WEBVTT 00:00:01.000 --> 00:00:04.000 Today is my last class with you. Awe, I'm sorry, too. You guys are 00:00:04.000 --> 00:00:08.000 a lot of fun. This has actually been the most interactive 7. 00:00:08.000 --> 00:00:12.000 1 I've ever had. Usually there are a couple of people who perk up and 00:00:12.000 --> 00:00:16.000 say things, but you guys are great because all sorts of people are 00:00:16.000 --> 00:00:20.000 willing to contribute. So, I've had a wonderful time and 00:00:20.000 --> 00:00:24.000 it certainly seems like you guys have learned a lot. 00:00:24.000 --> 00:00:28.000 What I'd like to do for my last lecture is pick up again a little 00:00:28.000 --> 00:00:32.000 bit like I did with genomics and try to give you a sense of where 00:00:32.000 --> 00:00:36.000 things are going. I always like doing this because I 00:00:36.000 --> 00:00:40.000 get to talk about things that are in none of the textbooks that, 00:00:40.000 --> 00:00:44.000 well, I mean, it's just stuff that many people working in the field 00:00:44.000 --> 00:00:48.000 don't necessarily know. And that's what's so much fun about 00:00:48.000 --> 00:00:52.000 teaching introductory biology is because it only takes a semester for 00:00:52.000 --> 00:00:56.000 you guys to get up to the point of at least being able to understand 00:00:56.000 --> 00:01:01.000 what's getting done on the cutting-edge. 00:01:01.000 --> 00:01:05.000 Even if you might not yet be able to go off and practice it, 00:01:05.000 --> 00:01:09.000 you might need a little more experience for that, 00:01:09.000 --> 00:01:13.000 but you'd be surprised, it's not that much more. 00:01:13.000 --> 00:01:17.000 Take maybe Project Lab and you'll be able to start doing it already. 00:01:17.000 --> 00:01:21.000 It's really wonderful that it's possible to grasp what's going on. 00:01:21.000 --> 00:01:25.000 And, in many ways, you guys may have an advantage in grasping what's 00:01:25.000 --> 00:01:29.000 going on because, as I've already hinted, 00:01:29.000 --> 00:01:33.000 biology's undergoing this remarkable transformation from being a purely 00:01:33.000 --> 00:01:37.000 laboratory-based science where each individual works on his or her own 00:01:37.000 --> 00:01:41.000 project to being an information-based science that 00:01:41.000 --> 00:01:45.000 involves an integration of vast amounts of data across the whole 00:01:45.000 --> 00:01:50.000 world and trying to learn things from this tremendous dataset. 00:01:50.000 --> 00:01:52.000 And, in that sense, I think the new students coming into 00:01:52.000 --> 00:01:55.000 the field have a distinct advantage over those who have been in it. 00:01:55.000 --> 00:01:58.000 And certainly the students who know mathematical and physical and 00:01:58.000 --> 00:02:01.000 chemical and other sorts of things, and aren't scared to write computer 00:02:01.000 --> 00:02:04.000 code when they need to write computer code have a really 00:02:04.000 --> 00:02:07.000 great advantage. So, anyway, all that by way of 00:02:07.000 --> 00:02:11.000 introduction. I want to talk about two subjects today of great interest 00:02:11.000 --> 00:02:14.000 to me. One is DNA variation and one is RNA variation. 00:02:14.000 --> 00:02:18.000 The variation of DNA sequence between individuals within a 00:02:18.000 --> 00:02:21.000 population, and in particular our population, and the other is RNA 00:02:21.000 --> 00:02:25.000 variation, the variation in RNA expression between different cell 00:02:25.000 --> 00:02:28.000 types, different tissues. And the work I'm going to talk 00:02:28.000 --> 00:02:32.000 about today is work that I, and my colleagues, have all been 00:02:32.000 --> 00:02:36.000 involved in. And it's stuff I know and love. 00:02:36.000 --> 00:02:40.000 So, feel free to ask questions about it. I may know the answers, 00:02:40.000 --> 00:02:44.000 but what's reasonably fun about these lectures is if I don't know 00:02:44.000 --> 00:02:48.000 the answers it's probably the case that the answers aren't known. 00:02:48.000 --> 00:02:52.000 So, that's good fun because it's stuff I really do know well, 00:02:52.000 --> 00:02:56.000 and I love. So, anyway, here's some DNA sequence. It's pretty boring. 00:02:56.000 --> 00:03:00.000 This is a chunk of sequence from, let's say, the human genome. 00:03:00.000 --> 00:03:04.000 How much does this differ between any two individuals? 00:03:04.000 --> 00:03:09.000 If I were to sequence any two chromosomes, any two copies of the 00:03:09.000 --> 00:03:14.000 chromosome from an individual in this class or two individuals on 00:03:14.000 --> 00:03:19.000 this planet, how much would they differ? The answer is that much. 00:03:19.000 --> 00:03:25.000 That's the average amount of difference between any two people on 00:03:25.000 --> 00:03:30.000 this planet. Not a lot. If you counted up, it is on average 00:03:30.000 --> 00:03:35.000 one nucleotide difference out of 1, 00 nucleotides on average, somewhat 00:03:35.000 --> 00:03:41.000 less than one part in 1, 00 or better than 99.9% identity 00:03:41.000 --> 00:03:46.000 between any two individuals. Now, that is a very small amount, 00:03:46.000 --> 00:03:51.000 not just in absolute terms, 99.9% identity is a lot, 00:03:51.000 --> 00:03:57.000 but in comparative terms with other species. If I take two chimpanzees 00:03:57.000 --> 00:04:02.000 in Africa, on average they will differ by about twice as much as any 00:04:02.000 --> 00:04:07.000 two random humans. And if I take two orangutans in 00:04:07.000 --> 00:04:12.000 Southeast Asia, they will on average differ by about 00:04:12.000 --> 00:04:17.000 eight times as much as any two humans on this planet. 00:04:17.000 --> 00:04:21.000 You guys think the orangutans all look the same. 00:04:21.000 --> 00:04:26.000 They think you all look the same, and they're right. So, why is this? 00:04:26.000 --> 00:04:31.000 Why are humans amongst mammalian 00:04:31.000 --> 00:04:36.000 species relatively limited in the amount of variation? 00:04:36.000 --> 00:04:40.000 Well, it's a direct result of our population history. 00:04:40.000 --> 00:04:45.000 It turns out that the amount of variation that can be sustained in a 00:04:45.000 --> 00:04:50.000 population depends on two things. At equilibrium, if population has 00:04:50.000 --> 00:04:55.000 constant size N for a very long time and a certain mutation rate, 00:04:55.000 --> 00:04:59.000 Mu, you can just write a piece of arithmetic that says, 00:04:59.000 --> 00:05:04.000 well, mutations are always arising due to new mutations in the 00:05:04.000 --> 00:05:09.000 population and mutations are being lost by genetic drift, 00:05:09.000 --> 00:05:14.000 just by random sampling from generation to generation. 00:05:14.000 --> 00:05:17.000 And those two processes, the creation of new mutations and 00:05:17.000 --> 00:05:21.000 the loss of mutations just due to random sampling in each generation, 00:05:21.000 --> 00:05:25.000 sets up an equilibrium, and the equilibrium defines an equation 00:05:25.000 --> 00:05:29.000 there, Pi equals one over one plus four and Mu reciprocal which 00:05:29.000 --> 00:05:33.000 equation you have no need to memorize whatsoever and possibly 00:05:33.000 --> 00:05:36.000 even no need to write down. The important point is the concept, 00:05:36.000 --> 00:05:40.000 that if you know the number of organisms in the population and you 00:05:40.000 --> 00:05:43.000 know the mutation rate, those set up the bounds of mutation 00:05:43.000 --> 00:05:47.000 and drift, and you can write down how polymorphic, 00:05:47.000 --> 00:05:51.000 how heterozygous random individuals should be at equilibrium. 00:05:51.000 --> 00:05:54.000 That is if the population has been at size N for a very long time. 00:05:54.000 --> 00:05:58.000 Well, the expected amount of heterozygosity for the 00:05:58.000 --> 00:06:02.000 human population -- Sorry. For a population of size 10, 00:06:02.000 --> 00:06:06.000 00 would be about one nucleotide in 1300. We have exactly the amount of 00:06:06.000 --> 00:06:11.000 heterozygosity you would expect for a population of about 10, 00:06:11.000 --> 00:06:15.000 00 individuals. Yeah, but wait, we're not a population of 10,000 00:06:15.000 --> 00:06:20.000 individuals. Why do we have the heterozygosity you would expect from 00:06:20.000 --> 00:06:25.000 a population of 10, 00 individuals? We're six billion. 