WEBVTT 00:00:00.000 --> 00:00:02.700 align:middle line:84% ZOLTAN SZALLASI: --like modeling and what modeling can achieve, 00:00:02.700 --> 00:00:06.210 align:middle line:84% or why people are doing it at all, or is it doable at all. 00:00:06.210 --> 00:00:08.220 align:middle line:84% Should one be interested in this, 00:00:08.220 --> 00:00:11.928 align:middle line:84% or this is completely harebrained idea? 00:00:11.928 --> 00:00:13.470 align:middle line:84% Reverse engineering, where you really 00:00:13.470 --> 00:00:16.020 align:middle line:84% want to reconstruct how the system looks like-- what you 00:00:16.020 --> 00:00:18.360 align:middle line:90% are going to use for modeling. 00:00:18.360 --> 00:00:22.020 align:middle line:84% And one more issue is that I really need to talk about-- 00:00:22.020 --> 00:00:25.080 align:middle line:90% but of course this is like-- 00:00:25.080 --> 00:00:32.940 align:middle line:84% nothing really exciting happened in biological modeling yet. 00:00:32.940 --> 00:00:38.520 align:middle line:84% People tried and applied the very same tools, methods, 00:00:38.520 --> 00:00:42.325 align:middle line:84% that have been around in modeling for decades. 00:00:42.325 --> 00:00:44.700 align:middle line:84% They're facing the very same issues-- very same problems. 00:00:44.700 --> 00:00:47.940 align:middle line:90% You have three-- couple ODEs. 00:00:47.940 --> 00:00:49.933 align:middle line:90% Well, what's going to happen? 00:00:49.933 --> 00:00:51.850 align:middle line:84% Very quickly you're going to run into trouble. 00:00:51.850 --> 00:00:54.600 align:middle line:90% So what's the way out? 00:00:54.600 --> 00:00:57.360 align:middle line:84% This is really, really the hope that the robustness-- 00:00:57.360 --> 00:00:59.130 align:middle line:84% as we know, I mean, life is robust-- 00:00:59.130 --> 00:01:02.160 align:middle line:84% can teach us some lessons that we 00:01:02.160 --> 00:01:05.580 align:middle line:84% can exploit if we want to do large scale 00:01:05.580 --> 00:01:06.750 align:middle line:90% meaningful modeling. 00:01:06.750 --> 00:01:08.550 align:middle line:84% So actually that's like the third point. 00:01:08.550 --> 00:01:11.812 align:middle line:84% Now for you guys, what's your background? 00:01:11.812 --> 00:01:13.841 align:middle line:90% AUDIENCE: Biology [INAUDIBLE]. 00:01:13.841 --> 00:01:15.461 align:middle line:90% ZOLTAN SZALLASI: OK, and yours? 00:01:15.461 --> 00:01:18.050 align:middle line:84% AUDIENCE: Bio and chem engineering. 00:01:18.050 --> 00:01:18.765 align:middle line:90% Biology and-- 00:01:18.765 --> 00:01:19.890 align:middle line:90% ZOLTAN SZALLASI: And-- and? 00:01:19.890 --> 00:01:20.770 align:middle line:84% AUDIENCE: And chemical engineering. 00:01:20.770 --> 00:01:21.430 align:middle line:90% ZOLTAN SZALLASI: Chemical-- OK. 00:01:21.430 --> 00:01:22.070 align:middle line:90% So good. 00:01:22.070 --> 00:01:28.037 align:middle line:84% So I mean for a pure biologist to carry the point home 00:01:28.037 --> 00:01:29.620 align:middle line:84% that robustness is something important 00:01:29.620 --> 00:01:33.310 align:middle line:84% and to look out for, that this might give the solution, 00:01:33.310 --> 00:01:34.900 align:middle line:90% is kind of new. 00:01:34.900 --> 00:01:36.880 align:middle line:84% For you guys, probably it's pretty obvious 00:01:36.880 --> 00:01:39.735 align:middle line:84% that if you do very large scale modeling, especially 00:01:39.735 --> 00:01:41.860 align:middle line:84% for chemical engineering, then these type of things 00:01:41.860 --> 00:01:42.873 align:middle line:90% are very important. 00:01:42.873 --> 00:01:44.290 align:middle line:84% I mean, you know how it is, right? 00:01:44.290 --> 00:01:47.770 align:middle line:84% If you think about all the chaotic systems like-- yeah? 00:01:47.770 --> 00:01:48.892 align:middle line:90% AUDIENCE: [INAUDIBLE]? 00:01:48.892 --> 00:01:51.190 align:middle line:84% I mean, I have some background, but it's [INAUDIBLE].. 00:01:51.190 --> 00:01:52.690 align:middle line:84% ZOLTAN SZALLASI: Oh, I will, I will. 00:01:52.690 --> 00:01:54.940 align:middle line:90% It's just a introductory slide. 00:01:54.940 --> 00:01:56.445 align:middle line:90% Is this important or not? 00:01:56.445 --> 00:01:57.820 align:middle line:84% For you guys, you will understand 00:01:57.820 --> 00:01:58.737 align:middle line:90% why this is important. 00:01:58.737 --> 00:02:00.740 align:middle line:84% For biologists, why would I care? 00:02:00.740 --> 00:02:02.620 align:middle line:84% So that's kind of the distinction. 00:02:02.620 --> 00:02:04.330 align:middle line:84% OK, so what's the goals of science? 00:02:04.330 --> 00:02:07.420 align:middle line:84% Of course the main goal of science-- and one 00:02:07.420 --> 00:02:10.449 align:middle line:84% can open a whole discussion-- this is predictive power. 00:02:10.449 --> 00:02:12.550 align:middle line:84% And the other goal is understanding [INAUDIBLE] 00:02:12.550 --> 00:02:14.800 align:middle line:84% and playing in a playground for social [INAUDIBLE] 00:02:14.800 --> 00:02:15.710 align:middle line:90% and so forth. 00:02:15.710 --> 00:02:18.800 align:middle line:84% But what we want to do is predictions. 00:02:18.800 --> 00:02:21.730 align:middle line:84% So how did biology evolve or how did it work? 00:02:21.730 --> 00:02:23.140 align:middle line:90% Obviously you had a black box. 00:02:23.140 --> 00:02:24.610 align:middle line:90% Something happened, right? 00:02:24.610 --> 00:02:26.440 align:middle line:84% You set certain conditions, perturbations. 00:02:26.440 --> 00:02:28.210 align:middle line:84% You get something that you can describe 00:02:28.210 --> 00:02:30.200 align:middle line:90% that's the living organism. 00:02:30.200 --> 00:02:32.830 align:middle line:90% And there are some few readouts. 00:02:32.830 --> 00:02:38.020 align:middle line:84% For example, you get a drug and the patient responded or not. 00:02:38.020 --> 00:02:41.500 align:middle line:84% Or like if there is no oxygen around, then for most organisms 00:02:41.500 --> 00:02:42.685 align:middle line:90% this means death. 00:02:42.685 --> 00:02:44.560 align:middle line:84% You do not really know-- and we didn't really 00:02:44.560 --> 00:02:46.960 align:middle line:84% know for a long time-- what's inside the box. 00:02:46.960 --> 00:02:48.760 align:middle line:90% We just play with this. 00:02:48.760 --> 00:02:51.550 align:middle line:84% We had certain inputs and we had certain outputs. 00:02:51.550 --> 00:02:55.060 align:middle line:84% Of course later-- biology during the past several hundred years 00:02:55.060 --> 00:02:55.935 align:middle line:90% or 100 years-- 00:02:55.935 --> 00:02:57.310 align:middle line:84% we started to learn that there is 00:02:57.310 --> 00:03:02.860 align:middle line:84% a very complicated intricate network inside the cell, which 00:03:02.860 --> 00:03:04.760 align:middle line:90% is like the unit of life. 00:03:04.760 --> 00:03:07.787 align:middle line:84% And if you want to do modeling of course, 00:03:07.787 --> 00:03:09.370 align:middle line:84% you can do black box things and that's 00:03:09.370 --> 00:03:12.970 align:middle line:84% what brute force or exhaustive reverse engineering is about. 00:03:12.970 --> 00:03:15.010 align:middle line:84% That you do all sorts of inputs and outputs, 00:03:15.010 --> 00:03:16.725 align:middle line:84% and you can kind of guess what's inside. 00:03:16.725 --> 00:03:17.350 align:middle line:90% [CROWD CHATTER] 00:03:17.350 --> 00:03:20.090 align:middle line:84% But it's much more efficient if you-- 00:03:20.090 --> 00:03:20.590 align:middle line:90% I'm sorry. 00:03:20.590 --> 00:03:21.465 align:middle line:90% It really bothers me. 00:03:21.465 --> 00:03:22.180 align:middle line:90% [DOOR CLOSES] 00:03:22.180 --> 00:03:24.528 align:middle line:84% It's much more efficient if you have a pretty good idea 00:03:24.528 --> 00:03:25.570 align:middle line:90% of what's inside the box. 00:03:25.570 --> 00:03:28.660 align:middle line:84% Because then the modeling description 00:03:28.660 --> 00:03:30.640 align:middle line:84% is going to be much more efficient. 00:03:30.640 --> 00:03:32.080 align:middle line:90% So what's inside? 00:03:32.080 --> 00:03:32.870 align:middle line:90% Whoops. 00:03:32.870 --> 00:03:33.370 align:middle line:90% Sorry. 00:03:33.370 --> 00:03:33.870 align:middle line:90% Come in. 00:03:33.870 --> 00:03:36.250 align:middle line:90% 00:03:36.250 --> 00:03:38.830 align:middle line:84% Now you all have some biological background. 00:03:38.830 --> 00:03:41.800 align:middle line:84% So there are genes, RNA, proteins-- proteins 00:03:41.800 --> 00:03:43.300 align:middle line:90% have certain activation states. 00:03:43.300 --> 00:03:45.280 align:middle line:90% These things are translocated. 00:03:45.280 --> 00:03:47.930 align:middle line:84% I mean, there is temporal and spatial information 00:03:47.930 --> 00:03:50.578 align:middle line:84% and that's going to set up your entire network. 00:03:50.578 --> 00:03:52.120 align:middle line:84% So what systems biology-- or at least 00:03:52.120 --> 00:03:54.400 align:middle line:84% some field of systems biology-- is aiming to do 00:03:54.400 --> 00:03:57.580 align:middle line:84% is creating a predictive mathematical logical 00:03:57.580 --> 00:03:59.500 align:middle line:84% representation of the living organism. 00:03:59.500 --> 00:04:02.990 align:middle line:84% And that leads to the whole issue that what is-- 00:04:02.990 --> 00:04:04.510 align:middle line:84% again, for you guys it's obvious. 00:04:04.510 --> 00:04:07.210 align:middle line:84% For biologists, not that obvious whether modeling 00:04:07.210 --> 00:04:07.840 align:middle line:90% is good or bad. 00:04:07.840 --> 00:04:10.060 align:middle line:84% Biologists do not really like the whole idea 00:04:10.060 --> 00:04:11.290 align:middle line:90% of in silico modeling. 00:04:11.290 --> 00:04:13.300 align:middle line:90% And that's kind of obvious-- 00:04:13.300 --> 00:04:15.280 align:middle line:90% I mean, that's understandable. 00:04:15.280 --> 00:04:18.760 align:middle line:84% But for you guys acquainted with the backgrounds, 00:04:18.760 --> 00:04:21.640 align:middle line:84% you know very well that all we are doing in science in most 00:04:21.640 --> 00:04:22.840 align:middle line:90% cases, is actually modeling. 00:04:22.840 --> 00:04:26.020 align:middle line:84% We are creating some sort of a mathematical representation 00:04:26.020 --> 00:04:27.010 align:middle line:90% of a system. 00:04:27.010 --> 00:04:28.130 align:middle line:90% Think about anything. 00:04:28.130 --> 00:04:29.770 align:middle line:84% I mean of course, if you describe 00:04:29.770 --> 00:04:31.600 align:middle line:84% the movement of the Earth and the sun, 00:04:31.600 --> 00:04:33.730 align:middle line:90% that's two mass points. 00:04:33.730 --> 00:04:37.290 align:middle line:84% That's kind of representation, and that's 00:04:37.290 --> 00:04:39.040 align:middle line:84% going to give you pretty good predictions. 00:04:39.040 --> 00:04:41.320 align:middle line:84% But all we are doing in science is essentially 00:04:41.320 --> 00:04:43.640 align:middle line:90% some sort of modeling. 00:04:43.640 --> 00:04:45.650 align:middle line:84% So again, let me just mention this very quickly. 00:04:45.650 --> 00:04:48.850 align:middle line:84% This is for biologists, because biologists are really 00:04:48.850 --> 00:04:51.460 align:middle line:84% abhorred by the whole idea of in silico modeling, 00:04:51.460 --> 00:04:54.070 align:middle line:84% although they started to warm up towards the whole concept 00:04:54.070 --> 00:04:54.875 align:middle line:90% recently. 00:04:54.875 --> 00:04:56.500 align:middle line:84% But biologists have been doing modeling 00:04:56.500 --> 00:05:00.708 align:middle line:84% for the very beginning of biology. 00:05:00.708 --> 00:05:02.500 align:middle line:84% If you had an experimental drug, you're not 00:05:02.500 --> 00:05:04.150 align:middle line:90% going to give it to the kid-- 00:05:04.150 --> 00:05:05.380 align:middle line:90% or little Johnny. 00:05:05.380 --> 00:05:07.150 align:middle line:84% We'd have to see whether it works or not. 00:05:07.150 --> 00:05:09.370 align:middle line:84% But you're going to use an experimental model. 00:05:09.370 --> 00:05:11.440 align:middle line:84% Now there's a very strong underlying assumption 00:05:11.440 --> 00:05:14.230 align:middle line:90% here, which is, this model-- 00:05:14.230 --> 00:05:18.130 align:middle line:84% this animal-- is going to react to the drug the very same way 00:05:18.130 --> 00:05:19.540 align:middle line:90% as a child will-- 00:05:19.540 --> 00:05:21.880 align:middle line:84% which might be correct or incorrect. 00:05:21.880 --> 00:05:24.370 align:middle line:84% But some of you may recall or you 00:05:24.370 --> 00:05:26.980 align:middle line:84% might have heard of the thalidomide issue, when 00:05:26.980 --> 00:05:29.440 align:middle line:84% they just simply chose the wrong animal model. 00:05:29.440 --> 00:05:33.280 align:middle line:84% There is an animal model that would have produced the birth 00:05:33.280 --> 00:05:34.330 align:middle line:90% malformations. 00:05:34.330 --> 00:05:35.630 align:middle line:90% They didn't choose that. 00:05:35.630 --> 00:05:38.500 align:middle line:84% But the bottom line is that biologists 00:05:38.500 --> 00:05:41.440 align:middle line:84% have been using models, and they have no problem 00:05:41.440 --> 00:05:45.100 align:middle line:84% when they are using an animal when they are trying to model, 00:05:45.100 --> 00:05:46.670 align:middle line:90% for example, the human response. 00:05:46.670 --> 00:05:49.617 align:middle line:84% So modeling can be an in vitro and in vivo model. 00:05:49.617 --> 00:05:51.700 align:middle line:84% And of course, the alternative is in silico model. 00:05:51.700 --> 00:05:53.680 align:middle line:84% OK, so this is like the intro for the biologist 00:05:53.680 --> 00:05:55.480 align:middle line:90% that you don't need to hear. 00:05:55.480 --> 00:05:59.710 align:middle line:84% So the first issue is that what is in the black box? 00:05:59.710 --> 00:06:02.050 align:middle line:84% And this is essentially, how do we represent 00:06:02.050 --> 00:06:04.300 align:middle line:84% biological knowledge in a way that 00:06:04.300 --> 00:06:05.740 align:middle line:90% could be used for modeling? 00:06:05.740 --> 00:06:07.570 align:middle line:84% And that's sort of reverse engineering 00:06:07.570 --> 00:06:10.270 align:middle line:84% of the intracellular regulatory network. 00:06:10.270 --> 00:06:11.740 align:middle line:84% So just to give you some estimate 00:06:11.740 --> 00:06:14.890 align:middle line:84% of how complex is this thing that's inside the black box, 00:06:14.890 --> 00:06:18.110 align:middle line:84% let me just give you some estimates 00:06:18.110 --> 00:06:21.145 align:middle line:90% that I was playing with. 00:06:21.145 --> 00:06:23.270 align:middle line:84% For a while, I've been giving a tutorial on systems 00:06:23.270 --> 00:06:24.740 align:middle line:90% biology for many years now. 00:06:24.740 --> 00:06:28.880 align:middle line:84% So it just tries to give people a feeling 00:06:28.880 --> 00:06:30.350 align:middle line:90% of how complex the system is. 00:06:30.350 --> 00:06:33.380 align:middle line:84% So the cautious estimate is that the number of interacting 00:06:33.380 --> 00:06:37.442 align:middle line:84% parameters is going to be on the order of a couple of hundred 00:06:37.442 --> 00:06:38.900 align:middle line:84% thousands, but less than a million. 00:06:38.900 --> 00:06:41.483 align:middle line:84% The way this comes out, you have 10,000 to 20,000 active genes 00:06:41.483 --> 00:06:42.320 align:middle line:90% per cell. 00:06:42.320 --> 00:06:44.210 align:middle line:84% Different post-translational modifications, 00:06:44.210 --> 00:06:46.250 align:middle line:84% different localizations, that will 00:06:46.250 --> 00:06:48.890 align:middle line:84% have independent regulatory input for something else. 00:06:48.890 --> 00:06:52.400 align:middle line:84% It's going to be, let's say that the three per yeast 00:06:52.400 --> 00:06:54.530 align:middle line:84% and three to six that's probably less than 10 00:06:54.530 --> 00:06:58.740 align:middle line:84% per protein per gene on average in humans. 00:06:58.740 --> 00:07:02.040 align:middle line:84% So of course the whole thing may be way off either direction 00:07:02.040 --> 00:07:04.640 align:middle line:84% because if you have splice variants for example, lots 00:07:04.640 --> 00:07:06.470 align:middle line:84% of antisense transcribed genes, this number 00:07:06.470 --> 00:07:08.130 align:middle line:90% might be way higher. 