WEBVTT 00:00:00.000 --> 00:00:00.610 align:middle line:90% 00:00:00.610 --> 00:00:02.440 align:middle line:84% The following content is provided 00:00:02.440 --> 00:00:06.340 align:middle line:84% by MIT OpenCourseWare under a Creative Commons license. 00:00:06.340 --> 00:00:08.470 align:middle line:84% Additional information about our license 00:00:08.470 --> 00:00:10.870 align:middle line:84% and MIT OpenCourseWare in general 00:00:10.870 --> 00:00:13.090 align:middle line:90% is available at ocw.mit.edu. 00:00:13.090 --> 00:00:19.090 align:middle line:90% 00:00:19.090 --> 00:00:22.210 align:middle line:84% GEORGE CHURCH: Use this slide to review 00:00:22.210 --> 00:00:24.400 align:middle line:84% how far the in-situ might go, and what 00:00:24.400 --> 00:00:26.640 align:middle line:84% its current limitations might be. 00:00:26.640 --> 00:00:29.650 align:middle line:84% And then we'll move on to arrays. 00:00:29.650 --> 00:00:33.730 align:middle line:84% One potential advantage to these sort of microscopic in-situ 00:00:33.730 --> 00:00:37.600 align:middle line:84% analyses is that if you use a non-destructive visualization 00:00:37.600 --> 00:00:40.450 align:middle line:84% rather than fixing the cells-- you actually monitor them real 00:00:40.450 --> 00:00:41.200 align:middle line:90% time-- 00:00:41.200 --> 00:00:43.450 align:middle line:84% you can sample this basically as quickly 00:00:43.450 --> 00:00:47.680 align:middle line:84% as a modern microscopic camera system can monitor, 00:00:47.680 --> 00:00:49.600 align:middle line:90% on the order of a millisecond. 00:00:49.600 --> 00:00:53.590 align:middle line:84% You can obtain a sensitivity on the order of a single molecule 00:00:53.590 --> 00:00:56.770 align:middle line:90% of fluorescence. 00:00:56.770 --> 00:00:58.190 align:middle line:90% This is very challenging. 00:00:58.190 --> 00:01:03.190 align:middle line:84% It requires very small pixel sizes, but it is possible. 00:01:03.190 --> 00:01:05.440 align:middle line:84% And it's the basis of some of the sequencing methods 00:01:05.440 --> 00:01:08.630 align:middle line:84% that we discussed a couple of classes ago. 00:01:08.630 --> 00:01:12.520 align:middle line:84% And so that resolution itself is typically 00:01:12.520 --> 00:01:15.310 align:middle line:84% on the order of a micron or a quarter of a micron, 00:01:15.310 --> 00:01:17.800 align:middle line:84% sort of set by the limit of the optics 00:01:17.800 --> 00:01:22.750 align:middle line:84% in terms of the diffraction that can typically occur. 00:01:22.750 --> 00:01:25.240 align:middle line:84% But you can get below that diffraction limit 00:01:25.240 --> 00:01:29.740 align:middle line:84% of 250 nanometers down as low as 10 nanometers 00:01:29.740 --> 00:01:34.570 align:middle line:84% by using tricks such as near-field optics 00:01:34.570 --> 00:01:44.530 align:middle line:84% and various deconvolution methods. 00:01:44.530 --> 00:01:49.690 align:middle line:84% Multiplicity is really the greatest limitation 00:01:49.690 --> 00:01:52.960 align:middle line:84% of the in-situ method right now, and it's certainly 00:01:52.960 --> 00:01:56.440 align:middle line:84% an opportunity for the creative ones in this group to address. 00:01:56.440 --> 00:01:58.150 align:middle line:84% How can we get through multiplicity, 00:01:58.150 --> 00:02:01.540 align:middle line:84% looking at all of the RNA simultaneously 00:02:01.540 --> 00:02:04.360 align:middle line:84% as we can do in microarrays, which is essentially 00:02:04.360 --> 00:02:09.020 align:middle line:84% a microscopic method as well, But still have 00:02:09.020 --> 00:02:12.190 align:middle line:84% the spatial advantages of an in-situ? 00:02:12.190 --> 00:02:14.020 align:middle line:90% This is an unsolved problem. 00:02:14.020 --> 00:02:17.770 align:middle line:84% Multiplicity now is typically around one or two or three 00:02:17.770 --> 00:02:18.910 align:middle line:90% colors. 00:02:18.910 --> 00:02:23.290 align:middle line:84% The colors can be deconvolved by using band filters. 00:02:23.290 --> 00:02:25.120 align:middle line:84% If you use combinations of colors, 00:02:25.120 --> 00:02:27.010 align:middle line:84% you can conveniently discriminate 00:02:27.010 --> 00:02:31.520 align:middle line:84% the 24 different types of human metaphase chromosomes. 00:02:31.520 --> 00:02:33.670 align:middle line:84% However, this depends, you should note-- 00:02:33.670 --> 00:02:37.370 align:middle line:84% don't get fooled into thinking you actually have 24 colors. 00:02:37.370 --> 00:02:39.273 align:middle line:84% These are combinations of ratios of color, 00:02:39.273 --> 00:02:41.440 align:middle line:84% considering them to be false-colored by the computer 00:02:41.440 --> 00:02:42.370 align:middle line:90% algorithm. 00:02:42.370 --> 00:02:46.270 align:middle line:84% But they depend on non-overlap or of being 00:02:46.270 --> 00:02:52.000 align:middle line:84% able to find objects in the visual field and extend them. 00:02:52.000 --> 00:02:55.180 align:middle line:84% If where they overlap, you now have mixtures of mixtures, 00:02:55.180 --> 00:02:58.990 align:middle line:84% this no longer is simply deconvolved. 00:02:58.990 --> 00:03:01.070 align:middle line:84% So for all and practical purposes, 00:03:01.070 --> 00:03:03.475 align:middle line:84% we're limited to around four or five colors. 00:03:03.475 --> 00:03:06.080 align:middle line:90% 00:03:06.080 --> 00:03:10.480 align:middle line:84% So in situ, I would not call immediately 00:03:10.480 --> 00:03:11.860 align:middle line:90% genomics-compatible. 00:03:11.860 --> 00:03:14.740 align:middle line:84% The systems biology, it takes a vast number 00:03:14.740 --> 00:03:17.950 align:middle line:84% of in-situ experiments to get the kind of comprehensive data 00:03:17.950 --> 00:03:19.970 align:middle line:84% you can get out of microarray experiments. 00:03:19.970 --> 00:03:24.040 align:middle line:84% So let's focus on the kind of experiments like microarrays 00:03:24.040 --> 00:03:29.530 align:middle line:84% that can get us full genome scale information, 00:03:29.530 --> 00:03:32.740 align:middle line:84% and what are the limitations in quality 00:03:32.740 --> 00:03:33.820 align:middle line:90% for these sort of things. 00:03:33.820 --> 00:03:37.330 align:middle line:84% Now, we can either lump or split various ways 00:03:37.330 --> 00:03:39.800 align:middle line:90% of measuring arrays in cells. 00:03:39.800 --> 00:03:46.330 align:middle line:84% The top two items on slide 28 either 00:03:46.330 --> 00:03:49.690 align:middle line:84% can be characterized-- microarrays might be associated 00:03:49.690 --> 00:03:52.630 align:middle line:84% with, sort of for historical reasons, 00:03:52.630 --> 00:03:57.100 align:middle line:84% longer probes, maybe the length of an entire gene 00:03:57.100 --> 00:04:00.610 align:middle line:90% or entire cDNA, messenger RNA. 00:04:00.610 --> 00:04:04.960 align:middle line:84% And affymetrix and other oligonucleotide-based methods 00:04:04.960 --> 00:04:09.640 align:middle line:84% typically use 25-nucleotide-long ligament oligomers. 00:04:09.640 --> 00:04:15.340 align:middle line:84% Typically, the long microarray probes 00:04:15.340 --> 00:04:18.670 align:middle line:84% are used as single probes, one probe per gene, 00:04:18.670 --> 00:04:21.380 align:middle line:84% while the short ones typically have 24 probes per gene. 00:04:21.380 --> 00:04:24.820 align:middle line:84% These are not necessary differences. 00:04:24.820 --> 00:04:26.710 align:middle line:84% You can imagine various combinations. 00:04:26.710 --> 00:04:29.950 align:middle line:84% Another difference is in the long probes, 00:04:29.950 --> 00:04:32.163 align:middle line:84% typically, you'll do an experiment and control 00:04:32.163 --> 00:04:33.580 align:middle line:84% with different colors, and they're 00:04:33.580 --> 00:04:37.870 align:middle line:84% mixed together to control for some of the variations that 00:04:37.870 --> 00:04:39.100 align:middle line:90% might occur. 00:04:39.100 --> 00:04:41.290 align:middle line:84% In the spotting of the long probes, 00:04:41.290 --> 00:04:43.090 align:middle line:84% this is typically spotted mechanically, 00:04:43.090 --> 00:04:46.090 align:middle line:84% while the short probes are developed 00:04:46.090 --> 00:04:52.120 align:middle line:84% by a photochemical method in which 100 mass, sort 00:04:52.120 --> 00:04:57.940 align:middle line:84% of black and white mass such as use in the Silicon Valley 00:04:57.940 --> 00:05:00.670 align:middle line:84% for making computer chips, which we introduced 00:05:00.670 --> 00:05:05.410 align:middle line:84% in the sequencing technology lecture, these kind of 100 00:05:05.410 --> 00:05:07.810 align:middle line:84% photomass will allow you to make 25-mer, four 00:05:07.810 --> 00:05:11.110 align:middle line:90% possibilities per base. 00:05:11.110 --> 00:05:13.960 align:middle line:84% And you'll make, say, 20 of these scattered 00:05:13.960 --> 00:05:18.610 align:middle line:84% along the gene, and a mismatch control for each of those 20 00:05:18.610 --> 00:05:22.150 align:middle line:84% perfect matches, PM and MM abbreviation. 00:05:22.150 --> 00:05:26.140 align:middle line:84% Those mismatched controls help you get at the possible cross 00:05:26.140 --> 00:05:28.600 align:middle line:84% hybridization by related sequences or even 00:05:28.600 --> 00:05:30.370 align:middle line:90% distantly related sequences. 00:05:30.370 --> 00:05:36.130 align:middle line:84% And then what they typically do is subtract the mismatch 00:05:36.130 --> 00:05:38.110 align:middle line:84% controls from the perfect masses, 00:05:38.110 --> 00:05:41.560 align:middle line:84% and then average across all 20 of them, 00:05:41.560 --> 00:05:46.660 align:middle line:84% or some statistically good sampling of those 20. 00:05:46.660 --> 00:05:47.160 align:middle line:90% OK. 00:05:47.160 --> 00:05:50.340 align:middle line:84% So you typically do ratios for the long probes, 00:05:50.340 --> 00:05:54.450 align:middle line:84% and you try to get absolute amounts from the short probes. 00:05:54.450 --> 00:05:56.550 align:middle line:84% And then there are two wildly different methods 00:05:56.550 --> 00:05:58.710 align:middle line:90% in the bottom of the slide. 00:05:58.710 --> 00:06:02.310 align:middle line:84% These are called SAGE, standing for Serial Analysis of Gene 00:06:02.310 --> 00:06:07.680 align:middle line:84% Expression, and NPSS, an acronym for a highly 00:06:07.680 --> 00:06:09.810 align:middle line:90% parallel bead-based method. 00:06:09.810 --> 00:06:12.870 align:middle line:84% Both of them essentially sequence, determine 00:06:12.870 --> 00:06:17.070 align:middle line:84% the sequence of somewhere between 14 and 22 nucleotides. 00:06:17.070 --> 00:06:21.600 align:middle line:84% This is the sort of minimum length sequence that's 00:06:21.600 --> 00:06:27.660 align:middle line:84% often but not always sufficient to identify an RNA molecule. 00:06:27.660 --> 00:06:30.270 align:middle line:84% You're basically counting individual RNA molecules 00:06:30.270 --> 00:06:31.980 align:middle line:84% with a tag that's just long enough 00:06:31.980 --> 00:06:34.170 align:middle line:84% to be able to recognize it in a database. 00:06:34.170 --> 00:06:37.770 align:middle line:84% At a 14-mer, you can recognize it and say, a human cDNA 00:06:37.770 --> 00:06:39.930 align:middle line:84% database, but it's not unique enough 00:06:39.930 --> 00:06:42.360 align:middle line:84% to identify it in a human genomic database. 