WEBVTT 00:00:00.000 --> 00:00:01.292 align:middle line:90% ZOLTAN SZALLASI: -technologies. 00:00:01.292 --> 00:00:05.340 align:middle line:84% And this is kind of the introductory slide. 00:00:05.340 --> 00:00:08.700 align:middle line:90% You heard a lot from Zack. 00:00:08.700 --> 00:00:10.860 align:middle line:84% I think that was last time, like a week ago 00:00:10.860 --> 00:00:12.870 align:middle line:84% or two weeks ago about microarray technology. 00:00:12.870 --> 00:00:15.450 align:middle line:84% And I'm sure that he gave an extremely 00:00:15.450 --> 00:00:18.960 align:middle line:84% inspirational and enthusiastic talk about the possibilities 00:00:18.960 --> 00:00:20.220 align:middle line:90% and scope of this technology. 00:00:20.220 --> 00:00:22.053 align:middle line:84% But let me give you just a little this line. 00:00:22.053 --> 00:00:24.120 align:middle line:84% So whenever a new technology appears 00:00:24.120 --> 00:00:26.610 align:middle line:84% or for example, microarray technology or the Genome 00:00:26.610 --> 00:00:30.900 align:middle line:84% Project, of course first, there is a general ebullience 00:00:30.900 --> 00:00:33.180 align:middle line:84% and optimism that all problems will be solved 00:00:33.180 --> 00:00:34.410 align:middle line:90% within a couple of years. 00:00:34.410 --> 00:00:36.690 align:middle line:84% There are lots of reasons for this optimism. 00:00:36.690 --> 00:00:38.910 align:middle line:84% One is that you want investors in company 00:00:38.910 --> 00:00:40.470 align:middle line:90% and you want public funding. 00:00:40.470 --> 00:00:44.940 align:middle line:84% But of course, in a couple of months or years, 00:00:44.940 --> 00:00:48.670 align:middle line:84% realistic expectations start to appear. 00:00:48.670 --> 00:00:51.420 align:middle line:84% And then you have to start to think about limitations 00:00:51.420 --> 00:00:53.080 align:middle line:90% of the actual technology. 00:00:53.080 --> 00:00:55.710 align:middle line:84% And actually, that's the reason that we have 00:00:55.710 --> 00:00:58.480 align:middle line:90% a talk about this very topic. 00:00:58.480 --> 00:01:01.230 align:middle line:84% So when we talk about limitations of the technology, 00:01:01.230 --> 00:01:04.379 align:middle line:84% we have to define what you actually want to do in science. 00:01:04.379 --> 00:01:08.970 align:middle line:84% Of course, there are lots of various definitions of science. 00:01:08.970 --> 00:01:11.940 align:middle line:84% But in a sense, you'd like to make predictions 00:01:11.940 --> 00:01:13.450 align:middle line:90% about some of a system. 00:01:13.450 --> 00:01:17.070 align:middle line:84% And we are going to talk about limitations in those terms. 00:01:17.070 --> 00:01:20.950 align:middle line:84% How is it going to limit your predictive power? 00:01:20.950 --> 00:01:23.070 align:middle line:84% So when you talk about limitations, 00:01:23.070 --> 00:01:25.470 align:middle line:84% you can talk about how accurate your measurement is. 00:01:25.470 --> 00:01:27.120 align:middle line:84% There are limitations of the accuracy, 00:01:27.120 --> 00:01:29.190 align:middle line:84% of the measurements, accuracy and noise. 00:01:29.190 --> 00:01:31.322 align:middle line:84% But there are limitations in terms of sensitivity. 00:01:31.322 --> 00:01:32.280 align:middle line:90% What are you measuring. 00:01:32.280 --> 00:01:35.490 align:middle line:84% How complete is your measurement? 00:01:35.490 --> 00:01:39.210 align:middle line:84% And of course, I'm going to be very briefly touching on this, 00:01:39.210 --> 00:01:43.330 align:middle line:84% that even if you measure everything very accurately, 00:01:43.330 --> 00:01:45.810 align:middle line:84% there are inherent limitations in your predictive power. 00:01:45.810 --> 00:01:47.410 align:middle line:90% You cannot predict everything. 00:01:47.410 --> 00:01:51.063 align:middle line:84% Think about unpredictability in terms of chaos. 00:01:51.063 --> 00:01:52.980 align:middle line:84% Even if you measure everything very precisely, 00:01:52.980 --> 00:01:56.730 align:middle line:84% there are systems that you simply cannot predict how 00:01:56.730 --> 00:02:01.590 align:middle line:84% it's going to behave in a analytic way. 00:02:01.590 --> 00:02:04.290 align:middle line:84% So noise-- first, I would like to define what is noise 00:02:04.290 --> 00:02:05.490 align:middle line:90% and what is signal. 00:02:05.490 --> 00:02:08.550 align:middle line:84% Noise is an inherent feature of complex systems. 00:02:08.550 --> 00:02:11.009 align:middle line:84% And noise in continuous and discrete measurements, noise 00:02:11.009 --> 00:02:12.780 align:middle line:84% is the limitations of the technology. 00:02:12.780 --> 00:02:14.520 align:middle line:84% And of course, we need to talk about what 00:02:14.520 --> 00:02:15.910 align:middle line:90% can be done about noise. 00:02:15.910 --> 00:02:19.380 align:middle line:84% That's what statistics was invented for. 00:02:19.380 --> 00:02:23.280 align:middle line:84% And I'm going to talk briefly about normalization. 00:02:23.280 --> 00:02:24.570 align:middle line:90% So what is noise? 00:02:24.570 --> 00:02:27.300 align:middle line:90% There are different definitions. 00:02:27.300 --> 00:02:30.420 align:middle line:84% This, I, of course, took from Webster. 00:02:30.420 --> 00:02:33.810 align:middle line:84% And let's just look at this point 00:02:33.810 --> 00:02:36.300 align:middle line:84% D, which is "an irrelevant or meaningless data 00:02:36.300 --> 00:02:39.720 align:middle line:84% output, occurring along with desired information." 00:02:39.720 --> 00:02:43.230 align:middle line:84% Now you should be aware of that noise is not always 00:02:43.230 --> 00:02:43.960 align:middle line:90% a bad thing. 00:02:43.960 --> 00:02:47.553 align:middle line:84% Sometimes noise may turn out to be a very important signal. 00:02:47.553 --> 00:02:48.720 align:middle line:90% And what is your background? 00:02:48.720 --> 00:02:50.220 align:middle line:90% I'm just-- 00:02:50.220 --> 00:02:51.397 align:middle line:90% AUDIENCE: Biology. 00:02:51.397 --> 00:02:53.110 align:middle line:90% ZOLTAN SZALLASI: Biology. 00:02:53.110 --> 00:02:53.610 align:middle line:90% Sorry? 00:02:53.610 --> 00:02:54.075 align:middle line:90% AUDIENCE: Medicine. 00:02:54.075 --> 00:02:55.410 align:middle line:90% ZOLTAN SZALLASI: Medicine, OK-- 00:02:55.410 --> 00:02:57.720 align:middle line:84% so this is probably not the best example for you. 00:02:57.720 --> 00:03:01.050 align:middle line:84% But there were two guys, two radial astronomers, 00:03:01.050 --> 00:03:02.020 align:middle line:90% many, many years ago. 00:03:02.020 --> 00:03:03.720 align:middle line:84% I think that was almost 50 years ago, 00:03:03.720 --> 00:03:05.430 align:middle line:90% who were looking for signals. 00:03:05.430 --> 00:03:09.330 align:middle line:84% And they just saw this noise coming 00:03:09.330 --> 00:03:12.540 align:middle line:84% from every direction in the universe as radial astronomers. 00:03:12.540 --> 00:03:14.640 align:middle line:84% And that turned out to be the cosmic background 00:03:14.640 --> 00:03:17.110 align:middle line:84% radiation, which is one of the most important discoveries. 00:03:17.110 --> 00:03:18.540 align:middle line:84% And these guys that actually got the Nobel Prize 00:03:18.540 --> 00:03:19.590 align:middle line:90% a couple of years later. 00:03:19.590 --> 00:03:21.333 align:middle line:90% That started with pure noise. 00:03:21.333 --> 00:03:23.250 align:middle line:84% And they were trying to get rid of that noise. 00:03:23.250 --> 00:03:24.220 align:middle line:90% They couldn't. 00:03:24.220 --> 00:03:27.540 align:middle line:84% And that led to discovery of the cosmic background radiation. 00:03:27.540 --> 00:03:30.790 align:middle line:84% But if you think about medicine, then for example, 00:03:30.790 --> 00:03:35.730 align:middle line:84% the way cisplatin was discovered as a chemotherapeutic agent, 00:03:35.730 --> 00:03:38.910 align:middle line:84% what happened is actually that the electrodes that they 00:03:38.910 --> 00:03:41.700 align:middle line:84% were using in those experiments contained platinum. 00:03:41.700 --> 00:03:44.010 align:middle line:90% And they saw the effect of-- 00:03:44.010 --> 00:03:44.730 align:middle line:90% come in. 00:03:44.730 --> 00:03:46.742 align:middle line:84% Oh, you're going to hang out, OK. 00:03:46.742 --> 00:03:48.450 align:middle line:84% Actually, I just wanted to close the door 00:03:48.450 --> 00:03:51.510 align:middle line:90% because I know that's annoying. 00:03:51.510 --> 00:03:53.048 align:middle line:84% And then they tried to figure out 00:03:53.048 --> 00:03:55.590 align:middle line:84% what was killing the cells, what was slowing down the growth. 00:03:55.590 --> 00:03:57.257 align:middle line:84% And then they realized that actually, it 00:03:57.257 --> 00:04:00.917 align:middle line:84% was the platinum containing those electrodes. 00:04:00.917 --> 00:04:03.000 align:middle line:84% And this is the way they discovered this platinum. 00:04:03.000 --> 00:04:07.750 align:middle line:84% So the point is that noise is not always a bad thing. 00:04:07.750 --> 00:04:11.670 align:middle line:84% Now what if you see as noise or error in measurements, 00:04:11.670 --> 00:04:14.220 align:middle line:84% in biological measurements, might be a key component 00:04:14.220 --> 00:04:15.780 align:middle line:90% of biological processes. 00:04:15.780 --> 00:04:18.990 align:middle line:84% So of course, mutations in evolution 00:04:18.990 --> 00:04:20.529 align:middle line:90% are extremely important. 00:04:20.529 --> 00:04:24.450 align:middle line:84% And when we will talk about discrete measurements, one form 00:04:24.450 --> 00:04:26.310 align:middle line:84% of that is actually when you are sequencing, 00:04:26.310 --> 00:04:28.620 align:middle line:84% you'll see lots of noise in the human world, 00:04:28.620 --> 00:04:29.970 align:middle line:90% all sorts of genomes. 00:04:29.970 --> 00:04:31.360 align:middle line:90% That's called junk DNA. 00:04:31.360 --> 00:04:37.140 align:middle line:84% Well, we do not really know what this junk DNA is for. 00:04:37.140 --> 00:04:40.140 align:middle line:90% Then-- no, just go ahead. 00:04:40.140 --> 00:04:40.640 align:middle line:90% There. 00:04:40.640 --> 00:04:41.307 align:middle line:90% AUDIENCE: Sorry. 00:04:41.307 --> 00:04:43.173 align:middle line:90% 00:04:43.173 --> 00:04:45.090 align:middle line:84% ZOLTAN SZALLASI: I should have said, go ahead. 00:04:45.090 --> 00:04:47.850 align:middle line:90% 00:04:47.850 --> 00:04:49.180 align:middle line:90% It's still the introduction. 00:04:49.180 --> 00:04:53.400 align:middle line:84% So we do not really pursue this junk DNA 00:04:53.400 --> 00:04:55.950 align:middle line:84% when you are trying to find genes or exons of introns 00:04:55.950 --> 00:04:58.650 align:middle line:84% or transcription factor binding site. 00:04:58.650 --> 00:05:02.020 align:middle line:90% And that is going to extremely-- 00:05:02.020 --> 00:05:04.240 align:middle line:84% it's going to bother you very much when 00:05:04.240 --> 00:05:06.790 align:middle line:84% you're trying to find these true signals in the genome. 00:05:06.790 --> 00:05:08.800 align:middle line:84% We do not really know what junk DNA is for. 00:05:08.800 --> 00:05:11.320 align:middle line:84% But there might be a good reason it's there. 00:05:11.320 --> 00:05:13.690 align:middle line:84% It might be determining the spatial distance 00:05:13.690 --> 00:05:16.300 align:middle line:90% of different genes or whatnot. 00:05:16.300 --> 00:05:20.620 align:middle line:84% Another type of noise which seems to be very important 00:05:20.620 --> 00:05:23.200 align:middle line:90% is during differentiation. 00:05:23.200 --> 00:05:25.760 align:middle line:84% Very often, you see an asymmetric cell division. 00:05:25.760 --> 00:05:30.190 align:middle line:84% So RNA or proteins are divided or distributed between the two 00:05:30.190 --> 00:05:31.690 align:middle line:90% daughters cells asymmetrically. 00:05:31.690 --> 00:05:35.440 align:middle line:84% And that is actually done or that happens more or less 00:05:35.440 --> 00:05:36.220 align:middle line:90% by chance. 00:05:36.220 --> 00:05:39.530 align:middle line:84% And the two daughters are going to go one way or another, 00:05:39.530 --> 00:05:43.900 align:middle line:84% depending on how much RNA or protein they got. 00:05:43.900 --> 00:05:45.400 align:middle line:84% This, you can perceive as something 00:05:45.400 --> 00:05:49.120 align:middle line:84% of a noise measurement if you do a single cell measurement. 00:05:49.120 --> 00:05:51.820 align:middle line:84% And stochastic fluctuations may be very important 00:05:51.820 --> 00:05:56.440 align:middle line:84% for the stability of complex, physical-chemical systems. 00:05:56.440 --> 00:06:00.340 align:middle line:84% I might be talking about stochastic genetic networks 00:06:00.340 --> 00:06:03.070 align:middle line:84% and robustness much later in April when 00:06:03.070 --> 00:06:04.390 align:middle line:90% we're talking about modeling. 00:06:04.390 --> 00:06:10.600 align:middle line:84% And let's suffice now, that stochasticity and noise 00:06:10.600 --> 00:06:14.067 align:middle line:84% in complex systems might be a very important feature 00:06:14.067 --> 00:06:15.775 align:middle line:84% to maintain the stability of that system. 00:06:15.775 --> 00:06:18.310 align:middle line:90% 00:06:18.310 --> 00:06:21.010 align:middle line:84% You should be aware, of course, that genetic networks 00:06:21.010 --> 00:06:25.030 align:middle line:84% and biological systems or stochastic systems, 00:06:25.030 --> 00:06:28.570 align:middle line:84% because you know that, for example, you have only a couple 00:06:28.570 --> 00:06:30.190 align:middle line:84% of 100 copies of a given transcription 00:06:30.190 --> 00:06:33.070 align:middle line:84% factor per nucleus or even less, sometimes you have only 50 00:06:33.070 --> 00:06:34.340 align:middle line:90% per nucleus. 00:06:34.340 --> 00:06:36.790 align:middle line:84% The intracellular environment is not a free solution. 00:06:36.790 --> 00:06:38.457 align:middle line:84% And the reaction kinetics is often slow. 00:06:38.457 --> 00:06:40.832 align:middle line:84% And what it means is that if you have a stochastic system 00:06:40.832 --> 00:06:43.210 align:middle line:84% in this case, that if you have a completely deterministic 00:06:43.210 --> 00:06:47.290 align:middle line:84% system, then from any given gene expression found at any given 00:06:47.290 --> 00:06:51.430 align:middle line:84% state, you can go to one state, to one another state. 00:06:51.430 --> 00:06:53.090 align:middle line:90% That's a deterministic system. 00:06:53.090 --> 00:06:55.480 align:middle line:84% Whereas in a stochastic system, from any given state, 00:06:55.480 --> 00:06:57.910 align:middle line:84% you can go to different states with a bit 00:06:57.910 --> 00:06:59.210 align:middle line:90% of certain probability. 00:06:59.210 --> 00:07:01.970 align:middle line:84% So that's what we mean by stochastic systems. 00:07:01.970 --> 00:07:05.740 align:middle line:84% Now, if you have this, you have a stochastic system in biology. 00:07:05.740 --> 00:07:09.280 align:middle line:84% Then when you are measuring gene expression levels or protein 00:07:09.280 --> 00:07:11.560 align:middle line:84% expression levels or any activity 00:07:11.560 --> 00:07:13.720 align:middle line:84% of any biological parameters, you 00:07:13.720 --> 00:07:17.140 align:middle line:84% will perceive that as a noise in your measurement. 