598 --> 00:50:24.890 align:middle line:84% So winner's curse is best described for auction theory, 00:50:24.890 --> 00:50:27.720 align:middle line:90% and it goes like this. 00:50:27.720 --> 00:50:29.390 align:middle line:84% Imagine that there are a bunch of people 00:50:29.390 --> 00:50:31.370 align:middle line:90% who are bidding on an item. 00:50:31.370 --> 00:50:33.530 align:middle line:90% And they all have an accurate-- 00:50:33.530 --> 00:50:36.110 align:middle line:84% I'm sorry, yes, an unbiased estimate 00:50:36.110 --> 00:50:38.570 align:middle line:90% of the value of that item. 00:50:38.570 --> 00:50:40.320 align:middle line:90% But it's not that precise. 00:50:40.320 --> 00:50:43.520 align:middle line:84% So people are just as likely to overestimate its value 00:50:43.520 --> 00:50:45.360 align:middle line:90% as underestimate it. 00:50:45.360 --> 00:50:46.743 align:middle line:90% And they all place bids. 00:50:46.743 --> 00:50:48.410 align:middle line:84% And those unbiased bids will therefore-- 00:50:48.410 --> 00:50:51.200 align:middle line:84% and let's say they bid what they think that the item is actually 00:50:51.200 --> 00:50:52.820 align:middle line:90% worth. 00:50:52.820 --> 00:50:54.590 align:middle line:84% Those unbiased bids will fluctuate around 00:50:54.590 --> 00:50:56.010 align:middle line:90% the true value. 00:50:56.010 --> 00:50:58.640 align:middle line:84% And then one of those will have fluctuated up the most, 00:50:58.640 --> 00:51:00.350 align:middle line:84% and that will be the winning bid. 00:51:00.350 --> 00:51:04.460 align:middle line:84% So the winning bid, conditional on it being the winning bid, 00:51:04.460 --> 00:51:06.530 align:middle line:84% will almost always have overestimated the value. 00:51:06.530 --> 00:51:09.920 align:middle line:84% And the best description of this is a Samuelson article 00:51:09.920 --> 00:51:13.820 align:middle line:84% called, "I Won the Auction but Don't Want the Prize." 00:51:13.820 --> 00:51:17.780 align:middle line:90% So beware on eBay. 00:51:17.780 --> 00:51:22.328 align:middle line:84% But so in association studies, basically the winning bid 00:51:22.328 --> 00:51:23.870 align:middle line:84% is equivalent to the first person who 00:51:23.870 --> 00:51:26.360 align:middle line:84% gets an exciting enough result that they can publish it 00:51:26.360 --> 00:51:29.090 align:middle line:84% quickly and in a prominent journal and that sort of thing. 00:51:29.090 --> 00:51:33.110 align:middle line:84% And conditional on them having found something interesting, 00:51:33.110 --> 00:51:35.553 align:middle line:84% it's likely that they overestimated the true value. 00:51:35.553 --> 00:51:37.970 align:middle line:84% And the true value might either be a weak effect, a weaker 00:51:37.970 --> 00:51:39.710 align:middle line:84% effect, or actually no effect at all. 00:51:39.710 --> 00:51:44.010 align:middle line:90% And only time can tell. 00:51:44.010 --> 00:51:46.340 align:middle line:84% And so we found that it was consistent-- 00:51:46.340 --> 00:51:48.260 align:middle line:84% that the degree by which they overestimated 00:51:48.260 --> 00:51:52.460 align:middle line:84% was generally consistent with the winner's curse phenomenon. 00:51:52.460 --> 00:51:55.715 align:middle line:84% So this is the take home messages 00:51:55.715 --> 00:51:57.590 align:middle line:84% for reading the association study literature. 00:51:57.590 --> 00:52:00.490 align:middle line:84% So if you pick up a paper and it says 00:52:00.490 --> 00:52:03.700 align:middle line:84% that something is associated with a disease, 00:52:03.700 --> 00:52:06.550 align:middle line:84% there is a possibility that it will be true. 00:52:06.550 --> 00:52:09.220 align:middle line:84% But it's I would say well under 50%. 00:52:09.220 --> 00:52:14.260 align:middle line:84% It's not 0% either, but it's well under 50%. 00:52:14.260 --> 00:52:16.090 align:middle line:84% The genetic effects for associations 00:52:16.090 --> 00:52:17.440 align:middle line:90% are likely to be quite modest. 00:52:17.440 --> 00:52:21.980 align:middle line:84% So if somebody is claiming a fivefold effect 00:52:21.980 --> 00:52:25.390 align:middle line:84% and their confidence intervals go from 1.1 to 40, 00:52:25.390 --> 00:52:31.390 align:middle line:84% it's probably closer to 1.1, if not 1. 00:52:31.390 --> 00:52:33.640 align:middle line:84% And then because of these modest effects, 00:52:33.640 --> 00:52:35.860 align:middle line:84% you need large study sites to detect these reliably 00:52:35.860 --> 00:52:39.390 align:middle line:84% and also to narrow those confidence intervals down. 00:52:39.390 --> 00:52:41.920 align:middle line:84% Now, I'll just give you an example, our favorite example, 00:52:41.920 --> 00:52:44.380 align:middle line:84% for association of a missense polymorphism, 00:52:44.380 --> 00:52:46.780 align:middle line:84% or we think that's the causal polymorphism. 00:52:46.780 --> 00:52:48.910 align:middle line:84% In type 2 diabetes, there is a gene 00:52:48.910 --> 00:52:51.760 align:middle line:84% called PPAR, which is important for fat cell differentiation 00:52:51.760 --> 00:52:54.040 align:middle line:84% and is a nuclear hormone receptor. 00:52:54.040 --> 00:52:55.180 align:middle line:90% Probably binds fatty acids. 00:52:55.180 --> 00:52:58.160 align:middle line:90% 00:52:58.160 --> 00:53:03.410 align:middle line:84% So the first study that was published is right here. 00:53:03.410 --> 00:53:04.990 align:middle line:90% So each line is a study. 00:53:04.990 --> 00:53:09.250 align:middle line:84% The point is the point estimate for the effect on diabetes. 00:53:09.250 --> 00:53:10.750 align:middle line:84% And actually, in this case, anything 00:53:10.750 --> 00:53:15.940 align:middle line:84% to the left of this line is an association with diabetes. 00:53:15.940 --> 00:53:18.400 align:middle line:84% And the line around the point represents the confidence 00:53:18.400 --> 00:53:20.240 align:middle line:90% interval around that estimate. 00:53:20.240 --> 00:53:22.080 align:middle line:84% So for example, this study up here, 00:53:22.080 --> 00:53:24.580 align:middle line:84% even though it shows the-- it trends towards association was 00:53:24.580 --> 00:53:27.670 align:middle line:84% considered negative because this line crosses-- 00:53:27.670 --> 00:53:30.520 align:middle line:84% the 95% confidence will cross as 1. 00:53:30.520 --> 00:53:32.020 align:middle line:84% So this is the first study which you 00:53:32.020 --> 00:53:34.930 align:middle line:84% can see the winner's curse in action here. 00:53:34.930 --> 00:53:39.700 align:middle line:84% Overestimates is the strongest estimate of risk. 00:53:39.700 --> 00:53:42.970 align:middle line:84% And it was followed by, I believe, these three 00:53:42.970 --> 00:53:45.010 align:middle line:84% and a bunch of other studies, all of which 00:53:45.010 --> 00:53:47.710 align:middle line:84% were considered to be negative because their 95% confidence 00:53:47.710 --> 00:53:49.990 align:middle line:90% intervals crossed 1. 00:53:49.990 --> 00:53:52.583 align:middle line:84% We then did a pretty big study where we showed an association. 00:53:52.583 --> 00:53:55.000 align:middle line:84% We actually pointed out that all of the literature to date 00:53:55.000 --> 00:53:57.550 align:middle line:84% was consistent with an association somewhere 00:53:57.550 --> 00:53:58.270 align:middle line:90% in this range. 00:53:58.270 --> 00:54:00.220 align:middle line:84% Much weaker than originally described, 00:54:00.220 --> 00:54:03.940 align:middle line:84% but modest enough that these studies didn't have power 00:54:03.940 --> 00:54:04.600 align:middle line:90% to pick it up. 00:54:04.600 --> 00:54:07.660 align:middle line:84% And in fact, as you get larger and larger samples, what 00:54:07.660 --> 00:54:10.510 align:middle line:84% you can see is that the confidence intervals-- 00:54:10.510 --> 00:54:12.333 align:middle line:84% is that things tend to really focus in. 00:54:12.333 --> 00:54:13.750 align:middle line:84% And you put all the data together, 00:54:13.750 --> 00:54:14.980 align:middle line:84% you get a pretty narrow confidence interval. 