WEBVTT 00:00:00.000 --> 00:00:02.730 align:middle line:84% TODD GOLUB: So I come at this actually 00:00:02.730 --> 00:00:05.660 align:middle line:84% originally from a pediatric oncology perspective. 00:00:05.660 --> 00:00:10.440 align:middle line:84% So I'm going to start by giving examples 00:00:10.440 --> 00:00:14.280 align:middle line:84% of a couple of patients that I saw in the Jimmy Fund clinic 00:00:14.280 --> 00:00:17.170 align:middle line:84% at the Dana-Farber that were typical. 00:00:17.170 --> 00:00:22.650 align:middle line:84% So the first patient was a nine-year-old girl who 00:00:22.650 --> 00:00:26.010 align:middle line:84% presented to her pediatrician with-- 00:00:26.010 --> 00:00:27.037 align:middle line:90% turn these off? 00:00:27.037 --> 00:00:30.170 align:middle line:90% 00:00:30.170 --> 00:00:32.009 align:middle line:90% With a fever and bruises. 00:00:32.009 --> 00:00:34.790 align:middle line:90% 00:00:34.790 --> 00:00:37.940 align:middle line:84% She got a blood test and then a bone marrow test 00:00:37.940 --> 00:00:41.300 align:middle line:84% that revealed that her bone marrow had been replaced 00:00:41.300 --> 00:00:44.660 align:middle line:90% by acute leukemia cells-- 00:00:44.660 --> 00:00:47.390 align:middle line:84% acute lymphoblastic leukemia cells, or ALL. 00:00:47.390 --> 00:00:51.470 align:middle line:84% And so she was enrolled on what was a standard chemotherapy 00:00:51.470 --> 00:00:56.030 align:middle line:84% protocol, which was nine different drugs in rotation 00:00:56.030 --> 00:00:57.500 align:middle line:90% and combination. 00:00:57.500 --> 00:00:59.000 align:middle line:84% She entered remission in three weeks 00:00:59.000 --> 00:01:01.040 align:middle line:90% and is still alive and well. 00:01:01.040 --> 00:01:03.710 align:middle line:84% And then, a few months later, there is a second patient-- 00:01:03.710 --> 00:01:06.380 align:middle line:84% this kid happened to be a boy, about the same age-- 00:01:06.380 --> 00:01:10.940 align:middle line:84% same presentation, same diagnosis-- 00:01:10.940 --> 00:01:13.188 align:middle line:84% acute lymphoblastic leukemia-- was enrolled 00:01:13.188 --> 00:01:14.480 align:middle line:90% on the same treatment protocol. 00:01:14.480 --> 00:01:17.810 align:middle line:90% So got the same drugs. 00:01:17.810 --> 00:01:19.490 align:middle line:90% Same hospital settings. 00:01:19.490 --> 00:01:21.590 align:middle line:84% So it was as close to a controlled experiment 00:01:21.590 --> 00:01:26.840 align:middle line:84% as one could do in a human being because, of course, response 00:01:26.840 --> 00:01:29.720 align:middle line:84% to any kind of therapeutic intervention 00:01:29.720 --> 00:01:32.630 align:middle line:84% is not only a measure of the treatment 00:01:32.630 --> 00:01:35.120 align:middle line:84% itself but how it's delivered. so that was controlled, 00:01:35.120 --> 00:01:40.730 align:middle line:84% But this patient didn't respond, unfortunately, 00:01:40.730 --> 00:01:43.260 align:middle line:90% and died about six months later. 00:01:43.260 --> 00:01:45.830 align:middle line:84% So the overarching question, that I'm 00:01:45.830 --> 00:01:47.900 align:middle line:84% sure you've been addressing in this course, also, 00:01:47.900 --> 00:01:52.640 align:middle line:84% is generally how to understand this clinical variability 00:01:52.640 --> 00:01:56.270 align:middle line:84% and to discern whether there's a molecular underpinning 00:01:56.270 --> 00:01:59.640 align:middle line:90% to that variability. 00:01:59.640 --> 00:02:04.160 align:middle line:84% So we're particularly interested in this patient who 00:02:04.160 --> 00:02:06.260 align:middle line:84% responded particularly well and so did 00:02:06.260 --> 00:02:09.020 align:middle line:84% what is pretty straightforward, which 00:02:09.020 --> 00:02:13.790 align:middle line:84% is to do a standard chromosome analysis, karyotyping 00:02:13.790 --> 00:02:17.330 align:middle line:84% experiment using some molecular techniques known 00:02:17.330 --> 00:02:19.400 align:middle line:84% as fluorescence in situ hybridization. 00:02:19.400 --> 00:02:23.210 align:middle line:84% Details don't really matter, except to say that, 00:02:23.210 --> 00:02:28.550 align:middle line:84% while it wasn't apparent at the routine morphological analysis 00:02:28.550 --> 00:02:31.460 align:middle line:84% of the chromosomes, if you look molecularly, 00:02:31.460 --> 00:02:33.830 align:middle line:84% it was clear that there is a translocation between 00:02:33.830 --> 00:02:39.410 align:middle line:84% chromosomes 12 and 21 that fused two genes-- 00:02:39.410 --> 00:02:40.970 align:middle line:84% actually two transcription factors-- 00:02:40.970 --> 00:02:42.980 align:middle line:90% one called TEL, one called AML-- 00:02:42.980 --> 00:02:46.898 align:middle line:84% to make this tell AML1 translocation. 00:02:46.898 --> 00:02:49.190 align:middle line:84% And it turned out that, while it hadn't been previously 00:02:49.190 --> 00:02:52.790 align:middle line:84% discovered, actually, the majority 00:02:52.790 --> 00:02:58.680 align:middle line:84% of known genetic abnormalities in childhood leukemia-- 00:02:58.680 --> 00:03:01.910 align:middle line:84% the lion's share of those actually 00:03:01.910 --> 00:03:03.920 align:middle line:90% have this translocation. 00:03:03.920 --> 00:03:06.200 align:middle line:84% And of those patients in retrospective and now 00:03:06.200 --> 00:03:12.440 align:middle line:84% perspective analyzes, about 90% of those patients survive. 00:03:12.440 --> 00:03:15.920 align:middle line:84% It turned out that patient number two had a different 00:03:15.920 --> 00:03:18.980 align:middle line:84% translocation-- a 922 translocation-- 00:03:18.980 --> 00:03:23.060 align:middle line:84% that fuses two genes BCR and ABL-- 00:03:23.060 --> 00:03:24.515 align:middle line:84% that you've probably talked about. 00:03:24.515 --> 00:03:26.390 align:middle line:84% And it's known that these patients, give them 00:03:26.390 --> 00:03:31.230 align:middle line:84% the same therapeutic regimen, only have about 10% survival. 00:03:31.230 --> 00:03:36.920 align:middle line:84% And so these are molecular tests done at the time of diagnosis. 00:03:36.920 --> 00:03:42.260 align:middle line:84% So what I think this really sets up-- which is now perhaps 00:03:42.260 --> 00:03:44.390 align:middle line:84% generally accepted but was really emerging 00:03:44.390 --> 00:03:46.460 align:middle line:84% at this time-- was the notion that cancer 00:03:46.460 --> 00:03:48.380 align:middle line:84% is a genetic disease, it seems like, 00:03:48.380 --> 00:03:53.210 align:middle line:84% and that the outcome clinical outcome is predictable based 00:03:53.210 --> 00:03:55.490 align:middle line:84% on molecular determinants at the time of diagnosis 00:03:55.490 --> 00:03:57.380 align:middle line:84% as long as you know what to look for. 00:03:57.380 --> 00:03:58.490 align:middle line:90% Yeah. 00:03:58.490 --> 00:04:00.110 align:middle line:84% AUDIENCE: Was this before Gleevec? 00:04:00.110 --> 00:04:01.568 align:middle line:84% TODD GOLUB: This is before Gleevec. 00:04:01.568 --> 00:04:04.970 align:middle line:84% Turns out that, even with Gleevec, which targets BCR-ABL, 00:04:04.970 --> 00:04:09.080 align:middle line:84% in these patients, it's not particularly effective. 00:04:09.080 --> 00:04:11.240 align:middle line:84% AUDIENCE: What is the one [INAUDIBLE]?? 00:04:11.240 --> 00:04:12.830 align:middle line:90% TODD GOLUB: Uncharacterized. 00:04:12.830 --> 00:04:18.990 align:middle line:84% [AUDIO OUT] balance translocations 00:04:18.990 --> 00:04:22.900 align:middle line:84% that identify a single oncogene at a breakpoint. 00:04:22.900 --> 00:04:26.040 align:middle line:84% So we started thinking about alternative ways 00:04:26.040 --> 00:04:29.640 align:middle line:84% to try to think about molecular classification of cancers 00:04:29.640 --> 00:04:30.570 align:middle line:90% more generally. 00:04:30.570 --> 00:04:35.670 align:middle line:84% And this is just the obvious experiment-- 00:04:35.670 --> 00:04:36.900 align:middle line:90% two biological states. 00:04:36.900 --> 00:04:40.770 align:middle line:84% They could be clinical states or biological states-- 00:04:40.770 --> 00:04:43.440 align:middle line:84% that you collect some kind of genomic information on, 00:04:43.440 --> 00:04:47.700 align:middle line:84% like microarray data, and then have this pesky problem 00:04:47.700 --> 00:04:50.280 align:middle line:84% of actually trying to figure out how to interpret 00:04:50.280 --> 00:04:51.390 align:middle line:90% the patterns that emerge. 00:04:51.390 --> 00:04:55.050 align:middle line:84% And I'll say quite a bit about that part because it's tough. 00:04:55.050 --> 00:04:58.900 align:middle line:90% 00:04:58.900 --> 00:05:02.860 align:middle line:84% Have you talked about the non-analytical 00:05:02.860 --> 00:05:11.270 align:middle line:84% but these laboratory aspects of microarrays at all? 00:05:11.270 --> 00:05:14.050 align:middle line:84% Let me just spend one slide talking about this. 00:05:14.050 --> 00:05:18.370 align:middle line:84% So all of the microarrays are based on the basic principle 00:05:18.370 --> 00:05:25.540 align:middle line:84% of somehow labeling mRNA or its derivatives from a cell 00:05:25.540 --> 00:05:29.030 align:middle line:84% and hybridizing to probes on some of a solid support, 00:05:29.030 --> 00:05:34.480 align:middle line:84% whether it's a microscope slide a silica wafer or something. 00:05:34.480 --> 00:05:37.420 align:middle line:84% They're cDNA arrays, and they're oligonucleotide arrays. 00:05:37.420 --> 00:05:39.220 align:middle line:84% And you can either make them yourself 00:05:39.220 --> 00:05:42.790 align:middle line:84% or you can buy them commercially. 00:05:42.790 --> 00:05:48.010 align:middle line:84% cDNAs generally take the form of two-color hybridizations, where 00:05:48.010 --> 00:05:55.510 align:middle line:84% you simultaneously hybridize a test sample and a reference 00:05:55.510 --> 00:05:59.380 align:middle line:84% or a control, whereas the oligonucleotide arrays 00:05:59.380 --> 00:06:01.840 align:middle line:90% generally use a single color. 00:06:01.840 --> 00:06:05.422 align:middle line:84% This is completely and entirely a historical artifact 00:06:05.422 --> 00:06:06.880 align:middle line:84% of how these things were developed. 00:06:06.880 --> 00:06:09.790 align:middle line:84% There's nothing intrinsic about cDNA arrays 00:06:09.790 --> 00:06:15.430 align:middle line:84% that makes them require two-color hybridization. 00:06:15.430 --> 00:06:18.400 align:middle line:84% The only reason for using two-color hybridization 00:06:18.400 --> 00:06:21.970 align:middle line:84% is, if the quality of arrays is so bad 00:06:21.970 --> 00:06:24.460 align:middle line:84% that you need an internal control for every spot. 00:06:24.460 --> 00:06:26.140 align:middle line:84% Otherwise you can't interpret the data. 00:06:26.140 --> 00:06:29.243 align:middle line:84% And such was the case for the earliest cDNA arrays, which 00:06:29.243 --> 00:06:30.535 align:middle line:90% were the first ones to be made. 00:06:30.535 --> 00:06:33.790 align:middle line:90% 00:06:33.790 --> 00:06:38.680 align:middle line:84% And I would venture to say that two arrays, in general, 00:06:38.680 --> 00:06:42.790 align:middle line:84% are going to become obsolete very quickly, in favor 00:06:42.790 --> 00:06:46.240 align:middle line:84% of oligonucleotide arrays-- probably commercial ones. 00:06:46.240 --> 00:06:47.740 align:middle line:90% The genome is finite. 00:06:47.740 --> 00:06:50.890 align:middle line:84% Once you have the genome represented and identified, 00:06:50.890 --> 00:06:52.925 align:middle line:84% probably not much advantage to making your own. 00:06:52.925 --> 00:06:54.550 align:middle line:84% And most of the arguments that say, oh, 00:06:54.550 --> 00:06:57.880 align:middle line:84% we can make them much cheaper making them ourselves are 00:06:57.880 --> 00:07:04.750 align:middle line:84% usually Enron-type accounting practices where you 00:07:04.750 --> 00:07:08.710 align:middle line:84% don't-- well, we didn't count the cost of all the people 00:07:08.710 --> 00:07:10.122 align:middle line:90% involved in making these things. 00:07:10.122 --> 00:07:12.580 align:middle line:84% We didn't include the fact that we spent three years trying 00:07:12.580 --> 00:07:14.500 align:middle line:84% to figure out how to do these and they didn't exactly work. 00:07:14.500 --> 00:07:16.875 align:middle line:84% When you really get right down to it, usually buying them 00:07:16.875 --> 00:07:18.535 align:middle line:84% is cheaper, and they're higher quality. 00:07:18.535 --> 00:07:21.310 align:middle line:84% We should also mention that there are non-microarray 00:07:21.310 --> 00:07:25.180 align:middle line:84% based expression profiling methods, 00:07:25.180 --> 00:07:28.660 align:middle line:84% like serial analysis of gene expression or something 00:07:28.660 --> 00:07:32.620 align:middle line:84% called MPSS, that are transcript-counting-based 00:07:32.620 --> 00:07:35.560 align:middle line:84% methods, where you use essentially, 00:07:35.560 --> 00:07:41.890 align:middle line:84% DNA, sequencing to enumerate the precise number of copies 00:07:41.890 --> 00:07:45.160 align:middle line:84% of a given transcript in a given cell. 00:07:45.160 --> 00:07:49.560 align:middle line:84% And the proponents of this method say-- 00:07:49.560 --> 00:07:51.310 align:middle line:84% and they're correct-- this is the only way 00:07:51.310 --> 00:07:55.390 align:middle line:84% to really know how many copies of a particular transcript 00:07:55.390 --> 00:07:57.910 align:middle line:90% that are in a particular cell. 00:07:57.910 --> 00:07:59.740 align:middle line:84% I would say that's absolutely true, 00:07:59.740 --> 00:08:02.050 align:middle line:84% but it's actually not that interesting 00:08:02.050 --> 00:08:04.540 align:middle line:84% because, for most biological questions, 00:08:04.540 --> 00:08:07.300 align:middle line:84% it actually doesn't matter how many absolute copies 00:08:07.300 --> 00:08:08.560 align:middle line:90% of transcript there are. 00:08:08.560 --> 00:08:10.210 align:middle line:84% It's not that useful information. 00:08:10.210 --> 00:08:13.930 align:middle line:84% What's useful is to have some kind of comparative experiment 00:08:13.930 --> 00:08:17.320 align:middle line:84% that tells you there are more with statistical significance-- 00:08:17.320 --> 00:08:19.090 align:middle line:84% more in this sample than that sample. 00:08:19.090 --> 00:08:25.360 align:middle line:84% And rarely do you want to know that there are 682 copies, 00:08:25.360 --> 00:08:27.080 align:middle line:90% but it is true. 00:08:27.080 --> 00:08:31.060 align:middle line:84% But I would say that the low throughput and cost 00:08:31.060 --> 00:08:33.700 align:middle line:84% of these counting based methods greatly 00:08:33.700 --> 00:08:35.260 align:middle line:84% outweighs the benefit that you get 00:08:35.260 --> 00:08:37.210 align:middle line:84% from this absolute counting-based method. 00:08:37.210 --> 00:08:38.470 align:middle line:90% Yes. 00:08:38.470 --> 00:08:42.835 align:middle line:84% AUDIENCE: If you have more copies of [INAUDIBLE] 00:08:42.835 --> 00:08:44.960 align:middle line:84% this is what we're talking about with [INAUDIBLE],, 00:08:44.960 --> 00:08:46.624 align:middle line:90% with the same gene. 00:08:46.624 --> 00:08:50.560 align:middle line:90% They are copies of the gene? 00:08:50.560 --> 00:08:55.640 align:middle line:84% TODD GOLUB: Well, copies of the transcript, not of the gene. 00:08:55.640 --> 00:08:56.620 align:middle line:90% So you could say-- 00:08:56.620 --> 00:09:01.600 align:middle line:84% we sequenced-- we identified a million total transcripts 00:09:01.600 --> 00:09:06.580 align:middle line:90% in a cell of which 684 came-- 00:09:06.580 --> 00:09:12.280 align:middle line:84% were transcripts from-- encoded by gene X. 00:09:12.280 --> 00:09:17.440 align:middle line:84% So that can be useful quantitative information, 00:09:17.440 --> 00:09:20.380 align:middle line:90% but not so useful to-- 00:09:20.380 --> 00:09:22.500 align:middle line:84% these experiments are still many thousands 00:09:22.500 --> 00:09:23.680 align:middle line:90% of dollars per sample. 00:09:23.680 --> 00:09:26.235 align:middle line:90% 00:09:26.235 --> 00:09:30.040 align:middle line:84% AUDIENCE: When it comes down to a pharmaceutical company trying 00:09:30.040 --> 00:09:33.700 align:middle line:84% to figure out the mechanism of the disease and actually how 00:09:33.700 --> 00:09:37.900 align:middle line:84% to go in and develop target drug [INAUDIBLE] that might 00:09:37.900 --> 00:09:39.490 align:middle line:90% interfere-- 00:09:39.490 --> 00:09:43.900 align:middle line:84% I'm assuming that this technique would be useful. 00:09:43.900 --> 00:09:48.940 align:middle line:84% For example, if you did one of the above two [INAUDIBLE] 00:09:48.940 --> 00:09:51.130 align:middle line:84% and you don't know the quantitative number 00:09:51.130 --> 00:09:53.320 align:middle line:84% of the transcripts you have present, 00:09:53.320 --> 00:09:54.970 align:middle line:84% you're just going to, qualitatively, 00:09:54.970 --> 00:09:59.035 align:middle line:84% more or less compared to disease state or not or [INAUDIBLE] 00:09:59.035 --> 00:10:00.910 align:middle line:90% people [INAUDIBLE]. 00:10:00.910 --> 00:10:03.190 align:middle line:90% TODD GOLUB: Yeah. 00:10:03.190 --> 00:10:07.240 align:middle line:84% AUDIENCE: If you know a number, doesn't it 00:10:07.240 --> 00:10:10.630 align:middle line:84% give you ability to make speculative speculation 00:10:10.630 --> 00:10:13.510 align:middle line:84% about how those transits are being processed 00:10:13.510 --> 00:10:20.160 align:middle line:84% and whether that's an important process and the technology 00:10:20.160 --> 00:10:21.285 align:middle line:90% becomes easier [INAUDIBLE]? 00:10:21.285 --> 00:10:23.160 align:middle line:84% TODD GOLUB: I would say it doesn't, actually, 00:10:23.160 --> 00:10:25.200 align:middle line:84% because if you tell me-- if God tells me 00:10:25.200 --> 00:10:29.700 align:middle line:84% that there are 684 copies of a given transcript, 00:10:29.700 --> 00:10:31.650 align:middle line:90% I don't know what that means. 