WEBVTT 00:00:00.000 --> 00:00:01.380 align:middle line:90% 00:00:01.380 --> 00:00:03.570 align:middle line:84% MARCO RAMONI: What I'm going to talk about 00:00:03.570 --> 00:00:06.390 align:middle line:84% after this little introduction about microarrays 00:00:06.390 --> 00:00:12.720 align:middle line:84% is how to analyze this BLAST data. 00:00:12.720 --> 00:00:18.210 align:middle line:84% And the principle that I try to present to you 00:00:18.210 --> 00:00:22.590 align:middle line:84% is that there is no such a thing as putting 00:00:22.590 --> 00:00:24.780 align:middle line:84% your data into a freaking machine 00:00:24.780 --> 00:00:26.790 align:middle line:90% and expecting to get an answer. 00:00:26.790 --> 00:00:30.120 align:middle line:84% The type of analysis you make is always 00:00:30.120 --> 00:00:32.130 align:middle line:84% related to the question you're asking. 00:00:32.130 --> 00:00:34.920 align:middle line:84% This has to be a completely stupid point. 00:00:34.920 --> 00:00:40.180 align:middle line:84% But the tragedy of a lot of this field is that it's not. 00:00:40.180 --> 00:00:44.700 align:middle line:84% And a lot of people usually try to answer the same question 00:00:44.700 --> 00:00:48.840 align:middle line:84% using different methods and different questions 00:00:48.840 --> 00:00:51.870 align:middle line:84% using the same methods, which is even more disturbing. 00:00:51.870 --> 00:00:56.490 align:middle line:84% What I'm trying to tell you today is what kind of problems 00:00:56.490 --> 00:01:03.390 align:middle line:84% you can tackle with this kind of data and what kind of analysis 00:01:03.390 --> 00:01:05.819 align:middle line:84% you need to answer every different question. 00:01:05.819 --> 00:01:08.370 align:middle line:90% 00:01:08.370 --> 00:01:09.840 align:middle line:90% It's going to be very basic. 00:01:09.840 --> 00:01:14.950 align:middle line:84% I will introduce some kind of advanced notions at the end. 00:01:14.950 --> 00:01:18.000 align:middle line:84% But most of the rest is very basic 00:01:18.000 --> 00:01:20.190 align:middle line:84% and is what is routinely done by people 00:01:20.190 --> 00:01:23.430 align:middle line:84% in papers, in genome centers, and things like this. 00:01:23.430 --> 00:01:27.960 align:middle line:90% 00:01:27.960 --> 00:01:30.360 align:middle line:90% And this is important for you. 00:01:30.360 --> 00:01:35.400 align:middle line:84% Because at the very end, I will tell you the bad news. 00:01:35.400 --> 00:01:36.540 align:middle line:90% You have an assignment. 00:01:36.540 --> 00:01:38.880 align:middle line:84% And you have to use a couple of programs 00:01:38.880 --> 00:01:41.620 align:middle line:90% that are describing this thing. 00:01:41.620 --> 00:01:44.760 align:middle line:84% So what I'm going to do is start from the microarrays, 00:01:44.760 --> 00:01:47.220 align:middle line:84% tell you what you do with supervised classification 00:01:47.220 --> 00:01:49.980 align:middle line:84% and differentiable analysis, argue the prediction 00:01:49.980 --> 00:01:51.420 align:middle line:90% and validate your results. 00:01:51.420 --> 00:01:54.570 align:middle line:84% How do you do unsupervised analysis using basically 00:01:54.570 --> 00:01:57.120 align:middle line:84% clustering for different type of methods 00:01:57.120 --> 00:01:59.730 align:middle line:84% and different types of experiments? 00:01:59.730 --> 00:02:01.890 align:middle line:84% And then, at the end, I'm going to talk to you 00:02:01.890 --> 00:02:06.540 align:middle line:84% about base networks, which are those things that a few of you 00:02:06.540 --> 00:02:08.429 align:middle line:90% know are my passion. 00:02:08.429 --> 00:02:11.670 align:middle line:90% 00:02:11.670 --> 00:02:15.900 align:middle line:84% This is exactly the same slide I had last week, two weeks ago. 00:02:15.900 --> 00:02:18.870 align:middle line:84% This is central to molecular biology. 00:02:18.870 --> 00:02:21.690 align:middle line:84% I will not recapitulate it here, except to say 00:02:21.690 --> 00:02:24.360 align:middle line:90% that DNA gets copied in RNA. 00:02:24.360 --> 00:02:25.980 align:middle line:90% RNA gets copied in proteins. 00:02:25.980 --> 00:02:27.930 align:middle line:90% Proteins do all the job. 00:02:27.930 --> 00:02:31.410 align:middle line:84% So last week, we talked about DNA. 00:02:31.410 --> 00:02:34.980 align:middle line:84% Today, we are talking about RNA, OK? 00:02:34.980 --> 00:02:39.840 align:middle line:84% So all your cells in your body come from the same cell. 00:02:39.840 --> 00:02:45.060 align:middle line:84% And they all have the same DNA code. 00:02:45.060 --> 00:02:49.260 align:middle line:84% What makes a neuron different from a fingernail 00:02:49.260 --> 00:02:54.280 align:middle line:84% is that the proteins that are made of this are different. 00:02:54.280 --> 00:02:57.930 align:middle line:84% So the same code will express different type 00:02:57.930 --> 00:03:02.280 align:middle line:84% of RNA, which in turn will be turned into different proteins. 00:03:02.280 --> 00:03:05.820 align:middle line:84% And these will give different cells their different 00:03:05.820 --> 00:03:08.470 align:middle line:90% in nature. 00:03:08.470 --> 00:03:12.480 align:middle line:84% The idea of studying this is called functional genomics. 00:03:12.480 --> 00:03:15.510 align:middle line:84% What we talked about last week was structural genomics. 00:03:15.510 --> 00:03:17.800 align:middle line:84% We were looking at the structure of the DNA. 00:03:17.800 --> 00:03:20.700 align:middle line:84% Now we're looking at the action of this DNA. 00:03:20.700 --> 00:03:22.320 align:middle line:84% And we're looking at the function 00:03:22.320 --> 00:03:27.660 align:middle line:84% that each different cell performs while it exists. 00:03:27.660 --> 00:03:30.960 align:middle line:84% The aim of the game is elucidate functions and interactions 00:03:30.960 --> 00:03:32.280 align:middle line:90% among genes. 00:03:32.280 --> 00:03:34.970 align:middle line:84% Now functional genomics is a very old thing. 00:03:34.970 --> 00:03:36.720 align:middle line:84% I mean, you don't need computer scientists 00:03:36.720 --> 00:03:38.460 align:middle line:90% to do functional genomics. 00:03:38.460 --> 00:03:45.480 align:middle line:84% Functional genomics means try to understand what is 00:03:45.480 --> 00:03:46.800 align:middle line:90% expressed in a particular cell. 00:03:46.800 --> 00:03:48.040 align:middle line:90% And you can do it by hand. 00:03:48.040 --> 00:03:49.770 align:middle line:84% You can do it one gene in the time. 00:03:49.770 --> 00:03:53.940 align:middle line:84% And people have been doing one gene at a time forever. 00:03:53.940 --> 00:03:57.420 align:middle line:84% What is changed in modern functional genomics 00:03:57.420 --> 00:03:59.280 align:middle line:84% is the introduction of microarrays, 00:03:59.280 --> 00:04:01.470 align:middle line:84% which are these platforms that allows 00:04:01.470 --> 00:04:05.790 align:middle line:84% us to look at the whole transcripts, all the RNA, 00:04:05.790 --> 00:04:08.670 align:middle line:84% every gene in a cell and see what is expressed 00:04:08.670 --> 00:04:10.170 align:middle line:90% and what is not. 00:04:10.170 --> 00:04:11.350 align:middle line:90% This is what is changed. 00:04:11.350 --> 00:04:14.400 align:middle line:84% And this is why you need, at this point, computer 00:04:14.400 --> 00:04:16.019 align:middle line:90% scientists. 00:04:16.019 --> 00:04:19.560 align:middle line:84% But there is a little different change in this. 00:04:19.560 --> 00:04:22.710 align:middle line:84% There is an intellectual change in this, a very dramatic 00:04:22.710 --> 00:04:24.210 align:middle line:90% intellectual change in this. 00:04:24.210 --> 00:04:28.440 align:middle line:90% 00:04:28.440 --> 00:04:32.370 align:middle line:84% If I have to pick up one gene to see if this gene is expressed 00:04:32.370 --> 00:04:36.150 align:middle line:84% in a cell, I have to go in and read papers and, at some point, 00:04:36.150 --> 00:04:39.450 align:middle line:84% decide how to allocate my next two weeks or two years 00:04:39.450 --> 00:04:41.820 align:middle line:84% to see the expression of this particular gene. 00:04:41.820 --> 00:04:43.320 align:middle line:90% Sometimes I need money. 00:04:43.320 --> 00:04:45.270 align:middle line:90% So I need to put in writing-- 00:04:45.270 --> 00:04:48.330 align:middle line:84% to somebody else that hopefully will give me money-- 00:04:48.330 --> 00:04:53.460 align:middle line:84% why this darn gene is important, right? 00:04:53.460 --> 00:04:56.790 align:middle line:84% Now with microarrays, they don't need to do that. 00:04:56.790 --> 00:04:59.340 align:middle line:84% With microarrays, I use one microarray. 00:04:59.340 --> 00:05:03.390 align:middle line:84% And for a tissue, what they observe are 40,000 genes. 00:05:03.390 --> 00:05:06.360 align:middle line:84% I don't need to justify which gene I'm interested in. 00:05:06.360 --> 00:05:08.520 align:middle line:90% I look at all of them. 00:05:08.520 --> 00:05:11.490 align:middle line:84% And this is a very interesting consequence. 00:05:11.490 --> 00:05:13.390 align:middle line:90% One-- one is cool. 00:05:13.390 --> 00:05:15.240 align:middle line:84% One, I can look everything in action. 00:05:15.240 --> 00:05:19.830 align:middle line:84% I can try to be surprised by my results. 00:05:19.830 --> 00:05:22.830 align:middle line:84% But the other thing is that I get very different type 00:05:22.830 --> 00:05:24.360 align:middle line:90% of information. 00:05:24.360 --> 00:05:28.440 align:middle line:84% That is, suppose I have been spending two years seeing 00:05:28.440 --> 00:05:31.980 align:middle line:84% if a particular gene is expressed in a particular cell. 00:05:31.980 --> 00:05:36.540 align:middle line:84% And what you get at the end are two pictures-- 00:05:36.540 --> 00:05:39.330 align:middle line:84% one of this gene in a normal cell 00:05:39.330 --> 00:05:41.670 align:middle line:84% and the other in the cell you're interested in-- 00:05:41.670 --> 00:05:44.040 align:middle line:84% and see if there is some kind of change. 00:05:44.040 --> 00:05:47.040 align:middle line:84% And on one side you see a ball this size. 00:05:47.040 --> 00:05:49.560 align:middle line:84% And on the other side you see a ball this size. 00:05:49.560 --> 00:05:55.470 align:middle line:84% You take pictures and you send in your paper, right? 00:05:55.470 --> 00:05:57.390 align:middle line:90% This is what people do. 00:05:57.390 --> 00:06:00.330 align:middle line:84% And probably, you are going to make in your-- 00:06:00.330 --> 00:06:02.350 align:middle line:90% after you spend two years-- 00:06:02.350 --> 00:06:05.690 align:middle line:84% maybe even if the ball is not really that big, 00:06:05.690 --> 00:06:08.190 align:middle line:84% you are going to make a very long argument that this ball is 00:06:08.190 --> 00:06:08.982 align:middle line:90% really, really big. 00:06:08.982 --> 00:06:11.430 align:middle line:84% And there is some particular reason 00:06:11.430 --> 00:06:16.380 align:middle line:84% why you shouldn't take this ball as small as it is, right? 00:06:16.380 --> 00:06:20.970 align:middle line:84% Now when you have microarrays, you're measuring all the genes. 00:06:20.970 --> 00:06:24.390 align:middle line:84% So even if the ball is not that small, 00:06:24.390 --> 00:06:27.270 align:middle line:84% maybe there are lot of other balls that are much bigger. 00:06:27.270 --> 00:06:29.355 align:middle line:84% So at that point, it's kind of difficult to say, 00:06:29.355 --> 00:06:31.920 align:middle line:84% you know, my gene is kind of interesting. 00:06:31.920 --> 00:06:34.950 align:middle line:84% Because it ends up that your gene is expressed 00:06:34.950 --> 00:06:39.780 align:middle line:90% as another 20,000 genes, right? 00:06:39.780 --> 00:06:41.880 align:middle line:84% So you get another piece of information 00:06:41.880 --> 00:06:44.190 align:middle line:90% by looking at the entire genome. 00:06:44.190 --> 00:06:47.130 align:middle line:84% You also get what are the most dramatic changes. 00:06:47.130 --> 00:06:49.410 align:middle line:84% What are the most dramatic things that 00:06:49.410 --> 00:06:51.880 align:middle line:90% happen in that particular cell? 00:06:51.880 --> 00:06:54.660 align:middle line:84% And this is kind of interesting because you produce 00:06:54.660 --> 00:06:55.950 align:middle line:90% a new intellectual style. 00:06:55.950 --> 00:06:59.790 align:middle line:84% The new intellectual style is not hypothesis-driven and has 00:06:59.790 --> 00:07:05.040 align:middle line:84% been a disaster for the biomedical research culture-- 00:07:05.040 --> 00:07:08.040 align:middle line:84% for us, basically-- and still is a big problem 00:07:08.040 --> 00:07:09.600 align:middle line:90% when you submit grants. 00:07:09.600 --> 00:07:13.410 align:middle line:84% When you submit grants, the way you write the grants is to say, 00:07:13.410 --> 00:07:14.820 align:middle line:90% this is my hypothesis. 00:07:14.820 --> 00:07:17.130 align:middle line:84% This is why I think this is true. 00:07:17.130 --> 00:07:20.250 align:middle line:84% This is why I think this is important and interesting. 00:07:20.250 --> 00:07:21.990 align:middle line:84% This is what other people have seen-- 00:07:21.990 --> 00:07:24.630 align:middle line:84% and myself, I have seen-- to support this hypothesis. 00:07:24.630 --> 00:07:26.670 align:middle line:90% And this is what I plan to do. 00:07:26.670 --> 00:07:29.790 align:middle line:90% Now here, the hypothesis-- 00:07:29.790 --> 00:07:31.450 align:middle line:90% what the hell is the hypothesis? 00:07:31.450 --> 00:07:32.160 align:middle line:90% I don't know. 00:07:32.160 --> 00:07:34.380 align:middle line:90% I have a very vague scheme. 00:07:34.380 --> 00:07:36.930 align:middle line:84% And I can say, well, you know, my hypothesis 00:07:36.930 --> 00:07:41.520 align:middle line:84% is that the genes expressed in prostate cancer 00:07:41.520 --> 00:07:46.320 align:middle line:84% are different from the genes not expressed in the normal tissue. 00:07:46.320 --> 00:07:47.760 align:middle line:90% Gee, what a hypothesis. 00:07:47.760 --> 00:07:49.560 align:middle line:84% I mean, you don't need to go to grad school 00:07:49.560 --> 00:07:51.660 align:middle line:84% to come up with an hypothesis like this, right? 00:07:51.660 --> 00:07:56.140 align:middle line:84% But then the intellectual styles is completely different. 00:07:56.140 --> 00:08:02.290 align:middle line:84% So we are heading into something that 00:08:02.290 --> 00:08:06.040 align:middle line:84% has some simple statistics and some simple technology 00:08:06.040 --> 00:08:10.600 align:middle line:84% but has a much broader impact on the way people 00:08:10.600 --> 00:08:13.180 align:middle line:90% think about biology. 00:08:13.180 --> 00:08:17.080 align:middle line:84% One of my dearest quotations is from a physicists 00:08:17.080 --> 00:08:19.570 align:middle line:84% of the 19th century that used to say 00:08:19.570 --> 00:08:23.380 align:middle line:84% that, "there are two types of science, physics 00:08:23.380 --> 00:08:26.440 align:middle line:90% and STEM collection." 00:08:26.440 --> 00:08:29.020 align:middle line:84% What this mean is that physics provides 00:08:29.020 --> 00:08:32.140 align:middle line:84% mathematical, quantitative models of phenomena. 00:08:32.140 --> 00:08:35.320 align:middle line:84% STEM collections is going around and measure animals, 00:08:35.320 --> 00:08:37.360 align:middle line:84% and put them in your collection, your album, 00:08:37.360 --> 00:08:41.679 align:middle line:84% and show them to friends, and maybe arrange them in some way. 00:08:41.679 --> 00:08:45.220 align:middle line:84% What this thing is doing to the intellectual landscape 00:08:45.220 --> 00:08:49.930 align:middle line:84% of modern biology is, hopefully, to turn STEM collection 00:08:49.930 --> 00:08:52.030 align:middle line:84% into highly quantitative science. 00:08:52.030 --> 00:08:54.960 align:middle line:90% 00:08:54.960 --> 00:08:57.180 align:middle line:84% The characters that are behind all these things 00:08:57.180 --> 00:08:58.950 align:middle line:90% are microarray technology-- 00:08:58.950 --> 00:09:00.750 align:middle line:90% microarrays. 00:09:00.750 --> 00:09:04.740 align:middle line:84% They are able to measure the expression of thousands 00:09:04.740 --> 00:09:08.520 align:middle line:90% of genes at the same time. 00:09:08.520 --> 00:09:11.790 align:middle line:84% And now we have microarrays that are able-- 00:09:11.790 --> 00:09:14.670 align:middle line:84% on a little piece of plastic this size-- 00:09:14.670 --> 00:09:18.840 align:middle line:84% they are able to measure the expression of 55,000 00:09:18.840 --> 00:09:23.250 align:middle line:84% transcripts, which includes all the estimated 35,000, 40,000 00:09:23.250 --> 00:09:26.280 align:middle line:90% genes in the human genome. 00:09:26.280 --> 00:09:30.570 align:middle line:84% Technically, a microarray is-- although it is called an array, 00:09:30.570 --> 00:09:32.380 align:middle line:90% it's actually a vector. 00:09:32.380 --> 00:09:36.810 align:middle line:84% So for each cell in my array, I have 00:09:36.810 --> 00:09:40.380 align:middle line:84% a label that tells me the name of the gene there. 00:09:40.380 --> 00:09:42.630 align:middle line:84% And then I have the value of expression 00:09:42.630 --> 00:09:44.320 align:middle line:90% for that particular cell. 00:09:44.320 --> 00:09:47.820 align:middle line:84% So at the end, when they put them together, 00:09:47.820 --> 00:09:51.090 align:middle line:84% they become a vector that associates to each gene-- 00:09:51.090 --> 00:09:52.770 align:middle line:84% its particular value of expression 00:09:52.770 --> 00:09:56.258 align:middle line:84% in a particular cell or a tissue. 00:09:56.258 --> 00:09:58.050 align:middle line:84% Therefore, the arrays-- because, of course, 00:09:58.050 --> 00:10:02.970 align:middle line:84% putting down a vector of 25,000 genes, of 25,000 little cells, 00:10:02.970 --> 00:10:05.970 align:middle line:84% is less convenient from a geometrical point of view 00:10:05.970 --> 00:10:07.700 align:middle line:84% than putting down a square thing. 00:10:07.700 --> 00:10:08.820 align:middle line:90% It takes less space. 00:10:08.820 --> 00:10:12.750 align:middle line:84% They are called arrays but that can be kind of misleading. 00:10:12.750 --> 00:10:15.840 align:middle line:84% And there are two types of arrays that are currently 00:10:15.840 --> 00:10:18.150 align:middle line:90% used more frequently. 00:10:18.150 --> 00:10:20.610 align:middle line:84% One are called cDNA and the other 00:10:20.610 --> 00:10:22.390 align:middle line:84% are called oligonucleotide microarrays. 00:10:22.390 --> 00:10:24.390 align:middle line:84% I'm going to tell you in a second what they are. 00:10:24.390 --> 00:10:28.210 align:middle line:90% 00:10:28.210 --> 00:10:30.900 align:middle line:90% How does these things work? 00:10:30.900 --> 00:10:34.410 align:middle line:84% It works by reversing the natural phenomenon 00:10:34.410 --> 00:10:37.500 align:middle line:90% of transcription, right? 00:10:37.500 --> 00:10:44.760 align:middle line:84% So the idea is that I have special glue in each cell that 00:10:44.760 --> 00:10:47.010 align:middle line:84% is shaped exactly as the transcript 00:10:47.010 --> 00:10:49.260 align:middle line:90% that they want to measure. 