00:07:17.140 --> 00:07:18.220 align:middle line:90% Now is it true? 00:07:18.220 --> 00:07:20.680 align:middle line:84% Or is this really relevant to biology 00:07:20.680 --> 00:07:24.650 align:middle line:84% that you have stochasticity present in the system? 00:07:24.650 --> 00:07:27.790 align:middle line:84% And this was a paper that came out now, almost two years ago. 00:07:27.790 --> 00:07:29.650 align:middle line:84% But actually, they wanted to measure this. 00:07:29.650 --> 00:07:32.620 align:middle line:84% This was done in bacteria, but recently, similar studies 00:07:32.620 --> 00:07:34.300 align:middle line:90% were published in yeast as well. 00:07:34.300 --> 00:07:38.470 align:middle line:84% And what they did, they took two proteins from E coli. 00:07:38.470 --> 00:07:43.060 align:middle line:84% And they put on a green, two different GFP, two 00:07:43.060 --> 00:07:47.200 align:middle line:84% different green fluorescent proteins 00:07:47.200 --> 00:07:48.760 align:middle line:84% so they could measure the expression 00:07:48.760 --> 00:07:50.020 align:middle line:90% level of two proteins. 00:07:50.020 --> 00:07:52.150 align:middle line:84% And they set up the system in a way 00:07:52.150 --> 00:07:55.390 align:middle line:84% that if it was deterministic, then 00:07:55.390 --> 00:07:58.690 align:middle line:84% it was under the very same promoter. 00:07:58.690 --> 00:08:00.667 align:middle line:84% Both genes were expressed or driven 00:08:00.667 --> 00:08:01.750 align:middle line:90% by the very same promoter. 00:08:01.750 --> 00:08:03.740 align:middle line:84% And it was a very carefully setup experiment. 00:08:03.740 --> 00:08:05.380 align:middle line:84% So if the system was deterministic, 00:08:05.380 --> 00:08:09.520 align:middle line:84% then what they expected that the expression ratio of those two 00:08:09.520 --> 00:08:13.600 align:middle line:84% proteins is going to be the same in every single cell. 00:08:13.600 --> 00:08:16.120 align:middle line:84% Now what they found, that despite all their efforts 00:08:16.120 --> 00:08:18.670 align:middle line:84% to set up the experiment as perfect as possible, 00:08:18.670 --> 00:08:22.150 align:middle line:84% they found that two different colors. 00:08:22.150 --> 00:08:24.175 align:middle line:84% That means in this here, it's red and green. 00:08:24.175 --> 00:08:26.050 align:middle line:84% But of course, these are only false coloring. 00:08:26.050 --> 00:08:28.990 align:middle line:84% The point is that you have two different wavelengths where 00:08:28.990 --> 00:08:32.200 align:middle line:90% these signals are emitted. 00:08:32.200 --> 00:08:36.220 align:middle line:84% Depending on the actual cell, you 00:08:36.220 --> 00:08:41.020 align:middle line:84% can have very green, very red, and some yellow 00:08:41.020 --> 00:08:48.160 align:middle line:84% cells as well, which means that despite the careful way 00:08:48.160 --> 00:08:51.610 align:middle line:84% of this system of this experiment was set up, 00:08:51.610 --> 00:08:57.460 align:middle line:84% these cells expressed the two proteins at a different ratio. 00:08:57.460 --> 00:09:00.860 align:middle line:84% And that was due to stochasticity. 00:09:00.860 --> 00:09:07.300 align:middle line:84% So the point is that stochastic fluctuations 00:09:07.300 --> 00:09:10.430 align:middle line:90% occur in living organisms. 00:09:10.430 --> 00:09:12.940 align:middle line:84% They are trying to understand, now pretty hard, 00:09:12.940 --> 00:09:14.230 align:middle line:90% what's the relevance of this. 00:09:14.230 --> 00:09:16.990 align:middle line:84% It seems that it has a lot of relevance. 00:09:16.990 --> 00:09:21.354 align:middle line:84% But of course, we are not quite sure what the implications are. 00:09:21.354 --> 00:09:23.380 align:middle line:84% - You showld be aware of that whenever 00:09:23.380 --> 00:09:27.190 align:middle line:84% you do a biological measurement, or what especially these days 00:09:27.190 --> 00:09:29.290 align:middle line:84% technology might develop, you're always measuring 00:09:29.290 --> 00:09:30.890 align:middle line:90% population average data. 00:09:30.890 --> 00:09:33.520 align:middle line:84% So this again, is going to add to your noise. 00:09:33.520 --> 00:09:35.410 align:middle line:84% When you are measuring gene expression levels 00:09:35.410 --> 00:09:37.690 align:middle line:84% or do proteomics, you'll of course, 00:09:37.690 --> 00:09:42.340 align:middle line:84% grind down millions of cells or tens of thousands of cells. 00:09:42.340 --> 00:09:47.200 align:middle line:84% And that is going to be giving you a certain level of noise 00:09:47.200 --> 00:09:48.230 align:middle line:90% as well. 00:09:48.230 --> 00:09:49.930 align:middle line:84% And this is even true, this is true 00:09:49.930 --> 00:09:52.097 align:middle line:84% even if single cells are quantified. 00:09:52.097 --> 00:09:54.680 align:middle line:84% The reason for that is that if you have a stochastic network-- 00:09:54.680 --> 00:09:57.010 align:middle line:84% and let's imagine that you can really 00:09:57.010 --> 00:09:59.440 align:middle line:90% measure gene expression levels. 00:09:59.440 --> 00:10:01.630 align:middle line:84% You can do this for of course, individual proteins 00:10:01.630 --> 00:10:03.670 align:middle line:90% in single cells. 00:10:03.670 --> 00:10:05.800 align:middle line:84% But whenever you still do your measurement, 00:10:05.800 --> 00:10:07.660 align:middle line:84% you usually interfere with the cell. 00:10:07.660 --> 00:10:09.123 align:middle line:90% Or you kill the cell. 00:10:09.123 --> 00:10:10.540 align:middle line:84% So you don't really know how would 00:10:10.540 --> 00:10:12.430 align:middle line:90% that cell have progressed. 00:10:12.430 --> 00:10:14.140 align:middle line:84% So since you interfered with the system 00:10:14.140 --> 00:10:15.940 align:middle line:84% and still have the system, you can't really 00:10:15.940 --> 00:10:18.023 align:middle line:84% figure out what would have happened in the system. 00:10:18.023 --> 00:10:21.340 align:middle line:84% So you're actually going to end up with a population 00:10:21.340 --> 00:10:24.620 align:middle line:90% average data again. 00:10:24.620 --> 00:10:27.130 align:middle line:84% So there is no measurement without noise. 00:10:27.130 --> 00:10:30.017 align:middle line:84% As you know, it is usually the accuracy, sensitivity 00:10:30.017 --> 00:10:30.850 align:middle line:90% of your measurement. 00:10:30.850 --> 00:10:38.620 align:middle line:84% And I'm sure that most of you would be extremely troubled 00:10:38.620 --> 00:10:41.020 align:middle line:84% if you did a microarray measurement or some 00:10:41.020 --> 00:10:45.040 align:middle line:84% of a chemical or biological measurement, 00:10:45.040 --> 00:10:48.700 align:middle line:84% and in three triplicates, you would get exactly 00:10:48.700 --> 00:10:50.350 align:middle line:90% the very same number. 00:10:50.350 --> 00:10:53.290 align:middle line:84% That would mean probably, to most of you, 00:10:53.290 --> 00:10:56.020 align:middle line:84% that there is some of a systematic error 00:10:56.020 --> 00:10:59.050 align:middle line:84% in whatever you're doing at your photometer or whatnot. 00:10:59.050 --> 00:11:02.260 align:middle line:84% Because you would expect some of a spread 00:11:02.260 --> 00:11:04.400 align:middle line:90% of your continuous data. 00:11:04.400 --> 00:11:06.910 align:middle line:84% So it is expected for continuous variables 00:11:06.910 --> 00:11:08.650 align:middle line:84% to have data with a certain spread. 00:11:08.650 --> 00:11:10.330 align:middle line:90% And that's OK. 00:11:10.330 --> 00:11:12.580 align:middle line:84% And that's why statistics was invented. 00:11:12.580 --> 00:11:14.950 align:middle line:84% But you know that there is some of a true value 00:11:14.950 --> 00:11:16.180 align:middle line:90% of your measurement. 00:11:16.180 --> 00:11:20.440 align:middle line:84% But due to little fluctuations of whatnot, 00:11:20.440 --> 00:11:24.730 align:middle line:84% you have a certain spread around that through variable. 00:11:24.730 --> 00:11:27.280 align:middle line:84% And of course, usually the question is-- 00:11:27.280 --> 00:11:29.680 align:middle line:84% and that's what's the best statistics at this frequency 00:11:29.680 --> 00:11:31.330 align:middle line:84% statistics is really conserved with, 00:11:31.330 --> 00:11:35.093 align:middle line:84% that this is variable change due to a given treatment whenever 00:11:35.093 --> 00:11:36.260 align:middle line:90% you have a spread like that. 00:11:36.260 --> 00:11:39.920 align:middle line:84% So if you have measurement here or here or here 00:11:39.920 --> 00:11:43.030 align:middle line:84% and this is your starting point, then what's 00:11:43.030 --> 00:11:46.150 align:middle line:84% the probability that your parameter really 00:11:46.150 --> 00:11:48.260 align:middle line:90% change its value? 00:11:48.260 --> 00:11:51.080 align:middle line:84% So what you need to do for this is need to have of course, 00:11:51.080 --> 00:11:53.650 align:middle line:84% lots of measurements and/or a fairly good idea 00:11:53.650 --> 00:11:55.000 align:middle line:90% about the nature of the noise. 00:11:55.000 --> 00:11:57.560 align:middle line:90% That's very important as well. 00:11:57.560 --> 00:12:00.040 align:middle line:84% We are not going to get into this now. 00:12:00.040 --> 00:12:01.550 align:middle line:84% But as you know, that, for example, 00:12:01.550 --> 00:12:04.055 align:middle line:84% the easiest or the most convenient assumption 00:12:04.055 --> 00:12:05.680 align:middle line:84% is that you have a normal distribution. 00:12:05.680 --> 00:12:06.638 align:middle line:90% It's good to have that. 00:12:06.638 --> 00:12:08.320 align:middle line:84% Because if you have that, then you 00:12:08.320 --> 00:12:10.630 align:middle line:84% can actually make very simple calculations 00:12:10.630 --> 00:12:13.450 align:middle line:84% about the probability that whether your mean 00:12:13.450 --> 00:12:17.030 align:middle line:84% or whether your parameter is actually changed or not. 00:12:17.030 --> 00:12:21.520 align:middle line:84% So statistics was invented a long time ago. 00:12:21.520 --> 00:12:25.300 align:middle line:84% And actually partly, it was due to biological measurements. 00:12:25.300 --> 00:12:28.330 align:middle line:84% And so statistics is concerned in biology 00:12:28.330 --> 00:12:29.750 align:middle line:84% with many, many different issues. 00:12:29.750 --> 00:12:32.200 align:middle line:84% One of it is what is the true value of a given parameter 00:12:32.200 --> 00:12:34.510 align:middle line:90% if there is one true value? 00:12:34.510 --> 00:12:37.930 align:middle line:84% There is a very frequently used analysis 00:12:37.930 --> 00:12:40.510 align:middle line:84% that people or biologists are not really aware of, which 00:12:40.510 --> 00:12:42.640 align:middle line:90% is kind of a Bayesian analysis. 00:12:42.640 --> 00:12:46.930 align:middle line:84% And actually, this is the most frequently done 00:12:46.930 --> 00:12:49.120 align:middle line:84% statistical analysis in biology this 00:12:49.120 --> 00:12:50.960 align:middle line:90% is the way all science works. 00:12:50.960 --> 00:12:52.930 align:middle line:84% I have a similar belief about whether something 00:12:52.930 --> 00:12:54.940 align:middle line:84% is going to happen, or for example, an oncogene 00:12:54.940 --> 00:12:57.070 align:middle line:84% is going to transform a cell or not. 00:12:57.070 --> 00:12:58.600 align:middle line:90% I make a statement. 00:12:58.600 --> 00:13:01.570 align:middle line:84% And my job is actually, to convince 00:13:01.570 --> 00:13:05.500 align:middle line:84% you or other biologists that this is actually true. 00:13:05.500 --> 00:13:07.210 align:middle line:84% Now what you can do, what you usually do, 00:13:07.210 --> 00:13:08.793 align:middle line:84% is actually you repeat the experiment. 00:13:08.793 --> 00:13:11.590 align:middle line:84% And if you see the same phenomenon, 00:13:11.590 --> 00:13:15.058 align:middle line:84% then you are kind of updating your-- 00:13:15.058 --> 00:13:19.750 align:middle line:84% you are going to update our common belief that what I was 00:13:19.750 --> 00:13:22.270 align:middle line:90% saying that was actually true. 00:13:22.270 --> 00:13:24.490 align:middle line:84% So Bayesian statistics is always there 00:13:24.490 --> 00:13:28.060 align:middle line:90% in a hidden way in all biology. 00:13:28.060 --> 00:13:30.970 align:middle line:84% We are trying to update each other's belief 00:13:30.970 --> 00:13:34.870 align:middle line:90% network regarding biology. 00:13:34.870 --> 00:13:36.940 align:middle line:84% The third type of statistical analysis 00:13:36.940 --> 00:13:39.430 align:middle line:84% is that you don't really believe the measurements. 00:13:39.430 --> 00:13:42.670 align:middle line:84% But you know that there is some of systematic error there. 00:13:42.670 --> 00:13:46.150 align:middle line:84% And then you try to correct for this systematic error, 00:13:46.150 --> 00:13:47.860 align:middle line:84% and that's what's called normalization. 00:13:47.860 --> 00:13:50.590 align:middle line:84% And I'm going to talk about it in detail there. 00:13:50.590 --> 00:13:54.460 align:middle line:84% And there is a fourth issue in statistics 00:13:54.460 --> 00:13:58.148 align:middle line:84% when you actually producing a lot of, a lot of measurements. 00:13:58.148 --> 00:13:59.690 align:middle line:84% You are looking for certain patterns. 00:13:59.690 --> 00:14:02.560 align:middle line:84% Imagine that you're looking for gene expression changes that 00:14:02.560 --> 00:14:04.940 align:middle line:90% is going to cause cancer. 00:14:04.940 --> 00:14:08.320 align:middle line:84% And you have two populations of your samples, normal 00:14:08.320 --> 00:14:09.280 align:middle line:90% and cancer. 00:14:09.280 --> 00:14:11.680 align:middle line:84% And you see that certain type of genes 00:14:11.680 --> 00:14:14.650 align:middle line:84% are always downregulated or upregulated or mutated 00:14:14.650 --> 00:14:16.300 align:middle line:90% in cancer. 00:14:16.300 --> 00:14:20.020 align:middle line:90% Now this could happen by chance. 00:14:20.020 --> 00:14:25.690 align:middle line:84% If you do not have a large number of samples 00:14:25.690 --> 00:14:29.420 align:middle line:84% and this change is actually simply random-- 00:14:29.420 --> 00:14:31.390 align:middle line:90% so in certain cells, it's up. 00:14:31.390 --> 00:14:32.740 align:middle line:90% In other cells, it's down. 00:14:32.740 --> 00:14:36.280 align:middle line:84% Then if you have the wrong number of samples, 00:14:36.280 --> 00:14:38.410 align:middle line:84% then it might happen that just by chance, 00:14:38.410 --> 00:14:40.712 align:middle line:84% with a certain probability, in all normal samples, 00:14:40.712 --> 00:14:41.920 align:middle line:90% that mutation is not present. 00:14:41.920 --> 00:14:45.700 align:middle line:84% And in all disease samples or cancer samples, 00:14:45.700 --> 00:14:46.940 align:middle line:90% it's going to be present. 00:14:46.940 --> 00:14:51.280 align:middle line:84% So what you need to ask is that pattern, that 00:14:51.280 --> 00:14:54.640 align:middle line:84% would explain your biology be present by chance? 00:14:54.640 --> 00:14:56.782 align:middle line:84% And this kind of, there are too many numbers. 00:14:56.782 --> 00:14:58.240 align:middle line:84% And what you can do is actually you 00:14:58.240 --> 00:14:59.782 align:middle line:84% can try to solve analytically, that's 00:14:59.782 --> 00:15:01.160 align:middle line:90% what combinatorics is about. 