00:10:31.650 --> 00:10:38.880 align:middle line:84% I do know what it means that-- if there are five times as much 00:10:38.880 --> 00:10:41.160 align:middle line:84% of a given transcript in a disease state compared 00:10:41.160 --> 00:10:44.790 align:middle line:84% to a normal state, I can at least say that having one fifth 00:10:44.790 --> 00:10:47.040 align:middle line:84% the number of transcripts isn't sufficient to give you 00:10:47.040 --> 00:10:49.980 align:middle line:84% the disease state, whereas, if you just give me an absolute 00:10:49.980 --> 00:10:54.840 align:middle line:84% number without a comparative thing, 00:10:54.840 --> 00:10:57.570 align:middle line:84% I don't know what that means, And I'm not sure that saying, 00:10:57.570 --> 00:11:02.670 align:middle line:84% there's 600 in this state and 3,000 in that state-- 00:11:02.670 --> 00:11:07.020 align:middle line:84% to me, that's not really more useful information 00:11:07.020 --> 00:11:10.080 align:middle line:90% than knowing relative abundance. 00:11:10.080 --> 00:11:12.210 align:middle line:84% AUDIENCE: So it's really a question 00:11:12.210 --> 00:11:19.260 align:middle line:84% of what the resolution is, in terms of common [INAUDIBLE].. 00:11:19.260 --> 00:11:23.375 align:middle line:84% When you get an exact number from the bottom [INAUDIBLE] up 00:11:23.375 --> 00:11:28.710 align:middle line:84% to the top get a sufficiently resolved difference between 00:11:28.710 --> 00:11:30.210 align:middle line:90% samples that you can say-- 00:11:30.210 --> 00:11:31.290 align:middle line:90% TODD GOLUB: That's right. 00:11:31.290 --> 00:11:33.540 align:middle line:90% That's right. 00:11:33.540 --> 00:11:35.018 align:middle line:90% AUDIENCE: 104 through 500. 00:11:35.018 --> 00:11:36.060 align:middle line:90% TODD GOLUB: That's right. 00:11:36.060 --> 00:11:39.420 align:middle line:84% And I would say, for most biological and clinical 00:11:39.420 --> 00:11:43.380 align:middle line:84% applications, knowing that the ratio is 1 to 5 00:11:43.380 --> 00:11:46.350 align:middle line:84% gives you about 95% of the information that's useful, 00:11:46.350 --> 00:11:50.625 align:middle line:84% compared to saying, it's 500 versus 100. 00:11:50.625 --> 00:11:53.702 align:middle line:84% You might be able to think of some special experiments 00:11:53.702 --> 00:11:55.410 align:middle line:84% where you really want to know the number. 00:11:55.410 --> 00:12:01.300 align:middle line:84% But usually, it doesn't help you very much, in my opinion. 00:12:01.300 --> 00:12:04.090 align:middle line:90% 00:12:04.090 --> 00:12:06.010 align:middle line:84% There are all sorts of sources of variability 00:12:06.010 --> 00:12:07.630 align:middle line:90% that we won't discuss. 00:12:07.630 --> 00:12:09.770 align:middle line:84% Except, there's only one that matters. 00:12:09.770 --> 00:12:14.893 align:middle line:84% And that's the biological and clinical variability that 00:12:14.893 --> 00:12:16.060 align:middle line:90% goes into these experiments. 00:12:16.060 --> 00:12:17.560 align:middle line:84% There are a lot of people that spend 00:12:17.560 --> 00:12:22.390 align:middle line:84% a lot of time hand-wringing over these very technical things, 00:12:22.390 --> 00:12:24.610 align:middle line:84% and none of them make any difference, really, 00:12:24.610 --> 00:12:27.190 align:middle line:84% as far as I can tell because they're overwhelmed 00:12:27.190 --> 00:12:29.330 align:middle line:90% by biological variability. 00:12:29.330 --> 00:12:34.210 align:middle line:84% So making microarrays that are somewhat more 00:12:34.210 --> 00:12:37.090 align:middle line:84% accurate or precise in their measurement 00:12:37.090 --> 00:12:40.150 align:middle line:84% won't actually make that much of a dent in the problem 00:12:40.150 --> 00:12:44.830 align:middle line:84% because that variability is exceeded so tremendously 00:12:44.830 --> 00:12:46.600 align:middle line:90% by biological variability. 00:12:46.600 --> 00:12:49.010 align:middle line:90% That's all I'll say about that. 00:12:49.010 --> 00:12:51.410 align:middle line:84% So let me give some examples of applying this. 00:12:51.410 --> 00:12:55.810 align:middle line:84% And I know you've discussed these general methods before, 00:12:55.810 --> 00:12:58.120 align:middle line:90% but that's OK. 00:12:58.120 --> 00:13:00.850 align:middle line:84% So here's an experiment where we are 00:13:00.850 --> 00:13:06.220 align:middle line:84% interested in the differentiable clinical outcome of children 00:13:06.220 --> 00:13:11.350 align:middle line:84% with a brain tumor called medulloblastoma. 00:13:11.350 --> 00:13:13.535 align:middle line:84% Have you discussed this particular example? 00:13:13.535 --> 00:13:15.160 align:middle line:84% AUDIENCE: [INAUDIBLE] but not this one. 00:13:15.160 --> 00:13:15.785 align:middle line:90% TODD GOLUB: OK. 00:13:15.785 --> 00:13:21.310 align:middle line:84% So we had 60 pre-treatment biopsies of patients 00:13:21.310 --> 00:13:22.360 align:middle line:90% with this brain tumor. 00:13:22.360 --> 00:13:23.920 align:middle line:90% Tumors were biopsied. 00:13:23.920 --> 00:13:26.290 align:middle line:84% And we knew the long-term clinical outcome 00:13:26.290 --> 00:13:30.160 align:middle line:84% of these patients-- whether they survived from their disease 00:13:30.160 --> 00:13:33.970 align:middle line:90% or they died despite therapy. 00:13:33.970 --> 00:13:36.910 align:middle line:84% They had the tumors removed and were treated with chemotherapy 00:13:36.910 --> 00:13:38.068 align:middle line:90% and radiation. 00:13:38.068 --> 00:13:39.610 align:middle line:84% And we knew that some of the patients 00:13:39.610 --> 00:13:41.582 align:middle line:84% survived and some of them did not. 00:13:41.582 --> 00:13:43.540 align:middle line:84% And based on that, we said, well, perhaps there 00:13:43.540 --> 00:13:45.160 align:middle line:90% are two classes-- 00:13:45.160 --> 00:13:46.670 align:middle line:90% subclasses of the disease. 00:13:46.670 --> 00:13:51.340 align:middle line:90% So we clustered the data. 00:13:51.340 --> 00:13:55.240 align:middle line:84% This was across, I think, 6,800 genes on a microarray. 00:13:55.240 --> 00:13:57.430 align:middle line:84% So each dot represents a different patient sample. 00:13:57.430 --> 00:13:59.680 align:middle line:84% We said, well, if there are two classes, let's cluster 00:13:59.680 --> 00:14:06.040 align:middle line:84% into two groups because we have a clinical suspicion there 00:14:06.040 --> 00:14:07.690 align:middle line:90% are two classes. 00:14:07.690 --> 00:14:11.572 align:middle line:84% And of course, if you ask the algorithm to cluster 00:14:11.572 --> 00:14:13.030 align:middle line:84% into two groups, in this case, it's 00:14:13.030 --> 00:14:14.590 align:middle line:84% something called a self-organizing map. 00:14:14.590 --> 00:14:15.550 align:middle line:90% But it doesn't matter. 00:14:15.550 --> 00:14:19.260 align:middle line:84% You could take the two major branches on a dendrogram 00:14:19.260 --> 00:14:21.010 align:middle line:84% from a hierarchical clustering dendrogram. 00:14:21.010 --> 00:14:21.642 align:middle line:90% Whatever. 00:14:21.642 --> 00:14:22.600 align:middle line:90% You get the same thing. 00:14:22.600 --> 00:14:23.920 align:middle line:90% You get two classes. 00:14:23.920 --> 00:14:26.860 align:middle line:84% And the patients are shown there-- about equally 00:14:26.860 --> 00:14:27.650 align:middle line:90% distributed. 00:14:27.650 --> 00:14:30.790 align:middle line:84% And now, if we fill in the labels of the patients-- 00:14:30.790 --> 00:14:32.260 align:middle line:84% that is, whether their survivors-- 00:14:32.260 --> 00:14:34.690 align:middle line:84% turned out to be survivors or not survivors-- 00:14:34.690 --> 00:14:37.015 align:middle line:90% you get this picture, where-- 00:14:37.015 --> 00:14:39.970 align:middle line:90% 00:14:39.970 --> 00:14:41.770 align:middle line:84% you don't need a statistician to tell you 00:14:41.770 --> 00:14:46.060 align:middle line:84% that there's no correlation between this class structure 00:14:46.060 --> 00:14:49.190 align:middle line:90% and survival. 00:14:49.190 --> 00:14:50.923 align:middle line:90% So what do you take? 00:14:50.923 --> 00:14:52.465 align:middle line:84% What do you think of this experiment? 00:14:52.465 --> 00:14:57.283 align:middle line:84% AUDIENCE: What do you guys think of this? 00:14:57.283 --> 00:14:58.700 align:middle line:84% TODD GOLUB: What can you conclude? 00:14:58.700 --> 00:14:59.400 align:middle line:84% AUDIENCE: What do you do with this? 00:14:59.400 --> 00:15:00.250 align:middle line:84% How are you going to rescue this? 00:15:00.250 --> 00:15:02.557 align:middle line:84% Are you saying [INAUDIBLE] was [INAUDIBLE] signature 00:15:02.557 --> 00:15:03.640 align:middle line:90% that [INAUDIBLE] survival? 00:15:03.640 --> 00:15:05.483 align:middle line:90% What do you say? 00:15:05.483 --> 00:15:07.400 align:middle line:84% AUDIENCE: There's a difference, but it may not 00:15:07.400 --> 00:15:09.620 align:middle line:90% be related to [INAUDIBLE]. 00:15:09.620 --> 00:15:12.380 align:middle line:84% AUDIENCE: But maybe there was [INAUDIBLE] would 00:15:12.380 --> 00:15:13.270 align:middle line:90% work somewhere else. 00:15:13.270 --> 00:15:16.170 align:middle line:90% [INAUDIBLE] 00:15:16.170 --> 00:15:16.920 align:middle line:90% TODD GOLUB: Right. 00:15:16.920 --> 00:15:21.090 align:middle line:84% I think you're both hitting on the relevant point 00:15:21.090 --> 00:15:27.160 align:middle line:84% here, which is that the unsupervised learning 00:15:27.160 --> 00:15:29.650 align:middle line:84% clustering algorithm found some structure-- 00:15:29.650 --> 00:15:31.805 align:middle line:84% dominant structure that made these two classes. 00:15:31.805 --> 00:15:33.430 align:middle line:84% They just don't happen to have anything 00:15:33.430 --> 00:15:38.350 align:middle line:84% to do with the question that we're interested in, which 00:15:38.350 --> 00:15:39.740 align:middle line:90% was one of survival. 00:15:39.740 --> 00:15:42.640 align:middle line:84% And so that gets back to this basic notion 00:15:42.640 --> 00:15:48.130 align:middle line:84% of two general approaches to data analysis, which 00:15:48.130 --> 00:15:54.050 align:middle line:84% you've probably discussed but I think often gets confused. 00:15:54.050 --> 00:15:56.140 align:middle line:84% So unsupervised learning-- which is not 00:15:56.140 --> 00:15:57.820 align:middle line:84% exactly synonymous with clustering, 00:15:57.820 --> 00:16:00.730 align:middle line:84% but it's a reasonable first approximation, 00:16:00.730 --> 00:16:04.240 align:middle line:84% or supervised learning classification. 00:16:04.240 --> 00:16:09.070 align:middle line:84% So here, you're interested in finding a dominant structure 00:16:09.070 --> 00:16:12.190 align:middle line:84% defined only by the intrinsic gene expression 00:16:12.190 --> 00:16:15.220 align:middle line:84% patterns in a given data set, irrespective of anything 00:16:15.220 --> 00:16:16.870 align:middle line:84% you happen to know about the samples, 00:16:16.870 --> 00:16:19.180 align:middle line:90% such as their clinical outcome. 00:16:19.180 --> 00:16:22.060 align:middle line:90% Here, you're saying, whatever. 00:16:22.060 --> 00:16:24.520 align:middle line:84% Maybe there's some other interesting biological 00:16:24.520 --> 00:16:25.070 align:middle line:90% structure. 00:16:25.070 --> 00:16:26.500 align:middle line:84% But I'm not interested in that right now. 00:16:26.500 --> 00:16:29.125 align:middle line:84% I want to know whether there's a gene expression pattern that's 00:16:29.125 --> 00:16:30.820 align:middle line:84% correlated with the thing I care about 00:16:30.820 --> 00:16:35.690 align:middle line:84% in this particular example outcome. 00:16:35.690 --> 00:16:39.970 align:middle line:84% So we take the same data set-- the same matrix of data samples 00:16:39.970 --> 00:16:48.260 align:middle line:84% by gene expression values and now apply supervised learning 00:16:48.260 --> 00:16:48.760 align:middle line:90% approaches. 00:16:48.760 --> 00:16:52.570 align:middle line:84% This happens to be a K nearest neighbor classifier. 00:16:52.570 --> 00:16:55.480 align:middle line:84% Again, it doesn't make any difference what you use here. 00:16:55.480 --> 00:16:58.435 align:middle line:84% And this happens to be an eight-gene model. 00:16:58.435 --> 00:17:02.660 align:middle line:90% 00:17:02.660 --> 00:17:08.140 align:middle line:84% You classify the samples using a leave-one-out cross-validation 00:17:08.140 --> 00:17:09.550 align:middle line:90% approach so that you don't-- 00:17:09.550 --> 00:17:13.510 align:middle line:84% you attempt not to overfit the data-- 00:17:13.510 --> 00:17:15.460 align:middle line:84% the model to the initial training set. 00:17:15.460 --> 00:17:20.859 align:middle line:84% And then you ask, well, , of the two classes that are predicted 00:17:20.859 --> 00:17:22.569 align:middle line:84% how do those patients actually fare? 00:17:22.569 --> 00:17:26.349 align:middle line:84% And here is a survival plot in terms of months-- 00:17:26.349 --> 00:17:28.080 align:middle line:84% months of survival for those patients 00:17:28.080 --> 00:17:29.830 align:middle line:84% who are predicted to be alive versus those 00:17:29.830 --> 00:17:31.617 align:middle line:90% that are predicted to be dead. 00:17:31.617 --> 00:17:32.950 align:middle line:90% So what do you think about this? 00:17:32.950 --> 00:17:38.490 align:middle line:90% 00:17:38.490 --> 00:17:39.130 align:middle line:90% AUDIENCE: This? 00:17:39.130 --> 00:17:40.487 align:middle line:90% So we'll pick those genes? 00:17:40.487 --> 00:17:41.820 align:middle line:90% TODD GOLUB: Is this significant? 00:17:41.820 --> 00:17:48.090 align:middle line:90% 00:17:48.090 --> 00:17:53.130 align:middle line:84% So if you look in a basic biostatistics 00:17:53.130 --> 00:17:57.150 align:middle line:84% textbook about how to calculate statistical significance 00:17:57.150 --> 00:17:59.910 align:middle line:84% of a Kaplan-Meier survival curve such as this is, 00:17:59.910 --> 00:18:02.400 align:middle line:84% they'll tell you to do the log rank test. 00:18:02.400 --> 00:18:05.460 align:middle line:84% And if you did that, you would get a p-value 00:18:05.460 --> 00:18:07.740 align:middle line:84% that, if you've looked at a lot of these things, 00:18:07.740 --> 00:18:14.070 align:middle line:84% would match your intuition for this degree of separation. 00:18:14.070 --> 00:18:16.290 align:middle line:84% So it would look something like that. 00:18:16.290 --> 00:18:18.720 align:middle line:90% Is that reasonable? 00:18:18.720 --> 00:18:20.460 align:middle line:84% Well, that looks quite significant. 00:18:20.460 --> 00:18:23.700 align:middle line:84% But you should ask, well, how come this model is 00:18:23.700 --> 00:18:25.590 align:middle line:90% eight genes, for example? 00:18:25.590 --> 00:18:28.433 align:middle line:90% How did you choose that number? 00:18:28.433 --> 00:18:29.350 align:middle line:90% Well, it's quite easy. 00:18:29.350 --> 00:18:31.050 align:middle line:84% We chose that because I worked the best. 00:18:31.050 --> 00:18:35.280 align:middle line:84% It worked better than six, worked better than 10, or 50. 00:18:35.280 --> 00:18:41.880 align:middle line:84% So we had to pay some penalty for overfitting 00:18:41.880 --> 00:18:48.040 align:middle line:84% a model, potentially, to this particular data set. 00:18:48.040 --> 00:18:50.310 align:middle line:84% So the ways that you could then really 00:18:50.310 --> 00:18:53.970 align:middle line:84% test the statistical significance of this model 00:18:53.970 --> 00:18:56.680 align:middle line:84% would be to apply it to another data set. 00:18:56.680 --> 00:18:58.290 align:middle line:90% That's the gold standard. 00:18:58.290 --> 00:18:59.940 align:middle line:84% But short of that, a reasonable thing 00:18:59.940 --> 00:19:03.150 align:middle line:84% to do to better approximate the significance 00:19:03.150 --> 00:19:08.910 align:middle line:84% is to take into account the fact that there 00:19:08.910 --> 00:19:13.560 align:middle line:84% are a number of parameters of this model that were optimized 00:19:13.560 --> 00:19:14.520 align:middle line:90% to fit this data set. 00:19:14.520 --> 00:19:19.260 align:middle line:84% And you can do this by doing a permutation test, where 00:19:19.260 --> 00:19:22.770 align:middle line:84% you don't scramble the gene expression values themselves, 00:19:22.770 --> 00:19:27.630 align:middle line:84% but you randomize the class labels in terms of, 00:19:27.630 --> 00:19:30.330 align:middle line:84% are the patients alive or dead, and you 00:19:30.330 --> 00:19:33.150 align:middle line:84% go through the same procedure of attempting 00:19:33.150 --> 00:19:36.960 align:middle line:84% to build an optimal classifier and including 00:19:36.960 --> 00:19:38.580 align:middle line:90% choosing the number of-- 00:19:38.580 --> 00:19:41.700 align:middle line:84% optimal number of genes to ask, if you really 00:19:41.700 --> 00:19:45.540 align:middle line:84% try hard with these machine learning methods, 00:19:45.540 --> 00:19:48.810 align:middle line:84% how often can you make a classifier that works 00:19:48.810 --> 00:19:50.620 align:middle line:90% as well as this one does? 00:19:50.620 --> 00:19:54.810 align:middle line:84% And when we did that 1,000 times, 9 times out of 1,000, 00:19:54.810 --> 00:19:57.970 align:middle line:90% we could do this well or better. 00:19:57.970 --> 00:20:01.900 align:middle line:84% So we estimated the significance of this model here, 00:20:01.900 --> 00:20:03.780 align:middle line:90% which is still decent. 00:20:03.780 --> 00:20:08.160 align:middle line:84% But you can see we took a hit of a couple of orders of magnitude 00:20:08.160 --> 00:20:09.130 align:middle line:90% on this p-value. 00:20:09.130 --> 00:20:13.140 align:middle line:84% So if you had a nominal p-value of 0.05 or something, 00:20:13.140 --> 00:20:17.160 align:middle line:84% that result would entirely vanish when you appropriately 00:20:17.160 --> 00:20:20.310 align:middle line:84% attempt to correct for such multiple hypothesis testing. 00:20:20.310 --> 00:20:26.460 align:middle line:84% And this is independent of what particular classifier you used. 00:20:26.460 --> 00:20:29.490 align:middle line:84% So I would say, much of the literature-- and everyone's 00:20:29.490 --> 00:20:31.330 align:middle line:84% figuring on how to do this as we go along. 00:20:31.330 --> 00:20:35.400 align:middle line:84% But much of the literature and worries about failure 00:20:35.400 --> 00:20:38.040 align:middle line:84% to reproduce an initial model are 00:20:38.040 --> 00:20:42.600 align:middle line:84% due to the problem of overestimating 00:20:42.600 --> 00:20:44.910 align:middle line:84% the significance of an initial model because 00:20:44.910 --> 00:20:46.560 align:middle line:84% of these overfitting types of problems. 00:20:46.560 --> 00:20:50.190 align:middle line:90% 00:20:50.190 --> 00:20:53.180 align:middle line:84% So let me just make a couple of general comments 00:20:53.180 --> 00:20:58.820 align:middle line:90% about supervised learning. 