00:06:42.360 --> 00:06:48.510 align:middle line:84% The 22-mer is large enough to get an acceptable rate 00:06:48.510 --> 00:06:50.250 align:middle line:84% of false positives, false negatives 00:06:50.250 --> 00:06:53.537 align:middle line:90% in a human genomic library. 00:06:53.537 --> 00:06:55.620 align:middle line:84% So that's kind of the range which you can do this. 00:06:55.620 --> 00:07:00.690 align:middle line:84% And it's just, people tend to take shorter tags 00:07:00.690 --> 00:07:03.810 align:middle line:84% because the cost goes up with the length of the tag. 00:07:03.810 --> 00:07:06.300 align:middle line:84% And so these were conveniently short tags. 00:07:06.300 --> 00:07:09.870 align:middle line:84% So the top ones, you get quantitation 00:07:09.870 --> 00:07:12.930 align:middle line:84% by integration of the fluorescent signals. 00:07:12.930 --> 00:07:15.150 align:middle line:84% And the bottom two, you get quantitation 00:07:15.150 --> 00:07:19.080 align:middle line:90% by counting individual tags. 00:07:19.080 --> 00:07:23.430 align:middle line:84% The bottom two methods have the opportunity for discovery, 00:07:23.430 --> 00:07:25.620 align:middle line:84% while the top three, you basically 00:07:25.620 --> 00:07:30.690 align:middle line:84% can quantitate any gene or segment of the genome 00:07:30.690 --> 00:07:33.630 align:middle line:84% that you care to put on the array. 00:07:33.630 --> 00:07:36.240 align:middle line:84% But you won't necessarily discover anything outside 00:07:36.240 --> 00:07:37.275 align:middle line:90% of those features. 00:07:37.275 --> 00:07:40.100 align:middle line:90% 00:07:40.100 --> 00:07:43.010 align:middle line:84% So these are four of the key methods that 00:07:43.010 --> 00:07:44.870 align:middle line:84% are used for quantitating RNAs right 00:07:44.870 --> 00:07:48.620 align:middle line:84% now on a genome scale, where you'll do hopefully 00:07:48.620 --> 00:07:51.410 align:middle line:84% multiple experiments for each type. 00:07:51.410 --> 00:07:53.760 align:middle line:84% Now, let's just zoom in a little bit more 00:07:53.760 --> 00:07:55.520 align:middle line:84% so you can have appreciation for where 00:07:55.520 --> 00:07:57.650 align:middle line:84% some of the systematic and random errors 00:07:57.650 --> 00:08:00.080 align:middle line:84% might creep in to these kind of experiments. 00:08:00.080 --> 00:08:05.630 align:middle line:84% And I'll just arbitrarily use the 25 oligonucleotide probe 00:08:05.630 --> 00:08:11.120 align:middle line:84% arrays as an example for the long arrays, the microarrays. 00:08:11.120 --> 00:08:15.890 align:middle line:84% You might have, say, 1,000-nucleotide-long probes. 00:08:15.890 --> 00:08:19.790 align:middle line:84% You might have 10,000 of them on a glass slide about this big. 00:08:19.790 --> 00:08:22.400 align:middle line:90% 00:08:22.400 --> 00:08:25.550 align:middle line:84% With the photolithography, you can 00:08:25.550 --> 00:08:28.280 align:middle line:84% have a more like closer to a million 00:08:28.280 --> 00:08:31.280 align:middle line:90% features in a square centimeter. 00:08:31.280 --> 00:08:34.850 align:middle line:84% And in each of those features, each of those million 00:08:34.850 --> 00:08:37.940 align:middle line:84% positions on the array, in the square array, 00:08:37.940 --> 00:08:42.260 align:middle line:84% you will have maybe 10 to the 5th and 10 to the 6th molecules 00:08:42.260 --> 00:08:46.340 align:middle line:84% all identical in position 1, and then a new set of 10 00:08:46.340 --> 00:08:49.640 align:middle line:84% to the 6th molecules in position 2, all 00:08:49.640 --> 00:08:53.540 align:middle line:84% aimed at a different RNA or different part of an RNA. 00:08:53.540 --> 00:08:57.770 align:middle line:84% Each of those probe cells is ready to accept, 00:08:57.770 --> 00:09:03.530 align:middle line:84% fluorescently labeled or using biotin as an intermediate 00:09:03.530 --> 00:09:05.240 align:middle line:90% in order to get fluorescence. 00:09:05.240 --> 00:09:07.700 align:middle line:84% So you take your RNA and you directly [? biotinylate ?] it, 00:09:07.700 --> 00:09:10.550 align:middle line:90% or you make a cDNA copy. 00:09:10.550 --> 00:09:12.950 align:middle line:84% Or in one way or another, you introduce 00:09:12.950 --> 00:09:18.410 align:middle line:84% a fluorescent or bioitn molecule into a copy of your RNA. 00:09:18.410 --> 00:09:20.840 align:middle line:84% And then you apply that to the chip, 00:09:20.840 --> 00:09:25.430 align:middle line:90% and they will bind kinetically. 00:09:25.430 --> 00:09:29.900 align:middle line:84% And the more the mass action that you 00:09:29.900 --> 00:09:36.650 align:middle line:84% get from the original RNA, the RNAs that are the most abundant 00:09:36.650 --> 00:09:40.430 align:middle line:84% will result in the largest number of biotins 00:09:40.430 --> 00:09:45.230 align:middle line:84% or fluorescent molecules on the array at a given element. 00:09:45.230 --> 00:09:48.920 align:middle line:84% Here, an indirect conjugate with a fluorescent 00:09:48.920 --> 00:09:51.560 align:middle line:84% stripped out of [INAUDIBLE] It's like a rift in this covalently 00:09:51.560 --> 00:09:53.960 align:middle line:84% attached [INAUDIBLE],, and you get a fluorescent signal, 00:09:53.960 --> 00:09:55.502 align:middle line:84% which you quantitate, which tells you 00:09:55.502 --> 00:09:58.640 align:middle line:84% the amount of original messenger RNA. 00:09:58.640 --> 00:10:01.430 align:middle line:90% 00:10:01.430 --> 00:10:05.330 align:middle line:84% If you have 20 different oligonucleotides per gene, 00:10:05.330 --> 00:10:07.100 align:middle line:84% you can scatter this about the array 00:10:07.100 --> 00:10:08.420 align:middle line:90% or you can have them in lines. 00:10:08.420 --> 00:10:12.670 align:middle line:84% This tends to be back from when they had them in lines. 00:10:12.670 --> 00:10:14.588 align:middle line:84% So you get the streakiness there. 00:10:14.588 --> 00:10:16.130 align:middle line:84% Now, one of the first things you want 00:10:16.130 --> 00:10:18.680 align:middle line:84% to do, much of the software from the companies 00:10:18.680 --> 00:10:21.110 align:middle line:84% is set up on the assumption that you will only 00:10:21.110 --> 00:10:22.940 align:middle line:90% do the experiment once. 00:10:22.940 --> 00:10:26.690 align:middle line:84% Now, this may have been appealing 00:10:26.690 --> 00:10:29.210 align:middle line:84% in the early days from a cost standpoint, 00:10:29.210 --> 00:10:31.220 align:middle line:84% but it's not really cost-effective, 00:10:31.220 --> 00:10:33.290 align:middle line:84% in that you will make mistakes and you 00:10:33.290 --> 00:10:35.690 align:middle line:84% will draw incorrect conclusions that 00:10:35.690 --> 00:10:36.890 align:middle line:90% will require you to go back. 00:10:36.890 --> 00:10:40.610 align:middle line:84% But this is an example of an early experiment 00:10:40.610 --> 00:10:43.220 align:middle line:84% to establish the reproducibility from one experiment 00:10:43.220 --> 00:10:46.640 align:middle line:84% to another, possibly to reassure people that they didn't 00:10:46.640 --> 00:10:49.470 align:middle line:90% need to repeat the experiment. 00:10:49.470 --> 00:10:51.320 align:middle line:84% But in any case, this is the thing 00:10:51.320 --> 00:10:52.940 align:middle line:84% that is now commonly done in order 00:10:52.940 --> 00:10:56.360 align:middle line:84% to assess that your experiments are indeed reproducible. 00:10:56.360 --> 00:10:58.430 align:middle line:84% And what you expect from this as you go along 00:10:58.430 --> 00:11:04.130 align:middle line:84% the horizontal axis towards higher and higher copies 00:11:04.130 --> 00:11:09.020 align:middle line:84% per cell, going off to the right or going up 00:11:09.020 --> 00:11:12.770 align:middle line:84% on the vertical axis, when you get high copies per cell, 00:11:12.770 --> 00:11:19.250 align:middle line:84% then you expect there to be very close similarity in the two 00:11:19.250 --> 00:11:21.650 align:middle line:84% measures from two different experiments done 00:11:21.650 --> 00:11:22.710 align:middle line:90% at two different days. 00:11:22.710 --> 00:11:25.520 align:middle line:84% And then as you get to the very rare transcripts, 00:11:25.520 --> 00:11:27.980 align:middle line:84% you expect the various sources of noise 00:11:27.980 --> 00:11:31.940 align:middle line:84% in the experiment to start to dominate the light scattering 00:11:31.940 --> 00:11:33.410 align:middle line:90% and the array. 00:11:33.410 --> 00:11:36.350 align:middle line:84% The background fluorescence of the glass, 00:11:36.350 --> 00:11:39.410 align:middle line:84% the non-specific cross-hybridization 00:11:39.410 --> 00:11:42.507 align:middle line:84% between different RNAs, start to dominate over the true signal 00:11:42.507 --> 00:11:44.090 align:middle line:84% because, the true signal is going down 00:11:44.090 --> 00:11:46.340 align:middle line:84% and all those background signals are staying constant. 00:11:46.340 --> 00:11:48.575 align:middle line:84% So you start to get spread at a low number 00:11:48.575 --> 00:11:50.570 align:middle line:90% of copy numbers per cell. 00:11:50.570 --> 00:11:54.200 align:middle line:84% You can see a huge fraction of the RNAs 00:11:54.200 --> 00:11:59.540 align:middle line:84% in the yeast cells that present a single copy, say one or less 00:11:59.540 --> 00:12:01.550 align:middle line:90% fewer RNAs per cell. 00:12:01.550 --> 00:12:04.970 align:middle line:84% Now, this can either indicate that most 00:12:04.970 --> 00:12:08.930 align:middle line:84% of the RNAs in the cell are not physiologically significant, 00:12:08.930 --> 00:12:13.490 align:middle line:84% or it could indicate that all it takes is a small burst of one 00:12:13.490 --> 00:12:15.650 align:middle line:84% or a few molecules of RNA to produce 00:12:15.650 --> 00:12:17.720 align:middle line:84% an even larger burst of proteins and even 00:12:17.720 --> 00:12:19.980 align:middle line:84% larger bursts of activities of those proteins. 00:12:19.980 --> 00:12:21.630 align:middle line:90% So you get this amplification. 00:12:21.630 --> 00:12:23.540 align:middle line:84% And so the stochastics that we will 00:12:23.540 --> 00:12:25.580 align:middle line:84% study in the systems biology part of this 00:12:25.580 --> 00:12:28.770 align:middle line:84% becomes a more significant consideration. 00:12:28.770 --> 00:12:35.420 align:middle line:84% So looking at one molecule per cell, 00:12:35.420 --> 00:12:38.440 align:middle line:84% it's important to start thinking about what 00:12:38.440 --> 00:12:43.960 align:middle line:84% the implications of that might be for the systems biology, 00:12:43.960 --> 00:12:46.990 align:middle line:84% and asking, can we accurately measure it down there, 00:12:46.990 --> 00:12:49.900 align:middle line:84% and do we believe that it's biologically significant. 00:12:49.900 --> 00:12:53.110 align:middle line:84% Now, there's a whole variety of microarray data analyses, 00:12:53.110 --> 00:12:59.950 align:middle line:84% ranging from the very hardware-oriented first data 00:12:59.950 --> 00:13:02.500 align:middle line:84% acquisition modules all the way up 00:13:02.500 --> 00:13:09.700 align:middle line:84% through analyzing single array data at a statistical level, 00:13:09.700 --> 00:13:14.320 align:middle line:84% to multiple related experiments, such as the one we showed 00:13:14.320 --> 00:13:15.910 align:middle line:84% in the previous slide, all the way up 00:13:15.910 --> 00:13:20.500 align:middle line:84% to clustering multiple examples from multiple different 00:13:20.500 --> 00:13:26.590 align:middle line:84% conditions to start asking the biological questions about why 00:13:26.590 --> 00:13:28.660 align:middle line:90% RNAs go up and down together. 