00:07:08.130 --> 00:07:11.270 align:middle line:84% On the other hand, if modularity really works well in biology, 00:07:11.270 --> 00:07:15.110 align:middle line:84% you don't really need to put every parameter in your model. 00:07:15.110 --> 00:07:17.450 align:middle line:84% You can work with appropriate modules, 00:07:17.450 --> 00:07:20.150 align:middle line:84% and that's actually one of the key hopes. 00:07:20.150 --> 00:07:22.250 align:middle line:84% But that's one of the main directions 00:07:22.250 --> 00:07:23.810 align:middle line:84% biologists try to move to these days, 00:07:23.810 --> 00:07:25.190 align:middle line:90% because that's still very large. 00:07:25.190 --> 00:07:27.360 align:middle line:90% The point here is the following. 00:07:27.360 --> 00:07:29.690 align:middle line:84% If you take these numbers and translate 00:07:29.690 --> 00:07:32.690 align:middle line:84% into actual numbers of parameters and networks, 00:07:32.690 --> 00:07:34.070 align:middle line:90% for bacteria-- 00:07:34.070 --> 00:07:38.420 align:middle line:84% E. coli-- that that number, the complexity of that network, 00:07:38.420 --> 00:07:44.390 align:middle line:84% is not much higher than the regulatory control 00:07:44.390 --> 00:07:47.542 align:middle line:90% mechanism for a jumbo jet. 00:07:47.542 --> 00:07:49.250 align:middle line:84% So actually the idea or the whole concept 00:07:49.250 --> 00:07:50.390 align:middle line:90% is coming from John Doyle. 00:07:50.390 --> 00:07:52.858 align:middle line:84% I'm just kind of stealing this sentence from him. 00:07:52.858 --> 00:07:53.900 align:middle line:90% John Doyle is at Caltech. 00:07:53.900 --> 00:07:55.275 align:middle line:84% He's a very famous mathematician. 00:07:55.275 --> 00:07:57.140 align:middle line:84% Actually, he's the one who worked out 00:07:57.140 --> 00:07:59.900 align:middle line:90% the theory behind stealth. 00:07:59.900 --> 00:08:01.610 align:middle line:90% Stealth is not supposed to fly. 00:08:01.610 --> 00:08:04.010 align:middle line:84% The way that looks like, it's not supposed to fly. 00:08:04.010 --> 00:08:06.740 align:middle line:84% The reason that flies is that the control mechanism 00:08:06.740 --> 00:08:09.950 align:middle line:84% is so well-designed, that actually it's 00:08:09.950 --> 00:08:11.220 align:middle line:90% going to remain in the air. 00:08:11.220 --> 00:08:15.180 align:middle line:84% But if you describe like the aerodynamic capabilities, 00:08:15.180 --> 00:08:16.430 align:middle line:90% that's not supposed to fly. 00:08:16.430 --> 00:08:19.010 align:middle line:84% So the point is that the control theory-- the control 00:08:19.010 --> 00:08:22.280 align:middle line:84% mechanism-- in a very complex engine 00:08:22.280 --> 00:08:27.380 align:middle line:84% these days is not much, much, much higher, but definitely 00:08:27.380 --> 00:08:29.400 align:middle line:84% not many orders of magnitude higher than, 00:08:29.400 --> 00:08:30.510 align:middle line:90% for example, for E. coli. 00:08:30.510 --> 00:08:33.500 align:middle line:84% So if you want to try very, very, very, very hard 00:08:33.500 --> 00:08:35.630 align:middle line:84% and work out all the details for E. coli, 00:08:35.630 --> 00:08:39.185 align:middle line:84% then one might try to do some sort 00:08:39.185 --> 00:08:43.730 align:middle line:84% of a meaningful dynamic modeling on at least the bacteria. 00:08:43.730 --> 00:08:47.265 align:middle line:84% And actually, there are many groups 00:08:47.265 --> 00:08:48.890 align:middle line:84% or many institutions-- some companies-- 00:08:48.890 --> 00:08:50.540 align:middle line:90% who are trying to do that. 00:08:50.540 --> 00:08:51.980 align:middle line:90% Now that's not an obvious thing. 00:08:51.980 --> 00:08:53.870 align:middle line:84% There will be still lots and lots of questions open. 00:08:53.870 --> 00:08:55.940 align:middle line:84% Like I'm going to talk about parameter optimization, 00:08:55.940 --> 00:08:57.690 align:middle line:84% which is-- of course, you cannot solve it. 00:08:57.690 --> 00:09:01.550 align:middle line:84% It's just a very, very, very computationally 00:09:01.550 --> 00:09:02.610 align:middle line:90% intractable problem. 00:09:02.610 --> 00:09:08.300 align:middle line:84% But the point is that if the way you design an airplane-- 00:09:08.300 --> 00:09:09.620 align:middle line:90% which has to be robust-- 00:09:09.620 --> 00:09:13.340 align:middle line:84% is very similar to the way an organism will work, 00:09:13.340 --> 00:09:15.410 align:middle line:84% then there are certain tricks that you 00:09:15.410 --> 00:09:21.290 align:middle line:84% can borrow from control theory that you can apply to bacteria. 00:09:21.290 --> 00:09:22.970 align:middle line:90% So representation of biology-- 00:09:22.970 --> 00:09:24.860 align:middle line:84% the simplest level, this is just-- 00:09:24.860 --> 00:09:28.790 align:middle line:84% stole the epidermal growth factor receptor pathway 00:09:28.790 --> 00:09:29.713 align:middle line:90% from the CAG database. 00:09:29.713 --> 00:09:31.130 align:middle line:84% So the simplest level is, you have 00:09:31.130 --> 00:09:32.365 align:middle line:90% some sort of directed graph. 00:09:32.365 --> 00:09:33.740 align:middle line:84% Now this is obvious for you guys. 00:09:33.740 --> 00:09:35.420 align:middle line:84% For biologists, I point out that, this 00:09:35.420 --> 00:09:41.570 align:middle line:84% is like the simplest level of biological representation. 00:09:41.570 --> 00:09:43.490 align:middle line:84% So of course you have the nodes that 00:09:43.490 --> 00:09:45.440 align:middle line:84% are proteins or genes or post-translational 00:09:45.440 --> 00:09:46.490 align:middle line:90% modifications or whatnot. 00:09:46.490 --> 00:09:49.860 align:middle line:84% Then there is the edges that are the regulatory interactions. 00:09:49.860 --> 00:09:51.650 align:middle line:84% Now you cannot do much with this. 00:09:51.650 --> 00:09:53.690 align:middle line:90% This is not a dynamic model yet. 00:09:53.690 --> 00:09:56.660 align:middle line:84% You need a more detailed description of actually 00:09:56.660 --> 00:10:01.070 align:middle line:84% what's happening between some parameters, 00:10:01.070 --> 00:10:06.548 align:middle line:84% and you need to make that dynamic by introducing time. 00:10:06.548 --> 00:10:08.840 align:middle line:84% And in this case, if you have a continuous differential 00:10:08.840 --> 00:10:10.310 align:middle line:84% equation, then you need some sort 00:10:10.310 --> 00:10:12.920 align:middle line:84% of a description of how things are going to change here. 00:10:12.920 --> 00:10:16.580 align:middle line:84% And of course you need certain type of parameters-- 00:10:16.580 --> 00:10:18.050 align:middle line:90% genetic parameters. 00:10:18.050 --> 00:10:19.645 align:middle line:90% You need all these ingredients. 00:10:19.645 --> 00:10:21.020 align:middle line:84% Now this is obvious for you guys. 00:10:21.020 --> 00:10:22.970 align:middle line:90% Yeah. 00:10:22.970 --> 00:10:25.322 align:middle line:84% So when you do reverse engineering-- and biologists 00:10:25.322 --> 00:10:26.780 align:middle line:84% have been doing reverse engineering 00:10:26.780 --> 00:10:29.030 align:middle line:84% for many, many decades essentially. 00:10:29.030 --> 00:10:30.650 align:middle line:84% When you try to identify a protein, 00:10:30.650 --> 00:10:32.750 align:middle line:84% that's essentially reverse engineering 00:10:32.750 --> 00:10:34.020 align:middle line:90% the individual nodes. 00:10:34.020 --> 00:10:36.590 align:middle line:84% So that's what biochemistry or genome and other large scale 00:10:36.590 --> 00:10:39.260 align:middle line:84% node identifying projects are doing. 00:10:39.260 --> 00:10:42.350 align:middle line:84% Biochemistry, protein interaction screening, 00:10:42.350 --> 00:10:45.320 align:middle line:84% two-hybrid co-IP, is trying to determine 00:10:45.320 --> 00:10:48.230 align:middle line:84% all the regulatory interactions-- 00:10:48.230 --> 00:10:50.540 align:middle line:90% all the edges in your graph. 00:10:50.540 --> 00:10:53.870 align:middle line:84% But you can actually start, and there 00:10:53.870 --> 00:10:56.307 align:middle line:84% are some efforts that try to do brute force 00:10:56.307 --> 00:10:57.140 align:middle line:90% reverse engineering. 00:10:57.140 --> 00:10:58.230 align:middle line:90% Why would you do that? 00:10:58.230 --> 00:11:00.680 align:middle line:84% I mean, these approaches are extremely time-consuming. 00:11:00.680 --> 00:11:04.310 align:middle line:84% Although there are smart ways-- two-hybrid-- 00:11:04.310 --> 00:11:05.720 align:middle line:90% you can speed up things. 00:11:05.720 --> 00:11:08.090 align:middle line:84% You can do things in a high throughput manner. 00:11:08.090 --> 00:11:11.250 align:middle line:84% But working out every single detail for a genome of 20,000, 00:11:11.250 --> 00:11:14.790 align:middle line:84% 30,000 things is going to take a while. 00:11:14.790 --> 00:11:17.300 align:middle line:84% So if you know parts of it and you're 00:11:17.300 --> 00:11:19.900 align:middle line:84% quite sure about some subnetworks here, 00:11:19.900 --> 00:11:21.910 align:middle line:84% you can start to apply at the same time 00:11:21.910 --> 00:11:25.370 align:middle line:84% brute force methods which is why it's simple. 00:11:25.370 --> 00:11:29.230 align:middle line:84% You have consecutive time steps, and you 00:11:29.230 --> 00:11:31.990 align:middle line:84% try to determine a set of regulatory rules 00:11:31.990 --> 00:11:36.610 align:middle line:84% that can produce a gene expression pattern or parameter 00:11:36.610 --> 00:11:39.400 align:middle line:84% setting at the latest time point based on the previous time 00:11:39.400 --> 00:11:39.950 align:middle line:90% point. 00:11:39.950 --> 00:11:44.920 align:middle line:84% And if you do have many, many sufficiently diverse 00:11:44.920 --> 00:11:48.490 align:middle line:84% consecutive time points, then you 00:11:48.490 --> 00:11:50.950 align:middle line:84% can have a fairly good idea how things are 00:11:50.950 --> 00:11:53.110 align:middle line:90% going to regulate each other. 00:11:53.110 --> 00:11:57.220 align:middle line:84% Now the most widely used assumption 00:11:57.220 --> 00:12:00.050 align:middle line:84% is that you have some sort of an additive model. 00:12:00.050 --> 00:12:03.790 align:middle line:84% In this case, the state of a gene at a later time point 00:12:03.790 --> 00:12:05.920 align:middle line:84% is going to be determined somehow 00:12:05.920 --> 00:12:10.330 align:middle line:84% by the weighted sum of the state of another set of genes-- 00:12:10.330 --> 00:12:11.740 align:middle line:90% some sort of bias factor. 00:12:11.740 --> 00:12:16.383 align:middle line:84% And you have some sort of skewing function there. 00:12:16.383 --> 00:12:18.550 align:middle line:84% It doesn't really matter what actually you're doing. 00:12:18.550 --> 00:12:20.320 align:middle line:84% There are different types of methods. 00:12:20.320 --> 00:12:23.200 align:middle line:84% The number crunching can be done different ways. 00:12:23.200 --> 00:12:26.320 align:middle line:90% The bottom line is that-- 00:12:26.320 --> 00:12:28.570 align:middle line:84% these are just individual methods have been used-- 00:12:28.570 --> 00:12:31.340 align:middle line:84% that you need to determine all the bias factors and weight 00:12:31.340 --> 00:12:31.840 align:middle line:90% values. 00:12:31.840 --> 00:12:33.673 align:middle line:84% And if you think about it, you can just very 00:12:33.673 --> 00:12:35.500 align:middle line:84% simply think about it that it's like, you 00:12:35.500 --> 00:12:38.110 align:middle line:84% have many, many, many, many linear equations. 00:12:38.110 --> 00:12:42.970 align:middle line:84% And you need as many equations as bias factors and weight 00:12:42.970 --> 00:12:46.330 align:middle line:84% factors you have there, in order to do reverse engineering. 00:12:46.330 --> 00:12:49.570 align:middle line:84% Which was actually-- these are the exhaustive solutions, 00:12:49.570 --> 00:12:52.457 align:middle line:84% and that was actually done on the bound here. 00:12:52.457 --> 00:12:54.790 align:middle line:84% And what [INAUDIBLE],, who's a very smart mathematician, 00:12:54.790 --> 00:12:57.337 align:middle line:84% actually showed [INAUDIBLE] are unknown parameters 00:12:57.337 --> 00:12:59.170 align:middle line:84% in a set of ordinary differential equations, 00:12:59.170 --> 00:13:02.500 align:middle line:84% then you will need it on the order of 2r, 00:13:02.500 --> 00:13:05.920 align:middle line:84% on the order of r, sufficiently diverse time points-- 00:13:05.920 --> 00:13:08.320 align:middle line:84% to do efficient reverse engineering. 00:13:08.320 --> 00:13:11.770 align:middle line:84% This kind of [INAUDIBLE],, this is the number one would expect. 00:13:11.770 --> 00:13:15.453 align:middle line:84% But actually he proved it, so that's good. 00:13:15.453 --> 00:13:16.870 align:middle line:84% The point on this slide is that we 00:13:16.870 --> 00:13:19.000 align:middle line:84% should think about reverse engineering based 00:13:19.000 --> 00:13:22.330 align:middle line:84% on the assumption how your network works-- 00:13:22.330 --> 00:13:23.950 align:middle line:84% or network representation if you think 00:13:23.950 --> 00:13:26.380 align:middle line:84% about like Boolean networks and different levels 00:13:26.380 --> 00:13:27.760 align:middle line:90% of connectivity. 00:13:27.760 --> 00:13:31.660 align:middle line:84% You can make estimates of how many different time points, how 00:13:31.660 --> 00:13:34.255 align:middle line:84% many different measurements, you need 00:13:34.255 --> 00:13:36.380 align:middle line:90% to do a reverse engineering. 00:13:36.380 --> 00:13:36.880 align:middle line:90% Yeah? 00:13:36.880 --> 00:13:39.100 align:middle line:84% AUDIENCE: [INAUDIBLE] equates to the [INAUDIBLE]?? 00:13:39.100 --> 00:13:39.370 align:middle line:90% ZOLTAN SZALLASI: Sorry? 00:13:39.370 --> 00:13:40.930 align:middle line:84% AUDIENCE: You said it equates to [INAUDIBLE]?? 00:13:40.930 --> 00:13:42.680 align:middle line:84% ZOLTAN SZALLASI: In this case, no, no, no. 00:13:42.680 --> 00:13:45.820 align:middle line:84% This is the only one that has continuous differential 00:13:45.820 --> 00:13:46.460 align:middle line:90% equations. 00:13:46.460 --> 00:13:48.470 align:middle line:84% These are like Boolean representations. 00:13:48.470 --> 00:13:49.600 align:middle line:90% This is like an old slide. 00:13:49.600 --> 00:13:51.620 align:middle line:84% The point here is just showing that, of course, 00:13:51.620 --> 00:13:56.170 align:middle line:84% if you have a fully connected network, you cannot do it. 00:13:56.170 --> 00:13:58.750 align:middle line:84% If you have Boolean connectivity K, that 00:13:58.750 --> 00:14:02.200 align:middle line:84% means that your average number of input per node is K, 00:14:02.200 --> 00:14:05.160 align:middle line:90% then the number is much smaller. 00:14:05.160 --> 00:14:07.800 align:middle line:84% AUDIENCE: Oh, no, I think you mentioned [INAUDIBLE].. 00:14:07.800 --> 00:14:08.800 align:middle line:90% ZOLTAN SZALLASI: Oh, no. 00:14:08.800 --> 00:14:11.500 align:middle line:84% The linearity was simply-- this is just an additive model. 00:14:11.500 --> 00:14:14.530 align:middle line:90% 00:14:14.530 --> 00:14:19.720 align:middle line:84% We do not-- no lab has produced any sufficiently good quality 00:14:19.720 --> 00:14:23.860 align:middle line:84% data and large enough data set to worry 00:14:23.860 --> 00:14:28.587 align:middle line:90% about what the function is. 00:14:28.587 --> 00:14:30.670 align:middle line:84% What people are trying to determine at this scale, 00:14:30.670 --> 00:14:33.400 align:middle line:84% is actually who might be regulating whom. 00:14:33.400 --> 00:14:38.350 align:middle line:84% The equation is simply out of question at the moment. 00:14:38.350 --> 00:14:41.467 align:middle line:84% This just gives you a feel of how to-- or the difficulties 00:14:41.467 --> 00:14:42.550 align:middle line:90% involved in what you need. 00:14:42.550 --> 00:14:45.290 align:middle line:90% 00:14:45.290 --> 00:14:47.860 align:middle line:84% If you do not have enough independent measurements, 00:14:47.860 --> 00:14:50.582 align:middle line:84% then one other way of going-- and actually 00:14:50.582 --> 00:14:52.540 align:middle line:84% biologists have been using it for a long time-- 00:14:52.540 --> 00:14:54.580 align:middle line:90% is actually using perturbations. 