00:20:58.820 --> 00:21:00.570 align:middle line:84% And some of them may seem obvious, 00:21:00.570 --> 00:21:03.500 align:middle line:84% but turn out to be actually problematic. 00:21:03.500 --> 00:21:04.590 align:middle line:90% And this is one example. 00:21:04.590 --> 00:21:06.470 align:middle line:90% So the first step-- 00:21:06.470 --> 00:21:11.690 align:middle line:84% establish the class labels of what you're trying to classify. 00:21:11.690 --> 00:21:13.165 align:middle line:84% So in one of our first experiments 00:21:13.165 --> 00:21:16.610 align:middle line:84% where we were trying to classify the two different basic types 00:21:16.610 --> 00:21:18.350 align:middle line:90% of acute leukemia-- 00:21:18.350 --> 00:21:23.065 align:middle line:84% acute lymphoblastic leukemia or acute myeloid leukemia-- 00:21:23.065 --> 00:21:24.440 align:middle line:84% the way you build a classifier is 00:21:24.440 --> 00:21:26.990 align:middle line:84% to choose examples of one class, examples of another class, 00:21:26.990 --> 00:21:30.800 align:middle line:84% and then find gene expression patterns that 00:21:30.800 --> 00:21:32.370 align:middle line:90% are correlated with that. 00:21:32.370 --> 00:21:37.250 align:middle line:84% Well, who's to say that we have these right, right? 00:21:37.250 --> 00:21:40.190 align:middle line:84% What should you use as the gold standard for these things? 00:21:40.190 --> 00:21:41.900 align:middle line:84% It's not always obvious, particularly 00:21:41.900 --> 00:21:44.570 align:middle line:84% for the very things that we want to build better 00:21:44.570 --> 00:21:49.280 align:middle line:84% molecular classifiers for because the current clinical 00:21:49.280 --> 00:21:51.270 align:middle line:90% diagnostics are so poor. 00:21:51.270 --> 00:21:53.270 align:middle line:84% It doesn't really feel like a good gold standard 00:21:53.270 --> 00:21:58.620 align:middle line:84% to go back, rely on the clinical labels as the gold standard. 00:21:58.620 --> 00:22:00.290 align:middle line:90% And so this is, in general, a. 00:22:00.290 --> 00:22:02.360 align:middle line:84% Real problem for survival studies, 00:22:02.360 --> 00:22:06.530 align:middle line:84% if you force the question into a simple two-class problem-- 00:22:06.530 --> 00:22:08.540 align:middle line:90% survivor or non-survivor-- 00:22:08.540 --> 00:22:10.440 align:middle line:84% well everyone's a non survivor at some point. 00:22:10.440 --> 00:22:15.290 align:middle line:84% So at what point do you declare that a patient is a survivor 00:22:15.290 --> 00:22:17.580 align:middle line:90% of their tumor, for example? 00:22:17.580 --> 00:22:20.030 align:middle line:84% That requires some judgment call in terms of which 00:22:20.030 --> 00:22:22.940 align:middle line:90% bin to put the samples into. 00:22:22.940 --> 00:22:27.590 align:middle line:84% And I would say that much of at least our effort and time 00:22:27.590 --> 00:22:30.567 align:middle line:84% has gone into trying to figure out how to get this right. 00:22:30.567 --> 00:22:32.150 align:middle line:84% And there are some approaches that one 00:22:32.150 --> 00:22:36.320 align:middle line:84% could take to not have to be so rigid on how 00:22:36.320 --> 00:22:39.350 align:middle line:84% you assign these labels, but it's a challenge. 00:22:39.350 --> 00:22:41.940 align:middle line:90% 00:22:41.940 --> 00:22:48.560 align:middle line:84% So the second general step in making the classifier 00:22:48.560 --> 00:22:51.110 align:middle line:84% is selecting the features that you're 00:22:51.110 --> 00:22:54.500 align:middle line:90% going to feed into a model-- 00:22:54.500 --> 00:22:56.300 align:middle line:84% features in this case being genes. 00:22:56.300 --> 00:22:59.360 align:middle line:84% The details don't matter-- their whole long list of ways 00:22:59.360 --> 00:23:02.480 align:middle line:90% that you can rank genes. 00:23:02.480 --> 00:23:06.800 align:middle line:84% This is a simple one that is based on the mean expression 00:23:06.800 --> 00:23:10.910 align:middle line:84% level in the two classes and their standard deviation. 00:23:10.910 --> 00:23:14.180 align:middle line:84% It's a relatively unsophisticated way 00:23:14.180 --> 00:23:16.640 align:middle line:84% to select genes because it assumes that there's 00:23:16.640 --> 00:23:20.810 align:middle line:84% a uniform behavior of these marker genes in the two 00:23:20.810 --> 00:23:25.890 align:middle line:84% classes, which, in many cases, is not at all the case. 00:23:25.890 --> 00:23:29.190 align:middle line:84% There are many other methods as well. 00:23:29.190 --> 00:23:31.710 align:middle line:84% And then I'm sure you talked about these at length. 00:23:31.710 --> 00:23:36.530 align:middle line:84% So I think maybe I'll skip this, that you then take these things 00:23:36.530 --> 00:23:39.430 align:middle line:90% and classify. 00:23:39.430 --> 00:23:44.920 align:middle line:84% So for unsupervised learning, there are all these methods. 00:23:44.920 --> 00:23:47.920 align:middle line:84% And I think they basically don't matter-- 00:23:47.920 --> 00:23:55.450 align:middle line:84% that, either for clustering or supervised machine learning 00:23:55.450 --> 00:23:58.300 align:middle line:84% methods, if you get a result that 00:23:58.300 --> 00:24:02.830 align:middle line:84% is obtainable with only one magical-- 00:24:02.830 --> 00:24:06.310 align:middle line:84% one person's really special algorithm, 00:24:06.310 --> 00:24:11.500 align:middle line:84% I would worry deeply that there is a problem with-- 00:24:11.500 --> 00:24:14.920 align:middle line:84% there's an information leak somehow 00:24:14.920 --> 00:24:17.620 align:middle line:84% or something's not right because, at least 00:24:17.620 --> 00:24:20.030 align:middle line:84% in my experience, when there's really biologically 00:24:20.030 --> 00:24:23.120 align:middle line:84% or clinically meaningful structure to be found, 00:24:23.120 --> 00:24:26.890 align:middle line:84% you can find it with a number of different approaches. 00:24:26.890 --> 00:24:29.620 align:middle line:84% And in fact, that's a reasonable sanity check 00:24:29.620 --> 00:24:33.730 align:middle line:84% to make sure that you can recover structure, 00:24:33.730 --> 00:24:35.530 align:middle line:84% whether you're using machine learning 00:24:35.530 --> 00:24:40.210 align:middle line:84% or unsupervised clustering algorithms-- that you 00:24:40.210 --> 00:24:42.880 align:middle line:84% can recover it with multiple different methods. 00:24:42.880 --> 00:24:45.550 align:middle line:84% There are some examples to that that can be interesting. 00:24:45.550 --> 00:24:51.400 align:middle line:84% But on the whole, I'm confident saying it doesn't really 00:24:51.400 --> 00:24:53.650 align:middle line:84% matter, and it's really the input to these data sets 00:24:53.650 --> 00:24:57.310 align:middle line:84% that matters the most-- that is, are you really sampling 00:24:57.310 --> 00:25:00.340 align:middle line:84% the diversity of, for example, the disease process 00:25:00.340 --> 00:25:02.222 align:middle line:84% that you're studying with the samples that 00:25:02.222 --> 00:25:03.430 align:middle line:90% are in your initial data set? 00:25:03.430 --> 00:25:07.450 align:middle line:90% 00:25:07.450 --> 00:25:10.390 align:middle line:84% What's more challenging is, how do you 00:25:10.390 --> 00:25:13.030 align:middle line:84% evaluate the output of these clusters 00:25:13.030 --> 00:25:15.400 align:middle line:84% in terms of their biology, in terms 00:25:15.400 --> 00:25:17.770 align:middle line:84% of really knowing how robust the structure is, 00:25:17.770 --> 00:25:21.130 align:middle line:90% given any algorithm? 00:25:21.130 --> 00:25:25.000 align:middle line:84% And then how do you actually know when you-- 00:25:25.000 --> 00:25:27.730 align:middle line:84% having seen a given structure once, 00:25:27.730 --> 00:25:29.950 align:middle line:84% how do you actually apply it to another data 00:25:29.950 --> 00:25:32.152 align:middle line:84% set to know whether you see it there as well? 00:25:32.152 --> 00:25:32.860 align:middle line:90% It's not obvious. 00:25:32.860 --> 00:25:37.970 align:middle line:90% 00:25:37.970 --> 00:25:39.650 align:middle line:84% Any questions about that general stuff? 00:25:39.650 --> 00:25:41.317 align:middle line:84% I wasn't going to say any more because I 00:25:41.317 --> 00:25:45.092 align:middle line:90% know you've covered it. 00:25:45.092 --> 00:25:46.550 align:middle line:84% Let me give a couple other examples 00:25:46.550 --> 00:25:51.233 align:middle line:84% of applying these principles to some data sets. 00:25:51.233 --> 00:25:52.400 align:middle line:90% Did you talk about this one? 00:25:52.400 --> 00:25:53.130 align:middle line:90% AUDIENCE: No. 00:25:53.130 --> 00:25:54.860 align:middle line:90% TODD GOLUB: OK, so here-- 00:25:54.860 --> 00:25:57.800 align:middle line:84% again, focusing on childhood leukemia-- 00:25:57.800 --> 00:26:03.650 align:middle line:84% most kids with childhood ALL respond to chemotherapy. 00:26:03.650 --> 00:26:05.930 align:middle line:84% I told you about this subgroup of BCR-ABL patients 00:26:05.930 --> 00:26:06.740 align:middle line:90% that does not. 00:26:06.740 --> 00:26:11.480 align:middle line:84% Another group that does not respond well 00:26:11.480 --> 00:26:15.140 align:middle line:84% are infants less than a year of age, 00:26:15.140 --> 00:26:17.180 align:middle line:90% who generally don't respond. 00:26:17.180 --> 00:26:18.920 align:middle line:84% It turns out that most of those patients 00:26:18.920 --> 00:26:21.365 align:middle line:84% have translocations into a gene called MLL. 00:26:21.365 --> 00:26:24.530 align:middle line:90% 00:26:24.530 --> 00:26:28.178 align:middle line:84% But it's clinically of interest because these patients don't 00:26:28.178 --> 00:26:29.720 align:middle line:84% respond to conventional chemotherapy, 00:26:29.720 --> 00:26:31.490 align:middle line:84% and this just shows you that, using 00:26:31.490 --> 00:26:35.180 align:middle line:84% standard clinical criteria, they're hard to distinguish. 00:26:35.180 --> 00:26:38.750 align:middle line:84% So what if we take conventional ALL samples, 00:26:38.750 --> 00:26:44.090 align:middle line:84% these infant MLL rearranged leukemias, and some AML-- 00:26:44.090 --> 00:26:49.850 align:middle line:84% the myeloid leukemias and we apply an unsupervised learning 00:26:49.850 --> 00:26:52.190 align:middle line:90% approach? 00:26:52.190 --> 00:26:54.290 align:middle line:84% This happens to be principal component analysis, 00:26:54.290 --> 00:26:56.720 align:middle line:84% but could be your favorite clustering algorithm. 00:26:56.720 --> 00:26:59.740 align:middle line:90% 00:26:59.740 --> 00:27:01.810 align:middle line:90% What do you see here? 00:27:01.810 --> 00:27:03.250 align:middle line:90% AUDIENCE: What do you see? 00:27:03.250 --> 00:27:06.730 align:middle line:90% 00:27:06.730 --> 00:27:09.130 align:middle line:84% AUDIENCE: Three different classes. 00:27:09.130 --> 00:27:11.050 align:middle line:84% TODD GOLUB: Three different classes. 00:27:11.050 --> 00:27:12.130 align:middle line:90% Why do you say that? 00:27:12.130 --> 00:27:16.810 align:middle line:90% 00:27:16.810 --> 00:27:18.480 align:middle line:90% AUDIENCE: Jose, speak up. 00:27:18.480 --> 00:27:22.710 align:middle line:84% AUDIENCE: I mean, you're maximizing the separation 00:27:22.710 --> 00:27:25.850 align:middle line:84% so you see some separation-- some MLL [INAUDIBLE].. 00:27:25.850 --> 00:27:27.400 align:middle line:90% TODD GOLUB: OK, so right here. 00:27:27.400 --> 00:27:30.450 align:middle line:84% So, yes, you see three classes, but only 00:27:30.450 --> 00:27:33.570 align:middle line:84% if you have the colors filled in. 00:27:33.570 --> 00:27:38.550 align:middle line:84% So if you imagine this is just a group of leukemias, 00:27:38.550 --> 00:27:42.930 align:middle line:84% you might get the sense that there was something going on 00:27:42.930 --> 00:27:44.380 align:middle line:90% over here. 00:27:44.380 --> 00:27:47.550 align:middle line:84% But if you imagine these are all black, 00:27:47.550 --> 00:27:50.145 align:middle line:90% it's not so obvious, maybe. 00:27:50.145 --> 00:27:51.780 align:middle line:84% AUDIENCE: [INAUDIBLE] analyzed or have 00:27:51.780 --> 00:27:53.160 align:middle line:90% been optimized [INAUDIBLE]? 00:27:53.160 --> 00:27:55.770 align:middle line:84% TODD GOLUB: No, this is completely unsupervised. 00:27:55.770 --> 00:27:58.380 align:middle line:84% So that's the first point, is that these things often 00:27:58.380 --> 00:28:02.220 align:middle line:84% look clearer when you actually impose knowledge on them. 00:28:02.220 --> 00:28:07.080 align:middle line:84% Even though the structure here is done in an unsupervised way, 00:28:07.080 --> 00:28:09.570 align:middle line:84% you get the impression that it's really clean result 00:28:09.570 --> 00:28:12.302 align:middle line:84% if you superimpose knowledge afterwards. 00:28:12.302 --> 00:28:13.260 align:middle line:90% That's the first thing. 00:28:13.260 --> 00:28:15.052 align:middle line:84% But let's say, yes, they are three classes. 00:28:15.052 --> 00:28:18.900 align:middle line:84% And I think you can appreciate that. 00:28:18.900 --> 00:28:22.350 align:middle line:84% One question was, well, maybe these infants with the MLL-- 00:28:22.350 --> 00:28:24.930 align:middle line:90% rearranged genomes in green-- 00:28:24.930 --> 00:28:28.230 align:middle line:84% maybe they don't respond to therapy because they're babies 00:28:28.230 --> 00:28:31.990 align:middle line:84% and you know that this is a metabolic post-metabolism 00:28:31.990 --> 00:28:32.490 align:middle line:90% problem. 00:28:32.490 --> 00:28:37.150 align:middle line:84% And their leukemias are the same as the conventional ALLs shown 00:28:37.150 --> 00:28:39.118 align:middle line:90% in dark blue. 00:28:39.118 --> 00:28:41.160 align:middle line:84% This would argue that it's actually not the case, 00:28:41.160 --> 00:28:43.890 align:middle line:84% that they're fundamentally different leukemia. 00:28:43.890 --> 00:28:44.580 align:middle line:90% Is this helpful? 00:28:44.580 --> 00:28:49.880 align:middle line:90% 00:28:49.880 --> 00:28:53.560 align:middle line:84% So it's helpful, maybe, from a taxonomy perspective, 00:28:53.560 --> 00:28:57.040 align:middle line:84% but does it tell you what to do for these patients? 00:28:57.040 --> 00:29:01.150 align:middle line:84% So what if you wanted to gain some biological insight 00:29:01.150 --> 00:29:08.950 align:middle line:84% into what was different about these MLL green infant 00:29:08.950 --> 00:29:09.490 align:middle line:90% leukemias? 00:29:09.490 --> 00:29:10.270 align:middle line:90% What might you do? 00:29:10.270 --> 00:29:12.260 align:middle line:90% You've got these data. 00:29:12.260 --> 00:29:17.800 align:middle line:84% You see that those patients define a different class. 00:29:17.800 --> 00:29:20.942 align:middle line:90% What could you do? 00:29:20.942 --> 00:29:22.400 align:middle line:84% AUDIENCE: Any of you have any idea? 00:29:22.400 --> 00:29:24.240 align:middle line:90% What would you do with this? 00:29:24.240 --> 00:29:25.990 align:middle line:90% AUDIENCE: Inspect the component? 00:29:25.990 --> 00:29:30.100 align:middle line:84% Look for the things that have higher weight? 00:29:30.100 --> 00:29:32.372 align:middle line:84% Try to define biological function related 00:29:32.372 --> 00:29:34.920 align:middle line:90% to the [INAUDIBLE]? 00:29:34.920 --> 00:29:36.420 align:middle line:84% AUDIENCE: So the genes whose weights 00:29:36.420 --> 00:29:40.758 align:middle line:90% explain the most is separation. 00:29:40.758 --> 00:29:41.800 align:middle line:90% TODD GOLUB: That's right. 00:29:41.800 --> 00:29:46.300 align:middle line:90% 00:29:46.300 --> 00:29:47.320 align:middle line:90% You could do that. 00:29:47.320 --> 00:29:48.850 align:middle line:84% As it turns out, in this case, there 00:29:48.850 --> 00:29:54.098 align:middle line:84% are a lot of genes that actually have relatively equal weight. 00:29:54.098 --> 00:29:55.390 align:middle line:90% So you still have a large list. 00:29:55.390 --> 00:30:01.120 align:middle line:84% And the three principal components 00:30:01.120 --> 00:30:03.700 align:middle line:84% don't perfectly separate the classes. 00:30:03.700 --> 00:30:05.350 align:middle line:84% So you could go back and say, well now 00:30:05.350 --> 00:30:08.740 align:middle line:84% I believe that these MLL leukemias are 00:30:08.740 --> 00:30:11.290 align:middle line:90% a distinct entity. 00:30:11.290 --> 00:30:12.907 align:middle line:90% That would be reasonable. 00:30:12.907 --> 00:30:14.240 align:middle line:90% That's a reasonable thing to do. 00:30:14.240 --> 00:30:15.160 align:middle line:84% But the other way that you could do 00:30:15.160 --> 00:30:16.780 align:middle line:84% is to say, well, now this tells me 00:30:16.780 --> 00:30:18.572 align:middle line:84% that I believe that these MLL leukemias are 00:30:18.572 --> 00:30:19.540 align:middle line:90% a distinct entity. 00:30:19.540 --> 00:30:22.450 align:middle line:84% Now let's use supervised learning types of methods 00:30:22.450 --> 00:30:28.000 align:middle line:84% to identify the genes that are most correlated with the class 00:30:28.000 --> 00:30:31.150 align:middle line:84% of interest-- for example, high in the MLL 00:30:31.150 --> 00:30:34.630 align:middle line:90% class versus the others. 00:30:34.630 --> 00:30:36.610 align:middle line:84% That would be a straightforward thing to do. 00:30:36.610 --> 00:30:40.640 align:middle line:84% So you could rank the genes according to that distinction. 00:30:40.640 --> 00:30:44.530 align:middle line:84% So we did that and did what I think-- 00:30:44.530 --> 00:30:45.570 align:middle line:90% yeah. 00:30:45.570 --> 00:30:49.945 align:middle line:84% AUDIENCE: [INAUDIBLE] In the first case, 00:30:49.945 --> 00:30:57.390 align:middle line:84% [INAUDIBLE] compare between the different classifications-- 00:30:57.390 --> 00:30:59.432 align:middle line:84% the difference-- the different genes-- 00:30:59.432 --> 00:31:00.890 align:middle line:84% [INAUDIBLE] expression or whatever? 00:31:00.890 --> 00:31:04.330 align:middle line:84% And the second case, you're just thinking comparing the one 00:31:04.330 --> 00:31:04.955 align:middle line:90% classification? 00:31:04.955 --> 00:31:06.280 align:middle line:90% Is that correct? 00:31:06.280 --> 00:31:08.535 align:middle line:90% AUDIENCE: No. 00:31:08.535 --> 00:31:09.910 align:middle line:84% TODD GOLUB: No, I think it's more 00:31:09.910 --> 00:31:13.060 align:middle line:84% that, if you didn't have these colors to look at and you said, 00:31:13.060 --> 00:31:15.660 align:middle line:90% ah, there's some structure here. 00:31:15.660 --> 00:31:16.660 align:middle line:90% I don't know what it is. 00:31:16.660 --> 00:31:19.420 align:middle line:84% What's the biological basis of this structure? 