00:13:28.660 --> 00:13:32.980 align:middle line:84% For intermediate analyses, where we'll be talking today, 00:13:32.980 --> 00:13:38.650 align:middle line:84% as introductory issues of data analysis, 00:13:38.650 --> 00:13:43.060 align:middle line:84% I'll illustrate dChip and a couple of other tools 00:13:43.060 --> 00:13:47.530 align:middle line:84% that indicate how reproducible experiments can be, 00:13:47.530 --> 00:13:50.770 align:middle line:84% and the kind of systematic errors that can creep in. 00:13:50.770 --> 00:13:54.310 align:middle line:84% The reproducibility helps you by repeating, helps 00:13:54.310 --> 00:13:56.080 align:middle line:90% you reduce the random errors. 00:13:56.080 --> 00:13:59.350 align:middle line:84% And here are four papers recently 00:13:59.350 --> 00:14:04.000 align:middle line:84% that talk about measurements from multiple measures 00:14:04.000 --> 00:14:07.360 align:middle line:84% from the same experiment, or multiple measures by using 00:14:07.360 --> 00:14:09.850 align:middle line:84% two completely different microarray technologies. 00:14:09.850 --> 00:14:12.520 align:middle line:84% And I urge you to take a look at these. 00:14:12.520 --> 00:14:16.760 align:middle line:90% 00:14:16.760 --> 00:14:20.210 align:middle line:84% When we compare two distributions 00:14:20.210 --> 00:14:25.550 align:middle line:84% from microarray experiments, you can think of these. 00:14:25.550 --> 00:14:27.800 align:middle line:84% Even if they're not perfectly normal distributions, 00:14:27.800 --> 00:14:30.050 align:middle line:84% they're going to be roughly bell-shaped shaped curves. 00:14:30.050 --> 00:14:31.760 align:middle line:84% So let's say that this is experiment 1 00:14:31.760 --> 00:14:33.260 align:middle line:90% and this is experiment 2. 00:14:33.260 --> 00:14:35.090 align:middle line:90% You say, oh, they look the same. 00:14:35.090 --> 00:14:37.550 align:middle line:84% This is experiment 1 under condition 1. 00:14:37.550 --> 00:14:39.920 align:middle line:90% This is under condition 2. 00:14:39.920 --> 00:14:41.120 align:middle line:90% OK, now they look different. 00:14:41.120 --> 00:14:42.650 align:middle line:90% But how do you quantitate that? 00:14:42.650 --> 00:14:47.270 align:middle line:84% And the way you ask that is, the means of those two roughly 00:14:47.270 --> 00:14:50.600 align:middle line:84% bell-shaped distributions are far apart from one other. 00:14:50.600 --> 00:14:51.830 align:middle line:90% How far apart? 00:14:51.830 --> 00:14:54.410 align:middle line:84% Well, they're farther apart from one another 00:14:54.410 --> 00:14:57.320 align:middle line:84% than the width of the distributions individually. 00:14:57.320 --> 00:15:00.790 align:middle line:84% And the distance between them, you 00:15:00.790 --> 00:15:02.810 align:middle line:84% can think of as the mean of the difference 00:15:02.810 --> 00:15:05.330 align:middle line:90% of the distributions. 00:15:05.330 --> 00:15:09.860 align:middle line:84% And then the width is a measure of the root 00:15:09.860 --> 00:15:12.770 align:middle line:84% mean square standard deviation, so [? versus ?] 00:15:12.770 --> 00:15:14.455 align:middle line:90% the combined width of the two. 00:15:14.455 --> 00:15:16.580 align:middle line:84% If one of them is wide and the other one is narrow, 00:15:16.580 --> 00:15:19.370 align:middle line:84% you have to have some way of combining those. 00:15:19.370 --> 00:15:22.220 align:middle line:84% So that's sometimes called a student t-test. 00:15:22.220 --> 00:15:25.520 align:middle line:84% And the t statistic itself is simply 00:15:25.520 --> 00:15:28.880 align:middle line:84% the mean over the standard deviation. 00:15:28.880 --> 00:15:32.210 align:middle line:84% In other words, how many standard deviation widths apart 00:15:32.210 --> 00:15:33.320 align:middle line:90% are these two means? 00:15:33.320 --> 00:15:35.870 align:middle line:84% Or if you take the mean of the difference, 00:15:35.870 --> 00:15:37.850 align:middle line:84% take the distribution of the difference, 00:15:37.850 --> 00:15:40.220 align:middle line:84% then you want your null hypothesis. 00:15:40.220 --> 00:15:43.930 align:middle line:84% H0 here on slide 33 is the null hypothesis. 00:15:43.930 --> 00:15:46.160 align:middle line:84% If the mean value of the difference is 0, 00:15:46.160 --> 00:15:48.500 align:middle line:84% there's no difference between the two distributions. 00:15:48.500 --> 00:15:52.910 align:middle line:84% If you can rule that out, then that 00:15:52.910 --> 00:15:54.945 align:middle line:90% would be the point of this test. 00:15:54.945 --> 00:15:56.570 align:middle line:84% So you can think, how many widths apart 00:15:56.570 --> 00:15:59.630 align:middle line:84% are the means of these two distributions. 00:15:59.630 --> 00:16:02.060 align:middle line:84% Now, this requires that actually, 00:16:02.060 --> 00:16:04.100 align:middle line:84% the distributions be very close to normal, 00:16:04.100 --> 00:16:06.950 align:middle line:84% not distinguishable from normal distribution 00:16:06.950 --> 00:16:09.170 align:middle line:90% with all its properties. 00:16:09.170 --> 00:16:11.535 align:middle line:84% If you are in serious doubt or can 00:16:11.535 --> 00:16:13.160 align:middle line:84% prove that they're not normal, then you 00:16:13.160 --> 00:16:15.440 align:middle line:90% should go to a non-parametric. 00:16:15.440 --> 00:16:16.720 align:middle line:90% Normal means it's parametric. 00:16:16.720 --> 00:16:18.290 align:middle line:84% It has a mean and standard deviation 00:16:18.290 --> 00:16:22.070 align:middle line:90% that characterize it well. 00:16:22.070 --> 00:16:23.713 align:middle line:84% Then you can use a non-parametric. 00:16:23.713 --> 00:16:25.130 align:middle line:84% Whenever you see the word "ranks," 00:16:25.130 --> 00:16:27.440 align:middle line:84% that's a tip-off that you're going into something where 00:16:27.440 --> 00:16:29.090 align:middle line:90% you're making fewer assumptions. 00:16:29.090 --> 00:16:30.380 align:middle line:90% This has lower power. 00:16:30.380 --> 00:16:33.150 align:middle line:84% That means you might miss some significant differences. 00:16:33.150 --> 00:16:36.890 align:middle line:84% But on the other hand, if you can convince yourself 00:16:36.890 --> 00:16:40.977 align:middle line:84% with the Wilcoxon matched pair sign ranks test, 00:16:40.977 --> 00:16:43.310 align:middle line:84% then you don't need to worry about whether it's normally 00:16:43.310 --> 00:16:44.300 align:middle line:90% distributed. 00:16:44.300 --> 00:16:51.620 align:middle line:84% In any case, we're going to look at some distributions 00:16:51.620 --> 00:16:57.320 align:middle line:84% and ask informally whether these are the same distribution 00:16:57.320 --> 00:16:58.970 align:middle line:90% or different. 00:16:58.970 --> 00:17:00.442 align:middle line:90% Yes. 00:17:00.442 --> 00:17:01.980 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:17:01.980 --> 00:17:28.930 align:middle line:90% 00:17:28.930 --> 00:17:32.040 align:middle line:84% GEORGE CHURCH: So the question is, 00:17:32.040 --> 00:17:36.050 align:middle line:84% how do you deal with the multiple hypothesis testing. 00:17:36.050 --> 00:17:40.160 align:middle line:84% And this basically is exactly the same answer 00:17:40.160 --> 00:17:44.030 align:middle line:84% that we would have given in the last lecture 00:17:44.030 --> 00:17:49.310 align:middle line:84% on multiple hypothesis testing in genotyping. 00:17:49.310 --> 00:17:52.010 align:middle line:84% If you apply exactly the same-- it's 00:17:52.010 --> 00:17:56.270 align:middle line:84% a very good question, very appropriate here. 00:17:56.270 --> 00:17:57.950 align:middle line:84% Just as before, where you would have 00:17:57.950 --> 00:18:02.780 align:middle line:84% multiple different phenotype-genotype combinations 00:18:02.780 --> 00:18:05.300 align:middle line:84% that you might want to test, essentially 00:18:05.300 --> 00:18:07.685 align:middle line:84% testing every possible single nucleotide polymorphism 00:18:07.685 --> 00:18:15.080 align:middle line:84% or combination in the genome, to a first approximation, whatever 00:18:15.080 --> 00:18:19.070 align:middle line:84% your significance is, it needs to be that much more 00:18:19.070 --> 00:18:22.230 align:middle line:84% significant if you have that many hypotheses. 00:18:22.230 --> 00:18:23.710 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:18:23.710 --> 00:18:28.370 align:middle line:90% 00:18:28.370 --> 00:18:31.850 align:middle line:84% GEORGE CHURCH: You either have to improve your data. 00:18:31.850 --> 00:18:33.980 align:middle line:84% Allows you to test more hypotheses. 00:18:33.980 --> 00:18:38.150 align:middle line:84% Or you need to reduce the hypothesis, that number 00:18:38.150 --> 00:18:43.040 align:middle line:84% at the outset, by having a sharp biological question 00:18:43.040 --> 00:18:43.760 align:middle line:90% at the beginning. 00:18:43.760 --> 00:18:45.385 align:middle line:84% It's an excellent question, but there's 00:18:45.385 --> 00:18:49.700 align:middle line:84% no magic wand except those two that I know of. 00:18:49.700 --> 00:18:54.198 align:middle line:84% OK, so here's some examples of independent experiments. 00:18:54.198 --> 00:18:56.240 align:middle line:84% Now, when someone says an independent experiment, 00:18:56.240 --> 00:18:59.620 align:middle line:84% you have to be clear about, is it 00:18:59.620 --> 00:19:06.950 align:middle line:84% that the same RNA sample split and then labeled independently. 00:19:06.950 --> 00:19:09.420 align:middle line:84% That's really not an independent experiment. 00:19:09.420 --> 00:19:13.250 align:middle line:84% On the other hand, you could take two completely, 00:19:13.250 --> 00:19:16.910 align:middle line:84% where you repeated the best of an independent experiment. 00:19:16.910 --> 00:19:19.340 align:middle line:84% If your objective is to ask how reproducible 00:19:19.340 --> 00:19:21.050 align:middle line:84% is the entire biological phenomenon, 00:19:21.050 --> 00:19:27.050 align:middle line:84% you should go back as early as possible, make a new cell line, 00:19:27.050 --> 00:19:30.320 align:middle line:84% try to get the conditions exactly the same, 00:19:30.320 --> 00:19:32.810 align:middle line:84% but completely independently executed, 00:19:32.810 --> 00:19:34.220 align:middle line:84% possibly by different researchers 00:19:34.220 --> 00:19:36.560 align:middle line:90% in different laboratories. 00:19:36.560 --> 00:19:40.040 align:middle line:84% In that extreme, you expect to have more scatter. 00:19:40.040 --> 00:19:41.810 align:middle line:84% Here, these are the regression lines. 00:19:41.810 --> 00:19:45.560 align:middle line:84% The R squared is the number that pops out 00:19:45.560 --> 00:19:48.053 align:middle line:84% as an indication of deviation from the linear, 00:19:48.053 --> 00:19:50.720 align:middle line:84% just like the linear correlation coefficient, which is basically 00:19:50.720 --> 00:19:53.330 align:middle line:90% a squared term. 00:19:53.330 --> 00:19:56.270 align:middle line:84% You can see that as instead of splitting one sample, 00:19:56.270 --> 00:19:58.610 align:middle line:84% and doing kind of a trivial differentiable labeling, 00:19:58.610 --> 00:20:00.740 align:middle line:84% if you have more independent samples, 00:20:00.740 --> 00:20:04.460 align:middle line:84% you get more scatter and a lower figure of merit 00:20:04.460 --> 00:20:06.860 align:middle line:90% for the regression line. 00:20:06.860 --> 00:20:12.200 align:middle line:84% OK, now, what are the guidelines for-- 00:20:12.200 --> 00:20:13.820 align:middle line:84% what are some of the considerations 00:20:13.820 --> 00:20:15.080 align:middle line:90% in RNA quantitation? 