00:06:25.000 --> 00:06:31.000 It's a reflection of our history. 00:06:31.000 --> 00:06:35.000 Because remember I said that was the statement about what the 00:06:35.000 --> 00:06:38.000 population heterozygosity should be at equilibrium? 00:06:38.000 --> 00:06:42.000 We haven't been six billion people except very recently. 00:06:42.000 --> 00:06:45.000 The human population has undergone an exponential expansion. 00:06:45.000 --> 00:06:49.000 It used to be a relatively small size, and then it very recently 00:06:49.000 --> 00:06:52.000 underwent this huge exponential expansion. If you actually write 00:06:52.000 --> 00:06:56.000 down the equations, the amount of variation in our 00:06:56.000 --> 00:07:00.000 population was determined by that constant size for a very long time. 00:07:00.000 --> 00:07:03.000 And then a rapid exponential expansion that's basically taken 00:07:03.000 --> 00:07:07.000 place in a mere 3, 00 generations, it's much too rapid 00:07:07.000 --> 00:07:11.000 to have any affect on the real variation in our population. 00:07:11.000 --> 00:07:15.000 What do I mean by that? What's the mutation rate per nucleotide in the 00:07:15.000 --> 00:07:18.000 human genome? It's on the order of two times ten to the minus eighth 00:07:18.000 --> 00:07:22.000 per generation. In a mere 3,000 generations, 00:07:22.000 --> 00:07:26.000 a tiny mutation rate like two times ten to the minus eighth is not going 00:07:26.000 --> 00:07:30.000 to be able to build up much more variation. 00:07:30.000 --> 00:07:32.000 So you might as well ignore the last 100,000 years or so. 00:07:32.000 --> 00:07:34.000 They're irrelevant to how much variation we have. 00:07:34.000 --> 00:07:36.000 The variation we have was set by our ancestral population size. 00:07:36.000 --> 00:07:38.000 Now, don't get me wrong. Eventually it will equilibrate. 00:07:38.000 --> 00:07:40.000 A couple million years from now we will have a much higher variation in 00:07:40.000 --> 00:07:42.000 the human population as a function of our size, but the population 00:07:42.000 --> 00:07:44.000 variation we have today is set by the fact that humans derive from a 00:07:44.000 --> 00:07:47.000 founding population of about 10, 00 individuals or so. 00:07:47.000 --> 00:07:52.000 And that means that the variation that you see in the human population 00:07:52.000 --> 00:07:57.000 is mostly ancestral variations, the variation that we all walked 00:07:57.000 --> 00:08:03.000 around with in Africa. And, in fact, that makes a 00:08:03.000 --> 00:08:08.000 prediction. That would say that if most of the variation in the human 00:08:08.000 --> 00:08:13.000 population is from the ancestral African founding population then if 00:08:13.000 --> 00:08:19.000 I go to any two villages around this world, in Japan or in Sweden or in 00:08:19.000 --> 00:08:24.000 Nigeria, the variance that I see will largely be identical. 00:08:24.000 --> 00:08:30.000 And that prediction has been well satisfied. 00:08:30.000 --> 00:08:34.000 Because when you go and look and you collect variation in Japan or Sweden 00:08:34.000 --> 00:08:38.000 or Africa and you compare it, 90% of the variance are common 00:08:38.000 --> 00:08:42.000 across the entire world. Most variation is common ancestral 00:08:42.000 --> 00:08:46.000 variation around the world, and only a minority of the variance 00:08:46.000 --> 00:08:50.000 are new local mutations restricted to individual populations. 00:08:50.000 --> 00:08:54.000 This is so contrary to what people think because there's a natural 00:08:54.000 --> 00:08:58.000 tendency to kind of xenophobia, to imagine that world populations 00:08:58.000 --> 00:09:02.000 are very different in their genetic background. 00:09:02.000 --> 00:09:05.000 But, in point of fact, they're extremely similar. 00:09:05.000 --> 00:09:09.000 So, anyway, there's a limited amount of variation. 00:09:09.000 --> 00:09:13.000 That's why we have such little variation in the human population. 00:09:13.000 --> 00:09:17.000 Now, that variation, humans have a low rate of genetic variation. 00:09:17.000 --> 00:09:20.000 Most of the variance that are out there are due to common genetic 00:09:20.000 --> 00:09:24.000 variance, not rare variance. If I take your genome and I find a 00:09:24.000 --> 00:09:28.000 site of genetic variation at the point of heterozygosity in your 00:09:28.000 --> 00:09:32.000 genome, what's the probability that somebody else in this class also is 00:09:32.000 --> 00:09:36.000 heterozygous for that spot? It turns out that the odds are about 00:09:36.000 --> 00:09:40.000 95% that someone else in this class will also share that variance. 00:09:40.000 --> 00:09:44.000 So that the variance are not mostly rare, they're mostly common. 00:09:44.000 --> 00:09:48.000 And it turns out that some of this common variation, 00:09:48.000 --> 00:09:52.000 that is most of this variation is likely to be important in the risk 00:09:52.000 --> 00:09:56.000 of human genetic diseases. So human geneticists have gotten 00:09:56.000 --> 00:10:00.000 very excited about the following paradigm. 00:10:00.000 --> 00:10:03.000 If there's only a limited amount of genetic variation in the human 00:10:03.000 --> 00:10:06.000 population, actually, if you do the arithmetic, 00:10:06.000 --> 00:10:09.000 there are only about ten million sites of common variation in the 00:10:09.000 --> 00:10:12.000 human population, where common might be defined as 00:10:12.000 --> 00:10:15.000 more than about 1% in the population. There are only ten million sites. 00:10:15.000 --> 00:10:18.000 Folks are saying, well, why not enumerate them all? 00:10:18.000 --> 00:10:22.000 Let's just know them all, and then let's test each one for its 00:10:22.000 --> 00:10:25.000 risk of, say, confirming susceptibility of diabetes or heart 00:10:25.000 --> 00:10:28.000 disease or whatever? After all, ten million is not as 00:10:28.000 --> 00:10:32.000 big a number as it used to be. We now have the whole sequence of 00:10:32.000 --> 00:10:36.000 the human genome. Why not layer on the sequence of 00:10:36.000 --> 00:10:40.000 the human genome all common human genetic polymorphism? 00:10:40.000 --> 00:10:44.000 Now, that's a fairly outrageous idea but could be a very useful one. 00:10:44.000 --> 00:10:48.000 Some of these variance are important, by the way. 00:10:48.000 --> 00:10:52.000 We know that there are two nucleotides that vary in the gene 00:10:52.000 --> 00:10:56.000 apolipoprotein E on chromosome number 19. Apolipoprotein E is also 00:10:56.000 --> 00:11:00.000 an apolipoprotein like we talked about before with familiar 00:11:00.000 --> 00:11:04.000 hypercholesterolemia. But, in fact, it turns out that 00:11:04.000 --> 00:11:08.000 apolipoprotein E is expressed in the brain. And it turns out, 00:11:08.000 --> 00:11:13.000 amongst other tissues, that it comes in three variances, 00:11:13.000 --> 00:11:18.000 the spelling T-T, T-C and C-C at those two particular spots. 00:11:18.000 --> 00:11:22.000 And if you happen to be homozygous for the E4 variant, 00:11:22.000 --> 00:11:27.000 homozygous for the E4 variant, you have about a 60% to 70% lifetime 00:11:27.000 --> 00:11:32.000 risk of Alzheimer's disease. In this class 13 of you are 00:11:32.000 --> 00:11:37.000 homozygous for E4 and have a high lifetime risk of Alzheimer's. 00:11:37.000 --> 00:11:42.000 And it would be fairly trivial to go across the street to anybody's 00:11:42.000 --> 00:11:47.000 lab and test that. Now, I don't particular recommend 00:11:47.000 --> 00:11:52.000 it, and I haven't tested myself for this variant because there happens 00:11:52.000 --> 00:11:57.000 to be no particular therapy available today to delay the onset 00:11:57.000 --> 00:12:01.000 of Alzheimer's disease. And, therefore, 00:12:01.000 --> 00:12:05.000 I don't recommend finding out about that. But a number of 00:12:05.000 --> 00:12:08.000 pharmaceutical companies, knowing that this is a very 00:12:08.000 --> 00:12:11.000 important gene in the pathogenesis of Alzheimer's disease, 00:12:11.000 --> 00:12:15.000 are working on drugs to try to delay the pathogenesis using this 00:12:15.000 --> 00:12:18.000 information. And it may be the case that five or ten years from now 00:12:18.000 --> 00:12:21.000 people will begin to offer drugs that will delay the onset of 00:12:21.000 --> 00:12:25.000 Alzheimer's disease by delaying the interaction of apolipoprotein E with 00:12:25.000 --> 00:12:29.000 a target protein called towe, etc. So, this is an example of where a 00:12:29.000 --> 00:12:33.000 common variant in the population points us to the basis of a common 00:12:33.000 --> 00:12:37.000 disease and has important therapeutic implications. 