00:31:19.420 --> 00:31:22.600 align:middle line:84% Looking at the weights of the genes that 00:31:22.600 --> 00:31:26.170 align:middle line:84% are driving this distinction would be a reasonable thing 00:31:26.170 --> 00:31:27.730 align:middle line:90% to do. 00:31:27.730 --> 00:31:29.860 align:middle line:84% In this case, we had a specific question. 00:31:29.860 --> 00:31:32.770 align:middle line:84% Are these leukemias unique, or are they 00:31:32.770 --> 00:31:34.420 align:middle line:90% admixed with the others? 00:31:34.420 --> 00:31:37.477 align:middle line:84% Having determined that they are unique, 00:31:37.477 --> 00:31:39.310 align:middle line:84% it's a little bit cleaner to say, all right, 00:31:39.310 --> 00:31:44.920 align:middle line:84% let's use supervised methods to find the genes that distinguish 00:31:44.920 --> 00:31:46.480 align:middle line:90% one class from the other. 00:31:46.480 --> 00:31:49.210 align:middle line:84% Of course, if you had perfect separation, 00:31:49.210 --> 00:31:52.550 align:middle line:84% it would reduce to the same experiment. 00:31:52.550 --> 00:31:55.090 align:middle line:84% But because it's imperfect, there 00:31:55.090 --> 00:31:57.310 align:middle line:84% are some advantages to using class labels here. 00:31:57.310 --> 00:32:01.450 align:middle line:84% I should mention also that-- you see this blue guy here sitting 00:32:01.450 --> 00:32:02.830 align:middle line:90% in a sea of green? 00:32:02.830 --> 00:32:05.830 align:middle line:84% So this is a patient that, based on gene expression, 00:32:05.830 --> 00:32:08.230 align:middle line:84% one would predict to be MLL rearranged. 00:32:08.230 --> 00:32:11.960 align:middle line:84% But the clinical record for this patient study was not. 00:32:11.960 --> 00:32:15.170 align:middle line:84% But when we went back and actually looked at this, 00:32:15.170 --> 00:32:16.810 align:middle line:84% it turns out that there was a missed 00:32:16.810 --> 00:32:20.380 align:middle line:84% translocation into the MLL gene that you could recover by FISH. 00:32:20.380 --> 00:32:26.650 align:middle line:84% So this is not a public health menace-- 00:32:26.650 --> 00:32:29.260 align:middle line:84% diagnosing these leukemias properly. 00:32:29.260 --> 00:32:42.220 align:middle line:84% But there are examples of missed diagnoses that I think can be-- 00:32:42.220 --> 00:32:44.920 align:middle line:84% I think looking at these multi-parameter gene expression 00:32:44.920 --> 00:32:49.600 align:middle line:84% readouts can serve as a unifier, an integrator of lots 00:32:49.600 --> 00:32:51.410 align:middle line:90% of upstream genetic activity. 00:32:51.410 --> 00:32:55.330 align:middle line:84% And so I think the power to detect those upstream events is 00:32:55.330 --> 00:32:59.090 align:middle line:84% going to be higher when you look at some downstream pattern, 00:32:59.090 --> 00:33:00.940 align:middle line:84% such as an RNA pattern, as opposed 00:33:00.940 --> 00:33:03.310 align:middle line:84% to developing specific tests for each 00:33:03.310 --> 00:33:06.490 align:middle line:84% of the individual genetic abnormalities 00:33:06.490 --> 00:33:09.100 align:middle line:84% that could cause the same phenotype because, in the end, 00:33:09.100 --> 00:33:12.310 align:middle line:84% all you care about is knowing whether the molecular program 00:33:12.310 --> 00:33:15.370 align:middle line:90% has been activated. 00:33:15.370 --> 00:33:18.550 align:middle line:84% So you rank the genes according to this distinction 00:33:18.550 --> 00:33:23.050 align:middle line:84% and just start at the top of the list-- 00:33:23.050 --> 00:33:27.070 align:middle line:84% here is a gene that was top of the list of 12,600 or whatever 00:33:27.070 --> 00:33:30.760 align:middle line:90% that we're on on the list. 00:33:30.760 --> 00:33:35.950 align:middle line:84% And any time a tyrosine kinase rears its head 00:33:35.950 --> 00:33:38.955 align:middle line:90% in a cancer classification-- 00:33:38.955 --> 00:33:41.080 align:middle line:84% cancer biology experiment, you pay attention to it, 00:33:41.080 --> 00:33:44.110 align:middle line:84% particularly given the Gleevec story. 00:33:44.110 --> 00:33:45.510 align:middle line:90% So what do you think about this? 00:33:45.510 --> 00:33:48.220 align:middle line:90% I tell you, oh, look at this. 00:33:48.220 --> 00:33:51.580 align:middle line:90% The RNA level of a kinase-- 00:33:51.580 --> 00:33:53.230 align:middle line:84% so a receptor tyrosine kinase called 00:33:53.230 --> 00:34:01.180 align:middle line:84% FLT3 is characteristically high in the MLLs, 00:34:01.180 --> 00:34:02.540 align:middle line:90% compared to the others. 00:34:02.540 --> 00:34:05.860 align:middle line:84% What do you think of that, in terms of therapeutic-- 00:34:05.860 --> 00:34:10.000 align:middle line:84% potential therapeutic significance? 00:34:10.000 --> 00:34:14.151 align:middle line:84% AUDIENCE: What would you do with that? 00:34:14.151 --> 00:34:15.117 align:middle line:90% AUDIENCE: [INAUDIBLE]. 00:34:15.117 --> 00:34:17.500 align:middle line:84% Or maybe this already popped out, 00:34:17.500 --> 00:34:20.159 align:middle line:84% but can't you check the levels in the other two 00:34:20.159 --> 00:34:22.120 align:middle line:90% classifications? 00:34:22.120 --> 00:34:24.070 align:middle line:84% AUDIENCE: Because these [INAUDIBLE] slower. 00:34:24.070 --> 00:34:24.820 align:middle line:90% TODD GOLUB: Right. 00:34:24.820 --> 00:34:26.770 align:middle line:84% So we define this list by virtue of the fact 00:34:26.770 --> 00:34:31.389 align:middle line:84% that it's high in the MLLs compared to the other two 00:34:31.389 --> 00:34:32.590 align:middle line:90% combined. 00:34:32.590 --> 00:34:35.480 align:middle line:84% AUDIENCE: So how do we get the therapeutic K out of this? 00:34:35.480 --> 00:34:36.897 align:middle line:84% AUDIENCE: For the patients patient 00:34:36.897 --> 00:34:40.170 align:middle line:90% are responsive [INAUDIBLE] 00:34:40.170 --> 00:34:44.346 align:middle line:90% 00:34:44.346 --> 00:34:46.489 align:middle line:84% Are they classified as far as the response, 00:34:46.489 --> 00:34:48.739 align:middle line:90% the FLT3 inhibitor? 00:34:48.739 --> 00:34:50.150 align:middle line:90% TODD GOLUB: To a FLT3 inhibitor. 00:34:50.150 --> 00:34:50.639 align:middle line:90% AUDIENCE: Yeah. 00:34:50.639 --> 00:34:52.880 align:middle line:84% TODD GOLUB: So you want to treat patients with a FLT3 inhibitor? 00:34:52.880 --> 00:34:55.429 align:middle line:84% Well, that's not an FDA-approved drug, so you can't do that. 00:34:55.429 --> 00:34:58.817 align:middle line:90% 00:34:58.817 --> 00:35:00.400 align:middle line:84% AUDIENCE: OK, so what else do we have? 00:35:00.400 --> 00:35:04.693 align:middle line:90% 00:35:04.693 --> 00:35:07.555 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:35:07.555 --> 00:35:14.340 align:middle line:90% 00:35:14.340 --> 00:35:15.340 align:middle line:90% TODD GOLUB: So there is. 00:35:15.340 --> 00:35:18.700 align:middle line:90% 00:35:18.700 --> 00:35:23.590 align:middle line:84% The hypothesis would be that MLL leukemia 00:35:23.590 --> 00:35:29.560 align:middle line:84% cells are dependent on FLT3 kinase activity for survival. 00:35:29.560 --> 00:35:34.600 align:middle line:84% If that's not the case, then you don't care. 00:35:34.600 --> 00:35:37.000 align:middle line:84% Unless that's the case, the overexpression of this thing 00:35:37.000 --> 00:35:40.660 align:middle line:84% is totally irrelevant from a therapeutic perspective. 00:35:40.660 --> 00:35:43.210 align:middle line:84% So you could do that genetically-- for example, 00:35:43.210 --> 00:35:47.200 align:middle line:84% using RNA interference to knock down the expression, 00:35:47.200 --> 00:35:49.180 align:middle line:84% or you could do it pharmacologically, 00:35:49.180 --> 00:35:52.330 align:middle line:84% if there was a compound in development-- not yet a drug-- 00:35:52.330 --> 00:35:58.330 align:middle line:84% but a compound in development that inhibits kinase activity. 00:35:58.330 --> 00:36:00.700 align:middle line:84% And so that's what this experiment is. 00:36:00.700 --> 00:36:04.780 align:middle line:90% AUDIENCE: So [INAUDIBLE] 00:36:04.780 --> 00:36:05.280 align:middle line:90% 00:36:05.280 --> 00:36:07.120 align:middle line:84% Doing RNA interference-- is that something 00:36:07.120 --> 00:36:09.790 align:middle line:84% that you can just do in a person? 00:36:09.790 --> 00:36:13.090 align:middle line:84% TODD GOLUB: You can't do it in vivo in a person, 00:36:13.090 --> 00:36:16.930 align:middle line:84% but you can do it in human-derived cell lines. 00:36:16.930 --> 00:36:20.840 align:middle line:84% AUDIENCE: So clinically, that would not be [INAUDIBLE].. 00:36:20.840 --> 00:36:22.840 align:middle line:84% AUDIENCE: You could test the hypothesis that you 00:36:22.840 --> 00:36:25.190 align:middle line:90% want to go down that path. 00:36:25.190 --> 00:36:26.110 align:middle line:90% AUDIENCE: I see. 00:36:26.110 --> 00:36:27.152 align:middle line:90% TODD GOLUB: That's right. 00:36:27.152 --> 00:36:29.440 align:middle line:84% Now, you can make the argument that, well, 00:36:29.440 --> 00:36:31.990 align:middle line:84% doing these things in cell lines in mice-- 00:36:31.990 --> 00:36:34.280 align:middle line:90% that's not real disease. 00:36:34.280 --> 00:36:36.820 align:middle line:84% And so I don't care what any of this stuff shows. 00:36:36.820 --> 00:36:42.190 align:middle line:84% But still, if your hypothesis is that a given gene-- 00:36:42.190 --> 00:36:46.510 align:middle line:84% the overexpression of a given gene is important 00:36:46.510 --> 00:36:48.010 align:middle line:84% and you do the experiments to ablate 00:36:48.010 --> 00:36:51.910 align:middle line:84% the expression of that gene and nothing happens, 00:36:51.910 --> 00:36:55.280 align:middle line:84% that should deflate your enthusiasm a little bit. 00:36:55.280 --> 00:36:56.930 align:middle line:84% So here's the experiment, though. 00:36:56.930 --> 00:37:00.670 align:middle line:84% Here, now, taking patient-derived human cells 00:37:00.670 --> 00:37:05.380 align:middle line:84% that have been engineered to express firefly luciferase 00:37:05.380 --> 00:37:07.820 align:middle line:90% genes so that they glow-- 00:37:07.820 --> 00:37:10.420 align:middle line:84% and you can monitor in vivo tumor burden. 00:37:10.420 --> 00:37:13.120 align:middle line:84% So here, mice that, on week one, you 00:37:13.120 --> 00:37:20.980 align:middle line:84% inject in the tail vein infant leukemia-derived tumor cells. 00:37:20.980 --> 00:37:23.890 align:middle line:84% And you see, over 4 weeks time, the amount 00:37:23.890 --> 00:37:29.020 align:middle line:84% of luciferase activity increases as the cells grow 00:37:29.020 --> 00:37:32.260 align:middle line:84% and the mice start to die around week 4. 00:37:32.260 --> 00:37:38.050 align:middle line:84% And here is a cohort of mice also injected but treated 00:37:38.050 --> 00:37:45.400 align:middle line:84% with a drug once a day by mouth that functions as a FLT3 kinase 00:37:45.400 --> 00:37:46.550 align:middle line:90% inhibitor. 00:37:46.550 --> 00:37:50.170 align:middle line:84% And you can see that the development of the leukemia 00:37:50.170 --> 00:37:53.680 align:middle line:84% is significantly abrogated, which, 00:37:53.680 --> 00:37:55.180 align:middle line:84% at least to our first approximation, 00:37:55.180 --> 00:38:00.040 align:middle line:84% validates the hypothesis that FLT3 overexpression is not just 00:38:00.040 --> 00:38:01.660 align:middle line:84% a diagnostic marker of this class, 00:38:01.660 --> 00:38:05.590 align:middle line:84% but it's actually a potential therapeutic target. 00:38:05.590 --> 00:38:09.723 align:middle line:84% And so, based on this and some other preclinical data, 00:38:09.723 --> 00:38:11.140 align:middle line:84% the clinical trial that you wanted 00:38:11.140 --> 00:38:14.290 align:middle line:84% to do with a FLT3 inhibitor is being 00:38:14.290 --> 00:38:16.630 align:middle line:90% planned to treat patients. 00:38:16.630 --> 00:38:17.320 align:middle line:90% Yeah. 00:38:17.320 --> 00:38:20.200 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:38:20.200 --> 00:38:21.640 align:middle line:90% 00:38:21.640 --> 00:38:26.530 align:middle line:84% TODD GOLUB: So the cells are infected with a retrovirus that 00:38:26.530 --> 00:38:33.730 align:middle line:84% contains the cDNA for the firefly luciferase gene so 00:38:33.730 --> 00:38:40.000 align:middle line:84% that, if you inject these mice with the compound luciferin, 00:38:40.000 --> 00:38:46.330 align:middle line:84% they will emit the same enzyme that fireflies do, 00:38:46.330 --> 00:38:48.850 align:middle line:90% and they will glow. 00:38:48.850 --> 00:38:51.050 align:middle line:84% So usually, this is done in vitro in test tubes. 00:38:51.050 --> 00:38:54.580 align:middle line:84% But here, you introduce it into the cells and the animal 00:38:54.580 --> 00:38:57.730 align:middle line:90% so that you can monitor. 00:38:57.730 --> 00:38:59.110 align:middle line:84% What you used to have to do would 00:38:59.110 --> 00:39:02.350 align:middle line:84% be to inject a whole bunch of mice, kill some of them 00:39:02.350 --> 00:39:05.290 align:middle line:84% here, kill some of them here, and then examine 00:39:05.290 --> 00:39:08.870 align:middle line:84% the bone marrows to evaluate the progression of the disease. 00:39:08.870 --> 00:39:11.740 align:middle line:84% What's nice here is that you can follow a cohort of mice 00:39:11.740 --> 00:39:12.850 align:middle line:90% non-invasively. 00:39:12.850 --> 00:39:14.070 align:middle line:84% AUDIENCE: Let me ask a dumb question, because I've never 00:39:14.070 --> 00:39:14.980 align:middle line:90% actually done this. 00:39:14.980 --> 00:39:16.522 align:middle line:84% When you actually look at these mice, 00:39:16.522 --> 00:39:19.450 align:middle line:84% can you tell that they're fluorescent? 00:39:19.450 --> 00:39:20.410 align:middle line:90% TODD GOLUB: No. 00:39:20.410 --> 00:39:21.295 align:middle line:90% No. 00:39:21.295 --> 00:39:23.620 align:middle line:84% AUDIENCE: They don't actually look-- no. 00:39:23.620 --> 00:39:26.290 align:middle line:84% TODD GOLUB: No, you need to use a special device that 00:39:26.290 --> 00:39:31.150 align:middle line:84% can measure, I think, in the near-infrared range. 00:39:31.150 --> 00:39:34.000 align:middle line:84% There are green fluorescent protein mice 00:39:34.000 --> 00:39:35.740 align:middle line:90% that actually do glow. 00:39:35.740 --> 00:39:37.240 align:middle line:84% And you can tell that they're green. 00:39:37.240 --> 00:39:40.870 align:middle line:90% 00:39:40.870 --> 00:39:44.810 align:middle line:84% AUDIENCE: The mice that we use-- are they immune 00:39:44.810 --> 00:39:46.480 align:middle line:84% [INAUDIBLE],, which means you don't 00:39:46.480 --> 00:39:49.400 align:middle line:90% have a massive immune response? 00:39:49.400 --> 00:39:52.030 align:middle line:84% TODD GOLUB: So you have to do this in immunodeficient mice 00:39:52.030 --> 00:39:54.655 align:middle line:84% so that they don't reject the human tumors. 00:39:54.655 --> 00:39:58.155 align:middle line:90% 00:39:58.155 --> 00:39:59.530 align:middle line:84% AUDIENCE: Does that factor at all 00:39:59.530 --> 00:40:05.580 align:middle line:84% in your determination of the degree of proposed-- 00:40:05.580 --> 00:40:08.320 align:middle line:84% the spread of the tumor cells and whatnot 00:40:08.320 --> 00:40:11.410 align:middle line:90% because they're just there? 00:40:11.410 --> 00:40:15.320 align:middle line:84% The immune system can direct attacks against-- 00:40:15.320 --> 00:40:17.780 align:middle line:84% so when you're considering these experiments and saying, 00:40:17.780 --> 00:40:21.430 align:middle line:84% OK, I see this spread across the entire mouse 00:40:21.430 --> 00:40:24.233 align:middle line:84% and this level of [INAUDIBLE] how do you factor that in? 00:40:24.233 --> 00:40:25.150 align:middle line:90% TODD GOLUB: You don't. 00:40:25.150 --> 00:40:29.560 align:middle line:84% You factor that in by saying that there are many things that 00:40:29.560 --> 00:40:35.260 align:middle line:84% are occurring in these models that don't recapitulate 00:40:35.260 --> 00:40:36.460 align:middle line:90% what happens in the mouse. 00:40:36.460 --> 00:40:40.660 align:middle line:84% Most people don't get cancer by having intravenous injection 00:40:40.660 --> 00:40:42.400 align:middle line:90% of a tumor into them. 00:40:42.400 --> 00:40:45.820 align:middle line:84% Most patients have an immune system. 00:40:45.820 --> 00:40:48.955 align:middle line:84% So I think it's just one of the limitations that-- 00:40:48.955 --> 00:40:51.970 align:middle line:90% 00:40:51.970 --> 00:40:54.070 align:middle line:84% it would not be worth the time and expense 00:40:54.070 --> 00:40:57.220 align:middle line:84% to have a drug development project 00:40:57.220 --> 00:41:00.430 align:middle line:84% around every little inkling that comes out 00:41:00.430 --> 00:41:01.555 align:middle line:90% of a microarray experiment. 00:41:01.555 --> 00:41:04.810 align:middle line:84% So you need to do something, even though it's deficient 00:41:04.810 --> 00:41:07.660 align:middle line:84% in many ways-- and you've hit on some of them-- 00:41:07.660 --> 00:41:11.680 align:middle line:84% to say, is this interesting or is it not? 00:41:11.680 --> 00:41:13.420 align:middle line:84% I think-- not yet at the point where 00:41:13.420 --> 00:41:15.700 align:middle line:84% one can do this entirely computationally 00:41:15.700 --> 00:41:19.240 align:middle line:90% and have any kind of confidence. 00:41:19.240 --> 00:41:22.120 align:middle line:84% That being said, these so-called xenograft models, 00:41:22.120 --> 00:41:28.090 align:middle line:84% where you put a human tumor into a mouse model, 00:41:28.090 --> 00:41:33.220 align:middle line:84% are not particularly predictive of efficacy 00:41:33.220 --> 00:41:35.920 align:middle line:84% of a drug in the human clinical trial. 00:41:35.920 --> 00:41:38.170 align:middle line:84% But in the absence of anything better, it's 00:41:38.170 --> 00:41:39.730 align:middle line:90% still what most people do first. 00:41:39.730 --> 00:41:41.510 align:middle line:84% AUDIENCE: [INAUDIBLE] more robust, in other words, 00:41:41.510 --> 00:41:43.135 align:middle line:84% if there's no effect on the xenografts, 00:41:43.135 --> 00:41:45.340 align:middle line:84% then you're really a loser if you 00:41:45.340 --> 00:41:46.840 align:middle line:90% go to the human [INAUDIBLE]? 00:41:46.840 --> 00:41:47.920 align:middle line:90% TODD GOLUB: No. 00:41:47.920 --> 00:41:51.490 align:middle line:84% If anything, it's the opposite, that, if you 00:41:51.490 --> 00:41:54.640 align:middle line:84% show some activity in the xenografts, 00:41:54.640 --> 00:41:56.320 align:middle line:90% you often see activity. 