00:20:15.080 --> 00:20:16.940 align:middle line:84% I think we've touched upon this before, 00:20:16.940 --> 00:20:18.560 align:middle line:84% but I just want to drive it home, 00:20:18.560 --> 00:20:20.540 align:middle line:84% that some people will say, I'm only 00:20:20.540 --> 00:20:24.800 align:middle line:84% going to look at things that are more than a three-fold effect. 00:20:24.800 --> 00:20:28.340 align:middle line:84% This is sort of the ratio limits that you 00:20:28.340 --> 00:20:32.360 align:middle line:84% might perceive in the early RNA [? TIP ?] experiments. 00:20:32.360 --> 00:20:34.140 align:middle line:84% But I think we're getting better at it 00:20:34.140 --> 00:20:36.870 align:middle line:84% and the biological motivation is high. 00:20:36.870 --> 00:20:41.210 align:middle line:84% We've seen that human trisomies, where you just have a 1.5-fold 00:20:41.210 --> 00:20:44.210 align:middle line:84% increase in dosage, every single one of them has a huge 00:20:44.210 --> 00:20:45.500 align:middle line:90% phenotypic consequence. 00:20:45.500 --> 00:20:49.700 align:middle line:84% Many of them result in lethality. 00:20:49.700 --> 00:20:53.840 align:middle line:84% We should set as a goal to be able to monitor most 00:20:53.840 --> 00:20:57.740 align:middle line:84% of the RNAs of biological significance down to this 00:20:57.740 --> 00:21:03.740 align:middle line:84% 1.5-fold effect, which can have these dramatic implications. 00:21:03.740 --> 00:21:09.320 align:middle line:84% We mentioned the Oligonucleotides, 00:21:09.320 --> 00:21:11.300 align:middle line:84% we might be able to get more of them per gene. 00:21:11.300 --> 00:21:14.000 align:middle line:84% How cam we utilize this, not only 00:21:14.000 --> 00:21:16.700 align:middle line:84% the number that we can get, but the specificity? 00:21:16.700 --> 00:21:21.020 align:middle line:84% If you have a gene length oligonucleotide, or cDNA, 00:21:21.020 --> 00:21:24.470 align:middle line:84% then you're going to pick up not only the gene of interest, 00:21:24.470 --> 00:21:28.460 align:middle line:84% but every related gene, all the alternative splice forms, all 00:21:28.460 --> 00:21:30.900 align:middle line:84% the very, very close family members. 00:21:30.900 --> 00:21:33.680 align:middle line:84% So with oligonucleotides, you can then go and target 00:21:33.680 --> 00:21:35.810 align:middle line:84% individual splice forms, but then 00:21:35.810 --> 00:21:38.660 align:middle line:84% when you apply your algorithms, you 00:21:38.660 --> 00:21:40.790 align:middle line:84% have to be careful not to lump them all together 00:21:40.790 --> 00:21:41.732 align:middle line:90% as if it's one gene. 00:21:41.732 --> 00:21:43.190 align:middle line:84% You have to say, OK, this is splice 00:21:43.190 --> 00:21:44.780 align:middle line:90% form number one, number two. 00:21:44.780 --> 00:21:49.130 align:middle line:84% And just having oligonucleotides aimed at particular exons 00:21:49.130 --> 00:21:52.820 align:middle line:84% is not sufficient to tell you which exons 00:21:52.820 --> 00:21:54.800 align:middle line:90% insist in particular RNAs. 00:21:54.800 --> 00:22:00.620 align:middle line:84% You can have present in the population exons 1, 2, 4, 6 12, 00:22:00.620 --> 00:22:01.378 align:middle line:90% and so forth. 00:22:01.378 --> 00:22:03.920 align:middle line:84% But you don't know whether 1 and 12 are on the same molecule, 00:22:03.920 --> 00:22:04.420 align:middle line:90% however. 00:22:04.420 --> 00:22:09.110 align:middle line:84% That requires a more specialized method, possibly 00:22:09.110 --> 00:22:10.100 align:middle line:90% high-throughput method. 00:22:10.100 --> 00:22:13.070 align:middle line:90% 00:22:13.070 --> 00:22:18.140 align:middle line:84% There is another set of economic forces pushing towards just 00:22:18.140 --> 00:22:21.500 align:middle line:90% doing a subset of the genome. 00:22:21.500 --> 00:22:25.172 align:middle line:84% Just like not repeating the experiment, 00:22:25.172 --> 00:22:27.380 align:middle line:84% you probably don't want to give in to economic forces 00:22:27.380 --> 00:22:29.880 align:middle line:84% unless you absolutely have to, because if one person studies 00:22:29.880 --> 00:22:35.590 align:middle line:84% a cancer subset, another studies a blood-related subset, 00:22:35.590 --> 00:22:40.682 align:middle line:84% and another one studies these little pieces of the genome, 00:22:40.682 --> 00:22:42.640 align:middle line:84% then when they want to pool their data in order 00:22:42.640 --> 00:22:45.610 align:middle line:84% to ask questions about what genes are cluster together 00:22:45.610 --> 00:22:47.860 align:middle line:84% because they're in their proliferative cells, 00:22:47.860 --> 00:22:49.900 align:middle line:84% and which ones cluster together because they're 00:22:49.900 --> 00:22:52.822 align:middle line:84% in this developmental stage or another, 00:22:52.822 --> 00:22:54.280 align:middle line:84% they can't do it because they don't 00:22:54.280 --> 00:22:55.960 align:middle line:84% share enough genes on their arrays 00:22:55.960 --> 00:22:58.070 align:middle line:90% to do this meta analysis. 00:22:58.070 --> 00:23:00.610 align:middle line:84% So that's a consideration when you're 00:23:00.610 --> 00:23:02.135 align:middle line:84% in the experimental design phase. 00:23:02.135 --> 00:23:03.760 align:middle line:84% And hopefully, computational biologists 00:23:03.760 --> 00:23:06.130 align:middle line:84% are involved not only in the interpretation of the data, 00:23:06.130 --> 00:23:09.530 align:middle line:84% but in the design of the experiments as well. 00:23:09.530 --> 00:23:12.970 align:middle line:84% Here's yet another way of looking at the variation 00:23:12.970 --> 00:23:14.590 align:middle line:90% that you have in the experiment. 00:23:14.590 --> 00:23:18.350 align:middle line:84% We're introducing, I think, the coefficient of variation here, 00:23:18.350 --> 00:23:21.610 align:middle line:84% which is simply the standard deviation normalized 00:23:21.610 --> 00:23:22.750 align:middle line:90% to the mean. 00:23:22.750 --> 00:23:24.930 align:middle line:90% So you can just phrase it. 00:23:24.930 --> 00:23:28.690 align:middle line:84% It's a way of sharing, in a generic sense, 00:23:28.690 --> 00:23:30.100 align:middle line:90% how much variation you have. 00:23:30.100 --> 00:23:34.570 align:middle line:84% So you can say the coefficient of variation is, say, 10%. 00:23:34.570 --> 00:23:40.750 align:middle line:84% And that's independent of what units you're measuring. 00:23:40.750 --> 00:23:46.300 align:middle line:84% And so we have is on the horizontal axis, the x-axis 00:23:46.300 --> 00:23:48.790 align:middle line:84% here, number of messenger RNAs per cell, 00:23:48.790 --> 00:23:52.010 align:middle line:84% and in the vertical axis, the coefficient of variation. 00:23:52.010 --> 00:23:58.090 align:middle line:84% And you can see that when you get up above, say, 00:23:58.090 --> 00:24:07.950 align:middle line:84% 20% coefficient of variation, you getting less trustworthy, 00:24:07.950 --> 00:24:10.560 align:middle line:84% because here, we've used the algorithms that 00:24:10.560 --> 00:24:12.690 align:middle line:84% are built in to the [INAUDIBLE] software 00:24:12.690 --> 00:24:15.870 align:middle line:84% for asking whether it thinks an RNA is present 00:24:15.870 --> 00:24:18.030 align:middle line:84% or not if the intensity is very low, 00:24:18.030 --> 00:24:19.980 align:middle line:90% and a variety of other criteria. 00:24:19.980 --> 00:24:24.270 align:middle line:84% For a single experiment, it will classify whether it thinks 00:24:24.270 --> 00:24:25.680 align:middle line:90% the RNA is present or not. 00:24:25.680 --> 00:24:29.640 align:middle line:84% But if you use a large number of different experiments-- 00:24:29.640 --> 00:24:33.480 align:middle line:84% each of these dots being a different RNA-- 00:24:33.480 --> 00:24:35.310 align:middle line:84% you use a large number of experiments, 00:24:35.310 --> 00:24:38.128 align:middle line:84% you can now beat the company software, 00:24:38.128 --> 00:24:40.170 align:middle line:84% because it's made the assumption that you're just 00:24:40.170 --> 00:24:41.650 align:middle line:90% doing one experiment. 00:24:41.650 --> 00:24:45.210 align:middle line:84% And so here in the dark blue are examples 00:24:45.210 --> 00:24:47.610 align:middle line:84% where in 3 experiments, all three experiments 00:24:47.610 --> 00:24:50.880 align:middle line:90% was called f one by one. 00:24:50.880 --> 00:24:54.660 align:middle line:84% But you can see that even with cases 00:24:54.660 --> 00:24:57.150 align:middle line:84% where it's not called present in any of the three 00:24:57.150 --> 00:25:00.300 align:middle line:84% experiments, these magenta ones, you can still 00:25:00.300 --> 00:25:03.900 align:middle line:84% find very high reproducibility, that 00:25:03.900 --> 00:25:07.350 align:middle line:84% is to say, very low coefficient of variation, down around 10%. 00:25:07.350 --> 00:25:10.920 align:middle line:84% There's some pink dots all in this region around 10%, 00:25:10.920 --> 00:25:14.250 align:middle line:84% and these are just as reliable as the blue dots. 00:25:14.250 --> 00:25:16.890 align:middle line:84% Even though they're not called present by the software, 00:25:16.890 --> 00:25:18.972 align:middle line:84% collectively, they're very reproducible, 00:25:18.972 --> 00:25:20.430 align:middle line:84% and therefore, they're trustworthy. 00:25:20.430 --> 00:25:23.430 align:middle line:84% So actually, reproducing your experiment 00:25:23.430 --> 00:25:32.730 align:middle line:84% is not just something you do to appease nature and lower 00:25:32.730 --> 00:25:33.960 align:middle line:90% your statistical noise. 00:25:33.960 --> 00:25:37.290 align:middle line:84% It actually allows you to get data for RNAs that otherwise 00:25:37.290 --> 00:25:39.420 align:middle line:90% might be inaccessible. 00:25:39.420 --> 00:25:42.570 align:middle line:84% So there's immediate gratification there, 00:25:42.570 --> 00:25:44.055 align:middle line:90% even at the slight expense. 00:25:44.055 --> 00:25:46.600 align:middle line:90% 00:25:46.600 --> 00:25:49.410 align:middle line:84% So now let's broaden back out a little bit 00:25:49.410 --> 00:25:53.460 align:middle line:84% on a number of different methods and their advantages 00:25:53.460 --> 00:25:54.300 align:middle line:90% and disadvantages. 00:25:54.300 --> 00:25:56.730 align:middle line:84% Each one has a set of advantages. 00:25:56.730 --> 00:25:58.810 align:middle line:84% We've already talked about two of them, 00:25:58.810 --> 00:26:04.020 align:middle line:84% which is the immobilized genes, labeled RNA scenario. 00:26:04.020 --> 00:26:07.350 align:middle line:84% That's basically the microarrays or chips. 00:26:07.350 --> 00:26:11.730 align:middle line:84% And the advantage here is that in a very high throughput 00:26:11.730 --> 00:26:14.310 align:middle line:84% manner, you can manufacture large numbers of these. 00:26:14.310 --> 00:26:16.320 align:middle line:84% And you can get high multiplicity, 00:26:16.320 --> 00:26:20.922 align:middle line:84% all the RNAs that we know of monitored simultaneously. 