00:54:14.980 --> 00:54:16.810 align:middle line:84% And now, the p value for association 00:54:16.810 --> 00:54:18.130 align:middle line:90% is 10 to the minus 9. 00:54:18.130 --> 00:54:21.820 align:middle line:84% So it's real, but it took 20,000 alleles 00:54:21.820 --> 00:54:25.660 align:middle line:84% from people with diabetes and controls to get there. 00:54:25.660 --> 00:54:29.290 align:middle line:84% AUDIENCE: Are all the studies in the same [INAUDIBLE].. 00:54:29.290 --> 00:54:32.455 align:middle line:84% Can you really put them all together and do the overall? 00:54:32.455 --> 00:54:33.830 align:middle line:84% JOEL HIRSCHHORN: So for diabetes, 00:54:33.830 --> 00:54:35.980 align:middle line:84% the definition of diabetes is pretty clear. 00:54:35.980 --> 00:54:37.600 align:middle line:90% There's a WHO consensus. 00:54:37.600 --> 00:54:40.840 align:middle line:84% And they're more or less the same definition. 00:54:40.840 --> 00:54:42.640 align:middle line:90% I would say close enough. 00:54:42.640 --> 00:54:45.250 align:middle line:84% The allele itself is clearly the same variant. 00:54:45.250 --> 00:54:50.980 align:middle line:84% And assuming that the quality of genotyping was good, 00:54:50.980 --> 00:54:53.512 align:middle line:84% probably they were looking at the same thing there. 00:54:53.512 --> 00:54:55.220 align:middle line:84% Obviously, the populations are different. 00:54:55.220 --> 00:54:56.990 align:middle line:84% So I talked about population differences. 00:54:56.990 --> 00:55:01.420 align:middle line:84% So it turns out, also one of the best examples of heterogeneity 00:55:01.420 --> 00:55:03.627 align:middle line:84% between populations is for this polymorphism, 00:55:03.627 --> 00:55:05.710 align:middle line:84% not so much in diabetes, although it hasn't really 00:55:05.710 --> 00:55:07.820 align:middle line:84% been looked at well, but actually, obesity. 00:55:07.820 --> 00:55:09.695 align:middle line:84% So this has been looked at a lot for obesity, 00:55:09.695 --> 00:55:11.420 align:middle line:90% and wildly divergent results. 00:55:11.420 --> 00:55:15.610 align:middle line:84% It turns out that how much trans fatty acids you eat, 00:55:15.610 --> 00:55:18.430 align:middle line:84% or saturated to unsaturated fatty acids, 00:55:18.430 --> 00:55:21.160 align:middle line:84% affects whether this polymorphism has 00:55:21.160 --> 00:55:22.840 align:middle line:90% an effect on obesity. 00:55:22.840 --> 00:55:24.790 align:middle line:84% I can't remember right now which way it goes. 00:55:24.790 --> 00:55:28.540 align:middle line:84% Basically, on one diet, which is much more 00:55:28.540 --> 00:55:31.895 align:middle line:84% common in Western countries, there's I think no effect. 00:55:31.895 --> 00:55:33.520 align:middle line:84% And on another diet, which is much more 00:55:33.520 --> 00:55:36.640 align:middle line:84% common in Asian countries, there is an effect on obesity. 00:55:36.640 --> 00:55:40.060 align:middle line:84% So you can imagine that there- and actually, if you look, 00:55:40.060 --> 00:55:45.230 align:middle line:84% it turns out that these three are all in Asian populations. 00:55:45.230 --> 00:55:47.470 align:middle line:84% So it could be that there is some heterogeneity going 00:55:47.470 --> 00:55:50.843 align:middle line:84% on there, although it's statistically marginally 00:55:50.843 --> 00:55:52.510 align:middle line:84% significant whether any of these studies 00:55:52.510 --> 00:55:55.010 align:middle line:84% are different than any of the others. 00:55:55.010 --> 00:55:57.880 align:middle line:84% So these are mostly consistent with if you 00:55:57.880 --> 00:56:03.075 align:middle line:84% took darts and threw them all at the same odds ratio, 00:56:03.075 --> 00:56:04.450 align:middle line:84% you might end up with points that 00:56:04.450 --> 00:56:06.770 align:middle line:90% looked more or less like that. 00:56:06.770 --> 00:56:12.290 align:middle line:84% So but it could well be that there are differences. 00:56:12.290 --> 00:56:14.810 align:middle line:84% For things like-- for less well-- so for Alzheimer's 00:56:14.810 --> 00:56:16.850 align:middle line:84% disease, there's a little more debate sometimes 00:56:16.850 --> 00:56:18.872 align:middle line:84% of whether it's pathology-confirmed or not 00:56:18.872 --> 00:56:19.830 align:middle line:90% and that sort of thing. 00:56:19.830 --> 00:56:22.070 align:middle line:90% So all right. 00:56:22.070 --> 00:56:23.698 align:middle line:84% So what to believe is the question. 00:56:23.698 --> 00:56:24.740 align:middle line:90% So should you believe it? 00:56:24.740 --> 00:56:26.573 align:middle line:84% I would say initial skepticism is warranted. 00:56:26.573 --> 00:56:28.550 align:middle line:84% Replication, especially with low p values-- 00:56:28.550 --> 00:56:31.400 align:middle line:84% and maybe we can discuss what that means-- is encouraging. 00:56:31.400 --> 00:56:32.850 align:middle line:90% Large sample sizes are crucial. 00:56:32.850 --> 00:56:35.180 align:middle line:84% So if you're going to take anything home, there you go. 00:56:35.180 --> 00:56:37.700 align:middle line:90% 00:56:37.700 --> 00:56:42.340 align:middle line:84% So I want to quickly-- so this is a quantitatively-oriented 00:56:42.340 --> 00:56:43.340 align:middle line:90% class-- 00:56:43.340 --> 00:56:45.460 align:middle line:84% talk about applying Bayes' theorem to interpreting 00:56:45.460 --> 00:56:46.610 align:middle line:90% association studies. 00:56:46.610 --> 00:56:51.040 align:middle line:84% So Bayes' theorem is one of the fundamental theorems 00:56:51.040 --> 00:56:52.880 align:middle line:90% of probability. 00:56:52.880 --> 00:56:54.700 align:middle line:84% And in this case, what I've done is 00:56:54.700 --> 00:56:59.050 align:middle line:84% I've substituted in the words causal 00:56:59.050 --> 00:57:02.573 align:middle line:84% for probability of observing association. 00:57:02.573 --> 00:57:04.990 align:middle line:84% And causal is the probability that the variant that you're 00:57:04.990 --> 00:57:07.160 align:middle line:90% looking at actually is causal. 00:57:07.160 --> 00:57:09.490 align:middle line:84% So if you go through just plugging it 00:57:09.490 --> 00:57:11.790 align:middle line:84% into Bayes' theorem, this is what you get. 00:57:11.790 --> 00:57:13.540 align:middle line:84% And what you're interested in is basically 00:57:13.540 --> 00:57:15.560 align:middle line:84% given that I've seen an association, what's 00:57:15.560 --> 00:57:16.810 align:middle line:90% the likelihood that it's true? 00:57:16.810 --> 00:57:20.210 align:middle line:84% That's what you want to know when you pick up the paper. 00:57:20.210 --> 00:57:21.770 align:middle line:84% So this has to do with three terms. 00:57:21.770 --> 00:57:24.040 align:middle line:84% One is the probability that something is associated 00:57:24.040 --> 00:57:25.280 align:middle line:90% given that it was causal. 00:57:25.280 --> 00:57:26.830 align:middle line:84% So that's the power of the study. 00:57:26.830 --> 00:57:31.085 align:middle line:84% So would you see an association if it were causal? 00:57:31.085 --> 00:57:32.710 align:middle line:84% Would I have seen this association data 00:57:32.710 --> 00:57:33.650 align:middle line:90% if it were not causal? 00:57:33.650 --> 00:57:35.900 align:middle line:84% Well, that's the p value, the probability of observing 00:57:35.900 --> 00:57:37.930 align:middle line:90% the data by chance. 00:57:37.930 --> 00:57:40.277 align:middle line:84% And then the probability that it's causal to begin with, 00:57:40.277 --> 00:57:41.610 align:middle line:90% which is your prior probability. 00:57:41.610 --> 00:57:44.020 align:middle line:84% So this is the key to Bayes' theorem, which 00:57:44.020 --> 00:57:47.280 align:middle line:84% is that you have to specify the priors, which is, of course, 00:57:47.280 --> 00:57:48.280 align:middle line:90% always the hardest part. 00:57:48.280 --> 00:57:49.660 align:middle line:84% It's very easy to write down Bayes' theorem, 00:57:49.660 --> 00:57:51.452 align:middle line:84% but then specifying your prior distribution 00:57:51.452 --> 00:57:53.030 align:middle line:90% can sometimes be quite tricky. 