00:41:56.320 --> 00:41:59.150 align:middle line:90% But failure to see activity-- 00:41:59.150 --> 00:42:01.000 align:middle line:84% failure to see activity in xenografts 00:42:01.000 --> 00:42:02.998 align:middle line:90% is not particularly-- 00:42:02.998 --> 00:42:04.540 align:middle line:84% particularly for molecularly targeted 00:42:04.540 --> 00:42:09.160 align:middle line:84% therapies, where it may be that you can show in the mouse 00:42:09.160 --> 00:42:11.660 align:middle line:84% that you've really shut down the pathway. 00:42:11.660 --> 00:42:15.970 align:middle line:84% Let's say you've inhibited FLT3 completely. 00:42:15.970 --> 00:42:18.970 align:middle line:84% Drug companies are starting to use graphs in that way to say, 00:42:18.970 --> 00:42:23.350 align:middle line:84% all the mouse is is a test tube so that I can ask-- 00:42:23.350 --> 00:42:26.860 align:middle line:84% have I inhibited FLT3 enzymatic activity? 00:42:26.860 --> 00:42:28.280 align:middle line:90% Yes or no. 00:42:28.280 --> 00:42:32.470 align:middle line:84% If I have and I believe that FLT3 is a good target, 00:42:32.470 --> 00:42:35.200 align:middle line:84% I don't care whether the tumor's actually shrunk or not. 00:42:35.200 --> 00:42:39.107 align:middle line:84% I'm going to bring it forward to clinical trial. 00:42:39.107 --> 00:42:40.690 align:middle line:84% But you need something to convince you 00:42:40.690 --> 00:42:42.640 align:middle line:84% that the target is reasonable, yeah. 00:42:42.640 --> 00:42:44.410 align:middle line:84% AUDIENCE: An interesting [INAUDIBLE].. 00:42:44.410 --> 00:42:50.905 align:middle line:84% So the small spots in the [INAUDIBLE] three mouses are-- 00:42:50.905 --> 00:42:54.248 align:middle line:84% the purple ones are the xenographs, right? 00:42:54.248 --> 00:42:55.748 align:middle line:84% The purple spots are the xenographs. 00:42:55.748 --> 00:42:57.230 align:middle line:90% AUDIENCE: Are the tumors. 00:42:57.230 --> 00:42:58.100 align:middle line:90% TODD GOLUB: Yeah. 00:42:58.100 --> 00:43:00.410 align:middle line:84% AUDIENCE: But that means the spots are all very fixed 00:43:00.410 --> 00:43:02.777 align:middle line:90% in specific areas of the mouse. 00:43:02.777 --> 00:43:04.610 align:middle line:84% TODD GOLUB: Yeah, I think you see them there 00:43:04.610 --> 00:43:06.770 align:middle line:90% because the cells-- 00:43:06.770 --> 00:43:10.100 align:middle line:84% they're injected intravenously in the tail vein, 00:43:10.100 --> 00:43:12.170 align:middle line:84% but they home to the bone marrow and you're 00:43:12.170 --> 00:43:16.010 align:middle line:84% seeing large bone marrow cavities, which is why you 00:43:16.010 --> 00:43:18.260 align:middle line:90% see them over the flank there. 00:43:18.260 --> 00:43:20.928 align:middle line:90% 00:43:20.928 --> 00:43:22.220 align:middle line:90% I think I'm going to skip this. 00:43:22.220 --> 00:43:32.320 align:middle line:90% 00:43:32.320 --> 00:43:34.390 align:middle line:90% OK, did you talk about this? 00:43:34.390 --> 00:43:35.050 align:middle line:90% AUDIENCE: No. 00:43:35.050 --> 00:43:35.758 align:middle line:90% TODD GOLUB: Good. 00:43:35.758 --> 00:43:38.800 align:middle line:90% 00:43:38.800 --> 00:43:43.580 align:middle line:84% So when you do these experiments, 00:43:43.580 --> 00:43:47.290 align:middle line:84% the data usually present you with two-- 00:43:47.290 --> 00:43:49.510 align:middle line:84% you either have one or two problems 00:43:49.510 --> 00:43:52.570 align:middle line:84% after you do all this appropriate correction 00:43:52.570 --> 00:43:56.602 align:middle line:84% for multiple hypothesis testing that I told you about. 00:43:56.602 --> 00:43:58.060 align:middle line:84% Despite having done that, you still 00:43:58.060 --> 00:44:00.730 align:middle line:84% have a list of genes that's too impossibly 00:44:00.730 --> 00:44:04.030 align:middle line:84% long to bring biological understanding to, 00:44:04.030 --> 00:44:11.048 align:middle line:84% or you've corrected away everything 00:44:11.048 --> 00:44:13.090 align:middle line:84% and you have the impression that there's actually 00:44:13.090 --> 00:44:18.430 align:middle line:84% nothing that is differentially expressed in your two cases. 00:44:18.430 --> 00:44:21.520 align:middle line:84% And so let me show some of the more recent approaches 00:44:21.520 --> 00:44:23.770 align:middle line:84% to dealing with this because it's really substantially 00:44:23.770 --> 00:44:26.650 align:middle line:84% changing our thinking about how to do 00:44:26.650 --> 00:44:27.880 align:middle line:90% these kinds of experiments. 00:44:27.880 --> 00:44:33.010 align:middle line:84% So this is not a cancer example but a diabetes experiment 00:44:33.010 --> 00:44:37.570 align:middle line:84% where there were patients who were either-- 00:44:37.570 --> 00:44:42.700 align:middle line:84% adult patients who either had type 2 diabetes or they 00:44:42.700 --> 00:44:47.650 align:middle line:84% were normal, as defined by having a normal glucose 00:44:47.650 --> 00:44:48.760 align:middle line:90% tolerance test. 00:44:48.760 --> 00:44:53.380 align:middle line:84% And they underwent voluntary skeletal muscle biopsies 00:44:53.380 --> 00:44:56.590 align:middle line:90% under a euglycemic clamp. 00:44:56.590 --> 00:44:59.330 align:middle line:84% 18 of these patients, 17 of these patients-- 00:44:59.330 --> 00:45:01.270 align:middle line:90% it's a simple two-class problem. 00:45:01.270 --> 00:45:03.940 align:middle line:84% Get the expression data to identify 00:45:03.940 --> 00:45:05.440 align:middle line:84% those genes that are differentially 00:45:05.440 --> 00:45:08.470 align:middle line:90% expressed in these two classes. 00:45:08.470 --> 00:45:10.210 align:middle line:84% Do the appropriate permutation testing 00:45:10.210 --> 00:45:13.330 align:middle line:84% to make sure that you correct for multiple hypothesis 00:45:13.330 --> 00:45:13.960 align:middle line:90% testing. 00:45:13.960 --> 00:45:19.390 align:middle line:84% And here's the result. Nothing meets significance. 00:45:19.390 --> 00:45:21.490 align:middle line:84% So out of the 20,000 genes on the array, 00:45:21.490 --> 00:45:28.030 align:middle line:84% even the top-ranked gene doesn't meet statistical significance. 00:45:28.030 --> 00:45:31.370 align:middle line:90% Possible that this is the case. 00:45:31.370 --> 00:45:34.450 align:middle line:84% But the question is, are there other ways 00:45:34.450 --> 00:45:38.380 align:middle line:84% that you might go about recovering 00:45:38.380 --> 00:45:40.220 align:middle line:90% a biological story here? 00:45:40.220 --> 00:45:45.400 align:middle line:84% And so the way that Vamsi Mootha and graduate student, Arvind 00:45:45.400 --> 00:45:54.610 align:middle line:84% Subramanian took to this was to define groups of genes or gene 00:45:54.610 --> 00:46:00.430 align:middle line:84% sets whose activity as a collection of genes 00:46:00.430 --> 00:46:03.550 align:middle line:84% could be interrogated in these data sets. 00:46:03.550 --> 00:46:05.920 align:middle line:84% And we could have a rich discussion about, 00:46:05.920 --> 00:46:07.593 align:middle line:84% how does one define such gene sets? 00:46:07.593 --> 00:46:09.260 align:middle line:84% You could do it based on the literature. 00:46:09.260 --> 00:46:14.260 align:middle line:84% So ask Zach what genes are important in some pathway 00:46:14.260 --> 00:46:16.450 align:middle line:90% that he knows something about. 00:46:16.450 --> 00:46:17.350 align:middle line:90% That could be a list. 00:46:17.350 --> 00:46:21.100 align:middle line:84% Or you could say, we don't trust that about-- 00:46:21.100 --> 00:46:22.870 align:middle line:90% that'll bring Zach's bias. 00:46:22.870 --> 00:46:25.960 align:middle line:84% We're not interested in that or anyone else's intuition. 00:46:25.960 --> 00:46:30.250 align:middle line:84% Let's just experimentally derive lists of genes 00:46:30.250 --> 00:46:32.050 align:middle line:84% by one way or the other-- perturb cells, 00:46:32.050 --> 00:46:35.800 align:middle line:84% get the gene expression change, and that makes a gene set. 00:46:35.800 --> 00:46:38.710 align:middle line:84% And you can collect as many of these as you could stand. 00:46:38.710 --> 00:46:44.520 align:middle line:84% For these experiments, we made 150 of these gene sets. 00:46:44.520 --> 00:46:48.390 align:middle line:84% AUDIENCE: Before we go on, [INAUDIBLE] twice on the slide. 00:46:48.390 --> 00:46:49.740 align:middle line:90% TODD GOLUB: Oh, yeah. 00:46:49.740 --> 00:46:51.577 align:middle line:90% AUDIENCE: So he's enriched one. 00:46:51.577 --> 00:46:52.660 align:middle line:90% TODD GOLUB: He's enriched. 00:46:52.660 --> 00:46:53.452 align:middle line:90% Yeah, he should be. 00:46:53.452 --> 00:46:56.850 align:middle line:90% 00:46:56.850 --> 00:46:58.930 align:middle line:90% So then how do you do this? 00:46:58.930 --> 00:47:04.110 align:middle line:84% So first thing you do is rank all the genes on the array-- 00:47:04.110 --> 00:47:07.500 align:middle line:84% 1 through 20,000-- according to how well they're correlated 00:47:07.500 --> 00:47:09.580 align:middle line:90% with the distinction. 00:47:09.580 --> 00:47:11.310 align:middle line:84% I already told you this top one-- 00:47:11.310 --> 00:47:14.250 align:middle line:84% even that one doesn't meet significance as a single gene. 00:47:14.250 --> 00:47:18.240 align:middle line:84% And then you interrogate each of these gene sets and ask, 00:47:18.240 --> 00:47:18.990 align:middle line:90% are they enriched? 00:47:18.990 --> 00:47:23.440 align:middle line:84% And so here would be an example of a hypothetical gene set, 00:47:23.440 --> 00:47:27.330 align:middle line:84% so each gene in the gene set of a dozen genes or whatever, 00:47:27.330 --> 00:47:33.990 align:middle line:84% that is not enriched towards the top of this rank-ordered set 00:47:33.990 --> 00:47:38.700 align:middle line:84% list, whereas here is a hypothetical gene set. 00:47:38.700 --> 00:47:40.650 align:middle line:90% It's not perfect. 00:47:40.650 --> 00:47:43.300 align:middle line:84% But it's non-randomly distributed on this rank order 00:47:43.300 --> 00:47:43.800 align:middle line:90% list. 00:47:43.800 --> 00:47:45.730 align:middle line:90% It's enriched towards the top. 00:47:45.730 --> 00:47:47.730 align:middle line:84% AUDIENCE: Do you see this as a similar operation 00:47:47.730 --> 00:47:50.055 align:middle line:90% to the following-- 00:47:50.055 --> 00:47:52.680 align:middle line:84% there's a bunch of proteins that they would look at [INAUDIBLE] 00:47:52.680 --> 00:47:55.770 align:middle line:84% giving microarray result at gene ontology, and they'd say, 00:47:55.770 --> 00:48:04.660 align:middle line:84% what classes of gene ontology are overrepresented given-- 00:48:04.660 --> 00:48:07.020 align:middle line:90% in this set of genes? 00:48:07.020 --> 00:48:08.310 align:middle line:90% TODD GOLUB: Yes. 00:48:08.310 --> 00:48:12.210 align:middle line:84% So you can define these gene sets based on a gene ontology 00:48:12.210 --> 00:48:12.930 align:middle line:90% annotation. 00:48:12.930 --> 00:48:13.900 align:middle line:90% That's an example. 00:48:13.900 --> 00:48:15.870 align:middle line:84% The important part is to make sure 00:48:15.870 --> 00:48:20.890 align:middle line:84% that you appropriately correct for testing all the gene sets. 00:48:20.890 --> 00:48:24.420 align:middle line:84% So now, instead of 22,000 genes, we have 150 gene sets. 00:48:24.420 --> 00:48:29.100 align:middle line:84% But that you should think of 150 hypotheses. 00:48:29.100 --> 00:48:35.970 align:middle line:84% So you should do the same permutation type of testing 00:48:35.970 --> 00:48:38.260 align:middle line:84% and say, if I randomize-- in this case-- 00:48:38.260 --> 00:48:40.320 align:middle line:84% the diabetes versus normal distinction, 00:48:40.320 --> 00:48:43.950 align:middle line:84% is my favorite gene ontology class still enriched? 00:48:43.950 --> 00:48:47.310 align:middle line:84% And that's what some of the current approaches to that kind 00:48:47.310 --> 00:48:49.180 align:middle line:90% of annotation don't do. 00:48:49.180 --> 00:48:51.180 align:middle line:84% And so you can codify this in something 00:48:51.180 --> 00:48:53.250 align:middle line:84% called a Kolmogorov-Smirnov statistic. 00:48:53.250 --> 00:48:54.190 align:middle line:90% It doesn't matter. 00:48:54.190 --> 00:48:56.182 align:middle line:84% You can come up with an enrichment score 00:48:56.182 --> 00:48:56.890 align:middle line:90% for these things. 00:48:56.890 --> 00:48:58.320 align:middle line:84% And if you do that in this example 00:48:58.320 --> 00:49:04.100 align:middle line:84% you essentially get 1 gene set which meets-- 00:49:04.100 --> 00:49:06.400 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:49:06.400 --> 00:49:08.450 align:middle line:90% 00:49:08.450 --> 00:49:13.520 align:middle line:84% TODD GOLUB: Which gets quite high statistical significance 00:49:13.520 --> 00:49:15.950 align:middle line:90% for a set of genes. 00:49:15.950 --> 00:49:19.010 align:middle line:90% So how do you reconcile this? 00:49:19.010 --> 00:49:22.640 align:middle line:84% How do you reconcile this thing and this thing? 00:49:22.640 --> 00:49:23.487 align:middle line:90% How could that be? 00:49:23.487 --> 00:49:24.570 align:middle line:90% AUDIENCE: That's my point. 00:49:24.570 --> 00:49:28.210 align:middle line:84% I want to understand what you just were trying to explain. 00:49:28.210 --> 00:49:30.900 align:middle line:84% So the first thing-- you're saying that you didn't pick up 00:49:30.900 --> 00:49:33.010 align:middle line:84% any difference in the expression-- 00:49:33.010 --> 00:49:34.680 align:middle line:84% AUDIENCE: On a gene-by-gene basis. 00:49:34.680 --> 00:49:35.250 align:middle line:90% TODD GOLUB: That's right. 00:49:35.250 --> 00:49:36.510 align:middle line:84% AUDIENCE: But then, when you group 00:49:36.510 --> 00:49:38.130 align:middle line:84% a couple of them together, all of a sudden, 00:49:38.130 --> 00:49:38.970 align:middle line:90% there is a difference. 00:49:38.970 --> 00:49:39.540 align:middle line:90% TODD GOLUB: Right. 00:49:39.540 --> 00:49:40.415 align:middle line:90% So how could that be? 00:49:40.415 --> 00:49:42.960 align:middle line:90% AUDIENCE: How could that be? 00:49:42.960 --> 00:49:45.485 align:middle line:90% [INTERPOSING VOICES] 00:49:45.485 --> 00:49:47.760 align:middle line:84% AUDIENCE: --getting more information out of that. 00:49:47.760 --> 00:49:50.775 align:middle line:84% Maybe a weak signal [INAUDIBLE] sample [INAUDIBLE] 00:49:50.775 --> 00:49:53.102 align:middle line:90% the coherence. 00:49:53.102 --> 00:49:55.560 align:middle line:84% AUDIENCE: [INAUDIBLE] a couple things things being combined 00:49:55.560 --> 00:49:56.400 align:middle line:90% [INAUDIBLE]. 00:49:56.400 --> 00:49:58.020 align:middle line:90% Make it one up and one down. 00:49:58.020 --> 00:50:01.170 align:middle line:90% [INAUDIBLE] 00:50:01.170 --> 00:50:05.310 align:middle line:84% TODD GOLUB: So the microarrays themselves-- 00:50:05.310 --> 00:50:08.867 align:middle line:84% the precision of these arrays is not so fantastic. 00:50:08.867 --> 00:50:10.950 align:middle line:84% And so you can imagine if there's a subtle signal. 00:50:10.950 --> 00:50:16.410 align:middle line:84% On a gene by gene basis, it's difficult to detect it. 00:50:16.410 --> 00:50:19.950 align:middle line:84% But if you consider the coordinate regulation 00:50:19.950 --> 00:50:24.022 align:middle line:84% of a group of genes all in the same direction, as a group, 00:50:24.022 --> 00:50:25.230 align:middle line:90% this might be quite striking. 00:50:25.230 --> 00:50:28.770 align:middle line:84% And this is shown right here, which is really quite amazing 00:50:28.770 --> 00:50:30.070 align:middle line:90% when you think about it. 00:50:30.070 --> 00:50:32.400 align:middle line:84% So here, look at the mean expression 00:50:32.400 --> 00:50:34.980 align:middle line:84% level of all the diabetic patients 00:50:34.980 --> 00:50:36.990 align:middle line:90% versus all the normal patients. 00:50:36.990 --> 00:50:39.210 align:middle line:84% All the genes on the array are shown in gray. 00:50:39.210 --> 00:50:42.930 align:middle line:84% And so you would expect there are no outliers. 00:50:42.930 --> 00:50:45.660 align:middle line:84% There's nothing really way off the diagonal. 00:50:45.660 --> 00:50:47.910 align:middle line:84% If they were, those would show up 00:50:47.910 --> 00:50:52.560 align:middle line:84% as single genes that were differentially expressed. 00:50:52.560 --> 00:50:54.570 align:middle line:84% Of course, you could have one massive outlier 00:50:54.570 --> 00:50:56.653 align:middle line:84% that could screw you up with looking at the means. 00:50:56.653 --> 00:50:58.260 align:middle line:90% But still, you get my point. 00:50:58.260 --> 00:51:01.590 align:middle line:84% And here are these oxidative phosphorylation-- 00:51:01.590 --> 00:51:05.520 align:middle line:84% the gene set that was defined by those genes that are involved 00:51:05.520 --> 00:51:07.110 align:middle line:90% in oxidative phosphorylation. 00:51:07.110 --> 00:51:09.900 align:middle line:84% And you can see that, with only a few exceptions, 00:51:09.900 --> 00:51:13.140 align:middle line:84% they're all lined up just below the diagonal. 00:51:13.140 --> 00:51:19.710 align:middle line:84% Their change in gene expression is only about 20% 00:51:19.710 --> 00:51:23.320 align:middle line:84% compared to normal, but it's all in the same direction. 00:51:23.320 --> 00:51:32.040 align:middle line:84% So 20% change in this number of genes is quite significant. 00:51:32.040 --> 00:51:34.240 align:middle line:90% AUDIENCE: Are you getting that? 00:51:34.240 --> 00:51:35.670 align:middle line:90% Let me try to-- 00:51:35.670 --> 00:51:38.490 align:middle line:84% because it's an incredibly important point. 00:51:38.490 --> 00:51:42.570 align:middle line:84% The chance-- if you look at any given dot of thing, 00:51:42.570 --> 00:51:44.400 align:middle line:90% what does it means to be 1-- 00:51:44.400 --> 00:51:46.970 align:middle line:84% near the diagonal, one side or the other. 00:51:46.970 --> 00:51:48.750 align:middle line:84% I'm not going to make [INAUDIBLE] story. 00:51:48.750 --> 00:51:49.773 align:middle line:90% [INAUDIBLE] about. 00:51:49.773 --> 00:51:51.510 align:middle line:90% It's going to be a quick one-- 00:51:51.510 --> 00:51:53.920 align:middle line:84% the fact that it's on one side or the other. 00:51:53.920 --> 00:51:56.070 align:middle line:84% And then, just by dumb luck, they're 00:51:56.070 --> 00:51:59.640 align:middle line:84% all on one side of the diagonal if they're put diagonal. 