00:12:37.000 --> 00:12:41.000 There are some other ones, for example. 5% of you carry a 00:12:41.000 --> 00:12:45.000 particular variant in your factor 5 gene which is the clotting cascade. 00:12:45.000 --> 00:12:49.000 It's called the leiden variant. Those 5% of you are going to account 00:12:49.000 --> 00:12:53.000 for 50% of the admissions to emergency rooms for deep venous 00:12:53.000 --> 00:12:57.000 clots, for example. The much higher risk of deep venous 00:12:57.000 --> 00:13:02.000 clots. And, in particular, 00:13:02.000 --> 00:13:06.000 there are significant issues if you have that variant and you are a 00:13:06.000 --> 00:13:11.000 woman with taking birth control pills. Some of you were at higher 00:13:11.000 --> 00:13:16.000 risk for diabetes, type 2 adult onset diabetes. 00:13:16.000 --> 00:13:20.000 There's a particular variant in the population that increased your risk 00:13:20.000 --> 00:13:25.000 for type 2 diabetes by about 30%. 85% of you have the high-risk 00:13:25.000 --> 00:13:30.000 factor, so you might as well figure you do. 00:13:30.000 --> 00:13:35.000 15% of you have a lower risk, et cetera. And one I'm particularly 00:13:35.000 --> 00:13:40.000 interested in here, it turns out that HIV virus gets 00:13:40.000 --> 00:13:46.000 into cells with a co-receptor encoded by a gene called CCR5. 00:13:46.000 --> 00:13:51.000 Well, it turns out that if we go across the European population, 00:13:51.000 --> 00:13:57.000 10% of all chromosomes of European ancestry have a deletion 00:13:57.000 --> 00:14:02.000 within the CCR5 gene. If 10% of all chromosomes have that 00:14:02.000 --> 00:14:06.000 deletion then 10% times 10%, 1% of all individuals are homozygous 00:14:06.000 --> 00:14:10.000 for that deletion. Those individuals are essentially 00:14:10.000 --> 00:14:15.000 immune to infection from HIV. They are not susceptible. It's not 00:14:15.000 --> 00:14:19.000 through immunity, it's through lack of a receptor. 00:14:19.000 --> 00:14:23.000 Yes? You certainly can. It's not hard. It's a specific known variant. 00:14:23.000 --> 00:14:28.000 You could test for it. Absolutely. 00:14:28.000 --> 00:14:31.000 Now, of course, that only helps the 1% of people who 00:14:31.000 --> 00:14:34.000 have that variant. But what it did do was point to the 00:14:34.000 --> 00:14:37.000 pharmaceutical industry that the interaction between the virus and 00:14:37.000 --> 00:14:41.000 that variant is essential. And now companies are developing 00:14:41.000 --> 00:14:44.000 drugs to block the interaction with that particular protein. 00:14:44.000 --> 00:14:48.000 And that tells you that it's an important protein. Yes? 00:14:48.000 --> 00:14:56.000 Over the whole world? 00:14:56.000 --> 00:15:00.000 I just specified European population for that one. 00:15:00.000 --> 00:15:03.000 That one, interestingly, is not found at as high a frequency 00:15:03.000 --> 00:15:06.000 outside of Europe, and no one knows why, 00:15:06.000 --> 00:15:09.000 whether that might have been due to an ancient selective event or a 00:15:09.000 --> 00:15:13.000 genetic drift. By contrast, the apolipoprotein E 00:15:13.000 --> 00:15:16.000 variant, at that frequency of about 3% of people being homozygous and 00:15:16.000 --> 00:15:19.000 being at risk for Alzheimer's, is about the same frequency 00:15:19.000 --> 00:15:23.000 everywhere in the world. So, there's a little bit of 00:15:23.000 --> 00:15:26.000 population variation in frequency. Now, the HIV variant is found 00:15:26.000 --> 00:15:30.000 elsewhere but at considerably lower frequencies there. 00:15:30.000 --> 00:15:33.000 And that's an interesting question as to what causes that variation. 00:15:33.000 --> 00:15:36.000 So the notion would be, I've given you a couple of interesting examples, 00:15:36.000 --> 00:15:40.000 but, look, there's only ten million variants. Just write them all down. 00:15:40.000 --> 00:15:43.000 Make one big Excel spreadsheet with ten million variants along the top 00:15:43.000 --> 00:15:47.000 and all the diseases along the rows, and let's just fill in the matrix 00:15:47.000 --> 00:15:50.000 and then we'll really, you know, this is the way people 00:15:50.000 --> 00:15:54.000 think in a post-genomic era. Now, could you do something like 00:15:54.000 --> 00:15:57.000 that? You would have to enumerate all of the single nucleotide 00:15:57.000 --> 00:16:01.000 polymorphisms, or SNPs we call them, 00:16:01.000 --> 00:16:05.000 single nucleotide polymorphisms. Now, to give you an idea of the 00:16:05.000 --> 00:16:09.000 magnitude of this problem, as recently as 1998, the number of 00:16:09.000 --> 00:16:13.000 SNPs that were known in the human genome was a couple hundred. 00:16:13.000 --> 00:16:17.000 But then a project has taken off. In 1998 an initial SNP map of the 00:16:17.000 --> 00:16:21.000 human genome was built here at MIT that had about 4, 00:16:21.000 --> 00:16:25.000 00 of these variants. Then within the next year or so an 00:16:25.000 --> 00:16:29.000 international consortium was organized here and elsewhere to 00:16:29.000 --> 00:16:34.000 begin to collect more of these genetic variants. 00:16:34.000 --> 00:16:38.000 The goal was going to be to find 300, 00 of them within a period of two 00:16:38.000 --> 00:16:42.000 years. In fact, that goal was blown away and within 00:16:42.000 --> 00:16:46.000 three years two million of the SNPs in the human population were found. 00:16:46.000 --> 00:16:51.000 And as of today, if you go on the Web, you'll find the database with 00:16:51.000 --> 00:16:55.000 about 7.8 million of the roughly ten million SNPs in the human population 00:16:55.000 --> 00:17:00.000 already known. Now, that isn't all ten million. 00:17:00.000 --> 00:17:03.000 And it takes a while to collect the last ones, you know, 00:17:03.000 --> 00:17:07.000 collecting the last ones are hard, but we're already the hump of 00:17:07.000 --> 00:17:10.000 knowing the majority of common variation in the human population. 00:17:10.000 --> 00:17:14.000 Not just a sequence of the genome, but a database that already contains 00:17:14.000 --> 00:17:17.000 more than half of all common variation in the population. 00:17:17.000 --> 00:17:21.000 So, we could start building that Excel spreadsheet. 00:17:21.000 --> 00:17:24.000 Now, it turns out that it's even a little bit better than that because 00:17:24.000 --> 00:17:28.000 if we look at many chromosomes in the population, 00:17:28.000 --> 00:17:31.000 here are chromosomes in the population, it turns out that the 00:17:31.000 --> 00:17:35.000 common variance on each of those chromosomes tend to be correlated 00:17:35.000 --> 00:17:38.000 with each other. If I know your genotype at one 00:17:38.000 --> 00:17:41.000 variant, like over at this locus, I know your genotype at the next 00:17:41.000 --> 00:17:45.000 locus with reasonably high probability. There's a lot of local 00:17:45.000 --> 00:17:48.000 correlation. So, instead of looking like a scattered 00:17:48.000 --> 00:17:51.000 picture like that, it's more like this. 00:17:51.000 --> 00:17:55.000 If I know that you're red, red, red you're probably red, 00:17:55.000 --> 00:17:58.000 red, red over here. In other words, these variations occur in blocks 00:17:58.000 --> 00:18:01.000 that we called haplotypes. Here's real data. 00:18:01.000 --> 00:18:04.000 Across 111 kilobases of DNA there's a bunch of variants, 00:18:04.000 --> 00:18:08.000 but it turns out that the variants come in two basic flavors. 