00:57:53.030 --> 00:57:53.530 align:middle line:90% All right. 00:57:53.530 --> 00:57:55.072 align:middle line:84% So I'm going to take you through what 00:57:55.072 --> 00:57:57.250 align:middle line:84% I believe prior distributions look 00:57:57.250 --> 00:58:00.920 align:middle line:84% like within a couple of orders of magnitude. 00:58:00.920 --> 00:58:04.760 align:middle line:84% So what are the prior probabilities? 00:58:04.760 --> 00:58:07.672 align:middle line:84% So I'm going to say there are about 600,000 00:58:07.672 --> 00:58:09.130 align:middle line:84% independent common variants, and it 00:58:09.130 --> 00:58:11.320 align:middle line:84% has to do with the degree of redundancy that's 00:58:11.320 --> 00:58:13.445 align:middle line:84% been determined by the hap map and stuff like that. 00:58:13.445 --> 00:58:14.705 align:middle line:90% There are 10 million in total. 00:58:14.705 --> 00:58:16.330 align:middle line:84% Some of them aren't that common, so I'm 00:58:16.330 --> 00:58:19.460 align:middle line:84% going to shove those under the rug for a little bit. 00:58:19.460 --> 00:58:21.430 align:middle line:84% And of the rest, there's enough correlation 00:58:21.430 --> 00:58:23.972 align:middle line:84% that I'm going to say there's about 600,000 variants that you 00:58:23.972 --> 00:58:24.753 align:middle line:90% can type. 00:58:24.753 --> 00:58:26.920 align:middle line:84% The other assumption is that at least a few of those 00:58:26.920 --> 00:58:27.740 align:middle line:90% are going to be causal. 00:58:27.740 --> 00:58:29.198 align:middle line:84% So I'm assuming the common variants 00:58:29.198 --> 00:58:31.510 align:middle line:90% play some role in disease. 00:58:31.510 --> 00:58:34.390 align:middle line:84% So I'm going to say somewhere between 6 and 60. 00:58:34.390 --> 00:58:38.555 align:middle line:84% It may be more, but I would say causal to a degree 00:58:38.555 --> 00:58:40.180 align:middle line:84% where you have any power to detect them 00:58:40.180 --> 00:58:41.500 align:middle line:84% at all because obviously, if they have no power 00:58:41.500 --> 00:58:43.000 align:middle line:84% to detect them, they may as well not 00:58:43.000 --> 00:58:45.190 align:middle line:84% exist from the point of view of this study 00:58:45.190 --> 00:58:47.590 align:middle line:84% because if you plug in 0 for your power, 00:58:47.590 --> 00:58:51.790 align:middle line:84% then everything starts canceling out. 00:58:51.790 --> 00:58:55.000 align:middle line:84% So your prior probability is 1 in 10,000 to 1 00:58:55.000 --> 00:58:57.427 align:middle line:90% in 100,000 more or less. 00:58:57.427 --> 00:58:58.510 align:middle line:90% How about candidate genes? 00:58:58.510 --> 00:59:00.153 align:middle line:84% Let's say, well, I'm not studying-- 00:59:00.153 --> 00:59:01.570 align:middle line:84% you say this paper is not studying 00:59:01.570 --> 00:59:04.090 align:middle line:90% some stinking random variant. 00:59:04.090 --> 00:59:07.010 align:middle line:84% This is the greatest candidate gene in the whole world. 00:59:07.010 --> 00:59:09.455 align:middle line:84% Well, so we've gone through the exercise of writing down 00:59:09.455 --> 00:59:11.830 align:middle line:84% all the candidate genes that we think are great candidate 00:59:11.830 --> 00:59:12.760 align:middle line:90% genes for diseases. 00:59:12.760 --> 00:59:15.400 align:middle line:84% And you usually come up with somewhere between, say, 00:59:15.400 --> 00:59:19.330 align:middle line:84% 100 and 500 genes, so let's say 300 that you come up with. 00:59:19.330 --> 00:59:21.010 align:middle line:84% And I would argue that you could tell 00:59:21.010 --> 00:59:23.980 align:middle line:84% as equally beautiful a story about almost all 300 00:59:23.980 --> 00:59:26.110 align:middle line:84% of those genes once you knew post hoc that there 00:59:26.110 --> 00:59:27.970 align:middle line:90% was an association. 00:59:27.970 --> 00:59:29.690 align:middle line:90% So you can say, oh, of course. 00:59:29.690 --> 00:59:33.610 align:middle line:84% That's why protein kinase C theta is the key to diabetes. 00:59:33.610 --> 00:59:34.870 align:middle line:90% And you write your story. 00:59:34.870 --> 00:59:36.287 align:middle line:84% So I'm going to say that there are 00:59:36.287 --> 00:59:38.560 align:middle line:84% 300 candidate genes, all of them about equally 00:59:38.560 --> 00:59:40.390 align:middle line:90% likely to be associated. 00:59:40.390 --> 00:59:42.040 align:middle line:84% Again, using some things about there 00:59:42.040 --> 00:59:45.027 align:middle line:84% are about three or four blocks of linkage disequilibrium 00:59:45.027 --> 00:59:47.110 align:middle line:84% per gene, about three or four haplotypes per gene. 00:59:47.110 --> 00:59:49.443 align:middle line:84% I'm going to say there are about 12 independent variants 00:59:49.443 --> 00:59:50.380 align:middle line:90% per gene. 00:59:50.380 --> 00:59:52.510 align:middle line:84% So there's about 3,600 candidate variants 00:59:52.510 --> 00:59:55.610 align:middle line:84% on your list of 300 genes that you've just made. 00:59:55.610 --> 00:59:57.940 align:middle line:84% Let's assume-- and this is I think somewhat generous, 00:59:57.940 --> 01:00:00.760 align:middle line:84% but let's assume that all that biology that went into making 01:00:00.760 --> 01:00:05.005 align:middle line:84% that list is worthwhile and that half of all the causal variants 01:00:05.005 --> 01:00:06.130 align:middle line:90% are in the candidate genes. 01:00:06.130 --> 01:00:08.113 align:middle line:84% Well, then if you do all the multiplication, 01:00:08.113 --> 01:00:10.030 align:middle line:84% the prior probability is about 10 times higher 01:00:10.030 --> 01:00:12.250 align:middle line:84% for a candidate gene as it is for a random gene, 1 01:00:12.250 --> 01:00:16.360 align:middle line:90% in 100 to 1 in 1,000. 01:00:16.360 --> 01:00:19.622 align:middle line:84% And then what are the probabilities you suggested 01:00:19.622 --> 01:00:20.830 align:middle line:90% looking under a linkage peak? 01:00:20.830 --> 01:00:22.780 align:middle line:84% Let's go do all the genes under a linkage peak. 01:00:22.780 --> 01:00:24.400 align:middle line:84% Well, what are the prior probabilities there? 01:00:24.400 --> 01:00:26.020 align:middle line:84% Well, let's say actually not 200 genes. 01:00:26.020 --> 01:00:28.353 align:middle line:84% Let's say there are only 100 genes under a linkage peak. 01:00:28.353 --> 01:00:30.310 align:middle line:84% Again, the same 12 variants per gene. 01:00:30.310 --> 01:00:31.998 align:middle line:84% But this time, there's only one-- 01:00:31.998 --> 01:00:34.540 align:middle line:84% the assumption is there's only one gene or one variant that's 01:00:34.540 --> 01:00:36.490 align:middle line:84% under that linkage peak that's actually contributing. 01:00:36.490 --> 01:00:38.198 align:middle line:84% So again, the prior probability works out 01:00:38.198 --> 01:00:41.350 align:middle line:84% to be about the same as a candidate gene, 1 in 1,000. 01:00:41.350 --> 01:00:45.370 align:middle line:84% So I would say you should probably, at this point, 01:00:45.370 --> 01:00:47.140 align:middle line:84% be agnostic as to whether you're focusing 01:00:47.140 --> 01:00:48.640 align:middle line:84% all your effort under a linkage peak 01:00:48.640 --> 01:00:51.970 align:middle line:84% or on a list of candidate genes that you've 01:00:51.970 --> 01:00:55.330 align:middle line:90% come up with through biology. 01:00:55.330 --> 01:00:58.120 align:middle line:84% So let's assume that you now do your study, 01:00:58.120 --> 01:01:00.880 align:middle line:84% and you achieve some magical p value of 0.05. 01:01:00.880 --> 01:01:03.310 align:middle line:84% What's the likelihood that there was an association that 01:01:03.310 --> 01:01:04.210 align:middle line:90% was correct? 