WEBVTT 00:00:00.000 --> 00:00:00.282 align:middle line:90% 00:00:00.282 --> 00:00:01.740 align:middle line:84% ISAAC SAMUEL KOHANE: Also, I forgot 00:00:01.740 --> 00:00:05.400 align:middle line:84% to mention at this point, the output of microarray studies 00:00:05.400 --> 00:00:07.800 align:middle line:84% is foreign to basic biology researchers. 00:00:07.800 --> 00:00:11.700 align:middle line:84% They're used to looking at three or four or five or 20 numbers 00:00:11.700 --> 00:00:16.200 align:middle line:84% and performing some easy analysis in an Excel 00:00:16.200 --> 00:00:17.440 align:middle line:90% spreadsheet. 00:00:17.440 --> 00:00:20.280 align:middle line:84% But the point-- or doing a BLAST of one gene at a time. 00:00:20.280 --> 00:00:21.990 align:middle line:84% But the data from these microarrays 00:00:21.990 --> 00:00:24.360 align:middle line:84% does not easily load into standard tools. 00:00:24.360 --> 00:00:26.310 align:middle line:90% The analysis is non-standard. 00:00:26.310 --> 00:00:27.510 align:middle line:90% Excel does not cut it. 00:00:27.510 --> 00:00:29.850 align:middle line:84% Excel can barely load some of these data sets. 00:00:29.850 --> 00:00:33.690 align:middle line:84% And so all this is actually representing 00:00:33.690 --> 00:00:34.822 align:middle line:90% an interesting turn. 00:00:34.822 --> 00:00:36.280 align:middle line:84% So this is the other point of view, 00:00:36.280 --> 00:00:40.020 align:middle line:84% which is now that the standard biologist tools are 00:00:40.020 --> 00:00:44.130 align:middle line:84% no longer fitting nicely onto the desktop productivity suite 00:00:44.130 --> 00:00:45.900 align:middle line:90% that they previously had. 00:00:45.900 --> 00:00:48.253 align:middle line:84% And this is, in some way, a full circle. 00:00:48.253 --> 00:00:49.920 align:middle line:84% It used to be, of course, that computers 00:00:49.920 --> 00:00:52.740 align:middle line:84% were run by the high priests in the computer center. 00:00:52.740 --> 00:00:54.562 align:middle line:90% And not only that were they-- 00:00:54.562 --> 00:00:56.520 align:middle line:84% did they know about how to boot-up the computer 00:00:56.520 --> 00:00:58.353 align:middle line:84% and to make the computer do whatever it did, 00:00:58.353 --> 00:01:00.390 align:middle line:84% but they were also pretty steeped in the math 00:01:00.390 --> 00:01:03.370 align:middle line:84% and in the computational sciences, 00:01:03.370 --> 00:01:05.489 align:middle line:84% so that they could actually interpret 00:01:05.489 --> 00:01:07.950 align:middle line:90% the results of a study. 00:01:07.950 --> 00:01:09.900 align:middle line:84% With the advent of the personal computer 00:01:09.900 --> 00:01:13.080 align:middle line:84% and incredibly useful tools, like Excel, 00:01:13.080 --> 00:01:16.650 align:middle line:84% there was a lot that was now democratized and brought back 00:01:16.650 --> 00:01:21.540 align:middle line:84% to the desk of the individual biology researcher. 00:01:21.540 --> 00:01:25.260 align:middle line:84% But then when we scaled back up to large data sets that are not 00:01:25.260 --> 00:01:30.420 align:middle line:84% trivially analyzable with current desktop tools, 00:01:30.420 --> 00:01:33.720 align:middle line:84% then the biologist's education is-- 00:01:33.720 --> 00:01:36.330 align:middle line:84% then the biologist's computational education 00:01:36.330 --> 00:01:37.290 align:middle line:90% is lacking. 00:01:37.290 --> 00:01:40.470 align:middle line:84% The biologists desktop tools are lacking. 00:01:40.470 --> 00:01:46.140 align:middle line:84% And again, they are not in the loop. 00:01:46.140 --> 00:01:47.942 align:middle line:84% And let me tell you this-- and I'll 00:01:47.942 --> 00:01:50.400 align:middle line:84% get to more interesting details in a second-- but something 00:01:50.400 --> 00:01:51.567 align:middle line:90% that you have to understand. 00:01:51.567 --> 00:01:53.640 align:middle line:84% This is creating a lot of resentment 00:01:53.640 --> 00:01:57.240 align:middle line:84% among standard biologists around this whole genomic revolution. 00:01:57.240 --> 00:01:59.760 align:middle line:84% Because not only-- I mean, what you'll hear mostly 00:01:59.760 --> 00:02:05.040 align:middle line:84% is voice of skepticism about the methodologies. 00:02:05.040 --> 00:02:08.850 align:middle line:84% But what's really going on is that they're justifiably 00:02:08.850 --> 00:02:11.560 align:middle line:84% feeling quite threatened for a big part of what 00:02:11.560 --> 00:02:12.810 align:middle line:90% they've been doing previously. 00:02:12.810 --> 00:02:15.750 align:middle line:84% Because they're just going to be disintermediated 00:02:15.750 --> 00:02:20.250 align:middle line:84% by the commoditization and the quantification 00:02:20.250 --> 00:02:24.300 align:middle line:84% and computational expertise that is required. 00:02:24.300 --> 00:02:26.820 align:middle line:84% So what are the characteristics of a microarray? 00:02:26.820 --> 00:02:29.070 align:middle line:84% How do you recognize a microarray 00:02:29.070 --> 00:02:33.390 align:middle line:90% when you are being marketed one? 00:02:33.390 --> 00:02:36.640 align:middle line:84% The first is that it should have a small form factor. 00:02:36.640 --> 00:02:38.070 align:middle line:90% So one part of it's -- 00:02:38.070 --> 00:02:41.880 align:middle line:84% capable of measuring a significant fraction of some 00:02:41.880 --> 00:02:47.130 align:middle line:84% "-ome," the genome, the transcriptome, or the proteome. 00:02:47.130 --> 00:02:50.500 align:middle line:84% So I think small form factor is an obvious one. 00:02:50.500 --> 00:02:54.780 align:middle line:84% But it also needs to measure a large part of the total output 00:02:54.780 --> 00:02:56.140 align:middle line:90% of the system. 00:02:56.140 --> 00:02:58.920 align:middle line:84% Otherwise, it's not a microarray in the sense 00:02:58.920 --> 00:03:00.930 align:middle line:90% that we commonly understand it. 00:03:00.930 --> 00:03:04.240 align:middle line:84% There has to be minimal labor in data acquisition, 00:03:04.240 --> 00:03:06.480 align:middle line:84% so that it's an industrial process and not 00:03:06.480 --> 00:03:07.710 align:middle line:90% a labor-intensive process. 00:03:07.710 --> 00:03:10.520 align:middle line:90% 00:03:10.520 --> 00:03:12.200 align:middle line:84% You need to have an automated data path 00:03:12.200 --> 00:03:16.190 align:middle line:90% to a digital electronic format. 00:03:16.190 --> 00:03:21.470 align:middle line:84% And it has to be sustainable, high throughput processing. 00:03:21.470 --> 00:03:24.740 align:middle line:84% And that begs the question then, is this a microarray? 00:03:24.740 --> 00:03:27.740 align:middle line:84% Well, this is a very nice piece of technology from Sciomics, 00:03:27.740 --> 00:03:31.970 align:middle line:84% which has, on these pillars, these antibodies 00:03:31.970 --> 00:03:35.630 align:middle line:84% that allows you to measure, in a very consistent and systematic 00:03:35.630 --> 00:03:39.320 align:middle line:84% fashion, 200 protein measurements or typically 00:03:39.320 --> 00:03:42.470 align:middle line:84% 30 proteins measured in sextuplets, 00:03:42.470 --> 00:03:45.690 align:middle line:90% in sextuple for redundancy. 00:03:45.690 --> 00:03:48.560 align:middle line:90% So it's kind of a microarray. 00:03:48.560 --> 00:03:50.040 align:middle line:90% It's small. 00:03:50.040 --> 00:03:51.650 align:middle line:90% It's an array. 00:03:51.650 --> 00:03:57.110 align:middle line:84% But as typically used, it's measuring between 30 and 200 00:03:57.110 --> 00:03:58.880 align:middle line:84% proteins, and so, therefore, I would 00:03:58.880 --> 00:04:01.190 align:middle line:90% argue it's not a microarray. 00:04:01.190 --> 00:04:04.640 align:middle line:84% And basically, they're trying to sell their venture capitalists 00:04:04.640 --> 00:04:08.600 align:middle line:84% that this going to be as revolutionary as the standard 00:04:08.600 --> 00:04:10.593 align:middle line:84% Affymetrix RNA expression microarrays 00:04:10.593 --> 00:04:11.510 align:middle line:90% that I told you about. 00:04:11.510 --> 00:04:12.385 align:middle line:90% But it doesn't scale. 00:04:12.385 --> 00:04:14.135 align:middle line:84% It doesn't give you the comprehensiveness. 00:04:14.135 --> 00:04:15.860 align:middle line:84% And, therefore, you know in advance 00:04:15.860 --> 00:04:21.890 align:middle line:84% that this is only a somewhat more efficient version 00:04:21.890 --> 00:04:25.610 align:middle line:84% of the ELISA test, the antibody-based test. 00:04:25.610 --> 00:04:26.870 align:middle line:90% It's not leveraged. 00:04:26.870 --> 00:04:32.060 align:middle line:84% Similarly, when you're looking at two-dimensional gels, 00:04:32.060 --> 00:04:34.790 align:middle line:84% there you are capturing a large percent of the proteome. 00:04:34.790 --> 00:04:39.080 align:middle line:84% But it's neither small nor is there an automatic data 00:04:39.080 --> 00:04:39.830 align:middle line:90% path to analysis. 00:04:39.830 --> 00:04:42.590 align:middle line:84% Because you basically have to pick out these different blots 00:04:42.590 --> 00:04:46.040 align:middle line:84% and apply some intelligence about which ones 00:04:46.040 --> 00:04:48.240 align:middle line:84% you pick, and then feed them into the mass spec, 00:04:48.240 --> 00:04:49.490 align:middle line:90% and then do the deconvolution. 00:04:49.490 --> 00:04:54.620 align:middle line:84% It's actually not a high throughput data path. 00:04:54.620 --> 00:04:56.600 align:middle line:84% And when you speak, as I have, to those 00:04:56.600 --> 00:04:59.960 align:middle line:84% who are really in the know about this area, 00:04:59.960 --> 00:05:04.160 align:middle line:84% and they don't have an immediate sales pitch to the venture 00:05:04.160 --> 00:05:05.840 align:middle line:84% capitalists on the horizon, they'll 00:05:05.840 --> 00:05:11.360 align:middle line:84% acknowledge to you that this is not ready for prime time. 00:05:11.360 --> 00:05:14.090 align:middle line:84% Which explains why I had so much cognitive dissonance, 00:05:14.090 --> 00:05:16.800 align:middle line:84% in the late 1990s, when everybody is saying, 00:05:16.800 --> 00:05:18.612 align:middle line:84% forget this genomics, proteomics is it. 00:05:18.612 --> 00:05:19.820 align:middle line:90% And I'm saying, that's great. 00:05:19.820 --> 00:05:21.530 align:middle line:90% I'd love to go to proteomics. 00:05:21.530 --> 00:05:25.280 align:middle line:84% But I don't see these cheap, high throughput, 00:05:25.280 --> 00:05:27.440 align:middle line:84% comprehensive, electronically data 00:05:27.440 --> 00:05:31.130 align:middle line:84% acquired sources available today. 00:05:31.130 --> 00:05:31.940 align:middle line:90% That's the goal. 00:05:31.940 --> 00:05:33.170 align:middle line:90% I believe we'll reach it. 00:05:33.170 --> 00:05:36.170 align:middle line:84% But I think we're actually a full five years, at least, 00:05:36.170 --> 00:05:39.005 align:middle line:84% away from a high throughput version of this. 00:05:39.005 --> 00:05:42.490 align:middle line:90% 00:05:42.490 --> 00:05:46.080 align:middle line:84% And what do we actually use these microarrays to look for? 00:05:46.080 --> 00:05:48.900 align:middle line:84% We rarely use it to find a single gene 00:05:48.900 --> 00:05:51.930 align:middle line:90% responsible for a process. 00:05:51.930 --> 00:05:54.300 align:middle line:84% But we frequently use it to find a pathway. 00:05:54.300 --> 00:05:57.900 align:middle line:84% We frequently use it to find a set of genes working 00:05:57.900 --> 00:05:59.408 align:middle line:90% in a coordinated fashion. 00:05:59.408 --> 00:06:01.200 align:middle line:84% The assumption being that there's structure 00:06:01.200 --> 00:06:03.120 align:middle line:90% at the biological scale. 00:06:03.120 --> 00:06:06.540 align:middle line:90% And what do I mean by that? 00:06:06.540 --> 00:06:08.392 align:middle line:84% That genes behave in coordinated fashion. 00:06:08.392 --> 00:06:10.350 align:middle line:84% And so the questions that immediately are asked 00:06:10.350 --> 00:06:12.180 align:middle line:84% is why should this assumption hold 00:06:12.180 --> 00:06:14.235 align:middle line:84% and what evidence do we have that it does? 00:06:14.235 --> 00:06:23.840 align:middle line:90% 00:06:23.840 --> 00:06:26.060 align:middle line:84% So why does clustering work in functional genomics? 00:06:26.060 --> 00:06:29.140 align:middle line:90% 00:06:29.140 --> 00:06:32.510 align:middle line:84% Well, what do we see in the night sky when we look up? 00:06:32.510 --> 00:06:37.850 align:middle line:84% We see a fairly, apparently uniform star field. 00:06:37.850 --> 00:06:39.770 align:middle line:84% But when we look with the right instruments, 00:06:39.770 --> 00:06:43.430 align:middle line:84% we see clusters of stars, from which we 00:06:43.430 --> 00:06:49.590 align:middle line:84% have been able to apply our classical and then Einsteinian 00:06:49.590 --> 00:06:53.570 align:middle line:84% mechanics to figure out what are the forces that 00:06:53.570 --> 00:06:57.260 align:middle line:84% hold these clusters together and make them behave 00:06:57.260 --> 00:06:59.270 align:middle line:90% the way we see them behaving. 00:06:59.270 --> 00:07:05.900 align:middle line:84% Similarly, when the Genome Project essentially 00:07:05.900 --> 00:07:07.820 align:middle line:84% came to a conclusion in this year, 00:07:07.820 --> 00:07:10.220 align:middle line:84% what we had was the gene universe. 00:07:10.220 --> 00:07:14.250 align:middle line:84% We had the list of genes that form part of this universe. 00:07:14.250 --> 00:07:19.790 align:middle line:84% And here, for instance, is a bunch of such genes. 00:07:19.790 --> 00:07:21.620 align:middle line:84% However, this really doesn't tell us 00:07:21.620 --> 00:07:23.570 align:middle line:84% anything about how these genes interrelate. 00:07:23.570 --> 00:07:26.028 align:middle line:84% And although there's a bunch of hoopla made about the Human 00:07:26.028 --> 00:07:30.320 align:middle line:84% Genome Project, it doesn't start to become really interesting 00:07:30.320 --> 00:07:32.000 align:middle line:84% until we can do something like this-- 00:07:32.000 --> 00:07:34.730 align:middle line:84% a figure that I borrowed from the Kyoto Encyclopedia 00:07:34.730 --> 00:07:36.950 align:middle line:84% of Genes and Genomes, KEGG, which 00:07:36.950 --> 00:07:38.360 align:middle line:90% is a pathway representation. 00:07:38.360 --> 00:07:40.880 align:middle line:84% Understanding that this is how the stars are 00:07:40.880 --> 00:07:43.880 align:middle line:84% interrelated in a galaxy, how these genes interrelate 00:07:43.880 --> 00:07:45.050 align:middle line:90% in pathways. 00:07:45.050 --> 00:07:49.145 align:middle line:90% So first of all, why should-- 00:07:49.145 --> 00:07:53.810 align:middle line:84% why do I see this structure in the transcriptome? 00:07:53.810 --> 00:07:56.210 align:middle line:84% It's because, in fact, in order to get things done, like, 00:07:56.210 --> 00:07:58.730 align:middle line:84% in this case, make sure this cell kills itself, 00:07:58.730 --> 00:08:02.810 align:middle line:84% there has to be coordinated expression of genes. 00:08:02.810 --> 00:08:04.730 align:middle line:84% As we explained to you previously, 00:08:04.730 --> 00:08:07.610 align:middle line:84% there are many reasons why gene expression and protein 00:08:07.610 --> 00:08:09.620 align:middle line:84% expression may not be correlated. 00:08:09.620 --> 00:08:16.280 align:middle line:84% But if there was no coordination at all of gene expression, 00:08:16.280 --> 00:08:17.690 align:middle line:90% two things would be true. 00:08:17.690 --> 00:08:23.315 align:middle line:84% One, the genes would be behaving randomly together-- 00:08:23.315 --> 00:08:25.190 align:middle line:84% in other words, would not be working together 00:08:25.190 --> 00:08:27.090 align:middle line:84% to make a process, which seems unlikely. 00:08:27.090 --> 00:08:29.510 align:middle line:84% And then clustering would never work. 00:08:29.510 --> 00:08:32.278 align:middle line:84% Because there would be no coordinated action of genes. 00:08:32.278 --> 00:08:34.070 align:middle line:84% And the reason this coordinated action is-- 00:08:34.070 --> 00:08:36.169 align:middle line:84% in order to get a job done, these, 00:08:36.169 --> 00:08:38.599 align:middle line:84% out of principles of parsimony-- that is, 00:08:38.599 --> 00:08:41.780 align:middle line:84% you don't want to have a gene product unless you need it-- 00:08:41.780 --> 00:08:44.660 align:middle line:84% these things are only generated when, in fact, you 00:08:44.660 --> 00:08:46.730 align:middle line:90% need them for this process-- 00:08:46.730 --> 00:08:49.250 align:middle line:90% in this case, cell death. 00:08:49.250 --> 00:08:53.783 align:middle line:84% And that allows people to do these clustering analysis. 00:08:53.783 --> 00:08:55.700 align:middle line:84% And there'll be another lecture on clustering. 00:08:55.700 --> 00:08:57.320 align:middle line:84% And I just want to give you a taste 00:08:57.320 --> 00:08:59.610 align:middle line:90% of what I mean by clustering. 00:08:59.610 --> 00:09:02.390 align:middle line:84% So here's the world's smallest microarray. 00:09:02.390 --> 00:09:05.120 align:middle line:90% You have three experiments. 00:09:05.120 --> 00:09:06.980 align:middle line:84% And on this little, small microarray 00:09:06.980 --> 00:09:10.640 align:middle line:84% you're only measuring a gene three times. 00:09:10.640 --> 00:09:12.590 align:middle line:84% And across these three experiments, 00:09:12.590 --> 00:09:15.800 align:middle line:84% you therefore have three measurements of each gene. 00:09:15.800 --> 00:09:18.110 align:middle line:84% And the most standard type of clustering 00:09:18.110 --> 00:09:21.920 align:middle line:84% is the so-called phylogenetic type tree or the dendrogram, 00:09:21.920 --> 00:09:24.950 align:middle line:84% where you calculate all the correlation coefficients 00:09:24.950 --> 00:09:28.430 align:middle line:84% between every column, in this case, every set of genes, 00:09:28.430 --> 00:09:31.700 align:middle line:84% and the correlation coefficients are shown, as here. 00:09:31.700 --> 00:09:34.400 align:middle line:84% You then figure out which genes are the closest together, 00:09:34.400 --> 00:09:37.250 align:middle line:84% based on their correlation metric, which, in this case, 00:09:37.250 --> 00:09:38.600 align:middle line:90% is 0.88. 00:09:38.600 --> 00:09:39.890 align:middle line:90% And you bring them together. 00:09:39.890 --> 00:09:42.650 align:middle line:84% And then you can summarize, in a variety of ways, 00:09:42.650 --> 00:09:47.810 align:middle line:84% that pair, let's say, by their median. 00:09:47.810 --> 00:09:54.440 align:middle line:84% And from there, you add on the next closest gene. 00:09:54.440 --> 00:09:57.890 align:middle line:84% So very specifically, gene one and gene two 00:09:57.890 --> 00:09:59.900 align:middle line:84% get put together, because they're the closest. 00:09:59.900 --> 00:10:06.320 align:middle line:84% And then the next closest to the joint branch 00:10:06.320 --> 00:10:07.730 align:middle line:90% there is gene three. 00:10:07.730 --> 00:10:09.860 align:middle line:84% And recursively, this gets built up 00:10:09.860 --> 00:10:14.420 align:middle line:84% until you have one of these classical, large dendrograms, 00:10:14.420 --> 00:10:18.800 align:middle line:84% where, again, each row is a gene. 00:10:18.800 --> 00:10:21.140 align:middle line:84% The columns are different experiments or samples, 00:10:21.140 --> 00:10:23.390 align:middle line:90% in this case, a time course. 00:10:23.390 --> 00:10:25.460 align:middle line:84% But they were different, distinct samples. 