00:26:20.922 --> 00:26:23.550 align:middle line:84% The in-situ, we've also talked about. 00:26:23.550 --> 00:26:29.160 align:middle line:84% The major advantage is retaining the spatial relationships. 00:26:29.160 --> 00:26:33.450 align:middle line:84% Some of these other methods, if instead of immobilizing 00:26:33.450 --> 00:26:36.660 align:middle line:84% the probes on a solid surface, you immobilize 00:26:36.660 --> 00:26:41.580 align:middle line:84% the RNAs, and then one by one, you label the probes, 00:26:41.580 --> 00:26:47.010 align:middle line:84% this will allow you to first, say, separate 00:26:47.010 --> 00:26:51.660 align:middle line:84% the entire transcriptome of RNAs in electrophoretic separation. 00:26:51.660 --> 00:26:53.820 align:middle line:84% And so in a highly parallel method, 00:26:53.820 --> 00:26:55.920 align:middle line:84% you've now immobilized them after they've 00:26:55.920 --> 00:26:57.070 align:middle line:90% been separated by size. 00:26:57.070 --> 00:26:59.250 align:middle line:84% So if you want to know the size of the RNA, which 00:26:59.250 --> 00:27:06.090 align:middle line:84% is a big hint as to its exon composition and so on, 00:27:06.090 --> 00:27:09.000 align:middle line:84% measuring the size of RNA, this is one of the few ways 00:27:09.000 --> 00:27:10.290 align:middle line:90% to do it. 00:27:10.290 --> 00:27:14.010 align:middle line:84% Very hard to do with arrays or in situ. 00:27:14.010 --> 00:27:16.080 align:middle line:90% OK. 00:27:16.080 --> 00:27:18.570 align:middle line:84% If, on the other hand, you want sensitivity, 00:27:18.570 --> 00:27:22.560 align:middle line:84% where you want to really detect at the noise level, which say, 00:27:22.560 --> 00:27:25.860 align:middle line:84% for mammalian veterbrate RNAs is around 10 00:27:25.860 --> 00:27:28.500 align:middle line:84% to the minus 4 copies per cell, that's 00:27:28.500 --> 00:27:30.330 align:middle line:84% the level at which if you look for almost 00:27:30.330 --> 00:27:33.090 align:middle line:84% any part of the genome, any kind of RNA, 00:27:33.090 --> 00:27:35.010 align:middle line:84% even things that shouldn't be expressed, 00:27:35.010 --> 00:27:37.530 align:middle line:84% you will find them down in the minus 4 [? per cell. ?] 00:27:37.530 --> 00:27:39.960 align:middle line:84% That probably is not biologically significant, 00:27:39.960 --> 00:27:42.870 align:middle line:90% but it's a biological fact. 00:27:42.870 --> 00:27:46.920 align:middle line:84% Getting down to that level, or if you have a mixed tissue 00:27:46.920 --> 00:27:50.730 align:middle line:84% and you want to detect 1 part in 10 to the 10th, 00:27:50.730 --> 00:27:55.540 align:middle line:84% you might have 5 times 10 to the 5th messenger RNAs per cell. 00:27:55.540 --> 00:27:58.350 align:middle line:84% But if you have 10 to the 5th cells, 00:27:58.350 --> 00:28:02.770 align:middle line:84% then a single copy messenger RNA would be down to the minus 10. 00:28:02.770 --> 00:28:05.550 align:middle line:84% That's feasible with reverse transcriptase, 00:28:05.550 --> 00:28:07.680 align:middle line:84% quantitative reverse transcriptase, [INAUDIBLE].. 00:28:07.680 --> 00:28:12.865 align:middle line:84% And it has it is the standard which all the others can barely 00:28:12.865 --> 00:28:13.365 align:middle line:90% match. 00:28:13.365 --> 00:28:16.640 align:middle line:90% 00:28:16.640 --> 00:28:19.490 align:middle line:84% Reporter constructs are something 00:28:19.490 --> 00:28:22.250 align:middle line:84% we do not consider generally a high throughput 00:28:22.250 --> 00:28:25.430 align:middle line:84% method, although there are genomic constructs of reporter 00:28:25.430 --> 00:28:28.310 align:middle line:84% constructs for an entire genome like yeast. 00:28:28.310 --> 00:28:31.280 align:middle line:84% But here, the real advantage of that method 00:28:31.280 --> 00:28:33.560 align:middle line:84% is there's no worry of cross-hybridization. 00:28:33.560 --> 00:28:36.380 align:middle line:84% With in-situs, with northerns, with arrays, 00:28:36.380 --> 00:28:40.400 align:middle line:84% there is a chance that if you probe for RNA x, 00:28:40.400 --> 00:28:42.800 align:middle line:84% it will happen to hybridize, especially 00:28:42.800 --> 00:28:44.690 align:middle line:84% if it's present in high abundance, 00:28:44.690 --> 00:28:46.970 align:middle line:84% it will happen to hybridize to one or the other ones. 00:28:46.970 --> 00:28:48.590 align:middle line:84% But with a reporter construct, we 00:28:48.590 --> 00:28:53.720 align:middle line:84% will take a fluorescent protein array, a luminescent protein 00:28:53.720 --> 00:28:57.170 align:middle line:84% and hook it up to the gene, and insist with the gene 00:28:57.170 --> 00:28:59.060 align:middle line:90% you're interested in. 00:28:59.060 --> 00:29:01.610 align:middle line:84% And that will directly or indirectly 00:29:01.610 --> 00:29:06.380 align:middle line:84% monitor the expression of your favorite gene. 00:29:06.380 --> 00:29:10.370 align:middle line:84% That has no possibility for cross-hybridization. 00:29:10.370 --> 00:29:13.310 align:middle line:84% We've talked about the advantages of counting. 00:29:13.310 --> 00:29:16.070 align:middle line:84% The disadvantage is of course cost. 00:29:16.070 --> 00:29:18.110 align:middle line:84% It allows you to do gene discovery. 00:29:18.110 --> 00:29:20.810 align:middle line:84% It doesn't address alternative splicing. 00:29:20.810 --> 00:29:26.800 align:middle line:90% 00:29:26.800 --> 00:29:29.470 align:middle line:84% Here's an example of comparing two of those methods. 00:29:29.470 --> 00:29:32.530 align:middle line:84% As microarrays are being introduced, 00:29:32.530 --> 00:29:36.940 align:middle line:84% one needs to validate them to ask whether you're measuring 00:29:36.940 --> 00:29:39.430 align:middle line:84% one RNA or multiple RNAs of different sizes 00:29:39.430 --> 00:29:42.490 align:middle line:84% to ask whether quantifying a northern blot 00:29:42.490 --> 00:29:44.680 align:middle line:84% correlates with quantitating an array. 00:29:44.680 --> 00:29:50.830 align:middle line:84% And here, you can see a fairly acceptable linear relationship 00:29:50.830 --> 00:29:52.810 align:middle line:84% between the two quantitative measures. 00:29:52.810 --> 00:29:55.420 align:middle line:84% And this has been played out many times. 00:29:55.420 --> 00:29:59.830 align:middle line:90% 00:29:59.830 --> 00:30:03.550 align:middle line:84% The opportunity that you have when you make an array, 00:30:03.550 --> 00:30:08.290 align:middle line:84% we said that the SAGE and NPSS allow you to do discovery. 00:30:08.290 --> 00:30:14.080 align:middle line:84% But another way of doing it is putting down 00:30:14.080 --> 00:30:16.630 align:middle line:84% lots of oligonucleotides, even oligonucleotides 00:30:16.630 --> 00:30:19.120 align:middle line:84% in regions where your genome annotation may not have 00:30:19.120 --> 00:30:20.890 align:middle line:90% indicated that there's a gene. 00:30:20.890 --> 00:30:24.430 align:middle line:84% So you can see here the bottom 60% of this array 00:30:24.430 --> 00:30:27.070 align:middle line:84% was in so-called non-protein-coding regions. 00:30:27.070 --> 00:30:29.560 align:middle line:84% And you can just see what you get when you do that. 00:30:29.560 --> 00:30:32.530 align:middle line:84% It doesn't cost that much more to put down some 00:30:32.530 --> 00:30:34.000 align:middle line:90% of these non-coding regions. 00:30:34.000 --> 00:30:39.400 align:middle line:84% And you can ask in these untranslated regions 00:30:39.400 --> 00:30:42.580 align:middle line:84% whether there are maybe antisense RNAs that 00:30:42.580 --> 00:30:44.800 align:middle line:84% will overlap in the translated regions. 00:30:44.800 --> 00:30:47.320 align:middle line:84% Or you can look for DNA protein interactions 00:30:47.320 --> 00:30:49.070 align:middle line:90% in certain kinds of experiments. 00:30:49.070 --> 00:30:50.920 align:middle line:84% And you can look for RNA fine structure. 00:30:50.920 --> 00:30:52.870 align:middle line:84% Where does the gene actually end? 00:30:52.870 --> 00:30:55.060 align:middle line:84% You may annotate that the RNA ends here, 00:30:55.060 --> 00:30:57.620 align:middle line:84% but you need ways of actually measuring them. 00:30:57.620 --> 00:31:01.390 align:middle line:84% So there's a lot of uses for nucleic acid probes 00:31:01.390 --> 00:31:05.140 align:middle line:84% in so-called non-protein-coding regions, which 00:31:05.140 --> 00:31:08.380 align:middle line:84% can range from 12% of the genome in simple prokaryotes 00:31:08.380 --> 00:31:13.710 align:middle line:84% to up to 98% of the genome in humans. 00:31:13.710 --> 00:31:17.120 align:middle line:84% So what are the sources of random systematic errors? 00:31:17.120 --> 00:31:20.240 align:middle line:84% We have a secondary structure that we 00:31:20.240 --> 00:31:22.670 align:middle line:84% talked about at the beginning of this lecture. 00:31:22.670 --> 00:31:25.130 align:middle line:84% It can cause different parts of the array 00:31:25.130 --> 00:31:28.280 align:middle line:84% to have different hybridization efficiencies. 00:31:28.280 --> 00:31:30.320 align:middle line:84% The position on our array to have an effect, 00:31:30.320 --> 00:31:34.640 align:middle line:84% for example, poor mixing, if you're making your array 00:31:34.640 --> 00:31:37.010 align:middle line:84% by a non-reproducible method, the amount 00:31:37.010 --> 00:31:41.270 align:middle line:84% of target nucleic acid immobilized on the array 00:31:41.270 --> 00:31:41.930 align:middle line:90% can vary. 00:31:41.930 --> 00:31:44.220 align:middle line:84% And you need to control for that, for example 00:31:44.220 --> 00:31:47.293 align:middle line:90% by having an internal standard. 00:31:47.293 --> 00:31:48.710 align:middle line:84% Cross-hybridization hybridization, 00:31:48.710 --> 00:31:50.900 align:middle line:90% we've talked about. 00:31:50.900 --> 00:31:52.430 align:middle line:84% The unanticipated transcripts, you 00:31:52.430 --> 00:31:54.560 align:middle line:84% can handle by tiling, by basically 00:31:54.560 --> 00:31:57.450 align:middle line:84% putting oligonucleotides throughout the genome. 00:31:57.450 --> 00:31:59.930 align:middle line:84% So here's an example of spatial effects. 00:31:59.930 --> 00:32:04.580 align:middle line:84% What you do is spike in known amounts of known RNAs 00:32:04.580 --> 00:32:06.980 align:middle line:84% which are present throughout the array. 00:32:06.980 --> 00:32:10.010 align:middle line:84% And so these are internally spiked in addition 00:32:10.010 --> 00:32:14.030 align:middle line:84% to your unknown fluorescently labeled arm probes. 00:32:14.030 --> 00:32:19.240 align:middle line:84% And you can ask whether you're getting a perfectly uniform, 00:32:19.240 --> 00:32:24.750 align:middle line:84% edge-to-edge hybridization with the known for the answer. 00:32:24.750 --> 00:32:29.030 align:middle line:84% And if you're getting peaks and troughs, 00:32:29.030 --> 00:32:31.370 align:middle line:84% then you can use these internal standards 00:32:31.370 --> 00:32:34.670 align:middle line:84% to calibrate that particular hybridization experiment 00:32:34.670 --> 00:32:38.065 align:middle line:84% and correct for this kind of systematic error. 00:32:38.065 --> 00:32:39.440 align:middle line:84% This could occur again and again. 00:32:39.440 --> 00:32:42.050 align:middle line:84% Here's two different experiments giving roughly similar edge 00:32:42.050 --> 00:32:43.400 align:middle line:90% effects. 