WEBVTT 00:00:00.000 --> 00:00:01.890 align:middle line:90% 00:00:01.890 --> 00:00:05.440 align:middle line:84% ISAAC SAMUEL KOHANE: So let me give you a case history. 00:00:05.440 --> 00:00:08.220 align:middle line:84% This is work that I've done with one of my former interns, 00:00:08.220 --> 00:00:12.540 align:middle line:84% actually, who's a superb scientist at the Dana-Farber 00:00:12.540 --> 00:00:16.740 align:middle line:84% Cancer Institute, David Rowitch, who also is a part time 00:00:16.740 --> 00:00:18.120 align:middle line:90% neonatologist. 00:00:18.120 --> 00:00:21.120 align:middle line:84% I call this finding the needle in the haystack. 00:00:21.120 --> 00:00:24.400 align:middle line:84% And it's about studying the cerebellum. 00:00:24.400 --> 00:00:28.500 align:middle line:84% So I impressed, I hope, upon you on the last lecture, 00:00:28.500 --> 00:00:30.180 align:middle line:90% just how bad microarrays are. 00:00:30.180 --> 00:00:32.759 align:middle line:84% How noisy they are, and how difficult 00:00:32.759 --> 00:00:34.120 align:middle line:90% it is to interpret them. 00:00:34.120 --> 00:00:36.180 align:middle line:90% So now, I'll show you how-- 00:00:36.180 --> 00:00:41.550 align:middle line:84% darn it-- how, if you actually do the computation right, 00:00:41.550 --> 00:00:46.560 align:middle line:84% you can actually extract useful biomedical information out 00:00:46.560 --> 00:00:48.473 align:middle line:90% of it. 00:00:48.473 --> 00:00:49.890 align:middle line:84% So the cerebellum is a nice system 00:00:49.890 --> 00:00:53.083 align:middle line:84% to study because it has this nicely layered structure. 00:00:53.083 --> 00:00:54.750 align:middle line:84% The cerebellum is this part of the brain 00:00:54.750 --> 00:00:58.170 align:middle line:84% that sits in the posterior fossa, at the back. 00:00:58.170 --> 00:01:00.930 align:middle line:84% And has this very nicely layered structure 00:01:00.930 --> 00:01:02.700 align:middle line:84% which is detailed in this cross section 00:01:02.700 --> 00:01:07.770 align:middle line:84% here so that each layer has this distinct group of cells. 00:01:07.770 --> 00:01:10.050 align:middle line:84% And often, these group of cells are 00:01:10.050 --> 00:01:15.120 align:middle line:84% distinguished by different genes expressed at different times, 00:01:15.120 --> 00:01:18.810 align:middle line:84% specifically, for that cell type or cell layer. 00:01:18.810 --> 00:01:23.790 align:middle line:84% And the cerebellum has a pivotal role in locomotion and posture, 00:01:23.790 --> 00:01:26.160 align:middle line:84% but the fact that I can stand approximately straight 00:01:26.160 --> 00:01:30.600 align:middle line:84% and move around is almost solely a function of the activity 00:01:30.600 --> 00:01:31.950 align:middle line:90% of the cerebellum. 00:01:31.950 --> 00:01:34.872 align:middle line:84% And because of its nicely layered structure, 00:01:34.872 --> 00:01:36.330 align:middle line:84% the cerebellum has been the subject 00:01:36.330 --> 00:01:38.190 align:middle line:84% of a lot of neuroscience because it's 00:01:38.190 --> 00:01:42.780 align:middle line:84% much less complex than the other central neural tissues, 00:01:42.780 --> 00:01:45.880 align:middle line:90% in that regard. 00:01:45.880 --> 00:01:50.400 align:middle line:84% So there was a particular system that David was interested in. 00:01:50.400 --> 00:01:54.390 align:middle line:84% Sonic Hedgehog, it's one of these supposedly amusing names 00:01:54.390 --> 00:01:57.632 align:middle line:84% developed by fly biologists because fly biologists may 00:01:57.632 --> 00:01:59.340 align:middle line:84% frequently have a very bad sense of humor 00:01:59.340 --> 00:02:02.580 align:middle line:84% and they apply that to all the genes that they discover. 00:02:02.580 --> 00:02:04.440 align:middle line:84% Sonic Hedgehog is one such gene, turns out 00:02:04.440 --> 00:02:08.070 align:middle line:84% to be a very important signaling molecule 00:02:08.070 --> 00:02:10.320 align:middle line:84% that's very important in proliferation 00:02:10.320 --> 00:02:12.060 align:middle line:90% and differentiation of cells. 00:02:12.060 --> 00:02:15.690 align:middle line:84% And David wants to know, what was 00:02:15.690 --> 00:02:21.240 align:middle line:84% the role of Sonic Hedgehog in this outermost 00:02:21.240 --> 00:02:28.020 align:middle line:84% of the layer of the cerebellum, the external granule layer 00:02:28.020 --> 00:02:31.230 align:middle line:90% A, EGLA. 00:02:31.230 --> 00:02:35.425 align:middle line:84% It was known that cyclin D1 was expressed in this layer. 00:02:35.425 --> 00:02:37.800 align:middle line:84% And different other genes were expressed in other layers. 00:02:37.800 --> 00:02:39.300 align:middle line:84% It was known that Sonic Hedgehog was 00:02:39.300 --> 00:02:41.410 align:middle line:84% important in a variety of different ways. 00:02:41.410 --> 00:02:44.370 align:middle line:84% For instance, genes that had previously 00:02:44.370 --> 00:02:46.830 align:middle line:84% been implicated in a Sonic Hedgehog signaling 00:02:46.830 --> 00:02:55.410 align:middle line:84% pathway when a mutant cause in the mouse model here-- 00:02:55.410 --> 00:02:57.030 align:middle line:84% a posterior cerebellar tumor-- which 00:02:57.030 --> 00:03:01.800 align:middle line:84% is very much akin to the medulloblastomas that 00:03:01.800 --> 00:03:05.610 align:middle line:84% are a big cause of juvenile brain cancer, 00:03:05.610 --> 00:03:07.990 align:middle line:90% and a lethal one at that. 00:03:07.990 --> 00:03:11.370 align:middle line:84% And it's certainly been shown that medulloblastomas is 00:03:11.370 --> 00:03:13.530 align:middle line:84% are involved in the Sonic Hedgehog 00:03:13.530 --> 00:03:16.240 align:middle line:84% pathways, a large percentage of them. 00:03:16.240 --> 00:03:18.750 align:middle line:84% So the question was, could we find other culprits, 00:03:18.750 --> 00:03:24.240 align:middle line:84% other genes that are involved in the Sonic Hedgehog pathway? 00:03:24.240 --> 00:03:25.800 align:middle line:84% And it's a big haystack that we're 00:03:25.800 --> 00:03:30.730 align:middle line:84% looking into because we're looking in two large haystacks, 00:03:30.730 --> 00:03:31.230 align:middle line:90% actually. 00:03:31.230 --> 00:03:34.770 align:middle line:84% One is a large number of probes on the microarray. 00:03:34.770 --> 00:03:37.500 align:middle line:84% Two, we're looking at an entire cerebellum. 00:03:37.500 --> 00:03:41.400 align:middle line:84% A complex tissue-- even though it's one of the simpler neural 00:03:41.400 --> 00:03:42.060 align:middle line:90% tissues-- 00:03:42.060 --> 00:03:45.060 align:middle line:84% when in fact, we're trying to extract a signal out 00:03:45.060 --> 00:03:47.310 align:middle line:84% of one part of this complex tissue, 00:03:47.310 --> 00:03:52.170 align:middle line:84% there's a very superficial layer of cells called the EGLA. 00:03:52.170 --> 00:03:55.540 align:middle line:84% So can we find the signal in time and space? 00:03:55.540 --> 00:03:58.710 align:middle line:84% So we're not only looking for the right number of probes 00:03:58.710 --> 00:04:01.470 align:middle line:84% in the right location, but the right point in development. 00:04:01.470 --> 00:04:03.060 align:middle line:84% And this was quite a challenge for us 00:04:03.060 --> 00:04:07.110 align:middle line:84% because David is a personal friend and former intern. 00:04:07.110 --> 00:04:11.160 align:middle line:84% And he heard me pontificate about the importance 00:04:11.160 --> 00:04:12.120 align:middle line:90% of bioinformatics. 00:04:12.120 --> 00:04:15.300 align:middle line:84% And here was this closed friend who is now engaging 00:04:15.300 --> 00:04:16.584 align:middle line:90% in this type of collaboration. 00:04:16.584 --> 00:04:18.209 align:middle line:84% And if I gave him the wrong candidates, 00:04:18.209 --> 00:04:21.690 align:middle line:84% that means I'd be wasting the time of his post-doc, of which 00:04:21.690 --> 00:04:24.600 align:middle line:84% he had only one at that time, and delaying him by six months 00:04:24.600 --> 00:04:27.160 align:middle line:84% if I gave him some wrong things to look at. 00:04:27.160 --> 00:04:29.175 align:middle line:84% So it was a relatively high stakes game. 00:04:29.175 --> 00:04:31.680 align:middle line:90% 00:04:31.680 --> 00:04:34.470 align:middle line:84% So the experiment was very simple. 00:04:34.470 --> 00:04:38.010 align:middle line:84% To harvest, initially, just day five-- 00:04:38.010 --> 00:04:40.650 align:middle line:84% postnatal day five cerebellum, expose it 00:04:40.650 --> 00:04:44.187 align:middle line:84% to Sonic Hedgehog vehicle and growth arrest media. 00:04:44.187 --> 00:04:46.270 align:middle line:84% Forget about the growth arrest media for a second. 00:04:46.270 --> 00:04:49.500 align:middle line:84% But basically, exposing it to Sonic Hedgehog or vehicle, 00:04:49.500 --> 00:04:55.350 align:middle line:84% hybridize it to the now obsolete affymetrix mouse array. 00:04:55.350 --> 00:04:57.780 align:middle line:84% The mouse arrays are obsolete in this generation, 00:04:57.780 --> 00:04:59.820 align:middle line:90% it's no longer available. 00:04:59.820 --> 00:05:03.900 align:middle line:84% And we did what we refer to in the grant as noise models 00:05:03.900 --> 00:05:07.150 align:middle line:84% in our bioinformatic analysis, which I'll detail shortly. 00:05:07.150 --> 00:05:10.980 align:middle line:84% And then we gave a list of nine significantly up-regulated 00:05:10.980 --> 00:05:13.440 align:middle line:84% genes and 10 genes not significantly up-regulated 00:05:13.440 --> 00:05:15.990 align:middle line:84% and these will be subsequently screened with insight 00:05:15.990 --> 00:05:18.750 align:middle line:84% through hybridization to see whether, in fact, 00:05:18.750 --> 00:05:21.330 align:middle line:90% these predictions were correct. 00:05:21.330 --> 00:05:23.770 align:middle line:84% And so what we did was the very simple thing. 00:05:23.770 --> 00:05:26.790 align:middle line:84% If you recall, I showed you in the last lecture 00:05:26.790 --> 00:05:32.280 align:middle line:84% that I gave you how noise grew dramatically 00:05:32.280 --> 00:05:34.530 align:middle line:84% at lower levels of expression because you could flip 00:05:34.530 --> 00:05:37.920 align:middle line:84% the sign of the up or down regulation just 00:05:37.920 --> 00:05:40.320 align:middle line:84% by virtue of having low levels of expression 00:05:40.320 --> 00:05:42.450 align:middle line:90% affected by low levels of noise. 00:05:42.450 --> 00:05:43.980 align:middle line:84% So clearly, you need to have some 00:05:43.980 --> 00:05:46.260 align:middle line:84% of statistical measure of the separation of the two 00:05:46.260 --> 00:05:47.080 align:middle line:90% conditions. 00:05:47.080 --> 00:05:50.010 align:middle line:84% So in this case, the signal was the expression 00:05:50.010 --> 00:05:53.910 align:middle line:84% of a gene in response to the signal, namely, Sonic Hedgehog. 00:05:53.910 --> 00:05:57.000 align:middle line:84% And the other is the vehicle treated ones. 00:05:57.000 --> 00:05:59.250 align:middle line:84% And the question is, each such gene, 00:05:59.250 --> 00:06:04.470 align:middle line:84% could we find those genes that were clearly separated 00:06:04.470 --> 00:06:10.020 align:middle line:90% across those two conditions? 00:06:10.020 --> 00:06:13.720 align:middle line:84% And there's a variety of ways that you could have done it. 00:06:13.720 --> 00:06:15.420 align:middle line:84% And in fact, as you promise it, I'll 00:06:15.420 --> 00:06:18.780 align:middle line:84% have you explore some of the standard methods that, 00:06:18.780 --> 00:06:21.090 align:middle line:84% for instance, are things like a modified t-test, 00:06:21.090 --> 00:06:24.450 align:middle line:84% like the significance analysis of microarrays. 00:06:24.450 --> 00:06:28.260 align:middle line:84% But I will make the following point, which is, 00:06:28.260 --> 00:06:32.730 align:middle line:84% try to understand what you're doing by looking at ratios. 00:06:32.730 --> 00:06:35.070 align:middle line:84% For instance, just looking at a t-test, where you're 00:06:35.070 --> 00:06:36.540 align:middle line:84% looking at the difference in means, 00:06:36.540 --> 00:06:38.700 align:middle line:84% is fundamentally misunderstanding 00:06:38.700 --> 00:06:43.110 align:middle line:84% the computational challenge in a very simple way. 00:06:43.110 --> 00:06:45.963 align:middle line:84% What average do we want to calculate? 00:06:45.963 --> 00:06:48.380 align:middle line:84% And I realize not all of you are computationally involved, 00:06:48.380 --> 00:06:51.223 align:middle line:90% so I'll try to keep it simple. 00:06:51.223 --> 00:06:53.390 align:middle line:84% There's two types of averages that we might consider 00:06:53.390 --> 00:06:55.070 align:middle line:90% among many calculating. 00:06:55.070 --> 00:06:57.890 align:middle line:84% One is the simple arithmetic average, 00:06:57.890 --> 00:07:00.675 align:middle line:84% where you take where you sum the signal 00:07:00.675 --> 00:07:02.300 align:middle line:84% and divide it by the number of samples. 00:07:02.300 --> 00:07:05.360 align:middle line:84% And you sum the control divide it by the number of samples 00:07:05.360 --> 00:07:06.950 align:middle line:90% and you take that ratio. 00:07:06.950 --> 00:07:10.730 align:middle line:84% Or you take the geometric mean of them. 00:07:10.730 --> 00:07:14.160 align:middle line:84% And why would you do one or the other? 00:07:14.160 --> 00:07:16.970 align:middle line:84% Well, let me ask you the following thing. 00:07:16.970 --> 00:07:18.650 align:middle line:84% When I first thought about this, I 00:07:18.650 --> 00:07:21.050 align:middle line:84% was having a mortgage in my apartment, 00:07:21.050 --> 00:07:22.980 align:middle line:84% so it was incredibly important to me. 00:07:22.980 --> 00:07:28.220 align:middle line:84% If you have four different interest rates over 4 years, 00:07:28.220 --> 00:07:31.490 align:middle line:84% such that, the amount you have to pay out 00:07:31.490 --> 00:07:37.968 align:middle line:84% is 1 plus R1 times 1 plus R2, to 1 plus R3, 1 plus R4 00:07:37.968 --> 00:07:40.010 align:middle line:84% and you're going to have to pay different amounts 00:07:40.010 --> 00:07:41.090 align:middle line:90% on that principle. 00:07:41.090 --> 00:07:44.050 align:middle line:90% 00:07:44.050 --> 00:07:48.160 align:middle line:84% How am I going to estimate the average interest 00:07:48.160 --> 00:07:51.520 align:middle line:90% rate I'm paying per year? 00:07:51.520 --> 00:07:55.240 align:middle line:84% Is it the arithmetic average of these Rs? 00:07:55.240 --> 00:07:57.070 align:middle line:90% No because it's compounding. 00:07:57.070 --> 00:08:00.970 align:middle line:84% I know that's a simple point, but few people 00:08:00.970 --> 00:08:05.530 align:middle line:84% realize it's actually the geometric mean. 00:08:05.530 --> 00:08:09.377 align:middle line:84% And the same error is being made if you look at the arithmetic 00:08:09.377 --> 00:08:10.960 align:middle line:84% mean versus the geometric mean to look 00:08:10.960 --> 00:08:12.730 align:middle line:84% at ratios because, fundamentally, we're 00:08:12.730 --> 00:08:14.110 align:middle line:90% looking at a ratio. 