01:01:04.210 --> 01:01:08.165 align:middle line:84% So the greatest candidate gene in the world. 01:01:08.165 --> 01:01:09.040 align:middle line:90% So this is actually-- 01:01:09.040 --> 01:01:11.665 align:middle line:84% I'm going to give you 10 times more likelihood that this 01:01:11.665 --> 01:01:13.540 align:middle line:84% was correct than your typical candidate gene. 01:01:13.540 --> 01:01:15.505 align:middle line:90% Prior probability of 0.01. 01:01:15.505 --> 01:01:16.990 align:middle line:90% P value, 0.05. 01:01:16.990 --> 01:01:17.920 align:middle line:90% Posterior probability. 01:01:17.920 --> 01:01:19.378 align:middle line:84% This is assuming also perfect power 01:01:19.378 --> 01:01:21.003 align:middle line:84% to detect your association, which means 01:01:21.003 --> 01:01:22.300 align:middle line:90% you did this in a big study. 01:01:22.300 --> 01:01:24.050 align:middle line:84% If you do it in a small study, this number 01:01:24.050 --> 01:01:27.045 align:middle line:84% goes down because this thing scales with power. 01:01:27.045 --> 01:01:28.420 align:middle line:84% So the posterior probably, chance 01:01:28.420 --> 01:01:31.060 align:middle line:90% that you're right, about 15%. 01:01:31.060 --> 01:01:35.152 align:middle line:84% So that's a little bit disheartening, maybe. 01:01:35.152 --> 01:01:36.610 align:middle line:84% How about a typical candidate gene? 01:01:36.610 --> 01:01:38.693 align:middle line:84% This is the typical association study that's done. 01:01:38.693 --> 01:01:41.710 align:middle line:84% There's about a 1% chance that a p value of 0.05 01:01:41.710 --> 01:01:44.590 align:middle line:90% represents a true association. 01:01:44.590 --> 01:01:47.830 align:middle line:84% Candidate or linkage peak, so Calpain-10, 01:01:47.830 --> 01:01:51.130 align:middle line:84% 1% chance with a p value of 0.05. 01:01:51.130 --> 01:01:53.515 align:middle line:84% Now, their p value was lower, but still, that's 01:01:53.515 --> 01:01:55.060 align:middle line:90% not much lower. 01:01:55.060 --> 01:01:57.220 align:middle line:90% Random gene in the genome. 01:01:57.220 --> 01:02:01.990 align:middle line:84% Basically guaranteed to be wrong with a p value of 0.05. 01:02:01.990 --> 01:02:04.510 align:middle line:84% So what kind of p values do you need 01:02:04.510 --> 01:02:06.860 align:middle line:84% to actually have pretty good posterior problem? 01:02:06.860 --> 01:02:10.540 align:middle line:84% Because again, this is assuming good power. 01:02:10.540 --> 01:02:12.502 align:middle line:84% So basically, what you need is p values that 01:02:12.502 --> 01:02:14.710 align:middle line:84% are somewhere in the range of 10 to the minus 4 to 10 01:02:14.710 --> 01:02:15.580 align:middle line:90% to the minus 6. 01:02:15.580 --> 01:02:18.580 align:middle line:84% And this could be a single study or all studies put together. 01:02:18.580 --> 01:02:21.760 align:middle line:90% 01:02:21.760 --> 01:02:23.650 align:middle line:84% So basically, most report associations 01:02:23.650 --> 01:02:26.600 align:middle line:84% are probably incorrect but not all of them. 01:02:26.600 --> 01:02:28.130 align:middle line:84% Some will turn out to be correct. 01:02:28.130 --> 01:02:30.527 align:middle line:84% And I would say that if you have low p values replicated 01:02:30.527 --> 01:02:32.860 align:middle line:84% or there is a really, really, really, really good reason 01:02:32.860 --> 01:02:36.220 align:middle line:84% for plausibility, then that makes it much more likely 01:02:36.220 --> 01:02:36.850 align:middle line:90% to be right. 01:02:36.850 --> 01:02:38.920 align:middle line:84% And then just this last point, genes under linkage peaks 01:02:38.920 --> 01:02:40.690 align:middle line:84% are about the same as any other kind of candidate. 01:02:40.690 --> 01:02:42.520 align:middle line:84% Obviously, if you have a candidate that's 01:02:42.520 --> 01:02:44.520 align:middle line:84% in a good pathway and it's under a linkage peak, 01:02:44.520 --> 01:02:47.360 align:middle line:84% probably raises its prior probability. 01:02:47.360 --> 01:02:48.120 align:middle line:90% Oh, sorry. 01:02:48.120 --> 01:02:57.310 align:middle line:84% AUDIENCE: [INAUDIBLE] you're not just looking 01:02:57.310 --> 01:03:07.942 align:middle line:84% at the state that's [INAUDIBLE] independent SNPs being present, 01:03:07.942 --> 01:03:09.423 align:middle line:90% contributing to [INAUDIBLE]. 01:03:09.423 --> 01:03:10.340 align:middle line:90% JOEL HIRSCHHORN: Yeah. 01:03:10.340 --> 01:03:12.230 align:middle line:90% So I haven't talked at all-- 01:03:12.230 --> 01:03:14.557 align:middle line:84% I only alluded briefly to gene environment interaction 01:03:14.557 --> 01:03:16.140 align:middle line:84% with the PPARgamma and the diet thing. 01:03:16.140 --> 01:03:18.750 align:middle line:84% And haven't talked at all about gene-gene interaction. 01:03:18.750 --> 01:03:21.590 align:middle line:84% So there's a hot debate going on in the field, I would say, 01:03:21.590 --> 01:03:25.340 align:middle line:84% about whether one should look for gene-gene interaction 01:03:25.340 --> 01:03:28.650 align:middle line:84% without any evidence that either gene is involved, 01:03:28.650 --> 01:03:31.230 align:middle line:84% which basically involves doing all the pairwise at least, 01:03:31.230 --> 01:03:34.550 align:middle line:84% if not three-way or four-way or n-way combinations. 01:03:34.550 --> 01:03:36.650 align:middle line:84% And the challenge for that becomes apparent 01:03:36.650 --> 01:03:38.150 align:middle line:84% when you say that there are actually 01:03:38.150 --> 01:03:40.740 align:middle line:84% 600,000 independent variants across the genome. 01:03:40.740 --> 01:03:42.350 align:middle line:84% So even if you're just doing pairwise, 01:03:42.350 --> 01:03:43.830 align:middle line:90% that's 6 times 10 to the fifth. 01:03:43.830 --> 01:03:47.630 align:middle line:84% So it ends up being 3.6 times 10 the 11th variants-- 01:03:47.630 --> 01:03:49.940 align:middle line:84% sorry, pairwise combinations that you have to look for. 01:03:49.940 --> 01:03:51.170 align:middle line:84% And so you know that you're actually 01:03:51.170 --> 01:03:53.753 align:middle line:84% going to get a p value of 10 to the minus 11th for interaction 01:03:53.753 --> 01:03:56.000 align:middle line:84% at some point just because that's 01:03:56.000 --> 01:03:57.230 align:middle line:90% how many pairs you've tested. 01:03:57.230 --> 01:04:00.450 align:middle line:84% In fact, you can get 3.6 of them. 01:04:00.450 --> 01:04:05.690 align:middle line:84% So it becomes a issue of power again 01:04:05.690 --> 01:04:08.120 align:middle line:84% in that to correct for all the different hypotheses 01:04:08.120 --> 01:04:11.030 align:middle line:84% that you'd be testing, you need an enormous sample size, 01:04:11.030 --> 01:04:14.870 align:middle line:84% or you need to hope that the interaction effect is just 01:04:14.870 --> 01:04:17.295 align:middle line:84% so enormous that you actually get a p value that survives 01:04:17.295 --> 01:04:18.920 align:middle line:84% that correction for multiple hypothesis 01:04:18.920 --> 01:04:21.420 align:middle line:84% testing, which might be the case. 01:04:21.420 --> 01:04:23.420 align:middle line:84% But then the question is, if that interaction is 01:04:23.420 --> 01:04:27.830 align:middle line:84% so enormous, you have to draw very strange models that you 01:04:27.830 --> 01:04:31.340 align:middle line:84% wouldn't also pick up at least one of the two as being causal. 01:04:31.340 --> 01:04:34.667 align:middle line:84% Now, what should be done but isn't is that once 01:04:34.667 --> 01:04:36.500 align:middle line:84% you have something that you really believe-- 01:04:36.500 --> 01:04:38.417 align:middle line:84% so for example, we work in the diabetes field, 01:04:38.417 --> 01:04:40.760 align:middle line:84% and we really believe people are gamma Pro12Ala. 01:04:40.760 --> 01:04:43.220 align:middle line:84% But yet, we do not yet-- and we should-- 01:04:43.220 --> 01:04:46.070 align:middle line:84% test for pairwise interactions for every single thing 01:04:46.070 --> 01:04:48.260 align:middle line:84% we genotype in Pro12Ala because you 01:04:48.260 --> 01:04:50.480 align:middle line:84% can draw plenty of reasonable models 01:04:50.480 --> 01:04:52.635 align:middle line:84% where you would only detect one of the two things 01:04:52.635 --> 01:04:53.510 align:middle line:90% that are interacting. 