00:18:08.000 --> 00:18:11.000 98% of all chromosomes are either this, this, this, 00:18:11.000 --> 00:18:14.000 this, this or this, this, this, this, this. 00:18:14.000 --> 00:18:18.000 Then there tends to be sites of recombination that are actually 00:18:18.000 --> 00:18:21.000 hotspots of recombination where most of the recombination of the 00:18:21.000 --> 00:18:24.000 population is concentrated. And you get a couple of 00:18:24.000 --> 00:18:28.000 possibilities here. So, the human genome can kind of be 00:18:28.000 --> 00:18:31.000 broken up into these haplotypes. Blocks that might be 20, 00:18:31.000 --> 00:18:35.000 30, 40, sometimes 100 kilobases long in which within the block you tend 00:18:35.000 --> 00:18:39.000 to have a small number of haplotypes, or flavors as you might think of 00:18:39.000 --> 00:18:43.000 them, that define most of the chromosomes in the population. 00:18:43.000 --> 00:18:46.000 So, in fact, I don't actually need to know all the variants. 00:18:46.000 --> 00:18:50.000 If they're so well correlated within a block, 00:18:50.000 --> 00:18:54.000 if I knew this block structure I would be able to pick a small number 00:18:54.000 --> 00:18:58.000 of SNPs that would serve as a proxy for that entire block of inheritance 00:18:58.000 --> 00:19:01.000 in the population. So, what you might want to do is 00:19:01.000 --> 00:19:04.000 determine that entire haplotype block structure of hwo they're 00:19:04.000 --> 00:19:08.000 related to each other, and pick out tag snips. 00:19:08.000 --> 00:19:11.000 And it turns out that in theory, a mere 300,000 or so of them would 00:19:11.000 --> 00:19:14.000 suffice to proxy for most of the genome. So, you might want to 00:19:14.000 --> 00:19:18.000 declare an international project, and international haplotype map 00:19:18.000 --> 00:19:21.000 project to create a haplotype map of the human genome. 00:19:21.000 --> 00:19:24.000 And indeed, such a project was declared about a year and a half ago 00:19:24.000 --> 00:19:28.000 through some instigation of scientists and a number of places, 00:19:28.000 --> 00:19:31.000 including here. And this is $100 million project 00:19:31.000 --> 00:19:35.000 involving six different countries. And, it is already more than 00:19:35.000 --> 00:19:39.000 halfway done with the task, and it's very likely that by the 00:19:39.000 --> 00:19:42.000 middle of next year, we will have a pretty good haplotype 00:19:42.000 --> 00:19:46.000 map, not just knowing all the variation, but knowing the 00:19:46.000 --> 00:19:50.000 correlation between that variation, being able to break up the genome 00:19:50.000 --> 00:19:53.000 into these blocks. By the next time I teach 701, 00:19:53.000 --> 00:19:57.000 I should be able to show a haplotype map of the whole human genome 00:19:57.000 --> 00:20:01.000 already. That will allow you to start undertaking systematic studies 00:20:01.000 --> 00:20:05.000 of inheritance for different diseases across populations. 00:20:05.000 --> 00:20:08.000 And in fact, people are already doing things like that. 00:20:08.000 --> 00:20:12.000 Here's an example of a study done here at MIT like this, 00:20:12.000 --> 00:20:15.000 where to study inflammatory bowel disease, there was evidence that 00:20:15.000 --> 00:20:19.000 there might be a particular region of the genome that contained it, 00:20:19.000 --> 00:20:22.000 and haplotypes were determined across this, and blah, 00:20:22.000 --> 00:20:26.000 blah, blah, blah, blah, blah, blah. And this red haplotype 00:20:26.000 --> 00:20:29.000 here turns out to confer high risk, about a two and a half or higher 00:20:29.000 --> 00:20:33.000 risk of inflammatory bowel disease. 00:20:33.000 --> 00:20:36.000 And it sits over some genes involved in immune responses, 00:20:36.000 --> 00:20:40.000 certain cytokine genes and all that. And, things like this have been 00:20:40.000 --> 00:20:44.000 done for type 2 diabetes, schizophrenia, cardiovascular 00:20:44.000 --> 00:20:47.000 disease, just right now at the moment, a dozen or two examples. 00:20:47.000 --> 00:20:51.000 But I think we're set for an explosion in this kind of work. 00:20:51.000 --> 00:20:55.000 In addition, you can use this information to do things beyond 00:20:55.000 --> 00:20:59.000 medical genetics. You can use it for history and 00:20:59.000 --> 00:21:03.000 anthropology as well. It turns out rather interestingly, 00:21:03.000 --> 00:21:07.000 that since the human population originated in Africa and spread out 00:21:07.000 --> 00:21:12.000 from Africa all the way around the world arriving at different places 00:21:12.000 --> 00:21:17.000 in different times, you can trace those migrations by 00:21:17.000 --> 00:21:21.000 virtue of rare genetic variants that arose along the way, 00:21:21.000 --> 00:21:26.000 and let you, like a trail of break crumbs, see the migrations. 00:21:26.000 --> 00:21:30.000 So, for example, there are certain rare genetic variants that we can 00:21:30.000 --> 00:21:35.000 see in a South American Indian tribe, and we can actually see that they 00:21:35.000 --> 00:21:40.000 came along this route because we can see that residual of that. 00:21:40.000 --> 00:21:45.000 In fact, we can do things with this like take a look at Native American 00:21:45.000 --> 00:21:50.000 individuals and determine that they cluster into three distinct genetic 00:21:50.000 --> 00:21:55.000 groups that represent three distinct migrations over the land bridge. 00:21:55.000 --> 00:22:00.000 And, you can assign them to these different migrations. 00:22:00.000 --> 00:22:03.000 You can do this on the basis of mitochondrial genotype, 00:22:03.000 --> 00:22:06.000 etc. You can also, for example, determine when people talk about the 00:22:06.000 --> 00:22:09.000 out of Africa migration, there's now increasing evidence that 00:22:09.000 --> 00:22:13.000 there really were two, one that went this way over the land, 00:22:13.000 --> 00:22:16.000 and one that went this way following along the coast into southeast Asia. 00:22:16.000 --> 00:22:19.000 And, it looks like we're now beginning to get enough evidence of 00:22:19.000 --> 00:22:22.000 these two separate migrations by virtue of the genetic breadcrumbs 00:22:22.000 --> 00:22:26.000 that they have left along the way. 00:22:26.000 --> 00:22:30.000 So, it's really a very fascinating thing of how much you can 00:22:30.000 --> 00:22:34.000 reconstruct from looking at genetic variation, both the common variation 00:22:34.000 --> 00:22:38.000 that allows us to recognize medical risk, and the rare genetic variation 00:22:38.000 --> 00:22:43.000 that provides much more individual trails of things. 00:22:43.000 --> 00:22:47.000 None of this is perfect yet. There's lots to learn. But I think 00:22:47.000 --> 00:22:51.000 anthropologists are finding that the existing human population has a 00:22:51.000 --> 00:22:55.000 tremendous amount of its own history embedded in pattern of genetic 00:22:55.000 --> 00:23:00.000 variation across the world. You can do other things. 00:23:00.000 --> 00:23:04.000 I won't spend much time on this. Well, I'll take a moment on this, 00:23:04.000 --> 00:23:09.000 right? There's some very interesting work of a post-doctoral 00:23:09.000 --> 00:23:13.000 fellow here at MIT named Pardese Sebetti who has been trying to ask, 00:23:13.000 --> 00:23:18.000 can we see in the genetic variation in the population, 00:23:18.000 --> 00:23:22.000 signatures, patterns of ancient selection, or even recent selection 00:23:22.000 --> 00:23:27.000 in the human population? Now, hang onto your seats, 00:23:27.000 --> 00:23:32.000 because this will get just slightly tricky. 