00:15:01.160 --> 00:15:03.710 align:middle line:84% Or you can do some of permutation. 00:15:03.710 --> 00:15:06.500 align:middle line:84% And as you'll see later, that's actually a pretty nasty problem 00:15:06.500 --> 00:15:10.440 align:middle line:84% when you're trying to apply it for real-life problems. 00:15:10.440 --> 00:15:12.590 align:middle line:84% So biological measurements are often expensive. 00:15:12.590 --> 00:15:18.770 align:middle line:84% And something I'd like to point out to you, 00:15:18.770 --> 00:15:20.502 align:middle line:84% that if you follow the literature 00:15:20.502 --> 00:15:22.460 align:middle line:84% or when you will start to read the literature-- 00:15:22.460 --> 00:15:24.127 align:middle line:84% and I'm assuming you will because that's 00:15:24.127 --> 00:15:26.070 align:middle line:90% why you are taking this course. 00:15:26.070 --> 00:15:28.550 align:middle line:84% You going to see lots of nature science and high profile 00:15:28.550 --> 00:15:31.760 align:middle line:84% papers, in which they ran a single microarray 00:15:31.760 --> 00:15:35.780 align:middle line:84% measurement on a large number of different cancer samples. 00:15:35.780 --> 00:15:38.660 align:middle line:84% And then they are drawing all sorts of they 00:15:38.660 --> 00:15:41.810 align:middle line:84% are drawing all sorts of conclusions 00:15:41.810 --> 00:15:45.050 align:middle line:84% about which genes are important to cancer, which one is not. 00:15:45.050 --> 00:15:49.370 align:middle line:84% And these measurements have been still rather expensive. 00:15:49.370 --> 00:15:52.587 align:middle line:84% And it's not easy to come, not easy to obtain the samples. 00:15:52.587 --> 00:15:54.920 align:middle line:84% But you should be aware of that you cannot really do any 00:15:54.920 --> 00:15:55.748 align:middle line:90% statistics on that. 00:15:55.748 --> 00:15:58.040 align:middle line:84% You should do some sort of Bayesian type of statistics. 00:15:58.040 --> 00:16:01.277 align:middle line:84% But whatever they are doing on this is not really statistics. 00:16:01.277 --> 00:16:02.360 align:middle line:90% It's going to be Bayesian. 00:16:02.360 --> 00:16:04.880 align:middle line:84% I was going to say that I see this change very often 00:16:04.880 --> 00:16:06.320 align:middle line:84% and you either believe it or not. 00:16:06.320 --> 00:16:09.710 align:middle line:84% But you can't really get any hardcore numbers out of it 00:16:09.710 --> 00:16:12.350 align:middle line:84% that you can use for any statistical analysis 00:16:12.350 --> 00:16:14.150 align:middle line:90% or modeling. 00:16:14.150 --> 00:16:17.000 align:middle line:84% So reliable numbers cannot be produced without replicates, 00:16:17.000 --> 00:16:19.010 align:middle line:90% which is kind of obvious. 00:16:19.010 --> 00:16:21.650 align:middle line:84% So the central problem is that in massive biological 00:16:21.650 --> 00:16:23.872 align:middle line:84% measurements, quantitative and qualitative calls 00:16:23.872 --> 00:16:25.580 align:middle line:84% are supposed to be made on a large number 00:16:25.580 --> 00:16:28.940 align:middle line:84% of heterogeneous variables, using only a few replicates. 00:16:28.940 --> 00:16:31.710 align:middle line:84% That's what you're going to see over and over again 00:16:31.710 --> 00:16:35.960 align:middle line:84% if you work on a large scale or massively parallel biology. 00:16:35.960 --> 00:16:39.530 align:middle line:84% And this is one of the problems that technology 00:16:39.530 --> 00:16:42.080 align:middle line:84% and the analysis has to overcome. 00:16:42.080 --> 00:16:43.820 align:middle line:84% So where is the noise coming from 00:16:43.820 --> 00:16:45.090 align:middle line:90% in microarray measurements? 00:16:45.090 --> 00:16:49.460 align:middle line:84% So this is a slide, I think you saw some of variation 00:16:49.460 --> 00:16:51.290 align:middle line:90% of this in Zack's talk. 00:16:51.290 --> 00:16:56.900 align:middle line:84% So this is how an affymetrix DNA microarray chip works. 00:16:56.900 --> 00:17:01.730 align:middle line:84% So you start with tissue, and you extract RNA. 00:17:01.730 --> 00:17:07.609 align:middle line:84% And what you do is have to do an RT, 00:17:07.609 --> 00:17:10.670 align:middle line:84% reverse transcriptase treatment or step 00:17:10.670 --> 00:17:15.200 align:middle line:84% on it that is going to translate back the RNA into cDNA. 00:17:15.200 --> 00:17:16.730 align:middle line:84% And depending on how you do it, you 00:17:16.730 --> 00:17:19.069 align:middle line:90% can either produce cDNA or CRNA. 00:17:19.069 --> 00:17:21.109 align:middle line:84% Because during this process when you 00:17:21.109 --> 00:17:24.230 align:middle line:84% are producing the cDNA or the CRNA, 00:17:24.230 --> 00:17:30.150 align:middle line:84% fluorescence dyes will be incorporated into the polymers. 00:17:30.150 --> 00:17:32.840 align:middle line:84% And then these are going to be hybridized 00:17:32.840 --> 00:17:35.910 align:middle line:84% to the specific probes present on the chip. 00:17:35.910 --> 00:17:39.140 align:middle line:84% Now the underlying assumption or expectation 00:17:39.140 --> 00:17:42.170 align:middle line:84% is that ideally, one copy of a given RNA 00:17:42.170 --> 00:17:44.810 align:middle line:84% will produce one unit of a specific signal. 00:17:44.810 --> 00:17:49.100 align:middle line:84% If this were true, then you would have 00:17:49.100 --> 00:17:51.620 align:middle line:90% very accurate measurements. 00:17:51.620 --> 00:17:55.460 align:middle line:84% But now let's see what's actually happening in reality. 00:17:55.460 --> 00:17:59.450 align:middle line:84% When cDNA is produced from the RNA using the RT, 00:17:59.450 --> 00:18:02.060 align:middle line:84% this is an enzyme that has its own life, 00:18:02.060 --> 00:18:03.840 align:middle line:90% its own characteristics. 00:18:03.840 --> 00:18:07.430 align:middle line:84% So the initiation of the RT step is stochastic. 00:18:07.430 --> 00:18:10.020 align:middle line:84% Because as I'm sure you're aware of that, 00:18:10.020 --> 00:18:12.350 align:middle line:84% you need a starting primer that's 00:18:12.350 --> 00:18:15.230 align:middle line:84% going to be extended by the reverse transcriptase. 00:18:15.230 --> 00:18:17.540 align:middle line:84% And very often, the reverse transcriptase, 00:18:17.540 --> 00:18:19.377 align:middle line:90% the enzyme simply drops off. 00:18:19.377 --> 00:18:20.960 align:middle line:84% So that's why what you see is actually 00:18:20.960 --> 00:18:22.790 align:middle line:84% when you do a microarray measurement, 00:18:22.790 --> 00:18:25.820 align:middle line:84% you see usually a much stronger signal the three coming 00:18:25.820 --> 00:18:28.970 align:middle line:84% from the 3-prime end of the gene than from the 6-prime end. 00:18:28.970 --> 00:18:32.750 align:middle line:84% Because as you're transcribing and reverse transcribing 00:18:32.750 --> 00:18:36.210 align:middle line:84% the message, the RT starts to fall off. 00:18:36.210 --> 00:18:38.150 align:middle line:84% So you always have much stronger signal 00:18:38.150 --> 00:18:40.610 align:middle line:84% from where the RT is started, which is always the RT 00:18:40.610 --> 00:18:41.780 align:middle line:90% because that's where-- 00:18:41.780 --> 00:18:43.445 align:middle line:84% what you usually use is a poly(A). 00:18:43.445 --> 00:18:46.040 align:middle line:84% You're using the poly(A)-tag as your initiator. 00:18:46.040 --> 00:18:47.660 align:middle line:84% You can use random primers as well. 00:18:47.660 --> 00:18:50.090 align:middle line:84% And actually sometimes, it's used for some in bacteria. 00:18:50.090 --> 00:18:53.360 align:middle line:84% But most of the time you start with poly(A). 00:18:53.360 --> 00:18:57.950 align:middle line:84% Also, cRNA, which is used for the affymetrix chip, 00:18:57.950 --> 00:19:00.650 align:middle line:84% is produced in the presence of fluorescent dyes. 00:19:00.650 --> 00:19:03.470 align:middle line:84% And it's assumed that the dye incorporation. 00:19:03.470 --> 00:19:06.050 align:middle line:84% Or it was hoped or it is hoped the dye incorporation is 00:19:06.050 --> 00:19:07.560 align:middle line:90% going to be linear. 00:19:07.560 --> 00:19:09.585 align:middle line:84% And it's going to be incorporated 00:19:09.585 --> 00:19:10.460 align:middle line:90% in equal probability. 00:19:10.460 --> 00:19:12.980 align:middle line:90% But it's not the case. 00:19:12.980 --> 00:19:14.720 align:middle line:84% The cRNA production is not linear. 00:19:14.720 --> 00:19:18.980 align:middle line:84% There are messages that are transcribed into cRNA 00:19:18.980 --> 00:19:21.230 align:middle line:84% with a much higher probability, much higher efficiency 00:19:21.230 --> 00:19:21.920 align:middle line:90% than others. 00:19:21.920 --> 00:19:26.090 align:middle line:84% And the dye incorporation is not linear either. 00:19:26.090 --> 00:19:30.380 align:middle line:84% Also the affymetrix chip's involves a step, 00:19:30.380 --> 00:19:32.900 align:middle line:84% and you actually break down your cRNA. 00:19:32.900 --> 00:19:35.030 align:middle line:84% For whatever reason, this is the chip design. 00:19:35.030 --> 00:19:38.210 align:middle line:84% And breaking down the cRNA into small pieces 00:19:38.210 --> 00:19:41.305 align:middle line:84% is not going to be the same for all messages either. 00:19:41.305 --> 00:19:42.680 align:middle line:84% And of course, you have all sorts 00:19:42.680 --> 00:19:44.180 align:middle line:84% of other problems like hybridization 00:19:44.180 --> 00:19:46.430 align:middle line:90% or cross-hybridization. 00:19:46.430 --> 00:19:48.380 align:middle line:84% And one can go on and on and on what 00:19:48.380 --> 00:19:50.630 align:middle line:84% would give you the noise is just a couple of samples. 00:19:50.630 --> 00:19:53.000 align:middle line:84% But the point is that your final signal 00:19:53.000 --> 00:19:55.520 align:middle line:84% is going to be the sum of all the 00:19:55.520 --> 00:19:58.410 align:middle line:84% above, or all these things and others. 00:19:58.410 --> 00:20:00.270 align:middle line:84% So this is just the give you a feel 00:20:00.270 --> 00:20:04.470 align:middle line:84% for how many individual issues will arise when you're 00:20:04.470 --> 00:20:05.610 align:middle line:90% doing a micro measurement. 00:20:05.610 --> 00:20:07.735 align:middle line:84% Of course, the surface chemistry is very important, 00:20:07.735 --> 00:20:10.180 align:middle line:84% the background subtraction, and so forth. 00:20:10.180 --> 00:20:12.100 align:middle line:90% So let's see another example. 00:20:12.100 --> 00:20:14.190 align:middle line:84% This is the two-color microarray. 00:20:14.190 --> 00:20:16.080 align:middle line:84% The previous one with the affymetrix chip, 00:20:16.080 --> 00:20:19.440 align:middle line:84% you heard about last time, in which you are actually putting 00:20:19.440 --> 00:20:21.360 align:middle line:90% out a single cRNA per chip. 00:20:21.360 --> 00:20:24.090 align:middle line:84% And there was another competing technology 00:20:24.090 --> 00:20:25.680 align:middle line:90% invented at the same time. 00:20:25.680 --> 00:20:28.737 align:middle line:84% When you actually label cDNA of two different samples, 00:20:28.737 --> 00:20:30.570 align:middle line:84% you measure two samples, and you're actually 00:20:30.570 --> 00:20:34.420 align:middle line:84% measuring the ratio of for each individual gene. 00:20:34.420 --> 00:20:38.130 align:middle line:84% So in this case, what you do is have equal amounts 00:20:38.130 --> 00:20:39.690 align:middle line:90% of labeled cDNA samples. 00:20:39.690 --> 00:20:42.660 align:middle line:84% And what you're hoping for, what you're trying to achieve 00:20:42.660 --> 00:20:46.920 align:middle line:84% is that if a certain message is present at the same level 00:20:46.920 --> 00:20:50.430 align:middle line:84% in both samples, then the two intensity, 00:20:50.430 --> 00:20:52.210 align:middle line:84% the signals are going to be equal. 00:20:52.210 --> 00:20:55.200 align:middle line:84% So you're going to have a kind of red, yellow spot 00:20:55.200 --> 00:20:57.720 align:middle line:84% if a gene is overexpressed or underexpressed, 00:20:57.720 --> 00:21:00.940 align:middle line:84% you're going to have a stronger red or green color. 00:21:00.940 --> 00:21:03.870 align:middle line:84% Now what you are ending up with in these measurements 00:21:03.870 --> 00:21:04.885 align:middle line:90% is a ratio. 00:21:04.885 --> 00:21:06.510 align:middle line:84% And the problem is that actually, there 00:21:06.510 --> 00:21:10.000 align:middle line:90% is no truly blank spot. 00:21:10.000 --> 00:21:14.850 align:middle line:84% You always have some sort of a background noise there. 00:21:14.850 --> 00:21:17.130 align:middle line:90% And you are measuring the ratio. 00:21:17.130 --> 00:21:19.320 align:middle line:84% Then of course, that non-blank spot 00:21:19.320 --> 00:21:22.900 align:middle line:84% is going to give you some sort of a false pseudo signal. 00:21:22.900 --> 00:21:24.360 align:middle line:84% So if you're for example, if there 00:21:24.360 --> 00:21:29.935 align:middle line:84% is a gene that is not present at all in a given sample 00:21:29.935 --> 00:21:31.560 align:middle line:84% and it's expressed in the other sample, 00:21:31.560 --> 00:21:35.430 align:middle line:84% then the ratio would be of course, infinitely high. 00:21:35.430 --> 00:21:37.290 align:middle line:84% Or it would be very, very, very high. 00:21:37.290 --> 00:21:38.700 align:middle line:90% But you never see this. 00:21:38.700 --> 00:21:42.420 align:middle line:84% You always have, since you have a certain background intensity, 00:21:42.420 --> 00:21:44.100 align:middle line:84% what you see is some sort of, let's say, 00:21:44.100 --> 00:21:49.650 align:middle line:84% 100-fold upregulation, which in fact, or in true, in truth 00:21:49.650 --> 00:21:51.930 align:middle line:84% might be a complete downregulation 00:21:51.930 --> 00:21:57.750 align:middle line:84% or a complete lack of that gene in one of the samples. 00:21:57.750 --> 00:22:00.960 align:middle line:84% So this is perceived by the experimenter 00:22:00.960 --> 00:22:04.480 align:middle line:90% as compressing the signals. 00:22:04.480 --> 00:22:09.810 align:middle line:84% So you have a very wide dynamics of the ratios, 00:22:09.810 --> 00:22:12.565 align:middle line:84% from minus infinity to plus infinity, 00:22:12.565 --> 00:22:14.940 align:middle line:84% but what you actually you see-- and this is very usually, 00:22:14.940 --> 00:22:17.490 align:middle line:84% most of these measurements are, the ratios are cut-- 00:22:17.490 --> 00:22:21.975 align:middle line:84% is 100-fold up- or down-regulation on either side. 00:22:21.975 --> 00:22:24.490 align:middle line:90% 00:22:24.490 --> 00:22:27.240 align:middle line:84% There are lots of experimental issues 00:22:27.240 --> 00:22:34.240 align:middle line:84% that can also contribute to the [INAUDIBLE] noise. 00:22:34.240 --> 00:22:36.270 align:middle line:84% So this is how the Affymetrix chip is designed. 00:22:36.270 --> 00:22:37.775 align:middle line:90% You have seen this before. 00:22:37.775 --> 00:22:39.900 align:middle line:84% And what you have, so these are very short probes-- 00:22:39.900 --> 00:22:41.550 align:middle line:90% 25 base pair probes. 00:22:41.550 --> 00:22:44.220 align:middle line:84% The way Affymetrix tried to overcome this problem 00:22:44.220 --> 00:22:48.150 align:middle line:84% is that they designed a set of probes 00:22:48.150 --> 00:22:52.780 align:middle line:84% along a given gene using some sort of an algorithm. 