00:10:25.460 --> 00:10:29.090 align:middle line:84% And what you see here is that every row that's 00:10:29.090 --> 00:10:31.220 align:middle line:84% close to one another is close to one another 00:10:31.220 --> 00:10:35.113 align:middle line:84% because they have a higher correlation, between these two 00:10:35.113 --> 00:10:36.530 align:middle line:84% ones that are close to each other, 00:10:36.530 --> 00:10:39.500 align:middle line:84% than between two rows that are far apart. 00:10:39.500 --> 00:10:40.460 align:middle line:90% Now, there's something. 00:10:40.460 --> 00:10:43.430 align:middle line:90% I'm eliding some complexity. 00:10:43.430 --> 00:10:46.760 align:middle line:84% For instance, you can toggle around a whole branch, 00:10:46.760 --> 00:10:52.400 align:middle line:84% so that you might actually have some blocks, some rows, which 00:10:52.400 --> 00:10:55.940 align:middle line:84% might not be together in some rearrangements of this tree. 00:10:55.940 --> 00:10:58.940 align:middle line:84% Nonetheless, by and large, within a subbranch, 00:10:58.940 --> 00:11:02.593 align:middle line:84% you'll find, typically, that genes, 00:11:02.593 --> 00:11:04.010 align:middle line:84% with a similar expression pattern, 00:11:04.010 --> 00:11:07.730 align:middle line:84% are put together because they have the highest correlation 00:11:07.730 --> 00:11:08.270 align:middle line:90% coefficient. 00:11:08.270 --> 00:11:13.100 align:middle line:84% And the diagrammatic version of it is as follows. 00:11:13.100 --> 00:11:15.590 align:middle line:90% Higher expression level is red. 00:11:15.590 --> 00:11:17.280 align:middle line:90% Lower expression level is green. 00:11:17.280 --> 00:11:19.220 align:middle line:84% And you see that the red and the green bunches 00:11:19.220 --> 00:11:24.785 align:middle line:84% together, because, in fact, each of those genes, in each row, 00:11:24.785 --> 00:11:27.640 align:middle line:84% are clustered together by their expression pattern. 00:11:27.640 --> 00:11:28.790 align:middle line:90% That's all there is to it. 00:11:28.790 --> 00:11:31.240 align:middle line:84% And then when you look what these bunches actually 00:11:31.240 --> 00:11:35.153 align:middle line:84% represent and what people do is, essentially, they draw lines. 00:11:35.153 --> 00:11:36.820 align:middle line:84% And it's unfortunate, it's not much more 00:11:36.820 --> 00:11:38.950 align:middle line:84% fancy than that, at least in the publications 00:11:38.950 --> 00:11:41.770 align:middle line:90% that occurred from 2000-- 00:11:41.770 --> 00:11:43.750 align:middle line:90% 1999 to 2003. 00:11:43.750 --> 00:11:46.520 align:middle line:84% We draw lines next to these groups. 00:11:46.520 --> 00:11:49.960 align:middle line:84% You can actually find, for some of them, some functions, 00:11:49.960 --> 00:11:57.460 align:middle line:84% such as cell cycle or coagulation or apoptosis. 00:11:57.460 --> 00:12:00.100 align:middle line:84% You'll find different groups of genes clustered 00:12:00.100 --> 00:12:03.250 align:middle line:84% in that fashion that correspond to those different processes. 00:12:03.250 --> 00:12:07.390 align:middle line:84% Note, however, that for whatever reason, 00:12:07.390 --> 00:12:12.760 align:middle line:84% lines were not drawn against these other blocks of genes. 00:12:12.760 --> 00:12:17.350 align:middle line:84% And I'll argue, in another lecture, why that's a mistake. 00:12:17.350 --> 00:12:19.450 align:middle line:84% And furthermore, I'll argue in another lecture 00:12:19.450 --> 00:12:23.645 align:middle line:84% why this act of drawing lines, like these blocks, 00:12:23.645 --> 00:12:25.270 align:middle line:84% although it's ground in the basic truth 00:12:25.270 --> 00:12:28.030 align:middle line:84% that these genes more or less work together, where you draw 00:12:28.030 --> 00:12:31.930 align:middle line:84% the boundary of that line is an act of creative invagination 00:12:31.930 --> 00:12:35.300 align:middle line:84% and does not stand up well to statistical analysis. 00:12:35.300 --> 00:12:37.180 align:middle line:84% In other words, based on the knowledge 00:12:37.180 --> 00:12:38.800 align:middle line:84% of a particular biologist, you might 00:12:38.800 --> 00:12:43.330 align:middle line:84% decide to change the boundaries of that box. 00:12:43.330 --> 00:12:48.280 align:middle line:84% And there's well-grounded information theoretic methods 00:12:48.280 --> 00:12:52.360 align:middle line:84% which allow you to draw those functionally related groups 00:12:52.360 --> 00:12:54.430 align:middle line:90% more soundly. 00:12:54.430 --> 00:12:57.170 align:middle line:84% And this is just a close-up of that. 00:12:57.170 --> 00:12:59.590 align:middle line:84% So the point here is that, if you're not 00:12:59.590 --> 00:13:02.020 align:middle line:84% looking for a global pattern, as I just showed you, 00:13:02.020 --> 00:13:05.560 align:middle line:84% if you're not trying to see how things are working 00:13:05.560 --> 00:13:10.400 align:middle line:84% across a process, like coagulation or apoptosis, 00:13:10.400 --> 00:13:12.700 align:middle line:84% and if you're not looking for guilt by association, 00:13:12.700 --> 00:13:17.830 align:middle line:84% by which I mean, let's say you know 00:13:17.830 --> 00:13:21.220 align:middle line:84% the function of all these genes, and all these genes 00:13:21.220 --> 00:13:25.150 align:middle line:84% but not these genes, here, in the middle. 00:13:25.150 --> 00:13:27.307 align:middle line:84% Guilt by association says, essentially, 00:13:27.307 --> 00:13:29.140 align:middle line:84% that if you know the function of these genes 00:13:29.140 --> 00:13:30.700 align:middle line:84% and know the function of these genes, 00:13:30.700 --> 00:13:33.700 align:middle line:84% because these are coregulated in the same fashion, 00:13:33.700 --> 00:13:36.040 align:middle line:84% you can impute something about that process. 00:13:36.040 --> 00:13:39.520 align:middle line:84% And that heuristic course, as it may be, 00:13:39.520 --> 00:13:41.920 align:middle line:84% has been actually exploited very successfully 00:13:41.920 --> 00:13:45.490 align:middle line:90% by many investigators. 00:13:45.490 --> 00:13:48.583 align:middle line:84% So if you're not using guilt by association 00:13:48.583 --> 00:13:50.500 align:middle line:84% and if you're not looking for global patterns, 00:13:50.500 --> 00:13:52.300 align:middle line:84% then microarray technology may be 00:13:52.300 --> 00:13:54.640 align:middle line:90% inappropriate and misleading. 00:13:54.640 --> 00:13:56.680 align:middle line:84% And it may be inappropriate and misleading, 00:13:56.680 --> 00:13:59.770 align:middle line:84% because microarrays themselves are much more 00:13:59.770 --> 00:14:01.990 align:middle line:90% noisy than people understand. 00:14:01.990 --> 00:14:03.640 align:middle line:84% And therefore, if you're just looking-- 00:14:03.640 --> 00:14:05.765 align:middle line:84% specifically, if you trying to answer the question, 00:14:05.765 --> 00:14:07.910 align:middle line:84% do these three genes get expressed differentially, 00:14:07.910 --> 00:14:09.910 align:middle line:84% then you're probably using the wrong technology, 00:14:09.910 --> 00:14:14.170 align:middle line:84% for the wrong costs, and with much greater danger 00:14:14.170 --> 00:14:16.060 align:middle line:90% of being misled. 00:14:16.060 --> 00:14:18.050 align:middle line:90% And why is that a problem? 00:14:18.050 --> 00:14:21.940 align:middle line:84% Well, let me give you a decision theoretic view 00:14:21.940 --> 00:14:28.090 align:middle line:84% of microarray analysis as we should all, 00:14:28.090 --> 00:14:29.780 align:middle line:90% I think, think about it. 00:14:29.780 --> 00:14:32.770 align:middle line:84% So the most basic micro experiment 00:14:32.770 --> 00:14:36.190 align:middle line:84% asks the following question, across a set of experiments 00:14:36.190 --> 00:14:40.150 align:middle line:90% is a gene up or downregulated? 00:14:40.150 --> 00:14:41.860 align:middle line:84% Or more specifically, of the thousands 00:14:41.860 --> 00:14:44.590 align:middle line:84% of genes that we're measuring, which are up or downregulated? 00:14:44.590 --> 00:14:47.080 align:middle line:84% And for some, you'll use some statistic. 00:14:47.080 --> 00:14:49.600 align:middle line:84% And you find it, it is indeed upregulated. 00:14:49.600 --> 00:14:51.940 align:middle line:84% And for others, it is not upregulated. 00:14:51.940 --> 00:14:54.040 align:middle line:84% If a gene is said to be upregulated, 00:14:54.040 --> 00:14:56.170 align:middle line:84% is the gene relevant to the process being studied, 00:14:56.170 --> 00:14:57.130 align:middle line:90% let's say cancer? 00:14:57.130 --> 00:14:57.940 align:middle line:90% Yes. 00:14:57.940 --> 00:15:00.100 align:middle line:84% In that case, it's a true positive. 00:15:00.100 --> 00:15:02.750 align:middle line:84% It's perhaps a real target for cancer treatment. 00:15:02.750 --> 00:15:06.400 align:middle line:84% So therefore, big reward, because either you're 00:15:06.400 --> 00:15:08.560 align:middle line:84% a drug company or you're a researcher 00:15:08.560 --> 00:15:09.700 align:middle line:90% publishing a finding. 00:15:09.700 --> 00:15:12.400 align:middle line:84% If the gene is not relevant to the process being studied, 00:15:12.400 --> 00:15:14.010 align:middle line:90% it's a false positive. 00:15:14.010 --> 00:15:17.020 align:middle line:84% Therefore, you've wasted your money. 00:15:17.020 --> 00:15:19.360 align:middle line:84% If the gene is not significantly up or downregulated, 00:15:19.360 --> 00:15:21.985 align:middle line:84% in other words, it doesn't look interesting by this experiment, 00:15:21.985 --> 00:15:24.273 align:middle line:84% if the gene is relevant to the process being studied, 00:15:24.273 --> 00:15:25.190 align:middle line:90% then you've missed it. 00:15:25.190 --> 00:15:26.180 align:middle line:90% It's a false negative. 00:15:26.180 --> 00:15:29.350 align:middle line:84% So then, again, you've lost an opportunity. 00:15:29.350 --> 00:15:31.900 align:middle line:84% However, if the gene is not significantly up 00:15:31.900 --> 00:15:34.150 align:middle line:84% or downregulated, and it's not relevant to the process 00:15:34.150 --> 00:15:35.983 align:middle line:84% of being studied, then it's a true negative. 00:15:35.983 --> 00:15:37.090 align:middle line:90% So that's great. 00:15:37.090 --> 00:15:38.380 align:middle line:90% You're not wasting your time. 00:15:38.380 --> 00:15:40.220 align:middle line:90% And here's the problem. 00:15:40.220 --> 00:15:43.268 align:middle line:84% So you have 10,000 genes, let's say, or 20,000 genes 00:15:43.268 --> 00:15:45.310 align:middle line:84% that you're measuring in a microarray experiment. 00:15:45.310 --> 00:15:48.190 align:middle line:90% 00:15:48.190 --> 00:15:54.640 align:middle line:84% And the biologist and the big pharma 00:15:54.640 --> 00:15:57.850 align:middle line:84% can't bear the thought that they're 00:15:57.850 --> 00:16:02.440 align:middle line:84% going to miss the blockbuster false negative. 00:16:02.440 --> 00:16:05.680 align:middle line:90% 00:16:05.680 --> 00:16:09.940 align:middle line:84% And consequently, what they do is 00:16:09.940 --> 00:16:14.740 align:middle line:84% they change the threshold for declaring something to be a up 00:16:14.740 --> 00:16:20.500 align:middle line:84% or downregulated, so that more and more genes are considered 00:16:20.500 --> 00:16:23.140 align:middle line:90% to be up or downregulated. 00:16:23.140 --> 00:16:24.970 align:middle line:84% But the problem is that, not only does this 00:16:24.970 --> 00:16:28.100 align:middle line:84% increase, potentially, the number of true 00:16:28.100 --> 00:16:30.680 align:middle line:84% positivees you find, but it also hugely increases 00:16:30.680 --> 00:16:32.245 align:middle line:90% your number of false positives. 00:16:32.245 --> 00:16:33.620 align:middle line:84% And the problem again-- this gets 00:16:33.620 --> 00:16:34.700 align:middle line:84% back to the biologists don't understand 00:16:34.700 --> 00:16:36.283 align:middle line:84% the difference between a northern blot 00:16:36.283 --> 00:16:39.410 align:middle line:84% on a gene they understand and doing a microarray on 10,000 00:16:39.410 --> 00:16:40.970 align:middle line:84% genes that they know nothing about. 00:16:40.970 --> 00:16:43.610 align:middle line:84% Because they can be given literally 00:16:43.610 --> 00:16:45.260 align:middle line:90% thousands of false positives. 00:16:45.260 --> 00:16:47.000 align:middle line:84% And the typical interaction you'll 00:16:47.000 --> 00:16:48.980 align:middle line:84% have with the basic biologists, when 00:16:48.980 --> 00:16:53.780 align:middle line:84% you do this kind of analysis for them, they'll say, Oh, Zak, 00:16:53.780 --> 00:16:57.071 align:middle line:84% I see that you said, statistically significantly, 00:16:57.071 --> 00:16:59.180 align:middle line:84% we can only place bets about 10 genes. 00:16:59.180 --> 00:17:01.670 align:middle line:84% But I see down the list, that you 00:17:01.670 --> 00:17:04.040 align:middle line:84% said not to be significantly differentially expressed, 00:17:04.040 --> 00:17:07.369 align:middle line:84% a gene that I know from my own research 00:17:07.369 --> 00:17:09.050 align:middle line:84% is, in fact differentially expressed. 00:17:09.050 --> 00:17:10.760 align:middle line:84% But it just doesn't have a P value 00:17:10.760 --> 00:17:13.363 align:middle line:84% that reaches a level of significance. 00:17:13.363 --> 00:17:14.780 align:middle line:84% So therefore, couldn't we actually 00:17:14.780 --> 00:17:16.730 align:middle line:84% just push the whole threshold down 00:17:16.730 --> 00:17:21.119 align:middle line:84% to at least the same threshold as the gene 00:17:21.119 --> 00:17:22.619 align:middle line:84% that I already know something about? 00:17:22.619 --> 00:17:24.410 align:middle line:84% And the answer is, unfortunately, 00:17:24.410 --> 00:17:27.920 align:middle line:84% if you do that, then it'll still happen, 00:17:27.920 --> 00:17:29.600 align:middle line:84% you also-- there's nothing to say 00:17:29.600 --> 00:17:32.360 align:middle line:84% that these other genes, that are now included in the list, 00:17:32.360 --> 00:17:33.920 align:middle line:84% are anything at all like this gene 00:17:33.920 --> 00:17:35.150 align:middle line:84% that you happen to know a lot about. 00:17:35.150 --> 00:17:36.380 align:middle line:90% So you're using prior knowledge. 00:17:36.380 --> 00:17:37.290 align:middle line:90% You're not using the microarray. 00:17:37.290 --> 00:17:39.450 align:middle line:84% You're using prior knowledge to say something about it. 00:17:39.450 --> 00:17:41.010 align:middle line:84% And therefore, you get a lot of false positives. 00:17:41.010 --> 00:17:43.052 align:middle line:84% And the problem is each one of the false positive 00:17:43.052 --> 00:17:46.880 align:middle line:84% is one post-doc going off to do a validation study. 00:17:46.880 --> 00:17:49.100 align:middle line:84% And it turns out, with tens of thousands of genes, 00:17:49.100 --> 00:17:50.600 align:middle line:84% you can launch a lot of post-docs. 00:17:50.600 --> 00:17:54.080 align:middle line:84% So the pharma industry, with it's big, deep pockets, 00:17:54.080 --> 00:17:57.858 align:middle line:84% actually got exhausted by this and, actually, quite turned off 00:17:57.858 --> 00:17:59.150 align:middle line:90% by a lot of these technologies. 00:17:59.150 --> 00:18:02.412 align:middle line:84% Because a lot of the genes that they thought they were going 00:18:02.412 --> 00:18:04.370 align:middle line:84% to look after, these top genes, which look kind 00:18:04.370 --> 00:18:06.680 align:middle line:84% of interesting-- you can always make a story and create a story 00:18:06.680 --> 00:18:08.270 align:middle line:84% about why this gene may be involved in your favorite 00:18:08.270 --> 00:18:09.200 align:middle line:90% disease process-- 00:18:09.200 --> 00:18:10.680 align:middle line:90% turned out not to be the case. 00:18:10.680 --> 00:18:16.430 align:middle line:84% So by not understanding this decision theoretic view, 00:18:16.430 --> 00:18:19.070 align:middle line:84% a lot of money was wasted and a lot 00:18:19.070 --> 00:18:20.960 align:middle line:84% of disappointment with this technology, which 00:18:20.960 --> 00:18:22.460 align:middle line:84% is one of the reasons I wrote it, 00:18:22.460 --> 00:18:25.550 align:middle line:84% in the first sentence of the first chapter of my book, 00:18:25.550 --> 00:18:27.350 align:middle line:84% Microarrays for an Integrative Genomics, 00:18:27.350 --> 00:18:30.620 align:middle line:84% the functional genomic meltdown is imminent. 00:18:30.620 --> 00:18:32.870 align:middle line:90% And I first wrote that in 2000. 00:18:32.870 --> 00:18:34.968 align:middle line:90% And everything that I have-- 00:18:34.968 --> 00:18:37.010 align:middle line:84% everything that's happened since shows it's true, 00:18:37.010 --> 00:18:38.840 align:middle line:84% that, although there's more interesting science being 00:18:38.840 --> 00:18:40.970 align:middle line:84% developed, the fact is a lot of the companies that were built 00:18:40.970 --> 00:18:43.370 align:middle line:84% around this, and some of the bioinformatics companies 00:18:43.370 --> 00:18:47.870 align:middle line:84% that are built around this, are just going down the tubes. 00:18:47.870 --> 00:18:51.500 align:middle line:84% So let's just remind ourselves about-- 00:18:51.500 --> 00:18:54.750 align:middle line:84% by the way, I'm very optimistic about the whole field. 00:18:54.750 --> 00:18:58.640 align:middle line:84% But I'm optimistic about the whole field when we're careful 00:18:58.640 --> 00:19:03.440 align:middle line:84% in our use of the technology, both in understanding 00:19:03.440 --> 00:19:05.330 align:middle line:84% the actual engineering part of the technology 00:19:05.330 --> 00:19:08.190 align:middle line:84% and understanding the quantitative analysis. 00:19:08.190 --> 00:19:11.360 align:middle line:84% So I know that we've gone through this, with you, 00:19:11.360 --> 00:19:14.240 align:middle line:84% several times, but I want to revisit. 00:19:14.240 --> 00:19:20.540 align:middle line:84% Actually, is everybody comfortable in the way 00:19:20.540 --> 00:19:21.920 align:middle line:90% the basic microarrays work? 00:19:21.920 --> 00:19:25.195 align:middle line:90% Do I have to revisit it? 00:19:25.195 --> 00:19:25.820 align:middle line:90% If you want me? 00:19:25.820 --> 00:19:27.000 align:middle line:90% I do not have to revisit it? 00:19:27.000 --> 00:19:27.500 align:middle line:90% Thank you. 00:19:27.500 --> 00:19:29.780 align:middle line:90% All right. 00:19:29.780 --> 00:19:32.780 align:middle line:84% So now we can just remind ourselves 00:19:32.780 --> 00:19:37.460 align:middle line:84% about the difference between the two platforms. 00:19:37.460 --> 00:19:40.190 align:middle line:84% Affymetrix, as you know, does not 00:19:40.190 --> 00:19:44.000 align:middle line:84% represent, on each spot of the microarray, the entire gene. 00:19:44.000 --> 00:19:48.170 align:middle line:84% It just represents a few oligonucleotides 00:19:48.170 --> 00:19:50.690 align:middle line:84% that they have chosen to be representative 00:19:50.690 --> 00:19:52.130 align:middle line:90% of part of the gene-- 00:19:52.130 --> 00:19:56.300 align:middle line:90% more on the 3' end of the gene. 00:19:56.300 --> 00:19:58.970 align:middle line:84% And what they will calculate, as we will get to, 00:19:58.970 --> 00:20:03.140 align:middle line:84% is some measure of the hybridization 00:20:03.140 --> 00:20:06.200 align:middle line:84% of these perfectly matched probes to this gene. 00:20:06.200 --> 00:20:10.940 align:middle line:84% Versus a probe, which is exactly the same to this one 00:20:10.940 --> 00:20:16.820 align:middle line:84% but has a central base which is a mismatch to the designated 00:20:16.820 --> 00:20:18.210 align:middle line:90% area in that gene. 00:20:18.210 --> 00:20:22.965 align:middle line:84% So, there, the measure of the gene 00:20:22.965 --> 00:20:25.340 align:middle line:84% is essentially the difference between the predict matches 00:20:25.340 --> 00:20:29.540 align:middle line:84% and the mismatches in aggregate across the entire probe set. 00:20:29.540 --> 00:20:35.690 align:middle line:84% Whereas, in the Pat Brown inspired 00:20:35.690 --> 00:20:38.150 align:middle line:84% do-it-yourself spotted industry, what you're doing 00:20:38.150 --> 00:20:46.370 align:middle line:84% is you're spotting these probes right onto the glass slide. 