00:23:32.000 --> 00:23:35.000 But, hang on. It's only a couple of slides. Here was her idea. 00:23:35.000 --> 00:23:39.000 You see, when a mutation arises in the population, 00:23:39.000 --> 00:23:43.000 it usually dies out, right? Any new mutation just 00:23:43.000 --> 00:23:47.000 typically dies out. But, sometimes by chance it drifts 00:23:47.000 --> 00:23:50.000 up to a high frequency. Random events happen. But it 00:23:50.000 --> 00:23:54.000 usually takes a long time to do that. If some random mutation happens, 00:23:54.000 --> 00:23:58.000 and it happens to drift up to high frequency with no selection on it, 00:23:58.000 --> 00:24:02.000 then on average it takes a long time to do so. 00:24:02.000 --> 00:24:05.000 If you want, I could write a stochastic differential equation 00:24:05.000 --> 00:24:09.000 that would say that, but just take your gut feeling that 00:24:09.000 --> 00:24:12.000 if something has no selection on it and it's a rare event that'll drift 00:24:12.000 --> 00:24:16.000 up, when it drifts up it's kind of a slow process. It was a slow process. 00:24:16.000 --> 00:24:20.000 Then over the course of time that it took to drift to high frequency, 00:24:20.000 --> 00:24:23.000 a lot of genetic recombination would have had to have occurred many 00:24:23.000 --> 00:24:27.000 generations. And the correlation between the genotype at that spot 00:24:27.000 --> 00:24:31.000 and genotypes at other loci would break down. 00:24:31.000 --> 00:24:34.000 And there would only be short-range correlation. So, 00:24:34.000 --> 00:24:38.000 in other words, the amount of correlation between knowing the 00:24:38.000 --> 00:24:41.000 genotype here and the genotype here, maybe allele A here and a C here. 00:24:41.000 --> 00:24:45.000 That is an indication of time. It's a clock almost. It's like 00:24:45.000 --> 00:24:49.000 radioactive decay, right, that genetic recombination 00:24:49.000 --> 00:24:52.000 scrambles up the correlations. And, if something's old, the 00:24:52.000 --> 00:24:56.000 correlations go over short distances. But suppose that something happened. 00:24:56.000 --> 00:25:00.000 Some mutation happened that was very advantageous. 00:25:00.000 --> 00:25:03.000 Then, it would have risen to high frequency quickly because it was 00:25:03.000 --> 00:25:07.000 under selection. If it did so quickly, 00:25:07.000 --> 00:25:11.000 then the long-range correlations would not have had time to break 00:25:11.000 --> 00:25:15.000 down, and we'd have a smoking gun. A smoking gun would be that there 00:25:15.000 --> 00:25:18.000 would be a long-range correlation around that locus, 00:25:18.000 --> 00:25:22.000 much longer than you would expect across the genome. 00:25:22.000 --> 00:25:26.000 Things even out of this distance would show correlation with that, 00:25:26.000 --> 00:25:30.000 indicating that this was a recent event. 00:25:30.000 --> 00:25:34.000 So, we just measure across the genome, and look for this telltale 00:25:34.000 --> 00:25:39.000 sign of common variance that have very long range correlation that 00:25:39.000 --> 00:25:44.000 indicate that they're very recent. So, a plot of the allele frequency, 00:25:44.000 --> 00:25:49.000 common variance, sorry, if something has a common high frequency and 00:25:49.000 --> 00:25:54.000 long-range correlation, you wouldn't expect that by chance. 00:25:54.000 --> 00:25:58.000 So, something that was common in its 00:25:58.000 --> 00:26:02.000 frequency and had long-range correlation would be a signature of 00:26:02.000 --> 00:26:06.000 positive selection. So anyway, Pardise had this idea, 00:26:06.000 --> 00:26:09.000 and she tried it out with some interesting mutations, 00:26:09.000 --> 00:26:13.000 some mutations that confer resistance to malaria, 00:26:13.000 --> 00:26:17.000 one well-known mutation causing resistance to malaria called G6 PD 00:26:17.000 --> 00:26:21.000 and another one that she herself had proposed as a mutation causing 00:26:21.000 --> 00:26:24.000 resistance to malaria, variants in the CD4 ligand gene. 00:26:24.000 --> 00:26:28.000 And to make a long story short, both the known and her newly 00:26:28.000 --> 00:26:32.000 predicted variant showed this telltale property of having a high 00:26:32.000 --> 00:26:36.000 frequency and very long range correlation. 00:26:36.000 --> 00:26:40.000 Well that's very interesting because she was able to show that each of 00:26:40.000 --> 00:26:44.000 these mutations probably were the result of positive selection. 00:26:44.000 --> 00:26:49.000 But what you could do in principle is test every variant in the human 00:26:49.000 --> 00:26:53.000 genome this way: take any variant, look at its frequency, and compare 00:26:53.000 --> 00:26:58.000 it to the long range correlation around it, and test every single 00:26:58.000 --> 00:27:02.000 variant in the human population to see which ones might be the result 00:27:02.000 --> 00:27:06.000 of long range correlation. Now, when she proposed this, 00:27:06.000 --> 00:27:09.000 this was about a year and a half ago or two years ago, 00:27:09.000 --> 00:27:12.000 this was a pretty nutty idea because you would need all the variants in 00:27:12.000 --> 00:27:15.000 the human population, and you would need all this 00:27:15.000 --> 00:27:18.000 correlation information. But in fact, as I say, that 00:27:18.000 --> 00:27:21.000 information's almost upon us, and I believed that this experiment, 00:27:21.000 --> 00:27:24.000 this analysis to look for all strong positive selection in the human 00:27:24.000 --> 00:27:27.000 genome will in fact be done in the course of the next 12 months. 00:27:27.000 --> 00:27:30.000 So, I'm hoping by next year I can actually report on a genome-wide 00:27:30.000 --> 00:27:33.000 search for all the signatures of positive selection. 00:27:33.000 --> 00:27:36.000 Now, this doesn't detect all positive selection. 00:27:36.000 --> 00:27:39.000 It will detect sufficiently strong positive selection going back pretty 00:27:39.000 --> 00:27:42.000 much only over the 10, 00 years. When you do the 00:27:42.000 --> 00:27:45.000 arithmetic, that's how much power you have. Of course, 00:27:45.000 --> 00:27:48.000 10,000 years has been a pretty interesting time for the human 00:27:48.000 --> 00:27:52.000 population, right? The time of civilization and 00:27:52.000 --> 00:27:55.000 population density, and infectious diseases, 00:27:55.000 --> 00:27:58.000 and all that, and I think we'll have an interesting window into 00:27:58.000 --> 00:28:02.000 the change in diet. All of that should come out of 00:28:02.000 --> 00:28:06.000 something like this. So, there's a lot of really cool 00:28:06.000 --> 00:28:10.000 information in DNA variation to be had. All right, 00:28:10.000 --> 00:28:14.000 that's one half. The other half of what I would like to talk about is 00:28:14.000 --> 00:28:18.000 totally different. It's not about inherited DNA 00:28:18.000 --> 00:28:22.000 variation. It's about somatic differences between tissues in RNA 00:28:22.000 --> 00:28:26.000 variation. So, let's shift gears. 00:28:26.000 --> 00:28:30.000 RNA variation: let me start by giving you an example here. 00:28:30.000 --> 00:28:36.000 These are cells from two different patients with acute leukemia. 00:28:36.000 --> 00:28:43.000 Can you spot the difference between these? Yep? More like bunches of 00:28:43.000 --> 00:28:49.000 grapes and all that. Yeah, it turns out that's just a 00:28:49.000 --> 00:28:56.000 reflection of the field of view you have if you move over 00:28:56.000 --> 00:29:02.000 to look like that. But I mean, that's good. 00:29:02.000 --> 00:29:07.000 It's just that it turns out that that isn't actually a distinction 00:29:07.000 --> 00:29:12.000 when you look at more fields. Anything else? Yep? White blood 00:29:12.000 --> 00:29:16.000 cells like different. They look broken. There's more of 00:29:16.000 --> 00:29:21.000 them in this field of view. But you look at 100 fields of view 00:29:21.000 --> 00:29:26.000 and it turns out that's not either. Well, the reason you're having 00:29:26.000 --> 00:29:31.000 trouble spotting any difference is that highly trained pathologists 00:29:31.000 --> 00:29:35.000 can't find any difference either. I generally agree there's no 00:29:35.000 --> 00:29:39.000 difference between these two if you look at enough fields of view. 00:29:39.000 --> 00:29:43.000 But you can convince yourself if you look that you see things there. 