00:10:49.260 --> 00:10:54.900 align:middle line:84% Then, I will have my tissues that go on this microarray. 00:10:54.900 --> 00:11:01.080 align:middle line:84% And by some diverse method, they will hybridize. 00:11:01.080 --> 00:11:04.890 align:middle line:84% They will attach to the cells that are specific to them. 00:11:04.890 --> 00:11:06.790 align:middle line:90% And then, I wash them away. 00:11:06.790 --> 00:11:09.540 align:middle line:84% And what they have to do at the end 00:11:09.540 --> 00:11:13.860 align:middle line:84% is simply to measure how many of this RNA 00:11:13.860 --> 00:11:19.470 align:middle line:84% or how much of this RNA is left of a particular cell. 00:11:19.470 --> 00:11:21.480 align:middle line:90% How do I do it practically? 00:11:21.480 --> 00:11:23.200 align:middle line:90% It works like this. 00:11:23.200 --> 00:11:24.250 align:middle line:90% I get a bunch of tissues. 00:11:24.250 --> 00:11:26.170 align:middle line:84% Let's say I have only one tissue, for the time being. 00:11:26.170 --> 00:11:28.230 align:middle line:84% I'm producing one, single microarray, right? 00:11:28.230 --> 00:11:29.790 align:middle line:90% I get my tissues. 00:11:29.790 --> 00:11:31.350 align:middle line:90% I extract the RNA. 00:11:31.350 --> 00:11:35.580 align:middle line:84% And I tag this RNA with pieces of transcribed RNA 00:11:35.580 --> 00:11:38.040 align:middle line:90% with some fluorescent dye. 00:11:38.040 --> 00:11:42.600 align:middle line:84% Then, I put them in the dishwasher, [INAUDIBLE] 00:11:42.600 --> 00:11:45.660 align:middle line:84% station, which is microarray washer actually. 00:11:45.660 --> 00:11:47.340 align:middle line:90% And I hybridize it. 00:11:47.340 --> 00:11:48.480 align:middle line:90% I put it there. 00:11:48.480 --> 00:11:51.073 align:middle line:90% And then I scan it. 00:11:51.073 --> 00:11:52.740 align:middle line:84% So what happened is that-- you remember, 00:11:52.740 --> 00:11:57.960 align:middle line:84% this RNA is left attached to this particular cells. 00:11:57.960 --> 00:12:00.150 align:middle line:90% And because I have tagged it-- 00:12:00.150 --> 00:12:02.400 align:middle line:84% it's tagged with a fluorescent dye-- 00:12:02.400 --> 00:12:05.880 align:middle line:84% I would have more intensity in those places in which 00:12:05.880 --> 00:12:07.900 align:middle line:90% a lot of things are attached. 00:12:07.900 --> 00:12:09.780 align:middle line:84% So once I wash it, I can actually 00:12:09.780 --> 00:12:13.740 align:middle line:84% use a scanner, like the one for your picture at home, exactly 00:12:13.740 --> 00:12:14.700 align:middle line:90% a scanner. 00:12:14.700 --> 00:12:18.420 align:middle line:84% And that scanner will come up with a picture that 00:12:18.420 --> 00:12:22.170 align:middle line:84% looks like this one, in which each ball represents 00:12:22.170 --> 00:12:25.740 align:middle line:84% how much RNA is left in that particular spot. 00:12:25.740 --> 00:12:28.320 align:middle line:84% Because I have created or somebody else 00:12:28.320 --> 00:12:30.930 align:middle line:84% I pay has created that particular spot, 00:12:30.930 --> 00:12:33.360 align:middle line:84% knows exactly what is the transcript that is there. 00:12:33.360 --> 00:12:38.970 align:middle line:84% And they can tell you that the third spot from the left 00:12:38.970 --> 00:12:41.520 align:middle line:84% is gene [INAUDIBLE] [? alfalfa. ?] 00:12:41.520 --> 00:12:44.520 align:middle line:84% I can go measure the intensity of this gene 00:12:44.520 --> 00:12:47.430 align:middle line:84% and then turn everything into a database. 00:12:47.430 --> 00:12:51.060 align:middle line:84% I know for each probe what it represents. 00:12:51.060 --> 00:12:55.290 align:middle line:84% And for each microarray, I can measure how much of that 00:12:55.290 --> 00:12:59.685 align:middle line:90% is estimated in my sample. 00:12:59.685 --> 00:13:00.185 align:middle line:90% Question? 00:13:00.185 --> 00:13:03.140 align:middle line:90% 00:13:03.140 --> 00:13:04.710 align:middle line:90% This is how it looks like. 00:13:04.710 --> 00:13:06.920 align:middle line:84% This is cDNA microarrays. cDNA microarray-- 00:13:06.920 --> 00:13:09.995 align:middle line:90% I lied a little. 00:13:09.995 --> 00:13:12.530 align:middle line:90% You can't use scanners for this. 00:13:12.530 --> 00:13:16.460 align:middle line:84% You use some kind of laser stuff to read these points. 00:13:16.460 --> 00:13:20.000 align:middle line:84% But the idea is that in this microarrays, 00:13:20.000 --> 00:13:26.420 align:middle line:84% you copy the entire transcript of a gene. 00:13:26.420 --> 00:13:29.360 align:middle line:84% So I know how a particular gene-- 00:13:29.360 --> 00:13:31.100 align:middle line:84% what is the sequence that is transcribed 00:13:31.100 --> 00:13:32.300 align:middle line:90% for a particular gene. 00:13:32.300 --> 00:13:36.330 align:middle line:84% I make 1,000 clones or a million clones of this. 00:13:36.330 --> 00:13:38.390 align:middle line:90% And I put them in one spot. 00:13:38.390 --> 00:13:41.090 align:middle line:84% Now this microarray has two channels 00:13:41.090 --> 00:13:44.285 align:middle line:84% that I can read using laser scanning, OK? 00:13:44.285 --> 00:13:48.180 align:middle line:90% 00:13:48.180 --> 00:13:49.650 align:middle line:90% So I will have two samples. 00:13:49.650 --> 00:13:53.000 align:middle line:84% And I dye one in red and one in green. 00:13:53.000 --> 00:13:54.740 align:middle line:84% And I put them on this microarray. 00:13:54.740 --> 00:13:57.740 align:middle line:84% And they will compete competitively-- hybridize 00:13:57.740 --> 00:13:59.010 align:middle line:90% to this one. 00:13:59.010 --> 00:14:03.710 align:middle line:84% So if both things are highly expressed, what I will see 00:14:03.710 --> 00:14:05.640 align:middle line:90% is something that is yellow. 00:14:05.640 --> 00:14:08.060 align:middle line:84% If none of them is expressed, it would be some kind 00:14:08.060 --> 00:14:10.550 align:middle line:90% of black-grayish thing. 00:14:10.550 --> 00:14:13.310 align:middle line:84% And if the green is more expressed than the red-- 00:14:13.310 --> 00:14:14.330 align:middle line:90% will be greenish. 00:14:14.330 --> 00:14:16.430 align:middle line:84% And if the red is more expressed than the green, 00:14:16.430 --> 00:14:18.560 align:middle line:90% it would be reddish. 00:14:18.560 --> 00:14:20.240 align:middle line:84% And you can see here-- you see a lot 00:14:20.240 --> 00:14:23.600 align:middle line:84% of yellow balls, few green balls, and few red, 00:14:23.600 --> 00:14:25.040 align:middle line:90% and a lot of black. 00:14:25.040 --> 00:14:27.580 align:middle line:90% 00:14:27.580 --> 00:14:28.240 align:middle line:90% What's that? 00:14:28.240 --> 00:14:29.973 align:middle line:90% AUDIENCE: Just a basic question. 00:14:29.973 --> 00:14:31.390 align:middle line:84% What kind of different information 00:14:31.390 --> 00:14:33.223 align:middle line:84% will you get from the [? array ?] as opposed 00:14:33.223 --> 00:14:34.600 align:middle line:90% to [INAUDIBLE] intensity. 00:14:34.600 --> 00:14:35.710 align:middle line:90% Do you get any different-- 00:14:35.710 --> 00:14:37.930 align:middle line:84% MARCO RAMONI: No, this was just because it 00:14:37.930 --> 00:14:42.760 align:middle line:84% was the original-- the original one were made this way. 00:14:42.760 --> 00:14:44.140 align:middle line:90% To use a scanner-- 00:14:44.140 --> 00:14:47.770 align:middle line:84% which I guess what you're going to-- to use a scanner, 00:14:47.770 --> 00:14:51.010 align:middle line:84% you need to use silicon technology. 00:14:51.010 --> 00:14:52.900 align:middle line:90% And these are glass slides. 00:14:52.900 --> 00:14:56.630 align:middle line:84% So to build this kind of things is much easier. 00:14:56.630 --> 00:14:58.520 align:middle line:84% You can build these things at home. 00:14:58.520 --> 00:15:02.980 align:middle line:84% You can buy a robot that will spot the stains for you. 00:15:02.980 --> 00:15:04.990 align:middle line:84% The oligonucleotide microarrays-- 00:15:04.990 --> 00:15:07.500 align:middle line:90% I'll show you in second-- 00:15:07.500 --> 00:15:10.630 align:middle line:84% use another type of technology that requires really 00:15:10.630 --> 00:15:12.110 align:middle line:90% a production line. 00:15:12.110 --> 00:15:13.550 align:middle line:84% So it's not something you can do. 00:15:13.550 --> 00:15:15.850 align:middle line:84% So these things actually offer you flexibility. 00:15:15.850 --> 00:15:18.790 align:middle line:84% If you're interested in 1,000 genes rather than 40,000, 00:15:18.790 --> 00:15:19.640 align:middle line:90% you can do it. 00:15:19.640 --> 00:15:22.090 align:middle line:84% The other thing is that it costs much less. 00:15:22.090 --> 00:15:24.250 align:middle line:84% The problem is that because you copy 00:15:24.250 --> 00:15:29.320 align:middle line:84% the entire transcript, when you take your RNA out of a cell, 00:15:29.320 --> 00:15:31.030 align:middle line:90% you crash it. 00:15:31.030 --> 00:15:32.572 align:middle line:90% Then you dye it. 00:15:32.572 --> 00:15:34.030 align:middle line:84% Now what happened is that there may 00:15:34.030 --> 00:15:36.395 align:middle line:84% be a lot of crap floating around that is really not 00:15:36.395 --> 00:15:37.270 align:middle line:90% related to your gene. 00:15:37.270 --> 00:15:40.450 align:middle line:84% There may be very, very small fragments 00:15:40.450 --> 00:15:43.750 align:middle line:84% that are going hybridized to some random sequence 00:15:43.750 --> 00:15:45.760 align:middle line:84% in your clones, just because they are 00:15:45.760 --> 00:15:47.230 align:middle line:90% very small to hybridized there. 00:15:47.230 --> 00:15:49.930 align:middle line:84% But still, they will bring fluorescent dye 00:15:49.930 --> 00:15:51.460 align:middle line:90% to that particular spot. 00:15:51.460 --> 00:15:53.260 align:middle line:84% So the precision of this measurement 00:15:53.260 --> 00:15:55.690 align:middle line:84% is not really as great as it could be. 00:15:55.690 --> 00:15:59.290 align:middle line:84% You get a lot of random hybridization. 00:15:59.290 --> 00:16:02.150 align:middle line:84% There are tricks you can play for this. 00:16:02.150 --> 00:16:04.240 align:middle line:84% So if you want to make a comparative experiment, 00:16:04.240 --> 00:16:07.600 align:middle line:84% you can put one condition in one and one condition 00:16:07.600 --> 00:16:09.250 align:middle line:90% in the other channel. 00:16:09.250 --> 00:16:14.320 align:middle line:84% But you can use some kind of RNA soup 00:16:14.320 --> 00:16:17.150 align:middle line:84% that is not supposed hybridized to anything 00:16:17.150 --> 00:16:19.040 align:middle line:90% and put it on one channel. 00:16:19.040 --> 00:16:22.960 align:middle line:84% So if you get a lot of random hybridization on one side, 00:16:22.960 --> 00:16:27.820 align:middle line:84% this will pick it up and makes your point, your spot, yellow. 00:16:27.820 --> 00:16:30.115 align:middle line:84% In this case, you will treat yellow and black exactly 00:16:30.115 --> 00:16:31.190 align:middle line:90% in the same way. 00:16:31.190 --> 00:16:32.830 align:middle line:84% It's not like, no, I'm undecided. 00:16:32.830 --> 00:16:33.670 align:middle line:90% Both are up. 00:16:33.670 --> 00:16:36.220 align:middle line:84% You will say, well, this is up, but it's up because 00:16:36.220 --> 00:16:41.260 align:middle line:84% of some possible random hybridization. 00:16:41.260 --> 00:16:44.860 align:middle line:84% The resolution is to use a computational methods. 00:16:44.860 --> 00:16:50.770 align:middle line:84% And these microarrays produced by Affymetrix 00:16:50.770 --> 00:16:54.910 align:middle line:90% are like Microsoft Office. 00:16:54.910 --> 00:16:57.250 align:middle line:84% It's something that everybody uses. 00:16:57.250 --> 00:17:00.640 align:middle line:84% They are far more expensive than they should be. 00:17:00.640 --> 00:17:03.010 align:middle line:90% And everybody hates Affymetrix. 00:17:03.010 --> 00:17:05.589 align:middle line:84% But still, you can't live without Microsoft. 00:17:05.589 --> 00:17:07.839 align:middle line:84% You can't live without Affymetrix. 00:17:07.839 --> 00:17:10.240 align:middle line:84% AUDIENCE: Was the last slide a cDNA array? 00:17:10.240 --> 00:17:12.819 align:middle line:90% MARCO RAMONI: Yes, it was. 00:17:12.819 --> 00:17:15.099 align:middle line:84% So cDNA means that you put the entire transcript. 00:17:15.099 --> 00:17:18.160 align:middle line:84% And usually you have this two-channel dye. 00:17:18.160 --> 00:17:22.930 align:middle line:84% [INAUDIBLE] is a great expert in cDNA microarrays, which 00:17:22.930 --> 00:17:26.290 align:middle line:84% actually comes from one of the very first departments that 00:17:26.290 --> 00:17:28.720 align:middle line:90% made cDNA microarrays. 00:17:28.720 --> 00:17:35.200 align:middle line:84% And is [? Formica, right? ?] So oligonucleotide microarrays 00:17:35.200 --> 00:17:37.210 align:middle line:90% follow this idea. 00:17:37.210 --> 00:17:38.980 align:middle line:90% OK, I have my entire transcript. 00:17:38.980 --> 00:17:40.900 align:middle line:84% And my problem is this random hybridization. 00:17:40.900 --> 00:17:41.660 align:middle line:90% What can I do? 00:17:41.660 --> 00:17:45.610 align:middle line:84% Well, because I have the human genome-- 00:17:45.610 --> 00:17:47.290 align:middle line:84% the entire draft of the human genome-- 00:17:47.290 --> 00:17:49.990 align:middle line:84% I can take this gene and find out 00:17:49.990 --> 00:17:53.140 align:middle line:84% if there is a sequence, a small sequence in this gene, 00:17:53.140 --> 00:17:57.250 align:middle line:84% that is unique to this particular gene, right? 00:17:57.250 --> 00:18:00.190 align:middle line:84% So in this case, even if the sequence is small, 00:18:00.190 --> 00:18:03.580 align:middle line:84% if the broken part of RNA in my sample is small, 00:18:03.580 --> 00:18:05.650 align:middle line:84% it will not hybridize there because this sequence 00:18:05.650 --> 00:18:07.600 align:middle line:90% is too small and too specific. 00:18:07.600 --> 00:18:10.600 align:middle line:84% So the idea here for oligonucleotide microarrays 00:18:10.600 --> 00:18:13.097 align:middle line:84% is to say, I'm going to take my transcript-- 00:18:13.097 --> 00:18:14.680 align:middle line:84% the entire transcript, the subsequence 00:18:14.680 --> 00:18:16.055 align:middle line:84% of the transcript-- and I'm going 00:18:16.055 --> 00:18:19.570 align:middle line:84% to sample it 20 times, between 16 and 20 times, 00:18:19.570 --> 00:18:23.260 align:middle line:84% and find out these 25 small sequences that 00:18:23.260 --> 00:18:25.150 align:middle line:90% are specific to that thing. 00:18:25.150 --> 00:18:35.500 align:middle line:84% And, for good measure, I will create another sequence, 00:18:35.500 --> 00:18:39.970 align:middle line:84% a sequence that is exactly to this very specific sequence, 00:18:39.970 --> 00:18:44.740 align:middle line:84% except that I have the base in the middle that is flipped. 00:18:44.740 --> 00:18:48.550 align:middle line:84% And then we check that the sequence 00:18:48.550 --> 00:18:54.670 align:middle line:84% with that base in the middle is not specific to any other gene. 00:18:54.670 --> 00:18:59.110 align:middle line:84% So I had a positive control and a negative control. 00:18:59.110 --> 00:19:00.610 align:middle line:84% And this way, what I'm going to do 00:19:00.610 --> 00:19:03.040 align:middle line:84% is to have a very specific measurement 00:19:03.040 --> 00:19:06.520 align:middle line:84% and a very specific measurements of random hybridization. 00:19:06.520 --> 00:19:09.220 align:middle line:84% Then, once I have these 20 measurements, 00:19:09.220 --> 00:19:12.140 align:middle line:84% I will find some statistics to put them together. 00:19:12.140 --> 00:19:16.000 align:middle line:84% It's not easy because these measurements are not 00:19:16.000 --> 00:19:17.150 align:middle line:90% independent measurements. 00:19:17.150 --> 00:19:19.000 align:middle line:90% But it doesn't matter. 00:19:19.000 --> 00:19:21.250 align:middle line:84% I will put these things together somehow. 00:19:21.250 --> 00:19:25.210 align:middle line:84% And the measure I get at the end is going 00:19:25.210 --> 00:19:26.920 align:middle line:90% to be pretty accurate, right? 00:19:26.920 --> 00:19:28.530 align:middle line:84% This is why they cost a lot of money. 00:19:28.530 --> 00:19:30.900 align:middle line:84% This is why a lot of computational work 00:19:30.900 --> 00:19:32.790 align:middle line:84% goes into it, because you have to search 00:19:32.790 --> 00:19:34.150 align:middle line:90% for all these sequences. 00:19:34.150 --> 00:19:37.230 align:middle line:84% And this is why sometimes these microarrays get completely 00:19:37.230 --> 00:19:37.830 align:middle line:90% screwed up. 00:19:37.830 --> 00:19:39.540 align:middle line:84% There was a famous case a couple of years 00:19:39.540 --> 00:19:45.840 align:middle line:84% back in which they created a new mouse microarray. 00:19:45.840 --> 00:19:48.810 align:middle line:84% And somebody-- I don't remember where-- 00:19:48.810 --> 00:19:54.960 align:middle line:84% reanalyzed the sequences of their microarray-- that 00:19:54.960 --> 00:19:57.090 align:middle line:84% new edition of the mouse microarray-- 00:19:57.090 --> 00:20:00.542 align:middle line:84% and found out about 25% of those were screwed up. 00:20:00.542 --> 00:20:01.500 align:middle line:90% They were not specific. 00:20:01.500 --> 00:20:03.333 align:middle line:84% They were not following the standard design. 00:20:03.333 --> 00:20:06.210 align:middle line:90% 00:20:06.210 --> 00:20:08.370 align:middle line:90% Hello? 00:20:08.370 --> 00:20:11.620 align:middle line:84% OK, so this is how they look like. 00:20:11.620 --> 00:20:13.110 align:middle line:90% This is the scan microarray. 00:20:13.110 --> 00:20:14.790 align:middle line:90% This is Affymetrix microarray. 00:20:14.790 --> 00:20:16.590 align:middle line:90% This is how it scanned. 00:20:16.590 --> 00:20:19.380 align:middle line:84% The little spots here on the longer transcripts 00:20:19.380 --> 00:20:25.200 align:middle line:84% but are these probes that are sampling for one 00:20:25.200 --> 00:20:27.750 align:middle line:90% particular, specific sequence. 00:20:27.750 --> 00:20:32.370 align:middle line:84% And they are scatter in the microarray 00:20:32.370 --> 00:20:35.520 align:middle line:84% so that if something bad happened to a corner, 00:20:35.520 --> 00:20:37.393 align:middle line:84% it would not affect something else. 00:20:37.393 --> 00:20:39.810 align:middle line:84% Otherwise, you could have biases in the entire microarray. 00:20:39.810 --> 00:20:41.227 align:middle line:84% There were one next to each other. 00:20:41.227 --> 00:20:44.100 align:middle line:84% Your whole microarray will be screwed up. 00:20:44.100 --> 00:20:47.880 align:middle line:84% And that's how, in theory, one probe should 00:20:47.880 --> 00:20:49.950 align:middle line:90% look like when it's hybridized. 00:20:49.950 --> 00:20:54.150 align:middle line:84% Up there, all these transcripts are more or less hybridized. 00:20:54.150 --> 00:20:57.570 align:middle line:84% And down there, the random hybridization 00:20:57.570 --> 00:20:59.460 align:middle line:90% is not really hybridized. 