00:20:46.370 --> 00:20:49.250 align:middle line:84% And then you're going to have a competitive hybridization 00:20:49.250 --> 00:20:58.250 align:middle line:84% between a reference set of cDNAs and the test set of cDNAs. 00:20:58.250 --> 00:21:03.013 align:middle line:84% And that has some implications which we'll get to shortly. 00:21:03.013 --> 00:21:04.430 align:middle line:84% So we'll make a point, because I'm 00:21:04.430 --> 00:21:06.847 align:middle line:84% going to ask you a question later, that the spotted arrays 00:21:06.847 --> 00:21:09.330 align:middle line:84% have to end up on the slide through some physical process, 00:21:09.330 --> 00:21:13.220 align:middle line:90% such as a printing head. 00:21:13.220 --> 00:21:18.320 align:middle line:84% Whereas in the full lithographic techniques built by Affymetrix, 00:21:18.320 --> 00:21:22.730 align:middle line:84% what you do is, after you create the first layer, which 00:21:22.730 --> 00:21:26.330 align:middle line:84% is bonded to the glass or silicon substrate, 00:21:26.330 --> 00:21:29.060 align:middle line:84% you're going to chemically add, let's 00:21:29.060 --> 00:21:33.980 align:middle line:84% say, a nucleotide G masking off the rest of the array 00:21:33.980 --> 00:21:38.450 align:middle line:84% where you don't want to add that G nucleotide. 00:21:38.450 --> 00:21:41.480 align:middle line:84% And then let's say you want to add an A in the next step, 00:21:41.480 --> 00:21:46.730 align:middle line:84% you mask off everywhere that you don't want to have the A added, 00:21:46.730 --> 00:21:48.560 align:middle line:90% and so on. 00:21:48.560 --> 00:21:49.970 align:middle line:90% Clear so far? 00:21:49.970 --> 00:21:54.380 align:middle line:84% Now the problem, of course, is that to do this for a 25 miR 00:21:54.380 --> 00:21:55.460 align:middle line:90% requires-- 00:21:55.460 --> 00:21:58.040 align:middle line:84% in other words, on the nucleotide that's 25 long-- 00:21:58.040 --> 00:22:03.380 align:middle line:84% is on the order of 100 masking procedures. 00:22:03.380 --> 00:22:05.900 align:middle line:84% Anybody have an idea how many masking procedures 00:22:05.900 --> 00:22:07.100 align:middle line:90% you do for a Pentium 4? 00:22:07.100 --> 00:22:09.700 align:middle line:90% 00:22:09.700 --> 00:22:11.860 align:middle line:90% On the order of 20 to 30. 00:22:11.860 --> 00:22:14.800 align:middle line:84% So they're having to do four times as much 00:22:14.800 --> 00:22:19.690 align:middle line:84% photolithography as for a Pentium 4. 00:22:19.690 --> 00:22:25.840 align:middle line:84% So that's going to limit their costs, at some level. 00:22:25.840 --> 00:22:29.440 align:middle line:84% And furthermore, let me just say it right here, it 00:22:29.440 --> 00:22:32.290 align:middle line:84% turns out that 25 miRs are not the ideal length 00:22:32.290 --> 00:22:33.760 align:middle line:90% for hybridization. 00:22:33.760 --> 00:22:36.400 align:middle line:84% The longer-- if you get to a longer length, let's 00:22:36.400 --> 00:22:39.400 align:middle line:84% say a 60 miR or 70 miR, you get much more sensitivity 00:22:39.400 --> 00:22:41.320 align:middle line:90% and much more specificity. 00:22:41.320 --> 00:22:43.870 align:middle line:84% But they can't economically do it 00:22:43.870 --> 00:22:46.060 align:middle line:84% because of the limits of photolithography. 00:22:46.060 --> 00:22:48.100 align:middle line:84% But the advantage of photolithography 00:22:48.100 --> 00:22:52.363 align:middle line:84% is that you can get extremely compact and high resolution. 00:22:52.363 --> 00:22:54.280 align:middle line:84% And so you can put a lot of genes on one chip, 00:22:54.280 --> 00:22:57.070 align:middle line:90% like the entire transcriptome. 00:22:57.070 --> 00:23:04.405 align:middle line:84% So this is just a summary of what I just said here. 00:23:04.405 --> 00:23:08.210 align:middle line:90% 00:23:08.210 --> 00:23:11.330 align:middle line:84% Also, we like to think of these in some 00:23:11.330 --> 00:23:13.160 align:middle line:90% of computational abstraction. 00:23:13.160 --> 00:23:15.170 align:middle line:84% But it's important also to realize 00:23:15.170 --> 00:23:19.760 align:middle line:84% that these are real molecules that, during the hybridization 00:23:19.760 --> 00:23:22.730 align:middle line:84% reaction, are tethered to the substrate. 00:23:22.730 --> 00:23:25.910 align:middle line:84% And therefore, their secondary and tertiary structure 00:23:25.910 --> 00:23:30.590 align:middle line:84% may in fact influence how they're 00:23:30.590 --> 00:23:33.980 align:middle line:84% able to hybridize with the molecules from the sample. 00:23:33.980 --> 00:23:41.390 align:middle line:84% And it turns out that certain types of DNA probes 00:23:41.390 --> 00:23:44.270 align:middle line:84% just don't work as well as you'd expect, computationally, 00:23:44.270 --> 00:23:45.920 align:middle line:90% because of these effects. 00:23:45.920 --> 00:23:49.790 align:middle line:84% And if you look at Affymetrix probe sets, 00:23:49.790 --> 00:23:51.710 align:middle line:84% there are some probes that never work. 00:23:51.710 --> 00:23:53.570 align:middle line:84% And despite the fact that they're 00:23:53.570 --> 00:23:55.070 align:middle line:84% well-designed from the point of view 00:23:55.070 --> 00:23:58.940 align:middle line:84% of finding a matching hybridization sequence. 00:23:58.940 --> 00:24:01.310 align:middle line:84% But it's because of these kind of effects. 00:24:01.310 --> 00:24:03.710 align:middle line:84% And note, also, that they are not all of the same length. 00:24:03.710 --> 00:24:06.290 align:middle line:84% The efficiency of the masking reaction 00:24:06.290 --> 00:24:08.090 align:middle line:90% is not 100% by any means. 00:24:08.090 --> 00:24:11.900 align:middle line:84% And what you will find in each one of those probes 00:24:11.900 --> 00:24:15.950 align:middle line:84% is, in fact, millions of DNA molecules, which 00:24:15.950 --> 00:24:17.720 align:middle line:84% have a distribution, which hopefully 00:24:17.720 --> 00:24:21.590 align:middle line:84% is close to the desired 25 miR, but, in fact, often 00:24:21.590 --> 00:24:23.480 align:middle line:90% is considerably lower. 00:24:23.480 --> 00:24:26.570 align:middle line:84% And unfortunately, there's no way for the end user 00:24:26.570 --> 00:24:29.660 align:middle line:84% to be able to ascertain just how good the job had done. 00:24:29.660 --> 00:24:37.760 align:middle line:90% 00:24:37.760 --> 00:24:39.800 align:middle line:84% So given that fact, we can already 00:24:39.800 --> 00:24:43.340 align:middle line:84% see that, even though we can achieve a higher 00:24:43.340 --> 00:24:47.240 align:middle line:84% density with these Affymetrix microarrays, 00:24:47.240 --> 00:24:49.190 align:middle line:84% there is still a lot of variation 00:24:49.190 --> 00:24:51.560 align:middle line:84% that we can expect from different probes 00:24:51.560 --> 00:24:53.900 align:middle line:90% within a probe set. 00:24:53.900 --> 00:24:56.360 align:middle line:84% Which is why, for example, you might not 00:24:56.360 --> 00:25:00.020 align:middle line:84% see, in a single probe set, that the perfect matches have 00:25:00.020 --> 00:25:01.290 align:middle line:90% all the same intensity. 00:25:01.290 --> 00:25:04.550 align:middle line:84% You can see, for instance, if there's 11 perfect matched 00:25:04.550 --> 00:25:07.640 align:middle line:84% probes, five of them will be really lit up 00:25:07.640 --> 00:25:11.240 align:middle line:90% and six will be quite dark. 00:25:11.240 --> 00:25:15.770 align:middle line:84% And there's many other reasons why that might be the case. 00:25:15.770 --> 00:25:18.920 align:middle line:84% And I just want to make the following point. 00:25:18.920 --> 00:25:22.670 align:middle line:84% As I said, when you're looking, for instance, in a sporulation 00:25:22.670 --> 00:25:28.520 align:middle line:84% experiment in yeast, what we're looking at is, let's say, 00:25:28.520 --> 00:25:36.170 align:middle line:84% a test sample versus a reference sample. 00:25:36.170 --> 00:25:39.590 align:middle line:84% And we are looking for their competitive hybridization 00:25:39.590 --> 00:25:42.150 align:middle line:90% to the target. 00:25:42.150 --> 00:25:44.780 align:middle line:84% Now, a couple of problems immediately 00:25:44.780 --> 00:25:48.890 align:middle line:84% arise, such as, unfortunately, it 00:25:48.890 --> 00:25:52.202 align:middle line:84% turns out, for reasons that have yet to be made clear to me-- 00:25:52.202 --> 00:25:54.410 align:middle line:84% and in fact, if anybody in this room has any insight, 00:25:54.410 --> 00:25:57.770 align:middle line:90% I'd be very curious. 00:25:57.770 --> 00:26:01.940 align:middle line:84% Which one gets the C3 versus C5 dye actually 00:26:01.940 --> 00:26:03.000 align:middle line:90% makes a difference. 00:26:03.000 --> 00:26:04.750 align:middle line:84% So if you switch around the dyes, 00:26:04.750 --> 00:26:06.150 align:middle line:90% you'll get different ratios. 00:26:06.150 --> 00:26:08.900 align:middle line:84% So something sterically about the attachment 00:26:08.900 --> 00:26:11.390 align:middle line:84% of the different dye causes a difference 00:26:11.390 --> 00:26:13.790 align:middle line:90% in the competition reaction. 00:26:13.790 --> 00:26:16.580 align:middle line:84% And so people who are practitioners of this 00:26:16.580 --> 00:26:19.370 align:middle line:84% and who are fastidious about it, will, in fact, 00:26:19.370 --> 00:26:25.670 align:middle line:84% perform both labels, both the Cy3, Cy5, and Cy5, 00:26:25.670 --> 00:26:28.670 align:middle line:84% Cy3 label to actually better understand 00:26:28.670 --> 00:26:30.240 align:middle line:90% the limits of that technology. 00:26:30.240 --> 00:26:32.370 align:middle line:84% It's not clear to me why that's happening. 00:26:32.370 --> 00:26:33.950 align:middle line:84% The other point to note is, if you 00:26:33.950 --> 00:26:39.980 align:middle line:84% want to compare across different microarray experiments, 00:26:39.980 --> 00:26:43.550 align:middle line:84% you probably want to have the same reference sample. 00:26:43.550 --> 00:26:45.050 align:middle line:84% And that sounds like a trivial thing 00:26:45.050 --> 00:26:48.200 align:middle line:84% to say, because you'd like to see, for instance, let's say, 00:26:48.200 --> 00:26:49.370 align:middle line:90% this condition. 00:26:49.370 --> 00:26:52.460 align:middle line:84% The test condition is a bunch of different cancers. 00:26:52.460 --> 00:26:56.840 align:middle line:84% And you want to compare it against a reference RNA, which 00:26:56.840 --> 00:26:58.430 align:middle line:90% might be at Stanford. 00:26:58.430 --> 00:27:00.110 align:middle line:84% What they do is they pool together 00:27:00.110 --> 00:27:04.140 align:middle line:84% all their different lymph node tissue types. 00:27:04.140 --> 00:27:15.035 align:middle line:84% Problem, of course, is, if you deplete your supply, 00:27:15.035 --> 00:27:16.910 align:middle line:84% you're going to get a different set of ratios 00:27:16.910 --> 00:27:18.120 align:middle line:90% with the new supply. 00:27:18.120 --> 00:27:22.940 align:middle line:84% And because you're going to have different abundance of RNAs 00:27:22.940 --> 00:27:27.950 align:middle line:84% in the reference sample, you're going to get different ratios. 00:27:27.950 --> 00:27:30.680 align:middle line:84% And so you see drift in the ratios 00:27:30.680 --> 00:27:37.320 align:middle line:84% across experiments, over time, with this kind of setup. 00:27:37.320 --> 00:27:39.710 align:middle line:90% And it's highly problematic. 00:27:39.710 --> 00:27:43.010 align:middle line:84% At Mass General, they've developed a synthetic soup 00:27:43.010 --> 00:27:46.170 align:middle line:90% of spiked-in RNA controls. 00:27:46.170 --> 00:27:48.620 align:middle line:84% I think 256 different RNA controls. 00:27:48.620 --> 00:27:51.230 align:middle line:84% And that will give you a standard reference pool. 00:27:51.230 --> 00:27:52.940 align:middle line:84% But unfortunately, it's not quite 00:27:52.940 --> 00:27:56.510 align:middle line:84% the same as having the full eukaryotic reference pool. 00:27:56.510 --> 00:27:59.330 align:middle line:84% So the ratios you might get from that artificial pool 00:27:59.330 --> 00:28:02.262 align:middle line:84% might be quite a bit different from what 00:28:02.262 --> 00:28:03.470 align:middle line:90% you get in other experiments. 00:28:03.470 --> 00:28:08.660 align:middle line:84% So just want to make it clear to you that is not a done deal. 00:28:08.660 --> 00:28:12.530 align:middle line:84% So how do you use, actually, Affymetrix 00:28:12.530 --> 00:28:17.120 align:middle line:84% to determine the intensity of gene expression? 00:28:17.120 --> 00:28:19.107 align:middle line:84% I just said that, for the spotted arrays, 00:28:19.107 --> 00:28:20.690 align:middle line:84% all you're doing is actually reporting 00:28:20.690 --> 00:28:23.030 align:middle line:84% a ratio, and a ratio that typically 00:28:23.030 --> 00:28:24.980 align:middle line:90% has a normal distribution. 00:28:24.980 --> 00:28:27.900 align:middle line:84% What you're doing in Affymetrix is a multi-step process. 00:28:27.900 --> 00:28:29.420 align:middle line:84% First of all, you're going to take 00:28:29.420 --> 00:28:32.750 align:middle line:84% the average of the lowest 2% of the cell intensities 00:28:32.750 --> 00:28:33.950 align:middle line:90% in the sector. 00:28:33.950 --> 00:28:36.860 align:middle line:84% This is the background of each sector. 00:28:36.860 --> 00:28:38.360 align:middle line:84% And then you subtract the background 00:28:38.360 --> 00:28:42.650 align:middle line:84% from the average intensities of all cells in this sector. 00:28:42.650 --> 00:28:44.760 align:middle line:84% And you calculate a background noise, 00:28:44.760 --> 00:28:48.350 align:middle line:84% which is essentially a measure of variance 00:28:48.350 --> 00:28:50.570 align:middle line:84% of the pixels and the scaling factor 00:28:50.570 --> 00:28:52.520 align:middle line:84% and normalization factors, which are settings 00:28:52.520 --> 00:28:54.320 align:middle line:90% you comply to Affymetrix. 00:28:54.320 --> 00:28:56.300 align:middle line:84% And this is actually somewhat out of date. 00:28:56.300 --> 00:29:00.690 align:middle line:84% But I'll give you updates shortly. 00:29:00.690 --> 00:29:03.990 align:middle line:84% So what is reported, for instance, 00:29:03.990 --> 00:29:10.370 align:middle line:84% most often as the intensity of gene expression 00:29:10.370 --> 00:29:12.230 align:middle line:90% is the average difference. 00:29:12.230 --> 00:29:15.860 align:middle line:84% Which essentially is the difference 00:29:15.860 --> 00:29:18.500 align:middle line:84% between the perfect matches and the mismatches 00:29:18.500 --> 00:29:22.080 align:middle line:90% across the entire probe set. 00:29:22.080 --> 00:29:23.660 align:middle line:84% So a probe pair, first of all, is 00:29:23.660 --> 00:29:26.090 align:middle line:84% said to be positive if the difference 00:29:26.090 --> 00:29:28.700 align:middle line:84% between the perfect match and the mismatch 00:29:28.700 --> 00:29:34.820 align:middle line:84% is greater than the amount of noise that is computed 00:29:34.820 --> 00:29:36.020 align:middle line:90% based on the background. 00:29:36.020 --> 00:29:40.070 align:middle line:84% And a probe pair is negative if the mismatch 00:29:40.070 --> 00:29:43.855 align:middle line:84% is greater than the perfect match by the same amount. 00:29:43.855 --> 00:29:45.230 align:middle line:84% So therefore, not all probe pairs 00:29:45.230 --> 00:29:46.940 align:middle line:84% will be scored as positive or negative. 00:29:46.940 --> 00:29:50.090 align:middle line:84% And then, what you have, you calculate 00:29:50.090 --> 00:29:58.580 align:middle line:84% a trimmed mean across the positive or negative probes. 00:29:58.580 --> 00:30:00.890 align:middle line:84% That turned out, by the way, to be 00:30:00.890 --> 00:30:07.280 align:middle line:84% a highly fluky, non-robust measure of intensity. 00:30:07.280 --> 00:30:13.070 align:middle line:84% And that was ensconced in an algorithm called MAS, 00:30:13.070 --> 00:30:15.500 align:middle line:90% MicroArray analysis Suite 4.0. 00:30:15.500 --> 00:30:19.370 align:middle line:84% So when you read papers and you see Mass 4.0, 00:30:19.370 --> 00:30:22.200 align:middle line:84% that's the method that they're using to interpret the data. 00:30:22.200 --> 00:30:26.600 align:middle line:84% And the problem is, it was both oversensitive to outliers, 00:30:26.600 --> 00:30:28.100 align:middle line:84% because they're using the term mean, 00:30:28.100 --> 00:30:35.550 align:middle line:84% and it was also sensitive to the central moment. 00:30:35.550 --> 00:30:42.980 align:middle line:84% So instead, in MAS 5.0, they are using the Tukey biweight 00:30:42.980 --> 00:30:48.320 align:middle line:84% statistic, which is a much more robust statistic of outliers. 00:30:48.320 --> 00:30:52.730 align:middle line:84% And I will upload a paper today to the My Course website, which 00:30:52.730 --> 00:30:55.850 align:middle line:84% will allow you to read the analysis of how they approached 00:30:55.850 --> 00:30:56.780 align:middle line:90% this. 00:30:56.780 --> 00:30:58.490 align:middle line:84% And this was actually done by Affymetrix. 00:30:58.490 --> 00:31:01.550 align:middle line:84% And it's now part of their standard interpretation 00:31:01.550 --> 00:31:06.050 align:middle line:90% software, MAS 5.0. 00:31:06.050 --> 00:31:10.910 align:middle line:84% Another competing microarray or high throughput gene expression 00:31:10.910 --> 00:31:14.340 align:middle line:84% technology is a serial analysis of gene expression. 00:31:14.340 --> 00:31:16.730 align:middle line:84% Did Atul tell you about this at all? 00:31:16.730 --> 00:31:17.300 align:middle line:90% He did? 00:31:17.300 --> 00:31:17.900 align:middle line:90% All right. 00:31:17.900 --> 00:31:20.787 align:middle line:84% So suffice it to say that it looks digital, 00:31:20.787 --> 00:31:22.370 align:middle line:84% in that what you're doing is literally 00:31:22.370 --> 00:31:25.100 align:middle line:84% counting the number of tags with a particular sequence. 00:31:25.100 --> 00:31:27.290 align:middle line:84% But it turns out this has noise, as well, 00:31:27.290 --> 00:31:28.440 align:middle line:90% for a number of reasons. 00:31:28.440 --> 00:31:32.150 align:middle line:84% One, our sequencing technology is not 100%. 00:31:32.150 --> 00:31:34.160 align:middle line:90% It has error. 00:31:34.160 --> 00:31:38.090 align:middle line:84% And also, sometimes the tags are not perfectly chosen, either. 00:31:38.090 --> 00:31:42.170 align:middle line:90% So it's actually, in the right-- 00:31:42.170 --> 00:31:42.710 align:middle line:90% excuse me. 00:31:42.710 --> 00:31:44.730 align:middle line:84% In the right hands, with the right set of genes, 00:31:44.730 --> 00:31:45.650 align:middle line:90% it does a nice job. 00:31:45.650 --> 00:31:47.600 align:middle line:84% But in actual practice, I've seen 00:31:47.600 --> 00:31:52.010 align:middle line:84% it to be fairly noisy and certainly not cost competitive 00:31:52.010 --> 00:31:57.620 align:middle line:84% with the standard microarray technology. 