01:04:53.510 --> 01:04:55.343 align:middle line:84% And the other one would be completely masked 01:04:55.343 --> 01:04:56.960 align:middle line:90% or largely masked. 01:04:56.960 --> 01:04:59.240 align:middle line:84% So we don't do that, and we should. 01:04:59.240 --> 01:05:01.552 align:middle line:84% And but I would say that if you're-- 01:05:01.552 --> 01:05:03.260 align:middle line:84% again, if you're picking up a paper which 01:05:03.260 --> 01:05:05.843 align:middle line:84% says we didn't see anything with either one of these variants, 01:05:05.843 --> 01:05:08.700 align:middle line:84% but then we did the pairwise test and we see an interaction, 01:05:08.700 --> 01:05:10.117 align:middle line:84% think about your prior probability 01:05:10.117 --> 01:05:12.620 align:middle line:84% that that pairwise test was actually associated. 01:05:12.620 --> 01:05:15.255 align:middle line:84% The number of pairs in the genome versus the number that 01:05:15.255 --> 01:05:17.630 align:middle line:84% are actually going to be causal, your prior probabilities 01:05:17.630 --> 01:05:20.390 align:middle line:84% are probably several orders of magnitude lower. 01:05:20.390 --> 01:05:22.610 align:middle line:84% So unless there's, again, some really great reason to 01:05:22.610 --> 01:05:24.630 align:middle line:84% know that these genetic variants interact. 01:05:24.630 --> 01:05:28.040 align:middle line:84% So that's my incredibly biased answer to that 01:05:28.040 --> 01:05:30.875 align:middle line:84% or not biased but opinionated answer to that question. 01:05:30.875 --> 01:05:33.800 align:middle line:90% 01:05:33.800 --> 01:05:36.440 align:middle line:84% I won't go through why you get the same sort of winner's curse 01:05:36.440 --> 01:05:38.690 align:middle line:90% and things like that in linkage. 01:05:38.690 --> 01:05:42.260 align:middle line:84% But just to take you through rare variant association 01:05:42.260 --> 01:05:45.660 align:middle line:84% studies, again, just in case you pick up literature on this. 01:05:45.660 --> 01:05:48.080 align:middle line:84% So this is the typical rare variant association study 01:05:48.080 --> 01:05:51.710 align:middle line:90% as it is practiced in 2004. 01:05:51.710 --> 01:05:54.380 align:middle line:84% Genes are resequenced in affected individuals 01:05:54.380 --> 01:05:56.900 align:middle line:84% to try to find, say, missense variants. 01:05:56.900 --> 01:05:59.867 align:middle line:84% And then those missense variants are then taken and typed 01:05:59.867 --> 01:06:00.950 align:middle line:90% in unaffected individuals. 01:06:00.950 --> 01:06:04.220 align:middle line:84% So it's sequencing here, genotyping here. 01:06:04.220 --> 01:06:07.040 align:middle line:84% And what's often observed is that they're not there. 01:06:07.040 --> 01:06:09.980 align:middle line:84% So I'll take you through a possible resequencing 01:06:09.980 --> 01:06:10.730 align:middle line:90% association study. 01:06:10.730 --> 01:06:12.438 align:middle line:84% Let's resequence your favorite gene, gene 01:06:12.438 --> 01:06:15.707 align:middle line:84% X. 200 diabetic individuals, and, lo and behold, 01:06:15.707 --> 01:06:17.540 align:middle line:84% you actually find 10 rare missense variants, 01:06:17.540 --> 01:06:20.600 align:middle line:84% each seen in one person with diabetes. 01:06:20.600 --> 01:06:22.670 align:middle line:84% You then type 200 healthy individuals. 01:06:22.670 --> 01:06:24.230 align:middle line:84% Those variants are not seen at all. 01:06:24.230 --> 01:06:31.430 align:middle line:84% So you conclude that rare missense variants 01:06:31.430 --> 01:06:33.660 align:middle line:90% in gene X cause diabetes. 01:06:33.660 --> 01:06:37.040 align:middle line:84% So now I'm going to take you through an equally possible 01:06:37.040 --> 01:06:39.260 align:middle line:90% resequencing study. 01:06:39.260 --> 01:06:43.310 align:middle line:84% Resequenced gene X in 200 Red Sox fans. 01:06:43.310 --> 01:06:46.730 align:middle line:84% You will identify-- if this is a gene that 01:06:46.730 --> 01:06:49.260 align:middle line:84% happens to tolerate rare missense variation, 01:06:49.260 --> 01:06:51.410 align:middle line:84% you will identify 10 rare missense variants. 01:06:51.410 --> 01:06:52.670 align:middle line:84% You might end finding some common variants, 01:06:52.670 --> 01:06:53.450 align:middle line:90% but you might not. 01:06:53.450 --> 01:06:56.685 align:middle line:84% Many genes don't have any common missense variants. 01:06:56.685 --> 01:06:59.060 align:middle line:84% If you then take those 10 missense variants and type them 01:06:59.060 --> 01:07:00.800 align:middle line:84% in 200 other people, say Yankees fans, 01:07:00.800 --> 01:07:02.540 align:middle line:90% you will not see those variants. 01:07:02.540 --> 01:07:04.580 align:middle line:90% So this is very plausible. 01:07:04.580 --> 01:07:07.310 align:middle line:84% And you will conclude that rare variants in gene X 01:07:07.310 --> 01:07:10.190 align:middle line:84% make you root for the Red Sox, which of course is not 01:07:10.190 --> 01:07:12.750 align:middle line:90% the correct answer. 01:07:12.750 --> 01:07:16.970 align:middle line:84% And the reason for this has to do with the fact 01:07:16.970 --> 01:07:18.680 align:middle line:84% that in identifying these rare variants, 01:07:18.680 --> 01:07:20.720 align:middle line:84% you have to go to very great depth 01:07:20.720 --> 01:07:22.340 align:middle line:90% in doing the resequencing. 01:07:22.340 --> 01:07:25.097 align:middle line:84% So you get common variants if you sequence two chromosomes. 01:07:25.097 --> 01:07:27.680 align:middle line:84% So that's what I showed you at the very beginning of the talk. 01:07:27.680 --> 01:07:30.260 align:middle line:84% So if you just compare two chromosomes 01:07:30.260 --> 01:07:32.060 align:middle line:84% and then look, are those variants common 01:07:32.060 --> 01:07:33.440 align:middle line:84% in the rest of the population, the answer 01:07:33.440 --> 01:07:34.380 align:middle line:90% is almost always yes. 01:07:34.380 --> 01:07:36.980 align:middle line:84% And the average allele frequency is about 1 in 5. 01:07:36.980 --> 01:07:40.070 align:middle line:84% So average is 20% for the minor allele 01:07:40.070 --> 01:07:41.720 align:middle line:90% and 80% for the other allele. 01:07:41.720 --> 01:07:44.645 align:middle line:84% But if you then sequence a large group of people, 01:07:44.645 --> 01:07:46.520 align:middle line:84% you obviously find all these common variants. 01:07:46.520 --> 01:07:49.100 align:middle line:84% You find those in the first four chromosomes you do. 01:07:49.100 --> 01:07:51.200 align:middle line:84% You basically run through all the common variants. 01:07:51.200 --> 01:07:52.880 align:middle line:84% But you actually keep finding variants 01:07:52.880 --> 01:07:55.010 align:middle line:84% as you go every once in a while, and this 01:07:55.010 --> 01:07:56.690 align:middle line:84% has to do with population genetics 01:07:56.690 --> 01:07:58.430 align:middle line:84% and what the frequency distributions are 01:07:58.430 --> 01:07:59.340 align:middle line:90% and stuff like that. 01:07:59.340 --> 01:08:00.680 align:middle line:84% But most of the variants you find, 01:08:00.680 --> 01:08:03.320 align:middle line:84% if you do deep resequencing, end up being these rare variants. 01:08:03.320 --> 01:08:05.420 align:middle line:84% They don't account for most of the heterozygosity 01:08:05.420 --> 01:08:06.800 align:middle line:90% because they're so rare. 01:08:06.800 --> 01:08:08.370 align:middle line:84% And in fact, their average frequency 01:08:08.370 --> 01:08:09.990 align:middle line:90% is something like 1 in 10,000. 