00:20:59.460 --> 00:21:02.550 align:middle line:84% But the resulting hybridization will 00:21:02.550 --> 00:21:08.670 align:middle line:84% be the difference between the real hybridization 00:21:08.670 --> 00:21:12.420 align:middle line:84% and the random hybridization for each probe, 00:21:12.420 --> 00:21:15.630 align:middle line:84% for each probe pair, the negative and the positive, 00:21:15.630 --> 00:21:19.320 align:middle line:84% and then a global measure to put them together, which I will not 00:21:19.320 --> 00:21:20.220 align:middle line:90% bother you with. 00:21:20.220 --> 00:21:25.200 align:middle line:90% 00:21:25.200 --> 00:21:26.250 align:middle line:90% So what's the problem? 00:21:26.250 --> 00:21:31.650 align:middle line:84% The problem is that this stuff costs $1,000 a pop. 00:21:31.650 --> 00:21:34.790 align:middle line:90% $1,000 a pop is a lot of money. 00:21:34.790 --> 00:21:38.120 align:middle line:84% And you remember what I was talking 00:21:38.120 --> 00:21:40.700 align:middle line:84% about-- the hypothesis-driven thing. 00:21:40.700 --> 00:21:43.220 align:middle line:84% When people draw balls for a single microarray, 00:21:43.220 --> 00:21:47.540 align:middle line:84% usually they do it twice, at most, three times. 00:21:47.540 --> 00:21:52.030 align:middle line:84% But if you're measuring 40,000 genes at the same time-- 00:21:52.030 --> 00:21:54.530 align:middle line:84% well, measuring twice is going to be a little problem, also, 00:21:54.530 --> 00:21:58.220 align:middle line:84% because you don't have any hypothesis to prove, right? 00:21:58.220 --> 00:22:02.750 align:middle line:84% So this is where the major cultural clash comes in. 00:22:02.750 --> 00:22:08.420 align:middle line:84% That when people analyze data, even in medical domains, 00:22:08.420 --> 00:22:09.960 align:middle line:90% database looks like this. 00:22:09.960 --> 00:22:14.570 align:middle line:84% There are some variables and a lot of cases. 00:22:14.570 --> 00:22:17.310 align:middle line:84% The microarrays data set usually look like this. 00:22:17.310 --> 00:22:20.030 align:middle line:84% You have a lot of variables, thousands of variables, 00:22:20.030 --> 00:22:21.560 align:middle line:90% and very little measurements. 00:22:21.560 --> 00:22:23.780 align:middle line:84% And it's kind of funny to see these people that 00:22:23.780 --> 00:22:27.410 align:middle line:84% work on the genetics and define from a genomic side. 00:22:27.410 --> 00:22:30.860 align:middle line:84% So when they design a genetic experiment for SNPs, 00:22:30.860 --> 00:22:35.930 align:middle line:84% they collect 5,000 patients, 2000 patients, 3,000 patients 00:22:35.930 --> 00:22:38.090 align:middle line:84% because that sample size is required 00:22:38.090 --> 00:22:39.650 align:middle line:90% to analyze a couple of SNPs. 00:22:39.650 --> 00:22:41.082 align:middle line:84% But then when they do microarrays, 00:22:41.082 --> 00:22:42.540 align:middle line:84% they expect to find the [INAUDIBLE] 00:22:42.540 --> 00:22:49.430 align:middle line:84% 5 microarrays out of 45,000 probes. 00:22:49.430 --> 00:22:51.600 align:middle line:90% It's exactly the same people. 00:22:51.600 --> 00:22:54.420 align:middle line:90% So what you can do with this? 00:22:54.420 --> 00:22:57.590 align:middle line:84% Well, let me introduce you to a notion that will 00:22:57.590 --> 00:23:01.100 align:middle line:90% remain very precious to you. 00:23:01.100 --> 00:23:04.100 align:middle line:84% When people will be confused-- as some people are-- you 00:23:04.100 --> 00:23:10.040 align:middle line:84% will have a very easy and fast answer. 00:23:10.040 --> 00:23:13.400 align:middle line:84% What is the difference between supervised and supervised? 00:23:13.400 --> 00:23:19.580 align:middle line:84% Is exactly the difference between a normal movie and a PG 00:23:19.580 --> 00:23:21.110 align:middle line:90% movie. 00:23:21.110 --> 00:23:24.020 align:middle line:84% Supervised means that there is something or somebody 00:23:24.020 --> 00:23:25.490 align:middle line:90% supervising. 00:23:25.490 --> 00:23:27.470 align:middle line:90% They're telling you things. 00:23:27.470 --> 00:23:31.050 align:middle line:84% This is what a supervisor does, tells you things. 00:23:31.050 --> 00:23:34.970 align:middle line:84% So a supervised thing means that they 00:23:34.970 --> 00:23:38.450 align:middle line:84% have either a human or some kind of signal that 00:23:38.450 --> 00:23:42.920 align:middle line:84% will tell me what a particular sample means, right? 00:23:42.920 --> 00:23:45.290 align:middle line:84% A typical supervised problem is-- 00:23:45.290 --> 00:23:52.760 align:middle line:84% let's try to decide what characterized the people 00:23:52.760 --> 00:23:54.170 align:middle line:90% in a particular room. 00:23:54.170 --> 00:23:57.080 align:middle line:84% I get measurements from this room and that room. 00:23:57.080 --> 00:23:59.570 align:middle line:84% And in this room, this is a graded course 00:23:59.570 --> 00:24:01.220 align:middle line:90% in functional genomics. 00:24:01.220 --> 00:24:05.900 align:middle line:84% In the other room is a class from the dental school. 00:24:05.900 --> 00:24:07.910 align:middle line:90% And let's make differences. 00:24:07.910 --> 00:24:09.530 align:middle line:84% We have properties of these people. 00:24:09.530 --> 00:24:10.790 align:middle line:90% Let's see what is different. 00:24:10.790 --> 00:24:13.640 align:middle line:84% But I will tell you that people in this room 00:24:13.640 --> 00:24:16.160 align:middle line:84% are different from the people in that room. 00:24:16.160 --> 00:24:19.940 align:middle line:84% And this is what is called your training signal, 00:24:19.940 --> 00:24:21.770 align:middle line:84% the difference between two clusters. 00:24:21.770 --> 00:24:27.470 align:middle line:84% In an unsupervised thing I know I have no supervision. 00:24:27.470 --> 00:24:30.770 align:middle line:84% I'm old enough to go to a R-rate movie. 00:24:30.770 --> 00:24:34.190 align:middle line:84% So the question in this case is-- 00:24:34.190 --> 00:24:35.960 align:middle line:90% I get a bunch of people-- 00:24:35.960 --> 00:24:37.820 align:middle line:90% are there groups among them? 00:24:37.820 --> 00:24:40.580 align:middle line:84% There are people that look more like others. 00:24:40.580 --> 00:24:44.030 align:middle line:84% There are people that control other people. 00:24:44.030 --> 00:24:47.990 align:middle line:84% It looks more like, I can say, gossip, 00:24:47.990 --> 00:24:50.480 align:middle line:90% finding stories in these things. 00:24:50.480 --> 00:24:53.040 align:middle line:84% But they answer two very different questions. 00:24:53.040 --> 00:24:57.470 align:middle line:84% One is, what is different between these two 00:24:57.470 --> 00:24:58.220 align:middle line:90% groups of people? 00:24:58.220 --> 00:25:01.550 align:middle line:84% And the other is what is similar, or what is related, 00:25:01.550 --> 00:25:04.280 align:middle line:84% or what are the stratification, or what 00:25:04.280 --> 00:25:06.830 align:middle line:84% are the things that we have in common 00:25:06.830 --> 00:25:08.195 align:middle line:90% among these different people? 00:25:08.195 --> 00:25:11.330 align:middle line:90% 00:25:11.330 --> 00:25:12.500 align:middle line:90% It becomes clear. 00:25:12.500 --> 00:25:14.720 align:middle line:84% So what can we do with the microarray? 00:25:14.720 --> 00:25:18.020 align:middle line:84% With the microwave-- well, the first thing we can say 00:25:18.020 --> 00:25:23.150 align:middle line:84% is, OK, I have two experimental conditions. 00:25:23.150 --> 00:25:27.050 align:middle line:84% And my aim is to see which genes are expressed more 00:25:27.050 --> 00:25:30.320 align:middle line:84% and which genes are expressed less in this condition, right? 00:25:30.320 --> 00:25:34.550 align:middle line:90% So typical example is cancer. 00:25:34.550 --> 00:25:36.260 align:middle line:84% I get a bunch of people with cancer. 00:25:36.260 --> 00:25:38.240 align:middle line:84% I get a bunch of people without cancer. 00:25:38.240 --> 00:25:40.890 align:middle line:84% I run microarrays and then see what is different. 00:25:40.890 --> 00:25:42.030 align:middle line:90% What does it mean? 00:25:42.030 --> 00:25:45.740 align:middle line:84% Well, it means that they have tissues from healthy cells 00:25:45.740 --> 00:25:47.030 align:middle line:90% and from tumor cells. 00:25:47.030 --> 00:25:50.930 align:middle line:84% And for each sample, I will create a microarray. 00:25:50.930 --> 00:25:54.080 align:middle line:84% And this is how my database, at the end, will look like. 00:25:54.080 --> 00:25:57.810 align:middle line:90% 00:25:57.810 --> 00:26:01.670 align:middle line:84% So the first column represents the name 00:26:01.670 --> 00:26:05.190 align:middle line:84% of the gene, name of the transcript directly looking at. 00:26:05.190 --> 00:26:07.340 align:middle line:84% And the second column represents the value 00:26:07.340 --> 00:26:10.400 align:middle line:84% for that transcript for sample one, sample two, sample three, 00:26:10.400 --> 00:26:12.717 align:middle line:90% sample four, and sample five. 00:26:12.717 --> 00:26:14.300 align:middle line:84% And then, I'm going to tell you, well, 00:26:14.300 --> 00:26:17.705 align:middle line:84% sample one, sample two, sample three belongs to one category, 00:26:17.705 --> 00:26:19.460 align:middle line:90% are in this room. 00:26:19.460 --> 00:26:22.790 align:middle line:84% And sample four and five are in that room. 00:26:22.790 --> 00:26:26.850 align:middle line:84% Go and find what's different between these two things. 00:26:26.850 --> 00:26:32.540 align:middle line:84% Now, what does it really mean-- what is different? 00:26:32.540 --> 00:26:34.550 align:middle line:84% Remember, so if we do it by hand, 00:26:34.550 --> 00:26:37.060 align:middle line:84% we can take pictures of balls and say this ball is 00:26:37.060 --> 00:26:38.470 align:middle line:90% bigger than this one. 00:26:38.470 --> 00:26:41.920 align:middle line:84% But if I have 50,000, 40,000 balls, what 00:26:41.920 --> 00:26:44.720 align:middle line:90% I'm going to do about this? 00:26:44.720 --> 00:26:49.160 align:middle line:84% Well, what they want to do, in this case, 00:26:49.160 --> 00:26:52.640 align:middle line:84% is to find what is more expressed in one side rather 00:26:52.640 --> 00:26:53.340 align:middle line:90% than another. 00:26:53.340 --> 00:26:55.610 align:middle line:84% And the currency provision-- that I will tell you 00:26:55.610 --> 00:26:56.240 align:middle line:90% in a second-- 00:26:56.240 --> 00:26:59.060 align:middle line:90% 00:26:59.060 --> 00:27:02.720 align:middle line:84% the currency of these measurements are called folds. 00:27:02.720 --> 00:27:06.470 align:middle line:84% Fold is how many times one condition is more expressed 00:27:06.470 --> 00:27:08.090 align:middle line:90% than the other condition. 00:27:08.090 --> 00:27:10.010 align:middle line:84% Now the problem is that this is good 00:27:10.010 --> 00:27:11.840 align:middle line:90% when you have one single ball. 00:27:11.840 --> 00:27:16.680 align:middle line:84% If I have 50 patients, what the hell I do? 00:27:16.680 --> 00:27:18.440 align:middle line:90% Should I take the mean? 00:27:18.440 --> 00:27:19.010 align:middle line:90% Sure. 00:27:19.010 --> 00:27:21.830 align:middle line:84% I can take the mean, but then I don't have any measure 00:27:21.830 --> 00:27:23.610 align:middle line:90% of the variance in my data. 00:27:23.610 --> 00:27:30.110 align:middle line:84% Maybe I have two things that the mean are very far apart. 00:27:30.110 --> 00:27:33.020 align:middle line:84% But because the variance is very big, they will overlap. 00:27:33.020 --> 00:27:35.210 align:middle line:84% So there is not much evidence that they can collect. 00:27:35.210 --> 00:27:38.690 align:middle line:84% So other measures are things like standardized differences, 00:27:38.690 --> 00:27:41.060 align:middle line:84% and make the difference, and standardize them 00:27:41.060 --> 00:27:43.730 align:middle line:84% by the variance, which will somehow take into account 00:27:43.730 --> 00:27:49.640 align:middle line:84% the variance, that is under the assumption 00:27:49.640 --> 00:27:51.540 align:middle line:84% that these things are normally distributed, 00:27:51.540 --> 00:27:54.950 align:middle line:84% so that kind of variance has any statistical meaning. 00:27:54.950 --> 00:27:56.070 align:middle line:90% Then what do you do? 00:27:56.070 --> 00:27:59.300 align:middle line:84% Well, then I decide the threshold. 00:27:59.300 --> 00:28:01.140 align:middle line:84% I get the landscape of this thing. 00:28:01.140 --> 00:28:05.040 align:middle line:84% And I'm going to say the top 50 genes are what I actually like 00:28:05.040 --> 00:28:06.590 align:middle line:90% and the bottom 20 genes. 00:28:06.590 --> 00:28:09.170 align:middle line:84% The top 50 are the ones that are more change in one condition 00:28:09.170 --> 00:28:10.460 align:middle line:84% and the bottom are the ones that are more 00:28:10.460 --> 00:28:11.780 align:middle line:90% change in the other condition. 00:28:11.780 --> 00:28:15.230 align:middle line:84% I'm going to pick up this stuff and see if there 00:28:15.230 --> 00:28:16.460 align:middle line:90% is anything interesting. 00:28:16.460 --> 00:28:19.370 align:middle line:84% What people do typically is to make up stories 00:28:19.370 --> 00:28:22.160 align:middle line:90% or to pull up a protein-- 00:28:22.160 --> 00:28:25.430 align:middle line:84% like people are doing with their project-- pull out the protein 00:28:25.430 --> 00:28:28.850 align:middle line:84% and see if I can actually-- a gene, find out the protein, 00:28:28.850 --> 00:28:32.030 align:middle line:84% and actually find out if this protein does something 00:28:32.030 --> 00:28:34.790 align:middle line:90% to my particular phenotype. 00:28:34.790 --> 00:28:37.220 align:middle line:84% This project, which is a project about preeclampsia 00:28:37.220 --> 00:28:40.340 align:middle line:84% the investigator there-- you get only two microarrays 00:28:40.340 --> 00:28:42.470 align:middle line:84% from a preeclamptic-- preeclampsia 00:28:42.470 --> 00:28:45.980 align:middle line:84% is a disease that women get during pregnancy. 00:28:45.980 --> 00:28:49.400 align:middle line:90% It's a very bad disease-- 00:28:49.400 --> 00:28:55.670 align:middle line:84% and the normal placenta, compare them, and pull out the protein. 00:28:55.670 --> 00:28:59.117 align:middle line:84% Put the protein into mice and found out the mice 00:28:59.117 --> 00:29:00.200 align:middle line:90% were getting preeclampsia. 00:29:00.200 --> 00:29:02.270 align:middle line:84% There are this kind of an exploratory thing. 00:29:02.270 --> 00:29:03.950 align:middle line:84% In this case, what I'm interested in 00:29:03.950 --> 00:29:06.230 align:middle line:90% is find out new hypothesis. 00:29:06.230 --> 00:29:08.630 align:middle line:84% Then they can test in some kind of laboratory setting. 00:29:08.630 --> 00:29:13.730 align:middle line:90% 00:29:13.730 --> 00:29:18.300 align:middle line:84% As I said-- because, in this case, I have only two samples. 00:29:18.300 --> 00:29:21.980 align:middle line:84% But suppose I have several patients, what I can do? 00:29:21.980 --> 00:29:26.450 align:middle line:90% 00:29:26.450 --> 00:29:27.950 align:middle line:84% One other problem we have here is 00:29:27.950 --> 00:29:30.530 align:middle line:84% that we are not really sure what kind of distributions 00:29:30.530 --> 00:29:32.390 align:middle line:90% are running this microarray. 00:29:32.390 --> 00:29:34.880 align:middle line:84% So what people say is, well, because we didn't know 00:29:34.880 --> 00:29:38.600 align:middle line:84% the distribution, let's use some distribution-free method, which 00:29:38.600 --> 00:29:39.680 align:middle line:90% is a good idea. 00:29:39.680 --> 00:29:42.200 align:middle line:84% But it's an idea that rests on the hope 00:29:42.200 --> 00:29:44.600 align:middle line:84% that there is some free lunch of life. 00:29:44.600 --> 00:29:46.820 align:middle line:84% And there is no free lunch in life. 00:29:46.820 --> 00:29:49.220 align:middle line:84% No parametric method, distribution-free method 00:29:49.220 --> 00:29:52.290 align:middle line:84% requires a lot of data because you have to do two things. 00:29:52.290 --> 00:29:55.520 align:middle line:84% First, you have to decide what kind of distribution you have, 00:29:55.520 --> 00:29:56.390 align:middle line:90% implicitly. 00:29:56.390 --> 00:29:59.270 align:middle line:84% And then, you have to run your test. 00:29:59.270 --> 00:30:01.610 align:middle line:84% People use parametric method typically 00:30:01.610 --> 00:30:04.020 align:middle line:84% because they have an idea of the distribution. 00:30:04.020 --> 00:30:06.290 align:middle line:84% And so they need less data to fit this test. 00:30:06.290 --> 00:30:08.600 align:middle line:84% If you have few data and no idea of the distribution, 00:30:08.600 --> 00:30:10.680 align:middle line:90% you are screwed. 00:30:10.680 --> 00:30:14.390 align:middle line:84% And running this kind of test tends 00:30:14.390 --> 00:30:16.460 align:middle line:90% to be kind of a dangerous. 00:30:16.460 --> 00:30:19.010 align:middle line:84% One, because usually your sample size 00:30:19.010 --> 00:30:22.190 align:middle line:84% is too small to run a proper parametric test. 00:30:22.190 --> 00:30:24.973 align:middle line:84% Two, because frequentist people have these things 00:30:24.973 --> 00:30:25.640 align:middle line:90% called p-values. 00:30:25.640 --> 00:30:28.610 align:middle line:90% 00:30:28.610 --> 00:30:31.940 align:middle line:84% P-values are very interesting animals. 00:30:31.940 --> 00:30:33.908 align:middle line:90% What is the p-value? 00:30:33.908 --> 00:30:35.450 align:middle line:84% Who gives me a definition of p-value? 00:30:35.450 --> 00:30:38.720 align:middle line:90% 00:30:38.720 --> 00:30:42.050 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:30:42.050 --> 00:30:43.160 align:middle line:90% MARCO RAMONI: Speak up. 00:30:43.160 --> 00:30:45.090 align:middle line:90% AUDIENCE: The probability that-- 00:30:45.090 --> 00:30:46.715 align:middle line:84% MARCO RAMONI: Two things are different? 00:30:46.715 --> 00:30:49.870 align:middle line:84% AUDIENCE: --that the means [INAUDIBLE] very different. 00:30:49.870 --> 00:30:55.770 align:middle line:84% MARCO RAMONI: OK, so this is what patients believe. 00:30:55.770 --> 00:30:57.780 align:middle line:90% But to do that-- 00:30:57.780 --> 00:31:00.870 align:middle line:90% you work with patients too much. 00:31:00.870 --> 00:31:02.700 align:middle line:84% And that's a very reasonable measure. 00:31:02.700 --> 00:31:04.380 align:middle line:84% I'm interested in finding what is 00:31:04.380 --> 00:31:08.340 align:middle line:84% the probability that these two things are different, right? 00:31:08.340 --> 00:31:10.360 align:middle line:90% This is not the p-value. 00:31:10.360 --> 00:31:14.190 align:middle line:84% The p-value is the probability that you will make a mistake 00:31:14.190 --> 00:31:17.220 align:middle line:84% if you repeat the experiment N times 00:31:17.220 --> 00:31:20.850 align:middle line:84% and compute it as the number of times you will be mistaken 00:31:20.850 --> 00:31:23.010 align:middle line:84% by repeating this study, which is 00:31:23.010 --> 00:31:25.110 align:middle line:84% an extremely [? masturbational ?] measure. 00:31:25.110 --> 00:31:27.090 align:middle line:84% There is no relationship with the probability 00:31:27.090 --> 00:31:28.200 align:middle line:90% of your hypothesis. 