00:51:59.640 --> 00:52:03.090 align:middle line:84% That's going to be incredibly unlikely. 00:52:03.090 --> 00:52:08.820 align:middle line:84% And so each individual gene is one side of the diagonal. 00:52:08.820 --> 00:52:13.140 align:middle line:84% The fact that all genes that we have pre-assigned beforehand 00:52:13.140 --> 00:52:14.593 align:middle line:84% of the other type-- in this case, 00:52:14.593 --> 00:52:17.010 align:middle line:84% [INAUDIBLE] phosphorylation-- that will end up on one side 00:52:17.010 --> 00:52:18.425 align:middle line:90% of diagonal-- 00:52:18.425 --> 00:52:20.550 align:middle line:84% that's hugely unlikely, the fact that you can just, 00:52:20.550 --> 00:52:23.160 align:middle line:84% by some luck, put them all on one side of the diagonal. 00:52:23.160 --> 00:52:26.460 align:middle line:84% AUDIENCE: So it's less tied to the enrichment [INAUDIBLE] 00:52:26.460 --> 00:52:32.130 align:middle line:84% the probability that you define [INAUDIBLE] that, given 00:52:32.130 --> 00:52:34.350 align:middle line:84% that you've said these are genes that should 00:52:34.350 --> 00:52:36.800 align:middle line:90% be related to some [INAUDIBLE]. 00:52:36.800 --> 00:52:37.827 align:middle line:90% AUDIENCE: Yes. 00:52:37.827 --> 00:52:39.110 align:middle line:90% AUDIENCE: That makes sense. 00:52:39.110 --> 00:52:39.770 align:middle line:90% TODD GOLUB: That's right. 00:52:39.770 --> 00:52:41.480 align:middle line:84% AUDIENCE: [INAUDIBLE] together, the classes can be formed. 00:52:41.480 --> 00:52:42.522 align:middle line:90% TODD GOLUB: That's right. 00:52:42.522 --> 00:52:47.600 align:middle line:84% Because if you look-- if you take this point in isolation, 00:52:47.600 --> 00:52:50.930 align:middle line:84% there's no way that that's going to be significant 00:52:50.930 --> 00:52:54.560 align:middle line:84% because it's right in the middle-- 00:52:54.560 --> 00:52:55.910 align:middle line:90% in this thing. 00:52:55.910 --> 00:52:59.600 align:middle line:84% So this is an eye-opening experiment, 00:52:59.600 --> 00:53:02.960 align:middle line:84% and it's causing us to go back and reanalyze, 00:53:02.960 --> 00:53:06.200 align:middle line:84% using this methodology called gene set enrichment analysis, 00:53:06.200 --> 00:53:07.070 align:middle line:90% some old data sets. 00:53:07.070 --> 00:53:09.170 align:middle line:84% Let me give you a couple of other examples 00:53:09.170 --> 00:53:12.740 align:middle line:84% of unpublished examples that in a slightly different-- 00:53:12.740 --> 00:53:14.430 align:middle line:84% use it in a slightly different way. 00:53:14.430 --> 00:53:17.700 align:middle line:84% So I told you about our medulloblastoma outcome 00:53:17.700 --> 00:53:18.950 align:middle line:90% prediction experiments before. 00:53:18.950 --> 00:53:24.620 align:middle line:84% And around the same time, there was a paper published 00:53:24.620 --> 00:53:28.520 align:middle line:84% that looked at the same question, essentially. 00:53:28.520 --> 00:53:32.660 align:middle line:84% Non-metastatic versus metastatic medulloblastoma-- different 00:53:32.660 --> 00:53:35.180 align:middle line:84% patients, different arrays, different groups, whatever. 00:53:35.180 --> 00:53:39.800 align:middle line:84% They made a classifier that was centered around the PDGF 00:53:39.800 --> 00:53:42.230 align:middle line:90% receptor alpha gene-- 00:53:42.230 --> 00:53:44.930 align:middle line:84% was a predictor and also a number 00:53:44.930 --> 00:53:48.710 align:middle line:84% of the downstream players of PDGF receptor alpha. 00:53:48.710 --> 00:53:53.150 align:middle line:90% And so we asked, are any o-- 00:53:53.150 --> 00:53:55.400 align:middle line:84% when we look at our classifier of outcome, which 00:53:55.400 --> 00:54:00.650 align:middle line:84% I showed you is pretty decent, where's the PDGF receptor 00:54:00.650 --> 00:54:03.320 align:middle line:90% alpha pathway on there? 00:54:03.320 --> 00:54:08.300 align:middle line:84% And neither PDGF receptor alpha or the genes 00:54:08.300 --> 00:54:11.090 align:middle line:84% in that pathway were among the top predictors-- top 00:54:11.090 --> 00:54:15.290 align:middle line:84% 50 genes in our data set, one which 00:54:15.290 --> 00:54:18.890 align:middle line:84% I think would lead 1 to believe that one or both of those data 00:54:18.890 --> 00:54:24.140 align:middle line:84% sets is wrong or the models derived from them are wrong. 00:54:24.140 --> 00:54:32.270 align:middle line:84% But if you take this PDGF receptor alpha-related genes 00:54:32.270 --> 00:54:35.150 align:middle line:84% as a gene set and ask, is it enriched 00:54:35.150 --> 00:54:38.810 align:middle line:84% in our data set using this methodology shown schematically 00:54:38.810 --> 00:54:42.890 align:middle line:90% here, it's enriched. 00:54:42.890 --> 00:54:47.220 align:middle line:84% This list is 12,000 genes long or so. 00:54:47.220 --> 00:54:50.240 align:middle line:84% So you can see they're not all stacked up, like 1 through 50, 00:54:50.240 --> 00:54:52.590 align:middle line:84% but they're non-randomly distributed, 00:54:52.590 --> 00:54:56.150 align:middle line:84% which we take to mean that, actually, the two data 00:54:56.150 --> 00:54:57.720 align:middle line:90% sets are consistent. 00:54:57.720 --> 00:55:00.860 align:middle line:84% If you had data sets of infinite size, 00:55:00.860 --> 00:55:02.720 align:middle line:84% then you'd start to see convergence 00:55:02.720 --> 00:55:07.550 align:middle line:84% of the markers being overlapping at the very top of the list. 00:55:07.550 --> 00:55:12.597 align:middle line:84% But with these smaller data sets and the clinical variability-- 00:55:12.597 --> 00:55:14.690 align:middle line:84% AUDIENCE: This puts some formalism 00:55:14.690 --> 00:55:17.540 align:middle line:84% around what [INAUDIBLE] was waving his hands 00:55:17.540 --> 00:55:19.040 align:middle line:90% and yelling about. 00:55:19.040 --> 00:55:21.030 align:middle line:84% For five years, people said, well, 00:55:21.030 --> 00:55:22.260 align:middle line:90% this microarray [INAUDIBLE]. 00:55:22.260 --> 00:55:23.180 align:middle line:84% You can't tell the difference between them. 00:55:23.180 --> 00:55:23.888 align:middle line:90% TODD GOLUB: Yeah. 00:55:23.888 --> 00:55:25.560 align:middle line:90% AUDIENCE: This is the pattern. 00:55:25.560 --> 00:55:26.850 align:middle line:90% It's the overall pattern. 00:55:26.850 --> 00:55:29.090 align:middle line:90% This is a much more formal way-- 00:55:29.090 --> 00:55:31.243 align:middle line:84% a pattern that's released to pick [INAUDIBLE].. 00:55:31.243 --> 00:55:33.410 align:middle line:84% TODD GOLUB: So here's another example of that thing. 00:55:33.410 --> 00:55:34.850 align:middle line:90% It's also unpublished. 00:55:34.850 --> 00:55:40.370 align:middle line:90% So we looked at lung cancer-- 00:55:40.370 --> 00:55:44.930 align:middle line:84% human adenocarcinoma in the lung and identify some-- 00:55:44.930 --> 00:55:48.500 align:middle line:84% we just drew the line at 50 because it's a nice number-- 00:55:48.500 --> 00:55:53.750 align:middle line:84% predictors of outcome in the Boston lung cancer patients. 00:55:53.750 --> 00:55:55.760 align:middle line:84% University of Michigan did the same experiment, 00:55:55.760 --> 00:55:57.680 align:middle line:90% published around the same time. 00:55:57.680 --> 00:56:00.710 align:middle line:90% Overlap in gene of these two-- 00:56:00.710 --> 00:56:04.160 align:middle line:90% listed 50 genes-- zero-- 00:56:04.160 --> 00:56:05.870 align:middle line:90% concerning. 00:56:05.870 --> 00:56:10.850 align:middle line:84% But if you look in the space of gene set space and you ask, 00:56:10.850 --> 00:56:13.400 align:middle line:84% what gene sets are enriched in one data set? 00:56:13.400 --> 00:56:16.490 align:middle line:84% What gene sets-- which you can think of loosely as pathways. 00:56:16.490 --> 00:56:18.770 align:middle line:84% They're not really pathways, but it's 00:56:18.770 --> 00:56:21.650 align:middle line:84% reasonable think of them for this purpose. 00:56:21.650 --> 00:56:26.280 align:middle line:84% There's really quite significant overlap in gene set space. 00:56:26.280 --> 00:56:30.590 align:middle line:84% So I think what this is saying is that Botstein is right, 00:56:30.590 --> 00:56:34.467 align:middle line:84% that there is more biologic coherence in these data sets. 00:56:34.467 --> 00:56:36.050 align:middle line:84% It's just we haven't been smart enough 00:56:36.050 --> 00:56:38.390 align:middle line:90% to really know how to see it. 00:56:38.390 --> 00:56:40.360 align:middle line:84% AUDIENCE: How did you choose your gene set? 00:56:40.360 --> 00:56:41.870 align:middle line:90% [INAUDIBLE] What do you use-- 00:56:41.870 --> 00:56:43.816 align:middle line:84% [INAUDIBLE] function or is it a pathway? 00:56:43.816 --> 00:56:47.940 align:middle line:84% Or how do you actually [INAUDIBLE]?? 00:56:47.940 --> 00:56:52.170 align:middle line:84% TODD GOLUB: We now have about 450 or so such gene sets-- 00:56:52.170 --> 00:56:54.150 align:middle line:84% some of which are good, some of which 00:56:54.150 --> 00:56:55.920 align:middle line:90% aren't particularly useful. 00:56:55.920 --> 00:56:58.770 align:middle line:90% They include some go annotation. 00:56:58.770 --> 00:57:00.270 align:middle line:84% I don't think those are particularly 00:57:00.270 --> 00:57:03.990 align:middle line:84% useful because the granularity isn't fine enough. 00:57:03.990 --> 00:57:07.290 align:middle line:84% I think, in the end, the most useful types of gene sets 00:57:07.290 --> 00:57:11.250 align:middle line:84% are going to be those that are experimentally derived. 00:57:11.250 --> 00:57:14.160 align:middle line:84% But this is a mixture of those, and we're not yet at the point 00:57:14.160 --> 00:57:18.630 align:middle line:84% where we even started try to understand-- of these 35 00:57:18.630 --> 00:57:20.850 align:middle line:84% enriched sets of genes, what are they 00:57:20.850 --> 00:57:22.860 align:middle line:90% and what's the biological story? 00:57:22.860 --> 00:57:26.550 align:middle line:84% AUDIENCE: You threw at it on the order of 60 gene sets? 00:57:26.550 --> 00:57:27.628 align:middle line:90% TODD GOLUB: No. 00:57:27.628 --> 00:57:28.170 align:middle line:90% AUDIENCE: No? 00:57:28.170 --> 00:57:30.180 align:middle line:90% TODD GOLUB: No. 00:57:30.180 --> 00:57:33.870 align:middle line:84% We threw at it 400-and-something gene sets and asked, 00:57:33.870 --> 00:57:36.690 align:middle line:84% how many of those are enriched in the Boston data set? 00:57:36.690 --> 00:57:39.420 align:middle line:90% And the answer is 35 plus 18. 00:57:39.420 --> 00:57:43.870 align:middle line:84% And 35 plus 12 were enriched here. 00:57:43.870 --> 00:57:47.340 align:middle line:84% And so the majority of the sets enriched in one 00:57:47.340 --> 00:57:49.020 align:middle line:90% were also enriched in the other. 00:57:49.020 --> 00:57:49.680 align:middle line:90% AUDIENCE: That's helpful. 00:57:49.680 --> 00:57:51.472 align:middle line:84% AUDIENCE: Are those cancer-specific or just 00:57:51.472 --> 00:57:52.200 align:middle line:90% biological? 00:57:52.200 --> 00:57:54.510 align:middle line:84% TODD GOLUB: No, they're not cancer specific. 00:57:54.510 --> 00:57:58.380 align:middle line:84% AUDIENCE: And those gene sets are manually 00:57:58.380 --> 00:58:03.860 align:middle line:84% annotated by their group, or is the [INAUDIBLE]?? 00:58:03.860 --> 00:58:05.840 align:middle line:84% TODD GOLUB: There a combination, as I said. 00:58:05.840 --> 00:58:08.240 align:middle line:84% Some are these [INAUDIBLE] pathways 00:58:08.240 --> 00:58:10.700 align:middle line:90% that are so-so annotation. 00:58:10.700 --> 00:58:15.290 align:middle line:84% Some are entirely computationally derived. 00:58:15.290 --> 00:58:19.400 align:middle line:84% That is, they're the nearest neighbor genes of a given index 00:58:19.400 --> 00:58:22.070 align:middle line:90% gene in a data set. 00:58:22.070 --> 00:58:26.840 align:middle line:90% They are various things. 00:58:26.840 --> 00:58:29.870 align:middle line:84% And what the definitive collection of gene sets 00:58:29.870 --> 00:58:35.403 align:middle line:84% would actually look like isn't obvious to me. 00:58:35.403 --> 00:58:37.820 align:middle line:84% AUDIENCE: [INAUDIBLE] learn more and more about mechanism. 00:58:37.820 --> 00:58:38.862 align:middle line:90% TODD GOLUB: That's right. 00:58:38.862 --> 00:58:42.140 align:middle line:90% 00:58:42.140 --> 00:58:44.870 align:middle line:84% On the one hand, I think it will be useful to just 00:58:44.870 --> 00:58:50.660 align:middle line:84% not fret about it too much and worry about exactly 00:58:50.660 --> 00:58:52.310 align:middle line:90% how to define these things. 00:58:52.310 --> 00:58:54.410 align:middle line:90% Just get them in there. 00:58:54.410 --> 00:58:56.210 align:middle line:84% The nice thing about this GSCA methodology 00:58:56.210 --> 00:58:57.918 align:middle line:84% that I didn't really go through in detail 00:58:57.918 --> 00:59:00.830 align:middle line:90% is that it's forgiving-- 00:59:00.830 --> 00:59:03.980 align:middle line:84% how you calculate these enrichment scores is forgiving 00:59:03.980 --> 00:59:05.870 align:middle line:84% of the definition of the gene sets 00:59:05.870 --> 00:59:08.210 align:middle line:84% because you're looking for enriched-- non-random 00:59:08.210 --> 00:59:09.770 align:middle line:90% enrichment of the gene. 00:59:09.770 --> 00:59:13.700 align:middle line:84% Set so the fact that a third or a 1/2 of the gene set 00:59:13.700 --> 00:59:16.248 align:middle line:84% may actually be inappropriately there 00:59:16.248 --> 00:59:18.290 align:middle line:84% doesn't make any difference because there's still 00:59:18.290 --> 00:59:21.610 align:middle line:84% enough that significantly enriched to detect it. 00:59:21.610 --> 00:59:23.080 align:middle line:90% AUDIENCE: You also [INAUDIBLE]? 00:59:23.080 --> 00:59:25.527 align:middle line:90% 00:59:25.527 --> 00:59:26.360 align:middle line:90% TODD GOLUB: You can. 00:59:26.360 --> 00:59:28.970 align:middle line:84% We didn't happen to do it here, but you can. 00:59:28.970 --> 00:59:31.345 align:middle line:84% Again, like any of these other things, 00:59:31.345 --> 00:59:33.470 align:middle line:84% there are going to be a number of different metrics 00:59:33.470 --> 00:59:37.850 align:middle line:84% that you could apply to measure significant enrichment. 00:59:37.850 --> 00:59:40.040 align:middle line:84% The most important thing is just to make sure 00:59:40.040 --> 00:59:44.300 align:middle line:84% that you correct for the possibility of whatever metric 00:59:44.300 --> 00:59:44.990 align:middle line:90% you use-- 00:59:44.990 --> 00:59:48.230 align:middle line:84% that you're detecting something beyond what you'd 00:59:48.230 --> 00:59:50.580 align:middle line:90% expect by chance. 00:59:50.580 --> 00:59:55.880 align:middle line:84% AUDIENCE: So I have two questions about [INAUDIBLE].. 00:59:55.880 --> 01:00:02.750 align:middle line:84% One would be-- so I'm assuming that you can also detect-- 01:00:02.750 --> 01:00:06.610 align:middle line:84% you have a [INAUDIBLE] of a particular gene, which 01:00:06.610 --> 01:00:08.590 align:middle line:90% you haven't mentioned so far. 01:00:08.590 --> 01:00:10.300 align:middle line:84% So is that [INAUDIBLE] an example 01:00:10.300 --> 01:00:14.650 align:middle line:90% of an actual [INAUDIBLE]. 01:00:14.650 --> 01:00:16.750 align:middle line:84% Is that something you would find by going back 01:00:16.750 --> 01:00:20.380 align:middle line:84% to your healthy samples and comparing-- 01:00:20.380 --> 01:00:24.770 align:middle line:84% looking for enrichment relative to your disease? 01:00:24.770 --> 01:00:26.353 align:middle line:90% AUDIENCE: A positive score? 01:00:26.353 --> 01:00:28.270 align:middle line:84% TODD GOLUB: So actually, the way you calculate 01:00:28.270 --> 01:00:32.268 align:middle line:84% this is the metric doesn't specifically look 01:00:32.268 --> 01:00:33.560 align:middle line:90% for enrichment towards the top. 01:00:33.560 --> 01:00:35.210 align:middle line:84% It looks for a non-random distribution. 01:00:35.210 --> 01:00:38.590 align:middle line:90% 01:00:38.590 --> 01:00:41.500 align:middle line:90% You would find depletion. 01:00:41.500 --> 01:00:43.180 align:middle line:84% What you could also find-- which I think 01:00:43.180 --> 01:00:45.580 align:middle line:84% is-- this score will capture and is not 01:00:45.580 --> 01:00:48.610 align:middle line:84% desirable-- would be something that's 01:00:48.610 --> 01:00:51.910 align:middle line:84% concentrated in the middle, which is very uninteresting. 01:00:51.910 --> 01:00:54.310 align:middle line:84% So there are some false positives in there. 01:00:54.310 --> 01:00:58.120 align:middle line:84% AUDIENCE: That may not actually be so interesting-- 01:00:58.120 --> 01:01:01.300 align:middle line:84% for example, in T-cell activation, 01:01:01.300 --> 01:01:04.570 align:middle line:90% you get normal [INAUDIBLE]. 01:01:04.570 --> 01:01:06.910 align:middle line:84% You get upregulation of certain proteins 01:01:06.910 --> 01:01:09.080 align:middle line:84% and you get downregulation of others. 01:01:09.080 --> 01:01:13.450 align:middle line:84% And so what I haven't heard yet is how you account 01:01:13.450 --> 01:01:18.310 align:middle line:84% for, perhaps, enrichment of part of your genes [INAUDIBLE] 01:01:18.310 --> 01:01:19.670 align:middle line:90% increasing another with that-- 01:01:19.670 --> 01:01:20.740 align:middle line:84% TODD GOLUB: So there is another version 01:01:20.740 --> 01:01:22.630 align:middle line:84% of this that tries to dissect the gene 01:01:22.630 --> 01:01:25.843 align:middle line:84% sets into those components that move coherently 01:01:25.843 --> 01:01:27.260 align:middle line:84% in one direction versus the other, 01:01:27.260 --> 01:01:28.593 align:middle line:90% because you're absolutely right. 01:01:28.593 --> 01:01:29.842 align:middle line:90% AUDIENCE: [INAUDIBLE] 01:01:29.842 --> 01:01:30.550 align:middle line:90% TODD GOLUB: Yeah. 01:01:30.550 --> 01:01:33.820 align:middle line:90% If, for example, you take a-- 01:01:33.820 --> 01:01:37.600 align:middle line:84% use GO annotation or something like that, or some pathway, 01:01:37.600 --> 01:01:40.030 align:middle line:84% if half the genes in the pathway go up 01:01:40.030 --> 01:01:45.040 align:middle line:84% and half the genes go down, that could look like no enrichment 01:01:45.040 --> 01:01:48.910 align:middle line:84% at all, whereas, if you separate those somehow, 01:01:48.910 --> 01:01:49.750 align:middle line:90% you could see it. 01:01:49.750 --> 01:01:53.377 align:middle line:90% 01:01:53.377 --> 01:01:54.460 align:middle line:90% How are we doing for time? 01:01:54.460 --> 01:01:55.490 align:middle line:84% AUDIENCE: Well, you've got 20 minutes-- 01:01:55.490 --> 01:01:56.145 align:middle line:90% 18 minutes. 