00:08:14.110 --> 00:08:16.990 align:middle line:84% And we did the analysis both ways. 00:08:16.990 --> 00:08:18.890 align:middle line:84% And I can tell you, we knew in this case, 00:08:18.890 --> 00:08:21.310 align:middle line:84% the answer is because we did the insight to hybridization. 00:08:21.310 --> 00:08:24.230 align:middle line:90% And first of all, the good news. 00:08:24.230 --> 00:08:27.490 align:middle line:84% The good news that we published in PNAS a couple of years 00:08:27.490 --> 00:08:31.690 align:middle line:84% ago was that, essentially, 80% of the time that we said 00:08:31.690 --> 00:08:35.890 align:middle line:84% a gene was expressed in the EGLA based on this simple geometric 00:08:35.890 --> 00:08:39.100 align:middle line:84% mean, sure enough, by insight of staining, it was. 00:08:39.100 --> 00:08:43.750 align:middle line:84% And when we said it wasn't, it wasn't. 00:08:43.750 --> 00:08:47.560 align:middle line:84% But the point is, if we use that arithmetic mean, 00:08:47.560 --> 00:08:50.060 align:middle line:84% we had double the false positive rate. 00:08:50.060 --> 00:08:52.870 align:middle line:84% So just understanding something as simple as the difference 00:08:52.870 --> 00:08:54.670 align:middle line:84% in the arithmetic mean and geometric mean, 00:08:54.670 --> 00:08:57.640 align:middle line:84% which is a non-obvious notion to anybody 00:08:57.640 --> 00:09:00.670 align:middle line:84% who's just been used to doing t-tests, 00:09:00.670 --> 00:09:03.790 align:middle line:90% screws up the analysis. 00:09:03.790 --> 00:09:07.660 align:middle line:84% Let's make this more relevant to human disease. 00:09:07.660 --> 00:09:11.380 align:middle line:84% How can we leverage this developmental view 00:09:11.380 --> 00:09:13.390 align:middle line:90% of the mouse? 00:09:13.390 --> 00:09:14.800 align:middle line:90% This we're talking about mice. 00:09:14.800 --> 00:09:16.420 align:middle line:84% How do we actually try to understand 00:09:16.420 --> 00:09:20.770 align:middle line:84% how this is reflected in its relevance 00:09:20.770 --> 00:09:26.500 align:middle line:84% to these tumors that are talking about, medulloblastoma? 00:09:26.500 --> 00:09:30.490 align:middle line:84% So we had, actually, some medulloblastoma. 00:09:30.490 --> 00:09:32.770 align:middle line:84% And I want to say about microarrays 00:09:32.770 --> 00:09:35.830 align:middle line:84% is that this community, along with the SNP community 00:09:35.830 --> 00:09:42.550 align:middle line:84% has been unique in making freely available the data 00:09:42.550 --> 00:09:43.660 align:middle line:90% to its community. 00:09:43.660 --> 00:09:46.000 align:middle line:84% For those of us who have MDs-- or will have MDs-- 00:09:46.000 --> 00:09:49.750 align:middle line:84% let us be aware that still today, 00:09:49.750 --> 00:09:52.840 align:middle line:84% most clinical research papers do not 00:09:52.840 --> 00:09:56.080 align:middle line:84% publish the primary data from which the conclusions are made. 00:09:56.080 --> 00:09:56.860 align:middle line:90% Still do not. 00:09:56.860 --> 00:10:01.780 align:middle line:84% And therefore, you are unable to independently verify 00:10:01.780 --> 00:10:03.730 align:middle line:84% these results or come out with better results 00:10:03.730 --> 00:10:04.970 align:middle line:90% or with improved results. 00:10:04.970 --> 00:10:07.430 align:middle line:84% And that's why it was really incredibly important. 00:10:07.430 --> 00:10:08.860 align:middle line:84% I'm sorry I didn't emphasize this 00:10:08.860 --> 00:10:11.350 align:middle line:84% before, that the early practitioners of microarrays, 00:10:11.350 --> 00:10:13.450 align:middle line:90% such as Todd Golub. 00:10:13.450 --> 00:10:16.840 align:middle line:84% And Pat Brown published all the data right away 00:10:16.840 --> 00:10:17.950 align:middle line:90% after publication. 00:10:17.950 --> 00:10:21.820 align:middle line:84% In fact, I looked a half year ago at Todd's original paper 00:10:21.820 --> 00:10:28.510 align:middle line:90% in 1999, on AML versus ALL. 00:10:28.510 --> 00:10:31.540 align:middle line:84% And I think there were 700 citations of just that one 00:10:31.540 --> 00:10:32.230 align:middle line:90% paper. 00:10:32.230 --> 00:10:35.280 align:middle line:84% And it would have not been the same had they had not 00:10:35.280 --> 00:10:37.780 align:middle line:84% published data because there was a lot of secondary analyzes 00:10:37.780 --> 00:10:41.168 align:middle line:84% and a lot of problems found with the paper and the size 00:10:41.168 --> 00:10:41.710 align:middle line:90% of the paper. 00:10:41.710 --> 00:10:45.550 align:middle line:84% It sets up a whole scientific process of motion 00:10:45.550 --> 00:10:47.000 align:middle line:90% that you don't have otherwise. 00:10:47.000 --> 00:10:49.120 align:middle line:84% So we have this data on a variety 00:10:49.120 --> 00:10:51.010 align:middle line:84% of different human medulloblastomas 00:10:51.010 --> 00:10:54.310 align:middle line:84% that have been published a couple of years ago. 00:10:54.310 --> 00:10:57.497 align:middle line:84% And then, we could go to a database called Homology-- 00:10:57.497 --> 00:10:59.080 align:middle line:84% and by the way, your next problem set, 00:10:59.080 --> 00:11:00.497 align:middle line:84% not the one I'm describing to you, 00:11:00.497 --> 00:11:05.210 align:middle line:84% but your next problem set will be a treasure hunt where 00:11:05.210 --> 00:11:08.320 align:middle line:84% we'll make you hop across all the different databases that 00:11:08.320 --> 00:11:11.860 align:middle line:90% you should know about as basic-- 00:11:11.860 --> 00:11:15.130 align:middle line:84% not basic biologists, but basic consumers 00:11:15.130 --> 00:11:18.490 align:middle line:84% of the very simple kinds of information 00:11:18.490 --> 00:11:20.920 align:middle line:84% around about SNPs and the microarrays. 00:11:20.920 --> 00:11:22.840 align:middle line:84% But we'll get to treasure hunt later. 00:11:22.840 --> 00:11:25.880 align:middle line:84% Suffice it to say that there is a database called humology, 00:11:25.880 --> 00:11:30.070 align:middle line:84% which allows you to map the genes in humans 00:11:30.070 --> 00:11:34.780 align:middle line:84% to those that are present in mice. 00:11:34.780 --> 00:11:38.350 align:middle line:84% And it turns out that there were present on those two 00:11:38.350 --> 00:11:42.340 align:middle line:84% microarrays, there on the order of about 3,000 genes 00:11:42.340 --> 00:11:44.680 align:middle line:84% that are homolog from mice to men. 00:11:44.680 --> 00:11:48.010 align:middle line:84% And so what we did then was to do a principal component 00:11:48.010 --> 00:11:52.060 align:middle line:90% analysis of these data sets. 00:11:52.060 --> 00:11:54.790 align:middle line:84% Now hands up those of you who really understand 00:11:54.790 --> 00:11:57.820 align:middle line:90% principal component analysis? 00:11:57.820 --> 00:11:59.210 align:middle line:90% Two, two of you. 00:11:59.210 --> 00:12:01.840 align:middle line:84% So those who don't, let me try to give you 00:12:01.840 --> 00:12:05.410 align:middle line:90% the simple version of it. 00:12:05.410 --> 00:12:07.660 align:middle line:84% Principal component analysis, essentially, 00:12:07.660 --> 00:12:11.800 align:middle line:84% tries to refactor the data so that you're 00:12:11.800 --> 00:12:14.230 align:middle line:84% looking at these new variables that 00:12:14.230 --> 00:12:19.910 align:middle line:84% are linear sums of the original variables. 00:12:19.910 --> 00:12:21.230 align:middle line:90% In this case, the genes. 00:12:21.230 --> 00:12:24.130 align:middle line:84% These principal components are linear versions 00:12:24.130 --> 00:12:29.560 align:middle line:84% of these original measurements that 00:12:29.560 --> 00:12:34.720 align:middle line:84% capture, in successive components, 00:12:34.720 --> 00:12:39.320 align:middle line:84% the major amount of variation of the data set. 00:12:39.320 --> 00:12:42.070 align:middle line:84% So the first principle components 00:12:42.070 --> 00:12:45.270 align:middle line:84% is this linear combination that gets 00:12:45.270 --> 00:12:48.460 align:middle line:84% the largest amount of the variation in the data set. 00:12:48.460 --> 00:12:50.760 align:middle line:84% The second principle components are orthogonal. 00:12:50.760 --> 00:12:54.300 align:middle line:84% Essentially, at right angles to that first principal component 00:12:54.300 --> 00:12:56.830 align:middle line:84% and get the second most amount of variants. 00:12:56.830 --> 00:12:57.880 align:middle line:90% And so on. 00:12:57.880 --> 00:13:00.090 align:middle line:84% And in fact, have as many principal components 00:13:00.090 --> 00:13:02.640 align:middle line:90% as you have original variables. 00:13:02.640 --> 00:13:06.030 align:middle line:84% But the nice thing is that, heuristically speaking, 00:13:06.030 --> 00:13:09.780 align:middle line:84% by the time you've gone through about two or three or four 00:13:09.780 --> 00:13:11.370 align:middle line:84% principal components, you've typically 00:13:11.370 --> 00:13:14.610 align:middle line:84% captured on the order of 70% to 80% 00:13:14.610 --> 00:13:17.310 align:middle line:84% to 90% of all the variants in the data set. 00:13:17.310 --> 00:13:19.830 align:middle line:84% In other words, by doing these linear combinations 00:13:19.830 --> 00:13:22.890 align:middle line:84% of orthogonal vectors, you're able to capture 00:13:22.890 --> 00:13:23.790 align:middle line:90% most of the variance. 00:13:23.790 --> 00:13:27.108 align:middle line:84% And the nice thing about that is for data sets, 00:13:27.108 --> 00:13:29.400 align:middle line:84% such as gene expression, where you have a lot of highly 00:13:29.400 --> 00:13:30.610 align:middle line:90% correlated behavior. 00:13:30.610 --> 00:13:32.753 align:middle line:84% In other words, there's large numbers 00:13:32.753 --> 00:13:34.920 align:middle line:84% of genes that are really not that independent of one 00:13:34.920 --> 00:13:39.180 align:middle line:84% another, this allows you to reduce the number of variables 00:13:39.180 --> 00:13:42.030 align:middle line:84% that you have to look at an analysis from a very 00:13:42.030 --> 00:13:42.750 align:middle line:90% large number-- 00:13:42.750 --> 00:13:45.370 align:middle line:84% tens of thousands-- to adjust a handful-- 00:13:45.370 --> 00:13:48.030 align:middle line:84% the first principle components that 00:13:48.030 --> 00:13:50.280 align:middle line:84% allow you to capture the majority of variation. 00:13:50.280 --> 00:13:52.260 align:middle line:84% And that's, by the way, the intuition 00:13:52.260 --> 00:13:54.900 align:middle line:84% that some of these face recognition software that you 00:13:54.900 --> 00:13:56.820 align:middle line:84% may have heard about from Homeland Security 00:13:56.820 --> 00:13:57.570 align:middle line:90% has been inspired. 00:13:57.570 --> 00:13:59.650 align:middle line:84% By essentially, taking the eigenface, 00:13:59.650 --> 00:14:01.650 align:middle line:84% where they look at the principal components that 00:14:01.650 --> 00:14:04.395 align:middle line:84% make up the features of the face and reduce it-- reduce 00:14:04.395 --> 00:14:07.680 align:middle line:84% the complexity to these principal components of face 00:14:07.680 --> 00:14:11.460 align:middle line:84% so that not only can you generate an eigenface, 00:14:11.460 --> 00:14:13.140 align:middle line:84% this generic face, but you can also 00:14:13.140 --> 00:14:16.950 align:middle line:84% map people to which faces they're most like. 00:14:16.950 --> 00:14:20.460 align:middle line:84% So it allows you a little bit of play and slot 00:14:20.460 --> 00:14:24.967 align:middle line:84% between different angles and different ways 00:14:24.967 --> 00:14:26.550 align:middle line:84% that you might look in a different day 00:14:26.550 --> 00:14:30.750 align:middle line:84% because it's still capturing the major variants of your face. 00:14:30.750 --> 00:14:32.700 align:middle line:90% So what we did is the following. 00:14:32.700 --> 00:14:35.370 align:middle line:84% Not only was there day five of the mouse, 00:14:35.370 --> 00:14:38.578 align:middle line:84% but we actually captured multiple time points 00:14:38.578 --> 00:14:39.120 align:middle line:90% on the order. 00:14:39.120 --> 00:14:40.370 align:middle line:90% I can't remember now, exactly. 00:14:40.370 --> 00:14:41.790 align:middle line:90% Maybe 15 time points. 00:14:41.790 --> 00:14:45.930 align:middle line:84% And it's a publicly available data set, if you're interested. 00:14:45.930 --> 00:14:47.880 align:middle line:84% 15 time points, going from embryonic days 00:14:47.880 --> 00:14:52.440 align:middle line:90% to postnatal days. 00:14:52.440 --> 00:14:56.590 align:middle line:90% Day 60 I believe, of the mouse. 00:14:56.590 --> 00:15:00.060 align:middle line:84% And then when we did the principal components 00:15:00.060 --> 00:15:06.180 align:middle line:84% of the gene's expression pattern across this time series, 00:15:06.180 --> 00:15:07.560 align:middle line:84% we saw the following when we just 00:15:07.560 --> 00:15:09.690 align:middle line:84% plotted the genes by that position 00:15:09.690 --> 00:15:12.360 align:middle line:84% in this new coordinate system, the first 00:15:12.360 --> 00:15:15.900 align:middle line:90% and second principal components. 00:15:15.900 --> 00:15:21.270 align:middle line:84% And I've colored them by whether they 00:15:21.270 --> 00:15:27.730 align:middle line:84% fell into this section or that half section of the egg, 00:15:27.730 --> 00:15:28.530 align:middle line:90% so to speak. 00:15:28.530 --> 00:15:30.330 align:middle line:90% And why did I color it so? 00:15:30.330 --> 00:15:32.550 align:middle line:84% Because if you actually look at when 00:15:32.550 --> 00:15:36.210 align:middle line:84% these genes had their maximal period, 00:15:36.210 --> 00:15:39.120 align:middle line:84% their maximal time of expression. 00:15:39.120 --> 00:15:42.750 align:middle line:84% The ones in this component, in this area of the two 00:15:42.750 --> 00:15:49.950 align:middle line:84% components, had an early peak and the genes in this component 00:15:49.950 --> 00:15:51.120 align:middle line:90% had a late peak. 00:15:51.120 --> 00:15:56.220 align:middle line:84% So the late-- sorry, the early principal component 00:15:56.220 --> 00:15:59.760 align:middle line:84% and the late set of principal components. 00:15:59.760 --> 00:16:04.230 align:middle line:84% In plane English, by just doing this simple decomposition-- 00:16:04.230 --> 00:16:07.050 align:middle line:84% the simple factoring into two principal components, 00:16:07.050 --> 00:16:09.600 align:middle line:84% we've taken that time series and re-plotted the genes. 00:16:09.600 --> 00:16:11.100 align:middle line:84% And now, we have a set of genes that 00:16:11.100 --> 00:16:13.290 align:middle line:84% are clearly separating themselves out 00:16:13.290 --> 00:16:14.400 align:middle line:90% from another set of genes. 