00:31:28.200 --> 00:31:30.075 align:middle line:84% And it's very difficult to put into practice. 00:31:30.075 --> 00:31:31.658 align:middle line:84% First of all, people should explain me 00:31:31.658 --> 00:31:33.780 align:middle line:84% why I should repeat my experiment 100 times when 00:31:33.780 --> 00:31:35.700 align:middle line:90% I already repeat 20. 00:31:35.700 --> 00:31:37.590 align:middle line:90% And this is what I know, right? 00:31:37.590 --> 00:31:41.580 align:middle line:84% The rest is educated or uneducated guess. 00:31:41.580 --> 00:31:45.360 align:middle line:84% But the p-value, in this case, has this other little problem. 00:31:45.360 --> 00:31:48.750 align:middle line:84% That because I repeat the experiment a lot of times, 00:31:48.750 --> 00:31:51.910 align:middle line:84% sometimes things may come up just at random. 00:31:51.910 --> 00:31:56.430 align:middle line:84% So if I say, OK, I'm going to accept something 00:31:56.430 --> 00:32:01.430 align:middle line:90% if my probability of is 5%-- 00:32:01.430 --> 00:32:03.840 align:middle line:90% so the p-value is 0.01-- 00:32:03.840 --> 00:32:06.810 align:middle line:84% if I test two hypothesis, to maintain 00:32:06.810 --> 00:32:09.628 align:middle line:84% the same level of error, because I 00:32:09.628 --> 00:32:11.670 align:middle line:84% have the probability that something will come out 00:32:11.670 --> 00:32:13.890 align:middle line:90% at random-- 00:32:13.890 --> 00:32:15.870 align:middle line:84% assuming that these two tests are independent-- 00:32:15.870 --> 00:32:18.870 align:middle line:84% I have to multiply the probability of this p-value, 00:32:18.870 --> 00:32:19.410 align:middle line:90% right? 00:32:19.410 --> 00:32:24.540 align:middle line:84% So my real threshold to get a 5% evidence of a p-value would be 00:32:24.540 --> 00:32:29.640 align:middle line:84% the product of these two 5%s to maintain the same level 00:32:29.640 --> 00:32:31.350 align:middle line:90% of strength of evidence. 00:32:31.350 --> 00:32:35.470 align:middle line:84% Now imagine if you have to multiply 0.05 40,000 times. 00:32:35.470 --> 00:32:37.800 align:middle line:90% What kind of threshold you get? 00:32:37.800 --> 00:32:38.610 align:middle line:90% Nothing. 00:32:38.610 --> 00:32:40.750 align:middle line:84% Nothing will pass that particular test. 00:32:40.750 --> 00:32:43.530 align:middle line:84% This is called Bonferroni correction. 00:32:43.530 --> 00:32:45.000 align:middle line:90% Nothing will pass the test. 00:32:45.000 --> 00:32:46.440 align:middle line:84% I have a very dear friend of mine 00:32:46.440 --> 00:32:48.750 align:middle line:84% who is very frustrated by this and decided 00:32:48.750 --> 00:32:51.780 align:middle line:84% to be a biologist after trying to use p-test 00:32:51.780 --> 00:32:53.160 align:middle line:90% on this kind of experiments. 00:32:53.160 --> 00:32:56.310 align:middle line:84% Because the threshold, the accepted evidence 00:32:56.310 --> 00:33:01.950 align:middle line:84% is 0.05%, will turn against you when you are testing 00:33:01.950 --> 00:33:03.870 align:middle line:90% your hypothesis 40,000 times. 00:33:03.870 --> 00:33:09.630 align:middle line:84% Besides, this is under the most lenient condition. 00:33:09.630 --> 00:33:11.820 align:middle line:84% Because you'll assume that all your hypotheses 00:33:11.820 --> 00:33:13.260 align:middle line:90% are independent. 00:33:13.260 --> 00:33:15.280 align:middle line:84% But we know that this is not true. 00:33:15.280 --> 00:33:17.620 align:middle line:84% We know that these genes regulate each other. 00:33:17.620 --> 00:33:19.830 align:middle line:84% So the probability something is up 00:33:19.830 --> 00:33:21.900 align:middle line:84% is not independent of the probability something 00:33:21.900 --> 00:33:23.580 align:middle line:90% else is up, OK? 00:33:23.580 --> 00:33:30.180 align:middle line:84% So even under the most simple condition, 00:33:30.180 --> 00:33:34.180 align:middle line:84% we have a little problem with this. 00:33:34.180 --> 00:33:35.918 align:middle line:90% So what can we do? 00:33:35.918 --> 00:33:37.960 align:middle line:84% Well, I will tell you in a second what can we do. 00:33:37.960 --> 00:33:39.565 align:middle line:90% But what we can do further-- 00:33:39.565 --> 00:33:41.190 align:middle line:84% not for the experiment but in general-- 00:33:41.190 --> 00:33:43.650 align:middle line:90% so once I have the differences-- 00:33:43.650 --> 00:33:48.690 align:middle line:84% OK, I can go back to my lab and put the protein into a couple 00:33:48.690 --> 00:33:50.490 align:middle line:90% of mice and see what happened. 00:33:50.490 --> 00:33:52.920 align:middle line:84% But isn't there anything better that I 00:33:52.920 --> 00:33:55.690 align:middle line:90% can do using these differences? 00:33:55.690 --> 00:34:00.165 align:middle line:84% Well, maybe I can make predictive models. 00:34:00.165 --> 00:34:02.790 align:middle line:90% 00:34:02.790 --> 00:34:06.990 align:middle line:84% Predictive models, rather than using proteins-- 00:34:06.990 --> 00:34:09.960 align:middle line:84% one protein at a time, what are called markers-- 00:34:09.960 --> 00:34:12.960 align:middle line:84% are able to put together a batch of proteins 00:34:12.960 --> 00:34:15.389 align:middle line:84% and provide a profile, a prediction 00:34:15.389 --> 00:34:17.790 align:middle line:90% for a particular outcome. 00:34:17.790 --> 00:34:22.830 align:middle line:84% In this case, maybe I can predict if a particular tissue 00:34:22.830 --> 00:34:26.130 align:middle line:90% is a tumor or it's not. 00:34:26.130 --> 00:34:29.280 align:middle line:84% I can predict if a particular tissue is a type of tumor 00:34:29.280 --> 00:34:30.870 align:middle line:90% or is not a type of tumor. 00:34:30.870 --> 00:34:33.540 align:middle line:84% Maybe this type of tumor require different therapies. 00:34:33.540 --> 00:34:36.030 align:middle line:84% Maybe I can predict how long it will 00:34:36.030 --> 00:34:39.030 align:middle line:84% take for a particular tissue to come back 00:34:39.030 --> 00:34:42.929 align:middle line:84% as a cancer, because they find a particular signature. 00:34:42.929 --> 00:34:45.389 align:middle line:90% Now how do I find the signature? 00:34:45.389 --> 00:34:48.730 align:middle line:84% I have to run a game called feature selection. 00:34:48.730 --> 00:34:51.480 align:middle line:84% Feature selection is-- I have a class. 00:34:51.480 --> 00:34:55.620 align:middle line:84% I have all these predictors down together. 00:34:55.620 --> 00:34:58.575 align:middle line:84% And I'm going to select some of them as good predictors 00:34:58.575 --> 00:35:02.150 align:middle line:84% for [? C. ?] I cannot use all of that, right? 00:35:02.150 --> 00:35:03.510 align:middle line:90% Why cannot use all of them? 00:35:03.510 --> 00:35:06.880 align:middle line:90% 00:35:06.880 --> 00:35:13.620 align:middle line:84% Because good prediction comes from specificity, right? 00:35:13.620 --> 00:35:16.260 align:middle line:84% I'm glad you agree with this, because it's not really 00:35:16.260 --> 00:35:18.720 align:middle line:84% such a normal statement to accept. 00:35:18.720 --> 00:35:21.067 align:middle line:84% People believe that you use 40,000 variables, 00:35:21.067 --> 00:35:22.650 align:middle line:84% you're going to give better prediction 00:35:22.650 --> 00:35:25.152 align:middle line:84% if you use five variables independently of the quality 00:35:25.152 --> 00:35:25.860 align:middle line:90% of this variable. 00:35:25.860 --> 00:35:28.140 align:middle line:84% I mean, as long as these five variables are a subset 00:35:28.140 --> 00:35:29.940 align:middle line:90% of the 40,000 variables. 00:35:29.940 --> 00:35:33.090 align:middle line:84% But we all agree that this doesn't happen, right? 00:35:33.090 --> 00:35:34.320 align:middle line:90% Right? 00:35:34.320 --> 00:35:36.270 align:middle line:84% If somebody has a doubt, I have a joke. 00:35:36.270 --> 00:35:37.980 align:middle line:90% OK, no joke. 00:35:37.980 --> 00:35:40.600 align:middle line:90% So what do I do? 00:35:40.600 --> 00:35:45.060 align:middle line:84% Well, I want to identify those genes that predict my class, 00:35:45.060 --> 00:35:47.680 align:middle line:84% the set of genes that predict the class. 00:35:47.680 --> 00:35:52.300 align:middle line:84% So if I do feature selection, I typically 00:35:52.300 --> 00:35:53.770 align:middle line:84% increase the predictive accuracy. 00:35:53.770 --> 00:35:55.810 align:middle line:84% I get a more competitive presentation. 00:35:55.810 --> 00:35:59.290 align:middle line:84% I can get some insight in the process that may happen. 00:35:59.290 --> 00:36:03.370 align:middle line:84% Although, remember, this is not just differentiable analysis. 00:36:03.370 --> 00:36:05.260 align:middle line:84% It's something that I want to use 00:36:05.260 --> 00:36:10.810 align:middle line:84% as a prognostic or a diagnostic set of markers when combined. 00:36:10.810 --> 00:36:16.040 align:middle line:84% And why differences are important? 00:36:16.040 --> 00:36:17.800 align:middle line:84% Well, because we start from the assumption 00:36:17.800 --> 00:36:21.490 align:middle line:84% is that if two things are exactly the same across two 00:36:21.490 --> 00:36:23.650 align:middle line:84% samples, it's very difficult that they 00:36:23.650 --> 00:36:26.060 align:middle line:84% will be able to discriminate between them. 00:36:26.060 --> 00:36:31.810 align:middle line:84% So classification, which is this task in feature selection, 00:36:31.810 --> 00:36:36.220 align:middle line:84% looks sometimes very much like differentiable analysis, 00:36:36.220 --> 00:36:38.260 align:middle line:90% but it's not. 00:36:38.260 --> 00:36:39.820 align:middle line:90% I have a twist at the end. 00:36:39.820 --> 00:36:41.980 align:middle line:84% And the aim of my game is not really 00:36:41.980 --> 00:36:43.540 align:middle line:90% to find out what is different. 00:36:43.540 --> 00:36:45.520 align:middle line:90% It's finding what is predictive. 00:36:45.520 --> 00:36:49.660 align:middle line:84% And the example is-- supposed I give you two groups of people. 00:36:49.660 --> 00:36:52.030 align:middle line:84% And you don't know it, but one group are men 00:36:52.030 --> 00:36:54.800 align:middle line:90% and the other group are women. 00:36:54.800 --> 00:36:59.200 align:middle line:84% And then I give you a list of properties of these people. 00:36:59.200 --> 00:37:03.100 align:middle line:84% And there would be a lot of differences. 00:37:03.100 --> 00:37:07.090 align:middle line:84% Women tend to be slightly shorter than men. 00:37:07.090 --> 00:37:09.910 align:middle line:84% Women tend to have more hair at certain age. 00:37:09.910 --> 00:37:12.310 align:middle line:90% Women tend to make less money. 00:37:12.310 --> 00:37:14.620 align:middle line:84% But there are a couple of anatomical differences 00:37:14.620 --> 00:37:16.970 align:middle line:84% that are really good predictors of these differences. 00:37:16.970 --> 00:37:19.900 align:middle line:84% Doesn't mean that there is no other difference. 00:37:19.900 --> 00:37:23.050 align:middle line:84% But it means that particular anatomical feature 00:37:23.050 --> 00:37:26.478 align:middle line:84% is a perfect predictor between male and female. 00:37:26.478 --> 00:37:28.270 align:middle line:84% So if you're doing differentiable analysis, 00:37:28.270 --> 00:37:30.400 align:middle line:84% you may be also interested in the fact 00:37:30.400 --> 00:37:34.010 align:middle line:84% that these people have differences in income. 00:37:34.010 --> 00:37:39.580 align:middle line:84% But if you include these factors into your predictive model, 00:37:39.580 --> 00:37:42.370 align:middle line:84% maybe because I'm short and don't make a lot of money, 00:37:42.370 --> 00:37:45.400 align:middle line:84% you end up classifying me as a woman, OK? 00:37:45.400 --> 00:37:46.510 align:middle line:90% May confuse your ideas. 00:37:46.510 --> 00:37:49.670 align:middle line:90% 00:37:49.670 --> 00:37:52.370 align:middle line:84% So we were saying, non-parametric method 00:37:52.370 --> 00:37:54.800 align:middle line:84% has kind of little problems with this 00:37:54.800 --> 00:37:58.460 align:middle line:84% because we don't have enough samples. 00:37:58.460 --> 00:38:03.140 align:middle line:84% But we have classifiers that are parametric classifiers. 00:38:03.140 --> 00:38:05.000 align:middle line:84% In this case, we make an assumption 00:38:05.000 --> 00:38:10.340 align:middle line:84% about the distribution of our data. 00:38:10.340 --> 00:38:14.120 align:middle line:84% And then we try to fit our data into this distribution, 00:38:14.120 --> 00:38:17.210 align:middle line:84% thus saving us a lot of effort in collecting more data. 00:38:17.210 --> 00:38:20.030 align:middle line:84% Because data are very complicated and hypotheses 00:38:20.030 --> 00:38:20.690 align:middle line:90% are cheap. 00:38:20.690 --> 00:38:24.080 align:middle line:84% We can actually go and validate our hypotheses afterward. 00:38:24.080 --> 00:38:27.500 align:middle line:84% So this is, as many of you know, is called a Naive Bayes 00:38:27.500 --> 00:38:33.320 align:middle line:84% Classifier, in which I assume that each gene down there is 00:38:33.320 --> 00:38:36.980 align:middle line:84% conditionally independent given the class. 00:38:36.980 --> 00:38:40.460 align:middle line:84% It doesn't mean that it is independent, right? 00:38:40.460 --> 00:38:43.100 align:middle line:84% Like we were doing before-- we were doing independent tests, 00:38:43.100 --> 00:38:43.760 align:middle line:90% you remember? 00:38:43.760 --> 00:38:45.350 align:middle line:84% The independent tests assume that they 00:38:45.350 --> 00:38:47.150 align:middle line:90% are marginally independent. 00:38:47.150 --> 00:38:50.750 align:middle line:84% In this case, they are conditionally independent. 00:38:50.750 --> 00:38:54.080 align:middle line:84% Conditional independent is that once I know the class, 00:38:54.080 --> 00:38:57.470 align:middle line:84% I don't give a dime about the dependency between these two 00:38:57.470 --> 00:38:58.290 align:middle line:90% genes. 00:38:58.290 --> 00:39:00.770 align:middle line:84% Maybe there is a very complicated relationship 00:39:00.770 --> 00:39:02.810 align:middle line:90% between gene one and gene two. 00:39:02.810 --> 00:39:05.540 align:middle line:84% But because my interest, in this case, 00:39:05.540 --> 00:39:08.450 align:middle line:84% is to find a classification, I don't care. 00:39:08.450 --> 00:39:11.570 align:middle line:84% Because as far as the classification of the class 00:39:11.570 --> 00:39:15.830 align:middle line:84% is concerned, these things are not related, right? 00:39:15.830 --> 00:39:19.610 align:middle line:84% So it's like a weak independence assumption where 00:39:19.610 --> 00:39:21.050 align:middle line:90% weak, in this case, means good. 00:39:21.050 --> 00:39:23.660 align:middle line:84% Because we are not forcing an assumption 00:39:23.660 --> 00:39:26.420 align:middle line:84% that is too strong within your data analysis. 00:39:26.420 --> 00:39:30.410 align:middle line:90% Once I have that, I run-- 00:39:30.410 --> 00:39:32.150 align:middle line:84% let me go back a second to this-- 00:39:32.150 --> 00:39:35.700 align:middle line:90% 00:39:35.700 --> 00:39:38.010 align:middle line:90% the other one. 00:39:38.010 --> 00:39:40.120 align:middle line:90% I want the other picture. 00:39:40.120 --> 00:39:40.620 align:middle line:90% Come on. 00:39:40.620 --> 00:39:47.330 align:middle line:90% 00:39:47.330 --> 00:39:47.885 align:middle line:90% See this? 00:39:47.885 --> 00:39:51.890 align:middle line:84% In this case, my genes are marginally independent. 00:39:51.890 --> 00:39:54.980 align:middle line:84% The arrow is going the other direction, right? 00:39:54.980 --> 00:39:58.430 align:middle line:84% So all these genes [? cause ?] my class, 00:39:58.430 --> 00:40:01.620 align:middle line:90% but they are independent. 00:40:01.620 --> 00:40:04.520 align:middle line:84% And this is the structure of a standard classifier. 00:40:04.520 --> 00:40:08.110 align:middle line:90% 00:40:08.110 --> 00:40:10.713 align:middle line:84% In the other case, they are conditionally independent 00:40:10.713 --> 00:40:11.380 align:middle line:90% given the class. 00:40:11.380 --> 00:40:18.040 align:middle line:90% 00:40:18.040 --> 00:40:21.780 align:middle line:84% So once I have this particular model, 00:40:21.780 --> 00:40:24.210 align:middle line:84% I have selected which are the genes that I like. 00:40:24.210 --> 00:40:26.250 align:middle line:84% I have estimated the parametric model. 00:40:26.250 --> 00:40:28.780 align:middle line:90% Then, I can make predictions. 00:40:28.780 --> 00:40:33.480 align:middle line:84% So if I had used some kind of differentiable analysis using 00:40:33.480 --> 00:40:36.330 align:middle line:84% a non-parametric test, by definition, 00:40:36.330 --> 00:40:39.360 align:middle line:90% I don't have parameters. 00:40:39.360 --> 00:40:40.620 align:middle line:90% It's non-parametric. 00:40:40.620 --> 00:40:42.930 align:middle line:84% So I cannot really make a prediction with the parameter 00:40:42.930 --> 00:40:47.250 align:middle line:84% set [? lower. ?] What people use are things called mixture 00:40:47.250 --> 00:40:53.790 align:middle line:84% of expert, in which they assign some kind of arbitrary weight 00:40:53.790 --> 00:40:56.400 align:middle line:90% to different genes. 00:40:56.400 --> 00:40:58.770 align:middle line:84% And each gene will be like an expert, 00:40:58.770 --> 00:41:02.760 align:middle line:84% judging if this particular tissue is a cancer 00:41:02.760 --> 00:41:04.560 align:middle line:90% or is not a cancer. 00:41:04.560 --> 00:41:07.590 align:middle line:84% But these weights are actually embedded 00:41:07.590 --> 00:41:09.540 align:middle line:84% in any parametric model you derive, 00:41:09.540 --> 00:41:11.910 align:middle line:84% which is the probability of observing 00:41:11.910 --> 00:41:15.090 align:middle line:84% that particular gene expressed, given the fact 00:41:15.090 --> 00:41:19.680 align:middle line:84% that you have a change in your class, that the class is tumor 00:41:19.680 --> 00:41:20.680 align:middle line:90% or not tumor. 00:41:20.680 --> 00:41:23.940 align:middle line:84% So you can apply Bayes' theorem, and reverse those errors, 00:41:23.940 --> 00:41:26.010 align:middle line:84% and obtain the posterior probability 00:41:26.010 --> 00:41:32.590 align:middle line:84% that your particular sample is a tumor or is not a tumor. 00:41:32.590 --> 00:41:33.465 align:middle line:90% This is how it works. 00:41:33.465 --> 00:41:37.528 align:middle line:90% 00:41:37.528 --> 00:41:38.820 align:middle line:90% Well, this is what I just said. 00:41:38.820 --> 00:41:40.180 align:middle line:90% [INAUDIBLE] had this one. 00:41:40.180 --> 00:41:41.490 align:middle line:90% So I have a class. 00:41:41.490 --> 00:41:44.070 align:middle line:84% And I'm interested in the probability of the class, given 00:41:44.070 --> 00:41:45.570 align:middle line:84% the sample molecular profile, which 00:41:45.570 --> 00:41:47.350 align:middle line:90% is my new patient coming in. 00:41:47.350 --> 00:41:52.470 align:middle line:84% And by applying Bayes' rule, I can actually compute. 