00:29:43.000 --> 00:29:46.000 But these actually are two very different kinds of leukemia. 00:29:46.000 --> 00:29:50.000 And, these patients have to be treated very differently. 00:29:50.000 --> 00:29:54.000 But, pathologists cannot determine which leukemia it is just by looking 00:29:54.000 --> 00:29:57.000 at the microscope, it turns out. This is the work of this man, 00:29:57.000 --> 00:30:01.000 Sydney Farber, namesake of the Dana Farber Cancer Institute here in 00:30:01.000 --> 00:30:05.000 Boston, who in the 1950s began noticing that patients with 00:30:05.000 --> 00:30:08.000 leukemias, some of them seemed different in the way they responded 00:30:08.000 --> 00:30:12.000 to a certain treatment, and he said, look, I think there's 00:30:12.000 --> 00:30:16.000 some underlying classification of these leukemias, 00:30:16.000 --> 00:30:19.000 but I can't get any reliable way to tell it in the microscope. 00:30:19.000 --> 00:30:23.000 And he put many years into working this out, first by noticing certain 00:30:23.000 --> 00:30:27.000 difference in enzymes in the cells, and then people noticed certain 00:30:27.000 --> 00:30:31.000 things in cell surface markers, and some chromosomal rearrangements. 00:30:31.000 --> 00:30:34.000 And nowadays, there are a bunch of test that can be done by a 00:30:34.000 --> 00:30:38.000 pathologist when a patient comes in with acute leukemia to determine 00:30:38.000 --> 00:30:42.000 whether they have AML or ALL. But it turns out that you can't do 00:30:42.000 --> 00:30:46.000 it by looking. You have to do some kind of 00:30:46.000 --> 00:30:50.000 immunohystochemical test of some sort in order to do that. 00:30:50.000 --> 00:30:54.000 So this is a triumph of diagnosis. After 40 years of work, we can now 00:30:54.000 --> 00:30:58.000 correctly classify patients as AML or ALL. And they get the 00:30:58.000 --> 00:31:02.000 appropriate treatment. And if they don't get the right 00:31:02.000 --> 00:31:06.000 treatment, they have a much higher chance of dying. 00:31:06.000 --> 00:31:10.000 And if they do get the right treatment, they have a much higher 00:31:10.000 --> 00:31:14.000 chance of living. So, this is great. 00:31:14.000 --> 00:31:18.000 There's only one problem with the story. It took 40 years, 00:31:18.000 --> 00:31:22.000 40 years to sort this out. That's a long time. Couldn't we do 00:31:22.000 --> 00:31:26.000 better? Surely these cells know what they are. 00:31:26.000 --> 00:31:30.000 Surely we could just ask them if they are. Well, here's the idea. 00:31:30.000 --> 00:31:33.000 Suppose we could ask each cell, please tell us every gene that you 00:31:33.000 --> 00:31:37.000 have turned on, and the level to which you have that 00:31:37.000 --> 00:31:40.000 gene expressed. In other words, 00:31:40.000 --> 00:31:44.000 let us summarize each cell, each tumor by a description of its 00:31:44.000 --> 00:31:47.000 complete pattern of gene expression to 22,000 genes on the human genome. 00:31:47.000 --> 00:31:51.000 Let's write down the level of expression, X1 up to X22, 00:31:51.000 --> 00:31:54.000 00 for each of the 22,000 genes of the genome. So, 00:31:54.000 --> 00:31:58.000 ever tumor becomes a point in 22, 00 dimensional space, right? 00:31:58.000 --> 00:32:01.000 Now clearly, if we had every tumor described as a point in 22, 00:32:01.000 --> 00:32:05.000 00 dimensional space, we ought to be able to sort out which tumors are 00:32:05.000 --> 00:32:09.000 similar to each other, right? Well, it turns out you can 00:32:09.000 --> 00:32:13.000 do that now. These are gene chips, one of several technologies by which 00:32:13.000 --> 00:32:17.000 on a piece of glass are put little spots, each of which contains a 00:32:17.000 --> 00:32:21.000 piece of DNA, a unique DNA sequence. Actually, many copies of that DNA 00:32:21.000 --> 00:32:25.000 sequence are there. Each of these is a 25 base long DNA 00:32:25.000 --> 00:32:29.000 sequence, and I can design this so whatever DNA sequence you 00:32:29.000 --> 00:32:32.000 want is in each spot. The way that's done is with the same 00:32:32.000 --> 00:32:36.000 photolithographic techniques that are used to make microprocessors. 00:32:36.000 --> 00:32:40.000 People have worked out a chemistry where through a mask, 00:32:40.000 --> 00:32:44.000 you shine a light, photodeprotect certain pixels; the pixels that are 00:32:44.000 --> 00:32:48.000 photodeprotected you can chemically attach an A, then re-protect the 00:32:48.000 --> 00:32:52.000 surface. Use a light. Chemically photodeprotect certain 00:32:52.000 --> 00:32:56.000 spots. Wash on a C. And in this fashion, 00:32:56.000 --> 00:33:00.000 since you can randomly address the spots by light, 00:33:00.000 --> 00:33:04.000 and then chemically add bases to whatever spots are deprotected, 00:33:04.000 --> 00:33:08.000 you can simultaneously construct hundreds of thousands of spots each 00:33:08.000 --> 00:33:12.000 containing its own unique specified oligonucleotide sequence. 00:33:12.000 --> 00:33:16.000 And you can get them in little plastic chips. 00:33:16.000 --> 00:33:20.000 And then if you want, all you do is you take a tumor. 00:33:20.000 --> 00:33:24.000 You grind it up. You prepare RNA. You fluorescently label the RNA 00:33:24.000 --> 00:33:28.000 with some appropriate fluorescent dye. You squirt it into the chip. 00:33:28.000 --> 00:33:31.000 You wash it back and forth. You rock it back and forth, 00:33:31.000 --> 00:33:35.000 wash it out, and stick it in a laser scanner. And it'll see how much 00:33:35.000 --> 00:33:38.000 fluorescence is stuck to each spot. And bingo: you get a readout of the 00:33:38.000 --> 00:33:42.000 level of gene expression. I guess each spot, you should 00:33:42.000 --> 00:33:45.000 design it so that this spot has an oligonucleotide complementary to 00:33:45.000 --> 00:33:49.000 gene number one. And the next one, 00:33:49.000 --> 00:33:53.000 an oligonucleotide matching by Crick-Watson base pairing 00:33:53.000 --> 00:33:56.000 complementary to gene number two and gene number three. 00:33:56.000 --> 00:34:00.000 So, if I knew all the genes in the genome, I could make a detector spot 00:34:00.000 --> 00:34:03.000 for each gene in the genome. And of course we know essentially 00:34:03.000 --> 00:34:07.000 all the genes in the genome. So you can make those detector 00:34:07.000 --> 00:34:10.000 spots and you can buy them. So, you can now get a readout of 00:34:10.000 --> 00:34:13.000 all the, I mean, this is like so cool because when I 00:34:13.000 --> 00:34:17.000 started teaching 701, which wasn't that long ago because I 00:34:17.000 --> 00:34:20.000 ain't (sic) that old still, the way people did an analysis of 00:34:20.000 --> 00:34:23.000 gene expression is they used primitive technologies where they 00:34:23.000 --> 00:34:27.000 would analyze one gene at a time, certain things called northern blots 00:34:27.000 --> 00:34:30.000 and things like that, right? And, you know, 00:34:30.000 --> 00:34:34.000 you'd put in a lot of work and you get the expression level of a gene, 00:34:34.000 --> 00:34:37.000 whereas now you can get the expression of all the genes 00:34:37.000 --> 00:34:41.000 simultaneously, and it's pretty mind boggling that 00:34:41.000 --> 00:34:44.000 you can do that. How do you analyze data like that? 00:34:44.000 --> 00:34:48.000 So, we still use northern blots. It's true. So, 00:34:48.000 --> 00:34:51.000 every tumor becomes a vector, and we get a vector corresponding to 00:34:51.000 --> 00:34:55.000 each tumor. So, this line here is the first tumor, 00:34:55.000 --> 00:34:59.000 the second tumor, the third tumor, the fourth tumor. 00:34:59.000 --> 00:35:02.000 The columns here correspond to genes. There are 22, 00:35:02.000 --> 00:35:06.000 00 columns in this matrix, and I've shown a certain subset of 00:35:06.000 --> 00:35:10.000 the columns because these genes here have the interesting property that 00:35:10.000 --> 00:35:14.000 they tend to be high red in the ALL tumors, and they tend to be low blue 00:35:14.000 --> 00:35:18.000 in the AML tumors, whereas these genes here have the 00:35:18.000 --> 00:35:22.000 opposite property. They tend to be low blue in the ALL 00:35:22.000 --> 00:35:26.000 tumors and high red in the AML tumors. These genes do a pretty 00:35:26.000 --> 00:35:30.000 good job of telling apart these tumors. 