00:14:54.580 --> 00:14:56.180 align:middle line:84% So if you do individual perturbations 00:14:56.180 --> 00:14:58.180 align:middle line:84% in the system in a kind of directed perturbation 00:14:58.180 --> 00:15:00.700 align:middle line:84% that these are the part of the network I'm interested in, 00:15:00.700 --> 00:15:03.040 align:middle line:84% then you can do this type of measurement. 00:15:03.040 --> 00:15:05.542 align:middle line:84% And the principle is very simple. 00:15:05.542 --> 00:15:07.000 align:middle line:84% Imagine that you have three genes-- 00:15:07.000 --> 00:15:09.305 align:middle line:84% A regulating B, B is regulating C. 00:15:09.305 --> 00:15:11.740 align:middle line:84% And then you create a perturbation matrix. 00:15:11.740 --> 00:15:13.270 align:middle line:90% You knock out individual genes. 00:15:13.270 --> 00:15:15.040 align:middle line:84% And this is going to show you that based 00:15:15.040 --> 00:15:18.460 align:middle line:84% on this topology, how the individual downstream regulated 00:15:18.460 --> 00:15:20.060 align:middle line:90% genes are going to behave. 00:15:20.060 --> 00:15:21.670 align:middle line:84% So from this perturbation matrix, 00:15:21.670 --> 00:15:25.250 align:middle line:84% you can create an accessibility matrix, 00:15:25.250 --> 00:15:27.820 align:middle line:84% which means that who might be regulating whom. 00:15:27.820 --> 00:15:30.850 align:middle line:84% And of course a given accessibility matrix 00:15:30.850 --> 00:15:33.520 align:middle line:84% can be associated with different regulatory networks. 00:15:33.520 --> 00:15:35.710 align:middle line:84% Both of these networks can be described 00:15:35.710 --> 00:15:37.600 align:middle line:90% by this accessibility matrix. 00:15:37.600 --> 00:15:39.400 align:middle line:84% And then usually what people think 00:15:39.400 --> 00:15:42.040 align:middle line:84% is like, let's take the most parsimonious, 00:15:42.040 --> 00:15:43.690 align:middle line:90% and they tend to believe it. 00:15:43.690 --> 00:15:47.180 align:middle line:84% Now perturbation measurements are not working very well. 00:15:47.180 --> 00:15:50.110 align:middle line:84% So I mean, you can work on subnetworks, 00:15:50.110 --> 00:15:52.870 align:middle line:84% and it has been shown to work if you 00:15:52.870 --> 00:15:56.530 align:middle line:84% have a subnetwork that has been sufficiently well described. 00:15:56.530 --> 00:16:00.370 align:middle line:84% But nothing has been really produced by these methods. 00:16:00.370 --> 00:16:03.610 align:middle line:84% Nothing really new has been discovered by itself. 00:16:03.610 --> 00:16:06.730 align:middle line:84% But this is kind of the gist of it. 00:16:06.730 --> 00:16:09.400 align:middle line:84% This is kind of the underlying principle, how are 00:16:09.400 --> 00:16:13.460 align:middle line:90% people trying to approach this. 00:16:13.460 --> 00:16:16.870 align:middle line:84% So if you do reverse engineering with these type of methods, 00:16:16.870 --> 00:16:21.720 align:middle line:84% given good quality data and enough computational 00:16:21.720 --> 00:16:24.300 align:middle line:84% number-crunching power, most probably 00:16:24.300 --> 00:16:26.640 align:middle line:84% we are going to have some sort of a regulatory topology 00:16:26.640 --> 00:16:27.330 align:middle line:90% of a network. 00:16:27.330 --> 00:16:29.050 align:middle line:90% It's almost given. 00:16:29.050 --> 00:16:32.700 align:middle line:84% And there are lots of preps who are 00:16:32.700 --> 00:16:34.740 align:middle line:90% producing knockouts for yeast. 00:16:34.740 --> 00:16:38.820 align:middle line:84% And they started to produce some meaningful data sets. 00:16:38.820 --> 00:16:43.080 align:middle line:84% But if you do dynamic modeling, you 00:16:43.080 --> 00:16:46.530 align:middle line:84% would at least hope that you have some of kinetic parameters 00:16:46.530 --> 00:16:50.250 align:middle line:84% in that model that describe or approximate reality 00:16:50.250 --> 00:16:51.660 align:middle line:90% fairly well. 00:16:51.660 --> 00:16:55.807 align:middle line:84% So another aspect of reverse engineering 00:16:55.807 --> 00:16:58.140 align:middle line:84% is when you have some idea about the kinetic parameters, 00:16:58.140 --> 00:17:01.830 align:middle line:84% but you're trying to refine these measurements. 00:17:01.830 --> 00:17:04.119 align:middle line:84% And that's what [INAUDIBLE] lab did. 00:17:04.119 --> 00:17:05.369 align:middle line:90% That's what they did actually. 00:17:05.369 --> 00:17:09.630 align:middle line:84% They essentially tagged by green fluorescent protein, 00:17:09.630 --> 00:17:14.170 align:middle line:90% every single protein in E. coli. 00:17:14.170 --> 00:17:16.740 align:middle line:84% So green for us, protein is like a tag 00:17:16.740 --> 00:17:18.210 align:middle line:84% and it's on every single protein. 00:17:18.210 --> 00:17:20.670 align:middle line:84% And they just do time series measurements 00:17:20.670 --> 00:17:21.848 align:middle line:90% on every single protein. 00:17:21.848 --> 00:17:23.640 align:middle line:84% It can be done in a high throughput fashion 00:17:23.640 --> 00:17:25.057 align:middle line:84% because in every single well, they 00:17:25.057 --> 00:17:26.670 align:middle line:90% know which protein is tagged. 00:17:26.670 --> 00:17:29.040 align:middle line:84% And if you do this, then you can actually-- 00:17:29.040 --> 00:17:31.380 align:middle line:84% from the measurements-- you can refine 00:17:31.380 --> 00:17:34.350 align:middle line:84% the actual kinetic parameters pretty well. 00:17:34.350 --> 00:17:37.200 align:middle line:84% The reason you need this is quite obvious. 00:17:37.200 --> 00:17:40.530 align:middle line:84% When biologists or biochemists determine genetic parameters, 00:17:40.530 --> 00:17:43.140 align:middle line:84% that's usually done in a free solution. 00:17:43.140 --> 00:17:45.360 align:middle line:84% You grind up the bacteria, do something, 00:17:45.360 --> 00:17:48.420 align:middle line:84% and that can be quite different from the actual genetic 00:17:48.420 --> 00:17:50.550 align:middle line:90% parameter present in the cell. 00:17:50.550 --> 00:17:51.870 align:middle line:90% So for one thing-- 00:17:51.870 --> 00:17:54.240 align:middle line:84% think about it-- in a free solution, 00:17:54.240 --> 00:17:58.020 align:middle line:84% you approximate things with very large number of molecules. 00:17:58.020 --> 00:18:02.370 align:middle line:84% Whereas in cells, sometimes you have 50 to 100 proteins 00:18:02.370 --> 00:18:03.400 align:middle line:90% of the same kind. 00:18:03.400 --> 00:18:04.440 align:middle line:84% So you might have a very different-- 00:18:04.440 --> 00:18:05.857 align:middle line:84% and it's not a free solution, it's 00:18:05.857 --> 00:18:07.330 align:middle line:90% a much more dense solution. 00:18:07.330 --> 00:18:08.910 align:middle line:84% So the efficient kinetic parameter 00:18:08.910 --> 00:18:10.120 align:middle line:90% can be very different. 00:18:10.120 --> 00:18:13.660 align:middle line:84% So reverse engineering can give you much better estimate 00:18:13.660 --> 00:18:16.335 align:middle line:84% so we can refine the estimates on the actual category. 00:18:16.335 --> 00:18:18.930 align:middle line:90% AUDIENCE: [INAUDIBLE] localized? 00:18:18.930 --> 00:18:20.760 align:middle line:90% ZOLTAN SZALLASI: As well, yes. 00:18:20.760 --> 00:18:22.590 align:middle line:84% I mean the point is that if you do-- 00:18:22.590 --> 00:18:25.260 align:middle line:84% you want to start with the best parameter set. 00:18:25.260 --> 00:18:27.570 align:middle line:90% And in this case, it's-- 00:18:27.570 --> 00:18:28.320 align:middle line:90% yes, localization. 00:18:28.320 --> 00:18:29.160 align:middle line:90% Of course it's important. 00:18:29.160 --> 00:18:30.870 align:middle line:84% I think about it at membrane proteins. 00:18:30.870 --> 00:18:33.570 align:middle line:84% I mean things that will happen in the liquid phase 00:18:33.570 --> 00:18:36.970 align:middle line:84% is going to be very different from the free water solution. 00:18:36.970 --> 00:18:38.638 align:middle line:90% So [INAUDIBLE]. 00:18:38.638 --> 00:18:41.510 align:middle line:90% 00:18:41.510 --> 00:18:44.590 align:middle line:84% So let's assume that one went through all the difficulties 00:18:44.590 --> 00:18:47.800 align:middle line:84% and have a fairly good description of the topology 00:18:47.800 --> 00:18:52.708 align:middle line:84% and the regulatory interactions, and then you 00:18:52.708 --> 00:18:54.250 align:middle line:84% have good quality kinetic parameters. 00:18:54.250 --> 00:18:56.930 align:middle line:90% So should one try-- 00:18:56.930 --> 00:18:57.430 align:middle line:90% yeah. 00:18:57.430 --> 00:18:58.450 align:middle line:90% AUDIENCE: Just have a question. 00:18:58.450 --> 00:18:59.367 align:middle line:90% ZOLTAN SZALLASI: Yeah. 00:18:59.367 --> 00:19:00.190 align:middle line:90% 00:19:00.190 --> 00:19:01.690 align:middle line:84% AUDIENCE: Two things that stand out. 00:19:01.690 --> 00:19:04.870 align:middle line:84% [INAUDIBLE] the schematics slide you had before. 00:19:04.870 --> 00:19:05.787 align:middle line:90% That one, yeah. 00:19:05.787 --> 00:19:07.120 align:middle line:90% ZOLTAN SZALLASI: This one, yeah. 00:19:07.120 --> 00:19:10.260 align:middle line:84% AUDIENCE: There are two things about this that stand out. 00:19:10.260 --> 00:19:16.240 align:middle line:84% Assuming [INAUDIBLE] I'm sure that some of the GFP tagged 00:19:16.240 --> 00:19:20.428 align:middle line:84% proteins actually were lethal [INAUDIBLE] right? 00:19:20.428 --> 00:19:21.595 align:middle line:90% ZOLTAN SZALLASI: Some, yeah. 00:19:21.595 --> 00:19:25.330 align:middle line:84% AUDIENCE: Because they included function or [INAUDIBLE] 00:19:25.330 --> 00:19:28.840 align:middle line:90% critical [INAUDIBLE]. 00:19:28.840 --> 00:19:31.720 align:middle line:84% And other proteins, obviously because many of them 00:19:31.720 --> 00:19:38.230 align:middle line:84% [INAUDIBLE] large protein, will probably reduce-- 00:19:38.230 --> 00:19:40.370 align:middle line:84% or have reduced activity or whatever. 00:19:40.370 --> 00:19:42.290 align:middle line:84% And both of those actually go towards-- 00:19:42.290 --> 00:19:44.560 align:middle line:84% so if we want to understand parameters 00:19:44.560 --> 00:19:49.420 align:middle line:84% or the kinetic parameters for gene regulation or whatever, 00:19:49.420 --> 00:19:53.680 align:middle line:84% I'm going to seriously bias the information that you get out 00:19:53.680 --> 00:19:56.260 align:middle line:84% of the [INAUDIBLE] much the same way that 00:19:56.260 --> 00:20:00.140 align:middle line:84% having the solution [INAUDIBLE] parameters [INAUDIBLE].. 00:20:00.140 --> 00:20:04.900 align:middle line:90% 00:20:04.900 --> 00:20:06.130 align:middle line:90% How do they address that? 00:20:06.130 --> 00:20:06.595 align:middle line:90% ZOLTAN SZALLASI: They didn't. 00:20:06.595 --> 00:20:07.390 align:middle line:84% AUDIENCE: They didn't address that? 00:20:07.390 --> 00:20:07.810 align:middle line:90% Oh. 00:20:07.810 --> 00:20:08.770 align:middle line:84% ZOLTAN SZALLASI: At this stage, they 00:20:08.770 --> 00:20:11.162 align:middle line:84% were happy to see that they can produce something 00:20:11.162 --> 00:20:12.370 align:middle line:90% in a high throughput fashion. 00:20:12.370 --> 00:20:17.240 align:middle line:84% Your points are absolutely valid and correct and important. 00:20:17.240 --> 00:20:19.360 align:middle line:84% What they do is, they do another expression 00:20:19.360 --> 00:20:20.950 align:middle line:90% at an individual protein level. 00:20:20.950 --> 00:20:23.680 align:middle line:84% Where they say that if I rewire-- 00:20:23.680 --> 00:20:26.290 align:middle line:90% and they took a sub-- 00:20:26.290 --> 00:20:28.600 align:middle line:84% this was not done in the actual publication 00:20:28.600 --> 00:20:30.010 align:middle line:90% for the entire network. 00:20:30.010 --> 00:20:33.580 align:middle line:84% They just took that part of the network that describes 00:20:33.580 --> 00:20:36.040 align:middle line:90% how the flagella are assembled. 00:20:36.040 --> 00:20:44.320 align:middle line:84% And if you use these refined more these kinetic parameters-- 00:20:44.320 --> 00:20:47.260 align:middle line:84% that actually were significantly different from some of it that 00:20:47.260 --> 00:20:48.613 align:middle line:90% was measured before-- 00:20:48.613 --> 00:20:50.530 align:middle line:84% then you have a much more accurate description 00:20:50.530 --> 00:20:56.020 align:middle line:84% of the dynamics-- how the flagella are assembled. 00:20:56.020 --> 00:21:01.600 align:middle line:84% So they are kind of proving at the level 00:21:01.600 --> 00:21:04.030 align:middle line:84% of the overall functional module. 00:21:04.030 --> 00:21:08.720 align:middle line:84% AUDIENCE: Do you see a possibility in that case-- 00:21:08.720 --> 00:21:13.975 align:middle line:84% assuming that two proteins, particularly the network 00:21:13.975 --> 00:21:19.630 align:middle line:84% [INAUDIBLE],, are important for genetic function. 00:21:19.630 --> 00:21:24.370 align:middle line:84% Assuming that those are tweaked by the GFP in different ways 00:21:24.370 --> 00:21:29.380 align:middle line:84% so that maybe the implication of looking at the GFP 00:21:29.380 --> 00:21:30.800 align:middle line:90% tag on one protein [INAUDIBLE]. 00:21:30.800 --> 00:21:32.217 align:middle line:84% ZOLTAN SZALLASI: So the point is-- 00:21:32.217 --> 00:21:32.920 align:middle line:90% OK, I see. 00:21:32.920 --> 00:21:36.340 align:middle line:84% AUDIENCE: Can you [INAUDIBLE] de-convolute the amount 00:21:36.340 --> 00:21:36.910 align:middle line:90% of error? 00:21:36.910 --> 00:21:37.150 align:middle line:90% ZOLTAN SZALLASI: Yeah, yeah. 00:21:37.150 --> 00:21:38.800 align:middle line:84% So what they were actually doing was, 00:21:38.800 --> 00:21:42.910 align:middle line:84% the GFP is not really trying to measure the protein. 00:21:42.910 --> 00:21:45.860 align:middle line:84% It's measuring the activity of the translation. 00:21:45.860 --> 00:21:49.390 align:middle line:84% So when you have a GFP tagged protein for, 00:21:49.390 --> 00:21:52.660 align:middle line:84% let's say protein A, then what they are really interested in 00:21:52.660 --> 00:21:55.630 align:middle line:84% is how fast that protein is activated 00:21:55.630 --> 00:21:59.150 align:middle line:84% by the non-tagged transcription factors. 00:21:59.150 --> 00:22:01.720 align:middle line:84% And what they were interested in, 00:22:01.720 --> 00:22:04.270 align:middle line:84% the parameters they wanted to refine, 00:22:04.270 --> 00:22:09.380 align:middle line:90% is the activation of that gene. 00:22:09.380 --> 00:22:13.090 align:middle line:90% So the GFP protein is simply a-- 00:22:13.090 --> 00:22:14.590 align:middle line:84% AUDIENCE: So it says nothing about-- 00:22:14.590 --> 00:22:15.757 align:middle line:90% ZOLTAN SZALLASI: --a marker. 00:22:15.757 --> 00:22:17.615 align:middle line:84% AUDIENCE: It says nothing about the actual-- 00:22:17.615 --> 00:22:19.990 align:middle line:84% ZOLTAN SZALLASI: If you have a very quick, very immediate 00:22:19.990 --> 00:22:22.750 align:middle line:84% feedback loop, and it interferes with that, 00:22:22.750 --> 00:22:24.500 align:middle line:90% yeah, that's a problem. 00:22:24.500 --> 00:22:27.580 align:middle line:84% But if you do not have that, you have much larger feedback 00:22:27.580 --> 00:22:29.380 align:middle line:84% for many, many, many different proteins. 00:22:29.380 --> 00:22:33.190 align:middle line:84% All they wanted to see, how that actual protein is activated. 00:22:33.190 --> 00:22:37.120 align:middle line:84% And the activators of that transcription unit 00:22:37.120 --> 00:22:39.310 align:middle line:90% is unaffected. 00:22:39.310 --> 00:22:42.310 align:middle line:84% There is no GFP tag on those proteins. 00:22:42.310 --> 00:22:43.810 align:middle line:84% AUDIENCE: Well I think my question-- 00:22:43.810 --> 00:22:46.810 align:middle line:84% and I'll just modify it [INAUDIBLE] back and forth. 00:22:46.810 --> 00:22:48.430 align:middle line:84% Is there a way to-- so let's say there 00:22:48.430 --> 00:22:52.415 align:middle line:84% are two competing regulatory proteins for transcription. 