00:16:14.400 --> 00:16:17.190 align:middle line:84% And they're characterized by being early 00:16:17.190 --> 00:16:20.670 align:middle line:90% versus late peaking. 00:16:20.670 --> 00:16:24.000 align:middle line:90% Now, so what? 00:16:24.000 --> 00:16:30.660 align:middle line:84% We then took the human homologues of those genes 00:16:30.660 --> 00:16:32.760 align:middle line:84% and we asked in the medulloblastomas, 00:16:32.760 --> 00:16:36.000 align:middle line:84% which genes compared to non-medulloblastomas were 00:16:36.000 --> 00:16:41.220 align:middle line:84% upregulated and downregulated in the human tumors? 00:16:41.220 --> 00:16:43.020 align:middle line:90% Please? 00:16:43.020 --> 00:16:47.248 align:middle line:84% AUDIENCE: Describe the x-axis and y-axis, what's the-- 00:16:47.248 --> 00:16:49.290 align:middle line:84% ISAAC SAMUEL KOHANE: This is the first and second 00:16:49.290 --> 00:16:50.123 align:middle line:90% principal component. 00:16:50.123 --> 00:16:55.845 align:middle line:84% AUDIENCE: Yeah, but I mean, the number 0,1, 2, 3 and-- 00:16:55.845 --> 00:16:57.345 align:middle line:84% ISAAC SAMUEL KOHANE: It's, I can't-- 00:16:57.345 --> 00:17:00.670 align:middle line:90% 00:17:00.670 --> 00:17:01.920 align:middle line:90% I don't think it's zero based. 00:17:01.920 --> 00:17:03.630 align:middle line:84% I think zero is somewhere around here. 00:17:03.630 --> 00:17:06.460 align:middle line:90% 00:17:06.460 --> 00:17:07.530 align:middle line:90% AUDIENCE: So it's just-- 00:17:07.530 --> 00:17:09.930 align:middle line:84% I mean, I know, so what does that mean? 00:17:09.930 --> 00:17:12.510 align:middle line:90% 00:17:12.510 --> 00:17:14.760 align:middle line:84% ISAAC SAMUEL KOHANE: Its according to the coefficient. 00:17:14.760 --> 00:17:16.470 align:middle line:84% If you position that gene according 00:17:16.470 --> 00:17:20.152 align:middle line:84% to its position along those two principal components. 00:17:20.152 --> 00:17:21.735 align:middle line:84% So each one of these is a single gene. 00:17:21.735 --> 00:17:24.550 align:middle line:90% 00:17:24.550 --> 00:17:29.280 align:middle line:84% So then what we do is we plotted those up and down regulated 00:17:29.280 --> 00:17:33.870 align:middle line:84% genes by whether they fell into-- 00:17:33.870 --> 00:17:36.670 align:middle line:84% we just took those up and down regulated genes 00:17:36.670 --> 00:17:37.920 align:middle line:90% without knowing anything more. 00:17:37.920 --> 00:17:42.420 align:middle line:84% And we say, where do they fall into this late and early 00:17:42.420 --> 00:17:46.230 align:middle line:84% decomposition based on the principal component analysis? 00:17:46.230 --> 00:17:52.350 align:middle line:84% And the simple answer was essentially, 00:17:52.350 --> 00:17:58.380 align:middle line:84% 90% of the genes that were upregulated 00:17:58.380 --> 00:18:04.530 align:middle line:84% fell into the early period of expression. 00:18:04.530 --> 00:18:10.320 align:middle line:84% And similarly, around 90% of the genes that were downregulated-- 00:18:10.320 --> 00:18:11.790 align:middle line:90% in humans, mind you-- 00:18:11.790 --> 00:18:18.420 align:middle line:84% fell into the later phase of development of the mouse. 00:18:18.420 --> 00:18:19.332 align:middle line:90% Go ahead. 00:18:19.332 --> 00:18:20.790 align:middle line:84% AUDIENCE: I have a question the way 00:18:20.790 --> 00:18:23.370 align:middle line:84% you're presenting the data doesn't make apparent why 00:18:23.370 --> 00:18:24.950 align:middle line:90% you need the second component. 00:18:24.950 --> 00:18:34.450 align:middle line:90% 00:18:34.450 --> 00:18:38.020 align:middle line:84% ISAAC SAMUEL KOHANE: That's a very insightful question. 00:18:38.020 --> 00:18:44.860 align:middle line:84% And the answer is that for what I'm describing it, does 00:18:44.860 --> 00:18:48.460 align:middle line:84% not because the separation that you're seeing 00:18:48.460 --> 00:18:51.480 align:middle line:84% is just in one dimension and you're absolutely right. 00:18:51.480 --> 00:18:52.480 align:middle line:90% It doesn't add anything. 00:18:52.480 --> 00:18:54.970 align:middle line:84% Turns out, there were actually different processes 00:18:54.970 --> 00:18:57.280 align:middle line:84% that I'm not describing here that actually 00:18:57.280 --> 00:19:00.433 align:middle line:84% were separated out by going into the second principal component. 00:19:00.433 --> 00:19:02.600 align:middle line:84% But I think you're showing that you're understanding 00:19:02.600 --> 00:19:05.530 align:middle line:84% what the hell I'm talking about by saying, for this distinction 00:19:05.530 --> 00:19:08.590 align:middle line:84% early and late, all you need is the first principal component. 00:19:08.590 --> 00:19:11.020 align:middle line:90% You're absolutely right. 00:19:11.020 --> 00:19:12.670 align:middle line:90% So I just want to-- 00:19:12.670 --> 00:19:16.510 align:middle line:84% those of you who've taken basic pathology in medical school-- 00:19:16.510 --> 00:19:18.280 align:middle line:90% will hear the following. 00:19:18.280 --> 00:19:21.267 align:middle line:90% 00:19:21.267 --> 00:19:23.350 align:middle line:84% I don't know if I have a pointer to it or a slide, 00:19:23.350 --> 00:19:26.470 align:middle line:84% but it's been known by Cohnheim and others 00:19:26.470 --> 00:19:28.900 align:middle line:84% from the turn of the century have speculated 00:19:28.900 --> 00:19:35.020 align:middle line:84% that there's something about the embryonic program 00:19:35.020 --> 00:19:38.530 align:middle line:84% of development that is recapitulated by tumors. 00:19:38.530 --> 00:19:41.155 align:middle line:84% That tumors are similar-- not the same thing-- 00:19:41.155 --> 00:19:43.953 align:middle line:84% are similar to essentially a poorly controlled 00:19:43.953 --> 00:19:45.370 align:middle line:84% embryonic development because when 00:19:45.370 --> 00:19:52.630 align:middle line:84% you look under the microscope, a lot of the tissues in tumors 00:19:52.630 --> 00:19:55.660 align:middle line:84% are not as well differentiated as the tissue from which they 00:19:55.660 --> 00:19:56.350 align:middle line:90% came from. 00:19:56.350 --> 00:19:58.690 align:middle line:84% Liver cancer, if the liver cancer looked exactly 00:19:58.690 --> 00:20:00.670 align:middle line:84% like liver tissue, it wouldn't be liver cancer, 00:20:00.670 --> 00:20:01.840 align:middle line:90% it would be liver. 00:20:01.840 --> 00:20:04.790 align:middle line:84% And so they're actually more primitive versions. 00:20:04.790 --> 00:20:08.200 align:middle line:84% And this actually adds support to this. 00:20:08.200 --> 00:20:13.810 align:middle line:84% In fact, just to check on this, we 00:20:13.810 --> 00:20:16.760 align:middle line:84% did the following experiment, which I don't show here. 00:20:16.760 --> 00:20:20.110 align:middle line:84% So what we've done here, just to place 00:20:20.110 --> 00:20:21.820 align:middle line:84% this in an appropriate perspective, 00:20:21.820 --> 00:20:24.460 align:middle line:84% is take neural tissue from humans-- 00:20:24.460 --> 00:20:27.670 align:middle line:84% human tumors-- and projected it against the components 00:20:27.670 --> 00:20:30.220 align:middle line:90% of mouse development. 00:20:30.220 --> 00:20:32.380 align:middle line:84% And what we C is this nice separation 00:20:32.380 --> 00:20:36.910 align:middle line:84% between upregulated genes and the early phase. 00:20:36.910 --> 00:20:39.340 align:middle line:84% Now maybe all we're showing are general markers 00:20:39.340 --> 00:20:42.610 align:middle line:84% of hyper-proliferation and not something 00:20:42.610 --> 00:20:44.240 align:middle line:90% specific to the system. 00:20:44.240 --> 00:20:48.940 align:middle line:84% So what we did is we took a similar pair, namely, lung 00:20:48.940 --> 00:20:52.180 align:middle line:90% cancer and lung development. 00:20:52.180 --> 00:20:54.340 align:middle line:84% And lung cancer and lung development 00:20:54.340 --> 00:20:56.290 align:middle line:84% showed, essentially, the same relationship. 00:20:56.290 --> 00:20:58.450 align:middle line:84% But then when you took medulloblastoma 00:20:58.450 --> 00:21:01.750 align:middle line:84% against the lung development background, 00:21:01.750 --> 00:21:03.910 align:middle line:84% the separation was no longer as good. 00:21:03.910 --> 00:21:06.040 align:middle line:84% And when you took the lung cancer 00:21:06.040 --> 00:21:10.120 align:middle line:84% against the neural development, it was also not as good. 00:21:10.120 --> 00:21:12.430 align:middle line:84% So what that tells us, it's something much more 00:21:12.430 --> 00:21:20.170 align:middle line:84% nuanced than simply embryogenesis recapitulates 00:21:20.170 --> 00:21:23.990 align:middle line:84% generically angiogenesis or the other way around I should say. 00:21:23.990 --> 00:21:27.490 align:middle line:84% It says that there is part of the differentiation 00:21:27.490 --> 00:21:30.010 align:middle line:84% program of that particular tissue, which 00:21:30.010 --> 00:21:35.050 align:middle line:90% is recapitulated by the tumor. 00:21:35.050 --> 00:21:38.560 align:middle line:84% And that's already interesting because it gives us 00:21:38.560 --> 00:21:40.103 align:middle line:84% some insights and development, but it 00:21:40.103 --> 00:21:42.020 align:middle line:84% gets more interesting when you start thinking, 00:21:42.020 --> 00:21:44.410 align:middle line:84% well, I know a lot about different stages 00:21:44.410 --> 00:21:45.370 align:middle line:90% of development. 00:21:45.370 --> 00:21:47.672 align:middle line:84% Can I actually start pulling out my understanding 00:21:47.672 --> 00:21:49.630 align:middle line:84% of the mechanisms operating at different stages 00:21:49.630 --> 00:21:50.890 align:middle line:84% of the development to understand what's 00:21:50.890 --> 00:21:52.557 align:middle line:84% different about these different cancers? 00:21:52.557 --> 00:21:54.670 align:middle line:90% 00:21:54.670 --> 00:21:57.540 align:middle line:84% So this is just showing you in the time 00:21:57.540 --> 00:21:59.820 align:middle line:84% series, what the principal components were showing you 00:21:59.820 --> 00:22:00.480 align:middle line:90% before. 00:22:00.480 --> 00:22:04.348 align:middle line:84% And in red and green, we're distinguishing the genes 00:22:04.348 --> 00:22:05.640 align:middle line:90% that are up and down regulated. 00:22:05.640 --> 00:22:07.530 align:middle line:84% And you see they have, in the human tumors, 00:22:07.530 --> 00:22:10.530 align:middle line:84% they have different time courses in the mouse data 00:22:10.530 --> 00:22:12.090 align:middle line:90% versus in the-- 00:22:12.090 --> 00:22:13.590 align:middle line:84% different courses of the mouse data, 00:22:13.590 --> 00:22:16.890 align:middle line:84% depending on whether they were upregulated or downregulated 00:22:16.890 --> 00:22:20.580 align:middle line:90% in the human tumors. 00:22:20.580 --> 00:22:23.118 align:middle line:84% And as I said here, this is the slide I wanted to Lobstein 00:22:23.118 --> 00:22:24.660 align:middle line:84% and Cohnheim were among the first who 00:22:24.660 --> 00:22:26.880 align:middle line:84% theorized similarities among human embryogenesis 00:22:26.880 --> 00:22:28.350 align:middle line:90% and biology of cancer cells. 00:22:28.350 --> 00:22:30.180 align:middle line:84% And it's actually not until now that we've 00:22:30.180 --> 00:22:33.720 align:middle line:84% gotten really much more objective evidence 00:22:33.720 --> 00:22:35.460 align:middle line:90% that this is the case. 00:22:35.460 --> 00:22:38.430 align:middle line:84% And the brain tumor classification system that is 00:22:38.430 --> 00:22:41.850 align:middle line:84% used and that was devised by Daly and Cushing in 1926-- 00:22:41.850 --> 00:22:44.340 align:middle line:84% from which our modern taxonomy is derived-- 00:22:44.340 --> 00:22:45.330 align:middle line:90% emphasized this. 00:22:45.330 --> 00:22:48.360 align:middle line:84% But they are very crude taxonomies and descriptive 00:22:48.360 --> 00:22:48.900 align:middle line:90% taxonomies. 00:22:48.900 --> 00:22:51.380 align:middle line:84% And here we're providing perhaps-- 00:22:51.380 --> 00:22:53.760 align:middle line:90% and I'm saying perhaps-- 00:22:53.760 --> 00:22:56.190 align:middle line:90% a much finer classification. 00:22:56.190 --> 00:22:57.790 align:middle line:90% Is that true? 00:22:57.790 --> 00:22:59.920 align:middle line:84% Well, let's turn around the principal components. 00:22:59.920 --> 00:23:06.630 align:middle line:84% Let's look at each tumor as a function of the genes that 00:23:06.630 --> 00:23:10.290 align:middle line:90% are upregulated. 00:23:10.290 --> 00:23:15.630 align:middle line:84% Sorry, let's look at the position of the tumors, now-- 00:23:15.630 --> 00:23:16.560 align:middle line:90% not the genes-- 00:23:16.560 --> 00:23:20.970 align:middle line:84% by the position in the principal component analysis. 00:23:20.970 --> 00:23:23.580 align:middle line:84% And to make a long story short, we 00:23:23.580 --> 00:23:30.550 align:middle line:84% were able to dissect out different days of development. 00:23:30.550 --> 00:23:33.600 align:middle line:84% So these are different mouse samples, and projecting 00:23:33.600 --> 00:23:35.970 align:middle line:84% onto different human cancers onto 00:23:35.970 --> 00:23:40.290 align:middle line:84% the same principal component analysis of these mice. 00:23:40.290 --> 00:23:43.150 align:middle line:84% And this is very interesting in a number of different ways. 00:23:43.150 --> 00:23:46.770 align:middle line:84% One is, we're separating out different days of development 00:23:46.770 --> 00:23:48.520 align:middle line:84% into its principal component space. 00:23:48.520 --> 00:23:52.200 align:middle line:84% But also, separating in that same space, the tumors. 00:23:52.200 --> 00:23:53.700 align:middle line:84% Now that's fundamentally interesting 00:23:53.700 --> 00:23:55.950 align:middle line:84% because it's telling us and the reason why I'm I 00:23:55.950 --> 00:23:57.870 align:middle line:84% bring this up in the genomic medicine class 00:23:57.870 --> 00:24:00.570 align:middle line:84% because if one of the fundamentals of modern medicine 00:24:00.570 --> 00:24:04.420 align:middle line:90% is taxonomization of disease. 00:24:04.420 --> 00:24:10.530 align:middle line:84% And here, I'm providing a very, very quantitative measure 00:24:10.530 --> 00:24:15.610 align:middle line:84% of differentiation of these different tumors 00:24:15.610 --> 00:24:19.360 align:middle line:84% based on their position in the space of development 00:24:19.360 --> 00:24:21.790 align:middle line:84% from a data set that you can never get from humans, 00:24:21.790 --> 00:24:24.010 align:middle line:90% but from mouse. 00:24:24.010 --> 00:24:28.300 align:middle line:84% And this is truly exciting because I 00:24:28.300 --> 00:24:29.830 align:middle line:84% don't have time, well actually, I 00:24:29.830 --> 00:24:31.240 align:middle line:84% don't have a slide to show you because we just 00:24:31.240 --> 00:24:32.050 align:middle line:90% submitted this-- 00:24:32.050 --> 00:24:34.425 align:middle line:84% sorry, it just got accepted to the genes and development, 00:24:34.425 --> 00:24:38.180 align:middle line:84% which is an important journal in this space. 