01:08:09.990 --> 01:08:11.940 align:middle line:84% So if you type them in 200 people, 01:08:11.940 --> 01:08:14.560 align:middle line:84% you're never going to see any of them again. 01:08:14.560 --> 01:08:17.490 align:middle line:84% So what needs to get done is basically, 01:08:17.490 --> 01:08:20.385 align:middle line:84% you need to resequence the controls as well. 01:08:20.385 --> 01:08:21.760 align:middle line:84% There are other ways of doing it. 01:08:21.760 --> 01:08:23.177 align:middle line:84% For example, if you have families, 01:08:23.177 --> 01:08:25.920 align:middle line:84% can see that the rare variants actually 01:08:25.920 --> 01:08:27.930 align:middle line:90% segregate with disease. 01:08:27.930 --> 01:08:31.050 align:middle line:84% That's sort of equivalent to resequencing control. 01:08:31.050 --> 01:08:34.920 align:middle line:90% So that's one way around it. 01:08:34.920 --> 01:08:37.529 align:middle line:84% But that's basically what you need to do. 01:08:37.529 --> 01:08:41.399 align:middle line:84% And furthermore, what I said there 01:08:41.399 --> 01:08:43.290 align:middle line:84% in that previous slide was you identified 01:08:43.290 --> 01:08:44.340 align:middle line:90% 10 missense variants. 01:08:44.340 --> 01:08:48.689 align:middle line:84% Well, what if you identified 11, but one of them was common? 01:08:48.689 --> 01:08:50.904 align:middle line:84% And you said, well, that's a common one, 01:08:50.904 --> 01:08:52.529 align:middle line:84% and it's present in cases and controls, 01:08:52.529 --> 01:08:54.050 align:middle line:84% so that's probably not functional. 01:08:54.050 --> 01:08:56.550 align:middle line:84% That's probably-- and I'll just concentrate on the ones that 01:08:56.550 --> 01:08:57.633 align:middle line:90% are only present in cases. 01:08:57.633 --> 01:09:00.660 align:middle line:84% Well, that's obviously rigging the results beforehand. 01:09:00.660 --> 01:09:03.085 align:middle line:84% So if you limit yourself to just those variants that 01:09:03.085 --> 01:09:05.460 align:middle line:84% are present in cases and say those are the only ones that 01:09:05.460 --> 01:09:08.290 align:middle line:84% matter, obviously they're only going to be present in cases. 01:09:08.290 --> 01:09:10.950 align:middle line:84% So you have to have some way of grouping these rare variants 01:09:10.950 --> 01:09:13.710 align:middle line:84% for analysis before you know the results of the association 01:09:13.710 --> 01:09:16.620 align:middle line:84% study of which ones are in cases and which ones are in controls, 01:09:16.620 --> 01:09:18.930 align:middle line:84% so something like either I'll only look at frame shifts 01:09:18.930 --> 01:09:21.000 align:middle line:84% or nonsense mutations, or I'll only 01:09:21.000 --> 01:09:23.819 align:middle line:84% look at things that have this particular in vitro functional 01:09:23.819 --> 01:09:27.550 align:middle line:90% property or something like that. 01:09:27.550 --> 01:09:31.590 align:middle line:84% So just in the last couple of minutes-- 01:09:31.590 --> 01:09:34.250 align:middle line:90% or I'm actually way ahead. 01:09:34.250 --> 01:09:35.250 align:middle line:90% AUDIENCE: Seven minutes. 01:09:35.250 --> 01:09:35.479 align:middle line:90% JOEL HIRSCHHORN: Oh, no. 01:09:35.479 --> 01:09:36.090 align:middle line:90% Not way ahead. 01:09:36.090 --> 01:09:36.240 align:middle line:90% Good. 01:09:36.240 --> 01:09:36.800 align:middle line:90% 12:30. 01:09:36.800 --> 01:09:37.310 align:middle line:90% Good. 01:09:37.310 --> 01:09:39.090 align:middle line:84% For a minute, I thought I had 37 minutes left. 01:09:39.090 --> 01:09:41.010 align:middle line:84% I was like, I've been talking longer than that. 01:09:41.010 --> 01:09:41.510 align:middle line:90% OK. 01:09:41.510 --> 01:09:43.630 align:middle line:90% So what could we learn? 01:09:43.630 --> 01:09:45.960 align:middle line:84% And I think this is also important to think about 01:09:45.960 --> 01:09:48.600 align:middle line:84% in what are the goals, and what are the likely outcomes? 01:09:48.600 --> 01:09:50.933 align:middle line:84% There's a lot of hype around the whole field of genomics 01:09:50.933 --> 01:09:53.250 align:middle line:84% and SNPs and common disease and just 01:09:53.250 --> 01:09:57.730 align:middle line:84% to try to take a look at what is and is not likely. 01:09:57.730 --> 01:10:01.530 align:middle line:84% So I don't know if anybody's seen the movie Gattaca. 01:10:01.530 --> 01:10:03.187 align:middle line:84% So Gattaca, I think people actually 01:10:03.187 --> 01:10:04.770 align:middle line:84% have Gattaca in the back of their mind 01:10:04.770 --> 01:10:08.850 align:middle line:84% when they are thinking about this, which is that we should 01:10:08.850 --> 01:10:11.435 align:middle line:84% be able to tell by genotyping, if we 01:10:11.435 --> 01:10:12.810 align:middle line:84% could sequence somebody's genome, 01:10:12.810 --> 01:10:14.115 align:middle line:90% we will be able to predict-- 01:10:14.115 --> 01:10:16.073 align:middle line:84% if you look at the opening scene in that movie, 01:10:16.073 --> 01:10:18.750 align:middle line:84% it's like, probability of coronary artery disease, 98%. 01:10:18.750 --> 01:10:21.090 align:middle line:90% Probability of stroke, 100%. 01:10:21.090 --> 01:10:22.240 align:middle line:90% This sort of thing. 01:10:22.240 --> 01:10:24.960 align:middle line:84% So we already know that genetic variation, actually, 01:10:24.960 --> 01:10:26.490 align:middle line:84% even if we had perfect information, 01:10:26.490 --> 01:10:29.162 align:middle line:84% only explains 30% to 50% of common disease. 01:10:29.162 --> 01:10:30.870 align:middle line:84% But what's even the likelihood that we'll 01:10:30.870 --> 01:10:32.113 align:middle line:90% get that good a prediction? 01:10:32.113 --> 01:10:34.530 align:middle line:84% I think it's probably not that great because it's probably 01:10:34.530 --> 01:10:36.810 align:middle line:84% a lot of alleles that are going to interact in complex ways 01:10:36.810 --> 01:10:38.352 align:middle line:84% that we may not be able to figure out 01:10:38.352 --> 01:10:39.450 align:middle line:90% for a very long time. 01:10:39.450 --> 01:10:41.100 align:middle line:84% But nonetheless, there is some thought 01:10:41.100 --> 01:10:44.010 align:middle line:84% that we may be able to identify people who are at high risk 01:10:44.010 --> 01:10:47.130 align:middle line:84% or particularly, people who might be at low risk. 01:10:47.130 --> 01:10:50.010 align:middle line:84% If that were the case-- and I've shown here probably a very 01:10:50.010 --> 01:10:52.800 align:middle line:84% optimistic scenario, which is that you could enrich this much 01:10:52.800 --> 01:10:53.928 align:middle line:90% for the high-risk people-- 01:10:53.928 --> 01:10:55.470 align:middle line:84% you might say that these people would 01:10:55.470 --> 01:10:57.420 align:middle line:84% get a more aggressive intervention preventive 01:10:57.420 --> 01:11:00.960 align:middle line:84% measure, and these people might get the standard of care 01:11:00.960 --> 01:11:03.812 align:middle line:84% or might even be able to get slightly less 01:11:03.812 --> 01:11:05.520 align:middle line:84% aggressive interventions depending on how 01:11:05.520 --> 01:11:07.620 align:middle line:90% the risk-benefit played out. 01:11:07.620 --> 01:11:10.260 align:middle line:84% This obviously requires that there be effective preventive 01:11:10.260 --> 01:11:11.560 align:middle line:90% measures and things like that. 01:11:11.560 --> 01:11:13.560 align:middle line:84% But you could imagine, for things like diabetes, 01:11:13.560 --> 01:11:16.230 align:middle line:84% you could say, you really are at high risk for getting diabetes. 01:11:16.230 --> 01:11:18.540 align:middle line:84% You really need to exercise and lose weight, 01:11:18.540 --> 01:11:20.430 align:middle line:84% and maybe that there would be some-- 01:11:20.430 --> 01:11:22.830 align:middle line:84% you could hook that person up with a nutritionist and all 01:11:22.830 --> 01:11:25.493 align:middle line:84% the personal trainer and all that sort of stuff. 