00:22:52.415 --> 00:22:53.290 align:middle line:90% ZOLTAN SZALLASI: Yes. 00:22:53.290 --> 00:22:53.770 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:22:53.770 --> 00:22:54.940 align:middle line:84% ZOLTAN SZALLASI: Of a certain protein 00:22:54.940 --> 00:22:56.320 align:middle line:90% that is GFP labeled, right? 00:22:56.320 --> 00:22:59.080 align:middle line:90% 00:22:59.080 --> 00:23:08.290 align:middle line:84% AUDIENCE: And you get differing regulation of-- 00:23:08.290 --> 00:23:09.520 align:middle line:90% I'm sorry. 00:23:09.520 --> 00:23:10.510 align:middle line:90% I screwed that up. 00:23:10.510 --> 00:23:15.800 align:middle line:84% Two proteins that are GFP tagged that you know are [INAUDIBLE] 00:23:15.800 --> 00:23:18.220 align:middle line:84% and you get different regulation of those genes-- 00:23:18.220 --> 00:23:19.095 align:middle line:90% ZOLTAN SZALLASI: Yes. 00:23:19.095 --> 00:23:20.230 align:middle line:90% AUDIENCE: --from this. 00:23:20.230 --> 00:23:23.560 align:middle line:84% Do you assume that there's a way to actually de-convolute what 00:23:23.560 --> 00:23:26.620 align:middle line:90% the actual right-- 00:23:26.620 --> 00:23:28.712 align:middle line:84% like, so there's an error associated 00:23:28.712 --> 00:23:29.920 align:middle line:90% with the tagging [INAUDIBLE]. 00:23:29.920 --> 00:23:30.795 align:middle line:90% ZOLTAN SZALLASI: Yes. 00:23:30.795 --> 00:23:33.820 align:middle line:84% AUDIENCE: And you're going to get two different expression 00:23:33.820 --> 00:23:36.280 align:middle line:84% profiles for both of those proteins in each experiment 00:23:36.280 --> 00:23:37.730 align:middle line:90% even though only one is tagged. 00:23:37.730 --> 00:23:38.260 align:middle line:84% ZOLTAN SZALLASI: Well, the only way 00:23:38.260 --> 00:23:40.593 align:middle line:84% you would do it is actually you take two different tags. 00:23:40.593 --> 00:23:42.240 align:middle line:84% Probably that's the best way to go. 00:23:42.240 --> 00:23:45.630 align:middle line:84% From one set of measurements, you can attempt those things. 00:23:45.630 --> 00:23:47.040 align:middle line:90% But-- 00:23:47.040 --> 00:23:49.080 align:middle line:84% AUDIENCE: Would that give-- is it possible then 00:23:49.080 --> 00:23:55.200 align:middle line:84% to get information out about the degree of perturbation 00:23:55.200 --> 00:23:58.230 align:middle line:84% of the actual system from using that double tagged system? 00:23:58.230 --> 00:23:59.190 align:middle line:90% Do you understand-- 00:23:59.190 --> 00:24:00.065 align:middle line:90% ZOLTAN SZALLASI: Yes. 00:24:00.065 --> 00:24:00.960 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:24:00.960 --> 00:24:02.418 align:middle line:84% ZOLTAN SZALLASI: I'm not quite sure 00:24:02.418 --> 00:24:05.820 align:middle line:90% whether it's worth the effort. 00:24:05.820 --> 00:24:07.200 align:middle line:90% What they are doing is-- 00:24:07.200 --> 00:24:07.860 align:middle line:90% as I said-- 00:24:07.860 --> 00:24:11.550 align:middle line:90% 00:24:11.550 --> 00:24:15.030 align:middle line:84% AUDIENCE: Because right now they don't have an estimate on how 00:24:15.030 --> 00:24:15.720 align:middle line:90% accurate their-- 00:24:15.720 --> 00:24:16.110 align:middle line:90% ZOLTAN SZALLASI: Yes. 00:24:16.110 --> 00:24:16.800 align:middle line:84% AUDIENCE: --parameters are, right? 00:24:16.800 --> 00:24:18.600 align:middle line:90% ZOLTAN SZALLASI: Absolutely. 00:24:18.600 --> 00:24:21.330 align:middle line:84% So what they do is they try to estimate 00:24:21.330 --> 00:24:25.020 align:middle line:84% whether their effort is worth or not based 00:24:25.020 --> 00:24:30.130 align:middle line:84% on the outcome of the entire modeling effort, 00:24:30.130 --> 00:24:33.960 align:middle line:84% and with issues like whether the description is a robust network 00:24:33.960 --> 00:24:36.570 align:middle line:90% or not. 00:24:36.570 --> 00:24:39.660 align:middle line:84% Every single measurement will give you 00:24:39.660 --> 00:24:43.560 align:middle line:90% a somewhat biased parameter. 00:24:43.560 --> 00:24:46.290 align:middle line:84% So if you measure a single parameter 00:24:46.290 --> 00:24:49.170 align:middle line:84% in different settings, they will always be different. 00:24:49.170 --> 00:24:51.090 align:middle line:90% What is the error there? 00:24:51.090 --> 00:24:56.490 align:middle line:84% The point is that with this way, you can refine parameters 00:24:56.490 --> 00:24:57.720 align:middle line:90% to some extent. 00:24:57.720 --> 00:25:00.090 align:middle line:84% And that was a significant improvement. 00:25:00.090 --> 00:25:02.940 align:middle line:84% Sometimes it is like an order of two decimal difference 00:25:02.940 --> 00:25:06.090 align:middle line:84% from what they thought it was before. 00:25:06.090 --> 00:25:08.430 align:middle line:84% Based on that, they have a more accurate description 00:25:08.430 --> 00:25:10.360 align:middle line:90% of the system. 00:25:10.360 --> 00:25:15.130 align:middle line:84% Now within certain error range, very different parameters 00:25:15.130 --> 00:25:16.630 align:middle line:84% would give you the very same result. 00:25:16.630 --> 00:25:19.200 align:middle line:84% This is what you expect from robustness. 00:25:19.200 --> 00:25:20.917 align:middle line:90% We are getting the error. 00:25:20.917 --> 00:25:23.250 align:middle line:84% You just missed the first couple of slides when I said-- 00:25:23.250 --> 00:25:27.040 align:middle line:90% 00:25:27.040 --> 00:25:28.990 align:middle line:84% if you keep playing with parameters, 00:25:28.990 --> 00:25:31.850 align:middle line:84% you can very easily overfit your entire system. 00:25:31.850 --> 00:25:33.850 align:middle line:84% So I'm not quite sure whether that's-- 00:25:33.850 --> 00:25:36.140 align:middle line:84% yes, for individual process, you can do it. 00:25:36.140 --> 00:25:39.400 align:middle line:84% And obviously, if you do very different taggings and very 00:25:39.400 --> 00:25:41.770 align:middle line:84% different types of measurements for the same proteins, 00:25:41.770 --> 00:25:44.020 align:middle line:84% then in some way you can de-convolute-- 00:25:44.020 --> 00:25:47.890 align:middle line:84% or at least you can catch very obvious contradictions. 00:25:47.890 --> 00:25:49.480 align:middle line:84% And you say, well for this parameter, 00:25:49.480 --> 00:25:50.920 align:middle line:90% I have no idea what's going on. 00:25:50.920 --> 00:25:53.620 align:middle line:84% Or for this parameter, I need to measure it individually. 00:25:53.620 --> 00:25:55.330 align:middle line:84% Because if you develop an antibody 00:25:55.330 --> 00:25:57.740 align:middle line:84% or some other measurement technique, 00:25:57.740 --> 00:25:59.990 align:middle line:84% then probably you can measure it much more accurately. 00:25:59.990 --> 00:26:01.615 align:middle line:84% So probably that would be the best way. 00:26:01.615 --> 00:26:05.139 align:middle line:84% For most proteins, probably you won't have that many problems. 00:26:05.139 --> 00:26:05.806 align:middle line:90% AUDIENCE: I see. 00:26:05.806 --> 00:26:08.502 align:middle line:84% So within the network, the parameters are defined in a way 00:26:08.502 --> 00:26:10.585 align:middle line:84% that it doesn't matter to the individual parameter 00:26:10.585 --> 00:26:10.870 align:middle line:90% [INAUDIBLE]-- 00:26:10.870 --> 00:26:11.800 align:middle line:90% ZOLTAN SZALLASI: Exactly. 00:26:11.800 --> 00:26:12.730 align:middle line:84% AUDIENCE: --how they combine [INAUDIBLE].. 00:26:12.730 --> 00:26:13.150 align:middle line:90% ZOLTAN SZALLASI: Yes. 00:26:13.150 --> 00:26:15.192 align:middle line:84% AUDIENCE: And making sure that that [INAUDIBLE].. 00:26:15.192 --> 00:26:16.540 align:middle line:90% ZOLTAN SZALLASI: Yes, yes, yes. 00:26:16.540 --> 00:26:20.200 align:middle line:84% And that's kind of the whole gist of large scale modeling. 00:26:20.200 --> 00:26:21.770 align:middle line:90% How much effort do you need? 00:26:21.770 --> 00:26:23.410 align:middle line:84% How much information do you put in 00:26:23.410 --> 00:26:27.670 align:middle line:84% to get some sort of a reasonably accurate prediction? 00:26:27.670 --> 00:26:28.390 align:middle line:90% OK? 00:26:28.390 --> 00:26:29.900 align:middle line:90% So let's move on to that. 00:26:29.900 --> 00:26:32.140 align:middle line:84% So what do you need to do for modeling? 00:26:32.140 --> 00:26:37.630 align:middle line:84% The obvious steps, you have to ask-- this is like step zero. 00:26:37.630 --> 00:26:40.540 align:middle line:84% You are going to ask interesting questions. 00:26:40.540 --> 00:26:42.643 align:middle line:84% You need to collect good quality data. 00:26:42.643 --> 00:26:44.560 align:middle line:84% You need to create an appropriate mathematical 00:26:44.560 --> 00:26:46.102 align:middle line:84% logical representation of the system. 00:26:46.102 --> 00:26:48.070 align:middle line:84% That's what we have talked so far about. 00:26:48.070 --> 00:26:51.040 align:middle line:84% We need to run simulation and test how well the simulation is 00:26:51.040 --> 00:26:52.090 align:middle line:90% fitting the data. 00:26:52.090 --> 00:26:55.210 align:middle line:84% Test for robustness or for other network properties. 00:26:55.210 --> 00:26:58.240 align:middle line:84% You update the model based on how well your model 00:26:58.240 --> 00:27:03.320 align:middle line:84% description is describing the data and then use it. 00:27:03.320 --> 00:27:04.684 align:middle line:90% So if you-- 00:27:04.684 --> 00:27:05.726 align:middle line:90% AUDIENCE: Quick question. 00:27:05.726 --> 00:27:07.127 align:middle line:90% Can you define robustness? 00:27:07.127 --> 00:27:08.710 align:middle line:84% ZOLTAN SZALLASI: We are getting there. 00:27:08.710 --> 00:27:09.252 align:middle line:90% AUDIENCE: OK. 00:27:09.252 --> 00:27:12.610 align:middle line:84% ZOLTAN SZALLASI: If-- well, it's relative 00:27:12.610 --> 00:27:14.830 align:middle line:84% intersensitivity to individual parameter 00:27:14.830 --> 00:27:16.420 align:middle line:90% values for the entire network. 00:27:16.420 --> 00:27:21.530 align:middle line:90% That's kind of the definition. 00:27:21.530 --> 00:27:22.600 align:middle line:90% So you have a network. 00:27:22.600 --> 00:27:24.700 align:middle line:84% Now you still have to ask another question. 00:27:24.700 --> 00:27:27.430 align:middle line:84% How does this network really behave? 00:27:27.430 --> 00:27:31.760 align:middle line:84% So essentially your question is, what is this equation here? 00:27:31.760 --> 00:27:36.070 align:middle line:84% And in most cases, what people think about is-- 00:27:36.070 --> 00:27:43.084 align:middle line:84% and this is the only models that have been really employed 00:27:43.084 --> 00:27:46.630 align:middle line:84% in some sort of a realistic way to describe 00:27:46.630 --> 00:27:47.680 align:middle line:90% biological systems-- 00:27:47.680 --> 00:27:49.692 align:middle line:84% is just using ordinary differential equations. 00:27:49.692 --> 00:27:51.400 align:middle line:84% Of course, there are other ways of doing. 00:27:51.400 --> 00:27:54.790 align:middle line:84% You can use differential equations for discrete steps. 00:27:54.790 --> 00:27:57.830 align:middle line:84% The point here is these are deterministic systems. 00:27:57.830 --> 00:28:00.910 align:middle line:84% So the question is, is biology a deterministic system? 00:28:00.910 --> 00:28:04.770 align:middle line:84% Or is it true that a deterministic description 00:28:04.770 --> 00:28:06.520 align:middle line:84% of biological network is going to give you 00:28:06.520 --> 00:28:09.640 align:middle line:90% something meaningful? 00:28:09.640 --> 00:28:12.640 align:middle line:84% Deterministic system-- you all know what it is. 00:28:12.640 --> 00:28:17.620 align:middle line:84% There is only one outcome from a previous parameter setting. 00:28:17.620 --> 00:28:22.240 align:middle line:84% But biology, as we know, is most probably not 00:28:22.240 --> 00:28:24.910 align:middle line:90% that highly deterministic. 00:28:24.910 --> 00:28:27.490 align:middle line:84% Actually, we know for a fact that individual biology 00:28:27.490 --> 00:28:29.200 align:middle line:84% calculations are always stochastic. 00:28:29.200 --> 00:28:31.660 align:middle line:84% It's another question how the biological network behaves. 00:28:31.660 --> 00:28:32.830 align:middle line:84% But for individual reactions, you 00:28:32.830 --> 00:28:34.455 align:middle line:84% know that you have very few molecules-- 00:28:34.455 --> 00:28:37.930 align:middle line:84% less than 400 transcription factors in a nucleus. 00:28:37.930 --> 00:28:39.550 align:middle line:84% Fairly far from a quick solution. 00:28:39.550 --> 00:28:41.350 align:middle line:90% Reaction can be often slow. 00:28:41.350 --> 00:28:44.530 align:middle line:84% So there's all the hallmarks of stochastic reactions 00:28:44.530 --> 00:28:47.810 align:middle line:84% that chemical engineers are all too familiar with. 00:28:47.810 --> 00:28:52.298 align:middle line:84% So in that case, you have to do different type of modeling. 00:28:52.298 --> 00:28:54.340 align:middle line:84% You want to use stochastic differential equations 00:28:54.340 --> 00:28:56.048 align:middle line:84% where you have kind of a stochastic error 00:28:56.048 --> 00:28:58.580 align:middle line:84% in differentiable equations-- which is kind of cheating. 00:28:58.580 --> 00:29:01.780 align:middle line:84% The real thing is what a long time ago, Gillespie 00:29:01.780 --> 00:29:04.370 align:middle line:90% introduced-- like 30 years ago. 00:29:04.370 --> 00:29:07.300 align:middle line:84% They actually do a stochastic modeling of individual chemical 00:29:07.300 --> 00:29:08.120 align:middle line:90% reactions. 00:29:08.120 --> 00:29:10.940 align:middle line:84% This is the thing that is actually closest to reality. 00:29:10.940 --> 00:29:14.210 align:middle line:84% Of course, the problem is that computation is so difficult. 00:29:14.210 --> 00:29:15.010 align:middle line:90% It's so intensive. 00:29:15.010 --> 00:29:20.710 align:middle line:84% And there is-- unless there's a good reason 00:29:20.710 --> 00:29:22.720 align:middle line:84% that you need to do this type of modeling, 00:29:22.720 --> 00:29:25.155 align:middle line:84% then this is not really an economical approach. 00:29:25.155 --> 00:29:27.280 align:middle line:84% Now there are other tricks people are trying to do. 00:29:27.280 --> 00:29:29.260 align:middle line:84% Like John Doyle is trying at Caltech 00:29:29.260 --> 00:29:31.960 align:middle line:84% with [INAUDIBLE] who came back from retirement, to do 00:29:31.960 --> 00:29:33.340 align:middle line:90% something called a tall leap. 00:29:33.340 --> 00:29:36.490 align:middle line:84% So this is like a really detailed description 00:29:36.490 --> 00:29:37.990 align:middle line:84% of how things are hitting each other 00:29:37.990 --> 00:29:39.580 align:middle line:84% and how chemical reactions develop. 00:29:39.580 --> 00:29:44.290 align:middle line:84% Now can you guess where the reaction is going 00:29:44.290 --> 00:29:45.940 align:middle line:90% to be in, let's say, 10 steps? 00:29:45.940 --> 00:29:48.020 align:middle line:84% You can tell what's going to happen in one step 00:29:48.020 --> 00:29:49.930 align:middle line:90% if you do a stochastic modeling. 00:29:49.930 --> 00:29:51.910 align:middle line:84% Can you guess what happens in 10 steps? 00:29:51.910 --> 00:29:53.530 align:middle line:84% So they're trying to cheat and they 00:29:53.530 --> 00:29:55.450 align:middle line:90% are trying to see whether-- 00:29:55.450 --> 00:29:58.720 align:middle line:84% if you think about like, ODE is going 00:29:58.720 --> 00:30:00.940 align:middle line:84% to describe very simply what's going to happen. 00:30:00.940 --> 00:30:03.580 align:middle line:84% The stochastic equation is going to jump around 00:30:03.580 --> 00:30:06.400 align:middle line:84% this in some way, and you will have alternative outcomes 00:30:06.400 --> 00:30:07.448 align:middle line:90% of the same reaction. 00:30:07.448 --> 00:30:09.490 align:middle line:84% With the tall leap, what they are trying to see-- 00:30:09.490 --> 00:30:11.410 align:middle line:84% they just jump from here to here. 00:30:11.410 --> 00:30:13.325 align:middle line:90% And there has to be certain-- 00:30:13.325 --> 00:30:14.950 align:middle line:84% what is the probability that the system 00:30:14.950 --> 00:30:19.510 align:middle line:84% is going to be in this state, in let's say 10 time units, 00:30:19.510 --> 00:30:21.220 align:middle line:90% or something like that? 00:30:21.220 --> 00:30:23.350 align:middle line:84% So this is, again, a wide open question. 