00:22:52.780 --> 00:22:55.200 align:middle line:84% And what they hoped for that, if you 00:22:55.200 --> 00:22:57.180 align:middle line:84% have lots, and lots, and lots of probes-- 00:22:57.180 --> 00:23:02.820 align:middle line:90% 11 or 16 probes per gene-- 00:23:02.820 --> 00:23:06.300 align:middle line:84% then from this set of probes, you 00:23:06.300 --> 00:23:09.848 align:middle line:84% can somehow estimate the true gene expression level. 00:23:09.848 --> 00:23:11.640 align:middle line:84% So this is how they are actually designing. 00:23:11.640 --> 00:23:13.410 align:middle line:84% As you see here, this is the entire gene, 00:23:13.410 --> 00:23:18.990 align:middle line:84% and you are tiling the gene across unique enough sequence 00:23:18.990 --> 00:23:20.970 align:middle line:90% regions of the genes. 00:23:20.970 --> 00:23:23.820 align:middle line:84% Now, the problem here is coming from the real measurement, 00:23:23.820 --> 00:23:29.850 align:middle line:84% that these up here are the perfect mesh probes that are 00:23:29.850 --> 00:23:34.890 align:middle line:90% supposed to measure the same g. 00:23:34.890 --> 00:23:37.560 align:middle line:84% To some extent, you would expect that all these expression 00:23:37.560 --> 00:23:40.590 align:middle line:84% levels, all these intensities, would be equal. 00:23:40.590 --> 00:23:44.585 align:middle line:84% And very often, for most genes, for most purposes, 00:23:44.585 --> 00:23:45.460 align:middle line:90% this is not the case. 00:23:45.460 --> 00:23:47.440 align:middle line:84% You have very bright and very dark probes. 00:23:47.440 --> 00:23:50.920 align:middle line:84% There are lots of reasons for this. 00:23:50.920 --> 00:23:52.830 align:middle line:84% You have cRNA secondary structure, 00:23:52.830 --> 00:23:54.130 align:middle line:90% and so on, and so forth. 00:23:54.130 --> 00:23:56.520 align:middle line:84% But the point is that when you look a little bit harder, 00:23:56.520 --> 00:23:58.290 align:middle line:84% deeper, into what you are actually 00:23:58.290 --> 00:24:01.200 align:middle line:84% getting from these measurements, well, 00:24:01.200 --> 00:24:03.390 align:middle line:84% you're expected to estimate the true gene expression 00:24:03.390 --> 00:24:07.350 align:middle line:84% level from this set of intensities 00:24:07.350 --> 00:24:12.490 align:middle line:84% that can often vary by four or five orders of magnitude. 00:24:12.490 --> 00:24:15.738 align:middle line:84% So that's how reality works, these experiments. 00:24:15.738 --> 00:24:18.030 align:middle line:84% So this is just another additional piece of information 00:24:18.030 --> 00:24:20.910 align:middle line:84% that it's not that easy to design this. 00:24:20.910 --> 00:24:22.740 align:middle line:84% But, of course, one can improve a lot, 00:24:22.740 --> 00:24:24.780 align:middle line:84% and the gentleman sitting here could tell you 00:24:24.780 --> 00:24:27.720 align:middle line:84% lots of interesting stories about how these things are 00:24:27.720 --> 00:24:31.800 align:middle line:84% designed or how they are not designed by the manufacturer. 00:24:31.800 --> 00:24:33.420 align:middle line:90% But that's the story. 00:24:33.420 --> 00:24:36.130 align:middle line:90% 00:24:36.130 --> 00:24:39.910 align:middle line:84% So I was just trying to give you a couple of thoughts, 00:24:39.910 --> 00:24:42.490 align:middle line:84% a couple of pieces of data, about where 00:24:42.490 --> 00:24:44.560 align:middle line:84% noise is coming from in Affymetrix 00:24:44.560 --> 00:24:46.330 align:middle line:90% measurements in real life. 00:24:46.330 --> 00:24:51.790 align:middle line:84% But even if you had very good quality measurements, 00:24:51.790 --> 00:24:56.450 align:middle line:84% you have other conceptual issues in this field as well. 00:24:56.450 --> 00:25:00.130 align:middle line:84% So let's assume that you want to use your numbers to reverse 00:25:00.130 --> 00:25:02.350 align:middle line:84% engineer a system or to do forward 00:25:02.350 --> 00:25:04.030 align:middle line:84% modeling, more forward simulation, 00:25:04.030 --> 00:25:05.230 align:middle line:90% large genetic networks. 00:25:05.230 --> 00:25:07.840 align:middle line:84% But you'd like to have very good quality numbers. 00:25:07.840 --> 00:25:12.040 align:middle line:84% The problem is that when you do these measurements, you always 00:25:12.040 --> 00:25:14.410 align:middle line:84% measure a very heterogeneous solution, a very 00:25:14.410 --> 00:25:18.290 align:middle line:84% heterogeneous population of RNA and proteins. 00:25:18.290 --> 00:25:20.890 align:middle line:84% Now, when you started this measurement, how are you going 00:25:20.890 --> 00:25:22.900 align:middle line:90% to normalize your numbers? 00:25:22.900 --> 00:25:26.590 align:middle line:84% So how do you express your measurements, 00:25:26.590 --> 00:25:29.020 align:middle line:84% even if your measurement, your technology is very good? 00:25:29.020 --> 00:25:30.490 align:middle line:90% Per unit RNA? 00:25:30.490 --> 00:25:32.283 align:middle line:90% Per microgram RNA? 00:25:32.283 --> 00:25:34.450 align:middle line:84% The problem with this is that if you have a decrease 00:25:34.450 --> 00:25:35.950 align:middle line:90% in the level of a given gene-- 00:25:35.950 --> 00:25:38.140 align:middle line:84% and some genes are very highly expressed-- 00:25:38.140 --> 00:25:42.610 align:middle line:84% and then the message of other genes, 00:25:42.610 --> 00:25:44.860 align:middle line:84% unavoidably, is the relative increase 00:25:44.860 --> 00:25:46.480 align:middle line:90% of the level of other messages. 00:25:46.480 --> 00:25:49.510 align:middle line:84% Because let's say you have a million copies, or let's 00:25:49.510 --> 00:25:51.920 align:middle line:84% say 10 million copies for RNA per single cell. 00:25:51.920 --> 00:25:54.430 align:middle line:84% So if a highly expressed gene is downregulated, 00:25:54.430 --> 00:25:56.410 align:middle line:84% then what you perceive in your measurement is-- 00:25:56.410 --> 00:25:59.800 align:middle line:84% unless you're actually trying to normalize for the actual copy 00:25:59.800 --> 00:26:00.340 align:middle line:90% numbers-- 00:26:00.340 --> 00:26:03.130 align:middle line:84% that some of the genes are slightly upregulated. 00:26:03.130 --> 00:26:04.630 align:middle line:84% So there are other conceptual issues 00:26:04.630 --> 00:26:08.650 align:middle line:84% as well why you will have noise in your measurement. 00:26:08.650 --> 00:26:10.780 align:middle line:84% But as I mentioned, the real problem 00:26:10.780 --> 00:26:14.920 align:middle line:84% is, the real issue is the actual technology. 00:26:14.920 --> 00:26:17.740 align:middle line:84% Now, what you can do with that is, 00:26:17.740 --> 00:26:20.860 align:middle line:84% when you have a set of measurements, 00:26:20.860 --> 00:26:24.070 align:middle line:84% you want to take a good, hard look at your data 00:26:24.070 --> 00:26:27.130 align:middle line:84% to see whether you have some sort of a systematic error 00:26:27.130 --> 00:26:28.270 align:middle line:90% in measurement. 00:26:28.270 --> 00:26:31.840 align:middle line:84% These are a bunch of Affymetrix measurements-- real life, 00:26:31.840 --> 00:26:34.030 align:middle line:84% real measurements-- in which what you see 00:26:34.030 --> 00:26:37.295 align:middle line:84% is the intensity distribution of all probe sets. 00:26:37.295 --> 00:26:39.670 align:middle line:84% So what you have here is the measurement, gene expression 00:26:39.670 --> 00:26:45.160 align:middle line:84% measurements on about 10, 11,000 different genes, 00:26:45.160 --> 00:26:47.540 align:middle line:84% all covered by a different probe set. 00:26:47.540 --> 00:26:51.770 align:middle line:84% And this is what you see as a distribution. 00:26:51.770 --> 00:26:54.430 align:middle line:84% Now, what you see here is there is 00:26:54.430 --> 00:26:57.290 align:middle line:84% one measurement that's very strongly an outlier, 00:26:57.290 --> 00:26:58.550 align:middle line:90% and some other as well. 00:26:58.550 --> 00:27:00.577 align:middle line:90% These are pretty much the same. 00:27:00.577 --> 00:27:03.160 align:middle line:84% And imagine that you're actually running the very same sample. 00:27:03.160 --> 00:27:05.170 align:middle line:84% Let's say you have a single cell line, 00:27:05.170 --> 00:27:07.040 align:middle line:84% and you're treating it with different drugs. 00:27:07.040 --> 00:27:09.490 align:middle line:84% Now, what you'd expect is essentially 00:27:09.490 --> 00:27:12.700 align:middle line:84% the very same distribution for each of these RNA samples 00:27:12.700 --> 00:27:16.480 align:middle line:84% with a few differences, a few variations. 00:27:16.480 --> 00:27:18.290 align:middle line:90% And you have this guy here. 00:27:18.290 --> 00:27:19.780 align:middle line:90% So what you can assume-- 00:27:19.780 --> 00:27:23.020 align:middle line:84% and this is actually what people do and the Affymetrix algorithm 00:27:23.020 --> 00:27:23.710 align:middle line:90% does-- 00:27:23.710 --> 00:27:27.940 align:middle line:84% that for some reason during this measurement, 00:27:27.940 --> 00:27:29.980 align:middle line:84% the fluorescent dye incorporation 00:27:29.980 --> 00:27:33.490 align:middle line:84% wasn't as efficient, or your fluorescent reader 00:27:33.490 --> 00:27:36.460 align:middle line:84% was miscalibrated, or something else, 00:27:36.460 --> 00:27:39.890 align:middle line:90% but a systematic error occurred. 00:27:39.890 --> 00:27:42.760 align:middle line:84% So what you assume, then, is that the distribution 00:27:42.760 --> 00:27:45.980 align:middle line:84% for all these measures is actually the same. 00:27:45.980 --> 00:27:49.540 align:middle line:84% So what you can do is can start to shift your curves, 00:27:49.540 --> 00:27:52.000 align:middle line:84% because you have a good reason to assume 00:27:52.000 --> 00:27:54.832 align:middle line:84% that these are actually all very similar distributions. 00:27:54.832 --> 00:27:56.290 align:middle line:84% So what you can do is actually take 00:27:56.290 --> 00:27:58.207 align:middle line:84% the mean or the median of all of these curves, 00:27:58.207 --> 00:28:00.700 align:middle line:84% and you shift them to the very same mean or median. 00:28:00.700 --> 00:28:03.280 align:middle line:84% And you simply decide where you are 00:28:03.280 --> 00:28:04.880 align:middle line:90% going to shift everything else. 00:28:04.880 --> 00:28:08.452 align:middle line:84% And then, based on that, you re-normalize all the numbers. 00:28:08.452 --> 00:28:10.660 align:middle line:84% And when you look for differentiable expressed genes, 00:28:10.660 --> 00:28:12.730 align:middle line:84% you work with those re-normalized numbers. 00:28:12.730 --> 00:28:15.760 align:middle line:84% Because if you did not, if you hadn't done this, 00:28:15.760 --> 00:28:17.920 align:middle line:84% then you'd say that every gene is downregulated, 00:28:17.920 --> 00:28:20.810 align:middle line:90% which is obviously false. 00:28:20.810 --> 00:28:23.620 align:middle line:84% So that's what normalization is about. 00:28:23.620 --> 00:28:25.870 align:middle line:84% So normalization, in general, is that you don't really 00:28:25.870 --> 00:28:28.120 align:middle line:84% believe the numbers that come out of your experiments, 00:28:28.120 --> 00:28:30.460 align:middle line:84% and you hope or you assume that you're 00:28:30.460 --> 00:28:33.640 align:middle line:84% going to actually improve those numbers by assuming 00:28:33.640 --> 00:28:36.250 align:middle line:84% that you have a systematic error that you can correct. 00:28:36.250 --> 00:28:38.240 align:middle line:84% There are two ways of doing this. 00:28:38.240 --> 00:28:41.020 align:middle line:84% One is that you assume that most or certain things 00:28:41.020 --> 00:28:43.210 align:middle line:84% do not change, and the second one 00:28:43.210 --> 00:28:45.312 align:middle line:84% is that actually, you have an error model. 00:28:45.312 --> 00:28:47.020 align:middle line:84% So the first one are you assume that most 00:28:47.020 --> 00:28:48.430 align:middle line:84% are certain things that are changes to what you 00:28:48.430 --> 00:28:50.000 align:middle line:90% saw on the previous slide. 00:28:50.000 --> 00:28:51.790 align:middle line:84% So you say most of these distributions 00:28:51.790 --> 00:28:54.260 align:middle line:84% actually have to be very, very similar. 00:28:54.260 --> 00:28:57.610 align:middle line:84% And you can shape the means or medians of these curves, 00:28:57.610 --> 00:28:59.980 align:middle line:84% but sometimes the shape of the curve 00:28:59.980 --> 00:29:01.550 align:middle line:84% is going to be different as well. 00:29:01.550 --> 00:29:07.300 align:middle line:84% And, well, if you have this nonlinearity 00:29:07.300 --> 00:29:10.795 align:middle line:84% of the die incorporation, then you not only assume that-- 00:29:10.795 --> 00:29:12.670 align:middle line:84% you can assume that, if the curve is shifted, 00:29:12.670 --> 00:29:14.170 align:middle line:84% then the shape of the curve is going 00:29:14.170 --> 00:29:15.380 align:middle line:90% to be [INAUDIBLE] as well. 00:29:15.380 --> 00:29:18.250 align:middle line:84% So you can do some [INAUDIBLE] [? lowest, ?] 00:29:18.250 --> 00:29:24.520 align:middle line:84% and you can try to change the shape of the curves as well 00:29:24.520 --> 00:29:27.070 align:middle line:84% and shape the means that most of the curves, all the curves, 00:29:27.070 --> 00:29:28.520 align:middle line:90% would look very similar. 00:29:28.520 --> 00:29:32.620 align:middle line:84% And whatever remains as an outlier after all this is done 00:29:32.620 --> 00:29:37.350 align:middle line:84% is your true outlier, what you perceived as a real outlier. 00:29:37.350 --> 00:29:39.700 align:middle line:84% And in most cases, actually, that makes a lot of sense, 00:29:39.700 --> 00:29:44.340 align:middle line:84% and it provides differential cause 00:29:44.340 --> 00:29:50.500 align:middle line:84% that can be corroborated by independent measurements. 00:29:50.500 --> 00:29:53.970 align:middle line:84% This is the similar problem for cDNA microarray measurement. 00:29:53.970 --> 00:29:57.090 align:middle line:84% In this case, the red versus green ratios 00:29:57.090 --> 00:30:01.050 align:middle line:84% are not expected to show any intensity dependence. 00:30:01.050 --> 00:30:04.830 align:middle line:84% But in most cases, when you do two-color microarray, 00:30:04.830 --> 00:30:06.040 align:middle line:90% this is what you see. 00:30:06.040 --> 00:30:08.370 align:middle line:84% So these are the intensity, and this is the ratios. 00:30:08.370 --> 00:30:11.920 align:middle line:84% And you see that what you'd expect is a curve like this, 00:30:11.920 --> 00:30:12.990 align:middle line:90% and that is what you see. 00:30:12.990 --> 00:30:15.660 align:middle line:84% That means that the red and the green dye 00:30:15.660 --> 00:30:20.830 align:middle line:84% is not incorporated with the very same efficiency, 00:30:20.830 --> 00:30:24.720 align:middle line:84% especially depending on the concentration 00:30:24.720 --> 00:30:30.130 align:middle line:84% or the concentration or the individual gene species. 