00:35:30.000 --> 00:35:35.000 So, here's a new tumor. Patient came in. We analyzed the 00:35:35.000 --> 00:35:40.000 RNA, squirted it on the chip. Can somebody classify that? Louder? 00:35:40.000 --> 00:35:45.000 AML. Next? Next? Congratulations, you're 00:35:45.000 --> 00:35:50.000 pathologists. Very good. That's right, you can do that. 00:35:50.000 --> 00:35:56.000 It works. And in fact, in the study that was done that was 00:35:56.000 --> 00:36:01.000 published about this, the computer was able to get it 00:36:01.000 --> 00:36:05.000 right 100% of the time. Not bad. So now you say, 00:36:05.000 --> 00:36:09.000 wait, wait, wait, but you're cheating. 00:36:09.000 --> 00:36:12.000 You're giving it a whole bunch of knowns. Once I have a whole bunch 00:36:12.000 --> 00:36:15.000 of knowns it's not so hard to classify a new tumor. 00:36:15.000 --> 00:36:19.000 What Sydney Farber did was he discovered in the first place that 00:36:19.000 --> 00:36:22.000 there existed two subtypes. Surely that's harder than 00:36:22.000 --> 00:36:26.000 classifying when you're given a bunch of knowns. And 00:36:26.000 --> 00:36:29.000 that's true. So, suppose instead, 00:36:29.000 --> 00:36:33.000 I didn't tell you in advance which were AML's and which were ALL's, 00:36:33.000 --> 00:36:37.000 and I just gave you vectors corresponding to a large number of 00:36:37.000 --> 00:36:41.000 tumors, do you think you would be able to sort out that they actually 00:36:41.000 --> 00:36:49.000 fell into two clusters? 00:36:49.000 --> 00:36:53.000 Could you by computer tell that there's one class and the other 00:36:53.000 --> 00:36:57.000 class? Turns out that you can. Now, I've made it a little easier 00:36:57.000 --> 00:37:02.000 by not listing most of the 22,000 columns here. 00:37:02.000 --> 00:37:06.000 But think about it. Every tumor is a point in 22, 00:37:06.000 --> 00:37:10.000 00 dimensional space. If some of the tumors are similar, 00:37:10.000 --> 00:37:14.000 what can you say about those points in 22,000 dimensional space? 00:37:14.000 --> 00:37:18.000 They're going to be clumped together. They're near each other. 00:37:18.000 --> 00:37:22.000 So, just plot every tumor as a point in 22,000 dimensional space, 00:37:22.000 --> 00:37:26.000 and your question is, do the points tend to lie in two clumps up in 22, 00:37:26.000 --> 00:37:30.000 00 dimensional space? And there's simple arithmetic you 00:37:30.000 --> 00:37:34.000 can learn using linear algebra to get some separating hyperplane and 00:37:34.000 --> 00:37:38.000 ask, do tumors lie on one side or the other? And, 00:37:38.000 --> 00:37:42.000 it turns out the procedures like that will quickly tell you that 00:37:42.000 --> 00:37:46.000 these tumors clump into two very clear clumps. They're not randomly 00:37:46.000 --> 00:37:50.000 distributed. And so, if you get these tumors, 00:37:50.000 --> 00:37:54.000 and you do gene expression on them and put the data into a computer, 00:37:54.000 --> 00:37:58.000 the amount of time it takes the computer to discover that there were 00:37:58.000 --> 00:38:02.000 actually two types of acute leukemia is about three seconds marked down 00:38:02.000 --> 00:38:06.000 from 40 years. That's good. So, you can reproduce the discovery 00:38:06.000 --> 00:38:10.000 of AML and ALL in three seconds. Now you know what the pathologists 00:38:10.000 --> 00:38:14.000 say about this. They say, oh, give me a break. 00:38:14.000 --> 00:38:18.000 It's shooting fish in a barrel. We know there was a distinction. 00:38:18.000 --> 00:38:22.000 Big deal that the computer can find the distinction. 00:38:22.000 --> 00:38:26.000 We knew that there was distinction there. I know the computer didn't 00:38:26.000 --> 00:38:30.000 know it and all that. Tell us something we don't know. 00:38:30.000 --> 00:38:35.000 That's a fair question. So it turns out that you can ask 00:38:35.000 --> 00:38:40.000 some more questions. You can say, suppose I take now 00:38:40.000 --> 00:38:45.000 just the ALL's. Are they a homogeneous class, 00:38:45.000 --> 00:38:50.000 or did they fall into two classes? It turns out that extending this 00:38:50.000 --> 00:38:55.000 work, folks here were able to show that we can further split that ALL 00:38:55.000 --> 00:39:00.000 class. There was a hint that you might be able to do so because 00:39:00.000 --> 00:39:06.000 there's some ALL patients who have disruptions of a gene called MLL. 00:39:06.000 --> 00:39:09.000 And this tends to be a little more common in infants, 00:39:09.000 --> 00:39:13.000 and tends to be associated with a poor prognosis. 00:39:13.000 --> 00:39:16.000 But it was really very unclear whether this was simply one of a 00:39:16.000 --> 00:39:20.000 zillion factoids about some leukemia patients, whether this was a 00:39:20.000 --> 00:39:24.000 fundamental distinction. So, what happened was folks took a 00:39:24.000 --> 00:39:27.000 lot of ALL patients, got their expression profiles, 00:39:27.000 --> 00:39:31.000 and lo and behold it turned out that ALL itself broke into two very 00:39:31.000 --> 00:39:34.000 different clusters. This is an artist's rendition of a 00:39:34.000 --> 00:39:38.000 22,000 dimensional space. We can't afford a 22,000 00:39:38.000 --> 00:39:42.000 dimensional projector here, so we're just using two dimensions. 00:39:42.000 --> 00:39:46.000 But, the two forms of ALL were quite distinct from each other, 00:39:46.000 --> 00:39:50.000 and so actually ALL itself should be split up into two classes, 00:39:50.000 --> 00:39:54.000 ALL plus and minus, or ALL one and two, or MLL and ALL. 00:39:54.000 --> 00:39:58.000 And it turns out that these forms are quite different. 00:39:58.000 --> 00:40:02.000 They have different outcomes and should be treated differently. 00:40:02.000 --> 00:40:07.000 It also turns out that a particularly good distinction 00:40:07.000 --> 00:40:12.000 between these two subtypes of ALL is found by looking at this particular 00:40:12.000 --> 00:40:17.000 gene called the flit-3 kinase. The flit-3 kinase gene, whatever 00:40:17.000 --> 00:40:23.000 that is, was of great interest because people know that they can 00:40:23.000 --> 00:40:28.000 make inhibitors against certain kinases. And so, 00:40:28.000 --> 00:40:33.000 it turned out that an inhibitor against flit-3 kinases, 00:40:33.000 --> 00:40:39.000 against this flit-3 kinase gene product. 00:40:39.000 --> 00:40:44.000 If you treat cells with that inhibitor, cells of this type die, 00:40:44.000 --> 00:40:49.000 and cells of this type are not affected. So in fact, 00:40:49.000 --> 00:40:54.000 there's a potential drug use of flit-3 kinases in the MLL class of 00:40:54.000 --> 00:41:00.000 these leukemias, and folks are trying some clinical 00:41:00.000 --> 00:41:05.000 trials now. So, not only did the analysis of the 00:41:05.000 --> 00:41:09.000 gene expression point to two important sub-types of leukemias, 00:41:09.000 --> 00:41:14.000 but the analysis of the gene expression even suggested potential 00:41:14.000 --> 00:41:19.000 targets for therapy. So, I'll give you a bunch more 00:41:19.000 --> 00:41:23.000 examples. I have a bunch more examples like that there. 00:41:23.000 --> 00:41:28.000 They are examples of taking lymphomas and showing that they can 00:41:28.000 --> 00:41:33.000 be split into two different categories, examples of taking 00:41:33.000 --> 00:41:38.000 breast cancers into several categories, colon cancers. 00:41:38.000 --> 00:41:42.000 Basically what's going on right now is an attempt to reclassify cancers 00:41:42.000 --> 00:41:47.000 based not on what they look like in the microscope, 00:41:47.000 --> 00:41:51.000 and based not on what organ in the body they affect, 00:41:51.000 --> 00:41:56.000 but based on, molecularly, what their description is, because 00:41:56.000 --> 00:42:01.000 the molecular description, as Bob talked to you about with CML 00:42:01.000 --> 00:42:05.000 and with Gleveck, turns out to be a tremendously 00:42:05.000 --> 00:42:10.000 powerful way of classifying cancers because you're able to see what is 00:42:10.000 --> 00:42:15.000 the molecular defect and can make a molecular targeted therapy. 