01:11:25.493 --> 01:11:27.660 align:middle line:84% And it might not be something that we, as a society, 01:11:27.660 --> 01:11:30.160 align:middle line:84% could afford for everybody but maybe at least for the people 01:11:30.160 --> 01:11:32.310 align:middle line:84% that are the highest risk, we could afford it 01:11:32.310 --> 01:11:33.430 align:middle line:90% if we could get there. 01:11:33.430 --> 01:11:35.520 align:middle line:84% But I don't think that this is necessarily really 01:11:35.520 --> 01:11:38.590 align:middle line:90% where the true promise lies. 01:11:38.590 --> 01:11:40.480 align:middle line:90% There's another possibility. 01:11:40.480 --> 01:11:43.140 align:middle line:84% This is called pharmacogenomics, which 01:11:43.140 --> 01:11:46.493 align:middle line:84% basically reclassification of individuals to guide therapies. 01:11:46.493 --> 01:11:47.910 align:middle line:84% So you have this mix of people who 01:11:47.910 --> 01:11:50.040 align:middle line:84% have some disease, like type 2 diabetes, 01:11:50.040 --> 01:11:53.100 align:middle line:84% but it turns out it's actually type 2A, 2B, 2C, and 2D. 01:11:53.100 --> 01:11:56.010 align:middle line:84% And maybe we could classify them by DNA sequence or expression 01:11:56.010 --> 01:11:57.720 align:middle line:90% profile or some genomic method. 01:11:57.720 --> 01:11:59.220 align:middle line:84% And we could say, well, these people 01:11:59.220 --> 01:12:01.680 align:middle line:84% will do best with treatment A, these people with B, 01:12:01.680 --> 01:12:04.500 align:middle line:84% these with C, and these with D. And again, that may or may not 01:12:04.500 --> 01:12:06.450 align:middle line:84% come to pass, and certainly, people are looking at it. 01:12:06.450 --> 01:12:07.980 align:middle line:84% And I think it's going to partially come to pass. 01:12:07.980 --> 01:12:09.960 align:middle line:84% I'll just give you one real-life example 01:12:09.960 --> 01:12:11.740 align:middle line:84% that actually is not in the clinic yet, 01:12:11.740 --> 01:12:13.920 align:middle line:90% but I'm not sure why not. 01:12:13.920 --> 01:12:17.310 align:middle line:84% So there's a gene called CYP2C9, which actually metabolizes 01:12:17.310 --> 01:12:23.280 align:middle line:84% warfarin, which is used commonly to treat conditions involved 01:12:23.280 --> 01:12:25.800 align:middle line:84% with thrombosis or blood clotting. 01:12:25.800 --> 01:12:31.770 align:middle line:84% This is an anti-clotting agent, actually, vitamin K. 01:12:31.770 --> 01:12:33.040 align:middle line:90% And this didn't quite-- 01:12:33.040 --> 01:12:33.540 align:middle line:90% I'm sorry. 01:12:33.540 --> 01:12:35.498 align:middle line:84% The white on yellow is not a good color scheme. 01:12:35.498 --> 01:12:37.680 align:middle line:84% But basically, most of the population 01:12:37.680 --> 01:12:40.377 align:middle line:90% has zero low-activity alleles. 01:12:40.377 --> 01:12:42.460 align:middle line:84% 20% of the population has one low-activity allele, 01:12:42.460 --> 01:12:46.038 align:middle line:84% and 5% of the population have two low-activity alleles. 01:12:46.038 --> 01:12:47.580 align:middle line:84% And if you look at the average dosage 01:12:47.580 --> 01:12:49.953 align:middle line:84% that people end up on by their genotype-- 01:12:49.953 --> 01:12:51.870 align:middle line:84% so this dosage is not based on their genotype. 01:12:51.870 --> 01:12:54.600 align:middle line:84% It's like there's trial and error by the physician, 01:12:54.600 --> 01:12:58.500 align:middle line:84% and they find, oh, all of a sudden, 01:12:58.500 --> 01:13:00.150 align:middle line:84% you have a nosebleed that won't stop. 01:13:00.150 --> 01:13:01.650 align:middle line:84% Maybe your warfarin dose is too high 01:13:01.650 --> 01:13:03.180 align:middle line:90% or your PT value is too high. 01:13:03.180 --> 01:13:05.070 align:middle line:90% We have to cut the dose. 01:13:05.070 --> 01:13:09.880 align:middle line:84% You see that there's a very nice allele versus dose correlation 01:13:09.880 --> 01:13:12.610 align:middle line:84% here so that if you have two low-activity alleles 01:13:12.610 --> 01:13:15.275 align:middle line:84% and you get put on the typical 5 milligram dose, 01:13:15.275 --> 01:13:16.900 align:middle line:84% you're going to have bleeding problems. 01:13:16.900 --> 01:13:18.970 align:middle line:84% And in fact, you're going to be at six times 01:13:18.970 --> 01:13:20.868 align:middle line:84% the risk of serious complications. 01:13:20.868 --> 01:13:22.660 align:middle line:84% So personally, if I were going on warfarin, 01:13:22.660 --> 01:13:25.990 align:middle line:84% I would probably want to know what my CYP2C9 genotype was. 01:13:25.990 --> 01:13:28.690 align:middle line:84% But it's not yet standard practice. 01:13:28.690 --> 01:13:30.475 align:middle line:84% One of the only cases that I know that is 01:13:30.475 --> 01:13:32.230 align:middle line:84% is actually if you have cancer and you're 01:13:32.230 --> 01:13:35.020 align:middle line:84% getting six [INAUDIBLE] purine or related drugs, 01:13:35.020 --> 01:13:38.950 align:middle line:84% there's a gene called TPMT, which again metabolizes that 01:13:38.950 --> 01:13:40.157 align:middle line:90% into a toxic intermediate. 01:13:40.157 --> 01:13:42.490 align:middle line:84% And that has now become pretty much the standard of care 01:13:42.490 --> 01:13:45.168 align:middle line:84% because again, the risk of death is higher 01:13:45.168 --> 01:13:46.460 align:middle line:90% if you have the wrong genotype. 01:13:46.460 --> 01:13:48.623 align:middle line:84% So death gets people's attention quite-- 01:13:48.623 --> 01:13:49.540 align:middle line:90% AUDIENCE: And that's-- 01:13:49.540 --> 01:13:50.020 align:middle line:90% JOEL HIRSCHHORN: --vigorously. 01:13:50.020 --> 01:13:50.890 align:middle line:90% AUDIENCE: Average dose, right? 01:13:50.890 --> 01:13:51.340 align:middle line:90% So-- 01:13:51.340 --> 01:13:51.475 align:middle line:90% [INTERPOSING VOICES] 01:13:51.475 --> 01:13:51.610 align:middle line:90% JOEL HIRSCHHORN: Yes. 01:13:51.610 --> 01:13:52.450 align:middle line:90% So there's a range. 01:13:52.450 --> 01:13:53.320 align:middle line:90% AUDIENCE: --histogram on the-- 01:13:53.320 --> 01:13:54.100 align:middle line:90% JOEL HIRSCHHORN: Oh, absolutely. 01:13:54.100 --> 01:13:54.310 align:middle line:90% Yes. 01:13:54.310 --> 01:13:56.650 align:middle line:84% So I have to say, I should have put error bars on here. 01:13:56.650 --> 01:13:59.200 align:middle line:84% I was looking at that and just noticing I did not have them. 01:13:59.200 --> 01:13:59.528 align:middle line:90% [INTERPOSING VOICES] 01:13:59.528 --> 01:14:01.390 align:middle line:84% AUDIENCE: --because even if you did a histogram-- 01:14:01.390 --> 01:14:01.840 align:middle line:90% JOEL HIRSCHHORN: Right. 01:14:01.840 --> 01:14:02.170 align:middle line:90% But-- 01:14:02.170 --> 01:14:02.860 align:middle line:84% AUDIENCE: --it doesn't necessarily change-- 01:14:02.860 --> 01:14:03.220 align:middle line:90% [INTERPOSING VOICES] 01:14:03.220 --> 01:14:03.580 align:middle line:90% JOEL HIRSCHHORN: Ah. 01:14:03.580 --> 01:14:04.480 align:middle line:90% So right. 01:14:04.480 --> 01:14:06.140 align:middle line:84% So if this were-- if you could say, 01:14:06.140 --> 01:14:08.598 align:middle line:84% well, physicians can do this without knowing ahead of time, 01:14:08.598 --> 01:14:10.498 align:middle line:84% then this should not happen, right? 01:14:10.498 --> 01:14:11.290 align:middle line:90% AUDIENCE: Oh, yeah. 01:14:11.290 --> 01:14:12.207 align:middle line:90% JOEL HIRSCHHORN: Yeah. 01:14:12.207 --> 01:14:12.790 align:middle line:90% So yeah. 01:14:12.790 --> 01:14:14.770 align:middle line:84% So this is the argument is oh, well, physicians 01:14:14.770 --> 01:14:16.450 align:middle line:84% are going to be good at this anyway. 