00:31:57.620 --> 00:32:01.280 align:middle line:84% Quite interesting is this very nice technology 00:32:01.280 --> 00:32:02.450 align:middle line:90% out of Illumina. 00:32:02.450 --> 00:32:04.202 align:middle line:84% So they're using this for genotyping, 00:32:04.202 --> 00:32:05.660 align:middle line:84% but they're thinking about using it 00:32:05.660 --> 00:32:07.910 align:middle line:84% for doing expression microarrays. 00:32:07.910 --> 00:32:09.680 align:middle line:84% And it's as follows, you actually 00:32:09.680 --> 00:32:12.200 align:middle line:90% have a fiber optic bundle. 00:32:12.200 --> 00:32:16.790 align:middle line:84% And the tip of the fiber optic bundle 00:32:16.790 --> 00:32:24.410 align:middle line:84% is sculpted so that it fits in nicely. 00:32:24.410 --> 00:32:28.370 align:middle line:90% It has a groove that fits in. 00:32:28.370 --> 00:32:29.750 align:middle line:84% Yeah, it has a groove, which fits 00:32:29.750 --> 00:32:33.980 align:middle line:84% a bump, nicely, on these optically coated beads, 00:32:33.980 --> 00:32:35.340 align:middle line:90% so that we can do the following. 00:32:35.340 --> 00:32:40.700 align:middle line:84% We can attach to each of the different optically coated 00:32:40.700 --> 00:32:44.850 align:middle line:90% beads a different piece of DNA. 00:32:44.850 --> 00:32:47.240 align:middle line:84% So we could literally have millions 00:32:47.240 --> 00:32:51.920 align:middle line:84% of different optically coated beads, each 00:32:51.920 --> 00:32:54.750 align:middle line:90% with their different DNAs. 00:32:54.750 --> 00:32:59.660 align:middle line:84% And then they will settle into the groove on the end 00:32:59.660 --> 00:33:02.840 align:middle line:84% of the sculpted fiber optic cable. 00:33:02.840 --> 00:33:05.300 align:middle line:84% And then you can perform a hybridization reaction 00:33:05.300 --> 00:33:07.430 align:middle line:90% with a sample. 00:33:07.430 --> 00:33:13.070 align:middle line:84% And then, by sending down a laser, down that fiber 00:33:13.070 --> 00:33:16.820 align:middle line:84% optic cable, you can actually determine which of these beads 00:33:16.820 --> 00:33:21.133 align:middle line:84% has a hybridized piece of DNA versus which does not. 00:33:21.133 --> 00:33:23.300 align:middle line:84% And this allows you to do, in a very high throughput 00:33:23.300 --> 00:33:29.150 align:middle line:84% fashion, millions of different hybridization reactions. 00:33:29.150 --> 00:33:31.760 align:middle line:84% It's currently being used for doing millions 00:33:31.760 --> 00:33:32.970 align:middle line:90% of different genotypes. 00:33:32.970 --> 00:33:35.510 align:middle line:84% So it's recognizing a whole bunch of different genotypes 00:33:35.510 --> 00:33:37.822 align:middle line:84% with different pieces of DNA corresponding 00:33:37.822 --> 00:33:38.780 align:middle line:90% to different genotypes. 00:33:38.780 --> 00:33:41.510 align:middle line:84% But, also, it's been it can be used 00:33:41.510 --> 00:33:44.780 align:middle line:90% to do expression microarrays. 00:33:44.780 --> 00:33:47.000 align:middle line:84% And there are some groups, around the country, that 00:33:47.000 --> 00:33:48.188 align:middle line:90% are now beginning to use it. 00:33:48.188 --> 00:33:50.480 align:middle line:84% And for those of you who are in the enterprise program, 00:33:50.480 --> 00:33:52.520 align:middle line:84% looking at how Illumina was funded 00:33:52.520 --> 00:33:56.360 align:middle line:84% and how they got together with scientists to avoid being sued 00:33:56.360 --> 00:33:59.093 align:middle line:84% by Affymetrix, it's actually a great case history, 00:33:59.093 --> 00:34:00.260 align:middle line:90% which I don't have time for. 00:34:00.260 --> 00:34:02.780 align:middle line:84% But the point is that Affymetrix has a very broad patent 00:34:02.780 --> 00:34:07.340 align:middle line:84% around the measurement of expression in two dimensions, 00:34:07.340 --> 00:34:10.370 align:middle line:84% in a two dimensional form factor, 00:34:10.370 --> 00:34:14.120 align:middle line:84% essentially, with any way that you place these DNA spots down. 00:34:14.120 --> 00:34:18.320 align:middle line:84% And this gets away from that patent in a very creative way 00:34:18.320 --> 00:34:21.049 align:middle line:90% and also in an economical way. 00:34:21.049 --> 00:34:22.219 align:middle line:90% AUDIENCE: Two questions. 00:34:22.219 --> 00:34:22.790 align:middle line:90% ISAAC SAMUEL KOHANE: Sure. 00:34:22.790 --> 00:34:24.207 align:middle line:84% AUDIENCE: Is the main benefit here 00:34:24.207 --> 00:34:27.238 align:middle line:84% that you don't have the tags or the data input down the line? 00:34:27.238 --> 00:34:29.280 align:middle line:84% ISAAC SAMUEL KOHANE: There's a couple of factors. 00:34:29.280 --> 00:34:34.670 align:middle line:84% First of all, you're not limited in having this form 00:34:34.670 --> 00:34:37.250 align:middle line:84% factor, this big, where you're going 00:34:37.250 --> 00:34:38.540 align:middle line:90% to do this photolithograph. 00:34:38.540 --> 00:34:41.570 align:middle line:84% Even so, with the photolithographic process, 00:34:41.570 --> 00:34:44.659 align:middle line:84% because of the registration of scanning, 00:34:44.659 --> 00:34:48.230 align:middle line:84% the registration of the full lithography, 00:34:48.230 --> 00:34:53.120 align:middle line:84% there's limits to how many different probes 00:34:53.120 --> 00:34:54.230 align:middle line:90% you can have on one chip. 00:34:54.230 --> 00:35:01.730 align:middle line:84% Whereas here, because each bead can have its own unique ID, 00:35:01.730 --> 00:35:08.220 align:middle line:84% so you know which one it is, and can have its own piece of DNA, 00:35:08.220 --> 00:35:09.500 align:middle line:90% it's going to self-assemble. 00:35:09.500 --> 00:35:12.710 align:middle line:84% And this is a very microscopic view of this. 00:35:12.710 --> 00:35:15.080 align:middle line:84% You can have many more different, 00:35:15.080 --> 00:35:17.060 align:middle line:84% because it's essentially self-assembly. 00:35:17.060 --> 00:35:20.270 align:middle line:84% You swirl around the fiber-optic cable in the soup of beads. 00:35:20.270 --> 00:35:21.800 align:middle line:84% And they basically all self-assemble 00:35:21.800 --> 00:35:23.810 align:middle line:90% onto this tip of that cable. 00:35:23.810 --> 00:35:28.100 align:middle line:84% You get many, many more different types of beads 00:35:28.100 --> 00:35:31.040 align:middle line:84% and, therefore, of probes in one sample. 00:35:31.040 --> 00:35:38.930 align:middle line:84% And the process of labeling the beads and attaching the probe 00:35:38.930 --> 00:35:41.690 align:middle line:84% is actually much cheaper than the Affymetrix technique. 00:35:41.690 --> 00:35:44.545 align:middle line:84% So it's higher density, more economical, 00:35:44.545 --> 00:35:45.920 align:middle line:84% and it gets away from the patent. 00:35:45.920 --> 00:35:50.390 align:middle line:90% 00:35:50.390 --> 00:35:53.090 align:middle line:84% Another technology is inkjet technology. 00:35:53.090 --> 00:35:57.410 align:middle line:84% The same people who brought you high resolution family 00:35:57.410 --> 00:36:00.350 align:middle line:84% photos said, if we're able to put down picoliters 00:36:00.350 --> 00:36:03.830 align:middle line:84% in precisely the right spot so that your kid looks 00:36:03.830 --> 00:36:05.840 align:middle line:84% the way your kid looks, perhaps we 00:36:05.840 --> 00:36:12.950 align:middle line:84% can use that same technology to spot a piece of DNA 00:36:12.950 --> 00:36:15.300 align:middle line:90% onto a flat surface. 00:36:15.300 --> 00:36:17.510 align:middle line:90% And in fact, they use-- 00:36:17.510 --> 00:36:19.280 align:middle line:84% So a spin-off of Hewlett-Packard, 00:36:19.280 --> 00:36:21.080 align:middle line:84% before it split off into Agilent, 00:36:21.080 --> 00:36:24.198 align:middle line:84% so a spin-off of Hewlett-Packard printers, 00:36:24.198 --> 00:36:26.240 align:middle line:84% actually developed the system where they actually 00:36:26.240 --> 00:36:30.590 align:middle line:84% have four colors, CATG, and they spit the four 00:36:30.590 --> 00:36:33.290 align:middle line:90% colors onto the glass slide. 00:36:33.290 --> 00:36:38.930 align:middle line:84% And they do successive in-situ synthesis of oligomers, 00:36:38.930 --> 00:36:40.520 align:middle line:90% 50 to 60 long. 00:36:40.520 --> 00:36:43.550 align:middle line:84% And because your pictures of your baby 00:36:43.550 --> 00:36:47.090 align:middle line:84% have to look good and have much more high tolerances 00:36:47.090 --> 00:36:52.040 align:middle line:84% than anybody cares about in Pat Brown's group, 00:36:52.040 --> 00:36:55.580 align:middle line:84% let's say, these are very good looking spots, 00:36:55.580 --> 00:36:57.800 align:middle line:84% with very nice, consistent shapes. 00:36:57.800 --> 00:37:00.860 align:middle line:90% And it's also very cheap. 00:37:00.860 --> 00:37:04.280 align:middle line:84% So this is the kind of shapes that you 00:37:04.280 --> 00:37:08.930 align:middle line:84% get with mechanical pens, which is the normal way you spot. 00:37:08.930 --> 00:37:11.150 align:middle line:84% And these are much more consistent 00:37:11.150 --> 00:37:14.705 align:middle line:84% and precise spots that you get with the inkjet. 00:37:14.705 --> 00:37:17.930 align:middle line:90% 00:37:17.930 --> 00:37:20.510 align:middle line:84% But there's many other uses for microarrays. 00:37:20.510 --> 00:37:23.040 align:middle line:84% If you recall at the beginning of the lecture, 00:37:23.040 --> 00:37:26.510 align:middle line:84% I didn't describe a microarray as being about expression. 00:37:26.510 --> 00:37:30.290 align:middle line:84% I said it had to be able to have a small form factor, 00:37:30.290 --> 00:37:34.550 align:middle line:84% it had to be able to interrogate for a large amount of the -ome 00:37:34.550 --> 00:37:35.600 align:middle line:90% that you're measuring. 00:37:35.600 --> 00:37:39.420 align:middle line:84% And it had to have a direct data path to an electronic format, 00:37:39.420 --> 00:37:40.370 align:middle line:90% and so on. 00:37:40.370 --> 00:37:44.360 align:middle line:84% So can we use these same kind of microarrays 00:37:44.360 --> 00:37:50.150 align:middle line:84% to study, for instance, gene expression control? 00:37:50.150 --> 00:37:54.530 align:middle line:84% So zinc finger proteins are proteins 00:37:54.530 --> 00:37:57.770 align:middle line:90% that actually bind to DNA. 00:37:57.770 --> 00:38:03.320 align:middle line:84% As you see here, so this finger is around the DNA molecule. 00:38:03.320 --> 00:38:06.230 align:middle line:84% And there's different parts of this that do the recognition 00:38:06.230 --> 00:38:11.240 align:middle line:84% and binding to a consensus site on the DNA. 00:38:11.240 --> 00:38:14.410 align:middle line:84% And so for understanding that, understanding 00:38:14.410 --> 00:38:18.500 align:middle line:84% that it was a consensus sequence with a limited repertoire 00:38:18.500 --> 00:38:20.960 align:middle line:84% of possible values, you could actually 00:38:20.960 --> 00:38:26.870 align:middle line:84% create a oligonucleotide that had 00:38:26.870 --> 00:38:29.735 align:middle line:84% all the possible permutations for this recognition site. 00:38:29.735 --> 00:38:32.650 align:middle line:90% 00:38:32.650 --> 00:38:39.170 align:middle line:84% And then you then ligate that oligonucleotide to a surface, 00:38:39.170 --> 00:38:42.770 align:middle line:90% like the microarray. 00:38:42.770 --> 00:38:47.240 align:middle line:84% And this is the actual results that you 00:38:47.240 --> 00:38:53.420 align:middle line:84% obtain when you take different transcription factor mutants. 00:38:53.420 --> 00:38:57.410 align:middle line:84% And you see that they bind at different spots, 00:38:57.410 --> 00:39:02.270 align:middle line:84% on the microarray, based on their affinity 00:39:02.270 --> 00:39:03.770 align:middle line:84% with a different consensus sequence. 00:39:03.770 --> 00:39:07.100 align:middle line:84% And since you know which spot corresponds to which consensus 00:39:07.100 --> 00:39:09.140 align:middle line:84% sequence, you can quickly develop 00:39:09.140 --> 00:39:12.260 align:middle line:84% a notion of what are the consensus sequences, what 00:39:12.260 --> 00:39:14.965 align:middle line:84% is shared and what is not shared among these mutants, 00:39:14.965 --> 00:39:16.340 align:middle line:84% and, therefore, what is important 00:39:16.340 --> 00:39:18.650 align:middle line:90% for the binding activity. 00:39:18.650 --> 00:39:25.490 align:middle line:84% As shown by this cartoon that represents the conservation 00:39:25.490 --> 00:39:28.650 align:middle line:84% of picker binding consensus sequence. 00:39:28.650 --> 00:39:30.440 align:middle line:84% So a lot of this work actually was 00:39:30.440 --> 00:39:34.310 align:middle line:84% pioneered by one of our faculty members in HST, Martha Bulyk. 00:39:34.310 --> 00:39:37.130 align:middle line:84% And it's now been scaled up so you can actually 00:39:37.130 --> 00:39:39.290 align:middle line:84% buy a lot of microarrays that allow 00:39:39.290 --> 00:39:46.070 align:middle line:84% you to do much larger-scale assessments of thousands 00:39:46.070 --> 00:39:47.885 align:middle line:90% of different binding sites. 00:39:47.885 --> 00:39:50.510 align:middle line:90% 00:39:50.510 --> 00:39:53.600 align:middle line:84% These are the protein microarrays that I 00:39:53.600 --> 00:39:55.950 align:middle line:90% discussed with you previously. 00:39:55.950 --> 00:40:01.280 align:middle line:84% And you can have a variety of different baits. 00:40:01.280 --> 00:40:06.440 align:middle line:84% A small molecule present on a microarray, an antibody present 00:40:06.440 --> 00:40:08.270 align:middle line:84% on the microarray, or a different kind 00:40:08.270 --> 00:40:10.940 align:middle line:90% of protein baits. 00:40:10.940 --> 00:40:15.440 align:middle line:84% Or you could even use a phage as a bait. 00:40:15.440 --> 00:40:18.200 align:middle line:84% The problem with all of these is that, as I said, 00:40:18.200 --> 00:40:22.220 align:middle line:84% we don't have yet, particularly for the proteins, 00:40:22.220 --> 00:40:25.440 align:middle line:84% high throughput methods of doing this. 00:40:25.440 --> 00:40:29.210 align:middle line:84% So for instance, selecting and laying down 00:40:29.210 --> 00:40:31.910 align:middle line:84% the antibodies for the entire proteome 00:40:31.910 --> 00:40:36.260 align:middle line:90% is just not within our vision. 00:40:36.260 --> 00:40:38.420 align:middle line:84% We just don't know how to get there. 00:40:38.420 --> 00:40:41.820 align:middle line:84% Currently, in a reproducible fashion, 00:40:41.820 --> 00:40:46.770 align:middle line:84% we can do it for maybe hundreds of different proteins. 00:40:46.770 --> 00:40:48.770 align:middle line:84% And these are the kinds of results that you get. 00:40:48.770 --> 00:40:50.728 align:middle line:84% You do, in fact, get these good looking graphs, 00:40:50.728 --> 00:40:52.400 align:middle line:84% where you can show that you can reliably 00:40:52.400 --> 00:40:54.650 align:middle line:84% identify the differentiable expression 00:40:54.650 --> 00:40:56.270 align:middle line:90% of different proteins. 00:40:56.270 --> 00:41:02.730 align:middle line:84% But again, these are from very small numbers of genes. 00:41:02.730 --> 00:41:03.230 align:middle line:90% Whoops. 00:41:03.230 --> 00:41:06.540 align:middle line:90% 00:41:06.540 --> 00:41:11.430 align:middle line:84% Others have tried the notion of a universal microarray, where 00:41:11.430 --> 00:41:15.690 align:middle line:84% what you're doing is, rather than having to re-engineer 00:41:15.690 --> 00:41:17.508 align:middle line:84% the microarray all the time, you're 00:41:17.508 --> 00:41:19.050 align:middle line:84% going to create a generic microarray. 00:41:19.050 --> 00:41:23.100 align:middle line:84% And then you're going to create some custom technology 00:41:23.100 --> 00:41:25.890 align:middle line:84% to link that generic microarray to the system 00:41:25.890 --> 00:41:27.280 align:middle line:90% that you're interested in. 00:41:27.280 --> 00:41:30.990 align:middle line:84% So this universal microarray has these zip codes, 00:41:30.990 --> 00:41:36.060 align:middle line:84% which basically are standardized sequences that 00:41:36.060 --> 00:41:39.000 align:middle line:84% recognize a bit of DNA with a complementary sequence, which 00:41:39.000 --> 00:41:42.840 align:middle line:84% you then add chemically to the sequence 00:41:42.840 --> 00:41:45.150 align:middle line:84% that you want to interrogate for. 00:41:45.150 --> 00:41:49.260 align:middle line:84% And such universal arrays don't require re-engineering 00:41:49.260 --> 00:41:55.170 align:middle line:84% of the zip code part of the probe, 00:41:55.170 --> 00:41:57.720 align:middle line:84% but only the part of the probe that you 00:41:57.720 --> 00:42:01.020 align:middle line:90% want to interrogate for. 00:42:01.020 --> 00:42:03.180 align:middle line:84% And it works actually reasonably well, 00:42:03.180 --> 00:42:05.220 align:middle line:84% as is shown in these experiments. 00:42:05.220 --> 00:42:08.880 align:middle line:84% I want to point out that a lot of these technologies 00:42:08.880 --> 00:42:14.070 align:middle line:84% became popular when Affymetrix microarrays were more 00:42:14.070 --> 00:42:16.500 align:middle line:84% poorly performing and much more expensive. 00:42:16.500 --> 00:42:19.230 align:middle line:84% Four years ago, an Affymetrix microarray at Harvard 00:42:19.230 --> 00:42:23.340 align:middle line:84% cost, with their academic discount, about $2,000. 00:42:23.340 --> 00:42:30.600 align:middle line:84% Now it costs about $200 for eight times as many genes. 00:42:30.600 --> 00:42:34.200 align:middle line:84% So the price factor has gone way down. 00:42:34.200 --> 00:42:36.555 align:middle line:84% These other technologies are just going to go away. 00:42:36.555 --> 00:42:40.020 align:middle line:84% It's sort of a Microsoft effect, the technology 00:42:40.020 --> 00:42:42.870 align:middle line:84% is good enough and cheap enough that there's 00:42:42.870 --> 00:42:45.120 align:middle line:84% very little incentive to invest a lot of your own time 00:42:45.120 --> 00:42:47.010 align:middle line:84% to make these other technologies better. 00:42:47.010 --> 00:42:50.160 align:middle line:84% Plus, as I said, there's patent control 00:42:50.160 --> 00:42:54.870 align:middle line:84% by Affymetrix, which makes it harder to create a business out 00:42:54.870 --> 00:42:55.596 align:middle line:90% there. 00:42:55.596 --> 00:42:57.300 align:middle line:84% AUDIENCE: Isn't it for things that 00:42:57.300 --> 00:43:00.000 align:middle line:84% were being published three, four years ago, maybe, were 00:43:00.000 --> 00:43:02.703 align:middle line:90% [INAUDIBLE]. 00:43:02.703 --> 00:43:04.620 align:middle line:84% ISAAC SAMUEL KOHANE: The short answer is, yes. 00:43:04.620 --> 00:43:07.740 align:middle line:84% And then I'm going to show you, quite literally, 00:43:07.740 --> 00:43:12.570 align:middle line:84% just how bad that intergenerational difference 00:43:12.570 --> 00:43:15.003 align:middle line:90% is. 00:43:15.003 --> 00:43:16.920 align:middle line:84% You have heard a lot about tissue microarrays. 00:43:16.920 --> 00:43:18.087 align:middle line:90% What are tissue microarrays? 00:43:18.087 --> 00:43:20.220 align:middle line:84% They're essentially these collections 00:43:20.220 --> 00:43:24.610 align:middle line:84% of hopefully homogeneous samples. 00:43:24.610 --> 00:43:26.940 align:middle line:84% So for instance, you could take a tumor 00:43:26.940 --> 00:43:29.820 align:middle line:84% and could slice it and dice it into these small salami slices 00:43:29.820 --> 00:43:31.320 align:middle line:84% that you lay down on the microarray. 