00:24:38.180 --> 00:24:39.910 align:middle line:84% We can actually separate the metastatic 00:24:39.910 --> 00:24:42.130 align:middle line:84% from the non-metastatic medulloblastoma 00:24:42.130 --> 00:24:43.150 align:middle line:90% in the same way. 00:24:43.150 --> 00:24:47.380 align:middle line:84% They separate in the same individual space. 00:24:47.380 --> 00:24:49.630 align:middle line:84% So what does this tell us more generically? 00:24:49.630 --> 00:24:54.820 align:middle line:84% Principal component analysis is not a commonly used tool 00:24:54.820 --> 00:24:57.430 align:middle line:90% for clinicians, obviously. 00:24:57.430 --> 00:24:59.980 align:middle line:84% And yet, here we have a tool that 00:24:59.980 --> 00:25:03.070 align:middle line:84% is creating a much better taxonomy of disease 00:25:03.070 --> 00:25:06.580 align:middle line:84% with predictive power previously not available to us. 00:25:06.580 --> 00:25:08.740 align:middle line:84% And allows us to understand-- because we're 00:25:08.740 --> 00:25:10.000 align:middle line:90% saying what are the genes? 00:25:10.000 --> 00:25:12.000 align:middle line:84% Because remember I said the principal components 00:25:12.000 --> 00:25:14.590 align:middle line:84% are a linear component or linear combinations of genes? 00:25:14.590 --> 00:25:16.840 align:middle line:84% We can say why is this gene in this position? 00:25:16.840 --> 00:25:18.610 align:middle line:84% Which genes have the major weights 00:25:18.610 --> 00:25:20.560 align:middle line:84% which make it responsible for that tumor 00:25:20.560 --> 00:25:21.580 align:middle line:90% being in that position. 00:25:21.580 --> 00:25:24.560 align:middle line:84% Might give us some insights into the underlying tumor mechanism. 00:25:24.560 --> 00:25:26.920 align:middle line:84% And so here, we have a fine grain [INAUDIBLE] 00:25:26.920 --> 00:25:30.430 align:middle line:84% which was not previously available to us. 00:25:30.430 --> 00:25:35.170 align:middle line:84% And I think it's, therefore, not a leap to think that-- 00:25:35.170 --> 00:25:39.430 align:middle line:84% in similar kinds of processes-- that computational approach 00:25:39.430 --> 00:25:43.660 align:middle line:84% to these data sets will allow us to recast our very, very 00:25:43.660 --> 00:25:47.770 align:middle line:90% squishy current taxonomies. 00:25:47.770 --> 00:25:50.560 align:middle line:84% I think I told you in the first lecture, 00:25:50.560 --> 00:25:53.560 align:middle line:84% with a few exceptions, the most notable exceptions 00:25:53.560 --> 00:25:56.380 align:middle line:84% being microbiology, most of our disease classifications 00:25:56.380 --> 00:25:58.570 align:middle line:90% are not mechanism-based. 00:25:58.570 --> 00:26:00.910 align:middle line:84% In microbiology, this classification 00:26:00.910 --> 00:26:02.110 align:middle line:90% is mechanism-based. 00:26:02.110 --> 00:26:03.820 align:middle line:84% It's the organism that's infecting you 00:26:03.820 --> 00:26:05.180 align:middle line:90% that's the name of the disease. 00:26:05.180 --> 00:26:07.748 align:middle line:84% But most of the diseases like lupus, 00:26:07.748 --> 00:26:09.790 align:middle line:84% probably a mixed bag of inflammation inflammatory 00:26:09.790 --> 00:26:12.082 align:middle line:84% diseases that we don't know, polycystic ovary diseases, 00:26:12.082 --> 00:26:15.520 align:middle line:84% we're describing symptoms, not pathophysiology. 00:26:15.520 --> 00:26:17.750 align:middle line:84% And I think this is allowing us to get much, 00:26:17.750 --> 00:26:19.840 align:middle line:84% much closer and therefore, a much more 00:26:19.840 --> 00:26:22.210 align:middle line:84% robust and fine-grained understanding of that. 00:26:22.210 --> 00:26:24.780 align:middle line:90% 00:26:24.780 --> 00:26:27.020 align:middle line:84% So my interim summary here is that this computing 00:26:27.020 --> 00:26:28.788 align:middle line:84% allows us to identify some pathways 00:26:28.788 --> 00:26:31.080 align:middle line:84% and provides more importantly, a natural classification 00:26:31.080 --> 00:26:32.450 align:middle line:90% of disease with further insight. 00:26:32.450 --> 00:26:35.130 align:middle line:90% 00:26:35.130 --> 00:26:37.370 align:middle line:84% Apparently, I don't know how to open my book. 00:26:37.370 --> 00:26:38.120 align:middle line:90% Go ahead. 00:26:38.120 --> 00:26:41.620 align:middle line:84% AUDIENCE: Did you try to redefine bacterias 00:26:41.620 --> 00:26:49.630 align:middle line:84% that were [INAUDIBLE] against the optimized use 00:26:49.630 --> 00:26:53.620 align:middle line:84% of combination of genes that you expect 00:26:53.620 --> 00:26:58.620 align:middle line:84% to be a developmental program, or something to that level? 00:26:58.620 --> 00:27:01.260 align:middle line:84% ISAAC SAMUEL KOHANE: So Jose is asking 00:27:01.260 --> 00:27:03.657 align:middle line:84% a very interesting question which the short answer is no, 00:27:03.657 --> 00:27:05.490 align:middle line:84% but I think it's a very good question, which 00:27:05.490 --> 00:27:07.590 align:middle line:84% is-- but I wouldn't know exactly where to start. 00:27:07.590 --> 00:27:12.780 align:middle line:84% But surely there is a thoughtful way to respond to it. 00:27:12.780 --> 00:27:15.395 align:middle line:84% Could you take linear combinations of genes-- 00:27:15.395 --> 00:27:16.770 align:middle line:84% and I wouldn't know exactly which 00:27:16.770 --> 00:27:19.103 align:middle line:84% linear combination of genes-- but maybe just using those 00:27:19.103 --> 00:27:21.672 align:middle line:84% as a starting point, perhaps, genes that 00:27:21.672 --> 00:27:23.880 align:middle line:84% have been implicated at various stages of development 00:27:23.880 --> 00:27:25.230 align:middle line:90% and using those. 00:27:25.230 --> 00:27:27.180 align:middle line:84% And I think that's a good way to go. 00:27:27.180 --> 00:27:28.950 align:middle line:84% My understanding of developmental Biology 00:27:28.950 --> 00:27:31.920 align:middle line:84% is unfortunately, it's very sparse knowledge 00:27:31.920 --> 00:27:32.860 align:middle line:90% at different stages. 00:27:32.860 --> 00:27:34.318 align:middle line:84% So it'd be hard to know which genes 00:27:34.318 --> 00:27:35.520 align:middle line:90% are generally informative. 00:27:35.520 --> 00:27:37.620 align:middle line:84% Nonetheless, I'm a big fan of the approach 00:27:37.620 --> 00:27:40.120 align:middle line:84% that you're suggesting, which is a knowledge based approach. 00:27:40.120 --> 00:27:42.360 align:middle line:84% What we're doing here was clearly knowledge free. 00:27:42.360 --> 00:27:44.450 align:middle line:84% We were just taking the data and just 00:27:44.450 --> 00:27:46.200 align:middle line:84% saying, what explains the maximum variance 00:27:46.200 --> 00:27:47.710 align:middle line:90% and going from there. 00:27:47.710 --> 00:27:50.760 align:middle line:84% But if we could conceive that process or the set of genes 00:27:50.760 --> 00:27:52.260 align:middle line:84% that we knew were implicated, I bet 00:27:52.260 --> 00:27:54.820 align:middle line:84% we'd have much better resolution than we have currently. 00:27:54.820 --> 00:27:55.740 align:middle line:84% So I think it's a very good thought, 00:27:55.740 --> 00:27:56.823 align:middle line:90% but we have not done that. 00:27:56.823 --> 00:28:01.250 align:middle line:90% 00:28:01.250 --> 00:28:08.630 align:middle line:84% Which brings me now to another aspect of computing. 00:28:08.630 --> 00:28:11.283 align:middle line:90% I've shown you guys this slide. 00:28:11.283 --> 00:28:12.200 align:middle line:90% Does it look familiar? 00:28:12.200 --> 00:28:13.250 align:middle line:90% Good. 00:28:13.250 --> 00:28:17.630 align:middle line:84% So here's the fantasy that most deans of medical schools 00:28:17.630 --> 00:28:19.790 align:middle line:90% and heads of hospitals have. 00:28:19.790 --> 00:28:21.770 align:middle line:84% And, again, this is not a very technical point. 00:28:21.770 --> 00:28:23.187 align:middle line:84% It's more of a sociological point, 00:28:23.187 --> 00:28:25.460 align:middle line:84% but it will end up being technical because it can tell 00:28:25.460 --> 00:28:28.430 align:middle line:84% you about a whole other area of biocomputing that I think 00:28:28.430 --> 00:28:33.540 align:middle line:84% is ripe for the picking, but it's also a rate limiting step. 00:28:33.540 --> 00:28:35.023 align:middle line:90% So they all have the fantasy. 00:28:35.023 --> 00:28:36.440 align:middle line:84% Now that they've seen things like, 00:28:36.440 --> 00:28:40.490 align:middle line:84% oh, yes I can predict different outcomes of different tumors, 00:28:40.490 --> 00:28:45.770 align:middle line:84% let's say, I want to be able to develop a set of targets 00:28:45.770 --> 00:28:47.930 align:middle line:84% that a drug company is going to pay me a lot for. 00:28:47.930 --> 00:28:50.480 align:middle line:84% So the fantasy is the nice doctor talks 00:28:50.480 --> 00:28:53.420 align:middle line:90% to the nice patient. 00:28:53.420 --> 00:28:55.660 align:middle line:84% He gets consent in the nice hospital. 00:28:55.660 --> 00:28:58.330 align:middle line:90% 00:28:58.330 --> 00:29:00.520 align:middle line:90% They have a family history. 00:29:00.520 --> 00:29:03.580 align:middle line:90% They bank the tissue obtained. 00:29:03.580 --> 00:29:05.980 align:middle line:90% They do some genomic analyzes. 00:29:05.980 --> 00:29:08.050 align:middle line:90% They do a clinical annotation. 00:29:08.050 --> 00:29:11.170 align:middle line:84% They do some fancy mumbo jumbo bio-informatics. 00:29:11.170 --> 00:29:13.618 align:middle line:84% And lo and behold, they're going to have the target 00:29:13.618 --> 00:29:15.910 align:middle line:84% that a drug company is going to pay a lot of money for. 00:29:15.910 --> 00:29:18.040 align:middle line:84% That indeed is-- for those of you 00:29:18.040 --> 00:29:21.730 align:middle line:84% who are more entrepreneurial oriented, that's 00:29:21.730 --> 00:29:23.350 align:middle line:90% the fantasy that's launched. 00:29:23.350 --> 00:29:28.000 align:middle line:84% Many, many Venture Capital funded ships, 00:29:28.000 --> 00:29:30.700 align:middle line:84% like the decode project and others. 00:29:30.700 --> 00:29:33.830 align:middle line:84% And the problem is the following. 00:29:33.830 --> 00:29:38.200 align:middle line:84% It's that this central piece, the phenotypic annotation, 00:29:38.200 --> 00:29:41.757 align:middle line:90% ends up being the hardest part. 00:29:41.757 --> 00:29:43.840 align:middle line:84% Understanding what really happened to the patient. 00:29:43.840 --> 00:29:45.340 align:middle line:84% Did the patient really have a tumor? 00:29:45.340 --> 00:29:46.900 align:middle line:84% How old were when they had the tumor? 00:29:46.900 --> 00:29:48.430 align:middle line:84% What drugs were they responsive to? 00:29:48.430 --> 00:29:50.080 align:middle line:90% That is the hard part. 00:29:50.080 --> 00:29:56.800 align:middle line:84% And any of you have ever looked at a clinical chart 00:29:56.800 --> 00:29:58.960 align:middle line:90% can understand why that is. 00:29:58.960 --> 00:30:01.690 align:middle line:84% There's just very little useful machine-able data 00:30:01.690 --> 00:30:03.280 align:middle line:90% in the clinical chart. 00:30:03.280 --> 00:30:05.410 align:middle line:84% Because in fact, the real fantasy is the following. 00:30:05.410 --> 00:30:09.670 align:middle line:84% And what I call [? Ratwitchz ?] at the Medical 00:30:09.670 --> 00:30:12.100 align:middle line:90% College of Wisconsin-- 00:30:12.100 --> 00:30:15.940 align:middle line:84% what they do is they take rats, consomic rats, 00:30:15.940 --> 00:30:18.970 align:middle line:84% where they have systematically substituted 00:30:18.970 --> 00:30:22.240 align:middle line:90% one chromosome for another. 00:30:22.240 --> 00:30:26.170 align:middle line:84% So which would give them a very big efficiencies of identifying 00:30:26.170 --> 00:30:29.890 align:middle line:90% linkage with a different trait. 00:30:29.890 --> 00:30:32.590 align:middle line:84% What they do is they take these different chromosomic rats-- 00:30:32.590 --> 00:30:36.430 align:middle line:84% different strains with well-understood genotypes-- 00:30:36.430 --> 00:30:41.890 align:middle line:84% and they exposed them to very well characterized 00:30:41.890 --> 00:30:45.460 align:middle line:84% environmental exposures, such as hypoxia. 00:30:45.460 --> 00:30:47.050 align:middle line:84% They put them in the oxygen equivalent 00:30:47.050 --> 00:30:50.890 align:middle line:84% of top of the mountain for several weeks. 00:30:50.890 --> 00:30:53.350 align:middle line:84% Or they give them a high salt diet or high fat diet. 00:30:53.350 --> 00:30:55.870 align:middle line:90% Or they volume deplete them. 00:30:55.870 --> 00:30:57.760 align:middle line:84% And after that wonderful experience, 00:30:57.760 --> 00:31:00.880 align:middle line:84% they then physically reconstruct them and rip out their heart 00:31:00.880 --> 00:31:03.250 align:middle line:84% and put their hearts on these prep machines 00:31:03.250 --> 00:31:05.960 align:middle line:84% where they can look at the contractility curves 00:31:05.960 --> 00:31:12.370 align:middle line:84% of the heart in the perfused prep. 00:31:12.370 --> 00:31:14.500 align:middle line:84% And that's great because now you have 00:31:14.500 --> 00:31:18.745 align:middle line:84% a well-defined phenotypic environmental characterization. 00:31:18.745 --> 00:31:20.900 align:middle line:84% You know the genotype and you have the expression. 00:31:20.900 --> 00:31:22.567 align:middle line:84% And by the way, let me say and point you 00:31:22.567 --> 00:31:26.020 align:middle line:84% to the following site, PGA.mcu.edu. 00:31:26.020 --> 00:31:28.480 align:middle line:90% They have all that data online. 00:31:28.480 --> 00:31:30.580 align:middle line:84% The phenotypes, the genotypes and expression. 00:31:30.580 --> 00:31:34.248 align:middle line:84% And they do it for cardiac phenotypes, kidney phenotypes, 00:31:34.248 --> 00:31:36.790 align:middle line:84% lots of different phenotypes and different parts of medicine. 00:31:36.790 --> 00:31:38.320 align:middle line:90% And so this is the fantasy. 00:31:38.320 --> 00:31:41.440 align:middle line:84% But of course, patients would probably objectively 00:31:41.440 --> 00:31:43.910 align:middle line:84% subjected them to this kind of treatment. 00:31:43.910 --> 00:31:48.580 align:middle line:84% So this high throughput phenotype fantasy 00:31:48.580 --> 00:31:53.300 align:middle line:84% just is unlikely to happen anytime soon. 00:31:53.300 --> 00:31:58.450 align:middle line:84% So we took a very different approach 00:31:58.450 --> 00:32:02.620 align:middle line:84% when we were dealt with this problem. 00:32:02.620 --> 00:32:06.370 align:middle line:84% And that was how to get sufficient numbers of samples 00:32:06.370 --> 00:32:08.290 align:middle line:90% in the right amount. 