00:32:43.400 --> 00:32:45.650 align:middle line:84% You need to account for these things 00:32:45.650 --> 00:32:48.830 align:middle line:84% to avoid that particular source of systematic errors, 00:32:48.830 --> 00:32:52.880 align:middle line:84% especially if you put all of your oligonucleotides 00:32:52.880 --> 00:32:54.620 align:middle line:84% for a particular gene near one another. 00:32:54.620 --> 00:32:58.550 align:middle line:84% A better strategy for statistical experimental design 00:32:58.550 --> 00:33:02.640 align:middle line:84% is to put your oligos randomly throughout the array. 00:33:02.640 --> 00:33:05.420 align:middle line:84% Here's another one, unanticipated RNAs. 00:33:05.420 --> 00:33:08.570 align:middle line:84% Two examples, one an open reading frame 00:33:08.570 --> 00:33:10.640 align:middle line:90% of unknown function. 00:33:10.640 --> 00:33:12.410 align:middle line:90% You can sometimes mis-annotate. 00:33:12.410 --> 00:33:15.500 align:middle line:84% If you have two open reading frames on opposite strands, 00:33:15.500 --> 00:33:17.970 align:middle line:84% it could be one is used, generally speaking. 00:33:17.970 --> 00:33:19.160 align:middle line:90% One is used and one isn't. 00:33:19.160 --> 00:33:20.270 align:middle line:84% And you could pick the wrong one. 00:33:20.270 --> 00:33:21.740 align:middle line:84% You might pick the big one, and it 00:33:21.740 --> 00:33:24.110 align:middle line:84% could be the little one is the one that's actually used. 00:33:24.110 --> 00:33:26.270 align:middle line:84% And that's what happened in this case. 00:33:26.270 --> 00:33:28.100 align:middle line:90% And another one is-- 00:33:28.100 --> 00:33:29.825 align:middle line:90% so that was a translated RNA. 00:33:29.825 --> 00:33:32.180 align:middle line:84% We just happened to pick the wrong strand. 00:33:32.180 --> 00:33:34.910 align:middle line:84% Here is an untranslated RNA, such as the snow 00:33:34.910 --> 00:33:36.440 align:middle line:90% RNAs we saw before. 00:33:36.440 --> 00:33:41.330 align:middle line:84% This one is an untranslated RNA that 00:33:41.330 --> 00:33:43.565 align:middle line:84% was discovered in a so-called intergenic region. 00:33:43.565 --> 00:33:49.200 align:middle line:90% 00:33:49.200 --> 00:33:53.620 align:middle line:84% If you have a statistical test for the goodness, 00:33:53.620 --> 00:33:59.970 align:middle line:84% the quality of an individual oligonucleotide hybridization, 00:33:59.970 --> 00:34:02.190 align:middle line:84% based on, say, its reproducibility 00:34:02.190 --> 00:34:07.560 align:middle line:84% or its relative intensity that you expect-- 00:34:07.560 --> 00:34:09.690 align:middle line:84% if you have 20 different oligonucleotides 00:34:09.690 --> 00:34:13.260 align:middle line:84% all for one gene, and you expect number 1 typically 00:34:13.260 --> 00:34:15.239 align:middle line:84% is stronger than number 2, and then 00:34:15.239 --> 00:34:18.370 align:middle line:84% you find a case where number 1 is weaker than number 2, 00:34:18.370 --> 00:34:19.860 align:middle line:90% then you can flag that. 00:34:19.860 --> 00:34:22.570 align:middle line:84% You can say, I don't believe that particular spot. 00:34:22.570 --> 00:34:25.060 align:middle line:90% And if you color-code them all-- 00:34:25.060 --> 00:34:28.920 align:middle line:84% see, here is white spots, things that 00:34:28.920 --> 00:34:33.034 align:middle line:84% don't fit your statistical model for the array hybridization. 00:34:33.034 --> 00:34:35.159 align:middle line:84% This is the advantage of having a statistical model 00:34:35.159 --> 00:34:37.710 align:middle line:90% of the entire process. 00:34:37.710 --> 00:34:40.500 align:middle line:84% Then you can mark those as white, 00:34:40.500 --> 00:34:42.090 align:middle line:84% and you can look to see whether they 00:34:42.090 --> 00:34:47.159 align:middle line:84% have a [? statistically ?] significant spatial 00:34:47.159 --> 00:34:48.909 align:middle line:84% distribution, which they do in this case. 00:34:48.909 --> 00:34:50.820 align:middle line:84% They all seem to be clumping at this corner. 00:34:50.820 --> 00:34:52.790 align:middle line:90% Now, what could cause that? 00:34:52.790 --> 00:34:54.540 align:middle line:84% Well, we've already illustrated that there 00:34:54.540 --> 00:34:58.200 align:middle line:84% are ways that you can use internal standards 00:34:58.200 --> 00:34:59.040 align:middle line:90% to calibrate. 00:34:59.040 --> 00:35:03.660 align:middle line:84% This was not a case where we had poor hybridization efficiency 00:35:03.660 --> 00:35:06.280 align:middle line:84% or strong hybridization efficiency around the edges. 00:35:06.280 --> 00:35:09.480 align:middle line:84% This was something where the alignment of the grid 00:35:09.480 --> 00:35:12.390 align:middle line:84% is done by these little squares along the edge, 00:35:12.390 --> 00:35:14.940 align:middle line:84% and the computer algorithm that finds 00:35:14.940 --> 00:35:18.600 align:middle line:84% these spots was distracted by this little spot 00:35:18.600 --> 00:35:21.540 align:middle line:84% off the side, which is not part of the checkerboard. 00:35:21.540 --> 00:35:26.940 align:middle line:84% And once you manually correct that error, now you snap in. 00:35:26.940 --> 00:35:29.550 align:middle line:84% On the right-hand side of slide 43 00:35:29.550 --> 00:35:34.950 align:middle line:84% is now the statistical model of this after getting 00:35:34.950 --> 00:35:36.600 align:middle line:90% the alignment right. 00:35:36.600 --> 00:35:40.650 align:middle line:84% Here, you had been associating the wrong oligos 00:35:40.650 --> 00:35:42.290 align:middle line:90% with a particular signal. 00:35:42.290 --> 00:35:43.290 align:middle line:90% It didn't fit the model. 00:35:43.290 --> 00:35:46.470 align:middle line:84% Now it fits the model, and you see the little scattered strips 00:35:46.470 --> 00:35:50.190 align:middle line:84% of gray where you have individual genes which 00:35:50.190 --> 00:35:54.300 align:middle line:84% are misbehaving, rather than the entire corner of the array. 00:35:54.300 --> 00:35:56.910 align:middle line:84% OK, so now we get the very interesting interpretation 00:35:56.910 --> 00:36:03.470 align:middle line:84% issues where we're using the same kind of information. 00:36:03.470 --> 00:36:07.370 align:middle line:84% Once you have a model, a very sophisticated model 00:36:07.370 --> 00:36:10.250 align:middle line:84% of how the individual oligos in the array 00:36:10.250 --> 00:36:14.960 align:middle line:84% behave, here, what we do is we take genomic DNA 00:36:14.960 --> 00:36:20.090 align:middle line:84% as an example of a fairly equimolar calibration standard. 00:36:20.090 --> 00:36:22.520 align:middle line:84% If you take genomic DNA and label it, 00:36:22.520 --> 00:36:25.340 align:middle line:84% you expect every segment in the genome 00:36:25.340 --> 00:36:28.460 align:middle line:84% to be present at the same molarity, 00:36:28.460 --> 00:36:32.180 align:middle line:84% with the exception of repetitive elements, which we'll put aside 00:36:32.180 --> 00:36:33.180 align:middle line:90% for the moment. 00:36:33.180 --> 00:36:36.650 align:middle line:84% And so that means that any place, that is, 00:36:36.650 --> 00:36:39.410 align:middle line:84% any oligonucleotide that doesn't hybridize 00:36:39.410 --> 00:36:41.660 align:middle line:84% with the genomic DNA, such as these ones that go 00:36:41.660 --> 00:36:44.240 align:middle line:90% close to baseline here at 0-- 00:36:44.240 --> 00:36:47.120 align:middle line:84% remember this is perfect mass versus [? mismatch ?] 00:36:47.120 --> 00:36:49.250 align:middle line:90% that we're plotting on this. 00:36:49.250 --> 00:36:52.220 align:middle line:84% When you get close to 0 for the genomic DNA in black, 00:36:52.220 --> 00:36:54.650 align:middle line:84% that means that really doesn't hybridize well. 00:36:54.650 --> 00:36:56.610 align:middle line:84% It's not that it's missing from the genome. 00:36:56.610 --> 00:36:59.542 align:middle line:84% It's that it has some secondary structure. 00:36:59.542 --> 00:37:01.250 align:middle line:84% So this is the secondary structure that's 00:37:01.250 --> 00:37:03.710 align:middle line:90% been a theme for this plot. 00:37:03.710 --> 00:37:09.680 align:middle line:84% And that sort of secondary structure 00:37:09.680 --> 00:37:13.070 align:middle line:84% is actually a piece of data that you can do data mining on. 00:37:13.070 --> 00:37:15.270 align:middle line:84% You can go through the entire genome 00:37:15.270 --> 00:37:17.567 align:middle line:84% and you can look for secondary structures. 00:37:17.567 --> 00:37:18.650 align:middle line:90% And you can ask for those. 00:37:18.650 --> 00:37:22.760 align:middle line:84% Secondary structures depend on what part of the genome 00:37:22.760 --> 00:37:23.870 align:middle line:90% is transcribed. 00:37:23.870 --> 00:37:25.590 align:middle line:90% Now, here is a messenger RNA. 00:37:25.590 --> 00:37:27.770 align:middle line:84% This is one of the few messenger RNAs 00:37:27.770 --> 00:37:30.470 align:middle line:84% for which you have a plausible secondary structure. 00:37:30.470 --> 00:37:33.200 align:middle line:84% Most secondary structures are on structural RNAs 00:37:33.200 --> 00:37:35.780 align:middle line:84% or enzyme-related RNAs, the ribosomal RNAs, 00:37:35.780 --> 00:37:37.080 align:middle line:90% [INAUDIBLE] RNAs, and so forth. 00:37:37.080 --> 00:37:41.270 align:middle line:84% This is the messenger RNA for this gene product, LTT. 00:37:41.270 --> 00:37:45.770 align:middle line:84% And if you look where this black arrow is 00:37:45.770 --> 00:37:50.330 align:middle line:84% coming from the right-hand 0 you'll see a long helix. 00:37:50.330 --> 00:37:53.600 align:middle line:84% And that helix is at the 3' end of the messenger RNA, 00:37:53.600 --> 00:37:57.110 align:middle line:84% and it's very well characterized both structurally and 00:37:57.110 --> 00:37:57.920 align:middle line:90% functionally. 00:37:57.920 --> 00:38:01.940 align:middle line:84% And it's known to be involved in at least one 00:38:01.940 --> 00:38:03.530 align:middle line:84% important biological process, which 00:38:03.530 --> 00:38:05.330 align:middle line:84% is the termination of transcription. 00:38:05.330 --> 00:38:07.850 align:middle line:84% When you get close to the end of the RNA, 00:38:07.850 --> 00:38:09.920 align:middle line:84% that [? pair ?] can reform, and it sends a signal 00:38:09.920 --> 00:38:13.040 align:middle line:84% to the transcription apparatus to stop. 00:38:13.040 --> 00:38:15.710 align:middle line:84% So that is a believable hairpin of known function. 00:38:15.710 --> 00:38:19.400 align:middle line:84% And the interesting feature of this microarray 00:38:19.400 --> 00:38:21.410 align:middle line:84% is that's one of the places where 00:38:21.410 --> 00:38:24.950 align:middle line:84% both the genomic DNA in black and two completely 00:38:24.950 --> 00:38:27.770 align:middle line:84% different RNA samples fail to hybridize, 00:38:27.770 --> 00:38:34.070 align:middle line:84% consistent with it being a very strong hairpin with a dozen 00:38:34.070 --> 00:38:37.190 align:middle line:90% G, C and A base pairs. 00:38:37.190 --> 00:38:40.370 align:middle line:84% Another thing you can derive from this detailed model 00:38:40.370 --> 00:38:43.640 align:middle line:84% of the array, here, you have 60 different oligonucleotides 00:38:43.640 --> 00:38:46.895 align:middle line:84% along the gene and adjacent intergenic regions. 