00:30:23.350 --> 00:30:26.210 align:middle line:84% The big question is, is this really an important thing, 00:30:26.210 --> 00:30:29.210 align:middle line:84% that we think about biology and model it 00:30:29.210 --> 00:30:30.770 align:middle line:90% in a stochastic fashion? 00:30:30.770 --> 00:30:32.720 align:middle line:84% Now let me just show you an example 00:30:32.720 --> 00:30:35.500 align:middle line:84% from real experiments that shows that, yes, it's important. 00:30:35.500 --> 00:30:38.360 align:middle line:84% There was a very nice experiment in which they created 00:30:38.360 --> 00:30:39.830 align:middle line:90% a plasmid of two proteins. 00:30:39.830 --> 00:30:42.560 align:middle line:84% They were tagged with two different GFPs. 00:30:42.560 --> 00:30:45.590 align:middle line:84% And this was a very carefully controlled experiment, 00:30:45.590 --> 00:30:49.070 align:middle line:84% when two proteins were translated. 00:30:49.070 --> 00:30:52.520 align:middle line:84% On the same plasmid in every single bacteria, 00:30:52.520 --> 00:30:54.470 align:middle line:84% and everything was driven gene expression 00:30:54.470 --> 00:30:56.540 align:middle line:84% by the very, very same promoters. 00:30:56.540 --> 00:31:00.595 align:middle line:84% And everything was in principle, identical in every single cell. 00:31:00.595 --> 00:31:02.720 align:middle line:84% So what you would expect in the relative expression 00:31:02.720 --> 00:31:07.140 align:middle line:84% of these proteins should be the same in every single cell. 00:31:07.140 --> 00:31:10.860 align:middle line:84% So we should show the same kind of mixture of red and green. 00:31:10.860 --> 00:31:14.120 align:middle line:90% So that was the expectation. 00:31:14.120 --> 00:31:15.560 align:middle line:84% And instead of this, what you see 00:31:15.560 --> 00:31:18.590 align:middle line:84% is there are very red, very green, and yellow cells, 00:31:18.590 --> 00:31:25.430 align:middle line:84% showing that individual cells will choose very different 00:31:25.430 --> 00:31:26.540 align:middle line:90% pathway-- or path. 00:31:26.540 --> 00:31:30.440 align:middle line:84% So there will be cells that will have much more of one 00:31:30.440 --> 00:31:31.940 align:middle line:84% of the proteins than the other ones, 00:31:31.940 --> 00:31:34.190 align:middle line:84% although everything was identical-- as much 00:31:34.190 --> 00:31:37.160 align:middle line:84% as we can control experimentally. 00:31:37.160 --> 00:31:38.910 align:middle line:84% We know stochastics is important. 00:31:38.910 --> 00:31:40.460 align:middle line:84% We know that probably differentiation 00:31:40.460 --> 00:31:42.620 align:middle line:90% is driven by specificity. 00:31:42.620 --> 00:31:46.430 align:middle line:84% You have a parent cell, and after division you 00:31:46.430 --> 00:31:48.110 align:middle line:84% have two cells that are going to go 00:31:48.110 --> 00:31:49.340 align:middle line:90% two complete different paths. 00:31:49.340 --> 00:31:51.890 align:middle line:84% One remains a stem cell, the other guy 00:31:51.890 --> 00:31:53.240 align:middle line:90% goes the opposite direction. 00:31:53.240 --> 00:31:54.770 align:middle line:84% And that might be just simply driven 00:31:54.770 --> 00:31:58.850 align:middle line:84% how certain proteins are distributed randomly 00:31:58.850 --> 00:32:02.000 align:middle line:84% between the two daughter cells during the cell division. 00:32:02.000 --> 00:32:04.940 align:middle line:84% It is known actually that for many differentiation steps, 00:32:04.940 --> 00:32:08.030 align:middle line:84% probably the decision whether a cell goes one way or other 00:32:08.030 --> 00:32:10.310 align:middle line:84% is going to be a stochastic process. 00:32:10.310 --> 00:32:12.320 align:middle line:84% Now whether this is important for modeling 00:32:12.320 --> 00:32:19.500 align:middle line:84% the cell cycle, cell growth, and other features, we do not know. 00:32:19.500 --> 00:32:20.000 align:middle line:90% Yes? 00:32:20.000 --> 00:32:21.080 align:middle line:84% AUDIENCE: Might be a silly question. 00:32:21.080 --> 00:32:22.455 align:middle line:84% But how do you define stochastic? 00:32:22.455 --> 00:32:23.862 align:middle line:90% Is it just randomness or-- 00:32:23.862 --> 00:32:25.820 align:middle line:84% ZOLTAN SZALLASI: Oh, it's a very good question. 00:32:25.820 --> 00:32:33.310 align:middle line:90% So-- well in a way-- 00:32:33.310 --> 00:32:35.290 align:middle line:90% OK. 00:32:35.290 --> 00:32:37.780 align:middle line:84% How would you define stochasticity? 00:32:37.780 --> 00:32:39.490 align:middle line:84% Again, this is the representation 00:32:39.490 --> 00:32:41.590 align:middle line:90% of scientific knowledge. 00:32:41.590 --> 00:32:44.770 align:middle line:84% So we describe temperature with a single number, 00:32:44.770 --> 00:32:47.200 align:middle line:84% although you know very well that-- 00:32:47.200 --> 00:32:48.075 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:32:48.075 --> 00:32:50.533 align:middle line:84% ZOLTAN SZALLASI: Yeah, you have individual kinetic energies 00:32:50.533 --> 00:32:51.290 align:middle line:90% of all molecules. 00:32:51.290 --> 00:32:52.930 align:middle line:84% So if you have many molecules, it doesn't really matter. 00:32:52.930 --> 00:32:53.560 align:middle line:90% You give temperature. 00:32:53.560 --> 00:32:56.210 align:middle line:84% If you have 10 molecules, the whole issue is very different. 00:32:56.210 --> 00:32:58.300 align:middle line:84% So stochasticity is that you go down 00:32:58.300 --> 00:33:01.330 align:middle line:84% one more level of detail description 00:33:01.330 --> 00:33:04.270 align:middle line:84% of the individual reactions, that will happen 00:33:04.270 --> 00:33:05.780 align:middle line:90% with a certain probability. 00:33:05.780 --> 00:33:08.230 align:middle line:84% So you can call it randomness, but the point 00:33:08.230 --> 00:33:10.240 align:middle line:84% is that most of these things as we understand, 00:33:10.240 --> 00:33:12.260 align:middle line:84% will occur with a certain probability. 00:33:12.260 --> 00:33:15.400 align:middle line:84% So probabilistic descriptions of individual molecular reactions 00:33:15.400 --> 00:33:18.010 align:middle line:84% are much more accurate and much, much closer to reality 00:33:18.010 --> 00:33:22.030 align:middle line:84% than continuous-- continuous differential equations are 00:33:22.030 --> 00:33:25.810 align:middle line:84% nice, and they were invented by Newton-Leibniz 00:33:25.810 --> 00:33:27.370 align:middle line:90% to do useful things. 00:33:27.370 --> 00:33:29.718 align:middle line:84% But I don't think that even they thought that this was 00:33:29.718 --> 00:33:31.510 align:middle line:84% an accurate-- well, probably they thought-- 00:33:31.510 --> 00:33:34.060 align:middle line:84% but for a long time, we know that this is not a very 00:33:34.060 --> 00:33:35.620 align:middle line:90% accurate description of-- 00:33:35.620 --> 00:33:38.630 align:middle line:84% certainly not since quantum chemistry and physics 00:33:38.630 --> 00:33:39.130 align:middle line:90% is around. 00:33:39.130 --> 00:33:43.027 align:middle line:90% 00:33:43.027 --> 00:33:44.610 align:middle line:84% So once you have a description and you 00:33:44.610 --> 00:33:47.010 align:middle line:84% decide which way to go-- so this slide just showed you 00:33:47.010 --> 00:33:49.050 align:middle line:84% that we have evidence that stochasticity 00:33:49.050 --> 00:33:50.280 align:middle line:90% is present in biology. 00:33:50.280 --> 00:33:51.450 align:middle line:90% We have lots of evidence. 00:33:51.450 --> 00:33:53.670 align:middle line:84% I just showed like one experiment. 00:33:53.670 --> 00:33:56.040 align:middle line:84% Whether it's worth doing for modeling 00:33:56.040 --> 00:33:58.410 align:middle line:84% and not going with all these, that's an open question. 00:33:58.410 --> 00:34:00.690 align:middle line:84% Probably it's going to determine which-- 00:34:00.690 --> 00:34:02.010 align:middle line:90% not, what's your question. 00:34:02.010 --> 00:34:04.052 align:middle line:84% That's why it's very important that you formulate 00:34:04.052 --> 00:34:05.500 align:middle line:90% your question appropriately. 00:34:05.500 --> 00:34:07.120 align:middle line:90% Now, once you decided-- 00:34:07.120 --> 00:34:09.480 align:middle line:84% so you have the network, topology, kinetic parameters, 00:34:09.480 --> 00:34:10.590 align:middle line:90% equations, everything. 00:34:10.590 --> 00:34:12.420 align:middle line:84% Then you actually start to run simulations 00:34:12.420 --> 00:34:13.900 align:middle line:90% and see what happens. 00:34:13.900 --> 00:34:16.050 align:middle line:84% So let me show this example that-- actually, this 00:34:16.050 --> 00:34:19.170 align:middle line:84% can describe certain biological phenomena. 00:34:19.170 --> 00:34:21.900 align:middle line:84% And you can actually formulate useful hypotheses 00:34:21.900 --> 00:34:23.030 align:middle line:90% based on this. 00:34:23.030 --> 00:34:27.810 align:middle line:84% This was something called a nucleocytoplasmic transport. 00:34:27.810 --> 00:34:29.610 align:middle line:84% This is an active reaction itself. 00:34:29.610 --> 00:34:32.139 align:middle line:84% Proteins get in the nucleus and they can't come out. 00:34:32.139 --> 00:34:35.790 align:middle line:84% And that's actually driven by this regulatory-- 00:34:35.790 --> 00:34:38.460 align:middle line:90% or this dynamic network-- 00:34:38.460 --> 00:34:39.929 align:middle line:90% involving a group of proteins. 00:34:39.929 --> 00:34:42.370 align:middle line:84% Essentially this protein is Rand protein, 00:34:42.370 --> 00:34:46.510 align:middle line:90% which is a G protein reaction. 00:34:46.510 --> 00:34:48.940 align:middle line:90% So it uses GTP as an energy. 00:34:48.940 --> 00:34:50.070 align:middle line:90% It's an active reaction. 00:34:50.070 --> 00:34:51.690 align:middle line:84% This is what is actually moving things 00:34:51.690 --> 00:34:54.270 align:middle line:90% in and out from the cytoplasm. 00:34:54.270 --> 00:34:56.159 align:middle line:84% When people started to think about modeling, 00:34:56.159 --> 00:34:58.350 align:middle line:90% they have a strong assumption-- 00:34:58.350 --> 00:35:01.170 align:middle line:84% everybody does-- that if you have a network that's 00:35:01.170 --> 00:35:04.080 align:middle line:84% not very stochastic, that's sufficiently isolated 00:35:04.080 --> 00:35:06.690 align:middle line:84% from the rest of the network, and we understand it 00:35:06.690 --> 00:35:08.490 align:middle line:84% well enough, then probably that network 00:35:08.490 --> 00:35:13.170 align:middle line:84% can be described pretty well with a mathematical model. 00:35:13.170 --> 00:35:19.708 align:middle line:84% And you can probably make pretty good predictions based on that. 00:35:19.708 --> 00:35:21.750 align:middle line:84% So these people actually who did this experiment, 00:35:21.750 --> 00:35:24.860 align:middle line:84% they were experts in the nucleocytoplasmic transport 00:35:24.860 --> 00:35:25.360 align:middle line:90% system. 00:35:25.360 --> 00:35:27.180 align:middle line:84% That's why they knew that probably there 00:35:27.180 --> 00:35:30.190 align:middle line:84% were not that many blank spots in this network. 00:35:30.190 --> 00:35:30.978 align:middle line:90% So what they did-- 00:35:30.978 --> 00:35:32.520 align:middle line:84% they took all the kinetic parameters. 00:35:32.520 --> 00:35:34.145 align:middle line:84% They already measured it several times. 00:35:34.145 --> 00:35:35.910 align:middle line:84% This is not such a large network. 00:35:35.910 --> 00:35:39.240 align:middle line:84% They created that and they started to do measurements. 00:35:39.240 --> 00:35:41.490 align:middle line:84% They actually measured pretty well-- 00:35:41.490 --> 00:35:44.587 align:middle line:84% pretty accurately-- how quickly proteins enter the nucleus. 00:35:44.587 --> 00:35:45.420 align:middle line:90% This is the nucleus. 00:35:45.420 --> 00:35:47.890 align:middle line:84% It's empty and then it's getting into the nucleus. 00:35:47.890 --> 00:35:51.250 align:middle line:84% So they can measure it quite accurately. 00:35:51.250 --> 00:35:53.880 align:middle line:84% So the first question you have, does the model 00:35:53.880 --> 00:35:55.980 align:middle line:84% produce time series of charts that feed the data? 00:35:55.980 --> 00:35:57.640 align:middle line:84% And they did all these experiments. 00:35:57.640 --> 00:36:00.600 align:middle line:84% And these are the measurements and these are the predictions. 00:36:00.600 --> 00:36:02.700 align:middle line:84% And yes, it seemed that the model 00:36:02.700 --> 00:36:05.068 align:middle line:84% produced fairly good descriptions-- fairly accurate 00:36:05.068 --> 00:36:05.610 align:middle line:90% descriptions. 00:36:05.610 --> 00:36:07.158 align:middle line:90% So that is good news. 00:36:07.158 --> 00:36:08.700 align:middle line:84% Now the next question you ask is, OK. 00:36:08.700 --> 00:36:10.950 align:middle line:84% I describe a certain experimental setting, 00:36:10.950 --> 00:36:12.930 align:middle line:90% but is the model robust? 00:36:12.930 --> 00:36:14.850 align:middle line:84% And how sensitive is the parameters? 00:36:14.850 --> 00:36:20.010 align:middle line:84% That is, if you have the right ratio of parameters 00:36:20.010 --> 00:36:21.120 align:middle line:90% and measurements-- 00:36:21.120 --> 00:36:23.460 align:middle line:84% I'm sure you understand this-- you can always 00:36:23.460 --> 00:36:24.450 align:middle line:90% overfit the system. 00:36:24.450 --> 00:36:27.058 align:middle line:84% You can always find a set of parameters that 00:36:27.058 --> 00:36:28.350 align:middle line:90% will give you the right result. 00:36:28.350 --> 00:36:30.718 align:middle line:90% Now is this anything useful? 00:36:30.718 --> 00:36:31.260 align:middle line:90% Probably not. 00:36:31.260 --> 00:36:33.010 align:middle line:84% Because that's what you'd like to predict. 00:36:33.010 --> 00:36:35.740 align:middle line:84% What's going to happen if you take a very different parameter 00:36:35.740 --> 00:36:36.240 align:middle line:90% setting? 00:36:36.240 --> 00:36:39.880 align:middle line:84% Or you put in a new drug or combinatorial treatment? 00:36:39.880 --> 00:36:42.930 align:middle line:84% So what these people do-- and in this case you need to do-- 00:36:42.930 --> 00:36:44.010 align:middle line:90% is check for sensitivity. 00:36:44.010 --> 00:36:48.132 align:middle line:84% So robustness is, you check the outcome 00:36:48.132 --> 00:36:50.340 align:middle line:84% of the behavior of the system-- whatever you measure. 00:36:50.340 --> 00:36:52.770 align:middle line:84% For example, in this case, how quick was the dynamic 00:36:52.770 --> 00:36:54.780 align:middle line:84% that things get into the nucleus. 00:36:54.780 --> 00:36:57.990 align:middle line:84% And you start to play around with every single parameter 00:36:57.990 --> 00:37:03.002 align:middle line:84% and ask the question that, in what parameter range 00:37:03.002 --> 00:37:04.710 align:middle line:84% do you still have the same result-- let's 00:37:04.710 --> 00:37:06.880 align:middle line:90% say with a 10% error? 00:37:06.880 --> 00:37:08.712 align:middle line:84% So these are all the parameter settings-- 00:37:08.712 --> 00:37:10.170 align:middle line:84% these are individual perturbations. 00:37:10.170 --> 00:37:12.900 align:middle line:84% These are perturbations on individual parameters 00:37:12.900 --> 00:37:16.470 align:middle line:84% at the same time that still give you 00:37:16.470 --> 00:37:19.960 align:middle line:84% the same result with a 10% error. 00:37:19.960 --> 00:37:22.420 align:middle line:90% And so that's the next step. 00:37:22.420 --> 00:37:23.790 align:middle line:90% And if you are-- 00:37:23.790 --> 00:37:25.770 align:middle line:84% see that, well, it looks fairly good. 00:37:25.770 --> 00:37:28.570 align:middle line:84% Then you start to trust your description is fairly robust. 00:37:28.570 --> 00:37:30.000 align:middle line:90% That's cool. 00:37:30.000 --> 00:37:31.740 align:middle line:84% Now the next thing you need to do 00:37:31.740 --> 00:37:34.980 align:middle line:84% is actually, can the model produce something useful-- 00:37:34.980 --> 00:37:37.210 align:middle line:84% for example, a testable hypothesis? 