00:30:30.130 --> 00:30:33.180 align:middle line:84% So what you see is that for let's say 00:30:33.180 --> 00:30:36.850 align:middle line:84% low-copy-number genes, red dye is incorporated with a higher 00:30:36.850 --> 00:30:38.100 align:middle line:90% efficiency than the green one. 00:30:38.100 --> 00:30:40.228 align:middle line:84% What you do in this case is actually 00:30:40.228 --> 00:30:42.270 align:middle line:84% try to straighten it out, because we are assuming 00:30:42.270 --> 00:30:44.070 align:middle line:84% that [? we have, ?] for all genes, 00:30:44.070 --> 00:30:47.912 align:middle line:84% the red and green incorporation should be the same. 00:30:47.912 --> 00:30:49.620 align:middle line:84% So what you're trying to do is to correct 00:30:49.620 --> 00:30:50.940 align:middle line:90% for systematic errors. 00:30:50.940 --> 00:30:54.960 align:middle line:84% And in the case of when you assuming 00:30:54.960 --> 00:30:57.730 align:middle line:84% that your basic assumption is that most things do not change, 00:30:57.730 --> 00:31:00.780 align:middle line:84% then you can choose a set of elements that will be used. 00:31:00.780 --> 00:31:02.370 align:middle line:84% That is, sometimes there is a set 00:31:02.370 --> 00:31:07.740 align:middle line:84% of housekeeping genes, which is a very shaky concept. 00:31:07.740 --> 00:31:10.090 align:middle line:84% You're assuming that certain genes do not change-- 00:31:10.090 --> 00:31:12.330 align:middle line:84% let's say metabolic genes do not change 00:31:12.330 --> 00:31:14.427 align:middle line:84% or structural proteins, the genes 00:31:14.427 --> 00:31:16.510 align:middle line:84% associated with structural proteins do not change. 00:31:16.510 --> 00:31:17.940 align:middle line:84% Now, this is used very often, as I'm 00:31:17.940 --> 00:31:20.315 align:middle line:84% sure you've seen, in [INAUDIBLE] is that they [INAUDIBLE] 00:31:20.315 --> 00:31:23.547 align:middle line:84% forget the age or [? actin. ?] Well, it's OK. 00:31:23.547 --> 00:31:26.130 align:middle line:84% It's just very difficult to find a set of genes that is really 00:31:26.130 --> 00:31:27.330 align:middle line:90% not expected to change. 00:31:27.330 --> 00:31:30.000 align:middle line:84% Or you can choose a set of special control genes 00:31:30.000 --> 00:31:32.640 align:middle line:84% that, for some reason, that those genes never change 00:31:32.640 --> 00:31:34.073 align:middle line:90% in your system. 00:31:34.073 --> 00:31:35.490 align:middle line:84% And of course, then, the next step 00:31:35.490 --> 00:31:37.920 align:middle line:84% is you need to determine the normalization function, which 00:31:37.920 --> 00:31:39.750 align:middle line:84% is a global mean or median normalization, 00:31:39.750 --> 00:31:41.400 align:middle line:84% or some of an [? interdependency ?] 00:31:41.400 --> 00:31:42.900 align:middle line:90% normalization. 00:31:42.900 --> 00:31:45.300 align:middle line:84% If you want to learn about this more, then actually, 00:31:45.300 --> 00:31:49.060 align:middle line:84% there's a whole website, and a chatroom, and whatnot. 00:31:49.060 --> 00:31:51.010 align:middle line:84% And there's a whole cottage industry 00:31:51.010 --> 00:31:53.160 align:middle line:84% that's trying to figure out, what's 00:31:53.160 --> 00:31:56.460 align:middle line:84% the best way of normalizing a microarray? 00:31:56.460 --> 00:31:58.320 align:middle line:84% The alternative is that if you come up 00:31:58.320 --> 00:32:04.390 align:middle line:84% with some idea about how the error is actually generated. 00:32:04.390 --> 00:32:08.370 align:middle line:84% So this is the most popular error model, 00:32:08.370 --> 00:32:13.410 align:middle line:84% in which it is assumed that at low concentrations, 00:32:13.410 --> 00:32:15.030 align:middle line:90% you have an admitted error. 00:32:15.030 --> 00:32:17.370 align:middle line:84% You have just simply a normal-- a white noise 00:32:17.370 --> 00:32:18.390 align:middle line:90% around your measurement. 00:32:18.390 --> 00:32:21.420 align:middle line:84% At high concentrations, you have a multiplicative error. 00:32:21.420 --> 00:32:24.060 align:middle line:84% And actually, for all noise, you have the combination 00:32:24.060 --> 00:32:25.530 align:middle line:90% of the two. 00:32:25.530 --> 00:32:27.370 align:middle line:84% So if you make these assumptions, 00:32:27.370 --> 00:32:32.520 align:middle line:84% then you can generate very good error models. 00:32:32.520 --> 00:32:34.290 align:middle line:84% And the normalization based on that 00:32:34.290 --> 00:32:38.070 align:middle line:84% actually gives you a very similar result as 00:32:38.070 --> 00:32:39.900 align:middle line:90% with the previous assumption. 00:32:39.900 --> 00:32:42.840 align:middle line:84% So actually, these two methods, it seems, 00:32:42.840 --> 00:32:45.831 align:middle line:84% are interchangeable, at least for cDNA microarray. 00:32:45.831 --> 00:32:48.915 align:middle line:90% 00:32:48.915 --> 00:32:50.540 align:middle line:84% Noise will limit the useful information 00:32:50.540 --> 00:32:51.560 align:middle line:90% content of measurements. 00:32:51.560 --> 00:32:52.870 align:middle line:90% That's the problem. 00:32:52.870 --> 00:32:57.330 align:middle line:84% That's the issue why you need to be able to deliver that. 00:32:57.330 --> 00:33:00.320 align:middle line:84% So it seems that if you take all these microarray measurements, 00:33:00.320 --> 00:33:05.990 align:middle line:84% then a reliable detection of two or four differences 00:33:05.990 --> 00:33:08.040 align:middle line:90% seems to be the practical limit. 00:33:08.040 --> 00:33:10.190 align:middle line:84% So this is actually a very optimistic and not 00:33:10.190 --> 00:33:12.120 align:middle line:90% cross-platform comparison. 00:33:12.120 --> 00:33:16.010 align:middle line:84% So if you do a large number of Affymetrix measurements 00:33:16.010 --> 00:33:17.330 align:middle line:90% in all sorts of-- 00:33:17.330 --> 00:33:19.640 align:middle line:84% or cDNA micro measurements, or a large number 00:33:19.640 --> 00:33:23.660 align:middle line:84% of very, very different cancer samples, then 00:33:23.660 --> 00:33:27.110 align:middle line:84% it seems that if you take all the information-- or all 00:33:27.110 --> 00:33:28.580 align:middle line:84% the useful information, you extract 00:33:28.580 --> 00:33:30.705 align:middle line:84% all the useful information from your measurements-- 00:33:30.705 --> 00:33:32.543 align:middle line:84% since there are two-fold difference-- 00:33:32.543 --> 00:33:34.970 align:middle line:84% a reliable detection of two-fold difference-- 00:33:34.970 --> 00:33:42.110 align:middle line:84% is pretty much the limit, it's possible 00:33:42.110 --> 00:33:46.010 align:middle line:84% that certain genes are going to be measured reliably, 00:33:46.010 --> 00:33:47.120 align:middle line:90% with higher accuracy. 00:33:47.120 --> 00:33:49.880 align:middle line:84% But across all genes, probably, this 00:33:49.880 --> 00:33:52.040 align:middle line:90% is the experimental limitation. 00:33:52.040 --> 00:33:54.672 align:middle line:84% And why is it an important issue? 00:33:54.672 --> 00:33:56.380 align:middle line:84% Again, getting back to the issue that you 00:33:56.380 --> 00:33:58.797 align:middle line:84% were trying to predict how your system is going to behave, 00:33:58.797 --> 00:34:00.890 align:middle line:84% let's assume that you want to figure out 00:34:00.890 --> 00:34:04.498 align:middle line:84% who is regulating whom, starting with time series measurements. 00:34:04.498 --> 00:34:07.040 align:middle line:84% So you're going to measure gene expression changes or protein 00:34:07.040 --> 00:34:12.120 align:middle line:84% changes within a certain time frame. 00:34:12.120 --> 00:34:14.000 align:middle line:84% So how would you design your experiments? 00:34:14.000 --> 00:34:16.130 align:middle line:84% They were experiments done on the cell cycle 00:34:16.130 --> 00:34:19.790 align:middle line:90% of yeast or human fibroblasts. 00:34:19.790 --> 00:34:22.830 align:middle line:84% But, of course, you have to choose your timing correctly. 00:34:22.830 --> 00:34:27.620 align:middle line:84% So if this is the error of your measurement, this solid line, 00:34:27.620 --> 00:34:30.230 align:middle line:84% then, of course, you don't want to take measurements 00:34:30.230 --> 00:34:34.969 align:middle line:84% more often than the error of the experiment measurement 00:34:34.969 --> 00:34:36.080 align:middle line:90% allows you to do. 00:34:36.080 --> 00:34:38.810 align:middle line:84% So if you know how fast genes are changing 00:34:38.810 --> 00:34:40.969 align:middle line:84% and what's your experimental error, from that, 00:34:40.969 --> 00:34:46.739 align:middle line:84% you can determine a sensible reliable time window, 00:34:46.739 --> 00:34:49.520 align:middle line:84% which seems to be the case that, for example, in yeast, 00:34:49.520 --> 00:34:52.340 align:middle line:84% there is no point of taking more [? geno-spacial ?] measurements 00:34:52.340 --> 00:34:54.800 align:middle line:84% more often than every 5 to 10 minutes, 00:34:54.800 --> 00:34:58.857 align:middle line:84% and in mammalian cells more often than 15 to 30 minutes. 00:34:58.857 --> 00:35:00.440 align:middle line:84% If you measure more frequently, you're 00:35:00.440 --> 00:35:02.150 align:middle line:84% just simply going to run into noise, 00:35:02.150 --> 00:35:03.975 align:middle line:84% and you're just wasting your money. 00:35:03.975 --> 00:35:05.600 align:middle line:84% So that's the reason why you need to be 00:35:05.600 --> 00:35:07.340 align:middle line:90% aware of the noise limitations. 00:35:07.340 --> 00:35:09.830 align:middle line:84% And when you know what you're error or noise 00:35:09.830 --> 00:35:12.740 align:middle line:84% of your measurement, you can make some back of the envelope 00:35:12.740 --> 00:35:15.050 align:middle line:84% calculations of how much information you can actually 00:35:15.050 --> 00:35:17.180 align:middle line:84% extract from that measurement and what 00:35:17.180 --> 00:35:20.300 align:middle line:90% that could be enough for. 00:35:20.300 --> 00:35:22.370 align:middle line:84% So moving on to the other issue of sensitivity 00:35:22.370 --> 00:35:26.682 align:middle line:84% and completeness, when you're trying to predict what's 00:35:26.682 --> 00:35:28.890 align:middle line:84% going to happen to your system, then it's, of course, 00:35:28.890 --> 00:35:32.730 align:middle line:84% the question is that there is a trade-off 00:35:32.730 --> 00:35:35.900 align:middle line:84% or there is this issue of how many parameters 00:35:35.900 --> 00:35:39.062 align:middle line:84% are we measuring, and how many parameters should we measure? 00:35:39.062 --> 00:35:41.270 align:middle line:84% If you're trying to predict whether a certain cell is 00:35:41.270 --> 00:35:42.260 align:middle line:90% going to-- 00:35:42.260 --> 00:35:45.560 align:middle line:84% a certain cancer is going to metastasize or not, 00:35:45.560 --> 00:35:47.540 align:middle line:84% how many genes do you need for that? 00:35:47.540 --> 00:35:50.570 align:middle line:84% If you want to predict how a cell cycle is going 00:35:50.570 --> 00:35:53.600 align:middle line:84% to progress, how many genes do you need to measure for that, 00:35:53.600 --> 00:35:57.450 align:middle line:90% and how many, are we measuring? 00:35:57.450 --> 00:36:00.110 align:middle line:84% So for that, we need to have at least some impression of how 00:36:00.110 --> 00:36:01.710 align:middle line:90% large are these networks. 00:36:01.710 --> 00:36:04.310 align:middle line:84% So this is just showing that it's pretty large. 00:36:04.310 --> 00:36:06.950 align:middle line:84% This is a graph representation of all interacting proteins 00:36:06.950 --> 00:36:07.950 align:middle line:90% in yeast. 00:36:07.950 --> 00:36:13.340 align:middle line:84% So in this case, you have about 5,000 proteins. 00:36:13.340 --> 00:36:15.890 align:middle line:84% Proteins, genes, protein modifications 00:36:15.890 --> 00:36:18.290 align:middle line:84% are all independently regulated, so you can call them 00:36:18.290 --> 00:36:20.510 align:middle line:90% something like bio nodes. 00:36:20.510 --> 00:36:24.650 align:middle line:84% And the [? cautious ?] estimate would be that for in each cell, 00:36:24.650 --> 00:36:27.110 align:middle line:84% the number of bio nodes are going 00:36:27.110 --> 00:36:28.910 align:middle line:84% to be on the order of, let's say, 00:36:28.910 --> 00:36:30.188 align:middle line:90% a couple of hundred thousand. 00:36:30.188 --> 00:36:31.730 align:middle line:84% This is coming from the fact that you 00:36:31.730 --> 00:36:33.920 align:middle line:84% have 10,000 to 20,000 active genes per cell, 00:36:33.920 --> 00:36:38.330 align:middle line:84% and you have, let's say, less than 10 00:36:38.330 --> 00:36:41.930 align:middle line:84% post-translation modifications per gene, per protein. 00:36:41.930 --> 00:36:44.960 align:middle line:84% And that would give you roughly this number. 00:36:44.960 --> 00:36:47.030 align:middle line:84% Of course, that could be much more and much less 00:36:47.030 --> 00:36:49.400 align:middle line:84% in terms of whether you're working with spliced variants 00:36:49.400 --> 00:36:51.440 align:middle line:84% or you actually need to measure on the module, 00:36:51.440 --> 00:36:52.920 align:middle line:90% the activity of modules. 00:36:52.920 --> 00:36:55.790 align:middle line:84% But this is probably on this order 00:36:55.790 --> 00:36:58.040 align:middle line:90% to have such a complete picture. 00:36:58.040 --> 00:37:00.470 align:middle line:84% Now, we certainly don't have this so far, 00:37:00.470 --> 00:37:03.050 align:middle line:84% but this is the way technology has developed. 00:37:03.050 --> 00:37:06.023 align:middle line:84% And actually, this seems to be the easiest thing to achieve. 00:37:06.023 --> 00:37:08.690 align:middle line:84% You just simply move to more and more and more genes, especially 00:37:08.690 --> 00:37:11.510 align:middle line:84% as the genome projects are being completed, 00:37:11.510 --> 00:37:14.510 align:middle line:84% and probably the coverage of the microarray chips of proteomics 00:37:14.510 --> 00:37:17.330 align:middle line:84% is going to reach a complete genome 00:37:17.330 --> 00:37:18.630 align:middle line:90% in the next couple of years. 00:37:18.630 --> 00:37:21.770 align:middle line:84% There is no real reason why it couldn't have been achieved. 00:37:21.770 --> 00:37:24.210 align:middle line:84% All you need [? to have ?] is the sequence information 00:37:24.210 --> 00:37:27.620 align:middle line:90% and set up the technology. 00:37:27.620 --> 00:37:31.365 align:middle line:84% But so the completeness can be achieved in terms of-- 00:37:31.365 --> 00:37:33.740 align:middle line:84% if you work hard enough-- and there are tens of thousands 00:37:33.740 --> 00:37:36.710 align:middle line:84% of biochemists and biologists working on this-- 00:37:36.710 --> 00:37:41.210 align:middle line:84% you can sooner or later measure most 00:37:41.210 --> 00:37:45.830 align:middle line:84% of the biologically important parameters of the cell. 00:37:45.830 --> 00:37:47.460 align:middle line:84% At least in principle, that means 00:37:47.460 --> 00:37:50.640 align:middle line:84% that you can have a probe that would measure this. 00:37:50.640 --> 00:37:53.400 align:middle line:84% But do we actually see signals coming 00:37:53.400 --> 00:37:56.190 align:middle line:84% from these when we are using microarray measurements? 00:37:56.190 --> 00:37:57.480 align:middle line:84% And there was a gentleman, [? Michael ?] [? Holland, ?] 00:37:57.480 --> 00:37:59.480 align:middle line:84% who did these experiments a couple of years ago, 00:37:59.480 --> 00:38:03.240 align:middle line:84% when he just simply took microarray measurements 00:38:03.240 --> 00:38:07.290 align:middle line:84% and RT-PCR measurements as well, on a couple of genes in yeast. 