00:38:46.895 --> 00:38:50.540 align:middle line:84% The question is, where does the RNA transcription stop. 00:38:50.540 --> 00:38:57.270 align:middle line:84% Well, if you look in the places where the DNA control is high, 00:38:57.270 --> 00:39:01.880 align:middle line:84% you'll find the RNA is high, going from right to left. 00:39:01.880 --> 00:39:03.170 align:middle line:90% The RNA tracks the DNA. 00:39:03.170 --> 00:39:05.600 align:middle line:84% The red and blue tracks the black 00:39:05.600 --> 00:39:09.410 align:middle line:90% until you get to position -33. 00:39:09.410 --> 00:39:14.360 align:middle line:84% And there, the red and blue drop to baseline, and the Black 00:39:14.360 --> 00:39:17.870 align:middle line:84% stays going up and down at a higher level. 00:39:17.870 --> 00:39:21.170 align:middle line:84% And that happens to coincide with the known transcriptional 00:39:21.170 --> 00:39:21.710 align:middle line:90% start. 00:39:21.710 --> 00:39:23.720 align:middle line:84% And so that would be another way of mapping 00:39:23.720 --> 00:39:25.580 align:middle line:90% the transcriptional start. 00:39:25.580 --> 00:39:31.310 align:middle line:84% You'll notice that some of the hybridization intentially 00:39:31.310 --> 00:39:32.360 align:middle line:90% drops below 0. 00:39:32.360 --> 00:39:37.430 align:middle line:84% This is just an artifact of having the perfect match 00:39:37.430 --> 00:39:38.630 align:middle line:90% minus the mismatch. 00:39:38.630 --> 00:39:43.100 align:middle line:84% If it happens to be the case that your mismatch control is 00:39:43.100 --> 00:39:46.190 align:middle line:84% cross-reacting with some other DNA, 00:39:46.190 --> 00:39:48.830 align:middle line:84% say repetitive DNA in the genome or RNA, 00:39:48.830 --> 00:39:53.190 align:middle line:84% then it can get actually more intense than the perfect match. 00:39:53.190 --> 00:39:54.830 align:middle line:84% And so you can get a negative value. 00:39:54.830 --> 00:39:58.760 align:middle line:84% But otherwise, the negative intensity would be meaningless. 00:39:58.760 --> 00:40:01.520 align:middle line:90% Now, splice domains. 00:40:01.520 --> 00:40:04.800 align:middle line:84% In principle, you can go through the whole human genome 00:40:04.800 --> 00:40:06.920 align:middle line:84% and you predict where all the exons are, 00:40:06.920 --> 00:40:09.530 align:middle line:84% where all the splice junctions are, and in principle, even all 00:40:09.530 --> 00:40:10.640 align:middle line:90% the alternative splicing. 00:40:10.640 --> 00:40:13.110 align:middle line:90% In practice, is not that easy. 00:40:13.110 --> 00:40:16.670 align:middle line:84% And you can use all the hidden Markov models 00:40:16.670 --> 00:40:18.200 align:middle line:84% and so forth we've been developing. 00:40:18.200 --> 00:40:20.450 align:middle line:84% You can do multi-sequence alignments 00:40:20.450 --> 00:40:24.710 align:middle line:84% to get these motifs here, where two bits is a full scale. 00:40:24.710 --> 00:40:27.200 align:middle line:84% And you can find donors and acceptors 00:40:27.200 --> 00:40:30.560 align:middle line:84% in this kind of pattern, GT donor, AG acceptor. 00:40:30.560 --> 00:40:32.600 align:middle line:84% But when you come right down to it, 00:40:32.600 --> 00:40:35.900 align:middle line:84% you want to have some way of going through this empirically 00:40:35.900 --> 00:40:37.650 align:middle line:90% as well. 00:40:37.650 --> 00:40:44.315 align:middle line:84% And so what you can do is you can basically ignore or look 00:40:44.315 --> 00:40:48.290 align:middle line:84% sort of independently, do a tiling of the genome 00:40:48.290 --> 00:40:53.875 align:middle line:84% with oligonucleotides, as was done here by Shoemaker et al. 00:40:53.875 --> 00:40:56.000 align:middle line:84% And this was, I think, one of the nicer papers that 00:40:56.000 --> 00:41:03.400 align:middle line:84% came out in the Nature issue on the human genome [INAUDIBLE] 00:41:03.400 --> 00:41:04.360 align:middle line:90% sequence. 00:41:04.360 --> 00:41:06.820 align:middle line:84% Here, as the sequence was coming out, 00:41:06.820 --> 00:41:08.890 align:middle line:84% chromosome 22 was one of the first chromosomes 00:41:08.890 --> 00:41:11.980 align:middle line:90% nearly completed. 00:41:11.980 --> 00:41:17.380 align:middle line:84% At the top of slide 47, you see how the metaphase chromosome 00:41:17.380 --> 00:41:19.990 align:middle line:90% is banded and labeled. 00:41:19.990 --> 00:41:22.540 align:middle line:84% If you take a little [? 113-kit ?] kilobase chunk 00:41:22.540 --> 00:41:24.490 align:middle line:84% of that, it's the next line down. 00:41:24.490 --> 00:41:26.980 align:middle line:84% Then you blow that up further, and all 00:41:26.980 --> 00:41:29.320 align:middle line:84% the way down to oligonucleotide 60-mers, 00:41:29.320 --> 00:41:32.590 align:middle line:84% tiled every 10 base pairs as a starting point 00:41:32.590 --> 00:41:36.980 align:middle line:84% all the way along this 100-kilobit chromosome 22. 00:41:36.980 --> 00:41:39.130 align:middle line:84% And then you hybridize it with RNAs 00:41:39.130 --> 00:41:41.230 align:middle line:84% from a variety of different human tissues. 00:41:41.230 --> 00:41:45.410 align:middle line:84% Then you ask, in the vertical axis, 00:41:45.410 --> 00:41:48.160 align:middle line:84% what is the log of a normalized signal intensity 00:41:48.160 --> 00:41:50.200 align:middle line:90% for these various RNAs. 00:41:50.200 --> 00:41:51.970 align:middle line:84% And you'll get a little histogram here, 00:41:51.970 --> 00:41:55.420 align:middle line:84% where purple spike means there's a lot of hybridization 00:41:55.420 --> 00:41:57.250 align:middle line:84% under at least some of the conditions. 00:41:57.250 --> 00:41:59.320 align:middle line:84% And then there'll be a zone where there's almost 00:41:59.320 --> 00:42:00.280 align:middle line:90% no hybridization. 00:42:00.280 --> 00:42:04.180 align:middle line:84% And that's because those introns that we had, 00:42:04.180 --> 00:42:08.120 align:middle line:84% that we showed in the previous slide, are spliced out, 00:42:08.120 --> 00:42:10.150 align:middle line:90% and they're in low abundance. 00:42:10.150 --> 00:42:11.650 align:middle line:84% They're displaced out of the nucleus 00:42:11.650 --> 00:42:13.543 align:middle line:84% before the RNAs accumulate, so they 00:42:13.543 --> 00:42:14.710 align:middle line:90% tend to be in low abundance. 00:42:14.710 --> 00:42:18.040 align:middle line:84% And they're not found in the mature messenger RNA. 00:42:18.040 --> 00:42:20.890 align:middle line:84% And so when you label these up, you're 00:42:20.890 --> 00:42:23.380 align:middle line:84% selectively labeling the exons in [? CNA. ?] 00:42:23.380 --> 00:42:26.260 align:middle line:84% And if you can see, they coincide well 00:42:26.260 --> 00:42:29.115 align:middle line:84% with the little green exons in the annotation, 00:42:29.115 --> 00:42:30.490 align:middle line:84% except every now and then, you'll 00:42:30.490 --> 00:42:34.000 align:middle line:84% find something-- here's a case for exon 3 00:42:34.000 --> 00:42:38.230 align:middle line:84% where the green annotation in the original sequence 00:42:38.230 --> 00:42:39.550 align:middle line:90% is too short. 00:42:39.550 --> 00:42:42.970 align:middle line:84% And here's a blow-up near the bottom where the purple region 00:42:42.970 --> 00:42:46.360 align:middle line:84% clearly extends beyond the green annotation in the [INAUDIBLE] 00:42:46.360 --> 00:42:49.750 align:middle line:84% from sequence algorithm, sequence analysis 00:42:49.750 --> 00:42:51.910 align:middle line:84% algorithm, where 102 base pairs should 00:42:51.910 --> 00:42:54.880 align:middle line:84% be extended five times to that exon to make it 00:42:54.880 --> 00:42:58.150 align:middle line:90% a slightly larger exon. 00:42:58.150 --> 00:43:00.320 align:middle line:84% But when you extend it by that, you ask, 00:43:00.320 --> 00:43:02.117 align:middle line:84% well, does it still have the splice site, 00:43:02.117 --> 00:43:04.450 align:middle line:84% or does it have a new splice site that we can recognize. 00:43:04.450 --> 00:43:05.492 align:middle line:90% And sure enough, it does. 00:43:05.492 --> 00:43:08.740 align:middle line:84% It has an AG and a fairly good match to the motif 00:43:08.740 --> 00:43:11.320 align:middle line:90% we had in the previous slide. 00:43:11.320 --> 00:43:15.700 align:middle line:84% And you can see that the purple intensity drops close to 0 00:43:15.700 --> 00:43:19.750 align:middle line:84% here as soon as you get out of the exon as now properly 00:43:19.750 --> 00:43:20.780 align:middle line:90% defined. 00:43:20.780 --> 00:43:25.570 align:middle line:84% So this is a way of including additional data in addition 00:43:25.570 --> 00:43:31.210 align:middle line:84% to the sequence by tiling and by quantitative hybridization. 00:43:31.210 --> 00:43:34.210 align:middle line:90% 00:43:34.210 --> 00:43:36.660 align:middle line:84% Now, the last topic today is time series. 00:43:36.660 --> 00:43:39.300 align:middle line:84% This connects the quantitative data 00:43:39.300 --> 00:43:43.410 align:middle line:84% that we're collecting, where you're not just collecting 00:43:43.410 --> 00:43:45.540 align:middle line:84% an isolated condition and comparing 00:43:45.540 --> 00:43:48.510 align:middle line:90% it some other condition. 00:43:48.510 --> 00:43:51.000 align:middle line:84% It actually matters the order of the different conditions 00:43:51.000 --> 00:43:51.750 align:middle line:90% that you have. 00:43:51.750 --> 00:43:56.400 align:middle line:84% And this is a great advantage in analyzing causality, 00:43:56.400 --> 00:44:02.910 align:middle line:84% and we'll illustrate it in the context of messenger RNA decay, 00:44:02.910 --> 00:44:07.620 align:middle line:84% and finally in ways of aligning different time series data. 00:44:07.620 --> 00:44:09.900 align:middle line:84% Now, why do we want time [? courses? ?] 00:44:09.900 --> 00:44:14.130 align:middle line:84% If we do a gene knockout or we do a gene deletion, by the time 00:44:14.130 --> 00:44:18.490 align:middle line:84% you isolate that mutant and characterize and do the RNA, 00:44:18.490 --> 00:44:21.240 align:middle line:84% you've now gotten not just the primary effects, 00:44:21.240 --> 00:44:25.500 align:middle line:84% but all the downstream effects of that knockout. 00:44:25.500 --> 00:44:27.360 align:middle line:84% So the best would be to have some kind 00:44:27.360 --> 00:44:30.910 align:middle line:84% of conditional control of the transcription, 00:44:30.910 --> 00:44:34.890 align:middle line:84% so that when you first either turn it on or turn it off, 00:44:34.890 --> 00:44:40.800 align:middle line:84% the first events that occur are likely to be primary events. 00:44:40.800 --> 00:44:44.470 align:middle line:84% Now, the way that you control that needs to be, 00:44:44.470 --> 00:44:48.690 align:middle line:84% not have too many perturbing forces on the whole system. 00:44:48.690 --> 00:44:53.730 align:middle line:84% So temperature shift, it's an easy class of mutation to get, 00:44:53.730 --> 00:44:57.840 align:middle line:84% but it's not suitable because there's a huge temperature 00:44:57.840 --> 00:44:59.550 align:middle line:90% effect on the entire system. 00:44:59.550 --> 00:45:02.760 align:middle line:84% Chemical knockouts can be more specific, 00:45:02.760 --> 00:45:05.430 align:middle line:90% but you need to prove that. 00:45:05.430 --> 00:45:10.980 align:middle line:84% An example of a fairly time-honored chemical knockout 00:45:10.980 --> 00:45:14.760 align:middle line:84% is rifampicin, which fairly specifically affects 00:45:14.760 --> 00:45:18.670 align:middle line:90% just the RNA polymerase. 