00:43:31.320 --> 00:43:33.403 align:middle line:84% And then you can stain it for a variety of things. 00:43:33.403 --> 00:43:36.130 align:middle line:84% And then you can look under the microscope. 00:43:36.130 --> 00:43:38.760 align:middle line:84% And if you're clever, get an image recognition program, 00:43:38.760 --> 00:43:42.810 align:middle line:84% to detect a particular color for a particular stain, 00:43:42.810 --> 00:43:47.310 align:middle line:84% to see how much of a process is present, based on the staining 00:43:47.310 --> 00:43:51.420 align:middle line:84% characteristics of that tissue or the morphological change 00:43:51.420 --> 00:43:52.140 align:middle line:90% in that tissue. 00:43:52.140 --> 00:43:54.690 align:middle line:84% It's not yet a high throughput technique. 00:43:54.690 --> 00:43:57.450 align:middle line:84% And if sort of fails the microarray definition 00:43:57.450 --> 00:43:58.680 align:middle line:90% that I discussed previously. 00:43:58.680 --> 00:44:02.340 align:middle line:84% But I think that as we get better in the image analysis 00:44:02.340 --> 00:44:05.250 align:middle line:84% part of it, it does have a shot of becoming much 00:44:05.250 --> 00:44:08.940 align:middle line:90% more of a commodity microarray. 00:44:08.940 --> 00:44:10.320 align:middle line:90% AUDIENCE: What have-- 00:44:10.320 --> 00:44:11.570 align:middle line:90% ISAAC SAMUEL KOHANE: Go ahead. 00:44:11.570 --> 00:44:13.320 align:middle line:84% AUDIENCE: How could they ever do something 00:44:13.320 --> 00:44:15.933 align:middle line:84% like this, though, given the fact that 00:44:15.933 --> 00:44:20.140 align:middle line:84% in the chemical environments, themselves, [INAUDIBLE] 00:44:20.140 --> 00:44:21.000 align:middle line:90% incredible. 00:44:21.000 --> 00:44:22.840 align:middle line:84% Not only are you dealing with molecules, 00:44:22.840 --> 00:44:26.560 align:middle line:84% you're dealing with a very complex, poorly understood 00:44:26.560 --> 00:44:27.400 align:middle line:90% system. 00:44:27.400 --> 00:44:28.525 align:middle line:90% ISAAC SAMUEL KOHANE: Right. 00:44:28.525 --> 00:44:30.550 align:middle line:90% AUDIENCE: Well-- 00:44:30.550 --> 00:44:33.370 align:middle line:84% ISAAC SAMUEL KOHANE: Well, the answer is the following. 00:44:33.370 --> 00:44:35.350 align:middle line:84% The question that you just raised 00:44:35.350 --> 00:44:37.720 align:middle line:84% is almost always true in genomics. 00:44:37.720 --> 00:44:39.010 align:middle line:90% It's a more general question. 00:44:39.010 --> 00:44:41.330 align:middle line:84% With complex systems, how can we do this? 00:44:41.330 --> 00:44:43.450 align:middle line:84% And the answer is, low hanging fruit. 00:44:43.450 --> 00:44:46.340 align:middle line:84% In other words, there are some strong effects. 00:44:46.340 --> 00:44:48.760 align:middle line:84% In other words, if there is a lymphocytic infiltrate 00:44:48.760 --> 00:44:53.080 align:middle line:84% into a tissue, you'll have a lot of immunoglobulins 00:44:53.080 --> 00:44:54.880 align:middle line:90% present as opposed to not. 00:44:54.880 --> 00:44:57.460 align:middle line:84% So if you stain for immunoglobulins, 00:44:57.460 --> 00:44:59.470 align:middle line:90% you'll see it light up. 00:44:59.470 --> 00:45:02.470 align:middle line:84% If you are looking for genes that-- if you're 00:45:02.470 --> 00:45:05.110 align:middle line:84% going to do an in situ hybridization 00:45:05.110 --> 00:45:10.180 align:middle line:84% for an apoptotic gene, those samples 00:45:10.180 --> 00:45:13.690 align:middle line:84% that have more apoptosis in them will light up more. 00:45:13.690 --> 00:45:17.260 align:middle line:84% Now, the way you handle them is, without a doubt, 00:45:17.260 --> 00:45:18.620 align:middle line:90% going to influence it. 00:45:18.620 --> 00:45:20.500 align:middle line:84% Which part of a tumor of this piece of salami 00:45:20.500 --> 00:45:25.550 align:middle line:84% was originally sliced out will also have an effect. 00:45:25.550 --> 00:45:27.190 align:middle line:90% But if-- 00:45:27.190 --> 00:45:30.970 align:middle line:84% AUDIENCE: That was less what I was getting at. 00:45:30.970 --> 00:45:33.160 align:middle line:90% Let's take like a fibroblast. 00:45:33.160 --> 00:45:34.720 align:middle line:90% ISAAC SAMUEL KOHANE: Yeah. 00:45:34.720 --> 00:45:36.456 align:middle line:84% AUDIENCE: The degree to which you-- 00:45:36.456 --> 00:45:40.150 align:middle line:84% I mean, obviously, that isn't standardized in how 00:45:40.150 --> 00:45:41.380 align:middle line:90% you make these microarrays. 00:45:41.380 --> 00:45:42.390 align:middle line:84% ISAAC SAMUEL KOHANE: Right, go ahead. 00:45:42.390 --> 00:45:44.110 align:middle line:84% AUDIENCE: The degree to which you 00:45:44.110 --> 00:45:47.560 align:middle line:84% have confluent layer will dictate 00:45:47.560 --> 00:45:51.235 align:middle line:84% the morphology of the cell, just like under visual inspection. 00:45:51.235 --> 00:45:52.360 align:middle line:90% ISAAC SAMUEL KOHANE: I see. 00:45:52.360 --> 00:45:53.652 align:middle line:90% I see where you're coming from. 00:45:53.652 --> 00:45:54.580 align:middle line:90% These are not cells. 00:45:54.580 --> 00:45:56.500 align:middle line:90% This is not cell culture. 00:45:56.500 --> 00:46:01.360 align:middle line:84% This is a hunk of tissue that was taken out of a patient, 00:46:01.360 --> 00:46:07.600 align:middle line:84% fixed, either through freezing or through some other process, 00:46:07.600 --> 00:46:10.360 align:middle line:90% and laid down onto that slide. 00:46:10.360 --> 00:46:12.700 align:middle line:84% It's not a tissue culture experiment. 00:46:12.700 --> 00:46:14.528 align:middle line:90% AUDIENCE: OK. 00:46:14.528 --> 00:46:16.570 align:middle line:84% ISAAC SAMUEL KOHANE: That's a very good question. 00:46:16.570 --> 00:46:19.090 align:middle line:90% 00:46:19.090 --> 00:46:23.280 align:middle line:84% And that speaks to the next generation of microarrays, 00:46:23.280 --> 00:46:25.780 align:middle line:84% which I'm not willing to speak about, which is when you have 00:46:25.780 --> 00:46:28.570 align:middle line:84% a confluent layer of living cells, 00:46:28.570 --> 00:46:33.213 align:middle line:84% as opposed to this fixed tissue that you're going to stain, 00:46:33.213 --> 00:46:34.630 align:middle line:84% that's another type of microarray. 00:46:34.630 --> 00:46:38.650 align:middle line:84% You can then, actually, for instance, target drugs, 00:46:38.650 --> 00:46:40.900 align:middle line:84% different drugs at different points in this microarray 00:46:40.900 --> 00:46:42.910 align:middle line:90% and see how these cells react. 00:46:42.910 --> 00:46:48.130 align:middle line:84% But god knows what that means in a set of fibroblasts 00:46:48.130 --> 00:46:49.480 align:middle line:90% spread as a monolayer. 00:46:49.480 --> 00:46:52.167 align:middle line:90% 00:46:52.167 --> 00:46:53.500 align:middle line:90% I think it's a big open subject. 00:46:53.500 --> 00:46:54.042 align:middle line:90% It's a very-- 00:46:54.042 --> 00:46:55.928 align:middle line:84% AUDIENCE: I think I'd laugh if somebody 00:46:55.928 --> 00:46:57.970 align:middle line:84% they tried to make any conclusion about something 00:46:57.970 --> 00:47:00.517 align:middle line:84% like that by entering individual cells into a layer. 00:47:00.517 --> 00:47:03.100 align:middle line:84% ISAAC SAMUEL KOHANE: Well, I'm glad you're laughing, because-- 00:47:03.100 --> 00:47:04.142 align:middle line:90% AUDIENCE: The structure-- 00:47:04.142 --> 00:47:05.642 align:middle line:84% ISAAC SAMUEL KOHANE: I'm glad you're 00:47:05.642 --> 00:47:07.240 align:middle line:84% laughing, because I've been crying 00:47:07.240 --> 00:47:10.330 align:middle line:84% about the way these much more boring microarrays have 00:47:10.330 --> 00:47:11.585 align:middle line:90% been used in the past. 00:47:11.585 --> 00:47:13.960 align:middle line:84% And I'm going to start to give you the bad news, shortly. 00:47:13.960 --> 00:47:16.570 align:middle line:84% But I just want to tell you that people are publishing, 00:47:16.570 --> 00:47:19.480 align:middle line:84% in first class journals, the kind of thing I'm just 00:47:19.480 --> 00:47:20.500 align:middle line:90% describing now. 00:47:20.500 --> 00:47:22.190 align:middle line:84% And some of the luminaries in our field. 00:47:22.190 --> 00:47:23.898 align:middle line:84% And so what you have to understand, guys, 00:47:23.898 --> 00:47:26.110 align:middle line:84% there's a lot of legacity in the field of genomics. 00:47:26.110 --> 00:47:28.330 align:middle line:84% And part of what this course aims to do 00:47:28.330 --> 00:47:31.707 align:middle line:84% is for you to learn the limitations. 00:47:31.707 --> 00:47:33.790 align:middle line:84% When Joel, for instance, talked to you about SNPs, 00:47:33.790 --> 00:47:37.690 align:middle line:84% you'll learn that half of all SNP studies are just wrong. 00:47:37.690 --> 00:47:41.050 align:middle line:84% And what I'm about to tell you is at least half 00:47:41.050 --> 00:47:42.760 align:middle line:84% of all microarray studies are wrong. 00:47:42.760 --> 00:47:45.250 align:middle line:84% But big companies are being formed around this model 00:47:45.250 --> 00:47:48.520 align:middle line:90% they're targeting with drugs-- 00:47:48.520 --> 00:47:49.550 align:middle line:90% kind of stuff, Todd. 00:47:49.550 --> 00:47:51.610 align:middle line:84% So this is an opportunity for you 00:47:51.610 --> 00:47:56.540 align:middle line:84% to make a big name for yourself in the future. 00:47:56.540 --> 00:47:59.200 align:middle line:84% In fact, so here comes the debunking. 00:47:59.200 --> 00:48:03.040 align:middle line:84% So Pliny the Elder said the following. 00:48:03.040 --> 00:48:04.540 align:middle line:84% "Indeed, what is there that does not 00:48:04.540 --> 00:48:06.457 align:middle line:84% appear fabulous when it comes to our knowledge 00:48:06.457 --> 00:48:07.280 align:middle line:90% for the first time? 00:48:07.280 --> 00:48:09.970 align:middle line:84% How many things, too, are looked upon as quite impossible 00:48:09.970 --> 00:48:11.750 align:middle line:84% until they have been actually affected." 00:48:11.750 --> 00:48:13.930 align:middle line:84% So this is a very optimistic statement 00:48:13.930 --> 00:48:17.180 align:middle line:84% about the future of science, technology, engineering. 00:48:17.180 --> 00:48:18.737 align:middle line:90% Build it, they will come. 00:48:18.737 --> 00:48:20.320 align:middle line:84% And this is Pliny the Elder, who wrote 00:48:20.320 --> 00:48:23.050 align:middle line:84% this big tome called Natural History, which 00:48:23.050 --> 00:48:24.670 align:middle line:90% was very well regarded. 00:48:24.670 --> 00:48:28.300 align:middle line:90% He's a Roman patriarch. 00:48:28.300 --> 00:48:30.160 align:middle line:90% And note when he died-- 00:48:30.160 --> 00:48:31.420 align:middle line:90% 79 AD. 00:48:31.420 --> 00:48:33.970 align:middle line:84% And that there is significance to that date of death. 00:48:33.970 --> 00:48:36.520 align:middle line:84% Because what he was was, in fact, not 00:48:36.520 --> 00:48:39.730 align:middle line:84% a true experimentalist scientist, 00:48:39.730 --> 00:48:42.310 align:middle line:84% using the scientific method as we currently understand it. 00:48:42.310 --> 00:48:47.230 align:middle line:84% He was more of a description-based person, 00:48:47.230 --> 00:48:50.110 align:middle line:84% describing what he sees in incredible detail, 00:48:50.110 --> 00:48:52.600 align:middle line:84% annotating meticulously, taxonomizing it. 00:48:52.600 --> 00:48:55.360 align:middle line:84% But not really putting forward any 00:48:55.360 --> 00:48:58.365 align:middle line:84% hypotheses about how things work. 00:48:58.365 --> 00:48:59.740 align:middle line:84% And because he did not understand 00:48:59.740 --> 00:49:02.110 align:middle line:84% the mechanism of things, when he went 00:49:02.110 --> 00:49:05.530 align:middle line:84% to visit his last scientific investigation, which 00:49:05.530 --> 00:49:09.010 align:middle line:84% was look at a local volcano, which 00:49:09.010 --> 00:49:10.390 align:middle line:90% happened to be called Vesuvius. 00:49:10.390 --> 00:49:14.650 align:middle line:84% When he went to the boat to visit Vesuvius, 00:49:14.650 --> 00:49:16.460 align:middle line:84% he misunderstood its basic nature. 00:49:16.460 --> 00:49:19.770 align:middle line:84% So, in fact, he died, as his son was 00:49:19.770 --> 00:49:21.810 align:middle line:84% trying to drag him from the island, 00:49:21.810 --> 00:49:25.530 align:middle line:84% probably from asphyxia from the vapors from this volcano. 00:49:25.530 --> 00:49:29.580 align:middle line:84% And consequently, I ask the question, 00:49:29.580 --> 00:49:33.250 align:middle line:84% are we in danger of an imminent genomic Plinian eruption. 00:49:33.250 --> 00:49:35.790 align:middle line:84% Now Plinian eruption is actually a term 00:49:35.790 --> 00:49:40.560 align:middle line:84% of the trade used by geologists to describe a massive explosion 00:49:40.560 --> 00:49:42.750 align:middle line:84% that results in a lot of hot air and ash 00:49:42.750 --> 00:49:44.670 align:middle line:90% going up into the stratosphere. 00:49:44.670 --> 00:49:49.170 align:middle line:84% And I think we can argue that we're in similar danger because 00:49:49.170 --> 00:49:52.020 align:middle line:84% of over-promising, the slowness to acknowledge 00:49:52.020 --> 00:49:54.150 align:middle line:84% the limitations of our measurement techniques, 00:49:54.150 --> 00:49:55.920 align:middle line:84% the challenge of linking genomic data 00:49:55.920 --> 00:49:57.900 align:middle line:84% to biological and clinical significance, 00:49:57.900 --> 00:50:00.630 align:middle line:84% and a lack of formal hypothesis testing, 00:50:00.630 --> 00:50:06.780 align:middle line:84% and the lack of sufficient multidisciplinary expertise. 00:50:06.780 --> 00:50:10.080 align:middle line:84% For all these reasons, the fact that Todd might roll his eyes 00:50:10.080 --> 00:50:12.260 align:middle line:84% at this tissue culture experiment 00:50:12.260 --> 00:50:15.810 align:middle line:84% doesn't mean that probably all VCs will think it's great. 00:50:15.810 --> 00:50:19.120 align:middle line:84% Because they won't have that multidisciplinary expertise. 00:50:19.120 --> 00:50:21.810 align:middle line:84% And so are functional genomicists over-promising? 00:50:21.810 --> 00:50:24.030 align:middle line:90% Absolutely, yes. 00:50:24.030 --> 00:50:25.290 align:middle line:90% So let's go through it. 00:50:25.290 --> 00:50:29.130 align:middle line:84% Let's go back to our world's smallest microarray, 00:50:29.130 --> 00:50:32.610 align:middle line:84% three genes measured before and after an intervention. 00:50:32.610 --> 00:50:36.060 align:middle line:84% And in fact, the content of Nature Genetics 00:50:36.060 --> 00:50:39.930 align:middle line:84% science journal articles about microarrays, 00:50:39.930 --> 00:50:43.200 align:middle line:84% in the first four years of functional genomics, 00:50:43.200 --> 00:50:45.780 align:middle line:84% of lots of publications in functional genomics, which 00:50:45.780 --> 00:50:49.290 align:middle line:84% I think of as 1999 to 2003, so plenty 00:50:49.290 --> 00:50:51.300 align:middle line:84% of papers which did the following sophisticated 00:50:51.300 --> 00:50:52.200 align:middle line:90% calculation. 00:50:52.200 --> 00:50:55.260 align:middle line:84% You take each gene before and after intervention. 00:50:55.260 --> 00:50:56.340 align:middle line:90% Take the ratio. 00:50:56.340 --> 00:50:57.660 align:middle line:90% And you report the ratios. 00:50:57.660 --> 00:51:00.160 align:middle line:84% And you sort them from high to low. 00:51:00.160 --> 00:51:02.982 align:middle line:84% So that, in fact, gene 2, which is 2.1/0.3, 00:51:02.982 --> 00:51:04.740 align:middle line:84% that gives you a ratio of 7, therefore, 00:51:04.740 --> 00:51:06.840 align:middle line:90% it's the most upregulated gene. 00:51:06.840 --> 00:51:08.770 align:middle line:84% It's lists like that that you'd get. 00:51:08.770 --> 00:51:11.310 align:middle line:84% So for instance-- and I feel bad for this poor guy, 00:51:11.310 --> 00:51:15.810 align:middle line:84% because I always cite his article, because I just 00:51:15.810 --> 00:51:17.250 align:middle line:90% don't want to change my slides. 00:51:17.250 --> 00:51:18.640 align:middle line:90% But this is a good point. 00:51:18.640 --> 00:51:22.393 align:middle line:84% So this group, out of Yale, did the following interesting 00:51:22.393 --> 00:51:24.060 align:middle line:84% experiment, which you should think about 00:51:24.060 --> 00:51:25.200 align:middle line:90% as you have your next meal. 00:51:25.200 --> 00:51:27.690 align:middle line:84% Rats, who are starved, live twice as long 00:51:27.690 --> 00:51:29.700 align:middle line:90% as unstarved rats. 00:51:29.700 --> 00:51:35.280 align:middle line:84% And so what they did is they starved these rats 00:51:35.280 --> 00:51:37.320 align:middle line:84% and compared them to non-starved rats 00:51:37.320 --> 00:51:42.390 align:middle line:84% and looked at the expression profile of the muscle tissue. 00:51:42.390 --> 00:51:44.522 align:middle line:84% And what they saw, using Affymetrix technology, 00:51:44.522 --> 00:51:45.480 align:middle line:90% was pretty interesting. 00:51:45.480 --> 00:51:47.938 align:middle line:84% There were a bunch of genes that are involved in mopping up 00:51:47.938 --> 00:51:49.830 align:middle line:84% free radicals, these bad things that 00:51:49.830 --> 00:51:53.250 align:middle line:84% screw up the structure of your proteins and DNA. 00:51:53.250 --> 00:51:54.810 align:middle line:84% A bunch of genes that are involved 00:51:54.810 --> 00:52:00.030 align:middle line:84% in mopping up free radicals and combating the oxidative stress 00:52:00.030 --> 00:52:04.662 align:middle line:84% were downregulated in these starved mice. 00:52:04.662 --> 00:52:06.870 align:middle line:84% I want to point out that they are, for instance, down 00:52:06.870 --> 00:52:09.930 align:middle line:90% by 1.5 fold, down by 1.6 fold. 00:52:09.930 --> 00:52:12.848 align:middle line:90% Remember those numbers. 00:52:12.848 --> 00:52:14.140 align:middle line:90% Well, I won't be cute about it. 00:52:14.140 --> 00:52:18.120 align:middle line:84% The fact is those fold changes, especially when they 00:52:18.120 --> 00:52:20.430 align:middle line:84% had a limited number of the replicates, 00:52:20.430 --> 00:52:25.320 align:middle line:84% like three replicates, are just not sustainable 00:52:25.320 --> 00:52:29.790 align:middle line:84% in any kind of analysis, even today, 00:52:29.790 --> 00:52:31.320 align:middle line:84% unless you do a lot more replicates 00:52:31.320 --> 00:52:32.760 align:middle line:90% and use much better platforms. 00:52:32.760 --> 00:52:36.090 align:middle line:84% So here, for example, is a study that we 00:52:36.090 --> 00:52:39.670 align:middle line:90% did using a Incyte chip. 00:52:39.670 --> 00:52:42.540 align:middle line:84% Incyte used to make a spotted microarray. 00:52:42.540 --> 00:52:45.180 align:middle line:84% They're one of the technologies that 00:52:45.180 --> 00:52:49.140 align:middle line:84% went away with the genomic implosion or meltdown. 00:52:49.140 --> 00:52:51.390 align:middle line:84% And what you see, here, I've taken the microarray 00:52:51.390 --> 00:52:53.640 align:middle line:84% and I've strung out, into a line, 00:52:53.640 --> 00:52:56.130 align:middle line:90% into a vector, the 8,000 genes. 00:52:56.130 --> 00:53:02.250 align:middle line:84% And I'm showing you the ratio of these genes 00:53:02.250 --> 00:53:05.280 align:middle line:84% across two conditions in this case cardiac muscle of a mouse 00:53:05.280 --> 00:53:07.710 align:middle line:90% before and after cocaine. 