00:32:08.290 --> 00:32:11.500 align:middle line:84% And you should know that there are companies like Ardeas, 00:32:11.500 --> 00:32:15.580 align:middle line:84% which have been given millions of dollars of venture capital 00:32:15.580 --> 00:32:18.040 align:middle line:84% unsuccessfully to try to solve this problem. 00:32:18.040 --> 00:32:20.230 align:middle line:84% To get enough tissues in the right amount. 00:32:20.230 --> 00:32:26.290 align:middle line:84% So the National Cancer Institute put out 00:32:26.290 --> 00:32:29.110 align:middle line:84% a request for applications saying, 00:32:29.110 --> 00:32:31.480 align:middle line:84% listen, there's thousands of tissues 00:32:31.480 --> 00:32:36.370 align:middle line:84% available throughout the pathology repositories 00:32:36.370 --> 00:32:39.550 align:middle line:84% of our country that we want to be able to do genomic studies. 00:32:39.550 --> 00:32:40.640 align:middle line:90% How to go about that? 00:32:40.640 --> 00:32:46.360 align:middle line:84% So inspired by Napster and Gnutella-- 00:32:46.360 --> 00:32:48.340 align:middle line:84% and for those of you who don't know what the-- 00:32:48.340 --> 00:32:51.400 align:middle line:84% those of you who are not aficionados 00:32:51.400 --> 00:32:54.880 align:middle line:84% of file sharing, Napster is a file sharing service 00:32:54.880 --> 00:32:57.010 align:middle line:84% where there actually is a centralized. 00:32:57.010 --> 00:32:59.350 align:middle line:84% But the data themselves are decentralized 00:32:59.350 --> 00:33:00.400 align:middle line:90% in multiple directories. 00:33:00.400 --> 00:33:02.958 align:middle line:84% Guntella, in order to avoid getting anybody sued, 00:33:02.958 --> 00:33:04.000 align:middle line:90% it's fully decentralized. 00:33:04.000 --> 00:33:05.860 align:middle line:84% The Directories are decentralized 00:33:05.860 --> 00:33:09.970 align:middle line:90% and the data is decentralized. 00:33:09.970 --> 00:33:13.240 align:middle line:84% And what we built, the shared Pathology Informatics network, 00:33:13.240 --> 00:33:16.690 align:middle line:84% SPIN, that's funded by about $7 million from the NCI, 00:33:16.690 --> 00:33:18.230 align:middle line:90% is exactly the same thing. 00:33:18.230 --> 00:33:23.020 align:middle line:84% And what we did is to take advantage of the twin obsession 00:33:23.020 --> 00:33:25.390 align:middle line:90% compulsions of pathologists. 00:33:25.390 --> 00:33:29.350 align:middle line:84% Pathologists are probably the best first taxonomists 00:33:29.350 --> 00:33:31.130 align:middle line:90% of the medical profession. 00:33:31.130 --> 00:33:35.500 align:middle line:84% They actually categorize organs and specimens 00:33:35.500 --> 00:33:37.390 align:middle line:90% better than anybody else. 00:33:37.390 --> 00:33:41.180 align:middle line:84% Second, they're obsessive compulsive collectors 00:33:41.180 --> 00:33:41.680 align:middle line:90% of tissue. 00:33:41.680 --> 00:33:44.900 align:middle line:84% And they keep tissue around for years and years. 00:33:44.900 --> 00:33:47.110 align:middle line:84% And so we thought that if we could 00:33:47.110 --> 00:33:49.110 align:middle line:84% take advantage of those twin compulsions, 00:33:49.110 --> 00:33:50.780 align:middle line:90% we could actually be successful. 00:33:50.780 --> 00:33:52.570 align:middle line:84% So at a bird's eye view what we did 00:33:52.570 --> 00:33:54.980 align:middle line:84% was the following, which was to create a network whereby 00:33:54.980 --> 00:33:58.310 align:middle line:84% a user, with a web browser, could send a query to his Query 00:33:58.310 --> 00:34:00.620 align:middle line:84% Composer, which would then send out 00:34:00.620 --> 00:34:03.440 align:middle line:84% a query to this nebulous thing called SPIN network. 00:34:03.440 --> 00:34:07.460 align:middle line:84% Which would, using this peer to peer technology, 00:34:07.460 --> 00:34:12.020 align:middle line:84% respond which samples were out there. 00:34:12.020 --> 00:34:16.580 align:middle line:84% More specifically, each tissue bank or institution 00:34:16.580 --> 00:34:21.020 align:middle line:84% would have its own node on the network, on the internet. 00:34:21.020 --> 00:34:23.780 align:middle line:84% And we provide them these open source 00:34:23.780 --> 00:34:27.530 align:middle line:84% three tools which allow them to extract from these pathology 00:34:27.530 --> 00:34:29.960 align:middle line:84% databases because there's a textual report 00:34:29.960 --> 00:34:31.909 align:middle line:90% with each pathology sample. 00:34:31.909 --> 00:34:36.440 align:middle line:84% Both the anonymized-- because we have this anonymization 00:34:36.440 --> 00:34:36.980 align:middle line:90% program-- 00:34:36.980 --> 00:34:43.909 align:middle line:84% the anonymized textual report, and code a few data elements 00:34:43.909 --> 00:34:44.960 align:middle line:90% to enable search. 00:34:44.960 --> 00:34:48.500 align:middle line:84% So that when you put in the query, 00:34:48.500 --> 00:34:51.380 align:middle line:84% the query percolates in a fully distributed way. 00:34:51.380 --> 00:34:53.179 align:middle line:84% There's no center to this network. 00:34:53.179 --> 00:34:56.389 align:middle line:84% This query percolates throughout the system. 00:34:56.389 --> 00:35:00.410 align:middle line:84% All the nodes respond eventually, if they can, 00:35:00.410 --> 00:35:06.282 align:middle line:84% and you get back the sum of all specimens involved. 00:35:06.282 --> 00:35:08.240 align:middle line:84% And those of you even in the medical profession 00:35:08.240 --> 00:35:10.590 align:middle line:84% probably don't appreciate the following point. 00:35:10.590 --> 00:35:12.560 align:middle line:84% The reason we created a peer to peer system 00:35:12.560 --> 00:35:15.995 align:middle line:84% is that pathology tissue banks are jealously 00:35:15.995 --> 00:35:18.620 align:middle line:84% guarded by the pathologists and by the surgeons who built them. 00:35:18.620 --> 00:35:20.510 align:middle line:84% They just don't want to share data. 00:35:20.510 --> 00:35:23.480 align:middle line:84% And it's essentially impossible today to know-- 00:35:23.480 --> 00:35:26.660 align:middle line:84% even across the street and the Brigham-- what samples 00:35:26.660 --> 00:35:27.500 align:middle line:90% they have. 00:35:27.500 --> 00:35:31.490 align:middle line:84% And by allowing them to actually fully control their own node 00:35:31.490 --> 00:35:33.500 align:middle line:84% and control what they expose to the outside, 00:35:33.500 --> 00:35:36.860 align:middle line:84% we're able to overcome the sociological obstacles 00:35:36.860 --> 00:35:40.350 align:middle line:84% for data sharing so that we can actually do this. 00:35:40.350 --> 00:35:43.407 align:middle line:84% So for instance, we've done this now, 00:35:43.407 --> 00:35:44.990 align:middle line:84% this is actually an out-of-date slide, 00:35:44.990 --> 00:35:48.170 align:middle line:84% not only across Harvard and UCLA, 00:35:48.170 --> 00:35:51.350 align:middle line:84% but now we've gone live with Pittsburgh 00:35:51.350 --> 00:35:52.460 align:middle line:90% and a few other sites. 00:35:52.460 --> 00:35:59.610 align:middle line:84% So that today, if you're a registered user, 00:35:59.610 --> 00:36:04.760 align:middle line:84% you can actually send out a query to all those institutions 00:36:04.760 --> 00:36:09.890 align:middle line:90% that you see on the right. 00:36:09.890 --> 00:36:13.760 align:middle line:84% And the query then percolates across the country. 00:36:13.760 --> 00:36:17.180 align:middle line:84% And this query takes on the order of 10 seconds 00:36:17.180 --> 00:36:18.590 align:middle line:90% on a bad day to execute. 00:36:18.590 --> 00:36:22.190 align:middle line:84% So for instance, we want to know how many distinct specimens 00:36:22.190 --> 00:36:25.470 align:middle line:84% can we have of renal transplant patients? 00:36:25.470 --> 00:36:29.790 align:middle line:84% And that's typically hard to come by. 00:36:29.790 --> 00:36:31.030 align:middle line:90% And here is a response. 00:36:31.030 --> 00:36:32.490 align:middle line:84% So this is out of seven nodes, this 00:36:32.490 --> 00:36:36.060 align:middle line:84% is a query I did back in October of last year. 00:36:36.060 --> 00:36:40.080 align:middle line:84% Out of the seven nodes that we had up at that time, 00:36:40.080 --> 00:36:42.390 align:middle line:84% with only two nodes responding, we already 00:36:42.390 --> 00:36:50.460 align:middle line:84% had 20,000 specimens identified with an age distribution shown 00:36:50.460 --> 00:36:51.060 align:middle line:90% as here. 00:36:51.060 --> 00:36:52.710 align:middle line:84% By the way, I'm aghast at the fact 00:36:52.710 --> 00:36:56.563 align:middle line:84% that apparently 34 specimens of patients ages 90 to 99. 00:36:56.563 --> 00:36:58.980 align:middle line:84% I wish it was true that was an error, but I looked into it 00:36:58.980 --> 00:36:59.820 align:middle line:90% and it's not. 00:36:59.820 --> 00:37:02.250 align:middle line:84% So some weirdos are actually transplanting 00:37:02.250 --> 00:37:04.108 align:middle line:90% things in very old patients. 00:37:04.108 --> 00:37:05.400 align:middle line:90% Would not happening in England. 00:37:05.400 --> 00:37:07.560 align:middle line:84% In any case, here's the age distribution. 00:37:07.560 --> 00:37:10.230 align:middle line:84% And if you click on the full text, 00:37:10.230 --> 00:37:12.300 align:middle line:84% and you have the appropriate privileges, 00:37:12.300 --> 00:37:15.570 align:middle line:84% you can actually see the full anonymized text that 00:37:15.570 --> 00:37:17.470 align:middle line:90% show a report of that patient. 00:37:17.470 --> 00:37:22.600 align:middle line:84% And so I make this point for a number of reasons. 00:37:22.600 --> 00:37:28.470 align:middle line:84% First, we're successfully mining the obsessive compulsive 00:37:28.470 --> 00:37:29.880 align:middle line:90% feature of pathologists. 00:37:29.880 --> 00:37:35.670 align:middle line:84% Second, I hinted to you that some of these outcomes analysis 00:37:35.670 --> 00:37:38.460 align:middle line:84% using functional genomic measures, such as the lymphoma 00:37:38.460 --> 00:37:41.070 align:middle line:84% study, are essentially not well reproduced 00:37:41.070 --> 00:37:42.120 align:middle line:90% in different studies. 00:37:42.120 --> 00:37:46.050 align:middle line:84% Whatever the reason, they are many. 00:37:46.050 --> 00:37:48.360 align:middle line:84% A lot of the ills in the analysis 00:37:48.360 --> 00:37:51.280 align:middle line:90% are excused by large numbers. 00:37:51.280 --> 00:37:54.780 align:middle line:84% So if we went from having 100 patients to having 20,000 00:37:54.780 --> 00:37:57.720 align:middle line:84% patients, I assure you that the differences 00:37:57.720 --> 00:37:59.550 align:middle line:84% in different machine learning techniques 00:37:59.550 --> 00:38:01.810 align:middle line:90% would matter very little. 00:38:01.810 --> 00:38:03.330 align:middle line:84% And so this is how we're beginning 00:38:03.330 --> 00:38:06.180 align:middle line:84% to approach, how do we harness the phenotypes that 00:38:06.180 --> 00:38:07.890 align:middle line:84% are out there that are available, 00:38:07.890 --> 00:38:10.772 align:middle line:84% that are not available otherwise through a completely different 00:38:10.772 --> 00:38:11.730 align:middle line:90% kind of bioinformatics? 00:38:11.730 --> 00:38:15.450 align:middle line:84% This is more infrastructure bioinformatics. 00:38:15.450 --> 00:38:17.010 align:middle line:90% But it's the rate-limiting step. 00:38:17.010 --> 00:38:20.947 align:middle line:84% We can work on a dozen samples until we're blue in the face. 00:38:20.947 --> 00:38:23.280 align:middle line:84% We really won't know how much we're overfitting our data 00:38:23.280 --> 00:38:27.318 align:middle line:90% until we have large samples. 00:38:27.318 --> 00:38:29.610 align:middle line:84% AUDIENCE: But the very first thing that you can request 00:38:29.610 --> 00:38:31.162 align:middle line:90% is that for [INAUDIBLE]. 00:38:31.162 --> 00:38:33.370 align:middle line:84% ISAAC SAMUEL KOHANE: Yeah, I wish it was that simple. 00:38:33.370 --> 00:38:35.990 align:middle line:90% So that shows you're awake. 00:38:35.990 --> 00:38:38.137 align:middle line:84% The next question is, how do I get the samples? 00:38:38.137 --> 00:38:40.470 align:middle line:84% So I wish it was true that once you identified a sample, 00:38:40.470 --> 00:38:42.630 align:middle line:84% you just click the box, put your credit card 00:38:42.630 --> 00:38:45.550 align:middle line:84% and you get FedExed the tissue sample. 00:38:45.550 --> 00:38:47.940 align:middle line:84% So what happens at that point is that you 00:38:47.940 --> 00:38:49.800 align:middle line:84% start a dialogue with investigator 00:38:49.800 --> 00:38:54.120 align:middle line:84% and the Institutional Review Board of that tissue bank. 00:38:54.120 --> 00:38:55.740 align:middle line:90% So that's the bad news. 00:38:55.740 --> 00:38:57.990 align:middle line:84% The good news is, we've solved one part of the problem 00:38:57.990 --> 00:39:00.090 align:middle line:84% is figuring out where the specimens are to find out 00:39:00.090 --> 00:39:01.170 align:middle line:90% if your study is feasible. 00:39:01.170 --> 00:39:03.000 align:middle line:84% And people are already being used as tools right now 00:39:03.000 --> 00:39:04.542 align:middle line:84% to write grants because they can say, 00:39:04.542 --> 00:39:06.810 align:middle line:84% I know that there's [INAUDIBLE] specimens out 00:39:06.810 --> 00:39:09.540 align:middle line:84% and they don't necessarily get access to the tissue for that. 00:39:09.540 --> 00:39:11.340 align:middle line:84% But they get in-principle agreement 00:39:11.340 --> 00:39:13.350 align:middle line:84% from the collaborators that if they get funded, 00:39:13.350 --> 00:39:15.360 align:middle line:84% they'll do this as a collaboration. 00:39:15.360 --> 00:39:17.350 align:middle line:84% But you're absolutely right to do it. 00:39:17.350 --> 00:39:21.150 align:middle line:84% So this is just another example how we're using computation 00:39:21.150 --> 00:39:24.270 align:middle line:84% to overcome the obstacles to doing genomic medicine. 00:39:24.270 --> 00:39:27.750 align:middle line:84% Let me give you one last case history 00:39:27.750 --> 00:39:31.510 align:middle line:84% and then we'll go to the problem set. 00:39:31.510 --> 00:39:36.743 align:middle line:84% So let's talk about how we can use computing again 00:39:36.743 --> 00:39:38.160 align:middle line:84% to have a different understanding, 00:39:38.160 --> 00:39:40.110 align:middle line:84% and how we use computing allows to have 00:39:40.110 --> 00:39:45.310 align:middle line:84% a different understanding of for instance, gene regulation. 00:39:45.310 --> 00:39:49.500 align:middle line:90% So this is time on x-axis. 00:39:49.500 --> 00:39:53.960 align:middle line:84% And this is some arbitrary value of expression. 