00:37:37.210 --> 00:37:38.610 align:middle line:90% And that's what these guys did. 00:37:38.610 --> 00:37:40.260 align:middle line:84% So they said, well, the model would 00:37:40.260 --> 00:37:43.590 align:middle line:84% predict that if you knock down this RCC-1 protein that 00:37:43.590 --> 00:37:47.040 align:middle line:84% helps RAM to perform its function, then 00:37:47.040 --> 00:37:49.060 align:middle line:84% it has to behave in a certain way. 00:37:49.060 --> 00:37:50.220 align:middle line:90% And that's what they did. 00:37:50.220 --> 00:37:51.970 align:middle line:84% They used the temperature-sensitive mutant 00:37:51.970 --> 00:37:53.340 align:middle line:90% to knock that protein out. 00:37:53.340 --> 00:37:59.100 align:middle line:84% And this is the measured dynamic or kinetics 00:37:59.100 --> 00:38:02.460 align:middle line:84% of the proteins getting into the nucleus after knockdown. 00:38:02.460 --> 00:38:03.870 align:middle line:90% And this is the simulated. 00:38:03.870 --> 00:38:07.890 align:middle line:84% So in this case, actually this describes the system 00:38:07.890 --> 00:38:11.190 align:middle line:84% fairly well after you introduce a very profound perturbation-- 00:38:11.190 --> 00:38:13.290 align:middle line:90% you knock down a protein. 00:38:13.290 --> 00:38:16.380 align:middle line:84% So these are the key steps you want to do 00:38:16.380 --> 00:38:17.730 align:middle line:90% in the case of the following. 00:38:17.730 --> 00:38:20.700 align:middle line:84% And if you have all these things, then you can assume-- 00:38:20.700 --> 00:38:24.780 align:middle line:84% or you have actually-- you can have fairly high confidence 00:38:24.780 --> 00:38:29.400 align:middle line:84% that your model is describing at least the subnetwork 00:38:29.400 --> 00:38:29.910 align:middle line:90% fairly well. 00:38:29.910 --> 00:38:31.770 align:middle line:84% AUDIENCE: [INAUDIBLE] the prediction model [INAUDIBLE]?? 00:38:31.770 --> 00:38:33.938 align:middle line:84% ZOLTAN SZALLASI: Well this is the part of-- 00:38:33.938 --> 00:38:34.563 align:middle line:90% I'm losing you. 00:38:34.563 --> 00:38:35.130 align:middle line:90% Thank you. 00:38:35.130 --> 00:38:37.350 align:middle line:84% This is the simulation, which you have on the bottom. 00:38:37.350 --> 00:38:38.690 align:middle line:90% AUDIENCE: OK, I guess-- 00:38:38.690 --> 00:38:39.190 align:middle line:90% I guess-- 00:38:39.190 --> 00:38:41.607 align:middle line:84% ZOLTAN SZALLASI: So you have two simulation states, right? 00:38:41.607 --> 00:38:44.030 align:middle line:84% That RCC-1 protein is present or absent. 00:38:44.030 --> 00:38:45.520 align:middle line:90% So that's what you have there. 00:38:45.520 --> 00:38:48.160 align:middle line:84% And you actually measure these things 00:38:48.160 --> 00:38:50.350 align:middle line:84% in the presence or absence of RCC-1. 00:38:50.350 --> 00:38:53.000 align:middle line:90% 00:38:53.000 --> 00:38:55.320 align:middle line:84% AUDIENCE: So are you able to predict like-- 00:38:55.320 --> 00:38:57.550 align:middle line:84% without actually speaking to examiners-- what is also 00:38:57.550 --> 00:38:59.490 align:middle line:84% going to be before you actually do 00:38:59.490 --> 00:39:01.080 align:middle line:90% the experiment or [INAUDIBLE]? 00:39:01.080 --> 00:39:02.580 align:middle line:90% You kept saying, like-- 00:39:02.580 --> 00:39:05.490 align:middle line:84% can you actually predict that curve 00:39:05.490 --> 00:39:07.000 align:middle line:84% without re-tweaking the parameters? 00:39:07.000 --> 00:39:08.685 align:middle line:84% ZOLTAN SZALLASI: This is not re-tweaking the parameters. 00:39:08.685 --> 00:39:09.030 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:39:09.030 --> 00:39:10.770 align:middle line:84% ZOLTAN SZALLASI: You build the model, 00:39:10.770 --> 00:39:12.150 align:middle line:90% you check for robustness. 00:39:12.150 --> 00:39:14.430 align:middle line:84% And then you do not touch the model anymore. 00:39:14.430 --> 00:39:15.180 align:middle line:90% Right? 00:39:15.180 --> 00:39:18.480 align:middle line:84% Then you start to do in silico experiments, saying, 00:39:18.480 --> 00:39:20.557 align:middle line:84% with RCC-1, how does the system behave? 00:39:20.557 --> 00:39:22.890 align:middle line:84% And then actually you measure whether the prediction was 00:39:22.890 --> 00:39:23.505 align:middle line:90% correct. 00:39:23.505 --> 00:39:26.130 align:middle line:84% And after you are done with five of these experiments, you say, 00:39:26.130 --> 00:39:27.550 align:middle line:90% I have a fairly good model. 00:39:27.550 --> 00:39:32.250 align:middle line:84% So I can start to play with this thing and ask questions. 00:39:32.250 --> 00:39:34.110 align:middle line:84% Introduce random changes-- what's 00:39:34.110 --> 00:39:36.390 align:middle line:84% the model that's going to speed up? 00:39:36.390 --> 00:39:38.460 align:middle line:84% And then the in silico model is going 00:39:38.460 --> 00:39:41.648 align:middle line:84% to give you certain answers and then you can test those. 00:39:41.648 --> 00:39:44.190 align:middle line:84% But the point is-- and that's going to be like the last slide 00:39:44.190 --> 00:39:45.750 align:middle line:90% in this lecture-- 00:39:45.750 --> 00:39:49.182 align:middle line:84% that you're trying to speed up the discovery process here. 00:39:49.182 --> 00:39:51.640 align:middle line:84% Of course, every single time you need to do the experiment. 00:39:51.640 --> 00:39:53.790 align:middle line:84% But you can go through a much larger number 00:39:53.790 --> 00:39:58.200 align:middle line:84% of possible experiments in silico than in vitro 00:39:58.200 --> 00:40:00.198 align:middle line:90% or in vivo. 00:40:00.198 --> 00:40:00.990 align:middle line:90% So that's kind of-- 00:40:00.990 --> 00:40:03.620 align:middle line:84% I've introduced this concept of measuring the robustness. 00:40:03.620 --> 00:40:07.800 align:middle line:84% The robustness is the relative insensitivity 00:40:07.800 --> 00:40:09.600 align:middle line:90% of a network to parameters. 00:40:09.600 --> 00:40:11.580 align:middle line:84% And this will answer the question 00:40:11.580 --> 00:40:13.590 align:middle line:84% that this is a relatively small network. 00:40:13.590 --> 00:40:16.913 align:middle line:84% This was still a couple of panels of parameters. 00:40:16.913 --> 00:40:18.330 align:middle line:84% But what people are trying to do-- 00:40:18.330 --> 00:40:21.390 align:middle line:84% and there are companies that try to do this. 00:40:21.390 --> 00:40:25.858 align:middle line:84% There is some company that my friends started five years ago. 00:40:25.858 --> 00:40:27.150 align:middle line:90% It's called G Network Sciences. 00:40:27.150 --> 00:40:29.100 align:middle line:84% They are physicists-- solid state physicists. 00:40:29.100 --> 00:40:31.620 align:middle line:84% And they say, well, we can model very large things. 00:40:31.620 --> 00:40:33.270 align:middle line:84% And these guys actually build networks 00:40:33.270 --> 00:40:37.980 align:middle line:84% on the order of 1,000 or 2,000 interacting things-- 00:40:37.980 --> 00:40:40.500 align:middle line:90% parameters, proteins. 00:40:40.500 --> 00:40:42.720 align:middle line:84% They use ordinary differential equations 00:40:42.720 --> 00:40:46.607 align:middle line:84% and they will do forward modeling-- forward simulations. 00:40:46.607 --> 00:40:48.690 align:middle line:84% But if you think about this whole robustness issue 00:40:48.690 --> 00:40:52.380 align:middle line:84% or network fitting, parameter fitting, well, it's 00:40:52.380 --> 00:40:53.940 align:middle line:90% going to be a problem. 00:40:53.940 --> 00:40:57.222 align:middle line:84% These are the time series data predicted by the model. 00:40:57.222 --> 00:40:59.430 align:middle line:84% And you can actually do measurements on these things, 00:40:59.430 --> 00:40:59.940 align:middle line:90% right? 00:40:59.940 --> 00:41:03.030 align:middle line:84% So you can start to see how well the simulations fit 00:41:03.030 --> 00:41:04.720 align:middle line:90% the experimental data. 00:41:04.720 --> 00:41:08.550 align:middle line:84% Now if you think about it, you have very many parameters here. 00:41:08.550 --> 00:41:10.890 align:middle line:84% You have thousands of parameters here. 00:41:10.890 --> 00:41:14.880 align:middle line:84% And they can change in all sorts of combinations. 00:41:14.880 --> 00:41:18.750 align:middle line:84% So when you try to fit your model to the experimental data, 00:41:18.750 --> 00:41:21.550 align:middle line:84% if you loosen here, you have to tighten there. 00:41:21.550 --> 00:41:24.892 align:middle line:84% So this is a very, very complicated-- 00:41:24.892 --> 00:41:26.850 align:middle line:84% computer scientists will understand right away. 00:41:26.850 --> 00:41:29.472 align:middle line:84% This is an intractable condition. 00:41:29.472 --> 00:41:30.930 align:middle line:84% Well, actually, if you think about, 00:41:30.930 --> 00:41:33.060 align:middle line:84% these parameters have only two values that it's going to be, 00:41:33.060 --> 00:41:34.170 align:middle line:90% and you have n parameters. 00:41:34.170 --> 00:41:35.880 align:middle line:84% Then what you have to go through-- 00:41:35.880 --> 00:41:38.640 align:middle line:90% exhaustive [INAUDIBLE]. 00:41:38.640 --> 00:41:40.230 align:middle line:90% So it's not doable. 00:41:40.230 --> 00:41:43.920 align:middle line:84% So they use also the trace genetic algorithms or whatnot. 00:41:43.920 --> 00:41:46.950 align:middle line:84% But nobody has any idea right now about the landscape 00:41:46.950 --> 00:41:49.170 align:middle line:90% of this parameter space. 00:41:49.170 --> 00:41:53.490 align:middle line:84% Is this like-- really have certain optimal states? 00:41:53.490 --> 00:41:56.370 align:middle line:84% Or is it kind of local optima and we're just playing around 00:41:56.370 --> 00:41:57.120 align:middle line:90% with these things? 00:41:57.120 --> 00:41:58.578 align:middle line:84% And you'd like to know these things 00:41:58.578 --> 00:42:01.330 align:middle line:90% before you trust your model. 00:42:01.330 --> 00:42:03.990 align:middle line:84% So you can play and you can do relatively exhaustive searches 00:42:03.990 --> 00:42:05.230 align:middle line:90% for small networks. 00:42:05.230 --> 00:42:09.180 align:middle line:84% But then you start to enter the realm of realistic size 00:42:09.180 --> 00:42:10.710 align:middle line:90% networks of 1,000 parameters. 00:42:10.710 --> 00:42:12.700 align:middle line:90% You have no idea what you are-- 00:42:12.700 --> 00:42:13.750 align:middle line:90% where you are. 00:42:13.750 --> 00:42:16.830 align:middle line:84% So that's what answers the whole issue of robustness. 00:42:16.830 --> 00:42:19.380 align:middle line:84% If we understand what robustness means in this network, 00:42:19.380 --> 00:42:21.660 align:middle line:84% maybe you do not really need to worry 00:42:21.660 --> 00:42:24.210 align:middle line:84% about the actual parameter values that much. 00:42:24.210 --> 00:42:26.970 align:middle line:84% Because within certain ranges, the system 00:42:26.970 --> 00:42:28.500 align:middle line:90% has to behave the very same way. 00:42:28.500 --> 00:42:33.560 align:middle line:90% 00:42:33.560 --> 00:42:35.725 align:middle line:90% Yeah, that we all know. 00:42:35.725 --> 00:42:38.010 align:middle line:84% And there is some evidence for this. 00:42:38.010 --> 00:42:41.390 align:middle line:84% So there was a very nice paper from Gary O'Dell when 00:42:41.390 --> 00:42:44.390 align:middle line:84% they took all the knowledge we have 00:42:44.390 --> 00:42:48.608 align:middle line:84% about the [INAUDIBLE] priority network of [INAUDIBLE].. 00:42:48.608 --> 00:42:50.150 align:middle line:84% So this is all the knowledge we know. 00:42:50.150 --> 00:42:52.670 align:middle line:84% These are all the proteins that play a role in it and all 00:42:52.670 --> 00:42:56.500 align:middle line:90% the regulatory interactions. 00:42:56.500 --> 00:42:58.990 align:middle line:84% And they describe these things and they put all this 00:42:58.990 --> 00:43:01.220 align:middle line:84% into MATLAB and they start to play with this. 00:43:01.220 --> 00:43:04.190 align:middle line:84% And what they found-- they found two striking things. 00:43:04.190 --> 00:43:06.940 align:middle line:84% One is that if you have this network, 00:43:06.940 --> 00:43:09.968 align:middle line:84% parameter values essentially do not matter. 00:43:09.968 --> 00:43:12.010 align:middle line:84% In the range of two to three orders of magnitude, 00:43:12.010 --> 00:43:14.110 align:middle line:84% you could change any parameter setting. 00:43:14.110 --> 00:43:16.840 align:middle line:90% You still have the same outcome. 00:43:16.840 --> 00:43:20.140 align:middle line:84% Also, this network behaved in a robust way 00:43:20.140 --> 00:43:24.490 align:middle line:84% only if you introduced two regulatory interactions-- 00:43:24.490 --> 00:43:25.900 align:middle line:90% these two guys. 00:43:25.900 --> 00:43:28.450 align:middle line:84% And when they looked harder, they actually 00:43:28.450 --> 00:43:31.360 align:middle line:84% found evidence that those are there. 00:43:31.360 --> 00:43:36.070 align:middle line:84% So the important message of this work 00:43:36.070 --> 00:43:39.670 align:middle line:84% was that maybe it's mainly the topology that's 00:43:39.670 --> 00:43:42.370 align:middle line:84% going to determine how your network behaves. 00:43:42.370 --> 00:43:43.780 align:middle line:84% People tried-- like Andrew Murray 00:43:43.780 --> 00:43:48.443 align:middle line:84% here at Harvard-- tried to de-convolute the network. 00:43:48.443 --> 00:43:50.860 align:middle line:84% One of the questions is that, if you have a network that's 00:43:50.860 --> 00:43:54.040 align:middle line:84% robust, is this working the very same way as we 00:43:54.040 --> 00:43:55.730 align:middle line:90% understand in control theory? 00:43:55.730 --> 00:43:58.030 align:middle line:84% So you have models, redundancy, feedback loops. 00:43:58.030 --> 00:44:02.560 align:middle line:84% Or there is some of an other interesting topology at play 00:44:02.560 --> 00:44:07.390 align:middle line:84% that evolution developed or invented. 00:44:07.390 --> 00:44:11.950 align:middle line:84% So they try to do this, but they still-- the sense of this study 00:44:11.950 --> 00:44:14.950 align:middle line:84% was that the parameter values can change a lot 00:44:14.950 --> 00:44:17.540 align:middle line:84% and you still have the very same outcome. 00:44:17.540 --> 00:44:19.738 align:middle line:84% So if this is true for many biological networks, 00:44:19.738 --> 00:44:21.280 align:middle line:84% large networks-- and this is actually 00:44:21.280 --> 00:44:23.115 align:middle line:84% the way it's always designed, although it 00:44:23.115 --> 00:44:24.490 align:middle line:84% will take a lot of work and a lot 00:44:24.490 --> 00:44:26.110 align:middle line:84% of independent confirmation, then 00:44:26.110 --> 00:44:28.990 align:middle line:84% you do not really need to worry that much about 00:44:28.990 --> 00:44:31.072 align:middle line:84% individual parameter fitting once you understand 00:44:31.072 --> 00:44:32.530 align:middle line:84% the typology and you have some sort 00:44:32.530 --> 00:44:36.860 align:middle line:84% of approximate field about the individual parameters. 00:44:36.860 --> 00:44:38.270 align:middle line:90% So that's good news. 00:44:38.270 --> 00:44:40.300 align:middle line:84% And also that they could make predictive-- 00:44:40.300 --> 00:44:42.470 align:middle line:90% they could make predictions. 00:44:42.470 --> 00:44:46.810 align:middle line:84% So robustness one of the key properties of a link organism. 00:44:46.810 --> 00:44:51.550 align:middle line:84% And-- oh, this is the kind of semi-official definition. 00:44:51.550 --> 00:44:54.130 align:middle line:90% And OK, so how does this work? 00:44:54.130 --> 00:44:59.890 align:middle line:84% As we understand, robustness is derived in control theory. 00:44:59.890 --> 00:45:01.960 align:middle line:90% Jumbo jet, three things-- 00:45:01.960 --> 00:45:04.063 align:middle line:84% we have feedback loops, we have redundancy, 00:45:04.063 --> 00:45:05.230 align:middle line:90% and you have modular design. 00:45:05.230 --> 00:45:07.875 align:middle line:90% 00:45:07.875 --> 00:45:10.000 align:middle line:84% If you apply these things in the right combination, 00:45:10.000 --> 00:45:13.150 align:middle line:84% the right way, involving lots and lots of complicated math, 00:45:13.150 --> 00:45:15.700 align:middle line:84% then you have probably a very robust system. 