00:41:52.470 --> 00:41:55.020 align:middle line:84% Because the probabilities I have are the probability 00:41:55.020 --> 00:41:57.960 align:middle line:84% of each feature given the class, which 00:41:57.960 --> 00:42:01.050 align:middle line:90% is the direction of the arrow. 00:42:01.050 --> 00:42:06.990 align:middle line:84% Bayes' rule will allow me to flip this rule backward, apply 00:42:06.990 --> 00:42:11.910 align:middle line:84% this as a product, and put all these things together 00:42:11.910 --> 00:42:14.860 align:middle line:84% into a single posterior probability. 00:42:14.860 --> 00:42:18.480 align:middle line:84% It's just the sum of this probability. 00:42:18.480 --> 00:42:21.930 align:middle line:84% I have another interesting thing with this-- 00:42:21.930 --> 00:42:24.480 align:middle line:84% another goodie with this thing that I can actually 00:42:24.480 --> 00:42:25.625 align:middle line:90% validate my stuff. 00:42:25.625 --> 00:42:27.750 align:middle line:84% Well, validate this would mean to go back in my lab 00:42:27.750 --> 00:42:29.042 align:middle line:90% and look at a couple of things. 00:42:29.042 --> 00:42:33.480 align:middle line:84% Validate means to see how my model is good to fit 00:42:33.480 --> 00:42:36.060 align:middle line:90% the entire 40,000 genes. 00:42:36.060 --> 00:42:39.450 align:middle line:84% And the best way to validate something 00:42:39.450 --> 00:42:43.530 align:middle line:84% is to have an independent test set. 00:42:43.530 --> 00:42:45.390 align:middle line:84% I collect patients here at Harvard. 00:42:45.390 --> 00:42:46.710 align:middle line:90% I build my model. 00:42:46.710 --> 00:42:51.760 align:middle line:84% And then, I call up my friends in San Antonio and say, listen, 00:42:51.760 --> 00:42:52.845 align:middle line:90% I have this model. 00:42:52.845 --> 00:42:55.000 align:middle line:90% Do you have 50 patients. 00:42:55.000 --> 00:43:01.030 align:middle line:84% for me that I can classify, and you know diagnosis already? 00:43:01.030 --> 00:43:03.450 align:middle line:90% And if he has them, then cool. 00:43:03.450 --> 00:43:05.910 align:middle line:84% I can really say, this is the accuracy 00:43:05.910 --> 00:43:09.120 align:middle line:90% of my model from here to there. 00:43:09.120 --> 00:43:13.515 align:middle line:84% But sometimes, we don't have these things. 00:43:13.515 --> 00:43:15.390 align:middle line:84% Well, quite often we don't have these things. 00:43:15.390 --> 00:43:18.360 align:middle line:90% So how can we do this cheaply? 00:43:18.360 --> 00:43:20.710 align:middle line:84% Cheaply-- we can use cross-validation. 00:43:20.710 --> 00:43:24.690 align:middle line:84% Cross-validation means that I take my data set. 00:43:24.690 --> 00:43:27.540 align:middle line:90% I split it in five parts. 00:43:27.540 --> 00:43:31.110 align:middle line:84% And I use four parts to learn my model. 00:43:31.110 --> 00:43:33.360 align:middle line:84% And then, I predict the fifth part. 00:43:33.360 --> 00:43:36.600 align:middle line:84% And then, I take other four of these five parts. 00:43:36.600 --> 00:43:38.250 align:middle line:90% I build another model. 00:43:38.250 --> 00:43:40.740 align:middle line:84% And I predict the remaining fifth part. 00:43:40.740 --> 00:43:46.080 align:middle line:84% This decreases the sample size I originally I had already. 00:43:46.080 --> 00:43:47.580 align:middle line:84% So what happened is that people use 00:43:47.580 --> 00:43:50.640 align:middle line:84% a thing that is called leave one out cross-validation, 00:43:50.640 --> 00:43:57.910 align:middle line:84% where the number of sets is equal to the number of samples. 00:43:57.910 --> 00:44:00.815 align:middle line:84% So what means that they pull out one sample. 00:44:00.815 --> 00:44:02.190 align:middle line:84% I build a model on the other one. 00:44:02.190 --> 00:44:03.607 align:middle line:84% I try to predict the sample that's 00:44:03.607 --> 00:44:05.760 align:middle line:84% taken out of which I know the classification. 00:44:05.760 --> 00:44:09.680 align:middle line:90% 00:44:09.680 --> 00:44:11.300 align:middle line:90% This is an example. 00:44:11.300 --> 00:44:17.120 align:middle line:84% One of the first predictive models that came out in 1999. 00:44:17.120 --> 00:44:20.960 align:middle line:90% We have two types of leukemia-- 00:44:20.960 --> 00:44:23.540 align:middle line:84% ALL and AML, acute lymphoblastic leukemia, 00:44:23.540 --> 00:44:25.010 align:middle line:90% acute myeloid leukemia. 00:44:25.010 --> 00:44:28.500 align:middle line:84% And as you can see under the microscope, 00:44:28.500 --> 00:44:31.790 align:middle line:84% they are very difficult to diagnose. 00:44:31.790 --> 00:44:38.220 align:middle line:84% So what these people at [? Wycliffe ?] did was to say, 00:44:38.220 --> 00:44:43.010 align:middle line:84% well, let's collect, I think, 27 and 11 patients, right? 00:44:43.010 --> 00:44:48.290 align:middle line:84% And what they did was to create a dummy vector of zeros 00:44:48.290 --> 00:44:53.700 align:middle line:84% and ones and then correlate the gene expression-- 00:44:53.700 --> 00:44:55.880 align:middle line:90% sorry. 00:44:55.880 --> 00:44:58.880 align:middle line:90% The columns are patients. 00:44:58.880 --> 00:45:04.310 align:middle line:90% The rows are genes, right? 00:45:04.310 --> 00:45:09.020 align:middle line:84% And now I don't remember if the blue is underexpressed 00:45:09.020 --> 00:45:09.860 align:middle line:90% or overexpressed. 00:45:09.860 --> 00:45:14.120 align:middle line:84% But what does it mean is that they take some kind of average 00:45:14.120 --> 00:45:15.500 align:middle line:90% to represent this picture. 00:45:15.500 --> 00:45:20.570 align:middle line:84% And the positive distance of the point from this average 00:45:20.570 --> 00:45:22.760 align:middle line:90% is the intensity of the red. 00:45:22.760 --> 00:45:26.210 align:middle line:84% And the negative distance is the intensity of the blue. 00:45:26.210 --> 00:45:28.730 align:middle line:84% So the more intense is the color, 00:45:28.730 --> 00:45:32.900 align:middle line:84% the farther would be your point compared 00:45:32.900 --> 00:45:34.550 align:middle line:90% to the mean of these values. 00:45:34.550 --> 00:45:39.230 align:middle line:84% And at the same time, the direction of this distance 00:45:39.230 --> 00:45:40.830 align:middle line:90% would be given by the color. 00:45:40.830 --> 00:45:43.350 align:middle line:84% So if it's dark blue, it would be very negative. 00:45:43.350 --> 00:45:47.150 align:middle line:90% 00:45:47.150 --> 00:45:50.280 align:middle line:84% If it's dark red, it would be very positive. 00:45:50.280 --> 00:45:52.550 align:middle line:84% So what they did was to correlate these genes 00:45:52.550 --> 00:45:54.720 align:middle line:90% and pull out the top 50. 00:45:54.720 --> 00:45:57.770 align:middle line:84% So the 50 that correlated more with the gene, 00:45:57.770 --> 00:46:03.260 align:middle line:84% with this dummy vector with the positive correlation 00:46:03.260 --> 00:46:07.190 align:middle line:84% and with the negative correlation, 50 and 50. 00:46:07.190 --> 00:46:11.540 align:middle line:84% And what they did then was to make a mixture of expert 00:46:11.540 --> 00:46:14.300 align:middle line:84% prediction and see what was the accuracy they could 00:46:14.300 --> 00:46:16.350 align:middle line:90% get of their own patients. 00:46:16.350 --> 00:46:18.830 align:middle line:84% And since then, there have been gazillions 00:46:18.830 --> 00:46:21.590 align:middle line:90% of paper written like this. 00:46:21.590 --> 00:46:24.090 align:middle line:84% I want to stress the fact that, in this case, 00:46:24.090 --> 00:46:28.680 align:middle line:84% we are not interested, again, in what is really different. 00:46:28.680 --> 00:46:31.220 align:middle line:84% We are interested in finding a molecular classification 00:46:31.220 --> 00:46:32.660 align:middle line:90% for these things. 00:46:32.660 --> 00:46:36.740 align:middle line:84% The hope here is that one day you can build a little check-- 00:46:36.740 --> 00:46:40.790 align:middle line:84% and they are really doing this for literature-- 00:46:40.790 --> 00:46:43.550 align:middle line:84% on which you can put some specific genes 00:46:43.550 --> 00:46:47.060 align:middle line:84% and have a classification that will tell you 00:46:47.060 --> 00:46:49.730 align:middle line:84% this patient has this particular type of leukemia. 00:46:49.730 --> 00:46:53.720 align:middle line:84% This patient has this other type of leukemia. 00:46:53.720 --> 00:46:58.400 align:middle line:84% OK, so I have talked about something 00:46:58.400 --> 00:47:05.120 align:middle line:84% that is [INAUDIBLE] question, is controversial thing. 00:47:05.120 --> 00:47:08.030 align:middle line:84% But I thought last night about including this thing. 00:47:08.030 --> 00:47:12.590 align:middle line:84% But then I said to myself, yeah, as long 00:47:12.590 --> 00:47:14.450 align:middle line:84% as I tell you that what I'm going to tell 00:47:14.450 --> 00:47:16.070 align:middle line:84% you may be kind of controversial. 00:47:16.070 --> 00:47:17.690 align:middle line:84% It's OK if I tell you that, right? 00:47:17.690 --> 00:47:20.963 align:middle line:84% And this is why you want to go to school to be a professor. 00:47:20.963 --> 00:47:22.880 align:middle line:84% Because then you can say controversial things. 00:47:22.880 --> 00:47:26.600 align:middle line:90% They cannot fire you, hopefully. 00:47:26.600 --> 00:47:34.250 align:middle line:84% One of the things people do to identify differences easier, 00:47:34.250 --> 00:47:40.970 align:middle line:84% even with pass it down, cut the threshold, 00:47:40.970 --> 00:47:44.720 align:middle line:90% is to deflate the variance, OK? 00:47:44.720 --> 00:47:47.340 align:middle line:84% If I have two samples that are very far-- 00:47:47.340 --> 00:47:52.130 align:middle line:84% but if I find a way to squeeze the variance of these cases, 00:47:52.130 --> 00:47:55.190 align:middle line:84% then I will have a much smaller variance. 00:47:55.190 --> 00:48:00.110 align:middle line:84% And I have more chances that my changes will be significant, 00:48:00.110 --> 00:48:00.830 align:middle line:90% right? 00:48:00.830 --> 00:48:03.320 align:middle line:84% Because the variance would be smaller. 00:48:03.320 --> 00:48:06.740 align:middle line:84% Now this is something that for any other type of data 00:48:06.740 --> 00:48:10.640 align:middle line:84% analysis, will send you at least in disrepute, sometimes 00:48:10.640 --> 00:48:11.450 align:middle line:90% in jail. 00:48:11.450 --> 00:48:14.390 align:middle line:84% If you do this on a company budget 00:48:14.390 --> 00:48:18.380 align:middle line:84% or if you do this on a clinical trial, you go to jail. 00:48:18.380 --> 00:48:20.840 align:middle line:84% In microarrays, people don't go to jail. 00:48:20.840 --> 00:48:24.930 align:middle line:84% Because it is an original thing that made originally sense. 00:48:24.930 --> 00:48:29.210 align:middle line:84% Remember cDNA microarrays? cDNA microarrays have two channels. 00:48:29.210 --> 00:48:31.280 align:middle line:84% Now we know that, by design, there 00:48:31.280 --> 00:48:34.100 align:middle line:84% is an imbalance between these two channels. 00:48:34.100 --> 00:48:36.180 align:middle line:84% One channel is more intense than another. 00:48:36.180 --> 00:48:39.380 align:middle line:84% So if I'm comparing to samples, what I may come up with 00:48:39.380 --> 00:48:43.985 align:middle line:84% is something that looks like this. 00:48:43.985 --> 00:48:47.200 align:middle line:90% 00:48:47.200 --> 00:48:48.990 align:middle line:84% So I have the two microarrays that 00:48:48.990 --> 00:48:54.240 align:middle line:84% are lying on two parallel things, right? 00:48:54.240 --> 00:48:56.130 align:middle line:84% And you see that there is a bias. 00:48:56.130 --> 00:49:00.220 align:middle line:84% All the red on one side and all the blue are on the other side. 00:49:00.220 --> 00:49:06.720 align:middle line:84% So what people used to actually do for this kind of platform, 00:49:06.720 --> 00:49:09.240 align:middle line:84% because you have two channels, is 00:49:09.240 --> 00:49:13.410 align:middle line:84% to try to reconcile these two channels by studying 00:49:13.410 --> 00:49:15.540 align:middle line:84% the distribution of these two things 00:49:15.540 --> 00:49:18.190 align:middle line:84% and try to put them one over the other. 00:49:18.190 --> 00:49:20.250 align:middle line:84% So as a form of correction, you do. 00:49:20.250 --> 00:49:23.100 align:middle line:84% Because, by design, you know that your platform 00:49:23.100 --> 00:49:24.570 align:middle line:90% will introduce some bias. 00:49:24.570 --> 00:49:25.590 align:middle line:90% And this is fair. 00:49:25.590 --> 00:49:28.660 align:middle line:90% This is good. 00:49:28.660 --> 00:49:31.260 align:middle line:84% The problem is that when oligonucleotide microarrays 00:49:31.260 --> 00:49:34.830 align:middle line:84% were introduced, people just blindly took these things 00:49:34.830 --> 00:49:36.900 align:middle line:90% and try to apply to microarrays. 00:49:36.900 --> 00:49:39.820 align:middle line:84% And you start coming up with a couple of problems. 00:49:39.820 --> 00:49:43.590 align:middle line:84% First of all, oligonucleotide microarrays 00:49:43.590 --> 00:49:44.940 align:middle line:90% are not two channels. 00:49:44.940 --> 00:49:46.770 align:middle line:90% They're one channel. 00:49:46.770 --> 00:49:50.430 align:middle line:84% So suppose there have 50 patients. 00:49:50.430 --> 00:49:51.730 align:middle line:90% What do I do there? 00:49:51.730 --> 00:49:55.890 align:middle line:84% Which patient do I take to be my baseline? 00:49:55.890 --> 00:49:59.400 align:middle line:84% I'm going to reconcile all the patients with the first patient 00:49:59.400 --> 00:50:01.170 align:middle line:90% at the beginning? 00:50:01.170 --> 00:50:03.120 align:middle line:84% And what happens if I change this patient? 00:50:03.120 --> 00:50:05.430 align:middle line:90% Are my genes going to change? 00:50:05.430 --> 00:50:08.450 align:middle line:90% Yeah, you bet so. 00:50:08.450 --> 00:50:13.850 align:middle line:84% So now if you really want to have a great success 00:50:13.850 --> 00:50:16.670 align:middle line:84% talk with biologists, go and tell them 00:50:16.670 --> 00:50:19.310 align:middle line:90% that they shouldn't normalize. 00:50:19.310 --> 00:50:22.820 align:middle line:84% Because there are about 100 different normalization 00:50:22.820 --> 00:50:24.590 align:middle line:90% methods of this type. 00:50:24.590 --> 00:50:26.630 align:middle line:90% And people are confused. 00:50:26.630 --> 00:50:28.970 align:middle line:84% But people are confused because there is really no need. 00:50:28.970 --> 00:50:31.250 align:middle line:84% People are not confused on normalization and cDNAs. 00:50:31.250 --> 00:50:33.020 align:middle line:84% People are confused for normalization 00:50:33.020 --> 00:50:35.990 align:middle line:84% to squeeze your variance and get better results. 00:50:35.990 --> 00:50:39.050 align:middle line:84% Because in reality, even when you have design 00:50:39.050 --> 00:50:43.040 align:middle line:84% with two channels-- so I have a pair case and control 00:50:43.040 --> 00:50:44.240 align:middle line:90% with microarrays. 00:50:44.240 --> 00:50:48.530 align:middle line:84% You get actually results that look like this one. 00:50:48.530 --> 00:50:51.590 align:middle line:84% Now these are microarrays that come 00:50:51.590 --> 00:50:55.730 align:middle line:84% from an institution from this [? street ?] 00:50:55.730 --> 00:51:00.590 align:middle line:84% to which I am not affiliated and nobody here is affiliated to. 00:51:00.590 --> 00:51:02.850 align:middle line:90% So I can actually speak on them. 00:51:02.850 --> 00:51:10.130 align:middle line:84% And this is a good example of why not to do normalization. 00:51:10.130 --> 00:51:18.200 align:middle line:84% So these are people before and after treatment, OK? 00:51:18.200 --> 00:51:20.180 align:middle line:84% These are paired experiments because it's 00:51:20.180 --> 00:51:23.780 align:middle line:84% the same person that is sample before treatment 00:51:23.780 --> 00:51:26.430 align:middle line:90% and after treatment. 00:51:26.430 --> 00:51:28.340 align:middle line:84% So you remember those lines that were 00:51:28.340 --> 00:51:30.530 align:middle line:90% like going one after the other? 00:51:30.530 --> 00:51:36.260 align:middle line:84% Means that we were plotting the intensity 00:51:36.260 --> 00:51:39.330 align:middle line:84% of one channel against the intensity of the other channel. 00:51:39.330 --> 00:51:40.160 align:middle line:90% So look. 00:51:40.160 --> 00:51:44.180 align:middle line:84% We plot this microarray against this microarray, 00:51:44.180 --> 00:51:47.810 align:middle line:84% which is the microarray before and after, right? 00:51:47.810 --> 00:51:51.650 align:middle line:84% So in this case, yeah, more or less, 00:51:51.650 --> 00:51:53.510 align:middle line:90% it looks like the other one. 00:51:53.510 --> 00:51:55.220 align:middle line:90% You remember? 00:51:55.220 --> 00:51:56.420 align:middle line:90% Now look at this one. 00:51:56.420 --> 00:52:01.440 align:middle line:90% 00:52:01.440 --> 00:52:03.930 align:middle line:84% Can you imagine any transformation 00:52:03.930 --> 00:52:07.530 align:middle line:84% that will put those things along the same line? 00:52:07.530 --> 00:52:08.220 align:middle line:90% Yeah. 00:52:08.220 --> 00:52:09.150 align:middle line:90% Look at this one. 00:52:09.150 --> 00:52:14.380 align:middle line:90% 00:52:14.380 --> 00:52:15.960 align:middle line:84% So in this case, what happened is 00:52:15.960 --> 00:52:18.660 align:middle line:84% that there is something that is highly screwed up. 00:52:18.660 --> 00:52:21.870 align:middle line:84% And, again, these are following exactly the same design 00:52:21.870 --> 00:52:25.010 align:middle line:84% that cDNA followed, the experimental design, 00:52:25.010 --> 00:52:28.230 align:middle line:84% although the end practice is absolutely different. 00:52:28.230 --> 00:52:31.590 align:middle line:84% So my advice as far as normalization 00:52:31.590 --> 00:52:35.730 align:middle line:84% is concerned is don't change your data that may be useful. 00:52:35.730 --> 00:52:39.330 align:middle line:84% But try to look at your data because they may contain 00:52:39.330 --> 00:52:41.580 align:middle line:90% some important information. 00:52:41.580 --> 00:52:43.740 align:middle line:84% This microarray is completely screwed up 00:52:43.740 --> 00:52:48.420 align:middle line:84% and should be either removed, redone, or done something 00:52:48.420 --> 00:52:49.020 align:middle line:90% about it. 00:52:49.020 --> 00:52:52.890 align:middle line:90% 00:52:52.890 --> 00:52:54.180 align:middle line:90% OK, so what have we learned? 00:52:54.180 --> 00:53:02.700 align:middle line:84% We have learned that we can actually find differences 00:53:02.700 --> 00:53:05.640 align:middle line:84% among samples in different conditions. 00:53:05.640 --> 00:53:07.620 align:middle line:90% We can make predictors. 00:53:07.620 --> 00:53:10.140 align:middle line:84% Have we learn anything interesting 00:53:10.140 --> 00:53:14.430 align:middle line:90% about the genome cells? 00:53:14.430 --> 00:53:16.260 align:middle line:90% Not really. 00:53:16.260 --> 00:53:19.770 align:middle line:84% We have learned nothing about the relationship among genes. 