00:53:07.710 --> 00:53:10.650 align:middle line:84% And what you see is that most of the genes have a ratio of 1. 00:53:10.650 --> 00:53:13.200 align:middle line:84% And some are a little bit above and some a little bit below. 00:53:13.200 --> 00:53:15.930 align:middle line:90% 00:53:15.930 --> 00:53:17.490 align:middle line:90% What to make of it? 00:53:17.490 --> 00:53:19.560 align:middle line:90% Where do we draw the line? 00:53:19.560 --> 00:53:21.630 align:middle line:90% What is upregulated? 00:53:21.630 --> 00:53:25.170 align:middle line:84% Is this 1.4 fold upregulated and above, 00:53:25.170 --> 00:53:27.347 align:middle line:90% is that the significance level? 00:53:27.347 --> 00:53:28.680 align:middle line:90% Well, I did the following thing. 00:53:28.680 --> 00:53:34.710 align:middle line:84% I applied a Fourier transform to this data, where 00:53:34.710 --> 00:53:37.073 align:middle line:84% the dimensional cross, across which 00:53:37.073 --> 00:53:38.490 align:middle line:84% I was doing the Fourier transform, 00:53:38.490 --> 00:53:42.930 align:middle line:84% was that of the linear position on the microarray. 00:53:42.930 --> 00:53:45.210 align:middle line:84% And as you know, the Fourier transform 00:53:45.210 --> 00:53:50.280 align:middle line:84% identifies elements which have periodicity. 00:53:50.280 --> 00:53:52.830 align:middle line:84% And what I found was the following periodicity-- 00:53:52.830 --> 00:53:55.620 align:middle line:84% a periodicity of 4, a periodicity of 9, 00:53:55.620 --> 00:53:58.830 align:middle line:84% and some other periodicities, but a huge periodicity at 4. 00:53:58.830 --> 00:54:03.010 align:middle line:84% Anybody have any idea why there was this periodicity of 4 00:54:03.010 --> 00:54:03.510 align:middle line:90% in the data? 00:54:03.510 --> 00:54:07.530 align:middle line:84% Remember, regardless of what gene it was, across this chip, 00:54:07.530 --> 00:54:10.210 align:middle line:90% there's a periodicity of four. 00:54:10.210 --> 00:54:11.738 align:middle line:90% AUDIENCE: Chip. 00:54:11.738 --> 00:54:13.030 align:middle line:90% ISAAC SAMUEL KOHANE: Chip what? 00:54:13.030 --> 00:54:14.290 align:middle line:90% AUDIENCE: Chip [INAUDIBLE]. 00:54:14.290 --> 00:54:15.550 align:middle line:90% The printing? 00:54:15.550 --> 00:54:17.560 align:middle line:84% ISAAC SAMUEL KOHANE: Print chip, what about it? 00:54:17.560 --> 00:54:20.140 align:middle line:84% AUDIENCE: Maybe the form of the-- or in the tip, 00:54:20.140 --> 00:54:21.490 align:middle line:90% especially with the-- 00:54:21.490 --> 00:54:23.240 align:middle line:84% ISAAC SAMUEL KOHANE: You're getting close. 00:54:23.240 --> 00:54:25.630 align:middle line:84% AUDIENCE: [INAUDIBLE] maybe, like, 00:54:25.630 --> 00:54:27.760 align:middle line:90% could there be any element-- 00:54:27.760 --> 00:54:29.470 align:middle line:90% ISAAC SAMUEL KOHANE: You got it. 00:54:29.470 --> 00:54:31.815 align:middle line:84% This is what the inside chip looks like. 00:54:31.815 --> 00:54:33.190 align:middle line:84% And the reason it looks like that 00:54:33.190 --> 00:54:35.650 align:middle line:90% is because they used four pins. 00:54:35.650 --> 00:54:38.292 align:middle line:84% Each pin took care of a different quadrant. 00:54:38.292 --> 00:54:40.750 align:middle line:84% And so if there's even slightly different physical chemical 00:54:40.750 --> 00:54:44.740 align:middle line:84% properties of each pin, in the engineering of the pin, 00:54:44.740 --> 00:54:48.130 align:middle line:84% you would get a slight systematic change 00:54:48.130 --> 00:54:49.660 align:middle line:84% in the amount of probe laid down. 00:54:49.660 --> 00:54:52.390 align:middle line:84% Consequently, it turned out that you 00:54:52.390 --> 00:54:57.700 align:middle line:84% got on this chip much more than 1.5 fold changes in the readout 00:54:57.700 --> 00:55:00.610 align:middle line:84% just based on which quadrant you were in. 00:55:00.610 --> 00:55:03.190 align:middle line:84% And that was just based on the physicality. 00:55:03.190 --> 00:55:05.290 align:middle line:84% Now Affymetrix, in the past has not 00:55:05.290 --> 00:55:07.420 align:middle line:90% been guilt-free in this matter. 00:55:07.420 --> 00:55:10.510 align:middle line:84% For instance, depending on which way the hybridization solution 00:55:10.510 --> 00:55:12.430 align:middle line:84% was washed on, you can see different-- you 00:55:12.430 --> 00:55:15.710 align:middle line:84% can see intensity effects going this way or that way. 00:55:15.710 --> 00:55:18.520 align:middle line:84% And if there was a thermal gradient in the hybridization 00:55:18.520 --> 00:55:22.390 align:middle line:84% chamber, you could also see different intensities 00:55:22.390 --> 00:55:26.410 align:middle line:84% that were chip geometry dependent and independent 00:55:26.410 --> 00:55:27.260 align:middle line:90% of the gene. 00:55:27.260 --> 00:55:27.760 align:middle line:90% Yes? 00:55:27.760 --> 00:55:31.243 align:middle line:84% AUDIENCE: What did you think to look for [INAUDIBLE]?? 00:55:31.243 --> 00:55:33.910 align:middle line:84% ISAAC SAMUEL KOHANE: Because, of course, I knew those four pins. 00:55:33.910 --> 00:55:35.410 align:middle line:84% I said, I wonder if there's anything 00:55:35.410 --> 00:55:36.710 align:middle line:90% different about the pins. 00:55:36.710 --> 00:55:39.130 align:middle line:90% By the way, the second-- 00:55:39.130 --> 00:55:42.760 align:middle line:84% and now, by the way, I almost do that routinely. 00:55:42.760 --> 00:55:44.890 align:middle line:84% Any series I do, as a sanity check, 00:55:44.890 --> 00:55:47.410 align:middle line:84% I just run a Fourier transform to see if there's regularity. 00:55:47.410 --> 00:55:49.270 align:middle line:84% The other periodicity was due to how often 00:55:49.270 --> 00:55:50.830 align:middle line:90% they replaced the printhead. 00:55:50.830 --> 00:55:54.640 align:middle line:90% 00:55:54.640 --> 00:55:57.610 align:middle line:84% But of course, the manufacturer would never tell you this. 00:55:57.610 --> 00:55:59.610 align:middle line:84% Let's make things a little bit more interesting. 00:55:59.610 --> 00:56:02.920 align:middle line:84% So this is an experiment that Atul did with Morris White 00:56:02.920 --> 00:56:07.600 align:middle line:84% at the Joslin, using a now obsolete set of chips. 00:56:07.600 --> 00:56:10.990 align:middle line:84% They had four patients with glucose intolerance. 00:56:10.990 --> 00:56:14.800 align:middle line:84% So it's not quite diabetes but it's the inability 00:56:14.800 --> 00:56:17.530 align:middle line:84% to bring down your blood sugar rapidly enough after you give 00:56:17.530 --> 00:56:19.600 align:middle line:90% him a challenge of glucose. 00:56:19.600 --> 00:56:21.790 align:middle line:84% And using a collection of three chips, 00:56:21.790 --> 00:56:24.010 align:middle line:90% you measured 35,000 genes. 00:56:24.010 --> 00:56:28.000 align:middle line:84% So we did it on patient one, patient two, patient three 00:56:28.000 --> 00:56:29.230 align:middle line:90% and patient four. 00:56:29.230 --> 00:56:32.350 align:middle line:84% And we repeated the same experiment 00:56:32.350 --> 00:56:36.895 align:middle line:84% with the same extracts of RNA from the patient's muscles. 00:56:36.895 --> 00:56:38.590 align:middle line:84% These were from the patient's muscles. 00:56:38.590 --> 00:56:41.020 align:middle line:84% So let me ask you the following question. 00:56:41.020 --> 00:56:43.030 align:middle line:84% For gene five on this microarray, 00:56:43.030 --> 00:56:46.430 align:middle line:84% what should the ratio be between gene five, here, and gene five, 00:56:46.430 --> 00:56:46.930 align:middle line:90% here? 00:56:46.930 --> 00:56:47.430 align:middle line:90% AUDIENCE: 1. 00:56:47.430 --> 00:56:50.025 align:middle line:90% 00:56:50.025 --> 00:56:51.650 align:middle line:84% ISAAC SAMUEL KOHANE: Did someone say 1? 00:56:51.650 --> 00:56:52.620 align:middle line:90% Thank you. 00:56:52.620 --> 00:56:54.020 align:middle line:90% The answer is 1. 00:56:54.020 --> 00:56:57.300 align:middle line:84% What should the ratio be between gene five, here, and gene 5, 00:56:57.300 --> 00:56:57.800 align:middle line:90% here? 00:56:57.800 --> 00:57:00.710 align:middle line:90% 00:57:00.710 --> 00:57:01.970 align:middle line:90% Who knows? 00:57:01.970 --> 00:57:02.720 align:middle line:90% It shouldn't be 1. 00:57:02.720 --> 00:57:04.280 align:middle line:84% The answer is, who knows, because it depends 00:57:04.280 --> 00:57:05.330 align:middle line:90% on the individual, right? 00:57:05.330 --> 00:57:07.100 align:middle line:90% They are different individuals. 00:57:07.100 --> 00:57:08.480 align:middle line:90% Good. 00:57:08.480 --> 00:57:13.730 align:middle line:84% So now that we understand that, let's repeat. 00:57:13.730 --> 00:57:19.040 align:middle line:84% The intrapatient variation should be 1. 00:57:19.040 --> 00:57:21.475 align:middle line:84% The interpatient variation, we don't know. 00:57:21.475 --> 00:57:23.600 align:middle line:84% But let's say this ratio, for the sake of argument, 00:57:23.600 --> 00:57:25.600 align:middle line:84% between gene five in patient one and patient two 00:57:25.600 --> 00:57:27.950 align:middle line:90% is a ratio of 5, OK? 00:57:27.950 --> 00:57:31.245 align:middle line:84% What should this ratio be of gene five from patient one? 00:57:31.245 --> 00:57:32.120 align:middle line:90% AUDIENCE: Close to 5. 00:57:32.120 --> 00:57:33.090 align:middle line:84% ISAAC SAMUEL KOHANE: Close to 5 as well. 00:57:33.090 --> 00:57:35.340 align:middle line:84% So what should the ratio of the ratios be? 00:57:35.340 --> 00:57:35.840 align:middle line:90% AUDIENCE: 1. 00:57:35.840 --> 00:57:37.257 align:middle line:84% ISAAC SAMUEL KOHANE: 1, thank you. 00:57:37.257 --> 00:57:42.860 align:middle line:84% So to repeat, the intrapatient variation ratio should be 1. 00:57:42.860 --> 00:57:46.460 align:middle line:84% The ratio of the ratios of the interpatient variation 00:57:46.460 --> 00:57:47.415 align:middle line:90% should be 1. 00:57:47.415 --> 00:57:47.915 align:middle line:90% Yes? 00:57:47.915 --> 00:57:50.290 align:middle line:84% AUDIENCE: So the measurement's actually were pretty close 00:57:50.290 --> 00:57:51.410 align:middle line:90% in time and all that? 00:57:51.410 --> 00:57:52.850 align:middle line:84% ISAAC SAMUEL KOHANE: Well, it's actually the same RNA sample 00:57:52.850 --> 00:57:53.392 align:middle line:90% that we took. 00:57:53.392 --> 00:57:57.390 align:middle line:84% And then we hybridized it all at the same time. 00:57:57.390 --> 00:57:59.755 align:middle line:90% But all good questions to ask. 00:57:59.755 --> 00:58:01.130 align:middle line:84% In fact, never reported typically 00:58:01.130 --> 00:58:02.285 align:middle line:90% in the journal article. 00:58:02.285 --> 00:58:05.606 align:middle line:90% 00:58:05.606 --> 00:58:08.320 align:middle line:84% In fact, so now I have to give you an anecdote. 00:58:08.320 --> 00:58:10.960 align:middle line:84% You'll learn about clustering in the lecture later 00:58:10.960 --> 00:58:14.000 align:middle line:90% by Steve Greenberg. 00:58:14.000 --> 00:58:17.350 align:middle line:84% But Todd Golub did one of the first studies 00:58:17.350 --> 00:58:19.180 align:middle line:84% of the differences that you can find 00:58:19.180 --> 00:58:22.677 align:middle line:84% in two physiological conditions or pathologcal conditions 00:58:22.677 --> 00:58:23.510 align:middle line:90% based on expression. 00:58:23.510 --> 00:58:26.380 align:middle line:84% And he looked at Acute Myelogenous Leukemia 00:58:26.380 --> 00:58:29.500 align:middle line:84% versus Acute Lymphoblastic Leukemia, AML versus ALL. 00:58:29.500 --> 00:58:31.750 align:middle line:84% And sure enough, using a supervised learning algorithm 00:58:31.750 --> 00:58:36.640 align:middle line:84% that you'll learn about, he could do it without a problem. 00:58:36.640 --> 00:58:38.650 align:middle line:84% He also had a test set and a training set. 00:58:38.650 --> 00:58:40.240 align:middle line:84% When I put all that data together, 00:58:40.240 --> 00:58:43.540 align:middle line:84% I found out I could actually distinguish another group, not 00:58:43.540 --> 00:58:47.320 align:middle line:84% AML versus ALL, but a test set versus the training set. 00:58:47.320 --> 00:58:49.317 align:middle line:84% And I said to myself, what the heck is this? 00:58:49.317 --> 00:58:50.650 align:middle line:90% And I went to talk back to Todd. 00:58:50.650 --> 00:58:52.210 align:middle line:84% Because, after all, Todd's one my former interns, 00:58:52.210 --> 00:58:53.750 align:middle line:84% so I have no problem calling him up. 00:58:53.750 --> 00:58:54.792 align:middle line:90% I said, Todd, what gives? 00:58:54.792 --> 00:58:57.397 align:middle line:84% He says, they were hybridized on different days. 00:58:57.397 --> 00:58:59.230 align:middle line:84% And that wasn't published in the literature. 00:58:59.230 --> 00:59:02.020 align:middle line:84% But just doing it a different day-- same technician 00:59:02.020 --> 00:59:04.490 align:middle line:84% but different days, had slightly different hybridization 00:59:04.490 --> 00:59:04.990 align:middle line:90% reaction. 00:59:04.990 --> 00:59:07.810 align:middle line:84% And that could be picked up just by looking at the data. 00:59:07.810 --> 00:59:10.660 align:middle line:84% Anyway, to remember where we were here. 00:59:10.660 --> 00:59:12.430 align:middle line:90% The intrapatient ratio-- 00:59:12.430 --> 00:59:13.180 align:middle line:90% 1. 00:59:13.180 --> 00:59:15.910 align:middle line:84% The ratio of the interpatient ratios-- 00:59:15.910 --> 00:59:16.630 align:middle line:90% 1. 00:59:16.630 --> 00:59:18.560 align:middle line:90% What in fact happens? 00:59:18.560 --> 00:59:21.430 align:middle line:84% Well, this is actually kind of crummy by 2004 standards, 00:59:21.430 --> 00:59:22.960 align:middle line:84% but it was not bad when we did it. 00:59:22.960 --> 00:59:25.910 align:middle line:84% Here is one chip, the other chip, one chip, the other chip, 00:59:25.910 --> 00:59:27.910 align:middle line:84% one chip, the other chip, for the four patients. 00:59:27.910 --> 00:59:30.160 align:middle line:84% And you see, here, what looks like a 1-to-1 line, 00:59:30.160 --> 00:59:31.180 align:middle line:90% but it's kind of muddy. 00:59:31.180 --> 00:59:35.260 align:middle line:84% And the correlation coefficients were 0.76 to 0.84. 00:59:35.260 --> 00:59:39.250 align:middle line:84% Currently, anybody who does more poorly than a correlation 00:59:39.250 --> 00:59:43.720 align:middle line:84% coefficient of 0.97, I question how good their hybridization 00:59:43.720 --> 00:59:45.315 align:middle line:90% laboratory is. 00:59:45.315 --> 00:59:46.690 align:middle line:84% Nonetheless, that's what we have. 00:59:46.690 --> 00:59:51.140 align:middle line:84% But you still can bleed it to a good 1-to-1 ratio. 00:59:51.140 --> 00:59:57.280 align:middle line:84% However, what was the ratio of the ratios? 00:59:57.280 --> 01:00:00.640 align:middle line:84% Now, here's the ratio of the ratios in one 01:00:00.640 --> 01:00:02.450 align:middle line:90% case versus the other case. 01:00:02.450 --> 01:00:04.210 align:middle line:84% And you don't have to be a statistician 01:00:04.210 --> 01:00:06.310 align:middle line:84% to realize that there's no signal here. 01:00:06.310 --> 01:00:08.590 align:middle line:90% These are blobs. 01:00:08.590 --> 01:00:10.670 align:middle line:90% Now what does a blob mean? 01:00:10.670 --> 01:00:14.080 align:middle line:84% It means that, with one set of chips, 01:00:14.080 --> 01:00:16.600 align:middle line:84% this gene was 10 times higher, because of the log scale, 01:00:16.600 --> 01:00:18.075 align:middle line:84% in patient one than patient three. 01:00:18.075 --> 01:00:20.200 align:middle line:84% With the other set of chips, this gene was 10 times 01:00:20.200 --> 01:00:22.540 align:middle line:84% lower in patient one than patient three. 01:00:22.540 --> 01:00:25.180 align:middle line:90% So this is particularly heinous. 01:00:25.180 --> 01:00:27.760 align:middle line:84% We have not only the wrong magnitude, 01:00:27.760 --> 01:00:30.850 align:middle line:84% we have the wrong direction of regulation. 01:00:30.850 --> 01:00:33.860 align:middle line:84% And by inspection, you know this is happening a lot. 01:00:33.860 --> 01:00:36.320 align:middle line:84% So right away, you should be asking yourself a question, 01:00:36.320 --> 01:00:36.987 align:middle line:90% how can this be? 01:00:36.987 --> 01:00:39.612 align:middle line:84% I mean, after all Zak's told me about how these microarrays are 01:00:39.612 --> 01:00:40.270 align:middle line:90% wonderful. 01:00:40.270 --> 01:00:42.320 align:middle line:84% How can this be that this is so bad? 01:00:42.320 --> 01:00:45.010 align:middle line:84% And after all, in the end, there is some signal here. 01:00:45.010 --> 01:00:46.960 align:middle line:90% And why is it so bad? 01:00:46.960 --> 01:00:50.920 align:middle line:90% Any idea why it is so bad? 01:00:50.920 --> 01:00:54.640 align:middle line:84% Anybody have any ideas why these results are so bad? 01:00:54.640 --> 01:00:58.524 align:middle line:84% We were not particularly worse than anybody else at that time. 01:00:58.524 --> 01:01:01.792 align:middle line:84% AUDIENCE: [INAUDIBLE] different days. 01:01:01.792 --> 01:01:03.250 align:middle line:84% ISAAC SAMUEL KOHANE: But we weren't 01:01:03.250 --> 01:01:06.860 align:middle line:84% doing that on different days, so some other kind of variation. 01:01:06.860 --> 01:01:09.220 align:middle line:90% So let me ask a lead-- go ahead. 01:01:09.220 --> 01:01:12.015 align:middle line:84% AUDIENCE: All you have, [INAUDIBLE] 01:01:12.015 --> 01:01:13.390 align:middle line:84% if you're doing ratios to ratios. 01:01:13.390 --> 01:01:13.810 align:middle line:90% ISAAC SAMUEL KOHANE: Right. 01:01:13.810 --> 01:01:14.620 align:middle line:90% AUDIENCE: It compounds it. 01:01:14.620 --> 01:01:15.730 align:middle line:84% ISAAC SAMUEL KOHANE: It does-- definitely compounds it. 01:01:15.730 --> 01:01:18.460 align:middle line:84% But which kind of measurements are the most sensitive to 01:01:18.460 --> 01:01:21.680 align:middle line:90% that kind of compounding? 01:01:21.680 --> 01:01:22.973 align:middle line:90% What kind of measurements? 01:01:22.973 --> 01:01:27.690 align:middle line:90% 01:01:27.690 --> 01:01:31.740 align:middle line:84% What, in the denominator, makes a ratio change the most? 01:01:31.740 --> 01:01:33.900 align:middle line:90% Small numbers, right? 01:01:33.900 --> 01:01:38.320 align:middle line:84% So, when you have a gene expression, 01:01:38.320 --> 01:01:41.970 align:middle line:84% which goes from 0.6 to 0.3, that's a two-fold difference. 01:01:41.970 --> 01:01:46.140 align:middle line:84% When you go from 600 to 300, that's a two-fold difference. 01:01:46.140 --> 01:01:49.470 align:middle line:84% You add a little bit noise, this one flips on the left, 01:01:49.470 --> 01:01:51.480 align:middle line:84% and that one's pretty stable on the right. 01:01:51.480 --> 01:01:52.260 align:middle line:90% Well guess what? 01:01:52.260 --> 01:01:54.390 align:middle line:84% 2/3 of transcriptome is expressed 01:01:54.390 --> 01:01:56.235 align:middle line:90% at very, very low levels. 01:01:56.235 --> 01:01:58.980 align:middle line:84% You have one or two copies of RNA. 01:01:58.980 --> 01:02:00.930 align:middle line:84% One third or less of transcriptome, 01:02:00.930 --> 01:02:03.840 align:middle line:84% we have hundreds of copies of RNA, millions 01:02:03.840 --> 01:02:05.690 align:middle line:90% of copies of RNA of that gene. 01:02:05.690 --> 01:02:08.640 align:middle line:84% And those are easily picked up by these microarrays. 01:02:08.640 --> 01:02:10.950 align:middle line:84% But they're very noisy when you get down 01:02:10.950 --> 01:02:14.620 align:middle line:84% to these very low levels because of exactly that problem. 