00:45:15.700 --> 00:45:19.180 align:middle line:84% We do not know whether we have some other design 00:45:19.180 --> 00:45:22.330 align:middle line:90% at work in biology. 00:45:22.330 --> 00:45:23.950 align:middle line:90% In principle, that's possible. 00:45:23.950 --> 00:45:26.800 align:middle line:84% There are lots of issues raised like-- 00:45:26.800 --> 00:45:30.820 align:middle line:84% it was raised that if you have a parallel distribution 00:45:30.820 --> 00:45:33.700 align:middle line:84% of regulatory interactions, if you think about this graph 00:45:33.700 --> 00:45:34.782 align:middle line:90% description-- 00:45:34.782 --> 00:45:35.740 align:middle line:90% like small neural nets. 00:45:35.740 --> 00:45:37.180 align:middle line:84% You might have heard a lot about these things. 00:45:37.180 --> 00:45:39.430 align:middle line:84% Now, these things do not translate in any way 00:45:39.430 --> 00:45:40.570 align:middle line:90% into dynamic pressure. 00:45:40.570 --> 00:45:43.210 align:middle line:84% So nobody knows whether this makes any sense or not. 00:45:43.210 --> 00:45:46.090 align:middle line:84% But I'm just throwing something at you that, well maybe, yes. 00:45:46.090 --> 00:45:49.300 align:middle line:84% Maybe if you have some of parallel distribution 00:45:49.300 --> 00:45:54.250 align:middle line:84% or regulatory interactions, then maybe that's 00:45:54.250 --> 00:45:56.290 align:middle line:84% going to be the key-- or some sort of a design. 00:45:56.290 --> 00:45:59.950 align:middle line:84% Nobody really understands why the World Wide Web works. 00:45:59.950 --> 00:46:01.090 align:middle line:90% It's not supposed to work. 00:46:01.090 --> 00:46:04.360 align:middle line:84% It works but nobody understands why. 00:46:04.360 --> 00:46:06.820 align:middle line:84% There are some ideas, but nobody can describe really 00:46:06.820 --> 00:46:07.690 align:middle line:90% the dynamics. 00:46:07.690 --> 00:46:11.530 align:middle line:84% And it's not supposed to work the way it is described. 00:46:11.530 --> 00:46:16.270 align:middle line:84% Whether that could be de-convoluted to describe 00:46:16.270 --> 00:46:19.330 align:middle line:84% the work of World Wide Web, why it works-- 00:46:19.330 --> 00:46:21.670 align:middle line:84% into these categories, into these terms, 00:46:21.670 --> 00:46:26.290 align:middle line:84% or there is some of a self-organizing evolutionary 00:46:26.290 --> 00:46:28.192 align:middle line:84% principle at work there, nobody knows. 00:46:28.192 --> 00:46:30.400 align:middle line:84% Lots of people like [INAUDIBLE] and some other people 00:46:30.400 --> 00:46:31.840 align:middle line:84% who are betting on these think that there 00:46:31.840 --> 00:46:34.450 align:middle line:84% is some kind of self-organizing evolution principle at work 00:46:34.450 --> 00:46:35.800 align:middle line:90% here. 00:46:35.800 --> 00:46:38.770 align:middle line:84% Nobody understands in real terms, what that means. 00:46:38.770 --> 00:46:41.050 align:middle line:84% But as you understand, there is kind 00:46:41.050 --> 00:46:43.180 align:middle line:84% of a philosophical or like a religious issue 00:46:43.180 --> 00:46:44.950 align:middle line:84% here, that there is control theory 00:46:44.950 --> 00:46:47.080 align:middle line:84% as an engineer puts together a robust system, 00:46:47.080 --> 00:46:48.940 align:middle line:90% and there is evolution. 00:46:48.940 --> 00:46:51.730 align:middle line:84% Is the two things essentially the same or there 00:46:51.730 --> 00:46:53.720 align:middle line:90% is difference? 00:46:53.720 --> 00:46:56.280 align:middle line:90% So how can you use robustness? 00:46:56.280 --> 00:46:58.630 align:middle line:84% You can exploit this for biology. 00:46:58.630 --> 00:47:01.570 align:middle line:84% There very nice people from [INAUDIBLE] who did-- 00:47:01.570 --> 00:47:04.090 align:middle line:84% actually they took Drosophila again. 00:47:04.090 --> 00:47:05.860 align:middle line:90% It's kind of interesting. 00:47:05.860 --> 00:47:08.470 align:middle line:84% Bone morphogenic proteins that will tell your cell which 00:47:08.470 --> 00:47:12.940 align:middle line:84% way to develop or not to develop are distributed spatially 00:47:12.940 --> 00:47:14.563 align:middle line:90% in a well-determined way. 00:47:14.563 --> 00:47:15.980 align:middle line:84% So for example, this is a protein. 00:47:15.980 --> 00:47:17.680 align:middle line:84% This is the whole Drosophila egg. 00:47:17.680 --> 00:47:22.125 align:middle line:84% Some proteins are expressed only in the midline. 00:47:22.125 --> 00:47:23.500 align:middle line:84% So these proteins tell you what's 00:47:23.500 --> 00:47:28.450 align:middle line:84% the top, what's front, back, ventral, dorsal, midline, 00:47:28.450 --> 00:47:31.102 align:middle line:90% lateral, in an animal. 00:47:31.102 --> 00:47:33.310 align:middle line:84% So if you think about it-- if you look at this thing, 00:47:33.310 --> 00:47:34.720 align:middle line:90% this is odd. 00:47:34.720 --> 00:47:36.640 align:middle line:84% You'd expect things spreading out. 00:47:36.640 --> 00:47:38.710 align:middle line:90% So how can this thing happen? 00:47:38.710 --> 00:47:41.140 align:middle line:84% This is like a real distribution. 00:47:41.140 --> 00:47:44.622 align:middle line:84% This is the central lateral distribution. 00:47:44.622 --> 00:47:45.830 align:middle line:90% So how can this thing happen? 00:47:45.830 --> 00:47:47.890 align:middle line:84% So these are involving very, very few proteins-- 00:47:47.890 --> 00:47:49.030 align:middle line:90% three or four-- 00:47:49.030 --> 00:47:52.990 align:middle line:84% and they describe this in kind of a one dimensional way. 00:47:52.990 --> 00:47:55.710 align:middle line:84% So they kind of unfold in the animal. 00:47:55.710 --> 00:48:01.920 align:middle line:84% So this is the distribution as a central and lateral. 00:48:01.920 --> 00:48:02.730 align:middle line:90% And what they did-- 00:48:02.730 --> 00:48:06.150 align:middle line:84% well, this is what we know about this system. 00:48:06.150 --> 00:48:08.220 align:middle line:90% We do not know the parameters-- 00:48:08.220 --> 00:48:09.360 align:middle line:90% roughly no. 00:48:09.360 --> 00:48:12.990 align:middle line:84% But let's generate a very large number of random networks. 00:48:12.990 --> 00:48:14.910 align:middle line:84% They generate 60,000 random networks. 00:48:14.910 --> 00:48:16.910 align:middle line:84% You can do it, they're relatively small network. 00:48:16.910 --> 00:48:18.540 align:middle line:84% A couple of differential equations-- 00:48:18.540 --> 00:48:20.370 align:middle line:84% partial differentiable equations. 00:48:20.370 --> 00:48:23.130 align:middle line:84% And then ask a question-- what are the common characteristics 00:48:23.130 --> 00:48:27.810 align:middle line:84% of all the robust networks relative to those 00:48:27.810 --> 00:48:28.950 align:middle line:90% that are not robust? 00:48:28.950 --> 00:48:33.700 align:middle line:84% And they came up with a handful of observations. 00:48:33.700 --> 00:48:36.240 align:middle line:84% [INAUDIBLE],, which is a protein, does not diffuse. 00:48:36.240 --> 00:48:39.060 align:middle line:84% [INAUDIBLE],, which is another protein, is not chopped. 00:48:39.060 --> 00:48:41.010 align:middle line:90% This thing has to break down. 00:48:41.010 --> 00:48:43.860 align:middle line:84% And only the observed distribution is robust. 00:48:43.860 --> 00:48:47.130 align:middle line:84% There was no distribution that was like flat and robust. 00:48:47.130 --> 00:48:49.590 align:middle line:84% And we are talking about a very large number-- 00:48:49.590 --> 00:48:52.410 align:middle line:84% tens of thousands of randomly generated networks that 00:48:52.410 --> 00:48:54.450 align:middle line:90% retained the overall wiring-- 00:48:54.450 --> 00:48:58.200 align:middle line:84% the topology-- but the parameter values were randomly changed. 00:48:58.200 --> 00:49:02.970 align:middle line:84% So they came up with these two hypotheses. 00:49:02.970 --> 00:49:05.130 align:middle line:84% And it showed that it's only the-- 00:49:05.130 --> 00:49:07.650 align:middle line:84% robust networks must have these features. 00:49:07.650 --> 00:49:10.650 align:middle line:84% And actually they could experimentally 00:49:10.650 --> 00:49:12.790 align:middle line:84% verify this, which is very, very nice. 00:49:12.790 --> 00:49:14.850 align:middle line:84% So the idea again is that you can play around 00:49:14.850 --> 00:49:17.040 align:middle line:84% with these networks, ask what's the common feature 00:49:17.040 --> 00:49:19.470 align:middle line:84% of robust networks, and that will help you-- 00:49:19.470 --> 00:49:22.390 align:middle line:84% that will teach you something about the actual biology 00:49:22.390 --> 00:49:22.890 align:middle line:90% present. 00:49:22.890 --> 00:49:25.470 align:middle line:90% 00:49:25.470 --> 00:49:29.610 align:middle line:84% So this is kind of modeling based on all the differential 00:49:29.610 --> 00:49:31.710 align:middle line:90% equations real networks. 00:49:31.710 --> 00:49:33.930 align:middle line:84% Now you all know that until we understand-- 00:49:33.930 --> 00:49:37.580 align:middle line:84% you understand that until we really understand how 00:49:37.580 --> 00:49:39.390 align:middle line:84% the parameter fitting is going to work, 00:49:39.390 --> 00:49:43.920 align:middle line:84% and how the system works, we cannot really do very large 00:49:43.920 --> 00:49:44.970 align:middle line:90% scale modeling. 00:49:44.970 --> 00:49:48.150 align:middle line:84% But there were other people who took alternative approaches. 00:49:48.150 --> 00:49:50.970 align:middle line:84% And they said, well, maybe we can understand something 00:49:50.970 --> 00:49:53.070 align:middle line:84% about the overall topology of networks 00:49:53.070 --> 00:49:55.440 align:middle line:84% if we take very rough models-- like Boolean networks. 00:49:55.440 --> 00:49:59.290 align:middle line:84% Genes are on and off and they regulate each other. 00:49:59.290 --> 00:50:02.430 align:middle line:84% And [INAUDIBLE] and [INAUDIBLE] Glass, they started in the '60s 00:50:02.430 --> 00:50:03.837 align:middle line:84% and they played around with this. 00:50:03.837 --> 00:50:06.420 align:middle line:84% And they wrote interesting books and they became very popular. 00:50:06.420 --> 00:50:11.880 align:middle line:84% And they were not really part of real experimental science, 00:50:11.880 --> 00:50:14.250 align:middle line:84% but nevertheless they produced some very interesting 00:50:14.250 --> 00:50:17.260 align:middle line:84% suggestions that kind of keep coming back. 00:50:17.260 --> 00:50:21.480 align:middle line:84% So let me mention a couple of things here briefly. 00:50:21.480 --> 00:50:23.543 align:middle line:84% Because as I said, these things-- 00:50:23.543 --> 00:50:24.960 align:middle line:84% these Boolean networks-- never had 00:50:24.960 --> 00:50:27.455 align:middle line:84% anything to do with real description of biology. 00:50:27.455 --> 00:50:28.830 align:middle line:84% They're very, very far from that. 00:50:28.830 --> 00:50:32.250 align:middle line:90% And that was the '60s, '70s-- 00:50:32.250 --> 00:50:35.280 align:middle line:84% they made huge errors in their analysis. 00:50:35.280 --> 00:50:38.530 align:middle line:84% They very much undersampled the gene expression space 00:50:38.530 --> 00:50:39.820 align:middle line:90% and so forth. 00:50:39.820 --> 00:50:42.600 align:middle line:84% But they introduced a couple of concepts 00:50:42.600 --> 00:50:44.550 align:middle line:84% which were not exactly new, but they showed 00:50:44.550 --> 00:50:46.410 align:middle line:90% that it's an important thing. 00:50:46.410 --> 00:50:48.810 align:middle line:84% That if you have a Boolean network, 00:50:48.810 --> 00:50:51.430 align:middle line:84% things turn on or off each other, 00:50:51.430 --> 00:50:54.120 align:middle line:84% then it's a deterministic network. 00:50:54.120 --> 00:50:57.900 align:middle line:84% Then depending on what's the average number of input 00:50:57.900 --> 00:51:00.420 align:middle line:84% per gene or per parameter, the system 00:51:00.420 --> 00:51:03.210 align:middle line:84% can display chaotic or organized behavior. 00:51:03.210 --> 00:51:05.520 align:middle line:84% Chaotic behavior means that the system just 00:51:05.520 --> 00:51:09.000 align:middle line:84% keeps wandering about in the whole gene expression space. 00:51:09.000 --> 00:51:11.940 align:middle line:84% Organized behavior means that you have limit cycles 00:51:11.940 --> 00:51:15.000 align:middle line:84% or attractors-- limit cycles in continuous differential 00:51:15.000 --> 00:51:17.400 align:middle line:84% equations, and in Boolean networks, you have attractors. 00:51:17.400 --> 00:51:19.260 align:middle line:84% Which means that it doesn't really matter, 00:51:19.260 --> 00:51:21.480 align:middle line:84% what's the initial parameter setting, in terms 00:51:21.480 --> 00:51:24.390 align:middle line:90% of individual genes on or off. 00:51:24.390 --> 00:51:27.630 align:middle line:84% The system tends to fall towards attractors. 00:51:27.630 --> 00:51:29.250 align:middle line:90% This is not new. 00:51:29.250 --> 00:51:31.590 align:middle line:84% Because people knew that this thing exists. 00:51:31.590 --> 00:51:34.770 align:middle line:84% But the concept that these attractors 00:51:34.770 --> 00:51:39.180 align:middle line:84% can behave or can display very robust properties 00:51:39.180 --> 00:51:41.040 align:middle line:90% is still very important. 00:51:41.040 --> 00:51:42.580 align:middle line:90% Now is this important or not? 00:51:42.580 --> 00:51:44.430 align:middle line:84% And then at that time, [INAUDIBLE]---- 00:51:44.430 --> 00:51:45.970 align:middle line:90% he's a very interesting guy. 00:51:45.970 --> 00:51:47.580 align:middle line:84% I mean just-- you know, he didn't do any experiments. 00:51:47.580 --> 00:51:49.620 align:middle line:84% In the '70s he came up with a set of papers. 00:51:49.620 --> 00:51:55.950 align:middle line:84% And it's just intellectually very entertaining, what he did. 00:51:55.950 --> 00:52:03.588 align:middle line:84% He actually suggested that differentiation states 00:52:03.588 --> 00:52:05.130 align:middle line:84% are essentially different attractors. 00:52:05.130 --> 00:52:08.080 align:middle line:84% And during malignant transformation, 00:52:08.080 --> 00:52:11.310 align:middle line:84% development of cancer cells go from a normal attractor 00:52:11.310 --> 00:52:14.130 align:middle line:90% towards a malignant-- 00:52:14.130 --> 00:52:14.940 align:middle line:90% a new attractor. 00:52:14.940 --> 00:52:16.878 align:middle line:84% And then they will stay there whatever you do. 00:52:16.878 --> 00:52:18.420 align:middle line:84% Or you have to change a lot of things 00:52:18.420 --> 00:52:20.350 align:middle line:90% to drive it back to normal. 00:52:20.350 --> 00:52:25.125 align:middle line:84% Now there might be some evidence this is actually true. 00:52:25.125 --> 00:52:27.000 align:middle line:84% You heard about gene expression measurements. 00:52:27.000 --> 00:52:32.040 align:middle line:84% And what [INAUDIBLE] group did, they took cancer-- 00:52:32.040 --> 00:52:36.120 align:middle line:84% breast cancer samples from patients before treatment, 00:52:36.120 --> 00:52:40.860 align:middle line:84% after treatment, and from the metastasis of the same tumor 00:52:40.860 --> 00:52:42.013 align:middle line:90% in the same patient. 00:52:42.013 --> 00:52:43.680 align:middle line:84% And they measured gene expression pairs. 00:52:43.680 --> 00:52:45.758 align:middle line:84% And then-- well if you look at cancer-- 00:52:45.758 --> 00:52:48.300 align:middle line:84% you might have heard about it or you might be aware of this-- 00:52:48.300 --> 00:52:53.565 align:middle line:84% [INAUDIBLE] completely messed up genome-- 00:52:53.565 --> 00:52:56.260 align:middle line:90% 00:52:56.260 --> 00:53:00.748 align:middle line:84% chromosomes recombine in a completely random fashion. 00:53:00.748 --> 00:53:03.040 align:middle line:84% And you would expect that well, yeah, the same tumor is 00:53:03.040 --> 00:53:05.020 align:middle line:90% going to keep changing a lot. 00:53:05.020 --> 00:53:08.300 align:middle line:84% Well based on this, it seems that this is not the case. 00:53:08.300 --> 00:53:11.680 align:middle line:84% A tumor in a given patient is always 00:53:11.680 --> 00:53:14.980 align:middle line:84% going to be very, very similar to the tumor 00:53:14.980 --> 00:53:17.020 align:middle line:84% in the same patient given a different time point 00:53:17.020 --> 00:53:20.832 align:middle line:84% or during metastasis whatever you do. 00:53:20.832 --> 00:53:21.790 align:middle line:90% It's kind of the same-- 00:53:21.790 --> 00:53:22.990 align:middle line:84% AUDIENCE: In terms of gene expression. 