00:38:07.290 --> 00:38:11.040 align:middle line:84% And what he was interested in is, one thing, one, 00:38:11.040 --> 00:38:16.380 align:middle line:84% what's the dynamic range of gene expression changes in yeast? 00:38:16.380 --> 00:38:19.170 align:middle line:84% And what they found is that the [? transcriptome ?] [INAUDIBLE] 00:38:19.170 --> 00:38:22.440 align:middle line:84% in yeast carries varies over six orders of magnitude. 00:38:22.440 --> 00:38:25.470 align:middle line:84% What, actually, this means is that there are lots of genes. 00:38:25.470 --> 00:38:26.700 align:middle line:90% There are lots of cells. 00:38:26.700 --> 00:38:28.740 align:middle line:84% The [INAUDIBLE] [? packages ?] is very large, 00:38:28.740 --> 00:38:32.550 align:middle line:84% and you cannot see this in every single cell because the lowest 00:38:32.550 --> 00:38:34.537 align:middle line:84% number means there are 0.01 copies per cell. 00:38:34.537 --> 00:38:36.120 align:middle line:84% So what you see is that certain genes, 00:38:36.120 --> 00:38:39.870 align:middle line:84% certain cells will express a single copy of a gene 00:38:39.870 --> 00:38:43.560 align:middle line:84% due to stochastic noise, and only every 100 cells 00:38:43.560 --> 00:38:44.931 align:middle line:90% will express that. 00:38:44.931 --> 00:38:51.330 align:middle line:84% So this is the dynamic range of gene expression changes. 00:38:51.330 --> 00:38:57.420 align:middle line:84% He was also interested in that if he measures the gene 00:38:57.420 --> 00:38:59.820 align:middle line:84% expression level of these 300, 400 genes, 00:38:59.820 --> 00:39:03.520 align:middle line:84% and he chose important genes like transcription factors, 00:39:03.520 --> 00:39:05.730 align:middle line:84% and he compares the different technologies-- 00:39:05.730 --> 00:39:08.880 align:middle line:84% RT-PCR is fairly sensitive, although at very, very 00:39:08.880 --> 00:39:11.447 align:middle line:84% low concentrations you run into stochasticity, 00:39:11.447 --> 00:39:13.530 align:middle line:84% that is probably the most sensitive technology you 00:39:13.530 --> 00:39:14.355 align:middle line:90% can use. 00:39:14.355 --> 00:39:16.050 align:middle line:84% And you compare it to microarray, 00:39:16.050 --> 00:39:19.260 align:middle line:84% then how sensitive is microarray relative to RT-PCR? 00:39:19.260 --> 00:39:21.360 align:middle line:90% And that's what he saw. 00:39:21.360 --> 00:39:25.560 align:middle line:84% And what it shows you, that this range of gene 00:39:25.560 --> 00:39:28.530 align:middle line:84% expression levels is completely-- 00:39:28.530 --> 00:39:30.870 align:middle line:90% this is not seen by microarray. 00:39:30.870 --> 00:39:33.990 align:middle line:84% So this is well under the sensitivity of microarray. 00:39:33.990 --> 00:39:36.450 align:middle line:84% What you see is that you start to see something 00:39:36.450 --> 00:39:38.850 align:middle line:84% of a correlation between the microarray measurement 00:39:38.850 --> 00:39:42.760 align:middle line:84% and the RT-PCR at two copies per cell. 00:39:42.760 --> 00:39:45.760 align:middle line:84% So all these genes are actually expressed and changing, 00:39:45.760 --> 00:39:47.760 align:middle line:84% and probably they are doing something important. 00:39:47.760 --> 00:39:49.510 align:middle line:84% As I said, most of the genes were actually 00:39:49.510 --> 00:39:51.240 align:middle line:90% transcription factors. 00:39:51.240 --> 00:39:53.130 align:middle line:84% But they are not seen by microarray. 00:39:53.130 --> 00:39:55.138 align:middle line:84% So sensitivity is a very important issue when 00:39:55.138 --> 00:39:56.430 align:middle line:90% you do microarray measurements. 00:39:56.430 --> 00:39:58.620 align:middle line:84% Well, then, depending on your technology, 00:39:58.620 --> 00:40:00.960 align:middle line:84% you will have lots of genes that are 00:40:00.960 --> 00:40:04.920 align:middle line:84% going to be under the sensitivity of the technology. 00:40:04.920 --> 00:40:08.550 align:middle line:84% I'm sure-- and it's, I mean, as new technologies are coming out 00:40:08.550 --> 00:40:10.960 align:middle line:84% right now, this is going to be improved as well. 00:40:10.960 --> 00:40:12.460 align:middle line:84% But this is another issue you should 00:40:12.460 --> 00:40:15.408 align:middle line:84% be aware of-- that even if you do microarray measurements 00:40:15.408 --> 00:40:17.700 align:middle line:84% and you see lots of blank spots, it doesn't necessarily 00:40:17.700 --> 00:40:19.080 align:middle line:84% mean that those genes are not changing 00:40:19.080 --> 00:40:20.130 align:middle line:90% or they are not present. 00:40:20.130 --> 00:40:23.980 align:middle line:84% Simply, your measurement is not sensitive enough. 00:40:23.980 --> 00:40:26.190 align:middle line:84% So the utmost goal of the technology is going to be, 00:40:26.190 --> 00:40:30.720 align:middle line:84% of course, single measuring, single-copy-- 00:40:30.720 --> 00:40:33.970 align:middle line:90% sorry-- per single gene. 00:40:33.970 --> 00:40:37.092 align:middle line:84% But even if you are measuring everything accurately, 00:40:37.092 --> 00:40:38.800 align:middle line:84% there might be problems with predictions. 00:40:38.800 --> 00:40:40.592 align:middle line:84% And this is what I was referring to before. 00:40:40.592 --> 00:40:46.800 align:middle line:84% And just very quickly, OK, because you're a biologist, 00:40:46.800 --> 00:40:51.780 align:middle line:84% so many years ago-- actually, I think it happened here at MIT-- 00:40:51.780 --> 00:40:53.580 align:middle line:84% a gentleman, Edward Lorenz, was trying 00:40:53.580 --> 00:40:55.538 align:middle line:84% to predict how the weather was going to change. 00:40:55.538 --> 00:40:56.640 align:middle line:90% It was in '60s. 00:40:56.640 --> 00:41:00.480 align:middle line:84% And what he did is he took a few ordinary differential 00:41:00.480 --> 00:41:03.660 align:middle line:84% equations, a completely deterministic system, 00:41:03.660 --> 00:41:07.980 align:middle line:84% and he tried to predict how the outcome 00:41:07.980 --> 00:41:10.740 align:middle line:84% of this set of differentiable equations is going to change. 00:41:10.740 --> 00:41:12.870 align:middle line:84% And what he was really shocked to see-- 00:41:12.870 --> 00:41:15.540 align:middle line:84% and a little bit later, the entire scientific community 00:41:15.540 --> 00:41:17.310 align:middle line:90% was shocked to see-- 00:41:17.310 --> 00:41:20.430 align:middle line:84% that these three ordinary differential equations 00:41:20.430 --> 00:41:22.140 align:middle line:84% produced a behavior very sensitive 00:41:22.140 --> 00:41:23.890 align:middle line:90% to the initial conditions. 00:41:23.890 --> 00:41:26.760 align:middle line:84% Which means that if you just change a very little, just 00:41:26.760 --> 00:41:29.160 align:middle line:84% a smidgen of the starting parameters, 00:41:29.160 --> 00:41:32.850 align:middle line:84% the outcome of the measurements was completely different. 00:41:32.850 --> 00:41:36.600 align:middle line:84% And this ended up in scientific history as chaos theory, 00:41:36.600 --> 00:41:39.910 align:middle line:84% where you might have heard about bifurcations and so forth. 00:41:39.910 --> 00:41:41.820 align:middle line:84% The point is that even if you start 00:41:41.820 --> 00:41:45.610 align:middle line:84% with a seemingly completely deterministic system, 00:41:45.610 --> 00:41:48.330 align:middle line:84% you might not be able to predict how that system is going 00:41:48.330 --> 00:41:51.267 align:middle line:84% to behave because of this very fact-- 00:41:51.267 --> 00:41:53.100 align:middle line:84% that small changes in the initial conditions 00:41:53.100 --> 00:41:55.848 align:middle line:84% can cause huge changes at later time points. 00:41:55.848 --> 00:41:57.390 align:middle line:84% Now, we know that biology is not like 00:41:57.390 --> 00:41:59.220 align:middle line:84% that because biology is a robust system, 00:41:59.220 --> 00:42:01.830 align:middle line:84% because we are sitting here when we are talking. 00:42:01.830 --> 00:42:04.920 align:middle line:84% So many people think that a biological system is somewhere 00:42:04.920 --> 00:42:08.700 align:middle line:84% on the edge of the completely deterministic and chaotic 00:42:08.700 --> 00:42:09.960 align:middle line:90% systems. 00:42:09.960 --> 00:42:11.940 align:middle line:84% But the bottom line is that just because you 00:42:11.940 --> 00:42:13.523 align:middle line:84% can measure everything very accurately 00:42:13.523 --> 00:42:15.273 align:middle line:84% doesn't necessarily mean that you're going 00:42:15.273 --> 00:42:16.530 align:middle line:90% to have very high prediction. 00:42:16.530 --> 00:42:20.130 align:middle line:84% But let me give you a much simpler representation 00:42:20.130 --> 00:42:22.830 align:middle line:84% or example of the very same problem. 00:42:22.830 --> 00:42:25.410 align:middle line:84% Imagine that you already measured very accurately 00:42:25.410 --> 00:42:27.900 align:middle line:84% the gene expression level-- and at very high sensitivity-- 00:42:27.900 --> 00:42:32.130 align:middle line:84% of all genes, or many genes, in a variety of cancer samples. 00:42:32.130 --> 00:42:33.630 align:middle line:84% And what you're trying to figure out 00:42:33.630 --> 00:42:37.410 align:middle line:84% is, what are the genes that are causing cancer? 00:42:37.410 --> 00:42:40.350 align:middle line:84% Now, let's assume that you found this subset of cancer samples 00:42:40.350 --> 00:42:43.200 align:middle line:84% that-- these are actually real measurements from melanoma. 00:42:43.200 --> 00:42:45.930 align:middle line:84% And this is-- let's say this is a subset of samples 00:42:45.930 --> 00:42:49.350 align:middle line:84% that is extremely malignant, kills the patient very quickly. 00:42:49.350 --> 00:42:52.080 align:middle line:84% And you also think that you found 00:42:52.080 --> 00:42:53.820 align:middle line:84% a group of genes that is going to be 00:42:53.820 --> 00:42:57.893 align:middle line:84% responsible for that extremely malignant state. 00:42:57.893 --> 00:42:59.310 align:middle line:84% But you need to ask the question-- 00:42:59.310 --> 00:43:00.960 align:middle line:90% as I referred to this before-- 00:43:00.960 --> 00:43:03.540 align:middle line:84% can this be due to chance, because you have 00:43:03.540 --> 00:43:05.370 align:middle line:90% a limited number of samples? 00:43:05.370 --> 00:43:07.605 align:middle line:84% Well, just by chance, if you randomly 00:43:07.605 --> 00:43:11.533 align:middle line:84% put in those two values, you can see something like this. 00:43:11.533 --> 00:43:13.450 align:middle line:84% Sometimes you can find an analytical solution, 00:43:13.450 --> 00:43:14.492 align:middle line:90% but more often you can't. 00:43:14.492 --> 00:43:16.860 align:middle line:84% You need to do some sort of computational solution. 00:43:16.860 --> 00:43:21.600 align:middle line:84% So you permutate your data set and look for similar patterns. 00:43:21.600 --> 00:43:24.870 align:middle line:84% And if you never find a similar pattern, a similar group 00:43:24.870 --> 00:43:28.050 align:middle line:84% of genes, in the permutated gene expression matrix, 00:43:28.050 --> 00:43:30.540 align:middle line:84% then you say, well, this is not due to chance. 00:43:30.540 --> 00:43:33.850 align:middle line:84% But this is not that obvious how to do it. 00:43:33.850 --> 00:43:39.310 align:middle line:84% So analytical solutions can be sometimes found. 00:43:39.310 --> 00:43:41.770 align:middle line:84% So let me just give you this very simple example. 00:43:41.770 --> 00:43:47.310 align:middle line:84% So this-- I usually pose a problem 00:43:47.310 --> 00:43:50.760 align:middle line:90% that you can solve at home. 00:43:50.760 --> 00:43:55.170 align:middle line:84% We had this problem that, at the dawn of microarray analysis, 00:43:55.170 --> 00:43:57.420 align:middle line:84% my lab measured gene expression measurements 00:43:57.420 --> 00:43:59.920 align:middle line:84% in different breast cancer cell lines. 00:43:59.920 --> 00:44:00.990 align:middle line:90% And when we reached-- 00:44:00.990 --> 00:44:01.875 align:middle line:84% because this was very expensive-- 00:44:01.875 --> 00:44:04.020 align:middle line:84% when we reached six breast cancer cell lines, 00:44:04.020 --> 00:44:06.840 align:middle line:84% we found that 13 consistently misregulated genes, up 00:44:06.840 --> 00:44:08.250 align:middle line:90% or down-regulated genes. 00:44:08.250 --> 00:44:10.750 align:middle line:84% And what we asked is, can this be due to chance or not? 00:44:10.750 --> 00:44:13.067 align:middle line:84% So this was translated into a combinatorics problem 00:44:13.067 --> 00:44:15.150 align:middle line:84% that you have eight different cell lines in a gene 00:44:15.150 --> 00:44:17.550 align:middle line:84% microarray, it's the number of genes misregulated 00:44:17.550 --> 00:44:19.860 align:middle line:90% in the i-th cell line. 00:44:19.860 --> 00:44:23.827 align:middle line:84% And the question was, can we find [? K ?] consistently 00:44:23.827 --> 00:44:25.890 align:middle line:84% in misregulated genes across all these cell lines 00:44:25.890 --> 00:44:27.960 align:middle line:84% by [? chance? ?] So if you like combinatorics, 00:44:27.960 --> 00:44:31.500 align:middle line:84% this is a nice little home exercise if you want to solve. 00:44:31.500 --> 00:44:35.250 align:middle line:84% But so you can find an analytical solution for this. 00:44:35.250 --> 00:44:36.300 align:middle line:90% And this is very simple. 00:44:36.300 --> 00:44:39.120 align:middle line:84% And this could be solved quite easily. 00:44:39.120 --> 00:44:42.030 align:middle line:84% And you a fairly reliable number of this. 00:44:42.030 --> 00:44:44.760 align:middle line:84% But what if more genes are involved? 00:44:44.760 --> 00:44:47.850 align:middle line:84% And more importantly, what if genes are not 00:44:47.850 --> 00:44:49.670 align:middle line:90% independently regulated? 00:44:49.670 --> 00:44:51.420 align:middle line:84% The underlying assumption in combinatorics 00:44:51.420 --> 00:44:54.420 align:middle line:84% is that you're drawing your samples independently, 00:44:54.420 --> 00:44:55.920 align:middle line:90% randomly and independently. 00:44:55.920 --> 00:44:58.680 align:middle line:84% But in this case, genes are co-regulated. 00:44:58.680 --> 00:45:00.600 align:middle line:84% If a transcription factor is upregulated, 00:45:00.600 --> 00:45:02.610 align:middle line:84% well, the downstream genes are going 00:45:02.610 --> 00:45:04.860 align:middle line:84% to be upregulated as well, or some of them 00:45:04.860 --> 00:45:06.300 align:middle line:90% will be upregulated. 00:45:06.300 --> 00:45:08.670 align:middle line:84% And this is coming from real samples. 00:45:08.670 --> 00:45:12.060 align:middle line:84% So what you see here is when you do a complete permutation, 00:45:12.060 --> 00:45:14.100 align:middle line:84% then this is going to be the distribution 00:45:14.100 --> 00:45:18.250 align:middle line:84% of correlation coefficients for each gene pair. 00:45:18.250 --> 00:45:20.740 align:middle line:84% But in real samples, this is what you see. 00:45:20.740 --> 00:45:23.460 align:middle line:84% So there is a high correlation of gene expression changes 00:45:23.460 --> 00:45:24.310 align:middle line:90% up and down-- 00:45:24.310 --> 00:45:26.977 align:middle line:84% which is kind of obvious because this is a genetically regulated 00:45:26.977 --> 00:45:27.480 align:middle line:90% network. 