01:14:16.450 --> 01:14:17.110 align:middle line:90% They're not. 01:14:17.110 --> 01:14:18.730 align:middle line:90% They suck at it. 01:14:18.730 --> 01:14:26.570 align:middle line:84% So but I think the other thing that is potentially useful, 01:14:26.570 --> 01:14:31.000 align:middle line:84% so if you look at least the two most well-established risk 01:14:31.000 --> 01:14:32.980 align:middle line:84% alleles for type 2 diabetes, turns out, 01:14:32.980 --> 01:14:34.210 align:middle line:90% they're both in drug targets. 01:14:34.210 --> 01:14:35.320 align:middle line:84% Now, this is admittedly a little bit 01:14:35.320 --> 01:14:37.000 align:middle line:84% circular because the reason they were looked at 01:14:37.000 --> 01:14:38.320 align:middle line:84% was because it was known ahead of time 01:14:38.320 --> 01:14:39.487 align:middle line:90% that they were drug targets. 01:14:39.487 --> 01:14:41.950 align:middle line:84% So they made sense as being good candidate genes. 01:14:41.950 --> 01:14:44.770 align:middle line:84% But it is interesting that two of the major drugs that 01:14:44.770 --> 01:14:46.540 align:middle line:84% are used to treat type 2 diabetes, there's 01:14:46.540 --> 01:14:48.970 align:middle line:84% variants in their targets that affect 01:14:48.970 --> 01:14:50.950 align:middle line:84% the genetic susceptibility to diabetes. 01:14:50.950 --> 01:14:53.643 align:middle line:84% And of course, the promise here is that if you could find all 01:14:53.643 --> 01:14:55.810 align:middle line:84% of the other genetic pathways, even if they only had 01:14:55.810 --> 01:14:58.690 align:middle line:84% a 1.1-fold effect on disease-- so PPARgamma is only about 01:14:58.690 --> 01:15:02.245 align:middle line:84% a 25% risk effect on risk of disease. 01:15:02.245 --> 01:15:04.120 align:middle line:84% But you could easily imagine that you come up 01:15:04.120 --> 01:15:05.920 align:middle line:84% with a pharmacologic intervention that 01:15:05.920 --> 01:15:08.440 align:middle line:84% does a lot more than the genetic polymorphism does. 01:15:08.440 --> 01:15:11.650 align:middle line:84% So there's only so much tolerance for how much 01:15:11.650 --> 01:15:14.890 align:middle line:84% genetic variation evolution will let happen. 01:15:14.890 --> 01:15:18.280 align:middle line:84% But it might be that giving thiazolidinediones might 01:15:18.280 --> 01:15:21.010 align:middle line:84% be like having 10 times as much of an effect 01:15:21.010 --> 01:15:23.240 align:middle line:84% as having an alanine allele does. 01:15:23.240 --> 01:15:26.110 align:middle line:84% So the idea is that hopefully by identifying the genes that are 01:15:26.110 --> 01:15:29.470 align:middle line:84% responsible-- and this is true for severe forms of the disease 01:15:29.470 --> 01:15:30.160 align:middle line:90% as well-- 01:15:30.160 --> 01:15:33.110 align:middle line:90% you might identify drug targets. 01:15:33.110 --> 01:15:35.230 align:middle line:84% And finally, of course, this helps just 01:15:35.230 --> 01:15:38.380 align:middle line:84% understand human biology, and that's always a worthy goal. 01:15:38.380 --> 01:15:44.380 align:middle line:84% So should talk about potential difficulties from all these-- 01:15:44.380 --> 01:15:47.550 align:middle line:90% from all these advances. 01:15:47.550 --> 01:15:49.300 align:middle line:84% There are obviously concerns about privacy 01:15:49.300 --> 01:15:50.080 align:middle line:90% with genetic data. 01:15:50.080 --> 01:15:52.120 align:middle line:84% And don't really have time to go into it, 01:15:52.120 --> 01:15:54.820 align:middle line:84% but obviously, the big fear has to do 01:15:54.820 --> 01:15:58.090 align:middle line:84% with insurance discrimination and employers 01:15:58.090 --> 01:15:59.420 align:middle line:90% knowing about genetic diseases. 01:15:59.420 --> 01:16:01.270 align:middle line:84% And there's bills that keep making it 01:16:01.270 --> 01:16:03.700 align:middle line:84% almost all the way through Congress and not quite 01:16:03.700 --> 01:16:05.920 align:middle line:90% to deal with this. 01:16:05.920 --> 01:16:07.210 align:middle line:90% So lobby your congressman. 01:16:07.210 --> 01:16:09.880 align:middle line:90% 01:16:09.880 --> 01:16:13.510 align:middle line:84% There's the Gattaca mindset about improper interpretation 01:16:13.510 --> 01:16:16.180 align:middle line:84% of predictive information, how predictive genetic information 01:16:16.180 --> 01:16:17.055 align:middle line:90% is. 01:16:17.055 --> 01:16:18.430 align:middle line:84% There are companies that will try 01:16:18.430 --> 01:16:21.190 align:middle line:84% to sell you that you should be on diet A or diet B based 01:16:21.190 --> 01:16:24.430 align:middle line:84% on SNP gene types that are in those irreproducible 01:16:24.430 --> 01:16:27.130 align:middle line:90% associations. 01:16:27.130 --> 01:16:30.280 align:middle line:84% There can be psychological impacts of if you say you have 01:16:30.280 --> 01:16:32.710 align:middle line:84% the high-risk allele for diabetes, that might-- 01:16:32.710 --> 01:16:35.860 align:middle line:84% without an understanding that this is a correlative rather 01:16:35.860 --> 01:16:37.465 align:middle line:84% than a perfectly predictive measure, 01:16:37.465 --> 01:16:40.180 align:middle line:84% that could have significant harm as well as 01:16:40.180 --> 01:16:43.220 align:middle line:84% impact on reproductive choices, again, 01:16:43.220 --> 01:16:45.943 align:middle line:84% depending on how predictive people think things are. 01:16:45.943 --> 01:16:48.610 align:middle line:84% There are huge interactions with concepts of race and ethnicity. 01:16:48.610 --> 01:16:50.027 align:middle line:84% We've alluded to this a little bit 01:16:50.027 --> 01:16:51.430 align:middle line:84% with different allele frequencies 01:16:51.430 --> 01:16:53.380 align:middle line:90% and different populations. 01:16:53.380 --> 01:16:55.600 align:middle line:84% Most alleles are present around the world, 01:16:55.600 --> 01:16:58.240 align:middle line:84% and you can't use generally any one variant 01:16:58.240 --> 01:17:00.160 align:middle line:84% to tell different ethnic groups apart, 01:17:00.160 --> 01:17:01.660 align:middle line:84% but there are definitely differences 01:17:01.660 --> 01:17:04.030 align:middle line:84% in allele frequencies between populations. 01:17:04.030 --> 01:17:09.100 align:middle line:84% And that has significant implications. 01:17:09.100 --> 01:17:11.763 align:middle line:84% And then finally, I certainly don't 01:17:11.763 --> 01:17:13.180 align:middle line:84% study the genetics of performance, 01:17:13.180 --> 01:17:14.680 align:middle line:84% but you could imagine that if people 01:17:14.680 --> 01:17:18.070 align:middle line:84% started getting into that, that could be quite dicey. 01:17:18.070 --> 01:17:20.800 align:middle line:84% And even if the science were perfect 01:17:20.800 --> 01:17:27.470 align:middle line:84% from a technical standpoint, it might create a lot of trouble. 01:17:27.470 --> 01:17:32.130 align:middle line:84% So David Altshuler is a collaborator with me 01:17:32.130 --> 01:17:35.500 align:middle line:84% on the diabetes work and also very involved in the hap map 01:17:35.500 --> 01:17:37.990 align:middle line:84% with the work that was done by Stacey Gabriel, Mark 01:17:37.990 --> 01:17:40.000 align:middle line:84% Daly, Steve Schaffner on haplotype blocks. 01:17:40.000 --> 01:17:41.953 align:middle line:84% The diabetes group also is Noelle Lake, 01:17:41.953 --> 01:17:43.870 align:middle line:84% Cecilia [INAUDIBLE] did some expression stuff. 01:17:43.870 --> 01:17:47.140 align:middle line:84% And Kirk and Lee and Eric were on the meta analysis 01:17:47.140 --> 01:17:47.890 align:middle line:90% association study. 01:17:47.890 --> 01:17:50.350 align:middle line:84% And there are a lot of genome project-type things 01:17:50.350 --> 01:17:52.760 align:middle line:84% that are obviously giving us all the tools to do this. 01:17:52.760 --> 01:17:54.500 align:middle line:90% So thanks. 01:17:54.500 --> 01:17:55.000 align:middle line:90%