00:42:15.000 --> 00:42:20.000 So, these sorts of tools are quite cool, and I've got to say, 00:42:20.000 --> 00:42:25.000 in the last year we've begun using these expression tools not just to 00:42:25.000 --> 00:42:30.000 classify cancers, but to classify drugs. 00:42:30.000 --> 00:42:34.000 We've begun an interesting and somewhat crazy project to take all 00:42:34.000 --> 00:42:38.000 the FDA approved drugs, put them onto cell types, 00:42:38.000 --> 00:42:42.000 and see what they do, that is, get a signature, a fingerprint, 00:42:42.000 --> 00:42:46.000 a gene expression description of the action of a drug. 00:42:46.000 --> 00:42:50.000 And then we hope, here's the nutty idea, 00:42:50.000 --> 00:42:54.000 that we can look up in the computer which drugs do which things and 00:42:54.000 --> 00:42:58.000 might be useful for which diseases, because we'd put the diseases and 00:42:58.000 --> 00:43:02.000 the drugs on an equal footing. All of them would be described in 00:43:02.000 --> 00:43:06.000 terms of their gene expression patterns. So, 00:43:06.000 --> 00:43:10.000 I'll tell you one interesting example, OK? This is an interesting 00:43:10.000 --> 00:43:14.000 enough example. I don't even have slides for it yet. 00:43:14.000 --> 00:43:18.000 It turns out that these patients with ALL that I've been talking 00:43:18.000 --> 00:43:23.000 about, some of the patients with ALL will respond to the drug 00:43:23.000 --> 00:43:27.000 dexamethasone. Some won't. If you take patients 00:43:27.000 --> 00:43:31.000 who respond to dexamethasone, and patients who are resistant to 00:43:31.000 --> 00:43:35.000 dexamethasone, and you get their gene expression 00:43:35.000 --> 00:43:40.000 patterns, you can ask are there some genes that explain the difference? 00:43:40.000 --> 00:43:44.000 And you can get a certain gene signature, a list of, 00:43:44.000 --> 00:43:48.000 say, a dozen or so genes that do a pretty good job of classifying who's 00:43:48.000 --> 00:43:53.000 sensitive and who's resistant. Then you can go to this database I 00:43:53.000 --> 00:43:57.000 was telling you about of the action of many drugs and say, 00:43:57.000 --> 00:44:01.000 do we see any drugs whose effect would be to produce a signature 00:44:01.000 --> 00:44:06.000 of sensitivity? If we found a drug X, 00:44:06.000 --> 00:44:10.000 which when we put it on cells turned on those genes that correlate with 00:44:10.000 --> 00:44:14.000 being sensitive to dexamethasone, you could hallucinate the following 00:44:14.000 --> 00:44:18.000 really happy possibility that when you added that drug together with 00:44:18.000 --> 00:44:22.000 dexamethasone, you might be able to treat resistant 00:44:22.000 --> 00:44:26.000 patients because that drug could make them sensitive to dexamethasone, 00:44:26.000 --> 00:44:30.000 and that you could find that drug just by looking it up in 00:44:30.000 --> 00:44:35.000 a computer database. So, we tried it and we hit a drug. 00:44:35.000 --> 00:44:40.000 There was a certain drug that came up on the screen, 00:44:40.000 --> 00:44:45.000 yes? That's very much in the idea too. We found a drug that produced 00:44:45.000 --> 00:44:49.000 the signature sensitivity, and tested it in vitro. In vitro, 00:44:49.000 --> 00:44:54.000 if you take cells that are resistant and you add dexamethasone, 00:44:54.000 --> 00:44:59.000 nothing happens because they're resistant. If you add drug X, 00:44:59.000 --> 00:45:04.000 nothing happens. But if you add both drug X plus dexamethasone, 00:45:04.000 --> 00:45:08.000 the cells drop dead. It's now going into clinical trials 00:45:08.000 --> 00:45:12.000 in human patients. It turns out drug X is already a 00:45:12.000 --> 00:45:15.000 well FDA approved drug, so it can be tested in human 00:45:15.000 --> 00:45:19.000 patients right away, so it's going to be tested. 00:45:19.000 --> 00:45:22.000 So, the gene expression pattern was able to tell us to use a drug which 00:45:22.000 --> 00:45:26.000 actually had nothing to do with cancer uses in a cancer setting 00:45:26.000 --> 00:45:30.000 because it might do something helpful. 00:45:30.000 --> 00:45:33.000 Now, what's the point of all this? We can turn up the lights because I 00:45:33.000 --> 00:45:37.000 think I'm going to stop the slides there. The point of all of this, 00:45:37.000 --> 00:45:41.000 which is what I've made again, and I will make again, 00:45:41.000 --> 00:45:45.000 because you are the generation that's going to really live this, 00:45:45.000 --> 00:45:48.000 is that biology is becoming information. Now, 00:45:48.000 --> 00:45:52.000 don't get me wrong. It's not stopping being 00:45:52.000 --> 00:45:56.000 biochemistry. It's going to be biochemistry. It's not stopping 00:45:56.000 --> 00:46:00.000 being molecular biology. It's not stopping any of the things 00:46:00.000 --> 00:46:03.000 it was before. 45:57 But it is also becoming 00:46:03.000 --> 00:46:07.000 information, that for the first time we're entering a world where we can 00:46:07.000 --> 00:46:11.000 collect vast amounts of information: all the genetic variants in a 00:46:11.000 --> 00:46:15.000 patient, all of the gene expression pattern in a cell, 00:46:15.000 --> 00:46:18.000 or all of the gene expression pattern induced by a drug, 00:46:18.000 --> 00:46:22.000 and that whatever question you're asking will be informed by being 00:46:22.000 --> 00:46:26.000 able to access that whole database. In no way does it decrease the role 00:46:26.000 --> 00:46:30.000 of the individual smart scientist working on his or her problem. 00:46:30.000 --> 00:46:32.000 To the contrary, the goal is to empower the 00:46:32.000 --> 00:46:35.000 individual smart scientist so that you have all of that information at 00:46:35.000 --> 00:46:38.000 your fingertips. There are databases scattered 00:46:38.000 --> 00:46:41.000 around the web that have sequences from different species, 00:46:41.000 --> 00:46:44.000 variations from the human population, all of these drug database, 00:46:44.000 --> 00:46:47.000 etc., etc., etc., etc. It's a time of tremendous ferment, 00:46:47.000 --> 00:46:50.000 a little bit of chaos. You talk to people in the field, 00:46:50.000 --> 00:46:53.000 they say, we're getting deluged by data. We're getting crushed by the 00:46:53.000 --> 00:46:56.000 amount of data. I don't' know what to do with all 00:46:56.000 --> 00:46:59.000 the data. There's only one solution for a 00:46:59.000 --> 00:47:02.000 field in that condition, and that is young scientists because 00:47:02.000 --> 00:47:05.000 the young scientists who come into the field are the ones who take for 00:47:05.000 --> 00:47:08.000 granted, of course we're going to have all these data. 00:47:08.000 --> 00:47:11.000 We love having all these data. This is just great, couldn't be 00:47:11.000 --> 00:47:14.000 happier to have all these data. We're not put off by it in the 00:47:14.000 --> 00:47:17.000 least. That's what's going on. That's what's so important about 00:47:17.000 --> 00:47:20.000 your generation, and that's why I think it's really 00:47:20.000 --> 00:47:23.000 important that even though it's 701 and we're supposed to be teaching 00:47:23.000 --> 00:47:26.000 you the basics, it's important that you see this 00:47:26.000 --> 00:47:29.000 stuff because this is the change that's going on, 00:47:29.000 --> 00:47:32.000 and we're counting on this very much to drive a revolution in health, 00:47:32.000 --> 00:47:35.000 a revolution in biomedical research, and we're counting on you guys very 00:47:35.000 --> 00:47:39.000 much to drive that revolution. It has been a pleasure to teach you 00:47:39.000 --> 00:47:43.000 this term. I hope many of you will stay in touch, 00:47:43.000 --> 00:47:48.000 and some of you will go into biology, and even those of you who don't will 00:47:48.000 --> 00:47:53.000 know lots about it and enjoy it. Thank you very much. [APPLAUSE]