00:53:19.770 --> 00:53:22.420 align:middle line:84% Although we are measuring all of them at the same time, 00:53:22.420 --> 00:53:24.690 align:middle line:84% we have completely disregarded-- actually we 00:53:24.690 --> 00:53:28.635 align:middle line:84% have fight against the idea that these things could be related. 00:53:28.635 --> 00:53:30.510 align:middle line:84% We are simply interested in finding something 00:53:30.510 --> 00:53:33.250 align:middle line:84% that was different in two conditions or simply interested 00:53:33.250 --> 00:53:35.130 align:middle line:84% in finding something that, put together, 00:53:35.130 --> 00:53:37.140 align:middle line:90% could predict this condition. 00:53:37.140 --> 00:53:38.250 align:middle line:90% That's it. 00:53:38.250 --> 00:53:42.120 align:middle line:84% This is where supervised classification brings us. 00:53:42.120 --> 00:53:44.490 align:middle line:84% If we want to take advantage of the fact 00:53:44.490 --> 00:53:46.710 align:middle line:84% that we measured all these genes and we observe 00:53:46.710 --> 00:53:50.010 align:middle line:84% the genome in action to try to decode something 00:53:50.010 --> 00:53:53.310 align:middle line:84% about the genome, then we need some different method. 00:53:53.310 --> 00:53:56.430 align:middle line:84% And we don't need supervision all the time. 00:53:56.430 --> 00:53:58.830 align:middle line:90% It's like when you're a kid. 00:53:58.830 --> 00:54:02.250 align:middle line:84% If there is supervision, there is very little fun. 00:54:02.250 --> 00:54:05.730 align:middle line:84% So the easiest thing we can do is 00:54:05.730 --> 00:54:10.080 align:middle line:84% to say, well, OK, forget about supervision. 00:54:10.080 --> 00:54:13.740 align:middle line:84% I got this bunch of genes in different conditions. 00:54:13.740 --> 00:54:15.060 align:middle line:90% Forget about these conditions. 00:54:15.060 --> 00:54:16.830 align:middle line:84% I don't care about these conditions. 00:54:16.830 --> 00:54:20.160 align:middle line:84% What I want to see is which are the genes that 00:54:20.160 --> 00:54:24.030 align:middle line:84% behave more similarly across all these different conditions? 00:54:24.030 --> 00:54:26.310 align:middle line:90% It's like having a car, right? 00:54:26.310 --> 00:54:32.070 align:middle line:84% Try to understand how it works by kicking it, and kicking 00:54:32.070 --> 00:54:34.710 align:middle line:84% in different points of the car, and then 00:54:34.710 --> 00:54:39.580 align:middle line:84% see how the things go together under different stresses. 00:54:39.580 --> 00:54:48.060 align:middle line:84% So if I kick the wheel, if I keep the trunk, if I 00:54:48.060 --> 00:54:50.130 align:middle line:90% kick the door, what happens? 00:54:50.130 --> 00:54:51.810 align:middle line:90% How these things move together? 00:54:51.810 --> 00:54:54.640 align:middle line:84% What is the relationship among these things? 00:54:54.640 --> 00:54:57.020 align:middle line:84% Was a nice analogy some time ago. 00:54:57.020 --> 00:54:59.020 align:middle line:90% But you study these things. 00:54:59.020 --> 00:55:01.870 align:middle line:84% And the way in which you study these things-- 00:55:01.870 --> 00:55:05.520 align:middle line:84% this was for sequencing the genome. 00:55:05.520 --> 00:55:08.670 align:middle line:84% Well, it's like you have, in the future, 00:55:08.670 --> 00:55:10.920 align:middle line:84% somebody comes up with the Volkswagen. 00:55:10.920 --> 00:55:13.200 align:middle line:84% And they discover a Volkswagen [INAUDIBLE] somewhere. 00:55:13.200 --> 00:55:15.280 align:middle line:84% And they have no idea what it is. 00:55:15.280 --> 00:55:18.030 align:middle line:84% So to understand how this works, they take the Volkswagen 00:55:18.030 --> 00:55:20.460 align:middle line:90% and they throw it off the cliff. 00:55:20.460 --> 00:55:24.690 align:middle line:84% And then, when it's down, they try to put the pieces together 00:55:24.690 --> 00:55:27.000 align:middle line:90% again, right? 00:55:27.000 --> 00:55:30.290 align:middle line:84% This is what somehow we're trying to do. 00:55:30.290 --> 00:55:34.950 align:middle line:84% We are breaking this down with some kind of solicitations 00:55:34.950 --> 00:55:37.630 align:middle line:84% and trying to see which parts behave together. 00:55:37.630 --> 00:55:40.500 align:middle line:84% So, in this case, like we had 1,000 Volkswagens-- well, 00:55:40.500 --> 00:55:41.610 align:middle line:90% 100 Volkswagens. 00:55:41.610 --> 00:55:43.590 align:middle line:90% And we keep throwing them down. 00:55:43.590 --> 00:55:45.120 align:middle line:84% And at the end, when they are down-- 00:55:45.120 --> 00:55:49.333 align:middle line:84% because we don't know how to open the engine-- 00:55:49.333 --> 00:55:51.750 align:middle line:84% when they are down, we will see there are some pieces that 00:55:51.750 --> 00:55:52.860 align:middle line:90% are closer together. 00:55:52.860 --> 00:55:54.720 align:middle line:90% And they remain closer together. 00:55:54.720 --> 00:55:57.390 align:middle line:84% And this is independent of the fact 00:55:57.390 --> 00:55:59.730 align:middle line:84% that these two things fall to the left 00:55:59.730 --> 00:56:03.870 align:middle line:84% or to the right of the main body of the Volkswagen. 00:56:03.870 --> 00:56:07.800 align:middle line:84% So a simple thing is to say, well, let's measure 00:56:07.800 --> 00:56:09.540 align:middle line:90% correlation among these things. 00:56:09.540 --> 00:56:11.300 align:middle line:90% Genes [INAUDIBLE] supervision. 00:56:11.300 --> 00:56:13.050 align:middle line:84% They have a lot of solicitations. 00:56:13.050 --> 00:56:16.650 align:middle line:84% They could be different compounds 00:56:16.650 --> 00:56:18.510 align:middle line:84% that are treating a particular disease. 00:56:18.510 --> 00:56:20.910 align:middle line:84% These are maybe different type of cancer. 00:56:20.910 --> 00:56:21.630 align:middle line:90% I don't care. 00:56:21.630 --> 00:56:23.130 align:middle line:84% I don't want to find classification. 00:56:23.130 --> 00:56:26.880 align:middle line:84% I just want to find out which are the genes that go together 00:56:26.880 --> 00:56:29.640 align:middle line:90% around these conditions. 00:56:29.640 --> 00:56:32.050 align:middle line:84% If I use correlation, the only thing I can do, 00:56:32.050 --> 00:56:36.330 align:middle line:84% though, is to look at pairwise comparisons, right? 00:56:36.330 --> 00:56:38.940 align:middle line:84% I can only say that one gene go to another gene. 00:56:38.940 --> 00:56:41.040 align:middle line:84% A correlation is a distance between two points. 00:56:41.040 --> 00:56:45.100 align:middle line:84% I cannot have groups of three, or five, or 15. 00:56:45.100 --> 00:56:46.600 align:middle line:84% How can I put these things together? 00:56:46.600 --> 00:56:48.570 align:middle line:84% Well, I can use another type of clustering 00:56:48.570 --> 00:56:50.250 align:middle line:90% called hierarchical clustering. 00:56:50.250 --> 00:56:52.548 align:middle line:84% Hierarchical clustering start putting things together. 00:56:52.548 --> 00:56:54.090 align:middle line:84% But when it puts two things together, 00:56:54.090 --> 00:56:57.960 align:middle line:84% it creates some kind of a dummy gene, which 00:56:57.960 --> 00:57:00.330 align:middle line:84% make us feel like the average of these two genes 00:57:00.330 --> 00:57:01.450 align:middle line:90% or something like this. 00:57:01.450 --> 00:57:04.560 align:middle line:84% And then try to correlate this average profile, 00:57:04.560 --> 00:57:07.900 align:middle line:84% this average gene, with other genes. 00:57:07.900 --> 00:57:12.570 align:middle line:84% So, at the end, the result would be something like this. 00:57:12.570 --> 00:57:15.130 align:middle line:90% 00:57:15.130 --> 00:57:17.100 align:middle line:84% Again, it's like the blue and the red. 00:57:17.100 --> 00:57:21.990 align:middle line:90% In this case, is green and red. 00:57:21.990 --> 00:57:23.610 align:middle line:90% These are the Stanford color. 00:57:23.610 --> 00:57:29.910 align:middle line:90% Wycliffe uses the blue and pink. 00:57:29.910 --> 00:57:32.340 align:middle line:90% Duke, I think, use yellow. 00:57:32.340 --> 00:57:36.810 align:middle line:84% John Hopkins use green and blue-- 00:57:36.810 --> 00:57:38.580 align:middle line:84% well, a few combinations of this. 00:57:38.580 --> 00:57:41.040 align:middle line:84% But you can actually recognize at least the platform 00:57:41.040 --> 00:57:43.620 align:middle line:84% they're using by the color of their pictures. 00:57:43.620 --> 00:57:45.810 align:middle line:84% So in this case, this is a Stanford picture. 00:57:45.810 --> 00:57:49.780 align:middle line:84% Again, the green is down and the red is up or vice versa. 00:57:49.780 --> 00:57:52.380 align:middle line:84% And you can actually, by visual inspection, 00:57:52.380 --> 00:57:54.240 align:middle line:84% see that there are some points that 00:57:54.240 --> 00:57:56.940 align:middle line:84% are very highly expressed away from the mean, 00:57:56.940 --> 00:58:01.290 align:middle line:84% are very down expresses, very [? low ?] expressed 00:58:01.290 --> 00:58:04.200 align:middle line:90% from the mean. 00:58:04.200 --> 00:58:06.820 align:middle line:84% So this is the zoom of that picture. 00:58:06.820 --> 00:58:11.100 align:middle line:84% And you can see that these things are creating 00:58:11.100 --> 00:58:13.290 align:middle line:90% a tree or Venn diagram. 00:58:13.290 --> 00:58:16.590 align:middle line:84% And this tree will put together groups 00:58:16.590 --> 00:58:19.890 align:middle line:90% of genes, not only two genes. 00:58:19.890 --> 00:58:22.818 align:middle line:84% And the problem here is that you don't really 00:58:22.818 --> 00:58:25.110 align:middle line:84% have a good measure to decide when you've made a group. 00:58:25.110 --> 00:58:27.660 align:middle line:84% Because, again, you have one single tree 00:58:27.660 --> 00:58:30.550 align:middle line:84% that will combine all of them in different order. 00:58:30.550 --> 00:58:33.027 align:middle line:90% So, technically, this is not-- 00:58:33.027 --> 00:58:34.860 align:middle line:84% although it's called clustering-- clustering 00:58:34.860 --> 00:58:38.880 align:middle line:84% means to put things together and divide them. 00:58:38.880 --> 00:58:42.180 align:middle line:84% Technically, this is a sorting algorithm by which 00:58:42.180 --> 00:58:43.780 align:middle line:90% I put a particular order-- 00:58:43.780 --> 00:58:45.090 align:middle line:90% in this case, a partial order-- 00:58:45.090 --> 00:58:46.230 align:middle line:90% over these things. 00:58:46.230 --> 00:58:50.580 align:middle line:84% And then some knowledgeable biologists 00:58:50.580 --> 00:58:56.040 align:middle line:84% will come and say, oh, among these people here 00:58:56.040 --> 00:59:03.510 align:middle line:84% in this group, I see that there are among this group-- 00:59:03.510 --> 00:59:06.480 align:middle line:84% I see that these genes are all related 00:59:06.480 --> 00:59:08.440 align:middle line:90% to a particular process. 00:59:08.440 --> 00:59:11.850 align:middle line:84% So maybe also these genes that is right embedded between them 00:59:11.850 --> 00:59:13.920 align:middle line:90% is related to the same process. 00:59:13.920 --> 00:59:16.020 align:middle line:90% And maybe it's apoptosis. 00:59:16.020 --> 00:59:17.902 align:middle line:84% And these are five apoptotic genes. 00:59:17.902 --> 00:59:19.360 align:middle line:84% And then, they find something else. 00:59:19.360 --> 00:59:20.610 align:middle line:90% And we create another group. 00:59:20.610 --> 00:59:23.130 align:middle line:84% But these groups, these different coloring-- 00:59:23.130 --> 00:59:26.280 align:middle line:84% the pink, the purple, and the red 00:59:26.280 --> 00:59:30.630 align:middle line:84% there-- are handmade by somebody with a lot of patience 00:59:30.630 --> 00:59:31.800 align:middle line:90% that put them together. 00:59:31.800 --> 00:59:34.653 align:middle line:90% 00:59:34.653 --> 00:59:35.570 align:middle line:90% AUDIENCE: [INAUDIBLE]. 00:59:35.570 --> 00:59:36.780 align:middle line:90% MARCO RAMONI: Say again. 00:59:36.780 --> 00:59:37.280 align:middle line:90% What? 00:59:37.280 --> 00:59:37.940 align:middle line:84% AUDIENCE: The trees are made by hand? 00:59:37.940 --> 00:59:38.210 align:middle line:90% MARCO RAMONI: No, no. 00:59:38.210 --> 00:59:39.470 align:middle line:90% The trees-- sorry. 00:59:39.470 --> 00:59:42.350 align:middle line:90% 00:59:42.350 --> 00:59:48.480 align:middle line:84% The tree itself is built through some kind of metric. 00:59:48.480 --> 00:59:49.940 align:middle line:84% I don't know why it's not coming-- 00:59:49.940 --> 00:59:50.510 align:middle line:90% OK. 00:59:50.510 --> 00:59:57.560 align:middle line:84% So I compute the correlation between these two points, 00:59:57.560 --> 00:59:59.870 align:middle line:90% these two vectors of values. 00:59:59.870 --> 01:00:04.850 align:middle line:84% Then, I create, let's say, an average value here. 01:00:04.850 --> 01:00:08.030 align:middle line:84% And then, I draw these two points. 01:00:08.030 --> 01:00:11.180 align:middle line:84% And they consider this new value that I have created 01:00:11.180 --> 01:00:13.850 align:middle line:90% as a new member of my data set. 01:00:13.850 --> 01:00:15.920 align:middle line:84% I didn't see what this correlates to. 01:00:15.920 --> 01:00:19.980 align:middle line:84% In this case, this correlates to this one. 01:00:19.980 --> 01:00:23.210 align:middle line:84% So the highest correlated thing is a gene. 01:00:23.210 --> 01:00:26.310 align:middle line:84% And this creates a new hypothetical thing, 01:00:26.310 --> 01:00:30.360 align:middle line:84% which is the average of these two and this one. 01:00:30.360 --> 01:00:32.150 align:middle line:84% So what happens is that, at the end, 01:00:32.150 --> 01:00:33.930 align:middle line:90% I create a structure like this. 01:00:33.930 --> 01:00:36.110 align:middle line:84% But the problem is that because they 01:00:36.110 --> 01:00:40.130 align:middle line:84% are all measures, at the end, they will have one single tree. 01:00:40.130 --> 01:00:42.380 align:middle line:90% So how do I create blocks? 01:00:42.380 --> 01:00:44.715 align:middle line:84% The way in which blocks are created-- and I say, 01:00:44.715 --> 01:00:47.300 align:middle line:84% color this in purple and this in pink. 01:00:47.300 --> 01:00:49.382 align:middle line:90% These were handmade. 01:00:49.382 --> 01:00:50.840 align:middle line:84% I will tell you in a second how you 01:00:50.840 --> 01:00:52.350 align:middle line:90% can avoid to do this handmade. 01:00:52.350 --> 01:00:57.430 align:middle line:90% 01:00:57.430 --> 01:00:59.560 align:middle line:84% I can do something more interesting also. 01:00:59.560 --> 01:01:06.097 align:middle line:90% 01:01:06.097 --> 01:01:07.930 align:middle line:84% That was a temporary experiment, [INAUDIBLE] 01:01:07.930 --> 01:01:09.190 align:middle line:90% second temporary experiment. 01:01:09.190 --> 01:01:12.010 align:middle line:84% So I knew the order of these microarrays. 01:01:12.010 --> 01:01:14.020 align:middle line:84% But sometimes, I'm not really interested 01:01:14.020 --> 01:01:16.820 align:middle line:84% only in the way in which genes go together. 01:01:16.820 --> 01:01:22.240 align:middle line:84% I'm also interested in finding some new class among patients, 01:01:22.240 --> 01:01:22.810 align:middle line:90% right? 01:01:22.810 --> 01:01:25.780 align:middle line:84% This is a very interesting paper from 2000 01:01:25.780 --> 01:01:27.580 align:middle line:84% in which what these people did was 01:01:27.580 --> 01:01:35.650 align:middle line:84% to try to cluster simultaneously genes and patients. 01:01:35.650 --> 01:01:39.100 align:middle line:84% And what they came up with were groups. 01:01:39.100 --> 01:01:40.480 align:middle line:90% You see those groups up there. 01:01:40.480 --> 01:01:42.760 align:middle line:84% The groups up there is not a Venn diagram. 01:01:42.760 --> 01:01:44.740 align:middle line:84% Those are group of patients based 01:01:44.740 --> 01:01:47.620 align:middle line:84% on some selection of genes that are more expressed 01:01:47.620 --> 01:01:49.510 align:middle line:90% across the two conditions. 01:01:49.510 --> 01:01:52.600 align:middle line:84% And then what they did was to find out that-- 01:01:52.600 --> 01:01:54.875 align:middle line:84% if you look at the survival time-- 01:01:54.875 --> 01:01:59.030 align:middle line:84% how many of what is a Kaplan-Meier curve? 01:01:59.030 --> 01:01:59.680 align:middle line:90% Everybody. 01:01:59.680 --> 01:02:04.060 align:middle line:84% OK, so if you look at the Kaplan-Meier curves 01:02:04.060 --> 01:02:07.570 align:middle line:84% of those groups, you see that there are very significant 01:02:07.570 --> 01:02:10.390 align:middle line:90% difference in survival, OK? 01:02:10.390 --> 01:02:15.310 align:middle line:84% So in this way, I can discover not something really 01:02:15.310 --> 01:02:16.930 align:middle line:84% that is about genes but something 01:02:16.930 --> 01:02:19.750 align:middle line:84% that is about the overall classified disease. 01:02:19.750 --> 01:02:23.020 align:middle line:84% I find out new classification for diseases 01:02:23.020 --> 01:02:28.720 align:middle line:84% with interesting clinical consequences. 01:02:28.720 --> 01:02:32.500 align:middle line:84% Again, problem is I have to do this darn coloring by hand. 01:02:32.500 --> 01:02:35.260 align:middle line:84% Is there a way by which we can actually 01:02:35.260 --> 01:02:37.030 align:middle line:90% avoid coloring this stuff? 01:02:37.030 --> 01:02:37.750 align:middle line:90% Yeah. 01:02:37.750 --> 01:02:38.650 align:middle line:90% There is a way. 01:02:38.650 --> 01:02:41.200 align:middle line:90% 01:02:41.200 --> 01:02:42.310 align:middle line:90% And this is the idea. 01:02:42.310 --> 01:02:45.720 align:middle line:90% 01:02:45.720 --> 01:02:47.520 align:middle line:84% If you want to cluster, it means that you 01:02:47.520 --> 01:02:49.930 align:middle line:84% have to make differences among things. 01:02:49.930 --> 01:02:52.530 align:middle line:84% So you can decide arbitrarily the number of clusters. 01:02:52.530 --> 01:02:54.390 align:middle line:90% And say, OK, I have 50 clusters. 01:02:54.390 --> 01:02:56.520 align:middle line:84% And you divide everything in 50 parts. 01:02:56.520 --> 01:03:01.620 align:middle line:84% But why not 49, or 38, or 15, or two. 01:03:01.620 --> 01:03:08.320 align:middle line:84% So central notion of the clustering is similarity. 01:03:08.320 --> 01:03:10.780 align:middle line:84% If we have a definition of similarity 01:03:10.780 --> 01:03:14.290 align:middle line:84% that is specific enough, then this similarity 01:03:14.290 --> 01:03:17.560 align:middle line:84% will allow us to say when we can actually 01:03:17.560 --> 01:03:21.070 align:middle line:84% cluster without creating a threshold, just 01:03:21.070 --> 01:03:23.440 align:middle line:84% a conceptual definition of similarity. 01:03:23.440 --> 01:03:27.190 align:middle line:84% So I have to postulate this description of similarity. 01:03:27.190 --> 01:03:30.130 align:middle line:84% And I need a piece of theology before this. 01:03:30.130 --> 01:03:32.860 align:middle line:90% But let me postulate this. 01:03:32.860 --> 01:03:35.650 align:middle line:84% In statistics, you don't believe that what you observed 01:03:35.650 --> 01:03:38.200 align:middle line:90% was directly created by God. 01:03:38.200 --> 01:03:42.550 align:middle line:84% What you believe is that there are some processes 01:03:42.550 --> 01:03:45.820 align:middle line:84% that you don't observe that generate the data that you 01:03:45.820 --> 01:03:51.530 align:middle line:84% observe with some randomness, some measure of uncertainty. 