01:01:56.145 --> 01:01:56.770 align:middle line:90% TODD GOLUB: OK. 01:01:56.770 --> 01:02:00.010 align:middle line:90% 01:02:00.010 --> 01:02:02.530 align:middle line:90% So let me push this to-- 01:02:02.530 --> 01:02:06.155 align:middle line:84% not classification but some newer directions 01:02:06.155 --> 01:02:07.780 align:middle line:84% that we're thinking about-- how can you 01:02:07.780 --> 01:02:12.460 align:middle line:84% use these signatures for useful things, 01:02:12.460 --> 01:02:15.220 align:middle line:84% particularly to think about something that's 01:02:15.220 --> 01:02:17.090 align:middle line:90% closer to drug discovery. 01:02:17.090 --> 01:02:21.920 align:middle line:84% So this is the way the usual discovery 01:02:21.920 --> 01:02:23.760 align:middle line:90% pipeline would look like. 01:02:23.760 --> 01:02:26.240 align:middle line:84% You have some disease process or biological process 01:02:26.240 --> 01:02:26.870 align:middle line:90% you care about. 01:02:26.870 --> 01:02:29.240 align:middle line:84% You do some microarray experiments. 01:02:29.240 --> 01:02:33.140 align:middle line:84% And then a miracle's is supposed to occur whereby you develop 01:02:33.140 --> 01:02:36.140 align:middle line:84% sufficient molecular understanding of what 01:02:36.140 --> 01:02:41.930 align:middle line:84% the data are telling you, that you can identify the smoking 01:02:41.930 --> 01:02:43.010 align:middle line:90% gun target. 01:02:43.010 --> 01:02:46.550 align:middle line:84% And then you partner with a drug company and say, 01:02:46.550 --> 01:02:48.680 align:middle line:84% screen for a small molecule that inhibits 01:02:48.680 --> 01:02:50.690 align:middle line:84% this critical therapeutic target. 01:02:50.690 --> 01:02:54.210 align:middle line:84% The problem is that this part is really tough. 01:02:54.210 --> 01:02:56.270 align:middle line:84% And so what we've been thinking about 01:02:56.270 --> 01:02:59.690 align:middle line:84% is, well, could you bypass the understanding part, 01:02:59.690 --> 01:03:03.260 align:middle line:84% at least initially, whereby you screen 01:03:03.260 --> 01:03:05.930 align:middle line:84% for small molecules based on their ability 01:03:05.930 --> 01:03:09.440 align:middle line:84% simply to perturb a signature of interest. 01:03:09.440 --> 01:03:11.180 align:middle line:84% And then, once you have those in hand, 01:03:11.180 --> 01:03:14.870 align:middle line:84% you could use them to further dissect the biology 01:03:14.870 --> 01:03:17.510 align:middle line:84% or, if you're lucky, think about them like drugs. 01:03:17.510 --> 01:03:23.280 align:middle line:84% And so the proof-of-concept experiment is shown here, 01:03:23.280 --> 01:03:23.870 align:middle line:90% where-- 01:03:23.870 --> 01:03:26.720 align:middle line:84% here are two biological states, for example-- 01:03:26.720 --> 01:03:29.930 align:middle line:84% a leukemia cell, which is undifferentiated, 01:03:29.930 --> 01:03:33.950 align:middle line:84% and a normal blood cell, which is fully mature 01:03:33.950 --> 01:03:37.940 align:middle line:84% and is differentiated along the myeloid pathway-- 01:03:37.940 --> 01:03:40.400 align:middle line:90% a peripheral blood neutrophil. 01:03:40.400 --> 01:03:45.450 align:middle line:84% It's not known what the critical targets are of this pathway. 01:03:45.450 --> 01:03:47.120 align:middle line:84% So it's hard to do a small molecule 01:03:47.120 --> 01:03:49.910 align:middle line:84% screen to induce this process, which would be nice, 01:03:49.910 --> 01:03:53.670 align:middle line:84% if you could simply induce your leukemia cells to turn 01:03:53.670 --> 01:03:55.200 align:middle line:90% into normal cells. 01:03:55.200 --> 01:03:56.960 align:middle line:84% So the question is, could we define 01:03:56.960 --> 01:04:00.260 align:middle line:84% a signature of this state, a signature of this state, 01:04:00.260 --> 01:04:04.260 align:middle line:84% and then screen for compounds that trigger the signature? 01:04:04.260 --> 01:04:06.320 align:middle line:84% So the details don't matter here, 01:04:06.320 --> 01:04:09.230 align:middle line:84% but the concept is, define signatures 01:04:09.230 --> 01:04:12.570 align:middle line:90% of the two states of interest. 01:04:12.570 --> 01:04:16.280 align:middle line:84% So we call this thing GEHTS, for gene expression based 01:04:16.280 --> 01:04:18.890 align:middle line:90% high throughput screening. 01:04:18.890 --> 01:04:22.040 align:middle line:84% Define the signatures-- now standards-- 01:04:22.040 --> 01:04:24.800 align:middle line:84% what we've been talking about-- our microarrays, 01:04:24.800 --> 01:04:26.630 align:middle line:90% where the experiment would be-- 01:04:26.630 --> 01:04:31.040 align:middle line:84% treat cells with various different chemical compounds 01:04:31.040 --> 01:04:33.860 align:middle line:84% and ask whether any of those compounds 01:04:33.860 --> 01:04:36.500 align:middle line:84% trigger the signature of interest. 01:04:36.500 --> 01:04:38.720 align:middle line:84% And to make this feasible, we simplify 01:04:38.720 --> 01:04:43.460 align:middle line:84% these complex signatures into a handful of genes 01:04:43.460 --> 01:04:47.255 align:middle line:84% that you can measure by multiplexed PCR. 01:04:47.255 --> 01:04:48.380 align:middle line:90% AUDIENCE: I read the paper. 01:04:48.380 --> 01:04:51.320 align:middle line:84% Was just purely a cost issue as opposed to 01:04:51.320 --> 01:04:52.805 align:middle line:84% going directly to the microarrays? 01:04:52.805 --> 01:04:54.680 align:middle line:84% TODD GOLUB: It's a cost and throughput issue. 01:04:54.680 --> 01:04:56.030 align:middle line:90% Yeah, so if you wanted to-- 01:04:56.030 --> 01:04:56.905 align:middle line:90% AUDIENCE: We're good. 01:04:56.905 --> 01:04:57.610 align:middle line:90% [INAUDIBLE] 01:04:57.610 --> 01:04:59.900 align:middle line:90% 01:04:59.900 --> 01:05:02.390 align:middle line:84% TODD GOLUB: If you wanted to screen tens of thousands 01:05:02.390 --> 01:05:06.110 align:middle line:84% of compounds, not really feasible 01:05:06.110 --> 01:05:11.270 align:middle line:84% if it costs you $500 a pop in the throughput issue. 01:05:11.270 --> 01:05:14.960 align:middle line:84% So yeah, it's a practical matter here. 01:05:14.960 --> 01:05:16.160 align:middle line:90% This part doesn't matter. 01:05:16.160 --> 01:05:19.430 align:middle line:84% Suffice it to say there's a method for how 01:05:19.430 --> 01:05:22.461 align:middle line:84% to measure a simplified signature in high throughput. 01:05:22.461 --> 01:05:26.070 align:middle line:90% 01:05:26.070 --> 01:05:32.160 align:middle line:84% So we screened a couple thousand compounds and asked, 01:05:32.160 --> 01:05:35.060 align:middle line:84% do any of them trigger this little mini gene signature? 01:05:35.060 --> 01:05:37.370 align:middle line:90% And some of them did. 01:05:37.370 --> 01:05:39.930 align:middle line:90% Details don't matter. 01:05:39.930 --> 01:05:42.950 align:middle line:84% But then the question should be, well, maybe these things 01:05:42.950 --> 01:05:46.190 align:middle line:84% just trigger these five genes that are-- 01:05:46.190 --> 01:05:49.010 align:middle line:84% I'm sorry, these compounds trigger the five genes 01:05:49.010 --> 01:05:50.907 align:middle line:90% but actually don't do anything. 01:05:50.907 --> 01:05:53.240 align:middle line:84% So one way that you could sort out whether they actually 01:05:53.240 --> 01:05:56.180 align:middle line:84% do anything biologically is to now step back and look 01:05:56.180 --> 01:05:59.750 align:middle line:84% across the whole genome again, take cells, treat them 01:05:59.750 --> 01:06:01.580 align:middle line:84% with these candidate compounds, and ask, 01:06:01.580 --> 01:06:05.750 align:middle line:84% did you actually recapitulate the overall molecular program, 01:06:05.750 --> 01:06:08.840 align:middle line:84% not just of these five genes but of the whole thing-- 01:06:08.840 --> 01:06:10.260 align:middle line:90% of the whole molecular program? 01:06:10.260 --> 01:06:14.460 align:middle line:84% So if you turn back to genome-wide arrays, 01:06:14.460 --> 01:06:16.700 align:middle line:84% you can see that a number of these compounds 01:06:16.700 --> 01:06:22.370 align:middle line:84% recapitulated the molecular program of differentiation. 01:06:22.370 --> 01:06:23.730 align:middle line:90% Does that make sense? 01:06:23.730 --> 01:06:28.580 align:middle line:84% So you use the simplified high throughput assay just as 01:06:28.580 --> 01:06:32.853 align:middle line:84% a readout of whether you've triggered the signature or not. 01:06:32.853 --> 01:06:34.520 align:middle line:84% And then, with those candidates in hand, 01:06:34.520 --> 01:06:36.620 align:middle line:84% you go back and interrogate them. 01:06:36.620 --> 01:06:39.800 align:middle line:84% AUDIENCE: Does everybody follow that? 01:06:39.800 --> 01:06:43.202 align:middle line:84% TODD GOLUB: So the genes in this little signature itself-- 01:06:43.202 --> 01:06:45.410 align:middle line:84% AUDIENCE: We don't if they're actually doing anything 01:06:45.410 --> 01:06:47.510 align:middle line:90% to the rest of the cell. 01:06:47.510 --> 01:06:49.760 align:middle line:84% TODD GOLUB: I'm pretty confident they don't, actually. 01:06:49.760 --> 01:06:50.610 align:middle line:90% But it's irrelevant. 01:06:50.610 --> 01:06:53.150 align:middle line:84% So you define these signatures not based 01:06:53.150 --> 01:06:55.520 align:middle line:84% on their being important or somebody thinks, oh, 01:06:55.520 --> 01:06:58.338 align:middle line:84% that's a good target, and that's important for leukemia 01:06:58.338 --> 01:06:59.630 align:middle line:90% or differentiation or whatever. 01:06:59.630 --> 01:07:00.393 align:middle line:90% It simply-- 01:07:00.393 --> 01:07:01.685 align:middle line:90% AUDIENCE: Represents the class. 01:07:01.685 --> 01:07:02.960 align:middle line:84% TODD GOLUB: It represents the class. 01:07:02.960 --> 01:07:03.930 align:middle line:90% All you care about. 01:07:03.930 --> 01:07:05.370 align:middle line:84% And that you can measure it well. 01:07:05.370 --> 01:07:07.880 align:middle line:84% So sometimes you find a good candidate marker 01:07:07.880 --> 01:07:12.080 align:middle line:84% that, for some reason, it doesn't behave nicely 01:07:12.080 --> 01:07:13.682 align:middle line:84% in this assay so you chuck it out 01:07:13.682 --> 01:07:15.140 align:middle line:84% and replace it with something else. 01:07:15.140 --> 01:07:19.800 align:middle line:90% 01:07:19.800 --> 01:07:25.160 align:middle line:84% So here's just an example to say that then, as you might expect, 01:07:25.160 --> 01:07:28.910 align:middle line:84% when you treat leukemia cells with these candidate compounds, 01:07:28.910 --> 01:07:32.030 align:middle line:84% discovered solely based on their gene expression changes, 01:07:32.030 --> 01:07:34.940 align:middle line:84% they do the thing-- the cells do the things that maturing 01:07:34.940 --> 01:07:38.090 align:middle line:84% leukemia cells should do, like-- they become phagocytic. 01:07:38.090 --> 01:07:39.065 align:middle line:90% They start engulfing-- 01:07:39.065 --> 01:07:40.850 align:middle line:84% AUDIENCE: [INAUDIBLE] much more differentiated 01:07:40.850 --> 01:07:41.808 align:middle line:90% like in their behavior. 01:07:41.808 --> 01:07:46.580 align:middle line:84% TODD GOLUB: Yeah, so it's a promising idea. 01:07:46.580 --> 01:07:49.640 align:middle line:84% This is one unpublished example that says, well, 01:07:49.640 --> 01:07:58.670 align:middle line:84% it is known that blood cell differentiation is largely 01:07:58.670 --> 01:08:00.583 align:middle line:84% governed at the transcriptional level. 01:08:00.583 --> 01:08:02.000 align:middle line:84% So maybe that's why you can define 01:08:02.000 --> 01:08:06.440 align:middle line:84% these transcriptional signatures of the differentiation process 01:08:06.440 --> 01:08:07.550 align:middle line:90% and screen for things. 01:08:07.550 --> 01:08:11.630 align:middle line:84% But here's an example of defining a gene expression 01:08:11.630 --> 01:08:13.040 align:middle line:90% signature. 01:08:13.040 --> 01:08:15.170 align:middle line:84% Again, the signature itself is devoid 01:08:15.170 --> 01:08:20.120 align:middle line:84% of any real biological meaning, other than it reads out, 01:08:20.120 --> 01:08:23.130 align:middle line:84% in this case, activation of a signal transduction pathway. 01:08:23.130 --> 01:08:26.180 align:middle line:84% So the usual way of thinking about this is, well, 01:08:26.180 --> 01:08:28.490 align:middle line:84% if you want to look at signal transduction-- proteins 01:08:28.490 --> 01:08:31.399 align:middle line:84% talk to each other, RNA has no place in that. 01:08:31.399 --> 01:08:33.082 align:middle line:84% RNA profiling has no place in that. 01:08:33.082 --> 01:08:35.540 align:middle line:84% But here the idea is, if we stimulate a signaling pathway-- 01:08:35.540 --> 01:08:37.370 align:middle line:84% in this case, by stimulating cells 01:08:37.370 --> 01:08:39.590 align:middle line:84% with platelet-derived growth factor 01:08:39.590 --> 01:08:43.310 align:middle line:84% and then capture the transcriptional response 01:08:43.310 --> 01:08:46.399 align:middle line:84% at the RNA level, could we use an RNA signature 01:08:46.399 --> 01:08:51.170 align:middle line:84% as a readout for PDGF receptor activation and a screen 01:08:51.170 --> 01:08:54.260 align:middle line:84% for inhibitors of the signal transduction 01:08:54.260 --> 01:08:56.990 align:middle line:90% pathway using RNA as a readout? 01:08:56.990 --> 01:09:01.370 align:middle line:84% And here, this is work of a MIT graduate student-- 01:09:01.370 --> 01:09:04.880 align:middle line:84% chemistry graduate student, who pulled out a compound 01:09:04.880 --> 01:09:09.500 align:middle line:84% called aurintricarboxylic acid as an example, which turns out 01:09:09.500 --> 01:09:15.770 align:middle line:84% to be a previously unknown inhibitor of the PDGF receptor 01:09:15.770 --> 01:09:16.910 align:middle line:90% itself. 01:09:16.910 --> 01:09:20.960 align:middle line:84% But it was discovered by looking down here at a signature. 01:09:20.960 --> 01:09:25.700 align:middle line:84% So I think this is going to be useful for various-- 01:09:25.700 --> 01:09:28.760 align:middle line:84% being able to screen for things that you can't otherwise 01:09:28.760 --> 01:09:29.359 align:middle line:90% screen for. 01:09:29.359 --> 01:09:34.370 align:middle line:84% So this is a chemical structure that is not currently 01:09:34.370 --> 01:09:38.120 align:middle line:84% being explored in people who know much more than I 01:09:38.120 --> 01:09:42.439 align:middle line:84% do about kinase inhibitors because no one thought 01:09:42.439 --> 01:09:43.250 align:middle line:90% to look at it. 01:09:43.250 --> 01:09:45.859 align:middle line:84% And it was discovered simply by using as a signature, 01:09:45.859 --> 01:09:49.100 align:middle line:90% as a readout. 01:09:49.100 --> 01:09:54.680 align:middle line:84% So the last five minutes, let me push the signature idea perhaps 01:09:54.680 --> 01:09:56.233 align:middle line:90% further than I should. 01:09:56.233 --> 01:09:57.900 align:middle line:84% AUDIENCE: But it's the end of the class. 01:09:57.900 --> 01:09:58.830 align:middle line:84% TODD GOLUB: But it's the end of the class, 01:09:58.830 --> 01:10:00.350 align:middle line:90% and I'm excited about the idea. 01:10:00.350 --> 01:10:02.810 align:middle line:84% But it has less data surrounding it. 01:10:02.810 --> 01:10:08.960 align:middle line:84% And that's the idea of using these signatures-- 01:10:08.960 --> 01:10:10.715 align:middle line:90% so you can see, at least in-- 01:10:10.715 --> 01:10:14.420 align:middle line:84% my thinking has shifted over the past year or so, 01:10:14.420 --> 01:10:18.388 align:middle line:84% much away from finding, oh, what's the needle-- 01:10:18.388 --> 01:10:20.180 align:middle line:84% using these microarray types of experiments 01:10:20.180 --> 01:10:21.980 align:middle line:84% to find a needle in a haystack-- what's 01:10:21.980 --> 01:10:24.860 align:middle line:84% the gene that's responsible for something I care about? 01:10:24.860 --> 01:10:27.050 align:middle line:84% To thinking about the power of these signatures 01:10:27.050 --> 01:10:30.600 align:middle line:90% as readouts for various things. 01:10:30.600 --> 01:10:34.520 align:middle line:84% So the idea here is to use the signatures-- 01:10:34.520 --> 01:10:41.000 align:middle line:84% RNA signatures as a vehicle for establishing connectivity 01:10:41.000 --> 01:10:47.900 align:middle line:84% between components of the genome and each other, 01:10:47.900 --> 01:10:50.668 align:middle line:84% that you manipulate through perturbation. 01:10:50.668 --> 01:10:53.210 align:middle line:84% It's a little bit different than relevance networks, I think, 01:10:53.210 --> 01:10:55.070 align:middle line:90% but it's conceptually similar. 01:10:55.070 --> 01:10:57.440 align:middle line:90% AUDIENCE: [INAUDIBLE] 01:10:57.440 --> 01:10:59.150 align:middle line:84% TODD GOLUB: Establishing connections 01:10:59.150 --> 01:11:01.790 align:middle line:84% between drugs and drugs and drugs and genes. 01:11:01.790 --> 01:11:04.280 align:middle line:84% So the idea would be that, if you can define a gene 01:11:04.280 --> 01:11:08.960 align:middle line:84% expression signature of, let's say-- 01:11:08.960 --> 01:11:13.520 align:middle line:90% comprehensively of all drugs-- 01:11:13.520 --> 01:11:16.820 align:middle line:84% there are only about 2,000 FDA-approved drugs. 01:11:16.820 --> 01:11:18.770 align:middle line:84% It's actually amazing that that's just 01:11:18.770 --> 01:11:21.200 align:middle line:84% not publicly available information, what 01:11:21.200 --> 01:11:22.700 align:middle line:84% happens to cells when you treat them 01:11:22.700 --> 01:11:25.880 align:middle line:90% with drugs we give to patients. 01:11:25.880 --> 01:11:27.780 align:middle line:84% If you had that and, for example, 01:11:27.780 --> 01:11:33.140 align:middle line:84% you had a signature that was a result of ablating each gene 01:11:33.140 --> 01:11:36.410 align:middle line:84% in the genome sequentially-- only 26,000 genes 01:11:36.410 --> 01:11:39.800 align:middle line:84% in the genome, and there are now reagents coming online with RNA 01:11:39.800 --> 01:11:42.260 align:middle line:84% interference where you can actually do that experiment-- 01:11:42.260 --> 01:11:48.110 align:middle line:84% then you'd have a systematic matrix of perturbations 01:11:48.110 --> 01:11:53.450 align:middle line:84% whereby you use these gene expression signatures 01:11:53.450 --> 01:12:01.040 align:middle line:84% as the universal bio assay to connect genes with genes, 01:12:01.040 --> 01:12:04.560 align:middle line:84% genes with drugs, and drugs with drugs. 