00:39:53.960 --> 00:39:55.700 align:middle line:84% Are these two time courses the same? 00:39:55.700 --> 00:39:59.210 align:middle line:90% 00:39:59.210 --> 00:40:03.080 align:middle line:84% Are those two genes developed by the same process? 00:40:03.080 --> 00:40:03.920 align:middle line:90% Yes or no? 00:40:03.920 --> 00:40:07.670 align:middle line:90% 00:40:07.670 --> 00:40:10.070 align:middle line:84% You could make an argument either way, right? 00:40:10.070 --> 00:40:12.120 align:middle line:90% Maybe they're the same. 00:40:12.120 --> 00:40:12.980 align:middle line:90% How about this one? 00:40:12.980 --> 00:40:17.720 align:middle line:90% 00:40:17.720 --> 00:40:19.935 align:middle line:84% Maybe they're the same, maybe not. 00:40:19.935 --> 00:40:21.560 align:middle line:84% I tell you, the stuff was pretty noisy. 00:40:21.560 --> 00:40:25.340 align:middle line:84% Maybe that's just noise, a priori, who knows. 00:40:25.340 --> 00:40:27.640 align:middle line:90% Now does order matter? 00:40:27.640 --> 00:40:29.480 align:middle line:90% It does in our analysis. 00:40:29.480 --> 00:40:31.820 align:middle line:90% Does time matter? 00:40:31.820 --> 00:40:33.500 align:middle line:90% Yes, good. 00:40:33.500 --> 00:40:39.170 align:middle line:84% Because it should be true that the likelihood of me being here 00:40:39.170 --> 00:40:44.000 align:middle line:84% now is informed much more by where I was five minutes ago 00:40:44.000 --> 00:40:46.310 align:middle line:84% and where I was a day ago, right? 00:40:46.310 --> 00:40:51.260 align:middle line:84% That's a fundamental Markovian property of Zack, 00:40:51.260 --> 00:40:55.010 align:middle line:90% and most physical processes. 00:40:55.010 --> 00:40:57.140 align:middle line:84% Now if there's a correlation coefficient at all, 00:40:57.140 --> 00:40:59.360 align:middle line:90% address that Markovian process. 00:40:59.360 --> 00:41:02.990 align:middle line:84% In other words, if you take two stocks, and shuffle the timing. 00:41:02.990 --> 00:41:05.130 align:middle line:84% Let's say two stocks, Apple and Microsoft, 00:41:05.130 --> 00:41:06.080 align:middle line:84% let's say for the sake of argument, 00:41:06.080 --> 00:41:07.490 align:middle line:90% that they are highly correlated. 00:41:07.490 --> 00:41:09.920 align:middle line:84% They the high correlation coefficient. 00:41:09.920 --> 00:41:14.210 align:middle line:84% If you permute the days that they maintain 00:41:14.210 --> 00:41:18.470 align:middle line:84% the same pair of stock values for that day 00:41:18.470 --> 00:41:21.230 align:middle line:84% together, to shuffle the days, is the coefficient 00:41:21.230 --> 00:41:23.450 align:middle line:90% the same or different? 00:41:23.450 --> 00:41:24.510 align:middle line:90% Same. 00:41:24.510 --> 00:41:27.590 align:middle line:84% So correlation coefficient does not in fact 00:41:27.590 --> 00:41:31.370 align:middle line:90% capture the effects of time. 00:41:31.370 --> 00:41:36.750 align:middle line:84% Now, as I hope you're aware from our previous discussions, 00:41:36.750 --> 00:41:40.190 align:middle line:84% a lot of the clustering that's been done even 00:41:40.190 --> 00:41:42.680 align:middle line:84% on time series data, is done on the basis 00:41:42.680 --> 00:41:44.930 align:middle line:90% of correlation coefficients. 00:41:44.930 --> 00:41:51.450 align:middle line:84% People in those dendrograms bring together genes that share 00:41:51.450 --> 00:41:53.450 align:middle line:84% the same expression pattern as measured 00:41:53.450 --> 00:41:55.470 align:middle line:90% by the correlation coefficient. 00:41:55.470 --> 00:41:57.695 align:middle line:84% So what consequences does that have? 00:41:57.695 --> 00:42:00.290 align:middle line:90% 00:42:00.290 --> 00:42:02.440 align:middle line:84% Let's go back to a very old study. 00:42:02.440 --> 00:42:06.280 align:middle line:84% This is [INAUDIBLE] study back in 1999. 00:42:06.280 --> 00:42:09.970 align:middle line:84% A classical Jewish study where they snipped off the-- 00:42:09.970 --> 00:42:13.540 align:middle line:84% no, they took a fibroblast from foreskins 00:42:13.540 --> 00:42:17.740 align:middle line:84% and they measured the gene expression pattern 00:42:17.740 --> 00:42:24.820 align:middle line:84% over time of the transcriptome of these foreskins 00:42:24.820 --> 00:42:27.130 align:middle line:90% after being exposed to serum. 00:42:27.130 --> 00:42:32.170 align:middle line:84% So you see here, every column is zero hour, 50 minutes and so on 00:42:32.170 --> 00:42:36.310 align:middle line:90% till 24 hours of each gene. 00:42:36.310 --> 00:42:38.710 align:middle line:84% And then they did this act of creativity, 00:42:38.710 --> 00:42:42.610 align:middle line:84% which is they draw lines next to the dendrogram and then cut out 00:42:42.610 --> 00:42:45.790 align:middle line:84% the blocks each line and say, oh look, 00:42:45.790 --> 00:42:48.730 align:middle line:84% there is a bunch of these genes and they seem to -- 00:42:48.730 --> 00:42:50.440 align:middle line:84% for instance, coagulation hemostasis. 00:42:50.440 --> 00:42:53.800 align:middle line:84% There's a tissue factor, pathway inhibitor and so on. 00:42:53.800 --> 00:42:55.550 align:middle line:90% These are coagulation involved. 00:42:55.550 --> 00:42:57.620 align:middle line:84% These are cell cycle proliferations. 00:42:57.620 --> 00:42:59.500 align:middle line:90% This is inflammation. 00:42:59.500 --> 00:43:03.310 align:middle line:84% Oh look, these are angiogenesis involved genes. 00:43:03.310 --> 00:43:04.570 align:middle line:90% These are cytoskeletal. 00:43:04.570 --> 00:43:06.730 align:middle line:84% And they all seem to be clustered together. 00:43:06.730 --> 00:43:11.500 align:middle line:84% And it's again, I use this metaphor all too often, 00:43:11.500 --> 00:43:12.710 align:middle line:90% but it's so true. 00:43:12.710 --> 00:43:14.800 align:middle line:84% This is very much like the dog at the opera. 00:43:14.800 --> 00:43:16.720 align:middle line:84% It's a miracle it can sing at all 00:43:16.720 --> 00:43:20.350 align:middle line:84% and you don't criticize how well it's singing. 00:43:20.350 --> 00:43:22.927 align:middle line:84% Because what they've done is this act of creativity. 00:43:22.927 --> 00:43:24.760 align:middle line:84% They've actually looked at the data of genes 00:43:24.760 --> 00:43:27.775 align:middle line:84% and they draw these lines and say, these belong a chunk. 00:43:27.775 --> 00:43:29.650 align:middle line:84% Now there are other lines they could draw on. 00:43:29.650 --> 00:43:32.530 align:middle line:84% And in fact, if you look closely at their original paper, 00:43:32.530 --> 00:43:35.260 align:middle line:84% there's large numbers of genes that they do not 00:43:35.260 --> 00:43:36.640 align:middle line:90% draw those lines next to. 00:43:36.640 --> 00:43:38.390 align:middle line:90% They don't have anything to say. 00:43:38.390 --> 00:43:41.210 align:middle line:90% So what's really going on here? 00:43:41.210 --> 00:43:43.930 align:middle line:84% So again, these clusters are defined 00:43:43.930 --> 00:43:47.390 align:middle line:90% by correlation coefficient. 00:43:47.390 --> 00:43:48.910 align:middle line:84% So what if we try to take advantage 00:43:48.910 --> 00:43:52.657 align:middle line:84% of the Markovian property of gene expression, 00:43:52.657 --> 00:43:53.740 align:middle line:90% like any physical process. 00:43:53.740 --> 00:43:56.710 align:middle line:90% Namely, that in a time series. 00:43:56.710 --> 00:43:59.710 align:middle line:90% 00:43:59.710 --> 00:44:03.860 align:middle line:84% The recent past informs you more than the distant past. 00:44:03.860 --> 00:44:09.070 align:middle line:84% So you can approximate the conditional probability 00:44:09.070 --> 00:44:12.890 align:middle line:84% of this value upon the recent past rather than the whole time 00:44:12.890 --> 00:44:13.390 align:middle line:90% series. 00:44:13.390 --> 00:44:15.130 align:middle line:84% That is, basically the assumption 00:44:15.130 --> 00:44:19.660 align:middle line:84% is that T0 is independent of the remote past, given 00:44:19.660 --> 00:44:20.470 align:middle line:90% the recent past. 00:44:20.470 --> 00:44:23.380 align:middle line:84% And how many steps back and look is what you well 00:44:23.380 --> 00:44:25.670 align:middle line:90% know as the mark of order. 00:44:25.670 --> 00:44:28.540 align:middle line:84% So mark of order is two, that means 00:44:28.540 --> 00:44:30.310 align:middle line:90% you look back two time slices. 00:44:30.310 --> 00:44:33.250 align:middle line:84% Markov order is one, you look back one type slices. 00:44:33.250 --> 00:44:36.700 align:middle line:84% If time ordering does not matter-- 00:44:36.700 --> 00:44:38.590 align:middle line:84% just like the correlation coefficient-- 00:44:38.590 --> 00:44:40.270 align:middle line:90% then your mark order is zero. 00:44:40.270 --> 00:44:42.820 align:middle line:84% The recent past actually doesn't inform you at all. 00:44:42.820 --> 00:44:46.310 align:middle line:90% 00:44:46.310 --> 00:44:50.870 align:middle line:84% Now let's look instead at clustering as this fiscal model 00:44:50.870 --> 00:44:52.680 align:middle line:90% selection problem. 00:44:52.680 --> 00:44:57.290 align:middle line:84% So rather than looking for highly correlated pairs, 00:44:57.290 --> 00:45:00.590 align:middle line:84% we'll consider each of the time series of his genes 00:45:00.590 --> 00:45:05.090 align:middle line:90% as being examples of a process. 00:45:05.090 --> 00:45:12.840 align:middle line:84% And the process is going to be represented therefore, 00:45:12.840 --> 00:45:17.630 align:middle line:84% by one or more genes, which are driving that process. 00:45:17.630 --> 00:45:24.650 align:middle line:84% And the question is, for any given pair of genes, 00:45:24.650 --> 00:45:26.660 align:middle line:84% do they belong to the same process? 00:45:26.660 --> 00:45:29.760 align:middle line:84% And essentially, with a very simple Markovian analysis, 00:45:29.760 --> 00:45:34.145 align:middle line:84% which is just a pain in the butt to implement, 00:45:34.145 --> 00:45:37.400 align:middle line:84% you apply Bayes' theorem and you say, 00:45:37.400 --> 00:45:45.290 align:middle line:84% is it more likely that these two gene signatures 00:45:45.290 --> 00:45:48.110 align:middle line:84% were generated by the same process 00:45:48.110 --> 00:45:50.060 align:middle line:90% or by a different process? 00:45:50.060 --> 00:45:52.635 align:middle line:84% If they are generated by the same process, 00:45:52.635 --> 00:45:55.010 align:middle line:84% if the model that they were generated by the same process 00:45:55.010 --> 00:45:57.140 align:middle line:84% has a higher probability than the one 00:45:57.140 --> 00:45:59.090 align:middle line:84% then they have a different process 00:45:59.090 --> 00:46:00.920 align:middle line:90% then you cluster them together. 00:46:00.920 --> 00:46:03.530 align:middle line:84% Just as you would in a regular dendrogram. 00:46:03.530 --> 00:46:05.278 align:middle line:84% Of course, the threshold you pick-- 00:46:05.278 --> 00:46:07.070 align:middle line:84% the probabilistic threshold that you pick-- 00:46:07.070 --> 00:46:09.950 align:middle line:84% is going to determine exactly which branches 00:46:09.950 --> 00:46:14.150 align:middle line:84% are brought together, which gene traces are brought together. 00:46:14.150 --> 00:46:18.320 align:middle line:84% But nonetheless, you now have a strong probabilistically based 00:46:18.320 --> 00:46:21.260 align:middle line:84% and Markovian based reason for bringing together two gene 00:46:21.260 --> 00:46:22.970 align:middle line:84% expression signatures, so then based 00:46:22.970 --> 00:46:26.420 align:middle line:84% on the marginal likelihood of these two gene expression 00:46:26.420 --> 00:46:29.760 align:middle line:84% patterns being generated by one process versus the other. 00:46:29.760 --> 00:46:37.160 align:middle line:84% And you can thereby eliminate them appropriately. 00:46:37.160 --> 00:46:39.170 align:middle line:84% Since I'm running short of time, let me just 00:46:39.170 --> 00:46:42.200 align:middle line:84% tell you that the patients have a number of nice tricks 00:46:42.200 --> 00:46:45.440 align:middle line:90% in their armamentarium. 00:46:45.440 --> 00:46:47.750 align:middle line:84% Which essentially, depend on the following observation 00:46:47.750 --> 00:46:54.140 align:middle line:84% that if we use the same data for the models, 00:46:54.140 --> 00:46:58.260 align:middle line:84% if we assume all models are initially equally likely, 00:46:58.260 --> 00:47:02.330 align:middle line:84% then you can do that very nice equivalents shown at the bottom 00:47:02.330 --> 00:47:05.150 align:middle line:84% such that, you only have to calculate 00:47:05.150 --> 00:47:07.730 align:middle line:84% the marginal likelihood rather than absolute probability 00:47:07.730 --> 00:47:09.890 align:middle line:84% in calculating which models are most likely. 00:47:09.890 --> 00:47:13.980 align:middle line:84% And therefore, simplifies a lot miscalculation. 00:47:13.980 --> 00:47:17.090 align:middle line:84% And incidentally, this program that we called CAGE 00:47:17.090 --> 00:47:20.040 align:middle line:84% is available publicly for-- if you look at time series. 00:47:20.040 --> 00:47:21.420 align:middle line:90% But what does it do? 00:47:21.420 --> 00:47:24.110 align:middle line:84% So if we took the same time series 00:47:24.110 --> 00:47:26.930 align:middle line:84% that you looked at with the expression 00:47:26.930 --> 00:47:32.960 align:middle line:84% data from the foreskin through this same Markovian clustering 00:47:32.960 --> 00:47:36.320 align:middle line:84% program, we only found four clusters. 00:47:36.320 --> 00:47:40.280 align:middle line:84% Two small clusters and two large clusters. 00:47:40.280 --> 00:47:43.640 align:middle line:90% So what's with that? 00:47:43.640 --> 00:47:47.330 align:middle line:90% The two small clusters-- 00:47:47.330 --> 00:47:52.250 align:middle line:84% the two small clusters were one, a cytokine cluster, 00:47:52.250 --> 00:47:55.220 align:middle line:90% and then an apoptose cluster. 00:47:55.220 --> 00:47:59.000 align:middle line:84% The large two clusters had a bunch of different genes 00:47:59.000 --> 00:48:00.540 align:middle line:90% in them. 00:48:00.540 --> 00:48:05.360 align:middle line:84% And they contained essentially, all the other clusters 00:48:05.360 --> 00:48:09.180 align:middle line:84% that had been obtained in the original analysis. 00:48:09.180 --> 00:48:12.050 align:middle line:84% So you could say at this point, well, Zack, you guys 00:48:12.050 --> 00:48:14.600 align:middle line:84% didn't do a very good job because despite all 00:48:14.600 --> 00:48:18.410 align:middle line:84% this convincing talk about the Markovian nature of gene 00:48:18.410 --> 00:48:20.990 align:middle line:84% expression, they were able to really pull apart 00:48:20.990 --> 00:48:24.690 align:middle line:84% a lot of processes and you were not. 00:48:24.690 --> 00:48:27.290 align:middle line:84% So how do we know that we did a better job or not? 00:48:27.290 --> 00:48:30.710 align:middle line:84% Well it turns out that back when they 00:48:30.710 --> 00:48:33.410 align:middle line:84% did their original experiment, Unigene, 00:48:33.410 --> 00:48:36.290 align:middle line:84% which is a database that Alberta will tell you about 00:48:36.290 --> 00:48:37.850 align:middle line:90% was an earlier stage. 