00:45:18.670 --> 00:45:22.800 align:middle line:84% And so this is an interesting case, 00:45:22.800 --> 00:45:29.770 align:middle line:84% where the effect is to stop initiation of transcription. 00:45:29.770 --> 00:45:32.640 align:middle line:84% And so then, as we do our time series, what we see 00:45:32.640 --> 00:45:39.120 align:middle line:84% is the RNAs for LPP, which we showed a few slides ago, 00:45:39.120 --> 00:45:40.120 align:middle line:90% is very stable. 00:45:40.120 --> 00:45:43.680 align:middle line:84% It basically lasts longer than the lifetime, 00:45:43.680 --> 00:45:48.000 align:middle line:84% the doubling time of the cell, possibly many cell generations. 00:45:48.000 --> 00:45:53.130 align:middle line:84% And other RNAs, such as CSPE, have extremely short half-lives 00:45:53.130 --> 00:45:55.170 align:middle line:84% in the order of 2 and 1/2 minutes. 00:45:55.170 --> 00:45:58.545 align:middle line:84% And you can compare various methods for quantitation. 00:45:58.545 --> 00:46:02.760 align:middle line:84% Then you've come up with different half-lives here. 00:46:02.760 --> 00:46:06.540 align:middle line:90% 00:46:06.540 --> 00:46:10.440 align:middle line:84% OK, so that's an example of a very significant class 00:46:10.440 --> 00:46:16.530 align:middle line:84% of chemically manipulated knockouts. 00:46:16.530 --> 00:46:18.810 align:middle line:90% So you can precisely phase them. 00:46:18.810 --> 00:46:21.420 align:middle line:84% You have very few other consequences, 00:46:21.420 --> 00:46:23.430 align:middle line:84% and then you can measure a time series. 00:46:23.430 --> 00:46:26.160 align:middle line:84% It'd be nice to be able to do that for any particular RNA 00:46:26.160 --> 00:46:28.650 align:middle line:84% and see what the downstream consequences are. 00:46:28.650 --> 00:46:31.140 align:middle line:84% Now, whenever you do a perturbation where 00:46:31.140 --> 00:46:32.700 align:middle line:84% you have two time series, you want 00:46:32.700 --> 00:46:39.570 align:middle line:84% to know how all the RNAs occurred 00:46:39.570 --> 00:46:44.310 align:middle line:84% during heat shock or some other pulse of some chemical 00:46:44.310 --> 00:46:47.940 align:middle line:84% relative to pulse of a different chemical, or the time series 00:46:47.940 --> 00:46:49.905 align:middle line:84% as it would have occurred without any. 00:46:49.905 --> 00:46:53.640 align:middle line:84% You can see how they won't necessarily 00:46:53.640 --> 00:46:55.200 align:middle line:90% align up point by point. 00:46:55.200 --> 00:46:56.698 align:middle line:84% You can't just start them at time 0 00:46:56.698 --> 00:46:57.990 align:middle line:90% and expect them all to line up. 00:46:57.990 --> 00:46:59.948 align:middle line:84% In fact, you can't even expect them necessarily 00:46:59.948 --> 00:47:02.070 align:middle line:84% to line up where you have a uniform stretch. 00:47:02.070 --> 00:47:05.490 align:middle line:84% You might have to have piecewise stretch, where certain parts go 00:47:05.490 --> 00:47:06.730 align:middle line:90% faster than others. 00:47:06.730 --> 00:47:09.750 align:middle line:84% Now, this may hopefully click in your mind 00:47:09.750 --> 00:47:11.910 align:middle line:84% a connection to the dynamic programming, where 00:47:11.910 --> 00:47:15.870 align:middle line:84% we had two sequences of bases or amino acids. 00:47:15.870 --> 00:47:17.730 align:middle line:84% And you wanted to expand or contract 00:47:17.730 --> 00:47:24.450 align:middle line:84% different sections of those by inserting a placeholder. 00:47:24.450 --> 00:47:26.850 align:middle line:84% Well, it doesn't make quite as much sense here 00:47:26.850 --> 00:47:29.400 align:middle line:84% with time series to insert a placeholder. 00:47:29.400 --> 00:47:31.030 align:middle line:90% So you can do that. 00:47:31.030 --> 00:47:33.720 align:middle line:84% You can have a discrete block diagram. 00:47:33.720 --> 00:47:40.350 align:middle line:84% Just this, here's series A and B in middle upper diagram. 00:47:40.350 --> 00:47:43.080 align:middle line:84% Or you can have a more continuous function, 00:47:43.080 --> 00:47:46.650 align:middle line:84% where you've tried to more smoothly warp. 00:47:46.650 --> 00:47:49.230 align:middle line:84% Both of these are dynamic programming algorithms. 00:47:49.230 --> 00:47:52.770 align:middle line:84% The smooth warping is slightly more a little more complicated. 00:47:52.770 --> 00:47:54.960 align:middle line:84% The insertion deletion one is exactly 00:47:54.960 --> 00:47:57.240 align:middle line:84% the same three conditions that we went through 00:47:57.240 --> 00:48:01.730 align:middle line:84% for pairwise alignments in dynamic programming. 00:48:01.730 --> 00:48:09.440 align:middle line:84% But this is partly to drive home how many different ways you 00:48:09.440 --> 00:48:11.060 align:middle line:90% can use dynamic programming. 00:48:11.060 --> 00:48:14.570 align:middle line:84% You can use it in HMMs, which is that multi-sequence alignment, 00:48:14.570 --> 00:48:17.360 align:middle line:84% and now for time series and gene expression. 00:48:17.360 --> 00:48:23.360 align:middle line:84% And you can see, from the literature on cell cycle, 00:48:23.360 --> 00:48:25.760 align:middle line:84% almost all of the data time series 00:48:25.760 --> 00:48:29.000 align:middle line:84% that we have so far actually don't align perfectly point 00:48:29.000 --> 00:48:31.700 align:middle line:84% by point, because you use wildly physiologically 00:48:31.700 --> 00:48:36.080 align:middle line:84% different conditions to get cells to synchronize, say, 00:48:36.080 --> 00:48:41.300 align:middle line:84% for cell division, or to start an event here using a mating 00:48:41.300 --> 00:48:44.120 align:middle line:84% pheromone, a small peptide that's 00:48:44.120 --> 00:48:48.260 align:middle line:84% released in the media that kind of controls the cell cycle 00:48:48.260 --> 00:48:50.930 align:middle line:84% and allows you to arrest and then release from arrest, 00:48:50.930 --> 00:48:52.640 align:middle line:84% or a temperature-sensitive mutant, 00:48:52.640 --> 00:48:54.410 align:middle line:84% even though I malign temperature-sensitive 00:48:54.410 --> 00:48:55.590 align:middle line:90% in just a moment. 00:48:55.590 --> 00:48:58.910 align:middle line:84% It is one of the most precise ways of getting 00:48:58.910 --> 00:49:01.730 align:middle line:90% synchrony of cell division. 00:49:01.730 --> 00:49:05.612 align:middle line:84% Cell division is a particularly good illustrative notion, 00:49:05.612 --> 00:49:07.820 align:middle line:84% partly because we mentioned it earlier in the course. 00:49:07.820 --> 00:49:11.030 align:middle line:84% But also, if you think of any dividing set of cells, 00:49:11.030 --> 00:49:13.070 align:middle line:84% many of the cell types that you'd be interested 00:49:13.070 --> 00:49:19.670 align:middle line:84% are dividing-- stem cells, microbial cells and so on. 00:49:19.670 --> 00:49:23.400 align:middle line:84% That automatically is a mixture of cells. 00:49:23.400 --> 00:49:25.972 align:middle line:84% If you mush them up and extract RNA, 00:49:25.972 --> 00:49:27.680 align:middle line:84% you're kidding yourself if you think this 00:49:27.680 --> 00:49:29.300 align:middle line:90% was a homogeneous population. 00:49:29.300 --> 00:49:35.040 align:middle line:84% If on the other hand, you synchronize the cells, 00:49:35.040 --> 00:49:39.150 align:middle line:84% then you've removed one major variable that could confound. 00:49:39.150 --> 00:49:42.750 align:middle line:84% Now they are much more homogeneous cells 00:49:42.750 --> 00:49:44.820 align:middle line:84% that are in the same state, and the cell cycle 00:49:44.820 --> 00:49:47.607 align:middle line:84% can be synchronously isolated as a population. 00:49:47.607 --> 00:49:49.440 align:middle line:84% There may be other sources of heterogeneity, 00:49:49.440 --> 00:49:51.090 align:middle line:90% but you've eliminated a big one. 00:49:51.090 --> 00:49:53.850 align:middle line:84% In any case, you take these two data series. 00:49:53.850 --> 00:49:57.120 align:middle line:84% They have different time constants, different lengths, 00:49:57.120 --> 00:49:58.470 align:middle line:90% and even different warps. 00:49:58.470 --> 00:50:01.350 align:middle line:84% Now, you want to take the x's and superimpose them 00:50:01.350 --> 00:50:02.790 align:middle line:90% on the o's. 00:50:02.790 --> 00:50:05.040 align:middle line:84% And here's an example of that now. 00:50:05.040 --> 00:50:07.788 align:middle line:84% They're both put together, and even 00:50:07.788 --> 00:50:09.330 align:middle line:84% though there may be little deviations 00:50:09.330 --> 00:50:11.220 align:middle line:84% for any particular gene, when you 00:50:11.220 --> 00:50:14.040 align:middle line:84% talk about the thousands of different genes, very 00:50:14.040 --> 00:50:15.000 align:middle line:90% rich pattern. 00:50:15.000 --> 00:50:20.430 align:middle line:84% Lots of information, plenty of opportunity for smoothing out 00:50:20.430 --> 00:50:21.420 align:middle line:90% individual variations. 00:50:21.420 --> 00:50:24.000 align:middle line:84% But here, you get superimposed patterns. 00:50:24.000 --> 00:50:26.520 align:middle line:84% And here's the traceback that tells you exactly where 00:50:26.520 --> 00:50:31.020 align:middle line:84% the insertions and deletions or smooth warping might occur 00:50:31.020 --> 00:50:35.530 align:middle line:84% to align these two different cell cycles' data sets. 00:50:35.530 --> 00:50:40.530 align:middle line:84% So in summary, we've connected the multi-sequence alignments 00:50:40.530 --> 00:50:46.470 align:middle line:84% from last class to allow you to model RNA structure, 00:50:46.470 --> 00:50:48.360 align:middle line:84% how RNA structure helps you model it. 00:50:48.360 --> 00:50:52.740 align:middle line:84% An interesting class of RNA guide sequences 00:50:52.740 --> 00:50:56.310 align:middle line:84% involved in methylation as an illustration 00:50:56.310 --> 00:50:59.430 align:middle line:84% of finding genes that don't encode proteins. 00:50:59.430 --> 00:51:02.250 align:middle line:84% And we talked about various quantitation methods, errors 00:51:02.250 --> 00:51:04.860 align:middle line:84% that present and solutions of errors, 00:51:04.860 --> 00:51:07.800 align:middle line:84% statistical methods for asking whether two distributions are 00:51:07.800 --> 00:51:13.950 align:middle line:84% related or have no difference in their means, 00:51:13.950 --> 00:51:17.340 align:middle line:84% interpretation errors about where RNAs start and stop, 00:51:17.340 --> 00:51:19.290 align:middle line:84% how you get alternative splicing, 00:51:19.290 --> 00:51:21.210 align:middle line:84% and finally time series data, which 00:51:21.210 --> 00:51:25.830 align:middle line:84% we will find very useful for connecting RNA 00:51:25.830 --> 00:51:31.560 align:middle line:84% and protein measures over time series for analyzing causality 00:51:31.560 --> 00:51:33.450 align:middle line:90% and systems biologies. 00:51:33.450 --> 00:51:35.020 align:middle line:90% OK, thank you very much. 00:51:35.020 --> 00:51:36.400 align:middle line:90% See you next time. 00:51:36.400 --> 00:51:40.130 align:middle line:84% Be sure to get your problem sets in to your teaching fellows. 00:51:40.130 --> 00:51:41.000 align:middle line:90%