01:02:14.620 --> 01:02:17.370 align:middle line:84% And so if you don't have a method that 01:02:17.370 --> 01:02:20.700 align:middle line:84% looks at the variance and does not 01:02:20.700 --> 01:02:23.970 align:middle line:84% take into account expression intensity, your hosed. 01:02:23.970 --> 01:02:29.552 align:middle line:84% And that's why just putting a fold cutoff that says, 01:02:29.552 --> 01:02:32.010 align:middle line:84% I'm going to look at everything that's two-fold or greater, 01:02:32.010 --> 01:02:33.510 align:middle line:84% you know that you're going to find a bunch of things 01:02:33.510 --> 01:02:36.240 align:middle line:84% at two-fold or greater at lower expression levels, 01:02:36.240 --> 01:02:38.010 align:middle line:90% but they're just wrong. 01:02:38.010 --> 01:02:42.030 align:middle line:84% And thank goodness, in 2004, you can't 01:02:42.030 --> 01:02:43.680 align:middle line:84% get a journal article now published 01:02:43.680 --> 01:02:46.200 align:middle line:84% unless you do a variance analysis. 01:02:46.200 --> 01:02:54.020 align:middle line:84% But that was not true in 1999 to 2002, at the very least. 01:02:54.020 --> 01:02:57.047 align:middle line:84% Even more worrying, so, as bioinformaticians, 01:02:57.047 --> 01:02:58.630 align:middle line:84% we were spoiled by the following fact. 01:02:58.630 --> 01:03:02.290 align:middle line:90% 01:03:02.290 --> 01:03:04.930 align:middle line:90% Is it really 1:45? 01:03:04.930 --> 01:03:06.550 align:middle line:90% and I have until when? 01:03:06.550 --> 01:03:07.990 align:middle line:90% AUDIENCE: 2:00. 01:03:07.990 --> 01:03:09.880 align:middle line:90% ISAAC SAMUEL KOHANE: All right. 01:03:09.880 --> 01:03:11.785 align:middle line:84% OK, time flies when you're having fun. 01:03:11.785 --> 01:03:14.415 align:middle line:90% 01:03:14.415 --> 01:03:15.790 align:middle line:84% We were spoiled in bioinformatics 01:03:15.790 --> 01:03:17.373 align:middle line:84% by having this international resource, 01:03:17.373 --> 01:03:19.602 align:middle line:84% called GenBank, where we put all our gene sequences 01:03:19.602 --> 01:03:22.060 align:middle line:84% into an international resource, and allowed the researchers 01:03:22.060 --> 01:03:24.268 align:middle line:84% to compare different things across different systems. 01:03:24.268 --> 01:03:27.220 align:middle line:84% And we thought we could do the same thing for microarrays. 01:03:27.220 --> 01:03:30.880 align:middle line:84% So for the 60 cancer cell lines that the National Cancer 01:03:30.880 --> 01:03:35.288 align:middle line:84% Institute collects to test the thousands of drugs 01:03:35.288 --> 01:03:37.330 align:middle line:84% that they obtain, from pharmaceutical industries, 01:03:37.330 --> 01:03:40.030 align:middle line:84% from the rainforest, from China, and so on, to test 01:03:40.030 --> 01:03:44.320 align:middle line:84% for chemotherapeutic efficacy, they have these 60 cancer cell 01:03:44.320 --> 01:03:47.360 align:middle line:84% lines that they've been keeping around for years. 01:03:47.360 --> 01:03:50.380 align:middle line:84% And so Todd's group, at the Whitehead, 01:03:50.380 --> 01:03:53.050 align:middle line:84% did the state of the art hybridization with Affymetrix. 01:03:53.050 --> 01:03:55.360 align:middle line:84% And Pat Brown's group, the best practitioners 01:03:55.360 --> 01:03:59.830 align:middle line:84% of spotted arrays, did the same analysis with spotted arrays. 01:03:59.830 --> 01:04:01.750 align:middle line:84% And shown here is something that you, again, 01:04:01.750 --> 01:04:03.700 align:middle line:84% don't have to be a statistician to understand 01:04:03.700 --> 01:04:05.700 align:middle line:84% the correlation of the genes that were in common 01:04:05.700 --> 01:04:07.210 align:middle line:90% across both platforms. 01:04:07.210 --> 01:04:10.380 align:middle line:84% And the correlation was terrible. 01:04:10.380 --> 01:04:12.060 align:middle line:90% There was almost no correlation. 01:04:12.060 --> 01:04:14.610 align:middle line:84% That meant that you could not compare the results 01:04:14.610 --> 01:04:16.930 align:middle line:90% from one platform to the other. 01:04:16.930 --> 01:04:21.682 align:middle line:84% And when I first published this result a couple of years ago-- 01:04:21.682 --> 01:04:23.640 align:middle line:84% it was actually a result of a course like this. 01:04:23.640 --> 01:04:25.500 align:middle line:84% One of the students did it as a final project. 01:04:25.500 --> 01:04:27.780 align:middle line:84% I told him, why don't you compare these two platforms? 01:04:27.780 --> 01:04:29.610 align:middle line:84% And I was convinced that he was wrong. 01:04:29.610 --> 01:04:30.570 align:middle line:90% When I looked at him, he wasn't. 01:04:30.570 --> 01:04:32.130 align:middle line:84% But I thought it might be some fluke. 01:04:32.130 --> 01:04:34.440 align:middle line:84% Since then, there's been about four or five papers doing this 01:04:34.440 --> 01:04:34.940 align:middle line:90% again. 01:04:34.940 --> 01:04:38.280 align:middle line:84% And it's getting better, but it's still pretty bad. 01:04:38.280 --> 01:04:41.310 align:middle line:84% Correlation now is up to about 0.6 01:04:41.310 --> 01:04:43.920 align:middle line:90% across the different platforms. 01:04:43.920 --> 01:04:47.640 align:middle line:84% That's a real problem, obviously. 01:04:47.640 --> 01:04:49.560 align:middle line:84% Here's one of the many reasons, subsequently, 01:04:49.560 --> 01:04:51.300 align:middle line:84% that we've discovered for this problem. 01:04:51.300 --> 01:04:58.620 align:middle line:84% Shown in orange is the ref seqs, the reference sequence 01:04:58.620 --> 01:05:01.470 align:middle line:84% that the National Center for Biotechnology Information 01:05:01.470 --> 01:05:06.390 align:middle line:84% maintains, a curated resource of what is the definitive mRNA 01:05:06.390 --> 01:05:07.530 align:middle line:90% subsequence of that gene. 01:05:07.530 --> 01:05:10.080 align:middle line:90% 01:05:10.080 --> 01:05:13.590 align:middle line:84% Shown in Black are the Affymetrix probes, 01:05:13.590 --> 01:05:16.080 align:middle line:84% which part of the genome they interrogate. 01:05:16.080 --> 01:05:18.180 align:middle line:90% Well, watch it. 01:05:18.180 --> 01:05:20.770 align:middle line:84% They seem to be falling off the edge. 01:05:20.770 --> 01:05:23.160 align:middle line:84% So if you have, for instance, a cDNA that 01:05:23.160 --> 01:05:27.360 align:middle line:84% interrogates this part and an Affymetrix 01:05:27.360 --> 01:05:29.730 align:middle line:90% that is off the ref seq. 01:05:29.730 --> 01:05:33.930 align:middle line:84% It may have, in fact, a very poor reproducibility. 01:05:33.930 --> 01:05:36.307 align:middle line:84% And we've subsequently done a study of that. 01:05:36.307 --> 01:05:37.890 align:middle line:84% We've actually looked at the position. 01:05:37.890 --> 01:05:39.265 align:middle line:84% By the way, Affymetrix previously 01:05:39.265 --> 01:05:42.300 align:middle line:84% had considered the exact sequences of these probes 01:05:42.300 --> 01:05:43.360 align:middle line:90% as proprietary. 01:05:43.360 --> 01:05:45.540 align:middle line:84% So it's only last year that they've 01:05:45.540 --> 01:05:47.970 align:middle line:84% revealed what those oligonucleotide sequences are, 01:05:47.970 --> 01:05:50.400 align:middle line:84% so that we can actually position them 01:05:50.400 --> 01:05:52.980 align:middle line:84% on the gene to know where, in fact, they're 01:05:52.980 --> 01:05:54.000 align:middle line:90% interrogating the gene. 01:05:54.000 --> 01:05:57.690 align:middle line:84% And consequently, when we eliminated those Affymetrix 01:05:57.690 --> 01:06:00.870 align:middle line:84% probes which fell off the gene, we 01:06:00.870 --> 01:06:03.180 align:middle line:84% had much better correlation across platforms 01:06:03.180 --> 01:06:04.148 align:middle line:90% than previously. 01:06:04.148 --> 01:06:05.940 align:middle line:84% There's other reasons, but that's, I think, 01:06:05.940 --> 01:06:06.750 align:middle line:90% a major reason. 01:06:06.750 --> 01:06:11.850 align:middle line:84% AUDIENCE: Maybe try to find some correlations [INAUDIBLE]?? 01:06:11.850 --> 01:06:13.230 align:middle line:90% ISAAC SAMUEL KOHANE: No. 01:06:13.230 --> 01:06:16.830 align:middle line:84% So here is a study, that was done by one of our fellows, 01:06:16.830 --> 01:06:19.950 align:middle line:84% of Affymetrix going from HuGeneFL, which 01:06:19.950 --> 01:06:26.380 align:middle line:90% is one generation, so to U95A. 01:06:26.380 --> 01:06:31.530 align:middle line:84% So this is looking at human RNA for the same genes. 01:06:31.530 --> 01:06:34.740 align:middle line:84% We took the same RNA, from the same muscle, 01:06:34.740 --> 01:06:37.320 align:middle line:84% and hybridized it on the same day. 01:06:37.320 --> 01:06:38.970 align:middle line:84% Looking at correlation coefficients, 01:06:38.970 --> 01:06:42.030 align:middle line:90% we got 0.7 and 0.59. 01:06:42.030 --> 01:06:43.350 align:middle line:90% Why do you think that is? 01:06:43.350 --> 01:06:46.540 align:middle line:90% 01:06:46.540 --> 01:06:51.395 align:middle line:84% AUDIENCE: [INAUDIBLE] Again, you might be looking at different-- 01:06:51.395 --> 01:06:53.020 align:middle line:84% ISAAC SAMUEL KOHANE: The answer as they 01:06:53.020 --> 01:06:54.277 align:middle line:90% picked different subsets. 01:06:54.277 --> 01:06:56.110 align:middle line:84% As they learned more about the human genome, 01:06:56.110 --> 01:06:59.140 align:middle line:84% you could be more and more accurate about which 01:06:59.140 --> 01:07:03.520 align:middle line:84% oligonucleotides uniquely represent that gene, as opposed 01:07:03.520 --> 01:07:05.590 align:middle line:90% to matching another gene. 01:07:05.590 --> 01:07:12.220 align:middle line:84% And so what we found is that the larger number of probe pairs 01:07:12.220 --> 01:07:14.830 align:middle line:84% shared between the generations of Affymetrix microarrays, 01:07:14.830 --> 01:07:19.630 align:middle line:84% the better the correlation between those microarrays. 01:07:19.630 --> 01:07:22.450 align:middle line:84% But to answer your question, that 01:07:22.450 --> 01:07:24.700 align:middle line:84% means that if it was done on a previous generation 01:07:24.700 --> 01:07:28.270 align:middle line:84% of microarrays, even within the Affymetrix family, 01:07:28.270 --> 01:07:30.700 align:middle line:84% reproducibility is not good, at least 01:07:30.700 --> 01:07:34.990 align:middle line:84% for those genes that don't meet those criteria, which 01:07:34.990 --> 01:07:36.280 align:middle line:90% is a huge waste. 01:07:36.280 --> 01:07:39.670 align:middle line:84% Because, remember what I told you, 01:07:39.670 --> 01:07:43.000 align:middle line:84% we can do DNA sequencing on blood. 01:07:43.000 --> 01:07:45.730 align:middle line:84% But the expression analysis has to be done on the tissue 01:07:45.730 --> 01:07:46.640 align:middle line:90% that you care about. 01:07:46.640 --> 01:07:49.270 align:middle line:84% So if someone used up some precious brain specimens 01:07:49.270 --> 01:07:51.100 align:middle line:90% to do that, they're gone. 01:07:51.100 --> 01:07:53.437 align:middle line:84% And it'll take them several years to recreate them. 01:07:53.437 --> 01:07:55.270 align:middle line:84% When I actually first presented this at NIH, 01:07:55.270 --> 01:07:58.660 align:middle line:84% I actually heard a groan ripple through the whole audience. 01:07:58.660 --> 01:08:00.730 align:middle line:84% Because they had actually just blithely gone 01:08:00.730 --> 01:08:03.040 align:middle line:84% on from one generation to the other and assumed that 01:08:03.040 --> 01:08:05.860 align:middle line:84% they would just be able to analyze everything together. 01:08:05.860 --> 01:08:08.155 align:middle line:84% And that was millions of dollars down the tube. 01:08:08.155 --> 01:08:10.660 align:middle line:90% 01:08:10.660 --> 01:08:11.630 align:middle line:90% Please? 01:08:11.630 --> 01:08:15.040 align:middle line:84% AUDIENCE: So you assume that all the old stuff is just junk? 01:08:15.040 --> 01:08:16.689 align:middle line:90% ISAAC SAMUEL KOHANE: No. 01:08:16.689 --> 01:08:18.520 align:middle line:84% But what you do is try to figure out 01:08:18.520 --> 01:08:22.040 align:middle line:84% which stuff is reproducible, and you're careful about it. 01:08:22.040 --> 01:08:24.700 align:middle line:90% And it's actually quite doable. 01:08:24.700 --> 01:08:25.630 align:middle line:90% And we do it. 01:08:25.630 --> 01:08:27.262 align:middle line:90% But do we naively-- 01:08:27.262 --> 01:08:28.720 align:middle line:84% and sure, we can't do it wholesale, 01:08:28.720 --> 01:08:30.170 align:middle line:84% the way they thought they were going to do it. 01:08:30.170 --> 01:08:30.648 align:middle line:90% Yes? 01:08:30.648 --> 01:08:32.560 align:middle line:84% AUDIENCE: The good thing is you can do [INAUDIBLE] process. 01:08:32.560 --> 01:08:34.472 align:middle line:84% You can just repeat the whole experiment. 01:08:34.472 --> 01:08:35.350 align:middle line:84% AUDIENCE: If you have the tissue. 01:08:35.350 --> 01:08:37.100 align:middle line:84% ISAAC SAMUEL KOHANE: If you have a tissue. 01:08:37.100 --> 01:08:40.990 align:middle line:84% Plus it's not-- even at $200 a shot to put on a chip, 01:08:40.990 --> 01:08:44.290 align:middle line:84% plus labor costs may be $500 a shot, 01:08:44.290 --> 01:08:48.279 align:middle line:84% so it's 500 times 100 patients, so it's 01:08:48.279 --> 01:08:50.380 align:middle line:90% looking like real money. 01:08:50.380 --> 01:08:53.979 align:middle line:84% And but I think the real problem there 01:08:53.979 --> 01:08:57.950 align:middle line:90% is limited tissue resources. 01:08:57.950 --> 01:09:00.040 align:middle line:90% It turns out to be a real issue. 01:09:00.040 --> 01:09:01.720 align:middle line:84% For those of you have ever tried to do one of these experiments, 01:09:01.720 --> 01:09:03.887 align:middle line:84% getting the right tissue, with the right annotation, 01:09:03.887 --> 01:09:07.359 align:middle line:84% out of the medical system is very hard. 01:09:07.359 --> 01:09:09.880 align:middle line:84% In fact, that's why I really like the MEMP program, 01:09:09.880 --> 01:09:13.600 align:middle line:84% because it makes the engineers go and deal with doctors. 01:09:13.600 --> 01:09:16.090 align:middle line:84% So they understand what kind of social pressures 01:09:16.090 --> 01:09:18.850 align:middle line:84% are efficacious getting the right kind of biomaterials 01:09:18.850 --> 01:09:20.740 align:middle line:90% out of the medical system. 01:09:20.740 --> 01:09:22.960 align:middle line:84% It sounds like a trivial issue, but it's not. 01:09:22.960 --> 01:09:24.609 align:middle line:84% So let me show you the following graph. 01:09:24.609 --> 01:09:31.359 align:middle line:84% Shown here, on the x-axis, is the amount of spiked-in probe. 01:09:31.359 --> 01:09:33.340 align:middle line:84% So this is a probe of known quantity. 01:09:33.340 --> 01:09:36.760 align:middle line:84% And here's the readout from Affymetrix on intensity. 01:09:36.760 --> 01:09:41.620 align:middle line:84% Shown in red are the perfect match probes. 01:09:41.620 --> 01:09:46.630 align:middle line:84% And you see, as you add more and more probe, it increases. 01:09:46.630 --> 01:09:48.430 align:middle line:84% Just as you'd hope, the signal increases. 01:09:48.430 --> 01:09:51.640 align:middle line:84% But what you see here is that it saturates at the high levels, 01:09:51.640 --> 01:09:55.300 align:middle line:90% right, as you'd expect it to. 01:09:55.300 --> 01:09:57.590 align:middle line:84% Now, pretty interesting is the mismatched probes. 01:09:57.590 --> 01:10:00.340 align:middle line:84% These are probes that were designed not to hybridize 01:10:00.340 --> 01:10:01.900 align:middle line:90% with the target sequence. 01:10:01.900 --> 01:10:04.510 align:middle line:84% But they too, in fact, rise also. 01:10:04.510 --> 01:10:06.400 align:middle line:84% They lag, but they rise, as well, 01:10:06.400 --> 01:10:09.880 align:middle line:84% with the amount of spiked-in control. 01:10:09.880 --> 01:10:12.357 align:middle line:84% And in fact, they don't quite saturate. 01:10:12.357 --> 01:10:13.190 align:middle line:90% What does that mean? 01:10:13.190 --> 01:10:15.370 align:middle line:84% It means that the average difference actually 01:10:15.370 --> 01:10:16.450 align:middle line:90% starts dropping down. 01:10:16.450 --> 01:10:18.820 align:middle line:84% Because even though the perfect match is saturating, 01:10:18.820 --> 01:10:21.190 align:middle line:84% the thing that you're subtracting out, 01:10:21.190 --> 01:10:22.870 align:middle line:90% the mismatch keeps on rising. 01:10:22.870 --> 01:10:26.410 align:middle line:84% So at the higher levels, you actually get a dropping signal. 01:10:26.410 --> 01:10:27.820 align:middle line:90% That's lousy. 01:10:27.820 --> 01:10:32.170 align:middle line:84% And that's with a clear solution background. 01:10:32.170 --> 01:10:34.535 align:middle line:84% Here, with a eukaryotic background for the spiking, 01:10:34.535 --> 01:10:35.410 align:middle line:90% it gets even noisier. 01:10:35.410 --> 01:10:38.410 align:middle line:84% You get less sensitivity and even more marked effect. 01:10:38.410 --> 01:10:40.090 align:middle line:90% All of these things are true. 01:10:40.090 --> 01:10:42.520 align:middle line:84% And yet, you could actually do very, very good science 01:10:42.520 --> 01:10:45.130 align:middle line:84% with microarrays and actually discover lots of biology, 01:10:45.130 --> 01:10:49.090 align:middle line:84% but you just have to be aware of its limitations. 01:10:49.090 --> 01:10:54.130 align:middle line:84% Let's get back to the dangers of dimensionality. 01:10:54.130 --> 01:10:56.120 align:middle line:84% Remember how I said, in the very first lecture, 01:10:56.120 --> 01:10:58.390 align:middle line:84% I gave my somewhat lame metaphor about, 01:10:58.390 --> 01:11:04.828 align:middle line:84% if every base was a bead on a necklace 01:11:04.828 --> 01:11:07.120 align:middle line:84% and the necklace was worn by everybody in Shea Stadium, 01:11:07.120 --> 01:11:08.995 align:middle line:84% it's take on the order of 1,000 Shea stadiums 01:11:08.995 --> 01:11:14.415 align:middle line:84% to have as many beads as we have bases in a single human genome. 01:11:14.415 --> 01:11:15.790 align:middle line:84% I was trying to impress you with, 01:11:15.790 --> 01:11:19.000 align:middle line:84% even though we'd like to think that gigabytes are trackable, 01:11:19.000 --> 01:11:20.890 align:middle line:90% it's still a pretty big amount. 01:11:20.890 --> 01:11:23.260 align:middle line:90% So let me revisit that question. 01:11:23.260 --> 01:11:25.900 align:middle line:84% Given 1,000 stadia full of people, with necklaces, 01:11:25.900 --> 01:11:29.350 align:middle line:84% with beads of 10 colors, and let's say 01:11:29.350 --> 01:11:32.770 align:middle line:84% that 1 in 1,000 necklace beads are different every baseball 01:11:32.770 --> 01:11:33.430 align:middle line:90% season. 01:11:33.430 --> 01:11:35.080 align:middle line:84% And you notice that the third seat, 01:11:35.080 --> 01:11:37.210 align:middle line:84% on the fifth row of all games, has 01:11:37.210 --> 01:11:39.730 align:middle line:84% a yellow bead in the middle of the necklace, every year, 01:11:39.730 --> 01:11:41.650 align:middle line:84% in the season opener, in the largest New York 01:11:41.650 --> 01:11:45.130 align:middle line:84% stadium, in the last 26 wins, in the last 102 years 01:11:45.130 --> 01:11:47.660 align:middle line:84% that the New York Yankees have won the World Series. 