00:48:37.850 --> 00:48:39.500 align:middle line:90% And what do I mean by that? 00:48:39.500 --> 00:48:43.160 align:middle line:84% Every time GenBank grows by a certain amount, by which I 00:48:43.160 --> 00:48:48.320 align:middle line:84% mean, more sequences are deposited into the human genome 00:48:48.320 --> 00:48:53.450 align:middle line:84% database, there's a periodic reassembly 00:48:53.450 --> 00:48:57.170 align:middle line:84% of the putative genes based on the better 00:48:57.170 --> 00:49:00.230 align:middle line:84% and better populated jigsaw puzzle of GenBank sequences. 00:49:00.230 --> 00:49:03.410 align:middle line:84% And sometimes it means that two things that 00:49:03.410 --> 00:49:04.910 align:middle line:84% were supposed to be different genes, 00:49:04.910 --> 00:49:06.980 align:middle line:84% when you have enough bridging sequences, 00:49:06.980 --> 00:49:08.330 align:middle line:90% ends up being the same gene. 00:49:08.330 --> 00:49:09.920 align:middle line:84% And a gene that was previously thought 00:49:09.920 --> 00:49:12.710 align:middle line:84% to be the same with enough distinguishing sequences, 00:49:12.710 --> 00:49:14.130 align:middle line:90% ends up being split into two. 00:49:14.130 --> 00:49:17.960 align:middle line:84% So these Unigene builds are different. 00:49:17.960 --> 00:49:22.580 align:middle line:84% And consequently, when we look back in 1999, 00:49:22.580 --> 00:49:26.630 align:middle line:84% 238 out of 517 genes were unknown. 00:49:26.630 --> 00:49:28.280 align:middle line:84% And we relabeled the genes according 00:49:28.280 --> 00:49:32.270 align:middle line:84% to the current state of the art at which point, only 20 00:49:32.270 --> 00:49:33.410 align:middle line:90% were left unknown. 00:49:33.410 --> 00:49:36.830 align:middle line:84% And there were 19 genes that were present twice 00:49:36.830 --> 00:49:38.450 align:middle line:90% in the data set. 00:49:38.450 --> 00:49:40.850 align:middle line:84% And the original clustering puts four 00:49:40.850 --> 00:49:44.060 align:middle line:84% of these in completely different clusters. 00:49:44.060 --> 00:49:47.750 align:middle line:90% Whereas, we only did this once. 00:49:47.750 --> 00:49:50.240 align:middle line:84% If we put our marker order back to zero, 00:49:50.240 --> 00:49:52.430 align:middle line:84% that means we ignore the past, we 00:49:52.430 --> 00:49:54.830 align:middle line:84% get exactly the same misclassification 00:49:54.830 --> 00:49:56.660 align:middle line:90% of these identical genes. 00:49:56.660 --> 00:50:00.140 align:middle line:84% So the conclusion here is that the temporal ordering 00:50:00.140 --> 00:50:01.040 align:middle line:90% does matter. 00:50:01.040 --> 00:50:06.500 align:middle line:84% And doing just a simplistic correlational analysis will not 00:50:06.500 --> 00:50:07.970 align:middle line:84% cut the mustard when you're really 00:50:07.970 --> 00:50:09.830 align:middle line:84% try to dissect some of the processes. 00:50:09.830 --> 00:50:14.090 align:middle line:84% And using a statistical measure of clumpiness 00:50:14.090 --> 00:50:18.260 align:middle line:84% of these clusters, rather than a looks right test, which 00:50:18.260 --> 00:50:23.300 align:middle line:84% had been the standard for most of these papers from 1999 00:50:23.300 --> 00:50:28.423 align:middle line:84% to 2003 probably is also not a safe thing to do. 00:50:28.423 --> 00:50:29.840 align:middle line:84% So it turns out, those two choices 00:50:29.840 --> 00:50:32.480 align:middle line:84% I showed you before were in fact, the same gene. 00:50:32.480 --> 00:50:35.520 align:middle line:84% Although, it was not known to be the case when it was first 00:50:35.520 --> 00:50:36.020 align:middle line:90% published. 00:50:36.020 --> 00:50:39.590 align:middle line:84% And these two traces were put in two different clusters. 00:50:39.590 --> 00:50:42.170 align:middle line:90% 00:50:42.170 --> 00:50:46.550 align:middle line:84% And likewise, these two traces also 00:50:46.550 --> 00:50:48.290 align:middle line:90% ended up being the same gene. 00:50:48.290 --> 00:50:51.350 align:middle line:84% And they too were put into two different clusters. 00:50:51.350 --> 00:50:56.220 align:middle line:90% 00:50:56.220 --> 00:50:56.730 align:middle line:90% All right. 00:50:56.730 --> 00:51:02.010 align:middle line:84% Let me wrap up by talking about another aspect 00:51:02.010 --> 00:51:03.870 align:middle line:90% of genomic medicine. 00:51:03.870 --> 00:51:05.640 align:middle line:84% Those of you who've done clinical research 00:51:05.640 --> 00:51:07.015 align:middle line:84% know the following is true, which 00:51:07.015 --> 00:51:12.810 align:middle line:84% is most clinical trials are censored. 00:51:12.810 --> 00:51:16.590 align:middle line:84% By which I mean, either some patients drop out of the trial 00:51:16.590 --> 00:51:18.450 align:middle line:90% before it's completed. 00:51:18.450 --> 00:51:23.730 align:middle line:84% Or the patient has died by in the trial, 00:51:23.730 --> 00:51:25.330 align:middle line:90% if mortality is your endpoint. 00:51:25.330 --> 00:51:27.240 align:middle line:84% So it's unclear whether, for instance, 00:51:27.240 --> 00:51:30.270 align:middle line:84% a patient who just died right after in the trial 00:51:30.270 --> 00:51:32.410 align:middle line:84% should have been included or not. 00:51:32.410 --> 00:51:36.510 align:middle line:84% And the big picture is that by having a particular cutoff 00:51:36.510 --> 00:51:38.970 align:middle line:84% point on the study, you're not allowing 00:51:38.970 --> 00:51:42.660 align:middle line:84% yourself to see the full evolution of the patient's 00:51:42.660 --> 00:51:43.200 align:middle line:90% history. 00:51:43.200 --> 00:51:46.290 align:middle line:84% And therefore, depending on what the nature of the study, 00:51:46.290 --> 00:51:48.240 align:middle line:84% you may actually have a very strong bias. 00:51:48.240 --> 00:51:50.040 align:middle line:84% And it's been shown in clinical research 00:51:50.040 --> 00:51:53.850 align:middle line:84% again and again that if this censoring is not attended to, 00:51:53.850 --> 00:51:59.740 align:middle line:84% you'll have a misinterpretation of the clinical trials. 00:51:59.740 --> 00:52:03.660 align:middle line:84% So that's well known in the clinical research arena. 00:52:03.660 --> 00:52:08.700 align:middle line:84% And the question is, is it equally true in genomics 00:52:08.700 --> 00:52:12.700 align:middle line:90% as applied to clinical research? 00:52:12.700 --> 00:52:17.070 align:middle line:84% Well, let's understand what most of the studies published 00:52:17.070 --> 00:52:19.290 align:middle line:90% actually do. 00:52:19.290 --> 00:52:21.940 align:middle line:90% They often do the following. 00:52:21.940 --> 00:52:25.620 align:middle line:84% They take a set of gene expression patterns, 00:52:25.620 --> 00:52:30.300 align:middle line:84% and do some clustering operation such as what [? Eliza ?] did 00:52:30.300 --> 00:52:33.990 align:middle line:90% originally in their paper. 00:52:33.990 --> 00:52:38.140 align:middle line:84% And they say, oh look, there's two or more 00:52:38.140 --> 00:52:41.635 align:middle line:84% in actual clusters just based on the gene expression profile. 00:52:41.635 --> 00:52:43.260 align:middle line:84% And then they ask themselves after they 00:52:43.260 --> 00:52:46.800 align:middle line:84% found these expression profiles, is there 00:52:46.800 --> 00:52:50.850 align:middle line:84% anything different about these patient populations? 00:52:50.850 --> 00:52:52.570 align:middle line:90% And they rummage around. 00:52:52.570 --> 00:52:57.210 align:middle line:84% And in this case, they found a wonderfully different mortality 00:52:57.210 --> 00:52:58.770 align:middle line:84% between the two groups of patients 00:52:58.770 --> 00:53:01.840 align:middle line:84% as defined by the expression pattern. 00:53:01.840 --> 00:53:02.940 align:middle line:90% So let's review that. 00:53:02.940 --> 00:53:05.580 align:middle line:84% You first look for a difference in expression pattern. 00:53:05.580 --> 00:53:08.430 align:middle line:84% And then you say, what makes these patients different? 00:53:08.430 --> 00:53:11.490 align:middle line:84% And it makes for a very impressive publication 00:53:11.490 --> 00:53:15.355 align:middle line:84% because you're not looking at the patient, just saying, 00:53:15.355 --> 00:53:16.980 align:middle line:84% I'm looking at gene expression pattern. 00:53:16.980 --> 00:53:19.620 align:middle line:84% And then when I open my eyes and allow 00:53:19.620 --> 00:53:22.290 align:middle line:84% myself to see what's different in a patient in these two 00:53:22.290 --> 00:53:23.940 align:middle line:84% groups, I'm finding something that's 00:53:23.940 --> 00:53:26.070 align:middle line:90% really clinically different. 00:53:26.070 --> 00:53:29.350 align:middle line:84% Now what's the problem with that? 00:53:29.350 --> 00:53:32.910 align:middle line:84% The problem with that is that you 00:53:32.910 --> 00:53:37.740 align:middle line:84% may be looking at the wrong thing for an example. 00:53:37.740 --> 00:53:43.200 align:middle line:84% With 10,000 genes to look at, there 00:53:43.200 --> 00:53:45.757 align:middle line:84% might be a cluster of genes that for instance-- 00:53:45.757 --> 00:53:47.340 align:middle line:84% let me just actually, before I come up 00:53:47.340 --> 00:53:49.750 align:middle line:90% with a fanciful example-- 00:53:49.750 --> 00:53:52.620 align:middle line:84% it might be that the reason that these patients have 00:53:52.620 --> 00:53:55.800 align:middle line:84% different expression profiles is because they 00:53:55.800 --> 00:53:56.940 align:middle line:90% weigh a different amount. 00:53:56.940 --> 00:54:00.570 align:middle line:84% One group is fatter than the other 00:54:00.570 --> 00:54:02.790 align:middle line:84% because one group is sicker than the other, 00:54:02.790 --> 00:54:04.980 align:middle line:84% except it was not picked up clinically. 00:54:04.980 --> 00:54:08.045 align:middle line:84% But one is going down the tubes and one is not. 00:54:08.045 --> 00:54:09.420 align:middle line:84% And what we're really seeing here 00:54:09.420 --> 00:54:11.700 align:middle line:84% is not something intrinsic about the tumor, 00:54:11.700 --> 00:54:13.860 align:middle line:84% but something intrinsic about the weight. 00:54:13.860 --> 00:54:19.470 align:middle line:84% And so getting at is a very indirect measure of mortality. 00:54:19.470 --> 00:54:21.540 align:middle line:84% And it may be very much unrelated. 00:54:21.540 --> 00:54:24.480 align:middle line:84% Can we do better by going directly 00:54:24.480 --> 00:54:28.410 align:middle line:84% from the gene expression pattern to the clinical phenotype 00:54:28.410 --> 00:54:31.530 align:middle line:84% of interest rather than going through this two step process 00:54:31.530 --> 00:54:35.280 align:middle line:84% where we first cluster the genes without any knowledge 00:54:35.280 --> 00:54:38.040 align:middle line:84% of the phenotype and then see how the phenotype is different. 00:54:38.040 --> 00:54:41.250 align:middle line:90% 00:54:41.250 --> 00:54:43.620 align:middle line:84% Specifically, can we directly find 00:54:43.620 --> 00:54:45.810 align:middle line:84% genes or linear combination genes 00:54:45.810 --> 00:54:48.420 align:middle line:84% that are highly correlated with survival times? 00:54:48.420 --> 00:54:50.610 align:middle line:84% For example, gene A plus 0.5 times 00:54:50.610 --> 00:54:55.200 align:middle line:84% gene B, plus 2 times gene C equals some probability 00:54:55.200 --> 00:54:55.950 align:middle line:90% of survival. 00:54:55.950 --> 00:54:58.890 align:middle line:90% Can we do that? 00:54:58.890 --> 00:55:05.370 align:middle line:84% And can we do that in the context of the kind 00:55:05.370 --> 00:55:08.260 align:middle line:90% of censoring that I described? 00:55:08.260 --> 00:55:11.670 align:middle line:84% So in summary, how can we use the survival times directly 00:55:11.670 --> 00:55:13.650 align:middle line:90% to find good predictors? 00:55:13.650 --> 00:55:17.030 align:middle line:90% 00:55:17.030 --> 00:55:19.330 align:middle line:84% So here's the fundamental problem. 00:55:19.330 --> 00:55:20.740 align:middle line:90% We have gene expression data. 00:55:20.740 --> 00:55:22.150 align:middle line:90% We have phenotypic data. 00:55:22.150 --> 00:55:26.330 align:middle line:84% And we want to find out how one predicts the other. 00:55:26.330 --> 00:55:32.660 align:middle line:84% Now previously, if I asked you that question, 00:55:32.660 --> 00:55:35.960 align:middle line:84% you'd say oh Zack, let's just use 00:55:35.960 --> 00:55:39.110 align:middle line:84% linear regression, logistic regression, for instance. 00:55:39.110 --> 00:55:42.590 align:middle line:84% It turns out with 20,000 variables, it breaks down. 00:55:42.590 --> 00:55:44.610 align:middle line:90% It just does not work. 00:55:44.610 --> 00:55:46.460 align:middle line:84% And so a nice answer to something 00:55:46.460 --> 00:55:48.440 align:middle line:90% called partial least squares. 00:55:48.440 --> 00:55:50.845 align:middle line:84% And what is partial least squares? 00:55:50.845 --> 00:55:52.970 align:middle line:84% Well, let me give you the intuitive feeling for it. 00:55:52.970 --> 00:55:55.940 align:middle line:84% I explain to you what principal components were. 00:55:55.940 --> 00:55:59.150 align:middle line:84% Principal components, being these components 00:55:59.150 --> 00:56:03.728 align:middle line:84% that capture a large component of the variance. 00:56:03.728 --> 00:56:06.020 align:middle line:84% And subsequent components are orthogonal to each other. 00:56:06.020 --> 00:56:07.700 align:middle line:84% This is like principal components, 00:56:07.700 --> 00:56:13.170 align:middle line:84% but as they relate to a specific outcome variable like survival. 00:56:13.170 --> 00:56:15.050 align:middle line:84% So these are principal components 00:56:15.050 --> 00:56:20.910 align:middle line:84% that are correlated essentially, with an outcome variable. 00:56:20.910 --> 00:56:22.880 align:middle line:90% So that's interesting. 00:56:22.880 --> 00:56:26.480 align:middle line:84% But the problem with perfect least squares 00:56:26.480 --> 00:56:31.670 align:middle line:84% is that it does not actually allow you to use sensor data. 00:56:31.670 --> 00:56:34.100 align:middle line:84% it assumes that the data is complete throughout. 00:56:34.100 --> 00:56:37.920 align:middle line:84% And that's going to immediately bias your data. 00:56:37.920 --> 00:56:40.760 align:middle line:84% So how do we take advantage of censoring? 00:56:40.760 --> 00:56:43.460 align:middle line:84% Now in classical clinical research 00:56:43.460 --> 00:56:46.460 align:middle line:84% in medicine, what you use is something called the Cox 00:56:46.460 --> 00:56:50.090 align:middle line:84% model, which essentially is a probabilistic model that says, 00:56:50.090 --> 00:56:55.790 align:middle line:84% what is the hazard of dying for this population of patients? 