00:45:27.480 --> 00:45:31.890 align:middle line:84% Now, the problem is that if you need to do this analysis, 00:45:31.890 --> 00:45:35.640 align:middle line:84% and you ask the question, is my pattern random or not, 00:45:35.640 --> 00:45:39.480 align:middle line:84% or can this be present due to chance or not, well, 00:45:39.480 --> 00:45:43.530 align:middle line:84% if you use permutated, a randomly permutated gene 00:45:43.530 --> 00:45:45.960 align:middle line:84% expression matrix as your benchmark, 00:45:45.960 --> 00:45:51.240 align:middle line:84% then in that case, your analysis, or your result, 00:45:51.240 --> 00:45:54.510 align:middle line:84% or your statistical analysis can be off by orders of magnitude-- 00:45:54.510 --> 00:45:56.610 align:middle line:84% by six or seven orders of magnitude-- 00:45:56.610 --> 00:45:58.780 align:middle line:84% relative to an analysis where you say, 00:45:58.780 --> 00:46:01.710 align:middle line:84% well, I'm going to permutate the sample 00:46:01.710 --> 00:46:05.370 align:middle line:84% but retain the overall dependency of gene expression 00:46:05.370 --> 00:46:06.473 align:middle line:90% changes. 00:46:06.473 --> 00:46:08.640 align:middle line:84% If you do that-- which is not an obvious thing to do 00:46:08.640 --> 00:46:11.280 align:middle line:84% and takes some computational tricks-- now 00:46:11.280 --> 00:46:15.820 align:middle line:84% you have a very different result. 00:46:15.820 --> 00:46:17.320 align:middle line:90% Noise in discrete measurements. 00:46:17.320 --> 00:46:17.830 align:middle line:90% Yeah? 00:46:17.830 --> 00:46:18.550 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:46:18.550 --> 00:46:20.092 align:middle line:84% ZOLTAN SZALLASI: OK, so what you have 00:46:20.092 --> 00:46:23.860 align:middle line:84% is that you found a pattern that a certain number of genes 00:46:23.860 --> 00:46:26.140 align:middle line:90% are causing, let's say, cancer. 00:46:26.140 --> 00:46:28.780 align:middle line:84% And what people usually do is they 00:46:28.780 --> 00:46:30.730 align:middle line:90% do a complete randomization. 00:46:30.730 --> 00:46:33.370 align:middle line:84% Let's say you just swap everybody, everybody, 00:46:33.370 --> 00:46:36.910 align:middle line:84% and then you look for the same pattern, and you don't find it. 00:46:36.910 --> 00:46:38.500 align:middle line:84% You never find the five genes that 00:46:38.500 --> 00:46:41.500 align:middle line:84% would show the same pattern, and you are happy. 00:46:41.500 --> 00:46:45.220 align:middle line:84% Now, the problem is that this has completely destroyed-- 00:46:45.220 --> 00:46:48.085 align:middle line:84% destroyed this permutation, the codependence of genes. 00:46:48.085 --> 00:46:50.830 align:middle line:90% 00:46:50.830 --> 00:46:55.090 align:middle line:84% And in reality, that means that if you have codependence-- 00:46:55.090 --> 00:46:57.790 align:middle line:84% imagine that there are certain genes that are very strongly 00:46:57.790 --> 00:47:00.610 align:middle line:84% coregulated and other genes are never coregulated-- 00:47:00.610 --> 00:47:02.500 align:middle line:84% that the ones that are actually coregulated 00:47:02.500 --> 00:47:05.528 align:middle line:84% are not actually two independent genes, but in your analysis, 00:47:05.528 --> 00:47:07.570 align:middle line:84% that should be one, you could replace [INAUDIBLE] 00:47:07.570 --> 00:47:08.512 align:middle line:90% gene, right? 00:47:08.512 --> 00:47:09.970 align:middle line:84% And this is what you should retain. 00:47:09.970 --> 00:47:12.267 align:middle line:84% Of course, you don't have complete coregulation and 00:47:12.267 --> 00:47:14.350 align:middle line:84% complete independence, but you have a distribution 00:47:14.350 --> 00:47:15.700 align:middle line:90% of correlation coefficients. 00:47:15.700 --> 00:47:18.620 align:middle line:90% That's what you see here. 00:47:18.620 --> 00:47:21.460 align:middle line:84% So the way one should do this is create 00:47:21.460 --> 00:47:23.950 align:middle line:84% a large number of random matrices 00:47:23.950 --> 00:47:26.530 align:middle line:84% in which the distribution of correlation coefficients 00:47:26.530 --> 00:47:30.160 align:middle line:84% is something like this, but apart from that, it's random. 00:47:30.160 --> 00:47:32.560 align:middle line:84% And then ask the question, is my pattern 00:47:32.560 --> 00:47:34.070 align:middle line:90% present in this as well? 00:47:34.070 --> 00:47:36.520 align:middle line:84% Now, if you compare the statistical power 00:47:36.520 --> 00:47:39.790 align:middle line:84% or the statistical confidence between these two matrices, 00:47:39.790 --> 00:47:44.180 align:middle line:84% you can be off by five or six orders of magnitude. 00:47:44.180 --> 00:47:46.840 align:middle line:84% So something that is significant in this 00:47:46.840 --> 00:47:49.370 align:middle line:84% is way below significance in this. 00:47:49.370 --> 00:47:50.470 align:middle line:90% So that's the point. 00:47:50.470 --> 00:47:52.345 align:middle line:84% It's not that obvious how to do these things. 00:47:52.345 --> 00:47:54.030 align:middle line:84% There's just an important point that 00:47:54.030 --> 00:47:56.530 align:middle line:84% sometimes, even if you have good quality measurements, 00:47:56.530 --> 00:47:59.380 align:middle line:84% biology is going to present you with very difficult problems. 00:47:59.380 --> 00:48:01.780 align:middle line:84% And this is actually present in sequence measurements. 00:48:01.780 --> 00:48:05.120 align:middle line:84% I mean the whole BLAST issue is about this as well. 00:48:05.120 --> 00:48:07.610 align:middle line:84% So moving on to noise in discrete measurements, 00:48:07.610 --> 00:48:08.720 align:middle line:90% which is the best example. 00:48:08.720 --> 00:48:11.740 align:middle line:84% The easiest example is actually DNA sequences. 00:48:11.740 --> 00:48:13.420 align:middle line:84% So of course, you have measurement error 00:48:13.420 --> 00:48:14.680 align:middle line:90% there as well. 00:48:14.680 --> 00:48:16.930 align:middle line:84% You have sequencing errors with a certain probability. 00:48:16.930 --> 00:48:21.400 align:middle line:84% Let's say now it's probably down to 0.1%, 00:48:21.400 --> 00:48:23.607 align:middle line:84% but you use [INAUDIBLE] between 0.1% and 1%. 00:48:23.607 --> 00:48:25.690 align:middle line:84% Of course, the solution was [? sequence ?] [? a ?] 00:48:25.690 --> 00:48:26.350 align:middle line:90% [? lot. ?] 00:48:26.350 --> 00:48:29.470 align:middle line:84% Of course, if you see a difference in your sequencing 00:48:29.470 --> 00:48:31.473 align:middle line:84% and it's not done with a single individual, 00:48:31.473 --> 00:48:32.890 align:middle line:84% you are not quite sure whether you 00:48:32.890 --> 00:48:35.590 align:middle line:84% are seeing a single nucleotide polymorphism, a SNP, 00:48:35.590 --> 00:48:37.790 align:middle line:90% or a sequencing error. 00:48:37.790 --> 00:48:40.270 align:middle line:84% But if you work hard enough and sequence enough, 00:48:40.270 --> 00:48:43.210 align:middle line:84% you will have some sort of feel about the true subsequence 00:48:43.210 --> 00:48:46.220 align:middle line:90% of a DNA sequence. 00:48:46.220 --> 00:48:50.590 align:middle line:84% Now, you end up with a very, very, very long stretch 00:48:50.590 --> 00:48:52.180 align:middle line:90% of layers. 00:48:52.180 --> 00:48:54.580 align:middle line:84% In the case of humans, it's 3 billion. 00:48:54.580 --> 00:48:57.970 align:middle line:84% And what you need to achieve-- or what is expected from you-- 00:48:57.970 --> 00:49:00.737 align:middle line:84% is to find genes, introns, exons, 00:49:00.737 --> 00:49:02.320 align:middle line:84% the transcription factor binding sites 00:49:02.320 --> 00:49:06.070 align:middle line:90% in this sea of four letters. 00:49:06.070 --> 00:49:07.693 align:middle line:90% Now, how do you do that? 00:49:07.693 --> 00:49:09.610 align:middle line:84% This is going to be an issue of noise as well. 00:49:09.610 --> 00:49:13.180 align:middle line:84% If you had only genes, like exons and introns, 00:49:13.180 --> 00:49:15.520 align:middle line:84% or only exons, and transcription factor binding sites, 00:49:15.520 --> 00:49:16.990 align:middle line:90% it would be very easy to find. 00:49:16.990 --> 00:49:19.120 align:middle line:84% The problem is that you have lots and lots 00:49:19.120 --> 00:49:21.985 align:middle line:84% of junk DNA or intergenic regions, 00:49:21.985 --> 00:49:24.340 align:middle line:84% and you have no idea what they are doing. 00:49:24.340 --> 00:49:28.570 align:middle line:84% And in those, sometimes, seemingly intelligible 00:49:28.570 --> 00:49:32.650 align:middle line:84% information will show up just by chance. 00:49:32.650 --> 00:49:34.150 align:middle line:90% So how it can be found? 00:49:34.150 --> 00:49:38.040 align:middle line:84% That's why the real way of building genomes 00:49:38.040 --> 00:49:41.460 align:middle line:84% is not only DNA sequencing because from that, 00:49:41.460 --> 00:49:44.640 align:middle line:84% it's very difficult to find the number of genes. 00:49:44.640 --> 00:49:47.340 align:middle line:84% Actually, if you look hard in the literature about the number 00:49:47.340 --> 00:49:49.080 align:middle line:84% of genes, usually, the number of genes 00:49:49.080 --> 00:49:52.350 align:middle line:84% keeps falling with time because, actually, they 00:49:52.350 --> 00:49:54.780 align:middle line:84% see that there are lots of erroneous predictions. 00:49:54.780 --> 00:49:57.540 align:middle line:84% Usually, these gene prediction algorithms 00:49:57.540 --> 00:50:04.170 align:middle line:84% tend to err on the side that would give you-- 00:50:04.170 --> 00:50:05.280 align:middle line:90% on a more liberal side. 00:50:05.280 --> 00:50:08.080 align:middle line:84% It tends to give you more genes than actually is present. 00:50:08.080 --> 00:50:10.020 align:middle line:84% So what you are looking for is actually cDNA-- 00:50:10.020 --> 00:50:13.110 align:middle line:84% for example, cDNA libraries, for the same organism, 00:50:13.110 --> 00:50:15.450 align:middle line:84% because these are the truly expressed genes. 00:50:15.450 --> 00:50:20.020 align:middle line:84% So you try to bring together the two different databases. 00:50:20.020 --> 00:50:23.610 align:middle line:84% And if you find a cDNA, well, that cDNA 00:50:23.610 --> 00:50:26.250 align:middle line:84% can help you to find the actual genes. 00:50:26.250 --> 00:50:29.680 align:middle line:84% Now, the problem is that cDNA has to be expressed. 00:50:29.680 --> 00:50:36.810 align:middle line:84% And if you didn't happen to prepare a cDNA from the cell 00:50:36.810 --> 00:50:39.720 align:middle line:84% line in which that gene is expressed, well, then, 00:50:39.720 --> 00:50:41.880 align:middle line:84% you won't have that gene in your cDNA library. 00:50:41.880 --> 00:50:45.640 align:middle line:84% Therefore, you cannot find it in your genome. 00:50:45.640 --> 00:50:47.140 align:middle line:90% So how it can found? 00:50:47.140 --> 00:50:50.850 align:middle line:84% And this, the DNA sequence information, 00:50:50.850 --> 00:50:54.030 align:middle line:84% can be refined to large extent by all sorts 00:50:54.030 --> 00:50:55.860 align:middle line:84% of different databases, data sources. 00:50:55.860 --> 00:50:59.370 align:middle line:84% But there are lots of unexpected issues in biology 00:50:59.370 --> 00:51:00.870 align:middle line:90% which are truly amazing. 00:51:00.870 --> 00:51:03.210 align:middle line:84% They are completely unexpected, and you would have never 00:51:03.210 --> 00:51:07.440 align:middle line:84% been able to come up with that idea simply based 00:51:07.440 --> 00:51:09.550 align:middle line:90% on primary sequence information. 00:51:09.550 --> 00:51:12.090 align:middle line:84% And I think I'll just give you two really shocking pieces 00:51:12.090 --> 00:51:14.700 align:middle line:84% of data that are actually pretty reasonable. 00:51:14.700 --> 00:51:17.490 align:middle line:84% One is the widespread occurrence of antisense transcription 00:51:17.490 --> 00:51:18.960 align:middle line:90% in the human genome. 00:51:18.960 --> 00:51:22.110 align:middle line:84% Why do they get-- what do these guys do it, or why? 00:51:22.110 --> 00:51:27.510 align:middle line:84% It's a long story, but what they found actually that-- 00:51:27.510 --> 00:51:31.470 align:middle line:84% they found in the human genome about 1,600 00:51:31.470 --> 00:51:36.720 align:middle line:84% actually transcribed antisense transcription units. 00:51:36.720 --> 00:51:40.490 align:middle line:84% So you know usually how the sense-- 00:51:40.490 --> 00:51:43.430 align:middle line:84% how the genome is read and described in the sense way? 00:51:43.430 --> 00:51:45.530 align:middle line:84% Maybe just looking into whether things 00:51:45.530 --> 00:51:48.530 align:middle line:84% are transcribed in the antisense way. 00:51:48.530 --> 00:51:49.650 align:middle line:90% I mean, you learned a lot. 00:51:49.650 --> 00:51:53.242 align:middle line:84% I mean, you learned a lot about microRNA, SiRna regulatory RNA. 00:51:53.242 --> 00:51:55.700 align:middle line:84% So there was a good reason why they were looking into this. 00:51:55.700 --> 00:51:57.530 align:middle line:84% Nobody would have expected that there 00:51:57.530 --> 00:52:02.300 align:middle line:84% is such a high number of actual antisense transcription units. 00:52:02.300 --> 00:52:09.560 align:middle line:84% Also, when a group checked out what 00:52:09.560 --> 00:52:12.230 align:middle line:84% portion of a given chromosome is actually transcribed, 00:52:12.230 --> 00:52:14.240 align:middle line:84% they were surprised to see that it 00:52:14.240 --> 00:52:16.998 align:middle line:84% was about one order of magnitude more than expected. 00:52:16.998 --> 00:52:19.040 align:middle line:84% What people usually do is you take a chromosome-- 00:52:19.040 --> 00:52:21.860 align:middle line:84% in case they check chromosome 21 and 22-- 00:52:21.860 --> 00:52:24.980 align:middle line:84% you know where the majority of exons or introns are, 00:52:24.980 --> 00:52:27.110 align:middle line:84% and based on that, you expect that most-- 00:52:27.110 --> 00:52:31.670 align:middle line:84% well, the exons are going to be transcribed, and maybe 00:52:31.670 --> 00:52:33.540 align:middle line:90% a couple of regulatory RNAs. 00:52:33.540 --> 00:52:35.450 align:middle line:84% So you have an expectation that, let's say, 00:52:35.450 --> 00:52:38.960 align:middle line:84% a couple of your chromosome, of a given chromosome, 00:52:38.960 --> 00:52:40.400 align:middle line:90% is transcribed. 00:52:40.400 --> 00:52:43.010 align:middle line:84% Now, what they found when they actually 00:52:43.010 --> 00:52:45.740 align:middle line:84% covered the entire chromosome by an asymmetric shape 00:52:45.740 --> 00:52:52.420 align:middle line:84% is actually 10 times as much information 00:52:52.420 --> 00:52:57.580 align:middle line:84% was transcribed from the DNA than expected based on exons. 00:52:57.580 --> 00:53:01.390 align:middle line:84% Again, you will have to predict this just simply 00:53:01.390 --> 00:53:03.637 align:middle line:84% based on primary sequence information. 00:53:03.637 --> 00:53:05.610 align:middle line:90% 00:53:05.610 --> 00:53:06.510 align:middle line:90% But what can you do? 00:53:06.510 --> 00:53:09.765 align:middle line:84% You have this sea of information that seems to be noise. 