01:03:51.530 --> 01:03:54.260 align:middle line:90% Now let's make an example. 01:03:54.260 --> 01:03:56.410 align:middle line:84% Let's suppose we take the electrocardiograms 01:03:56.410 --> 01:03:58.670 align:middle line:90% of each of us. 01:03:58.670 --> 01:04:03.880 align:middle line:84% And, hopefully, especially for me, 01:04:03.880 --> 01:04:06.010 align:middle line:84% all these electrocardiograms would be different. 01:04:06.010 --> 01:04:08.170 align:middle line:84% But, hopefully, they would be coming 01:04:08.170 --> 01:04:11.170 align:middle line:84% from the same process, which is the process of a healthy heart, 01:04:11.170 --> 01:04:15.070 align:middle line:84% mine will be slightly different because it's small but probably 01:04:15.070 --> 01:04:18.550 align:middle line:84% will not be different enough from yours to say that this is 01:04:18.550 --> 01:04:20.620 align:middle line:90% a completely different stuff. 01:04:20.620 --> 01:04:23.590 align:middle line:84% Now suppose we go to Brigham, to cardiology at Brigham, 01:04:23.590 --> 01:04:27.430 align:middle line:84% and we take electrocardiograms of people there. 01:04:27.430 --> 01:04:31.870 align:middle line:84% There, I expect people to have differences between themselves 01:04:31.870 --> 01:04:34.570 align:middle line:84% that are great enough to be generated 01:04:34.570 --> 01:04:38.020 align:middle line:84% by different processes, different pathology 01:04:38.020 --> 01:04:39.740 align:middle line:90% of the heart. 01:04:39.740 --> 01:04:45.940 align:middle line:84% Now I will pose to you that two things are similar if they are 01:04:45.940 --> 01:04:47.770 align:middle line:90% generated by the same process. 01:04:47.770 --> 01:04:50.470 align:middle line:84% And two things are different if they are generated 01:04:50.470 --> 01:04:52.520 align:middle line:90% by two different processes. 01:04:52.520 --> 01:04:54.940 align:middle line:84% And if you buy this story, then I 01:04:54.940 --> 01:04:58.840 align:middle line:84% can give you a method to compute when something 01:04:58.840 --> 01:05:02.170 align:middle line:84% is generated by the same process and when something is not. 01:05:02.170 --> 01:05:02.890 align:middle line:90% How? 01:05:02.890 --> 01:05:05.740 align:middle line:84% Well, we know that these processes that we do not 01:05:05.740 --> 01:05:11.620 align:middle line:84% observe but they underpin the data that we actually observe, 01:05:11.620 --> 01:05:14.920 align:middle line:84% generate our data with some kind of uncertainty, 01:05:14.920 --> 01:05:20.260 align:middle line:84% that is a random process that is generating data from this. 01:05:20.260 --> 01:05:24.670 align:middle line:90% An example is aging, right? 01:05:24.670 --> 01:05:28.960 align:middle line:84% Aging has a particular effect on people, 01:05:28.960 --> 01:05:33.475 align:middle line:84% usually make you wealthier, usually, 01:05:33.475 --> 01:05:36.790 align:middle line:84% after at a certain point, make you stronger, 01:05:36.790 --> 01:05:39.150 align:middle line:84% after a certain point, make you weaker, 01:05:39.150 --> 01:05:41.980 align:middle line:84% has affect on your marital status. 01:05:41.980 --> 01:05:44.650 align:middle line:84% You tend to get married, and then divorced, or widowed, 01:05:44.650 --> 01:05:47.590 align:middle line:90% or whatever. 01:05:47.590 --> 01:05:50.110 align:middle line:84% When coupled with other variables like gender, 01:05:50.110 --> 01:05:52.760 align:middle line:84% can have other physical effects like you 01:05:52.760 --> 01:05:56.770 align:middle line:84% can lose your hair if you're male and so forth, right? 01:05:56.770 --> 01:06:00.220 align:middle line:84% So if I find somebody that at 13 is 01:06:00.220 --> 01:06:05.350 align:middle line:84% on the verge of his third divorce, that's not impossible. 01:06:05.350 --> 01:06:08.080 align:middle line:84% But I would find it kind of unlikely. 01:06:08.080 --> 01:06:08.800 align:middle line:90% Why? 01:06:08.800 --> 01:06:12.100 align:middle line:84% Because there is a process called aging that dictates, 01:06:12.100 --> 01:06:14.920 align:middle line:84% more or less, that people to be on their third divorce usually 01:06:14.920 --> 01:06:16.930 align:middle line:90% have to be at least 35. 01:06:16.930 --> 01:06:21.700 align:middle line:84% So if this guy is 13, it's difficult. It's not impossible, 01:06:21.700 --> 01:06:22.880 align:middle line:90% but it's difficult. 01:06:22.880 --> 01:06:25.083 align:middle line:84% So we have these general expectations 01:06:25.083 --> 01:06:26.500 align:middle line:84% that stem from the fact that there 01:06:26.500 --> 01:06:28.750 align:middle line:84% is these processes generating the observation 01:06:28.750 --> 01:06:32.230 align:middle line:84% that we have and is constrained by other things. 01:06:32.230 --> 01:06:34.990 align:middle line:84% Like we're saying losing her is constrained by gender, 01:06:34.990 --> 01:06:36.970 align:middle line:90% probabilities change by gender. 01:06:36.970 --> 01:06:39.980 align:middle line:84% But at the same time, once I observe the data, 01:06:39.980 --> 01:06:43.690 align:middle line:84% I can tell you that something is probable to be generated 01:06:43.690 --> 01:06:46.600 align:middle line:84% by a particular process and something less probable 01:06:46.600 --> 01:06:50.350 align:middle line:84% to be generated by a particular process, right? 01:06:50.350 --> 01:06:52.780 align:middle line:90% And this is what we want to do. 01:06:52.780 --> 01:06:56.680 align:middle line:84% We want to compute the perceived probability that a set 01:06:56.680 --> 01:06:59.080 align:middle line:84% of processes, as responsible of my data-- 01:06:59.080 --> 01:07:01.570 align:middle line:90% so M given D will is the data-- 01:07:01.570 --> 01:07:05.290 align:middle line:84% for each class we model, for each way 01:07:05.290 --> 01:07:06.760 align:middle line:90% of combining my clusters. 01:07:06.760 --> 01:07:08.260 align:middle line:84% And then, I can combine the score 01:07:08.260 --> 01:07:11.380 align:middle line:84% and find out what is the most probable way of combining 01:07:11.380 --> 01:07:12.130 align:middle line:90% these clusters. 01:07:12.130 --> 01:07:16.870 align:middle line:84% And at the very end, what I will have is a bunch of clusters, 01:07:16.870 --> 01:07:19.810 align:middle line:84% not simply a tree, not something that I 01:07:19.810 --> 01:07:21.190 align:middle line:90% have to cut with the threshold. 01:07:21.190 --> 01:07:24.730 align:middle line:84% But I would be able to tell you that if two things are 01:07:24.730 --> 01:07:28.600 align:middle line:84% put together, they are N times more probable to be generated 01:07:28.600 --> 01:07:30.280 align:middle line:84% by the same process than they are 01:07:30.280 --> 01:07:33.830 align:middle line:84% to be generated by two different processes. 01:07:33.830 --> 01:07:35.800 align:middle line:84% Interesting paper-- you're going to read it. 01:07:35.800 --> 01:07:37.940 align:middle line:90% This is how it works. 01:07:37.940 --> 01:07:39.880 align:middle line:84% The probability of the model given the data 01:07:39.880 --> 01:07:43.453 align:middle line:84% by Bayes' theorem is equal to the probability of the data 01:07:43.453 --> 01:07:45.370 align:middle line:84% given the model times [? provision ?] of model 01:07:45.370 --> 01:07:47.570 align:middle line:84% on the [? unprovision ?] of the data. 01:07:47.570 --> 01:07:54.350 align:middle line:84% Now I will not get delve into details. 01:07:54.350 --> 01:07:57.190 align:middle line:84% But at the end of the day, under some assumptions-- 01:07:57.190 --> 01:07:59.860 align:middle line:84% like the assumption that before looking at any data, 01:07:59.860 --> 01:08:02.290 align:middle line:84% all models are equally probable and the assumptions 01:08:02.290 --> 01:08:07.360 align:middle line:84% that we are trying our models on the same data, which is usually 01:08:07.360 --> 01:08:07.960 align:middle line:90% what we do. 01:08:07.960 --> 01:08:10.960 align:middle line:84% We have the same set of expression data 01:08:10.960 --> 01:08:13.030 align:middle line:90% and want to find the best model. 01:08:13.030 --> 01:08:15.920 align:middle line:84% What we can compute is that probability, 01:08:15.920 --> 01:08:18.729 align:middle line:84% which is the probability of the data given the model, which 01:08:18.729 --> 01:08:21.130 align:middle line:84% is proportional to the probability of the model given 01:08:21.130 --> 01:08:21.880 align:middle line:90% the data. 01:08:21.880 --> 01:08:24.160 align:middle line:84% And, therefore, we can use that as a score. 01:08:24.160 --> 01:08:27.040 align:middle line:84% These things is kind of compared with this to compute. 01:08:27.040 --> 01:08:28.660 align:middle line:90% It's called marginal likelihood. 01:08:28.660 --> 01:08:32.080 align:middle line:84% And so we can search all these combinations 01:08:32.080 --> 01:08:34.899 align:middle line:84% and find out which is the most likely combination, which 01:08:34.899 --> 01:08:38.229 align:middle line:84% is the most probable combination, of, in this case, 01:08:38.229 --> 01:08:40.990 align:middle line:84% genes, given the data that we observe. 01:08:40.990 --> 01:08:44.319 align:middle line:90% 01:08:44.319 --> 01:08:46.640 align:middle line:90% Now let me go-- 01:08:46.640 --> 01:08:49.450 align:middle line:84% so these are a couple advanced topics. 01:08:49.450 --> 01:08:54.955 align:middle line:84% From now on, this is not subject to examination, for the test. 01:08:54.955 --> 01:08:58.930 align:middle line:90% 01:08:58.930 --> 01:09:02.609 align:middle line:84% Suppose I'm interested in something like control. 01:09:02.609 --> 01:09:04.526 align:middle line:84% Have we learned anything about control so far? 01:09:04.526 --> 01:09:06.442 align:middle line:84% Well, we have learned that things go together, 01:09:06.442 --> 01:09:07.319 align:middle line:90% things are similar. 01:09:07.319 --> 01:09:09.210 align:middle line:84% But we haven't really learned anything 01:09:09.210 --> 01:09:12.220 align:middle line:90% about how things control things. 01:09:12.220 --> 01:09:15.149 align:middle line:84% To see how genes control other genes, 01:09:15.149 --> 01:09:17.729 align:middle line:84% we need a very important experiment design 01:09:17.729 --> 01:09:19.800 align:middle line:90% which is a temporal experiment. 01:09:19.800 --> 01:09:23.380 align:middle line:84% We need to see what happened from one point to another. 01:09:23.380 --> 01:09:25.229 align:middle line:84% And you say, well, it's kind of easy. 01:09:25.229 --> 01:09:27.180 align:middle line:90% I take this clustering method. 01:09:27.180 --> 01:09:30.060 align:middle line:84% I use this clustering method, and I put them together. 01:09:30.060 --> 01:09:32.609 align:middle line:84% And then, I will find some kind of similarities. 01:09:32.609 --> 01:09:35.040 align:middle line:90% Can I do that? 01:09:35.040 --> 01:09:35.790 align:middle line:90% No. 01:09:35.790 --> 01:09:37.590 align:middle line:90% Why? 01:09:37.590 --> 01:09:41.310 align:middle line:84% Because measures like correlation or distance 01:09:41.310 --> 01:09:44.340 align:middle line:84% measures assume that all the observations 01:09:44.340 --> 01:09:47.460 align:middle line:84% you have are marginally independent. 01:09:47.460 --> 01:09:49.859 align:middle line:84% What happened to patient one in gene 01:09:49.859 --> 01:09:54.330 align:middle line:84% one is completely unrelated to what happened to patient 01:09:54.330 --> 01:09:59.140 align:middle line:90% two on gene one, right? 01:09:59.140 --> 01:10:02.410 align:middle line:84% But when it's time-- well, when it's time, 01:10:02.410 --> 01:10:04.360 align:middle line:90% it's really, really different. 01:10:04.360 --> 01:10:10.420 align:middle line:84% Time means that where I am now depends 01:10:10.420 --> 01:10:12.730 align:middle line:84% on where it was five minutes ago, 10 minutes ago, 01:10:12.730 --> 01:10:15.110 align:middle line:90% 30 minutes ago, 100 minutes ago. 01:10:15.110 --> 01:10:17.890 align:middle line:84% So if I keep measuring the same system a long time, 01:10:17.890 --> 01:10:20.800 align:middle line:84% my observations will not be independent. 01:10:20.800 --> 01:10:23.320 align:middle line:90% Let's put it this way. 01:10:23.320 --> 01:10:25.930 align:middle line:84% If I measure things a long time, I 01:10:25.930 --> 01:10:29.470 align:middle line:84% don't have ground to safely assume that my assumptions are 01:10:29.470 --> 01:10:30.098 align:middle line:90% independent. 01:10:30.098 --> 01:10:32.140 align:middle line:84% Because assuming that assumptions are independent 01:10:32.140 --> 01:10:34.450 align:middle line:90% is a simplification, right? 01:10:34.450 --> 01:10:37.840 align:middle line:84% If I have a model that is able to account for dependency, 01:10:37.840 --> 01:10:40.570 align:middle line:84% I can always reduce it to a model of independence. 01:10:40.570 --> 01:10:42.940 align:middle line:84% But I cannot do the other way around. 01:10:42.940 --> 01:10:44.950 align:middle line:84% And let me give you a practical example. 01:10:44.950 --> 01:10:51.490 align:middle line:84% These are two pairs, two genes up and two genes down. 01:10:51.490 --> 01:10:55.270 align:middle line:84% So you are measuring the distance between these two 01:10:55.270 --> 01:10:56.590 align:middle line:90% genes. 01:10:56.590 --> 01:11:02.710 align:middle line:84% Now the correlation of the two genes 01:11:02.710 --> 01:11:05.260 align:middle line:90% up there is something like 0.6. 01:11:05.260 --> 01:11:11.950 align:middle line:84% And the correlation of the genes down there is about 0.8, right? 01:11:11.950 --> 01:11:15.490 align:middle line:84% But now consider the memory of time. 01:11:15.490 --> 01:11:17.500 align:middle line:90% And look at the first picture. 01:11:17.500 --> 01:11:20.230 align:middle line:84% Except for the first point, when the first gene 01:11:20.230 --> 01:11:22.300 align:middle line:84% goes in one direction, the second gene 01:11:22.300 --> 01:11:25.240 align:middle line:84% goes exactly the same direction, right? 01:11:25.240 --> 01:11:27.180 align:middle line:90% They never intersect each other. 01:11:27.180 --> 01:11:30.400 align:middle line:84% The second point-- it goes from one point, goes down. 01:11:30.400 --> 01:11:31.930 align:middle line:90% The second gene goes down. 01:11:31.930 --> 01:11:35.890 align:middle line:84% The third point-- the first goes up and the second goes up. 01:11:35.890 --> 01:11:38.330 align:middle line:84% And then it goes down, and goes down, and goes down again. 01:11:38.330 --> 01:11:39.580 align:middle line:90% And the other goes down again. 01:11:39.580 --> 01:11:40.705 align:middle line:90% It goes down a little less. 01:11:40.705 --> 01:11:42.580 align:middle line:90% It goes-- look at it. 01:11:42.580 --> 01:11:45.550 align:middle line:84% Now look at the other one, which has a higher correlation. 01:11:45.550 --> 01:11:48.850 align:middle line:84% These genes are always one against the other. 01:11:48.850 --> 01:11:52.490 align:middle line:84% Every time one gene goes up, the other gene goes down. 01:11:52.490 --> 01:11:56.890 align:middle line:84% So if I am actually interested in the dynamics of my system-- 01:11:56.890 --> 01:11:59.560 align:middle line:84% why correlation would put these things 01:11:59.560 --> 01:12:02.080 align:middle line:90% more similar to those ones-- 01:12:02.080 --> 01:12:05.770 align:middle line:84% my good measure-- by keeping in mind that I'm interested 01:12:05.770 --> 01:12:08.800 align:middle line:84% in the dynamics of change of this thing-- 01:12:08.800 --> 01:12:11.480 align:middle line:84% would actually require a different perspective, 01:12:11.480 --> 01:12:14.800 align:middle line:84% a different measure that takes into account what happened 01:12:14.800 --> 01:12:18.190 align:middle line:84% before and that we put those two together, those 01:12:18.190 --> 01:12:20.210 align:middle line:90% two closer than these pairs. 01:12:20.210 --> 01:12:21.970 align:middle line:90% How can I model these things? 01:12:21.970 --> 01:12:24.940 align:middle line:84% I can use a thing called autoregressive models. 01:12:24.940 --> 01:12:27.128 align:middle line:84% Autoregressive models-- it is very simple. 01:12:27.128 --> 01:12:28.670 align:middle line:84% There are a lot of way of doing this. 01:12:28.670 --> 01:12:31.735 align:middle line:84% This is just an example of how to take into account your past. 01:12:31.735 --> 01:12:34.790 align:middle line:90% 01:12:34.790 --> 01:12:35.690 align:middle line:90% How can I do this? 01:12:35.690 --> 01:12:39.280 align:middle line:84% Well, I have a time series of dependent observations. 01:12:39.280 --> 01:12:46.510 align:middle line:84% And what I can say is I assume that my observed 01:12:46.510 --> 01:12:53.080 align:middle line:84% point, at in this moment, is independent of the remote past, 01:12:53.080 --> 01:12:55.750 align:middle line:90% given its recent past, right? 01:12:55.750 --> 01:13:01.060 align:middle line:84% So to know that I'm here now, you 01:13:01.060 --> 01:13:05.200 align:middle line:84% don't really need to know where I was the day before yesterday. 01:13:05.200 --> 01:13:08.500 align:middle line:84% You need to know where I was 10 minutes ago, an hour ago, maybe 01:13:08.500 --> 01:13:09.910 align:middle line:90% two hours ago. 01:13:09.910 --> 01:13:13.660 align:middle line:84% But the predictive ability of two days ago, 01:13:13.660 --> 01:13:16.480 align:middle line:84% where I was five miles from here, 01:13:16.480 --> 01:13:18.700 align:middle line:90% is going to be very, very weak. 01:13:18.700 --> 01:13:21.670 align:middle line:84% So you can actually summarize your data, 01:13:21.670 --> 01:13:24.250 align:middle line:84% summarize your expectation on somebody 01:13:24.250 --> 01:13:27.310 align:middle line:84% being here by forgetting the remote past 01:13:27.310 --> 01:13:29.380 align:middle line:84% and considering only the recent past. 01:13:29.380 --> 01:13:31.630 align:middle line:84% The most recent could be one point. 01:13:31.630 --> 01:13:33.550 align:middle line:84% And, in this case, you can create a model 01:13:33.550 --> 01:13:37.480 align:middle line:84% like this in which you plot your present-- 01:13:37.480 --> 01:13:39.250 align:middle line:90% that is, my time now-- 01:13:39.250 --> 01:13:41.080 align:middle line:90% with your immediate past. 01:13:41.080 --> 01:13:43.930 align:middle line:84% And, in this case, you're assuming that everything-- 01:13:43.930 --> 01:13:47.320 align:middle line:84% my observation is independent of my past, 01:13:47.320 --> 01:13:51.020 align:middle line:84% given my most recent observation. 01:13:51.020 --> 01:13:54.940 align:middle line:84% This is the simplest autoregressive model. 01:13:54.940 --> 01:13:56.800 align:middle line:84% Now this kind of experiments, again, 01:13:56.800 --> 01:14:00.370 align:middle line:84% tell us something about the similarity of things. 01:14:00.370 --> 01:14:02.012 align:middle line:84% Actually this is a kind of analysis. 