01:11:47.660 --> 01:11:49.780 align:middle line:90% In fact, let me restate that. 01:11:49.780 --> 01:11:54.190 align:middle line:84% Every time in the last 102 years that you, the very old sports 01:11:54.190 --> 01:11:59.200 align:middle line:84% fan, have seen that this equally old person is wearing 01:11:59.200 --> 01:12:01.720 align:middle line:84% a yellow bead in the same spot in the necklace, 01:12:01.720 --> 01:12:08.650 align:middle line:84% each one of those 102 World Series, the Yankees have won. 01:12:08.650 --> 01:12:11.440 align:middle line:84% I'm just asserting that as a fact. 01:12:11.440 --> 01:12:14.500 align:middle line:84% How good a bet is it that the bead will 01:12:14.500 --> 01:12:18.160 align:middle line:84% be also yellow on that position next time 01:12:18.160 --> 01:12:21.810 align:middle line:84% the Yankees are in the World Series? 01:12:21.810 --> 01:12:24.490 align:middle line:84% Do any of you want to put money, their own money on that bet? 01:12:24.490 --> 01:12:32.490 align:middle line:90% 01:12:32.490 --> 01:12:38.490 align:middle line:84% Anybody give me $1 for that bet with a $1 million in exchange, 01:12:38.490 --> 01:12:40.023 align:middle line:90% with $1 billion in exchange? 01:12:40.023 --> 01:12:41.190 align:middle line:90% The answer is no, of course. 01:12:41.190 --> 01:12:45.810 align:middle line:84% Because it's all too easy, after the fact, 01:12:45.810 --> 01:12:51.540 align:middle line:84% to look through all the games and through all the millions 01:12:51.540 --> 01:12:55.650 align:middle line:84% of combinations of necklaces, to find one necklace that 01:12:55.650 --> 01:12:58.780 align:middle line:84% was highly, highly correlated with a particular outcome, 01:12:58.780 --> 01:13:03.780 align:middle line:84% whether it be the Yankees winning the World 01:13:03.780 --> 01:13:11.440 align:middle line:84% Series or the color of the Coke bottle tops that day. 01:13:11.440 --> 01:13:14.010 align:middle line:84% The point is, if you have enough opportunities 01:13:14.010 --> 01:13:17.590 align:middle line:84% to test something, you're always going to find it. 01:13:17.590 --> 01:13:22.860 align:middle line:84% Which brings us to the following point, 01:13:22.860 --> 01:13:27.330 align:middle line:84% if you have that much in the way of genes, and only hundreds 01:13:27.330 --> 01:13:29.430 align:middle line:84% of patients, it's going to be all too 01:13:29.430 --> 01:13:32.970 align:middle line:84% easy to find a correlation between the values of those set 01:13:32.970 --> 01:13:36.720 align:middle line:84% of genes and the outcome that you care about, 01:13:36.720 --> 01:13:43.293 align:middle line:90% let's say mortality. 01:13:43.293 --> 01:13:44.740 align:middle line:90% AUDIENCE: Can I ask a question? 01:13:44.740 --> 01:13:46.170 align:middle line:90% ISAAC SAMUEL KOHANE: Please do. 01:13:46.170 --> 01:13:47.698 align:middle line:90% AUDIENCE: I mean-- 01:13:47.698 --> 01:13:48.990 align:middle line:90% ISAAC SAMUEL KOHANE: Go for it. 01:13:48.990 --> 01:13:50.250 align:middle line:84% AUDIENCE: I actually need to think about it. 01:13:50.250 --> 01:13:51.542 align:middle line:90% ISAAC SAMUEL KOHANE: Go for it. 01:13:51.542 --> 01:13:53.100 align:middle line:90% Go for it. 01:13:53.100 --> 01:13:54.942 align:middle line:84% In fact, that would be a good project, 01:13:54.942 --> 01:13:56.400 align:middle line:84% just to calculate that probability. 01:13:56.400 --> 01:14:00.740 align:middle line:84% AUDIENCE: Well, I get what you're saying. 01:14:00.740 --> 01:14:02.983 align:middle line:84% And I would accept it wholeheartedly, 01:14:02.983 --> 01:14:05.400 align:middle line:84% without any reservations, if you were talking about just-- 01:14:05.400 --> 01:14:07.808 align:middle line:90% maybe I was misinterpreting-- 01:14:07.808 --> 01:14:09.270 align:middle line:90% ISAAC SAMUEL KOHANE: Yeah. 01:14:09.270 --> 01:14:12.060 align:middle line:84% AUDIENCE: --just about any necklace 01:14:12.060 --> 01:14:13.670 align:middle line:90% in the audience or whatever. 01:14:13.670 --> 01:14:15.780 align:middle line:90% But as you begin to-- 01:14:15.780 --> 01:14:17.250 align:middle line:84% I mean, the example that you gave 01:14:17.250 --> 01:14:20.340 align:middle line:84% was starting to get more and more and more and more 01:14:20.340 --> 01:14:20.940 align:middle line:90% specific. 01:14:20.940 --> 01:14:24.960 align:middle line:84% And if I remember, correctly, the more 01:14:24.960 --> 01:14:30.580 align:middle line:84% specifications that you put on a probability, 01:14:30.580 --> 01:14:31.695 align:middle line:90% the harder it is to-- 01:14:31.695 --> 01:14:32.820 align:middle line:90% ISAAC SAMUEL KOHANE: Right. 01:14:32.820 --> 01:14:33.737 align:middle line:90% AUDIENCE: --feel that. 01:14:33.737 --> 01:14:36.690 align:middle line:84% And so then you begin to get into correlations 01:14:36.690 --> 01:14:39.690 align:middle line:84% that actually are just-- are probabilistically 01:14:39.690 --> 01:14:40.523 align:middle line:90% significant right? 01:14:40.523 --> 01:14:41.940 align:middle line:84% ISAAC SAMUEL KOHANE: Except that-- 01:14:41.940 --> 01:14:43.080 align:middle line:90% you're absolutely right. 01:14:43.080 --> 01:14:43.860 align:middle line:90% Everything you said is true. 01:14:43.860 --> 01:14:44.940 align:middle line:84% But you missed the following fact, 01:14:44.940 --> 01:14:46.650 align:middle line:84% which is I had the opportunity now 01:14:46.650 --> 01:14:50.130 align:middle line:84% to look at all the people in all these thousand Stadia. 01:14:50.130 --> 01:14:57.872 align:middle line:84% And all I had to do was find one bead on one necklace that 01:14:57.872 --> 01:14:58.830 align:middle line:90% would predict the game. 01:14:58.830 --> 01:15:03.180 align:middle line:84% I had millions of opportunities to find that. 01:15:03.180 --> 01:15:04.680 align:middle line:84% But you see, that's very good, Todd. 01:15:04.680 --> 01:15:06.540 align:middle line:84% Because you just went through the same error 01:15:06.540 --> 01:15:08.730 align:middle line:84% that all the functional genomicists make. 01:15:08.730 --> 01:15:10.590 align:middle line:84% Because after the fact, I can always 01:15:10.590 --> 01:15:13.770 align:middle line:84% find, with the thousands of genes, some gene that 01:15:13.770 --> 01:15:17.880 align:middle line:84% was up or down by dumb luck across the phenotype 01:15:17.880 --> 01:15:22.470 align:middle line:84% that I care about, bad cancer versus good cancer. 01:15:22.470 --> 01:15:24.300 align:middle line:84% And that's an important intuition to have. 01:15:24.300 --> 01:15:26.100 align:middle line:84% And I'm glad you asked the question, 01:15:26.100 --> 01:15:29.620 align:middle line:90% because that's exactly it. 01:15:29.620 --> 01:15:32.910 align:middle line:90% If indeed, I had been able to-- 01:15:32.910 --> 01:15:34.680 align:middle line:84% if I had a hypothesis I kept updating. 01:15:34.680 --> 01:15:36.690 align:middle line:84% And every time, it got stronger and stronger, then, boy, 01:15:36.690 --> 01:15:38.580 align:middle line:84% was I a genius by finding that yellow bead. 01:15:38.580 --> 01:15:43.020 align:middle line:84% But if after the fact, after having done all the experiments 01:15:43.020 --> 01:15:48.480 align:middle line:84% with 102 World Series, I look and search all the attendants 01:15:48.480 --> 01:15:50.130 align:middle line:84% to find which bead that correlates, 01:15:50.130 --> 01:15:51.480 align:middle line:90% you know I'm going to find it. 01:15:51.480 --> 01:15:53.520 align:middle line:84% It's certain that I'm going to find such a bead. 01:15:53.520 --> 01:15:54.750 align:middle line:84% But what's the likelihood that that bead 01:15:54.750 --> 01:15:56.750 align:middle line:84% is going to be useful for the next World Series? 01:15:56.750 --> 01:15:59.220 align:middle line:90% 0 or something. 01:15:59.220 --> 01:16:01.110 align:middle line:84% AUDIENCE: Well, you can hypothesize. 01:16:01.110 --> 01:16:03.810 align:middle line:84% The correct approach for scientists-- 01:16:03.810 --> 01:16:05.670 align:middle line:84% and just keeping to your analogy-- 01:16:05.670 --> 01:16:07.680 align:middle line:90% would be to hypothesize that. 01:16:07.680 --> 01:16:09.220 align:middle line:90% You wouldn't say, it is. 01:16:09.220 --> 01:16:13.890 align:middle line:84% But you'd say, if this is truly a phenomenon that actually 01:16:13.890 --> 01:16:15.790 align:middle line:90% exists, then we should see it. 01:16:15.790 --> 01:16:17.370 align:middle line:84% And then if you saw it, we would be 01:16:17.370 --> 01:16:19.230 align:middle line:84% able to make a stronger statement about it. 01:16:19.230 --> 01:16:20.355 align:middle line:90% ISAAC SAMUEL KOHANE: Right. 01:16:20.355 --> 01:16:21.370 align:middle line:90% That is correct. 01:16:21.370 --> 01:16:23.490 align:middle line:90% But that's not what happens. 01:16:23.490 --> 01:16:25.480 align:middle line:84% AUDIENCE: That's not what people do. 01:16:25.480 --> 01:16:26.880 align:middle line:84% ISAAC SAMUEL KOHANE: So remember, 01:16:26.880 --> 01:16:29.280 align:middle line:84% this is my poster child for you, that I gave you, 01:16:29.280 --> 01:16:31.150 align:middle line:90% of large b-cell lymphoma. 01:16:31.150 --> 01:16:34.290 align:middle line:90% And I like this paper a lot. 01:16:34.290 --> 01:16:35.550 align:middle line:90% It's groundbreaking. 01:16:35.550 --> 01:16:37.260 align:middle line:84% And it says the list of genes that 01:16:37.260 --> 01:16:39.660 align:middle line:84% distinguish between low and high clinical risk. 01:16:39.660 --> 01:16:41.310 align:middle line:90% Do you remember this paper? 01:16:41.310 --> 01:16:41.970 align:middle line:90% All right. 01:16:41.970 --> 01:16:45.132 align:middle line:84% This appeared, I think, in Nature. 01:16:45.132 --> 01:16:48.090 align:middle line:90% 01:16:48.090 --> 01:16:54.660 align:middle line:84% 2002, same disease, different microarray platform-- 01:16:54.660 --> 01:16:56.610 align:middle line:90% it's out of Todd's group. 01:16:56.610 --> 01:17:02.070 align:middle line:84% And they, too, find out high and low risk cohort 01:17:02.070 --> 01:17:04.020 align:middle line:84% that are predicted by a set of genes. 01:17:04.020 --> 01:17:06.450 align:middle line:84% I have a process point, which is, how the hell 01:17:06.450 --> 01:17:09.848 align:middle line:84% do you use, essentially, the same technique 01:17:09.848 --> 01:17:12.390 align:middle line:84% on the same disease and get into another first class journal? 01:17:12.390 --> 01:17:18.340 align:middle line:84% This was, I think, Nature Medicine 01:17:18.340 --> 01:17:19.590 align:middle line:90% or was it New England Journal? 01:17:19.590 --> 01:17:20.490 align:middle line:90% I can't remember which. 01:17:20.490 --> 01:17:20.760 align:middle line:90% AUDIENCE: New England Journal. 01:17:20.760 --> 01:17:21.660 align:middle line:84% ISAAC SAMUEL KOHANE: New England Journal, I think. 01:17:21.660 --> 01:17:22.840 align:middle line:90% New England Journal. 01:17:22.840 --> 01:17:23.340 align:middle line:90% Yeah. 01:17:23.340 --> 01:17:25.350 align:middle line:84% No, no, no, that was Nature Medicine. 01:17:25.350 --> 01:17:27.930 align:middle line:90% Then yet a third paper-- 01:17:27.930 --> 01:17:29.550 align:middle line:84% and no, I fully don't understand how 01:17:29.550 --> 01:17:34.920 align:middle line:84% this can happen-- same disease, same question, also 01:17:34.920 --> 01:17:36.360 align:middle line:84% using microarrays, and they, too, 01:17:36.360 --> 01:17:38.370 align:middle line:90% find a high and low risk group. 01:17:38.370 --> 01:17:42.770 align:middle line:90% 01:17:42.770 --> 01:17:46.425 align:middle line:84% What was the overlap in the set of genes that predicted outcome 01:17:46.425 --> 01:17:46.925 align:middle line:90% in these? 01:17:46.925 --> 01:17:51.030 align:middle line:90% 01:17:51.030 --> 01:17:57.418 align:middle line:84% On the order of 20% to 30% depending on how you sliced it. 01:17:57.418 --> 01:17:58.388 align:middle line:90% AUDIENCE: Chance. 01:17:58.388 --> 01:18:00.930 align:middle line:84% ISAAC SAMUEL KOHANE: Which is getting pretty close to chance. 01:18:00.930 --> 01:18:02.120 align:middle line:84% Now, I think there's some signal there. 01:18:02.120 --> 01:18:03.578 align:middle line:84% There clearly is some signal there, 01:18:03.578 --> 01:18:07.380 align:middle line:84% but I don't know how much of this yellow bead phenomenon 01:18:07.380 --> 01:18:11.160 align:middle line:84% we have, which is basically the multiple hypothesis testing 01:18:11.160 --> 01:18:14.830 align:middle line:84% problem revisited in a very, very vicious way. 01:18:14.830 --> 01:18:16.878 align:middle line:90% And the problem is-- 01:18:16.878 --> 01:18:18.420 align:middle line:84% this is more than actually a problem. 01:18:18.420 --> 01:18:19.770 align:middle line:84% I'm just irked, of course, that the same thing 01:18:19.770 --> 01:18:21.930 align:middle line:84% got published three times in first class journals. 01:18:21.930 --> 01:18:23.388 align:middle line:84% But the real problem is that people 01:18:23.388 --> 01:18:25.980 align:middle line:84% are being stratified today in oncology protocols 01:18:25.980 --> 01:18:27.850 align:middle line:90% based on expression profiles. 01:18:27.850 --> 01:18:31.020 align:middle line:84% So you know that, if they're not taking this multiple hypothesis 01:18:31.020 --> 01:18:33.210 align:middle line:84% testing really to heart, patients 01:18:33.210 --> 01:18:39.550 align:middle line:84% are being stratified the wrong way based on a subset of genes. 01:18:39.550 --> 01:18:41.220 align:middle line:84% And remember, I told you that you 01:18:41.220 --> 01:18:44.100 align:middle line:84% could get together as groups of two for a final project? 01:18:44.100 --> 01:18:46.110 align:middle line:84% I would strongly urge at least one group 01:18:46.110 --> 01:18:48.888 align:middle line:84% to think about looking at something like this. 01:18:48.888 --> 01:18:51.180 align:middle line:84% Perhaps this very study, because no one to my knowledge 01:18:51.180 --> 01:18:54.170 align:middle line:90% yet has actually published this. 01:18:54.170 --> 01:18:57.990 align:middle line:84% The emperor is not, at least, wearing enough clothes. 01:18:57.990 --> 01:18:58.990 align:middle line:90% What was overlap? 01:18:58.990 --> 01:19:00.150 align:middle line:84% And what was the characteristic of the genes 01:19:00.150 --> 01:19:01.590 align:middle line:90% that overlapped across those? 01:19:01.590 --> 01:19:04.180 align:middle line:84% And could you, in fact, come up with a robust-- 01:19:04.180 --> 01:19:07.830 align:middle line:84% more robust set of predictors based on those three studies. 01:19:07.830 --> 01:19:09.143 align:middle line:90% But that's highly problematic. 01:19:09.143 --> 01:19:11.685 align:middle line:84% And I think most of the fault is in this overfitting problem. 01:19:11.685 --> 01:19:14.740 align:middle line:90% 01:19:14.740 --> 01:19:16.287 align:middle line:84% So why are these things inconsistent? 01:19:16.287 --> 01:19:18.120 align:middle line:84% Well, also, they were different populations. 01:19:18.120 --> 01:19:19.930 align:middle line:90% Not all humans are the same. 01:19:19.930 --> 01:19:21.430 align:middle line:84% And this brings me to another point, 01:19:21.430 --> 01:19:24.915 align:middle line:84% which is, why, all of a sudden, is it OK to do 100 patients 01:19:24.915 --> 01:19:27.540 align:middle line:84% and do a clinical study, whereas for everything else that we've 01:19:27.540 --> 01:19:28.890 align:middle line:84% ever did, with much fewer variables, 01:19:28.890 --> 01:19:30.285 align:middle line:84% we had to get thousands of patients? 01:19:30.285 --> 01:19:32.160 align:middle line:84% And the answer, of course, is because we have 01:19:32.160 --> 01:19:34.030 align:middle line:90% limited resources and so on. 01:19:34.030 --> 01:19:35.970 align:middle line:84% But I think it's somewhat delusional 01:19:35.970 --> 01:19:39.795 align:middle line:84% to think that we could get as good predictors, 01:19:39.795 --> 01:19:42.420 align:middle line:84% with only hundreds of patients, as we could get with thousands. 01:19:42.420 --> 01:19:44.507 align:middle line:84% Now it's true that we can characterize two groups. 01:19:44.507 --> 01:19:46.840 align:middle line:84% Because we have all these thousands of genes to measure. 01:19:46.840 --> 01:19:48.270 align:middle line:84% And we can measure broad patterns. 01:19:48.270 --> 01:19:52.140 align:middle line:84% In fact, Botstein, who was one of the leaders in this area, 01:19:52.140 --> 01:19:55.620 align:middle line:84% never pretends that any single gene is a reliable measurement. 01:19:55.620 --> 01:19:59.010 align:middle line:84% He says, just give me an overall impression of the biology, 01:19:59.010 --> 01:20:00.330 align:middle line:90% of the pattern that's going on. 01:20:00.330 --> 01:20:02.580 align:middle line:84% But when you start then taking a set of specific genes 01:20:02.580 --> 01:20:04.770 align:middle line:84% and then hanging someone's prognosis on that, 01:20:04.770 --> 01:20:07.440 align:middle line:84% then you're perhaps getting close to a line 01:20:07.440 --> 01:20:09.960 align:middle line:84% where you have to be a lot more methodologically strict 01:20:09.960 --> 01:20:13.590 align:middle line:84% and understand what are the power and significance 01:20:13.590 --> 01:20:14.988 align:middle line:90% of your studies. 01:20:14.988 --> 01:20:17.280 align:middle line:84% And again, some of them are using different measurement 01:20:17.280 --> 01:20:18.210 align:middle line:90% platforms. 01:20:18.210 --> 01:20:21.210 align:middle line:84% There's overfitting and some use of indirect measures. 01:20:21.210 --> 01:20:26.640 align:middle line:84% So that's exactly where I wanted to end up today. 01:20:26.640 --> 01:20:28.522 align:middle line:84% I want to give you a heads up that, 01:20:28.522 --> 01:20:30.730 align:middle line:84% soon, you going to be getting your first problem set. 01:20:30.730 --> 01:20:32.438 align:middle line:84% It's going to be a very simple one, which 01:20:32.438 --> 01:20:36.210 align:middle line:84% is it's going to be a treasure hunt through all 01:20:36.210 --> 01:20:41.303 align:middle line:84% the national databases, to allow you to just make sure you touch 01:20:41.303 --> 01:20:42.720 align:middle line:84% all the databases, and so that you 01:20:42.720 --> 01:20:45.727 align:middle line:84% can be a modern biological researcher 01:20:45.727 --> 01:20:47.310 align:middle line:84% and find out what you need to find out 01:20:47.310 --> 01:20:51.985 align:middle line:84% about the different biological databases. 01:20:51.985 --> 01:20:54.360 align:middle line:84% The other problem set's going to be one around clustering 01:20:54.360 --> 01:20:55.152 align:middle line:90% and classification. 01:20:55.152 --> 01:20:56.940 align:middle line:84% Just make sure you can do it correctly. 01:20:56.940 --> 01:21:00.300 align:middle line:84% But I'd like to start thinking about your final projects. 01:21:00.300 --> 01:21:02.610 align:middle line:90% And it's now mid-February. 01:21:02.610 --> 01:21:05.520 align:middle line:84% Again, I really would like to have 01:21:05.520 --> 01:21:10.050 align:middle line:84% settled on your final project no later by than mid-March. 01:21:10.050 --> 01:21:13.800 align:middle line:84% So, therefore, if you're having any doubts 01:21:13.800 --> 01:21:16.260 align:middle line:84% that you're honing in on a project, 01:21:16.260 --> 01:21:18.720 align:middle line:84% please talk to me earlier than later. 01:21:18.720 --> 01:21:19.947 align:middle line:90% And there's nothing wrong. 01:21:19.947 --> 01:21:21.280 align:middle line:90% In fact, it's just the opposite. 01:21:21.280 --> 01:21:23.100 align:middle line:84% It's great if you can think of a problem 01:21:23.100 --> 01:21:25.392 align:middle line:84% that you would actually like to do further research on, 01:21:25.392 --> 01:21:29.600 align:middle line:84% there's nothing more motivating than that to address.