00:53:09.765 --> 00:53:11.440 align:middle line:84% So is there a way to deal with this? 00:53:11.440 --> 00:53:14.430 align:middle line:84% So let's assume that you need to find a transcription factor 00:53:14.430 --> 00:53:15.498 align:middle line:90% binding site. 00:53:15.498 --> 00:53:17.790 align:middle line:84% It's going to be something like T-G-G-A-C-T. Of course, 00:53:17.790 --> 00:53:20.830 align:middle line:84% you don't know that this is T-G-G-A-C-T. And, of course, 00:53:20.830 --> 00:53:23.640 align:middle line:84% it's not always T-G-G-A-C-T. It can be T-G-C-A-C-T 00:53:23.640 --> 00:53:26.530 align:middle line:84% because transcription factor binding sites like to play with 00:53:26.530 --> 00:53:27.030 align:middle line:90% sequence. 00:53:27.030 --> 00:53:28.650 align:middle line:84% And actually, this is the way they 00:53:28.650 --> 00:53:31.155 align:middle line:84% can change their affinity of their given-- or specificity 00:53:31.155 --> 00:53:33.900 align:middle line:90% of their given sequence. 00:53:33.900 --> 00:53:37.020 align:middle line:84% So then, this is going to be your actual sequence 00:53:37.020 --> 00:53:39.240 align:middle line:90% that it can bind to. 00:53:39.240 --> 00:53:42.455 align:middle line:84% Now, so this is what you're ending up. 00:53:42.455 --> 00:53:44.580 align:middle line:84% This is what you're looking for that you don't know 00:53:44.580 --> 00:53:46.590 align:middle line:90% that this is your binding site. 00:53:46.590 --> 00:53:48.265 align:middle line:84% And you're trying to add constraints. 00:53:48.265 --> 00:53:49.140 align:middle line:90% So this is one trick. 00:53:49.140 --> 00:53:52.680 align:middle line:84% You say transcription factor binding sites are usually 00:53:52.680 --> 00:53:57.120 align:middle line:84% within 500 base pairs upstream from the start codon 00:53:57.120 --> 00:53:58.230 align:middle line:90% of a given gene. 00:53:58.230 --> 00:54:01.852 align:middle line:84% And you also know that it tends to cluster in the same region. 00:54:01.852 --> 00:54:03.810 align:middle line:84% So for most transcription factor binding sites. 00:54:03.810 --> 00:54:05.290 align:middle line:90% You have more than one. 00:54:05.290 --> 00:54:07.300 align:middle line:84% So what you might be doing is you say, 00:54:07.300 --> 00:54:10.170 align:middle line:84% I'm looking for certain, let's say, 00:54:10.170 --> 00:54:14.400 align:middle line:84% six base pair long sequences that tend to cluster within 500 00:54:14.400 --> 00:54:17.010 align:middle line:90% base pairs of A-T-Gs. 00:54:17.010 --> 00:54:18.720 align:middle line:84% And then you're going to find something. 00:54:18.720 --> 00:54:21.180 align:middle line:84% But still, this is going to be very, very weak. 00:54:21.180 --> 00:54:26.580 align:middle line:84% You have way more letters, way more information, and way more 00:54:26.580 --> 00:54:29.190 align:middle line:90% noise from which, then-- 00:54:29.190 --> 00:54:32.640 align:middle line:84% the one-- than the level from which you could extract 00:54:32.640 --> 00:54:34.960 align:middle line:90% the important information. 00:54:34.960 --> 00:54:37.507 align:middle line:84% So even if you do all this, you will 00:54:37.507 --> 00:54:39.840 align:middle line:84% find that many other of the transcription factor binding 00:54:39.840 --> 00:54:43.800 align:middle line:84% [INAUDIBLE] sequences do not function as such. 00:54:43.800 --> 00:54:44.850 align:middle line:90% Well, why? 00:54:44.850 --> 00:54:46.620 align:middle line:84% We do not quite understand yet do 00:54:46.620 --> 00:54:49.930 align:middle line:84% to higher level of DNA [? regularization, ?] whatnot. 00:54:49.930 --> 00:54:52.080 align:middle line:84% And, of course, the problem is that you do not know 00:54:52.080 --> 00:54:53.320 align:middle line:90% what sequence to start with. 00:54:53.320 --> 00:54:55.060 align:middle line:90% So what can you do? 00:54:55.060 --> 00:54:57.510 align:middle line:84% You can hope that your statistical representation 00:54:57.510 --> 00:54:58.770 align:middle line:90% will help. 00:54:58.770 --> 00:55:02.820 align:middle line:84% And one trick is, of course, provided by nature, which 00:55:02.820 --> 00:55:05.140 align:middle line:90% is cross-species conservation. 00:55:05.140 --> 00:55:08.292 align:middle line:84% So you have the extremely noisy-- we call them "noisy," 00:55:08.292 --> 00:55:09.750 align:middle line:84% but of course, they are not noisy-- 00:55:09.750 --> 00:55:14.610 align:middle line:84% extremely noisy genomes-- human, chimp, mouse, rat, yeast, 00:55:14.610 --> 00:55:15.930 align:middle line:90% whatnot. 00:55:15.930 --> 00:55:19.830 align:middle line:84% And you're assuming-- and this assumption is a fair 00:55:19.830 --> 00:55:20.820 align:middle line:90% assumption-- 00:55:20.820 --> 00:55:24.240 align:middle line:84% that you have a cross-species conservation 00:55:24.240 --> 00:55:26.140 align:middle line:90% of important sequences. 00:55:26.140 --> 00:55:28.770 align:middle line:84% So what you're looking for, are there sequences 00:55:28.770 --> 00:55:32.880 align:middle line:84% that are conserved across several species? 00:55:32.880 --> 00:55:37.830 align:middle line:84% And if you combine all this with some of the smart tools 00:55:37.830 --> 00:55:39.820 align:middle line:84% like using artificial intelligence, 00:55:39.820 --> 00:55:42.180 align:middle line:84% machine learning, HMM, hidden Markov models, 00:55:42.180 --> 00:55:44.760 align:middle line:84% were extremely useful to find from [INAUDIBLE] gene 00:55:44.760 --> 00:55:46.860 align:middle line:84% identification, then you might start 00:55:46.860 --> 00:55:49.680 align:middle line:90% to see some patterns emerging. 00:55:49.680 --> 00:55:52.530 align:middle line:84% And this was done for yeast by [? Alexander's ?] group. 00:55:52.530 --> 00:55:54.360 align:middle line:84% So this just gives you a concrete example 00:55:54.360 --> 00:55:56.610 align:middle line:84% that showed that this is actually a very efficient way 00:55:56.610 --> 00:55:57.600 align:middle line:90% to go. 00:55:57.600 --> 00:56:01.710 align:middle line:84% When they sequenced four years species-- 00:56:01.710 --> 00:56:05.910 align:middle line:84% four very closely-related sequenced yeast species-- 00:56:05.910 --> 00:56:11.010 align:middle line:84% the average number of genes in each of them were about 5,500. 00:56:11.010 --> 00:56:13.380 align:middle line:84% The reason they did it is because they knew that these 00:56:13.380 --> 00:56:14.508 align:middle line:90% were very similar species. 00:56:14.508 --> 00:56:16.050 align:middle line:84% So what they found is actually, there 00:56:16.050 --> 00:56:18.970 align:middle line:84% is a very high level of [INAUDIBLE] of genes. 00:56:18.970 --> 00:56:21.090 align:middle line:84% So what they found is that this shows 00:56:21.090 --> 00:56:25.140 align:middle line:84% that same gene is present at the same location in all 00:56:25.140 --> 00:56:26.850 align:middle line:90% of these species. 00:56:26.850 --> 00:56:30.060 align:middle line:84% The order changes, sometimes the gene is lost, 00:56:30.060 --> 00:56:33.090 align:middle line:84% gained, especially either the chromosomes around telomers 00:56:33.090 --> 00:56:36.880 align:middle line:84% or subtelomeric regions, there's a lot of turbulence going on. 00:56:36.880 --> 00:56:39.930 align:middle line:84% But for most of the chromosomes, things 00:56:39.930 --> 00:56:44.250 align:middle line:84% are-- or the information is retained to a large extent. 00:56:44.250 --> 00:56:47.310 align:middle line:84% Of course, there is a slow and rapid evolution. 00:56:47.310 --> 00:56:49.950 align:middle line:84% They found that for certain very important genes, 00:56:49.950 --> 00:56:52.500 align:middle line:84% there's 100% nucleotide conservation 00:56:52.500 --> 00:56:53.880 align:middle line:90% across all species. 00:56:53.880 --> 00:56:56.790 align:middle line:84% For others, there is a very low level of conservation. 00:56:56.790 --> 00:56:59.040 align:middle line:84% Probably, that's something that nature 00:56:59.040 --> 00:57:02.440 align:middle line:90% can afford to experiment with. 00:57:02.440 --> 00:57:05.310 align:middle line:84% But the bottom line is, what they were doing is actually, 00:57:05.310 --> 00:57:08.100 align:middle line:84% they found that important transcription factor 00:57:08.100 --> 00:57:11.490 align:middle line:84% binding sites are going to be present in the very same 00:57:11.490 --> 00:57:13.345 align:middle line:90% location across all species. 00:57:13.345 --> 00:57:14.970 align:middle line:84% So what they were looking for here is-- 00:57:14.970 --> 00:57:16.803 align:middle line:84% actually, this is [INAUDIBLE] binding site-- 00:57:16.803 --> 00:57:19.080 align:middle line:84% and it shows you the four different species, 00:57:19.080 --> 00:57:21.810 align:middle line:84% and it shows this is at the very same location 00:57:21.810 --> 00:57:23.460 align:middle line:90% in all different species. 00:57:23.460 --> 00:57:26.970 align:middle line:84% This is the [INAUDIBLE],, the [INAUDIBLE] binding site. 00:57:26.970 --> 00:57:28.170 align:middle line:90% It's another type of box. 00:57:28.170 --> 00:57:30.420 align:middle line:84% So it shows that you have a very high conservation 00:57:30.420 --> 00:57:34.050 align:middle line:84% of important regulatory information. 00:57:34.050 --> 00:57:37.080 align:middle line:84% Now, what you can do is actually turn this around and look 00:57:37.080 --> 00:57:38.520 align:middle line:90% for unknown information. 00:57:38.520 --> 00:57:42.090 align:middle line:84% So what you do is-- what they did is let's generate-- 00:57:42.090 --> 00:57:48.390 align:middle line:84% or they generated all random sequences, which was X-Y-Z. 00:57:48.390 --> 00:57:52.800 align:middle line:84% That means that X, Y, and Z stands for any of A-T-C, 00:57:52.800 --> 00:57:57.720 align:middle line:84% A-T-C-G, and [? A-T-C, ?] and there is any number of random 00:57:57.720 --> 00:58:01.020 align:middle line:84% A-T-Cs between them, between 0 and 21. 00:58:01.020 --> 00:58:01.845 align:middle line:90% You can do this. 00:58:01.845 --> 00:58:04.823 align:middle line:84% This is within the realm of scientific computation. 00:58:04.823 --> 00:58:06.240 align:middle line:84% This is actually not [INAUDIBLE].. 00:58:06.240 --> 00:58:08.660 align:middle line:90% So these are any combination. 00:58:08.660 --> 00:58:12.980 align:middle line:84% And you look for certain statistically significant 00:58:12.980 --> 00:58:14.360 align:middle line:90% patterns for these. 00:58:14.360 --> 00:58:17.330 align:middle line:84% One of them is intergenic conservation. 00:58:17.330 --> 00:58:21.950 align:middle line:84% Are there any sequences like this, 00:58:21.950 --> 00:58:23.990 align:middle line:84% when you go through all sequences, that 00:58:23.990 --> 00:58:27.950 align:middle line:84% tend to be conserved between genes and intergenic regions? 00:58:27.950 --> 00:58:30.920 align:middle line:84% You can check for intergenic versus genic conservation, 00:58:30.920 --> 00:58:34.130 align:middle line:84% or you can check for upstream versus downstream conservation. 00:58:34.130 --> 00:58:37.368 align:middle line:84% These are all statistical benchmarks 00:58:37.368 --> 00:58:39.410 align:middle line:84% they found for known transcription factor binding 00:58:39.410 --> 00:58:40.430 align:middle line:90% sites. 00:58:40.430 --> 00:58:43.640 align:middle line:84% So what they found, that for known transcription factor 00:58:43.640 --> 00:58:46.100 align:middle line:90% binding sites, all these have-- 00:58:46.100 --> 00:58:48.710 align:middle line:84% those are more conserved in intergenic conservation. 00:58:48.710 --> 00:58:50.960 align:middle line:84% You have a higher intergenic versus genic conservation 00:58:50.960 --> 00:58:55.000 align:middle line:84% and upstream versus downstream conservation. 00:58:55.000 --> 00:58:56.480 align:middle line:90% So recall the problem. 00:58:56.480 --> 00:58:59.780 align:middle line:84% You're starting with any random sequence. 00:58:59.780 --> 00:59:01.790 align:middle line:84% You're just trying to figure out that any 00:59:01.790 --> 00:59:05.180 align:middle line:84% of these random sequences have any biological significance. 00:59:05.180 --> 00:59:07.340 align:middle line:84% Now, even more importantly, what they 00:59:07.340 --> 00:59:11.120 align:middle line:84% found is that when you start to find statistically 00:59:11.120 --> 00:59:14.300 align:middle line:84% significantly retained or conserved sequences, 00:59:14.300 --> 00:59:17.090 align:middle line:84% then these motives were also arranged 00:59:17.090 --> 00:59:19.700 align:middle line:84% in front of genes that tended to share function, 00:59:19.700 --> 00:59:20.870 align:middle line:90% which is very important. 00:59:20.870 --> 00:59:22.820 align:middle line:84% Because you're assuming that there 00:59:22.820 --> 00:59:24.620 align:middle line:84% are certain functional modules, so genes 00:59:24.620 --> 00:59:26.090 align:middle line:84% that tend to do the same thing has 00:59:26.090 --> 00:59:28.220 align:middle line:84% to be turned on or off at the same time 00:59:28.220 --> 00:59:30.827 align:middle line:90% or under the same conditions. 00:59:30.827 --> 00:59:32.660 align:middle line:84% So that's when they came up with a long list 00:59:32.660 --> 00:59:35.240 align:middle line:84% of potential transcription factor binding sites, 00:59:35.240 --> 00:59:38.720 align:middle line:84% in which all these things were pulled together. 00:59:38.720 --> 00:59:41.510 align:middle line:84% And they found that these are sequences 00:59:41.510 --> 00:59:46.880 align:middle line:84% that tend to be conserved in front of genes 00:59:46.880 --> 00:59:48.420 align:middle line:90% that tend to share function. 00:59:48.420 --> 00:59:51.560 align:middle line:84% And many of these actually were confirmed independently 00:59:51.560 --> 00:59:55.400 align:middle line:84% by experiments as new through transcription factor binding 00:59:55.400 --> 00:59:56.240 align:middle line:90% sites. 00:59:56.240 --> 01:00:00.890 align:middle line:84% So the bottom line is that in these measurements, even 01:00:00.890 --> 01:00:05.360 align:middle line:84% in discrete measurements, this sequencing-- 01:00:05.360 --> 01:00:10.190 align:middle line:84% you will have to face a lot of noise. 01:00:10.190 --> 01:00:13.850 align:middle line:84% Biological organisms were built a long time ago, 01:00:13.850 --> 01:00:16.100 align:middle line:90% and the blueprints were lost. 01:00:16.100 --> 01:00:17.720 align:middle line:84% If you knew how it was built, then you 01:00:17.720 --> 01:00:19.490 align:middle line:84% could figure out what's important or not, 01:00:19.490 --> 01:00:22.350 align:middle line:84% but everything was experimented a long time ago. 01:00:22.350 --> 01:00:24.620 align:middle line:84% So it seems to you now that right now, 01:00:24.620 --> 01:00:30.410 align:middle line:84% the important information is hidden in a sea of irrelevant 01:00:30.410 --> 01:00:32.160 align:middle line:90% information. 01:00:32.160 --> 01:00:34.340 align:middle line:84% And it will be very difficult-- and usually, it's 01:00:34.340 --> 01:00:39.090 align:middle line:84% impossible-- to find based on solely computational ground. 01:00:39.090 --> 01:00:42.800 align:middle line:84% But if you look for help from actual biology-- 01:00:42.800 --> 01:00:45.590 align:middle line:84% in this case, cross-species conservation-- well, 01:00:45.590 --> 01:00:51.394 align:middle line:84% then, the important things, the gold nuggets, start to emerge. 01:00:51.394 --> 01:00:52.980 align:middle line:90% OK, and that's it. 01:00:52.980 --> 01:00:54.910 align:middle line:90% Any questions? 01:00:54.910 --> 01:00:57.000 align:middle line:90%