01:14:02.012 --> 01:14:03.220 align:middle line:90% The data are always the same. 01:14:03.220 --> 01:14:06.160 align:middle line:84% Once we have this temporal data, if we do some clustering, 01:14:06.160 --> 01:14:10.060 align:middle line:84% we may see that things are working in the same way 01:14:10.060 --> 01:14:14.410 align:middle line:84% a long time but are hardly going to tell us that something 01:14:14.410 --> 01:14:16.190 align:middle line:90% controls something else. 01:14:16.190 --> 01:14:18.970 align:middle line:84% In this case-- as I was saying in the beginning-- 01:14:18.970 --> 01:14:21.410 align:middle line:84% it's not really the data of the design of the experiment. 01:14:21.410 --> 01:14:23.360 align:middle line:84% It's the type of analysis you make. 01:14:23.360 --> 01:14:25.840 align:middle line:84% So if your interest is to find out 01:14:25.840 --> 01:14:29.920 align:middle line:84% which are functional clusters of genes that work together, 01:14:29.920 --> 01:14:32.720 align:middle line:84% well, clustering is your solution. 01:14:32.720 --> 01:14:35.230 align:middle line:84% But if you're interested in dissecting 01:14:35.230 --> 01:14:39.100 align:middle line:84% what is the regulation, the mechanism of regulation 01:14:39.100 --> 01:14:41.290 align:middle line:84% among genes, that will not tell you. 01:14:41.290 --> 01:14:44.470 align:middle line:84% I may have things that behave kind of similarly, 01:14:44.470 --> 01:14:48.970 align:middle line:84% but they not necessarily behave together equally. 01:14:48.970 --> 01:14:51.970 align:middle line:84% To be extreme, I would consider that something 01:14:51.970 --> 01:14:54.820 align:middle line:84% that controls something else will not really 01:14:54.820 --> 01:14:57.640 align:middle line:84% have exactly the same temporal behavior, right? 01:14:57.640 --> 01:14:59.820 align:middle line:84% So if I want to have you here today, 01:14:59.820 --> 01:15:03.270 align:middle line:84% I have to call you yesterday or I have to be here yesterday. 01:15:03.270 --> 01:15:05.610 align:middle line:84% I have to do something before you're here 01:15:05.610 --> 01:15:08.250 align:middle line:90% if I'm controlling you, right? 01:15:08.250 --> 01:15:14.960 align:middle line:84% So a way to use these things, to try 01:15:14.960 --> 01:15:17.570 align:middle line:84% to dissect this kind of information, if this 01:15:17.570 --> 01:15:21.590 align:middle line:84% is the question you have, is to use a thing called 01:15:21.590 --> 01:15:23.100 align:middle line:90% Bayesian networks. 01:15:23.100 --> 01:15:26.420 align:middle line:84% Bayesian networks are regulating genes-- 01:15:26.420 --> 01:15:29.720 align:middle line:84% in this case, relating variables in general-- 01:15:29.720 --> 01:15:34.820 align:middle line:84% by looking at how probable it is that one particular set 01:15:34.820 --> 01:15:38.480 align:middle line:84% of variables will control another set of variables. 01:15:38.480 --> 01:15:41.750 align:middle line:84% Originally, these things were built 01:15:41.750 --> 01:15:47.360 align:middle line:84% for humans, humans you want to clone information from, 01:15:47.360 --> 01:15:48.770 align:middle line:90% knowledge from. 01:15:48.770 --> 01:15:54.290 align:middle line:84% You are buying lunch or dinner to your physician friend, 01:15:54.290 --> 01:15:58.100 align:middle line:84% getting drunk, and distract the promise that will 01:15:58.100 --> 01:15:59.660 align:middle line:90% come to your lab the day after. 01:15:59.660 --> 01:16:02.270 align:middle line:84% And they will draw a network of this knowledge saying 01:16:02.270 --> 01:16:05.090 align:middle line:84% this gene versus other gene versus other gene 01:16:05.090 --> 01:16:09.890 align:middle line:84% and then add some probability that 01:16:09.890 --> 01:16:13.340 align:middle line:84% describe the function by which a particular gene controls 01:16:13.340 --> 01:16:14.120 align:middle line:90% another gene. 01:16:14.120 --> 01:16:17.430 align:middle line:90% This particular example-- 01:16:17.430 --> 01:16:17.930 align:middle line:90% I'm sorry. 01:16:17.930 --> 01:16:20.597 align:middle line:84% There are a couple of people who have seen this example at least 01:16:20.597 --> 01:16:22.190 align:middle line:90% 100 times. 01:16:22.190 --> 01:16:26.660 align:middle line:84% This is not about genes, just the intuition of what is there. 01:16:26.660 --> 01:16:30.890 align:middle line:84% This network tells you that your age your education 01:16:30.890 --> 01:16:34.260 align:middle line:90% affects your income. 01:16:34.260 --> 01:16:37.160 align:middle line:90% So this is easy to draw. 01:16:37.160 --> 01:16:41.480 align:middle line:84% The problem is how age and education affect your income? 01:16:41.480 --> 01:16:44.000 align:middle line:84% This is specified by that particular set 01:16:44.000 --> 01:16:44.840 align:middle line:90% of distributions. 01:16:44.840 --> 01:16:49.230 align:middle line:84% And those distributions tell you that if you are young 01:16:49.230 --> 01:16:52.310 align:middle line:84% and if you have a low education, your probability 01:16:52.310 --> 01:16:55.610 align:middle line:90% of having a low income is 0.9. 01:16:55.610 --> 01:17:01.040 align:middle line:84% And as you grow older and you get more educated, 01:17:01.040 --> 01:17:04.940 align:middle line:84% your probability of having a higher income increases. 01:17:04.940 --> 01:17:06.560 align:middle line:84% It's not one because you can always 01:17:06.560 --> 01:17:08.870 align:middle line:90% choose to be an academic. 01:17:08.870 --> 01:17:12.608 align:middle line:90% 01:17:12.608 --> 01:17:14.650 align:middle line:84% The problem is that we're not interested in doing 01:17:14.650 --> 01:17:15.670 align:middle line:90% these things by hand. 01:17:15.670 --> 01:17:19.240 align:middle line:84% We are interested in finding these things from data, right? 01:17:19.240 --> 01:17:21.310 align:middle line:84% And we can play exactly the same game 01:17:21.310 --> 01:17:23.290 align:middle line:84% we're playing with the clustering thing. 01:17:23.290 --> 01:17:30.910 align:middle line:84% We can find out what is the most probable set of nodes, which 01:17:30.910 --> 01:17:34.810 align:middle line:84% is the set of nodes that are most probable to control 01:17:34.810 --> 01:17:37.210 align:middle line:90% a particular gene. 01:17:37.210 --> 01:17:39.010 align:middle line:84% And we can do this for each gene. 01:17:39.010 --> 01:17:39.985 align:middle line:90% So the final picture-- 01:17:39.985 --> 01:17:42.790 align:middle line:90% 01:17:42.790 --> 01:17:44.050 align:middle line:90% lost it. 01:17:44.050 --> 01:17:49.280 align:middle line:84% Oh, the final picture of this is this one. 01:17:49.280 --> 01:17:49.780 align:middle line:90% Come on. 01:17:49.780 --> 01:17:50.530 align:middle line:90% Give me a picture. 01:17:50.530 --> 01:17:55.320 align:middle line:90% 01:17:55.320 --> 01:17:55.920 align:middle line:90% Here it is. 01:17:55.920 --> 01:17:58.810 align:middle line:90% 01:17:58.810 --> 01:18:02.890 align:middle line:84% Each ball represents a gene, except these three blue balls. 01:18:02.890 --> 01:18:07.660 align:middle line:90% 01:18:07.660 --> 01:18:11.650 align:middle line:84% OK, so these are about 40 patients, say, 01:18:11.650 --> 01:18:16.720 align:middle line:84% 41 patients, pediatric patients with leukemia. 01:18:16.720 --> 01:18:19.870 align:middle line:84% And for these patients, we have measured some phenotypes. 01:18:19.870 --> 01:18:22.720 align:middle line:84% But the most important thing we are interested in 01:18:22.720 --> 01:18:27.220 align:middle line:84% is the molecular classification, so the type of 01:18:27.220 --> 01:18:30.800 align:middle line:84% is called oncology status, oncogene status, 01:18:30.800 --> 01:18:34.160 align:middle line:84% which is the molecular classification of the tumor. 01:18:34.160 --> 01:18:39.140 align:middle line:90% And this is their survival. 01:18:39.140 --> 01:18:42.730 align:middle line:84% And this is [? finding, ?] it's how many days they've 01:18:42.730 --> 01:18:45.710 align:middle line:90% been in the hospital, OK? 01:18:45.710 --> 01:18:48.310 align:middle line:90% 01:18:48.310 --> 01:18:51.160 align:middle line:84% And what they're interested in is find out 01:18:51.160 --> 01:18:54.610 align:middle line:84% if there is a relationship between-- 01:18:54.610 --> 01:18:56.860 align:middle line:84% you remember when we were analyzing the other things 01:18:56.860 --> 01:18:58.330 align:middle line:90% into different conditions? 01:18:58.330 --> 01:19:01.960 align:middle line:84% We were doing one analysis for each different phenotype. 01:19:01.960 --> 01:19:04.990 align:middle line:84% We couldn't put the phenotypes together in a single picture. 01:19:04.990 --> 01:19:08.080 align:middle line:84% In this case, we can put the two phenotypes in a single picture 01:19:08.080 --> 01:19:10.840 align:middle line:84% and see, for instance, if there is any link that 01:19:10.840 --> 01:19:13.990 align:middle line:84% will go from oncogene to survival 01:19:13.990 --> 01:19:19.360 align:middle line:84% and how this process is mediated by other genes. 01:19:19.360 --> 01:19:24.130 align:middle line:84% And what we can also find out are dependencies among genes 01:19:24.130 --> 01:19:25.090 align:middle line:90% and other genes. 01:19:25.090 --> 01:19:27.370 align:middle line:84% And you see there are directions in those arrows. 01:19:27.370 --> 01:19:29.140 align:middle line:84% And those directions mean actually 01:19:29.140 --> 01:19:31.690 align:middle line:84% that one gene controls the other gene. 01:19:31.690 --> 01:19:33.640 align:middle line:84% Example I usually run is suppose we 01:19:33.640 --> 01:19:36.970 align:middle line:84% want to discover which of these flickers 01:19:36.970 --> 01:19:39.640 align:middle line:90% control these lights, right? 01:19:39.640 --> 01:19:41.890 align:middle line:90% So I can do it this way. 01:19:41.890 --> 01:19:43.060 align:middle line:90% I can change the flicker. 01:19:43.060 --> 01:19:46.600 align:middle line:90% I can change these things. 01:19:46.600 --> 01:19:50.300 align:middle line:84% And this will affect these lights to be on and off. 01:19:50.300 --> 01:19:52.810 align:middle line:84% But if I try to unscrew those lights, 01:19:52.810 --> 01:19:55.900 align:middle line:84% they will not change the state of this one, right? 01:19:55.900 --> 01:19:59.120 align:middle line:84% The metrics we use is very similar to this one. 01:19:59.120 --> 01:20:01.720 align:middle line:84% So the metrics will actually take into account 01:20:01.720 --> 01:20:03.700 align:middle line:84% the fact that you are measuring the influence 01:20:03.700 --> 01:20:08.500 align:middle line:84% of a directed influence from one gene to another gene. 01:20:08.500 --> 01:20:10.750 align:middle line:90% It's not just a simple distance. 01:20:10.750 --> 01:20:13.060 align:middle line:84% It's not just a pairwise measure. 01:20:13.060 --> 01:20:16.360 align:middle line:84% And actually, it's not pairwise because, as you can see, 01:20:16.360 --> 01:20:18.130 align:middle line:84% you can have more than one parent. 01:20:18.130 --> 01:20:22.690 align:middle line:84% This node here-- just to make an example-- 01:20:22.690 --> 01:20:30.730 align:middle line:84% is three parents, this one, this one, and this one. 01:20:30.730 --> 01:20:31.330 align:middle line:90% No, sorry. 01:20:31.330 --> 01:20:33.100 align:middle line:90% This is a child. 01:20:33.100 --> 01:20:35.680 align:middle line:90% And this is a parent. 01:20:35.680 --> 01:20:37.840 align:middle line:90% And this is another child. 01:20:37.840 --> 01:20:40.540 align:middle line:84% And he also has a grandchild, here. 01:20:40.540 --> 01:20:43.660 align:middle line:84% So you can actually use this kind of information 01:20:43.660 --> 01:20:50.830 align:middle line:84% to create a molecular landscape of the control 01:20:50.830 --> 01:20:53.335 align:middle line:90% mechanism of your things. 01:20:53.335 --> 01:20:56.390 align:middle line:90% 01:20:56.390 --> 01:20:58.520 align:middle line:84% And you remember what we were saying 01:20:58.520 --> 01:21:01.050 align:middle line:90% about how probable it is? 01:21:01.050 --> 01:21:05.090 align:middle line:84% I can actually measure how more probable is something 01:21:05.090 --> 01:21:08.240 align:middle line:84% to be affected by some variables than is 01:21:08.240 --> 01:21:10.100 align:middle line:84% to be affected by other variables 01:21:10.100 --> 01:21:12.830 align:middle line:84% through something that they I will not bother you with. 01:21:12.830 --> 01:21:15.200 align:middle line:84% But it's called basically base factor. 01:21:15.200 --> 01:21:17.570 align:middle line:84% Base factor is the ratio between the probability 01:21:17.570 --> 01:21:20.810 align:middle line:84% of two models, which tells you how more probable 01:21:20.810 --> 01:21:23.005 align:middle line:84% is one model compared to another. 01:21:23.005 --> 01:21:24.380 align:middle line:84% And these are the numbers we get. 01:21:24.380 --> 01:21:29.600 align:middle line:84% So we say that oncogene status, which had these three parents, 01:21:29.600 --> 01:21:32.210 align:middle line:90% we choose these three parents-- 01:21:32.210 --> 01:21:34.700 align:middle line:84% these are all the other possible combination 01:21:34.700 --> 01:21:37.350 align:middle line:90% of parents we have explored. 01:21:37.350 --> 01:21:40.680 align:middle line:84% And this picture tells you that the runner-up-- 01:21:40.680 --> 01:21:44.330 align:middle line:84% which is this other thing down there, the second one-- 01:21:44.330 --> 01:21:48.110 align:middle line:84% is seven times less probable than the top one 01:21:48.110 --> 01:21:51.470 align:middle line:84% to be responsible for the oncogene status. 01:21:51.470 --> 01:21:55.460 align:middle line:84% And the third one is going to be 56 times, 01:21:55.460 --> 01:21:58.610 align:middle line:84% and [INAUDIBLE] times, and down, down, down, down, down. 01:21:58.610 --> 01:22:01.340 align:middle line:84% And you see, basically, the runner-up, 01:22:01.340 --> 01:22:02.930 align:middle line:84% which tells you how more probable 01:22:02.930 --> 01:22:05.810 align:middle line:84% is the model you have compared to the best 01:22:05.810 --> 01:22:07.580 align:middle line:90% scenario of any other model. 01:22:07.580 --> 01:22:11.960 align:middle line:84% So it gives you some kind measure of confidence. 01:22:11.960 --> 01:22:14.678 align:middle line:84% OK, so this is an example how you can validate these things. 01:22:14.678 --> 01:22:16.220 align:middle line:84% You can actually do cross-validation. 01:22:16.220 --> 01:22:16.700 align:middle line:90% You remember? 01:22:16.700 --> 01:22:19.158 align:middle line:84% We were saying you pull out one case and make a prediction. 01:22:19.158 --> 01:22:22.340 align:middle line:90% And first validation is here. 01:22:22.340 --> 01:22:24.690 align:middle line:84% It was 100% and something like this. 01:22:24.690 --> 01:22:27.440 align:middle line:90% But the interesting thing is-- 01:22:27.440 --> 01:22:31.520 align:middle line:84% the take-home message for today and the thing that is important 01:22:31.520 --> 01:22:35.460 align:middle line:84% is that because there are no hypotheses here, 01:22:35.460 --> 01:22:38.840 align:middle line:84% the way you collect the data is important. 01:22:38.840 --> 01:22:41.540 align:middle line:84% But the way in which you analyze the data 01:22:41.540 --> 01:22:44.370 align:middle line:84% is the thing that is going to give you the answer. 01:22:44.370 --> 01:22:48.140 align:middle line:84% So if you are interested in mechanism of control, 01:22:48.140 --> 01:22:50.930 align:middle line:84% comparative analysis will tell you squat. 01:22:50.930 --> 01:22:53.480 align:middle line:84% If you're interested in molecular classification, 01:22:53.480 --> 01:22:54.902 align:middle line:84% clustering will tell you nothing. 01:22:54.902 --> 01:22:56.360 align:middle line:84% If you're interested in discovering 01:22:56.360 --> 01:23:02.120 align:middle line:84% new types of disease, these metrics will tell you nothing. 01:23:02.120 --> 01:23:07.220 align:middle line:84% Each type of analysis, as a particular type of answers-- 01:23:07.220 --> 01:23:09.210 align:middle line:90% is designed to answer them. 01:23:09.210 --> 01:23:11.660 align:middle line:84% And this is the most important thing you want to consider. 01:23:11.660 --> 01:23:15.790 align:middle line:90% 01:23:15.790 --> 01:23:17.380 align:middle line:90% There is a review up there. 01:23:17.380 --> 01:23:19.030 align:middle line:84% If you want to be bored to tears, 01:23:19.030 --> 01:23:21.700 align:middle line:84% then you can take it down from that website. 01:23:21.700 --> 01:23:27.190 align:middle line:84% But it was the state-of-the-art until six months ago, 01:23:27.190 --> 01:23:28.480 align:middle line:90% nothing has changed much. 01:23:28.480 --> 01:23:31.120 align:middle line:90% 01:23:31.120 --> 01:23:33.580 align:middle line:84% So the second is a [INAUDIBLE] book, 01:23:33.580 --> 01:23:40.060 align:middle line:84% which is part of your school equipment, right? 01:23:40.060 --> 01:23:42.880 align:middle line:84% Didn't you have to buy this book, yes, for the course? 01:23:42.880 --> 01:23:45.548 align:middle line:90% 01:23:45.548 --> 01:23:46.090 align:middle line:90% AUDIENCE: No. 01:23:46.090 --> 01:23:47.230 align:middle line:90% MARCO RAMONI: No? 01:23:47.230 --> 01:23:48.040 align:middle line:90% OK, go and-- 01:23:48.040 --> 01:23:48.400 align:middle line:90% AUDIENCE: Not that I know of. 01:23:48.400 --> 01:23:49.240 align:middle line:84% MARCO RAMONI: --see it because he's 01:23:49.240 --> 01:23:50.365 align:middle line:90% the director of the course. 01:23:50.365 --> 01:23:51.790 align:middle line:90% You may want to kiss some ass. 01:23:51.790 --> 01:23:55.270 align:middle line:90% 01:23:55.270 --> 01:23:57.760 align:middle line:90% Gene cluster and SAM are the-- 01:23:57.760 --> 01:23:59.980 align:middle line:84% you member The two non-parametric statistics 01:23:59.980 --> 01:24:01.300 align:middle line:90% I was describing before? 01:24:01.300 --> 01:24:05.860 align:middle line:84% Age is the thing that implements the Bayesian metrics 01:24:05.860 --> 01:24:07.660 align:middle line:90% and the temporal analysis. 01:24:07.660 --> 01:24:13.120 align:middle line:84% And what I'm going to do is send around an assignment, which 01:24:13.120 --> 01:24:15.040 align:middle line:84% will probably be a data [? study. ?] 01:24:15.040 --> 01:24:19.690 align:middle line:84% And you will do two different analysis for it. 01:24:19.690 --> 01:24:21.970 align:middle line:84% I don't remember if you have to do both of them 01:24:21.970 --> 01:24:24.760 align:middle line:84% or if you have to choose which one you want to do. 01:24:24.760 --> 01:24:27.580 align:middle line:84% And one is going to be a supervised analysis using 01:24:27.580 --> 01:24:31.360 align:middle line:84% either a gene cluster or SAM, two different statistics. 01:24:31.360 --> 01:24:35.110 align:middle line:84% And the other is going to be an unsupervised analysis using 01:24:35.110 --> 01:24:38.950 align:middle line:84% gene [? calssification. ?] Gene cluster is two components, one 01:24:38.950 --> 01:24:42.910 align:middle line:84% doing clustering and one doing a supervised differentiable 01:24:42.910 --> 01:24:45.200 align:middle line:90% analysis. 01:24:45.200 --> 01:24:48.450 align:middle line:90% OK, thank you. 01:24:48.450 --> 01:24:50.000 align:middle line:90%