01:12:04.560 --> 01:12:05.732 align:middle line:90% Does that make sense? 01:12:05.732 --> 01:12:06.690 align:middle line:90% Do you have a question? 01:12:06.690 --> 01:12:09.390 align:middle line:90% AUDIENCE: [INAUDIBLE] 01:12:09.390 --> 01:12:29.310 align:middle line:90% 01:12:29.310 --> 01:12:33.370 align:middle line:84% How do we prevent from [INAUDIBLE] to the [INAUDIBLE] 01:12:33.370 --> 01:12:38.100 align:middle line:84% cancer cells [INAUDIBLE] cancer cells. 01:12:38.100 --> 01:12:41.850 align:middle line:90% But because of [INAUDIBLE]-- 01:12:41.850 --> 01:12:42.990 align:middle line:90% because of [INAUDIBLE]. 01:12:42.990 --> 01:12:48.460 align:middle line:90% 01:12:48.460 --> 01:12:52.060 align:middle line:84% TODD GOLUB: Well, we're arguing that-- 01:12:52.060 --> 01:12:53.950 align:middle line:84% don't bother doing this in lower organisms 01:12:53.950 --> 01:12:56.180 align:middle line:84% because you can do the experiment in human cells. 01:12:56.180 --> 01:12:58.130 align:middle line:84% So we're going to do it in human cells. 01:12:58.130 --> 01:13:00.130 align:middle line:84% And I would say-- the idea, again, 01:13:00.130 --> 01:13:03.400 align:middle line:84% for establishing these connections is not to be able 01:13:03.400 --> 01:13:03.995 align:middle line:90% to-- 01:13:03.995 --> 01:13:04.870 align:middle line:90% AUDIENCE: Be perfect. 01:13:04.870 --> 01:13:06.370 align:middle line:84% TODD GOLUB: I think people who think 01:13:06.370 --> 01:13:10.160 align:middle line:84% that you can create a wiring diagram 01:13:10.160 --> 01:13:12.700 align:middle line:84% and reverse engineer a cell are out of their minds 01:13:12.700 --> 01:13:14.470 align:middle line:90% based on these data. 01:13:14.470 --> 01:13:16.670 align:middle line:90% It's just not feasible. 01:13:16.670 --> 01:13:19.090 align:middle line:84% So the question is, can you find-- 01:13:19.090 --> 01:13:23.260 align:middle line:84% can you find, for example, enrichment in the same GSCA 01:13:23.260 --> 01:13:25.120 align:middle line:90% kind of thinking? 01:13:25.120 --> 01:13:28.300 align:middle line:84% Can you find enrichment of one signature in another, 01:13:28.300 --> 01:13:32.050 align:middle line:84% thereby establishing connectivity, in which case 01:13:32.050 --> 01:13:34.810 align:middle line:84% yes, you'll get some of it wrong because the context isn't right 01:13:34.810 --> 01:13:36.910 align:middle line:84% or the species mapping isn't right. 01:13:36.910 --> 01:13:39.530 align:middle line:84% But will you be able to see enough connections 01:13:39.530 --> 01:13:40.810 align:middle line:90% to establish connectivity? 01:13:40.810 --> 01:13:44.320 align:middle line:90% 01:13:44.320 --> 01:13:46.890 align:middle line:90% AUDIENCE: [INAUDIBLE] 01:13:46.890 --> 01:13:47.663 align:middle line:90% 01:13:47.663 --> 01:13:48.330 align:middle line:90% TODD GOLUB: Yep. 01:13:48.330 --> 01:13:49.200 align:middle line:90% AUDIENCE: Sure. 01:13:49.200 --> 01:13:50.610 align:middle line:90% TODD GOLUB: Yep, certainly some. 01:13:50.610 --> 01:13:53.110 align:middle line:84% We're finding, actually-- I'll give you a couple of examples 01:13:53.110 --> 01:13:54.720 align:middle line:90% that-- 01:13:54.720 --> 01:13:57.270 align:middle line:84% while that is certain to be-- that context-dependence is 01:13:57.270 --> 01:13:58.350 align:middle line:90% certain to be the case-- 01:13:58.350 --> 01:13:59.340 align:middle line:84% AUDIENCE: There's lots that's shared. 01:13:59.340 --> 01:14:00.882 align:middle line:84% TODD GOLUB: There's lots that shared. 01:14:00.882 --> 01:14:04.030 align:middle line:84% And so, when we first started talking about this project, 01:14:04.030 --> 01:14:06.660 align:middle line:84% there were a lot of objections to the idea-- oh, still are, 01:14:06.660 --> 01:14:10.980 align:middle line:84% that, oh, what if you don't choose the right cell 01:14:10.980 --> 01:14:11.970 align:middle line:90% line to do this in? 01:14:11.970 --> 01:14:14.680 align:middle line:84% Ideally, you would do it in like 100 different cell types. 01:14:14.680 --> 01:14:17.550 align:middle line:84% But then it gets then it gets to be a serious experiment. 01:14:17.550 --> 01:14:21.432 align:middle line:84% Even in one cell line, it's a huge experiment. 01:14:21.432 --> 01:14:22.890 align:middle line:84% AUDIENCE: I think I told this group 01:14:22.890 --> 01:14:26.640 align:middle line:84% that I've seen the same thing in [INAUDIBLE] relationships. 01:14:26.640 --> 01:14:29.040 align:middle line:84% Many, many cell types that people wouldn't think they 01:14:29.040 --> 01:14:29.665 align:middle line:90% would be there. 01:14:29.665 --> 01:14:30.540 align:middle line:90% They're just there. 01:14:30.540 --> 01:14:33.720 align:middle line:90% TODD GOLUB: Yeah, exactly. 01:14:33.720 --> 01:14:38.070 align:middle line:84% So let me give a couple examples and then we'll end. 01:14:38.070 --> 01:14:41.370 align:middle line:84% And it also-- again, comes back to this idea of enrichment, 01:14:41.370 --> 01:14:45.120 align:middle line:84% looking for enrichment using a Kolmogorov-Smirnov enrichment 01:14:45.120 --> 01:14:47.350 align:middle line:90% test. 01:14:47.350 --> 01:14:50.130 align:middle line:84% So here is an experiment that was published 01:14:50.130 --> 01:14:53.610 align:middle line:84% not by us, but by a group at Abbott Laboratories, where they 01:14:53.610 --> 01:14:55.620 align:middle line:84% are interested in this class of drugs called 01:14:55.620 --> 01:14:58.020 align:middle line:90% histone deacetylase inhibitors. 01:14:58.020 --> 01:14:59.280 align:middle line:90% They took-- what is it? 01:14:59.280 --> 01:15:03.360 align:middle line:84% Five of these things, treated cells, 01:15:03.360 --> 01:15:06.540 align:middle line:84% took a common set of genes that were regulated, 01:15:06.540 --> 01:15:09.540 align:middle line:90% defined a signature of 22 genes. 01:15:09.540 --> 01:15:10.470 align:middle line:90% It's a gene set. 01:15:10.470 --> 01:15:14.400 align:middle line:84% It's the HDAC inhibitor gene set. 01:15:14.400 --> 01:15:17.130 align:middle line:90% We took breast cancer cells-- 01:15:17.130 --> 01:15:18.720 align:middle line:90% not the cell type they used-- 01:15:18.720 --> 01:15:21.820 align:middle line:84% and treated them with a bunch of different drugs, 01:15:21.820 --> 01:15:23.970 align:middle line:84% including one, valproic acid, which 01:15:23.970 --> 01:15:26.490 align:middle line:84% is actually used to treat seizures, as it turns out. 01:15:26.490 --> 01:15:29.820 align:middle line:84% It was later discovered valproic acid actually 01:15:29.820 --> 01:15:33.600 align:middle line:84% has histone deacetylase inhibitory activity. 01:15:33.600 --> 01:15:35.070 align:middle line:90% It's weak, but it's there. 01:15:35.070 --> 01:15:40.050 align:middle line:84% And we ask, can we see enrichment of this signature 01:15:40.050 --> 01:15:42.720 align:middle line:90% in any of those compounds? 01:15:42.720 --> 01:15:43.950 align:middle line:90% And the answer is, yes. 01:15:43.950 --> 01:15:47.282 align:middle line:84% So Trichostatin A was actually-- where is it? 01:15:47.282 --> 01:15:48.990 align:middle line:84% Here's the top one-- was one of the drugs 01:15:48.990 --> 01:15:51.030 align:middle line:84% that they used to define the signature. 01:15:51.030 --> 01:15:53.070 align:middle line:90% So we recovered it-- 01:15:53.070 --> 01:15:53.940 align:middle line:90% not surprisingly. 01:15:53.940 --> 01:15:55.950 align:middle line:84% But it's a little bit cheating because it 01:15:55.950 --> 01:15:57.060 align:middle line:90% wasn't a new example. 01:15:57.060 --> 01:15:59.820 align:middle line:84% But here, you see sodium valproate, third and fourth 01:15:59.820 --> 01:16:02.430 align:middle line:84% ranked on the list, which was not 01:16:02.430 --> 01:16:05.070 align:middle line:84% used to define the signature but simply based 01:16:05.070 --> 01:16:08.070 align:middle line:90% on the signature connectivity. 01:16:08.070 --> 01:16:11.340 align:middle line:84% Had we not known it, we could have rediscovered 01:16:11.340 --> 01:16:15.000 align:middle line:84% that-- we could have discovered that sodium valproate was 01:16:15.000 --> 01:16:18.150 align:middle line:84% an extract inhibitor because of this connectivity. 01:16:18.150 --> 01:16:19.230 align:middle line:90% You can also see here-- 01:16:19.230 --> 01:16:21.750 align:middle line:84% Trichostatin A-- so the signature is defined-- 01:16:21.750 --> 01:16:23.580 align:middle line:84% I can't even remember-- in one cell type. 01:16:23.580 --> 01:16:28.350 align:middle line:84% And we see it triggering the signature-- the HDAC inhibitor 01:16:28.350 --> 01:16:31.680 align:middle line:84% signature in breast cancer cells and in leukemia cells. 01:16:31.680 --> 01:16:35.790 align:middle line:84% So it's robust [INAUDIBLE] context. 01:16:35.790 --> 01:16:41.430 align:middle line:84% Interestingly, we also ran our little connectivity-- mini 01:16:41.430 --> 01:16:44.908 align:middle line:84% connectivity map across the oxidative phosphorylation 01:16:44.908 --> 01:16:45.950 align:middle line:90% signature that we define. 01:16:45.950 --> 01:16:48.270 align:middle line:84% AUDIENCE: Does the [INAUDIBLE] protease cause hypoglycemia? 01:16:48.270 --> 01:16:49.103 align:middle line:90% TODD GOLUB: It does. 01:16:49.103 --> 01:16:50.040 align:middle line:90% We didn't know that. 01:16:50.040 --> 01:16:51.912 align:middle line:84% We thought we discovered something new. 01:16:51.912 --> 01:16:53.370 align:middle line:84% AUDIENCE: It's the wrong specialty. 01:16:53.370 --> 01:16:54.600 align:middle line:84% TODD GOLUB: In the wrong specialty. 01:16:54.600 --> 01:16:56.430 align:middle line:84% Yeah, it was reported about 20 years ago 01:16:56.430 --> 01:17:00.510 align:middle line:84% that valproate causes hyperglycemia and modulates 01:17:00.510 --> 01:17:02.400 align:middle line:90% oxidative phosphorylation-- 01:17:02.400 --> 01:17:07.260 align:middle line:84% separable activities, separable from the anti seizure activity, 01:17:07.260 --> 01:17:09.660 align:middle line:84% separable from the HDAC inhibitor activity, 01:17:09.660 --> 01:17:13.240 align:middle line:84% and they trigger separate gene sets. 01:17:13.240 --> 01:17:15.860 align:middle line:90% So one last example-- 01:17:15.860 --> 01:17:21.240 align:middle line:84% a signature defined of a drug called rapamycin, 01:17:21.240 --> 01:17:26.100 align:middle line:84% which sits in this pathway simplified here of PI3 kinase 01:17:26.100 --> 01:17:31.920 align:middle line:90% AKT, and a protein called mTOR. 01:17:31.920 --> 01:17:36.640 align:middle line:84% And it turns out that, if you define 01:17:36.640 --> 01:17:39.280 align:middle line:84% the signature of rapamycin treatment 01:17:39.280 --> 01:17:45.110 align:middle line:84% here, its defining T-cells published by David Sabatini, 01:17:45.110 --> 01:17:50.060 align:middle line:84% and apply that signature to our connectivity matrix, 01:17:50.060 --> 01:17:53.030 align:middle line:84% you see that rapamycin itself is recovered in two different cell 01:17:53.030 --> 01:17:53.610 align:middle line:90% types. 01:17:53.610 --> 01:17:54.680 align:middle line:90% But also, this thing-- 01:17:54.680 --> 01:17:58.490 align:middle line:84% LY294002-- we didn't know what that was initially. 01:17:58.490 --> 01:17:59.660 align:middle line:90% But then you look it up. 01:17:59.660 --> 01:18:02.330 align:middle line:84% It's actually an inhibitor of PI3 kinase, 01:18:02.330 --> 01:18:05.060 align:middle line:90% which is upstream of mTOR. 01:18:05.060 --> 01:18:08.930 align:middle line:84% So it puts, in the same pathway, two drugs 01:18:08.930 --> 01:18:11.690 align:middle line:84% that act together and trigger the same signature. 01:18:11.690 --> 01:18:13.970 align:middle line:84% | again, I don't think it's going 01:18:13.970 --> 01:18:17.270 align:middle line:84% to be possible in the near term to actually reconstruct 01:18:17.270 --> 01:18:19.820 align:middle line:84% signal transduction pathways in their entirety. 01:18:19.820 --> 01:18:26.210 align:middle line:84% But to be able to put either genes or drugs together 01:18:26.210 --> 01:18:28.610 align:middle line:84% in a pathway not previously known 01:18:28.610 --> 01:18:31.400 align:middle line:84% to be in the same pathway, I think, 01:18:31.400 --> 01:18:33.240 align:middle line:84% is going to be possible with this approach. 01:18:33.240 --> 01:18:35.900 align:middle line:84% So we're committed to figuring out 01:18:35.900 --> 01:18:40.550 align:middle line:84% how to launch what would be a large scale public domain 01:18:40.550 --> 01:18:45.708 align:middle line:84% connectivity map project, where we would do these perturbations 01:18:45.708 --> 01:18:47.250 align:middle line:84% and put the data in the public domain 01:18:47.250 --> 01:18:49.910 align:middle line:84% so that people could use the data to find 01:18:49.910 --> 01:18:52.490 align:middle line:90% their own connections. 01:18:52.490 --> 01:18:55.000 align:middle line:90% AUDIENCE: [INAUDIBLE] 01:18:55.000 --> 01:18:57.070 align:middle line:90% 01:18:57.070 --> 01:18:59.670 align:middle line:84% There is a database that's recording 01:18:59.670 --> 01:19:02.395 align:middle line:90% some molecules [INAUDIBLE]. 01:19:02.395 --> 01:19:09.040 align:middle line:90% 01:19:09.040 --> 01:19:11.630 align:middle line:84% What will be the major difference between this project 01:19:11.630 --> 01:19:13.810 align:middle line:90% and that one? 01:19:13.810 --> 01:19:16.078 align:middle line:84% TODD GOLUB: I'm not sure I know what-- 01:19:16.078 --> 01:19:18.370 align:middle line:84% AUDIENCE: I think she's talking about the [INAUDIBLE].. 01:19:18.370 --> 01:19:20.350 align:middle line:90% An expression, right? 01:19:20.350 --> 01:19:22.390 align:middle line:84% AUDIENCE: Yeah, very expression when 01:19:22.390 --> 01:19:24.760 align:middle line:90% conditions on the molecules-- 01:19:24.760 --> 01:19:27.010 align:middle line:84% TODD GOLUB: On response to the molecules. 01:19:27.010 --> 01:19:29.750 align:middle line:84% This is actual response to the molecules. 01:19:29.750 --> 01:19:35.470 align:middle line:84% So that data set is the resting, untreated expression 01:19:35.470 --> 01:19:38.200 align:middle line:84% of the cell lines, which you can then correlate 01:19:38.200 --> 01:19:40.390 align:middle line:90% with how they respond. 01:19:40.390 --> 01:19:44.167 align:middle line:84% This is the acute changes of response, 01:19:44.167 --> 01:19:46.000 align:middle line:84% because the nice thing about that experiment 01:19:46.000 --> 01:19:48.250 align:middle line:84% is that you only have to measure each cell line once 01:19:48.250 --> 01:19:50.417 align:middle line:84% and then you can correlate it with all those things. 01:19:50.417 --> 01:19:53.858 align:middle line:84% Here, you actually have to do a microarray for each different-- 01:19:53.858 --> 01:19:55.150 align:middle line:90% AUDIENCE: Condition time point. 01:19:55.150 --> 01:19:56.020 align:middle line:90% TODD GOLUB: Yep. 01:19:56.020 --> 01:19:57.400 align:middle line:90% OK, so the last slide. 01:19:57.400 --> 01:19:58.660 align:middle line:90% Whoops, let's skip that. 01:19:58.660 --> 01:20:02.450 align:middle line:90% 01:20:02.450 --> 01:20:05.560 align:middle line:90% There's some future challenges. 01:20:05.560 --> 01:20:09.520 align:middle line:84% Just to remember that these prognostic signatures 01:20:09.520 --> 01:20:14.941 align:middle line:84% that everyone's developing are really a function of therapy. 01:20:14.941 --> 01:20:19.140 align:middle line:84% They're likely only to be useful if you continue 01:20:19.140 --> 01:20:20.140 align:middle line:90% to use the same therapy. 01:20:20.140 --> 01:20:22.490 align:middle line:84% So in this medulloblastoma example, 01:20:22.490 --> 01:20:24.703 align:middle line:84% we can predict things pretty well as long 01:20:24.703 --> 01:20:26.620 align:middle line:84% as we're trying to predict response to therapy 01:20:26.620 --> 01:20:28.690 align:middle line:90% that was given 10 years ago. 01:20:28.690 --> 01:20:32.680 align:middle line:84% But the therapy has evolved, and it's not certain 01:20:32.680 --> 01:20:35.995 align:middle line:84% that our classifier will still hold. 01:20:35.995 --> 01:20:38.620 align:middle line:84% This is going to be challenging because the clinical trials are 01:20:38.620 --> 01:20:42.420 align:middle line:84% generally small and underpowered to do this kind of thing. 01:20:42.420 --> 01:20:46.150 align:middle line:84% Still not clear how, once these signatures are actually 01:20:46.150 --> 01:20:49.062 align:middle line:84% vetted and validated, what form will they 01:20:49.062 --> 01:20:50.770 align:middle line:84% take when they actually go to the clinic? 01:20:50.770 --> 01:20:52.475 align:middle line:90% Will they be a microarray? 01:20:52.475 --> 01:20:52.975 align:middle line:90% Maybe. 01:20:52.975 --> 01:20:56.480 align:middle line:90% 01:20:56.480 --> 01:20:59.930 align:middle line:84% I've given you some ideas about how 01:20:59.930 --> 01:21:04.070 align:middle line:84% one might do signature-based chemical discovery. 01:21:04.070 --> 01:21:07.250 align:middle line:84% But turning a chemical into a drug 01:21:07.250 --> 01:21:10.050 align:middle line:90% is a big deal and not easy. 01:21:10.050 --> 01:21:13.257 align:middle line:84% And so for that reason, pharmaceutical companies 01:21:13.257 --> 01:21:14.840 align:middle line:84% aren't dropping their current approach 01:21:14.840 --> 01:21:17.390 align:middle line:84% to drug discovery in favor of this. 01:21:17.390 --> 01:21:19.790 align:middle line:84% And in general, how to integrate these kind of signatures 01:21:19.790 --> 01:21:21.290 align:middle line:84% into the drug development process 01:21:21.290 --> 01:21:23.610 align:middle line:84% is still something to think about. 01:21:23.610 --> 01:21:28.440 align:middle line:84% But I still think this notion of using these tools-- 01:21:28.440 --> 01:21:30.110 align:middle line:84% whatever personalized medicine means-- 01:21:30.110 --> 01:21:33.560 align:middle line:84% but to gain more insight into the particulars 01:21:33.560 --> 01:21:36.770 align:middle line:84% of a given individual's disease to better match them 01:21:36.770 --> 01:21:40.370 align:middle line:84% with an existing or new therapeutic 01:21:40.370 --> 01:21:42.890 align:middle line:84% is likely to be here to stay even 01:21:42.890 --> 01:21:46.850 align:middle line:84% though there aren't a lot of examples of it happening yet. 01:21:46.850 --> 01:21:49.120 align:middle line:90% That's all I have.