00:56:55.790 --> 00:57:02.330 align:middle line:84% And basically, imputes for the missing data fundamentally. 00:57:02.330 --> 00:57:05.600 align:middle line:84% And for a small number of variables or genes, 00:57:05.600 --> 00:57:08.600 align:middle line:84% you actually can use a Cox model to actually figure out, 00:57:08.600 --> 00:57:11.330 align:middle line:84% essentially, what would have happened 00:57:11.330 --> 00:57:13.490 align:middle line:90% if you'd had a complete data. 00:57:13.490 --> 00:57:15.410 align:middle line:84% But the Cox model, just like regression, 00:57:15.410 --> 00:57:20.775 align:middle line:84% does not work well for tens of thousands of variables. 00:57:20.775 --> 00:57:21.900 align:middle line:90% So what are we going to do? 00:57:21.900 --> 00:57:24.920 align:middle line:84% How do we get to use partial least squares? 00:57:24.920 --> 00:57:27.290 align:middle line:84% To make a long story short, I guess 00:57:27.290 --> 00:57:31.970 align:middle line:84% I'll post a paper to myCourses website. 00:57:31.970 --> 00:57:37.910 align:middle line:84% What you do is actually model the patient's data 00:57:37.910 --> 00:57:40.020 align:middle line:90% using a Poisson regression. 00:57:40.020 --> 00:57:44.720 align:middle line:84% So that you actually transform this data 00:57:44.720 --> 00:57:50.120 align:middle line:84% set from one with missing data, to one with complete data based 00:57:50.120 --> 00:57:51.680 align:middle line:90% on this Poisson regression. 00:57:51.680 --> 00:57:55.850 align:middle line:84% But the problem is now we can create essentially many more 00:57:55.850 --> 00:57:56.908 align:middle line:90% data points, essentially. 00:57:56.908 --> 00:57:58.700 align:middle line:84% And just to give you an intuition behind it 00:57:58.700 --> 00:58:00.650 align:middle line:84% because I'm going to run out of time to go 00:58:00.650 --> 00:58:03.710 align:middle line:84% into a full description is, this will allow us 00:58:03.710 --> 00:58:06.480 align:middle line:90% to estimate the fine problems. 00:58:06.480 --> 00:58:09.440 align:middle line:84% So let's say that a patient lasted until this point 00:58:09.440 --> 00:58:12.740 align:middle line:84% and dropped out of the study, where might they have died? 00:58:12.740 --> 00:58:16.520 align:middle line:84% What can we assume, based on this prior data? 00:58:16.520 --> 00:58:20.600 align:middle line:84% So by using a plus model, we can enter all these new data 00:58:20.600 --> 00:58:21.300 align:middle line:90% points. 00:58:21.300 --> 00:58:25.760 align:middle line:84% And so we can actually figure out precisely, 00:58:25.760 --> 00:58:28.130 align:middle line:84% or probabilistic-ally I should say, 00:58:28.130 --> 00:58:31.910 align:middle line:84% what the values were the various time points and completely 00:58:31.910 --> 00:58:34.370 align:middle line:84% in this probabilistic sense, so now we 00:58:34.370 --> 00:58:36.420 align:middle line:90% can apply partial least squares. 00:58:36.420 --> 00:58:39.350 align:middle line:84% So what does that mean, pragmatically? 00:58:39.350 --> 00:58:44.490 align:middle line:84% So Bhattacharjee, how do you pronounce that? 00:58:44.490 --> 00:58:45.740 align:middle line:90% AUDIENCE: Bhattacharjee, yeah. 00:58:45.740 --> 00:58:49.280 align:middle line:84% ISAAC SAMUEL KOHANE: Bhattacharjee, and basically 00:58:49.280 --> 00:58:52.890 align:middle line:84% working with Todd Golub, did a study of lung cancer, 00:58:52.890 --> 00:58:55.790 align:middle line:90% which had in fact censored data. 00:58:55.790 --> 00:58:59.150 align:middle line:84% And each patient had a survival time 00:58:59.150 --> 00:59:02.600 align:middle line:84% and it was marked whether they were censored or not. 00:59:02.600 --> 00:59:08.330 align:middle line:84% And the question was, could they identify different outcomes? 00:59:08.330 --> 00:59:11.090 align:middle line:84% And the short answer is, they were able to identify, 00:59:11.090 --> 00:59:13.490 align:middle line:84% for instance, metastatic versus non metastatic, 00:59:13.490 --> 00:59:15.770 align:middle line:84% but there are a lot of other outcomes 00:59:15.770 --> 00:59:18.740 align:middle line:84% that they could not distinguish from one another 00:59:18.740 --> 00:59:21.222 align:middle line:90% using these 125 samples. 00:59:21.222 --> 00:59:22.680 align:middle line:84% So we asked ourselves the question. 00:59:22.680 --> 00:59:25.370 align:middle line:84% If we took into account these partial least squares 00:59:25.370 --> 00:59:27.590 align:middle line:84% on top of this Poisson model, could 00:59:27.590 --> 00:59:30.770 align:middle line:84% we actually tease apart in a reliable fashion 00:59:30.770 --> 00:59:35.060 align:middle line:90% some new phenotypes. 00:59:35.060 --> 00:59:40.220 align:middle line:84% And the short answer is these are 00:59:40.220 --> 00:59:41.870 align:middle line:90% uncorrected P values, whoops. 00:59:41.870 --> 00:59:45.520 align:middle line:90% 00:59:45.520 --> 00:59:49.380 align:middle line:84% These are uncorrected P values, but tenth of -7, 00:59:49.380 --> 00:59:52.570 align:middle line:84% if you do correction for multiple hypothesis testing, 00:59:52.570 --> 00:59:57.500 align:middle line:84% can see that they're actually going to be highly significant. 00:59:57.500 --> 01:00:02.480 align:middle line:84% We were able to obtain components 01:00:02.480 --> 01:00:06.950 align:middle line:84% that were highly predictive from the partial squares analysis. 01:00:06.950 --> 01:00:12.110 align:middle line:84% Highly predictive of clinical outcome. 01:00:12.110 --> 01:00:14.540 align:middle line:90% Where previously, you could not. 01:00:14.540 --> 01:00:19.160 align:middle line:84% And to make that very clear, these 01:00:19.160 --> 01:00:22.760 align:middle line:84% were two groups of patients that previously were not 01:00:22.760 --> 01:00:23.720 align:middle line:90% distinguishable. 01:00:23.720 --> 01:00:27.650 align:middle line:84% But by using this method, we were able to distinguish them 01:00:27.650 --> 01:00:30.680 align:middle line:90% with a P value of this amount. 01:00:30.680 --> 01:00:32.570 align:middle line:84% And provide a very sharp distinction 01:00:32.570 --> 01:00:34.820 align:middle line:84% between these two groups of patients that was actually 01:00:34.820 --> 01:00:38.233 align:middle line:84% not resolvable using the standard analysis that 01:00:38.233 --> 01:00:38.900 align:middle line:90% done previously. 01:00:38.900 --> 01:00:41.360 align:middle line:84% And the reason is because I just want 01:00:41.360 --> 01:00:43.970 align:middle line:84% to bring this home to you, because we 01:00:43.970 --> 01:00:46.850 align:middle line:84% were able to go directly from gene expression 01:00:46.850 --> 01:00:50.090 align:middle line:90% to the outcome of interest. 01:00:50.090 --> 01:00:52.700 align:middle line:84% We didn't ask ourselves, let's cluster the patient data 01:00:52.700 --> 01:00:55.760 align:middle line:84% and then see how the patients are different after the fact. 01:00:55.760 --> 01:00:58.880 align:middle line:84% Who said, I know the different characters 01:00:58.880 --> 01:01:01.370 align:middle line:84% of the patients in this case mortality. 01:01:01.370 --> 01:01:04.410 align:middle line:84% What combination of genes best predicts that difference? 01:01:04.410 --> 01:01:06.763 align:middle line:84% And by looking at that direct signal, rather than 01:01:06.763 --> 01:01:08.930 align:middle line:84% indirect signal with all the complications of having 01:01:08.930 --> 01:01:11.030 align:middle line:84% to deal with sensored data, we're 01:01:11.030 --> 01:01:13.680 align:middle line:84% able to separate out very, very well 01:01:13.680 --> 01:01:14.880 align:middle line:90% these subgroups of patients. 01:01:14.880 --> 01:01:17.750 align:middle line:84% So, again, showing you how by using a little bit 01:01:17.750 --> 01:01:19.310 align:middle line:84% more sophisticated computation, we're 01:01:19.310 --> 01:01:22.535 align:middle line:84% able to identify subgroups of such patients. 01:01:22.535 --> 01:01:25.860 align:middle line:90% 01:01:25.860 --> 01:01:27.650 align:middle line:84% So this brings me to the problem and I 01:01:27.650 --> 01:01:29.067 align:middle line:84% want to spend a little bit of time 01:01:29.067 --> 01:01:31.200 align:middle line:84% to make sure you all understand the problem set. 01:01:31.200 --> 01:01:34.400 align:middle line:84% So I think it's very important for you 01:01:34.400 --> 01:01:38.910 align:middle line:84% to get your hands dirty with the data. 01:01:38.910 --> 01:01:43.940 align:middle line:84% So there is a data set that's available at the following URL. 01:01:43.940 --> 01:01:50.380 align:middle line:84% And if you just Google Kunkel, Kohane, Haslett, 01:01:50.380 --> 01:01:51.380 align:middle line:90% you'll probably find it. 01:01:51.380 --> 01:01:53.300 align:middle line:90% But that's the full URL. 01:01:53.300 --> 01:01:58.580 align:middle line:84% And I'll send out to the group the URL 01:01:58.580 --> 01:02:00.200 align:middle line:90% through our group email. 01:02:00.200 --> 01:02:03.350 align:middle line:84% And that's a data set from a paper that we published, 01:02:03.350 --> 01:02:06.680 align:middle line:84% I think, a year or two ago about Duchenne muscular dystrophy. 01:02:06.680 --> 01:02:11.090 align:middle line:84% Duchenne muscular dystrophy is a degenerative disease 01:02:11.090 --> 01:02:13.950 align:middle line:90% of the muscles. 01:02:13.950 --> 01:02:18.020 align:middle line:84% It's the major cause of what Jerry Lewis raises money for. 01:02:18.020 --> 01:02:21.470 align:middle line:84% And we have data on the muscles of patients 01:02:21.470 --> 01:02:26.360 align:middle line:84% who have this disease and people who don't have that disease. 01:02:26.360 --> 01:02:32.180 align:middle line:84% And what I want you to look at is that comparison. 01:02:32.180 --> 01:02:35.270 align:middle line:84% And I want you to identify those genes that are differentially 01:02:35.270 --> 01:02:36.140 align:middle line:90% expressed. 01:02:36.140 --> 01:02:38.190 align:middle line:84% Now, if you want to work in pairs, that's fine. 01:02:38.190 --> 01:02:40.370 align:middle line:90% But no more than pairs of you. 01:02:40.370 --> 01:02:43.350 align:middle line:84% And I'd like to use at least two different tests. 01:02:43.350 --> 01:02:47.000 align:middle line:84% One is SAM, Significance and Analysis of Microarrays. 01:02:47.000 --> 01:02:51.050 align:middle line:84% The other is the t-test, which I hope most of you know. 01:02:51.050 --> 01:02:53.420 align:middle line:84% There's many, many toolkits that you could 01:02:53.420 --> 01:02:55.310 align:middle line:90% use to do this comparison. 01:02:55.310 --> 01:03:00.320 align:middle line:84% Let me recommend to you the MEV, the Multiple Experiment Viewer 01:03:00.320 --> 01:03:01.720 align:middle line:90% at Tiger. 01:03:01.720 --> 01:03:04.520 align:middle line:84% Tiger is the Institute for genomic research 01:03:04.520 --> 01:03:05.660 align:middle line:90% and at Tiger.org. 01:03:05.660 --> 01:03:09.530 align:middle line:84% So if you Google MEV at Tiger, it does many more things 01:03:09.530 --> 01:03:11.510 align:middle line:90% than just these two tests. 01:03:11.510 --> 01:03:16.590 align:middle line:84% But I recommend it to you for these doing these two tests. 01:03:16.590 --> 01:03:19.070 align:middle line:84% So given these data, I want you to answer 01:03:19.070 --> 01:03:20.870 align:middle line:90% the following question. 01:03:20.870 --> 01:03:22.910 align:middle line:84% What is the difference between the top ranked 01:03:22.910 --> 01:03:25.625 align:middle line:84% 50 genes differentially expressed, 01:03:25.625 --> 01:03:27.125 align:middle line:84% both up and down by the two methods? 01:03:27.125 --> 01:03:29.750 align:middle line:90% 01:03:29.750 --> 01:03:35.810 align:middle line:84% And why are the lists of genes different? 01:03:35.810 --> 01:03:39.830 align:middle line:84% I'd like to use another program called Map Finder to classify 01:03:39.830 --> 01:03:42.750 align:middle line:90% these genes by their function. 01:03:42.750 --> 01:03:46.070 align:middle line:84% And so the output I'm expecting from you 01:03:46.070 --> 01:03:49.340 align:middle line:84% is two sets of differentially expressed genes 01:03:49.340 --> 01:03:52.250 align:middle line:84% based on these two different methodologies, your explanation 01:03:52.250 --> 01:03:56.240 align:middle line:84% why these sets of genes are different 01:03:56.240 --> 01:03:59.120 align:middle line:84% and the classification of these two sets of genes 01:03:59.120 --> 01:04:02.510 align:middle line:90% by functional annotation. 01:04:02.510 --> 01:04:07.443 align:middle line:90% 01:04:07.443 --> 01:04:08.860 align:middle line:84% It's a very simple exercise, but I 01:04:08.860 --> 01:04:13.552 align:middle line:84% think it's important because if I've learned anything 01:04:13.552 --> 01:04:15.010 align:middle line:84% about this area, it's just doing it 01:04:15.010 --> 01:04:19.700 align:middle line:84% by yourself a few times opens your eyes up 01:04:19.700 --> 01:04:21.700 align:middle line:84% to how bad the data is and what the problems are 01:04:21.700 --> 01:04:23.900 align:middle line:90% in analyzing it. 01:04:23.900 --> 01:04:26.307 align:middle line:84% And if you're not comfortable in downloading tools, then 01:04:26.307 --> 01:04:28.015 align:middle line:84% partner with someone in the group who is. 01:04:28.015 --> 01:04:33.200 align:middle line:90% 01:04:33.200 --> 01:04:33.740 align:middle line:90% All right. 01:04:33.740 --> 01:04:36.077 align:middle line:84% AUDIENCE: Your course is on the website? 01:04:36.077 --> 01:04:38.660 align:middle line:84% ISAAC SAMUEL KOHANE: [INAUDIBLE] will post this on the website 01:04:38.660 --> 01:04:45.470 align:middle line:84% and I will send an email out of this URL. 01:04:45.470 --> 01:04:47.574 align:middle line:90% Any other questions? 01:04:47.574 --> 01:04:48.927 align:middle line:90% AUDIENCE: When's the due date? 01:04:48.927 --> 01:04:51.010 align:middle line:84% ISAAC SAMUEL KOHANE: That's an excellent question. 01:04:51.010 --> 01:04:55.150 align:middle line:90% It's now February 24. 01:04:55.150 --> 01:05:00.595 align:middle line:90% How about-- when is March 15? 01:05:00.595 --> 01:05:02.470 align:middle line:84% Can you look up what day of the week that is? 01:05:02.470 --> 01:05:13.670 align:middle line:90% 01:05:13.670 --> 01:05:14.750 align:middle line:90% AUDIENCE: It's a Monday. 01:05:14.750 --> 01:05:16.730 align:middle line:84% ISAAC SAMUEL KOHANE: So March 16. 01:05:16.730 --> 01:05:20.720 align:middle line:84% And please let me if you're having problems. 01:05:20.720 --> 01:05:24.530 align:middle line:84% The guy who actually built MEV is a friend of mine at Tiger, 01:05:24.530 --> 01:05:26.750 align:middle line:84% so I can kick his butt if it doesn't work 01:05:26.750 --> 01:05:28.070 align:middle line:90% for you at a particular time. 01:05:28.070 --> 01:05:31.820 align:middle line:90% 01:05:31.820 --> 01:05:33.430 align:middle line:90% All right.