00:53:22.990 --> 00:53:23.770 align:middle line:84% ZOLTAN SZALLASI: In terms of gene expression, 00:53:23.770 --> 00:53:25.330 align:middle line:84% I'm sorry-- gene expression pairs. 00:53:25.330 --> 00:53:27.800 align:middle line:84% And they are going to be very different from somebody else. 00:53:27.800 --> 00:53:30.490 align:middle line:84% So it seems that once a tumor develops, 00:53:30.490 --> 00:53:33.318 align:middle line:84% it finds a new stable state, it's going to stay there. 00:53:33.318 --> 00:53:34.610 align:middle line:90% It's not going to wander about. 00:53:34.610 --> 00:53:36.670 align:middle line:84% And now there is lots of evidence-- well, 00:53:36.670 --> 00:53:38.587 align:middle line:84% different labs work with different cell lines. 00:53:38.587 --> 00:53:40.628 align:middle line:84% And certain cell lines that have been established 00:53:40.628 --> 00:53:42.460 align:middle line:84% for a long time, and unless you introduce 00:53:42.460 --> 00:53:44.382 align:middle line:84% very huge genetic changes, they are 00:53:44.382 --> 00:53:45.590 align:middle line:90% pretty similar to each other. 00:53:45.590 --> 00:53:46.965 align:middle line:84% They are always much more similar 00:53:46.965 --> 00:53:49.047 align:middle line:84% to each other than, say, anything else. 00:53:49.047 --> 00:53:50.630 align:middle line:84% So it seems that the genetic network-- 00:53:50.630 --> 00:53:52.060 align:middle line:90% and this kind of makes sense-- 00:53:52.060 --> 00:53:55.810 align:middle line:90% finds a new robust stable state. 00:53:55.810 --> 00:53:58.720 align:middle line:84% And then it tends to stay there because if you 00:53:58.720 --> 00:54:00.580 align:middle line:84% do those individual small perturbations, 00:54:00.580 --> 00:54:02.028 align:middle line:90% the system can't fix itself. 00:54:02.028 --> 00:54:04.570 align:middle line:84% So this pretty much is pretty similar to what they suggested. 00:54:04.570 --> 00:54:06.280 align:middle line:90% You have strong attractors. 00:54:06.280 --> 00:54:07.450 align:middle line:90% Now, is this is useful? 00:54:07.450 --> 00:54:09.190 align:middle line:84% Not this representation-- we do not know. 00:54:09.190 --> 00:54:11.572 align:middle line:84% But you can ask questions like, how many perturbations 00:54:11.572 --> 00:54:13.780 align:middle line:84% do you need to introduce to drive the system from one 00:54:13.780 --> 00:54:14.922 align:middle line:90% attractor to another one? 00:54:14.922 --> 00:54:16.630 align:middle line:84% And if it shows that for a large network, 00:54:16.630 --> 00:54:19.480 align:middle line:84% you cannot do it with less than 100 perturbations, 00:54:19.480 --> 00:54:22.810 align:middle line:84% well that would suggest a very different therapeutic strategy 00:54:22.810 --> 00:54:25.300 align:middle line:90% than if you can do it with one. 00:54:25.300 --> 00:54:27.250 align:middle line:84% Or which things you need to change-- 00:54:27.250 --> 00:54:30.160 align:middle line:84% things that are highly connected or intermediate connected? 00:54:30.160 --> 00:54:33.070 align:middle line:84% So you can actually-- these representations might teach you 00:54:33.070 --> 00:54:35.200 align:middle line:84% something about the robustness of the network. 00:54:35.200 --> 00:54:38.320 align:middle line:84% That's why I wanted to mention it. 00:54:38.320 --> 00:54:43.210 align:middle line:84% Also I want to warn you about a couple things 00:54:43.210 --> 00:54:45.940 align:middle line:84% like, when we talk about modeling, 00:54:45.940 --> 00:54:48.190 align:middle line:84% we always talk about a single cell. 00:54:48.190 --> 00:54:50.380 align:middle line:84% What's happening in that cell and which way it's 00:54:50.380 --> 00:54:51.550 align:middle line:90% going to develop? 00:54:51.550 --> 00:54:56.440 align:middle line:84% And we shouldn't forget that whatever measurements you do, 00:54:56.440 --> 00:54:58.420 align:middle line:84% you always work with population average data. 00:54:58.420 --> 00:55:00.800 align:middle line:84% This kind of leads back to this stochasticity issue. 00:55:00.800 --> 00:55:03.175 align:middle line:84% And let me just show you this very nice little experiment 00:55:03.175 --> 00:55:07.450 align:middle line:84% by James Ferrell, when they measured 00:55:07.450 --> 00:55:11.830 align:middle line:84% the Jun Kinase activity of individual xenopus oocyte. 00:55:11.830 --> 00:55:15.160 align:middle line:84% And actually he had to do this whole experiment himself, 00:55:15.160 --> 00:55:17.457 align:middle line:84% although he's like a well-established full 00:55:17.457 --> 00:55:19.540 align:middle line:84% professor, because there was no student in his lab 00:55:19.540 --> 00:55:21.980 align:middle line:84% who's willing to do the following experiment. 00:55:21.980 --> 00:55:25.420 align:middle line:84% So what they did is that they measured Jun Kinase activity 00:55:25.420 --> 00:55:26.800 align:middle line:90% on the progesterone treatment. 00:55:26.800 --> 00:55:29.590 align:middle line:84% This is the dose response curve in a population 00:55:29.590 --> 00:55:31.630 align:middle line:90% of xenopus oocytes. 00:55:31.630 --> 00:55:33.610 align:middle line:90% And what happened is that-- 00:55:33.610 --> 00:55:35.410 align:middle line:84% it shows this very nice dose response 00:55:35.410 --> 00:55:37.780 align:middle line:90% curve and the coefficient 1. 00:55:37.780 --> 00:55:40.390 align:middle line:84% Now the problem was that this was very far from what 00:55:40.390 --> 00:55:43.730 align:middle line:90% he thought was the case. 00:55:43.730 --> 00:55:46.960 align:middle line:84% And so he asked some of the students 00:55:46.960 --> 00:55:48.430 align:middle line:84% to actually measure the Jun Kinase 00:55:48.430 --> 00:55:50.235 align:middle line:90% activity on individual oocytes. 00:55:50.235 --> 00:55:51.110 align:middle line:90% That's a lot of work. 00:55:51.110 --> 00:55:52.840 align:middle line:84% Instead of running a single [INAUDIBLE],, 00:55:52.840 --> 00:55:54.760 align:middle line:84% you have to run 100 different [INAUDIBLE] 00:55:54.760 --> 00:55:56.560 align:middle line:90% for individual xenopus oocytes. 00:55:56.560 --> 00:56:00.250 align:middle line:84% So he started to take individual oocytes and experiment. 00:56:00.250 --> 00:56:03.173 align:middle line:84% And actually he had a very strong assumption 00:56:03.173 --> 00:56:04.090 align:middle line:90% what he was expecting. 00:56:04.090 --> 00:56:06.100 align:middle line:84% That's why he invested all the time. 00:56:06.100 --> 00:56:09.670 align:middle line:84% But what he found is that for an individual oocyte, 00:56:09.670 --> 00:56:13.420 align:middle line:84% they are either always inactivated or activated, 00:56:13.420 --> 00:56:16.120 align:middle line:84% with a very sharp dose response curve-- with a [INAUDIBLE] 00:56:16.120 --> 00:56:18.820 align:middle line:90% coefficient of close to 100. 00:56:18.820 --> 00:56:21.860 align:middle line:84% He had a strong positive feedback loop there. 00:56:21.860 --> 00:56:24.497 align:middle line:84% So the point is that what you see here is like an average. 00:56:24.497 --> 00:56:26.830 align:middle line:84% Because each oocyte is activated at a slightly different 00:56:26.830 --> 00:56:28.010 align:middle line:90% concentration. 00:56:28.010 --> 00:56:30.880 align:middle line:84% But what you have in a single system 00:56:30.880 --> 00:56:33.440 align:middle line:84% is a strong positive feedback loop. 00:56:33.440 --> 00:56:35.400 align:middle line:84% So when you do and describe a system and when 00:56:35.400 --> 00:56:37.150 align:middle line:84% you try to do reverse engineering, 00:56:37.150 --> 00:56:38.955 align:middle line:84% if you go with population average data, 00:56:38.955 --> 00:56:40.330 align:middle line:84% your reverse engineering is going 00:56:40.330 --> 00:56:42.080 align:middle line:84% to give you a [INAUDIBLE] coefficient of 1 00:56:42.080 --> 00:56:46.480 align:middle line:84% and some sort of a description that this is what reality is. 00:56:46.480 --> 00:56:48.490 align:middle line:84% And this is what you should do for your modeling 00:56:48.490 --> 00:56:50.380 align:middle line:84% if you want to describe the system correctly. 00:56:50.380 --> 00:56:51.280 align:middle line:90% Right? 00:56:51.280 --> 00:56:54.640 align:middle line:84% So this is kind of another warning how complicated biology 00:56:54.640 --> 00:56:57.460 align:middle line:84% is, that most measurements we are doing right now 00:56:57.460 --> 00:56:59.290 align:middle line:90% is grinding down a lot of cells. 00:56:59.290 --> 00:57:03.250 align:middle line:84% And you're doing some sort of a population average measurement. 00:57:03.250 --> 00:57:04.510 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:57:04.510 --> 00:57:06.031 align:middle line:90% ZOLTAN SZALLASI: Yep. 00:57:06.031 --> 00:57:09.997 align:middle line:84% AUDIENCE: Could it be hypothesized that [INAUDIBLE]?? 00:57:09.997 --> 00:57:11.080 align:middle line:90% ZOLTAN SZALLASI: Oh, yeah. 00:57:11.080 --> 00:57:13.538 align:middle line:84% Actually he de-convoluted and then published another paper. 00:57:13.538 --> 00:57:16.210 align:middle line:84% He showed exactly why you have that. 00:57:16.210 --> 00:57:20.710 align:middle line:84% And there is a feedback parameter that you can tweak 00:57:20.710 --> 00:57:23.470 align:middle line:84% and you can actually show that you can completely 00:57:23.470 --> 00:57:24.790 align:middle line:90% change the hysteresis curve. 00:57:24.790 --> 00:57:29.328 align:middle line:84% AUDIENCE: Again, that's very similar to what [INAUDIBLE] 00:57:29.328 --> 00:57:30.120 align:middle line:90% stable [INAUDIBLE]. 00:57:30.120 --> 00:57:30.550 align:middle line:90% ZOLTAN SZALLASI: Yes. 00:57:30.550 --> 00:57:32.210 align:middle line:84% AUDIENCE: And like high induction [INAUDIBLE].. 00:57:32.210 --> 00:57:32.580 align:middle line:90% ZOLTAN SZALLASI: Absolutely. 00:57:32.580 --> 00:57:33.710 align:middle line:84% AUDIENCE: And low induction [INAUDIBLE].. 00:57:33.710 --> 00:57:34.877 align:middle line:90% ZOLTAN SZALLASI: Absolutely. 00:57:34.877 --> 00:57:37.820 align:middle line:84% The difference there is that this is a real network. 00:57:37.820 --> 00:57:40.010 align:middle line:84% And what these guys are doing, they put it in. 00:57:40.010 --> 00:57:42.642 align:middle line:84% Yeah, and that's what Jim Collins is doing at BU. 00:57:42.642 --> 00:57:44.645 align:middle line:84% These guys-- James Ferrell did it 00:57:44.645 --> 00:57:47.390 align:middle line:84% for an actually existing network and [INAUDIBLE] 00:57:47.390 --> 00:57:52.720 align:middle line:84% tweaking that individual feedback parameter. 00:57:52.720 --> 00:57:53.733 align:middle line:90% We have seen that. 00:57:53.733 --> 00:57:55.400 align:middle line:84% So that was kind of modularity, and this 00:57:55.400 --> 00:57:57.230 align:middle line:84% is the positive feedback loop here. 00:57:57.230 --> 00:57:59.600 align:middle line:90% It's kind of de-convoluted here. 00:57:59.600 --> 00:58:01.590 align:middle line:84% Is it important to understand or not? 00:58:01.590 --> 00:58:03.470 align:middle line:84% Well, this is something very simple, 00:58:03.470 --> 00:58:06.380 align:middle line:84% but this might be very useful for drug development. 00:58:06.380 --> 00:58:09.900 align:middle line:90% Because the concept is simple. 00:58:09.900 --> 00:58:13.010 align:middle line:84% You do not want to design a drug that's 00:58:13.010 --> 00:58:15.740 align:middle line:84% going to interfere with something inside a feedback 00:58:15.740 --> 00:58:17.417 align:middle line:90% loop. 00:58:17.417 --> 00:58:18.750 align:middle line:90% That's why it's a feedback loop. 00:58:18.750 --> 00:58:20.990 align:middle line:84% So if you keep an input there, that's 00:58:20.990 --> 00:58:23.430 align:middle line:90% kind of going down the drain. 00:58:23.430 --> 00:58:24.830 align:middle line:84% So if you understand this network 00:58:24.830 --> 00:58:27.718 align:middle line:84% and you understand you have feedback loops, you want-- 00:58:27.718 --> 00:58:29.510 align:middle line:84% that would give you very strong indications 00:58:29.510 --> 00:58:32.510 align:middle line:84% where to design drugs-- where to interfere with the system, 00:58:32.510 --> 00:58:35.000 align:middle line:90% because it makes sense. 00:58:35.000 --> 00:58:37.460 align:middle line:84% In the other case, when you are inside a feedback loop, 00:58:37.460 --> 00:58:41.270 align:middle line:84% it doesn't make any sense to put in too much effort. 00:58:41.270 --> 00:58:43.970 align:middle line:84% OK, the last thing I want to mention 00:58:43.970 --> 00:58:45.390 align:middle line:90% is constraint-based models. 00:58:45.390 --> 00:58:48.200 align:middle line:84% So these things are very, very, very complicated 00:58:48.200 --> 00:58:52.190 align:middle line:84% dynamic networks, and there are lots of problems involved. 00:58:52.190 --> 00:58:57.147 align:middle line:84% And is it a kind of a simple way or smart way 00:58:57.147 --> 00:58:59.480 align:middle line:84% to make a shortcut, and learning something about biology 00:58:59.480 --> 00:59:00.650 align:middle line:90% with much less effort? 00:59:00.650 --> 00:59:02.870 align:middle line:84% And those are the constraint-based models. 00:59:02.870 --> 00:59:04.640 align:middle line:84% This essentially-- there's nothing-- 00:59:04.640 --> 00:59:06.950 align:middle line:84% they are like [INAUDIBLE] analysis 00:59:06.950 --> 00:59:09.590 align:middle line:84% and these things have been around for ages 00:59:09.590 --> 00:59:11.090 align:middle line:90% for chemical engineers. 00:59:11.090 --> 00:59:13.010 align:middle line:84% And they are useful for certain things. 00:59:13.010 --> 00:59:17.283 align:middle line:84% They are essentially the beefed up version of Kirchhoff's Law, 00:59:17.283 --> 00:59:19.700 align:middle line:84% stating that if you have a metabolic network, for example, 00:59:19.700 --> 00:59:24.003 align:middle line:84% then whatever goes into a node, it has to come out. 00:59:24.003 --> 00:59:25.670 align:middle line:84% Otherwise the system is going to explode 00:59:25.670 --> 00:59:28.790 align:middle line:84% or that metabolite is going to disappear. 00:59:28.790 --> 00:59:30.320 align:middle line:90% These are not dynamic things. 00:59:30.320 --> 00:59:33.770 align:middle line:84% All you are describing is that at every single metabolite 00:59:33.770 --> 00:59:40.310 align:middle line:84% node, the net flux has to be zero. 00:59:40.310 --> 00:59:43.670 align:middle line:84% And then you have a set of linear equations. 00:59:43.670 --> 00:59:45.620 align:middle line:84% And of course, usually you have much more 00:59:45.620 --> 00:59:47.270 align:middle line:84% parameters than linear equations. 00:59:47.270 --> 00:59:50.750 align:middle line:84% At least this can start a question. 00:59:50.750 --> 00:59:53.162 align:middle line:84% And that's what [INAUDIBLE] group 00:59:53.162 --> 00:59:54.620 align:middle line:84% has been doing for a while and they 00:59:54.620 --> 00:59:56.760 align:middle line:90% did very interesting things. 00:59:56.760 --> 01:00:02.040 align:middle line:84% For example, they showed that if this is like the flux cone-- 01:00:02.040 --> 01:00:04.180 align:middle line:90% which is, I'm sure that-- 01:00:04.180 --> 01:00:07.220 align:middle line:84% those of you in chemical engineering are familiar with-- 01:00:07.220 --> 01:00:09.161 align:middle line:90% this gives you the-- 01:00:09.161 --> 01:00:13.100 align:middle line:84% underneath the flux cone, you have all the possible 01:00:13.100 --> 01:00:14.840 align:middle line:90% or allowed solutions. 01:00:14.840 --> 01:00:16.205 align:middle line:90% I mean you can-- 01:00:16.205 --> 01:00:19.220 align:middle line:84% since you have more parameters than equations, 01:00:19.220 --> 01:00:21.950 align:middle line:84% there are many different combinations of the parameters 01:00:21.950 --> 01:00:23.220 align:middle line:90% that give you the solution. 01:00:23.220 --> 01:00:28.400 align:middle line:84% So the animal can live always underneath this-- 01:00:28.400 --> 01:00:32.205 align:middle line:84% anywhere underneath this flux [INAUDIBLE].. 01:00:32.205 --> 01:00:33.830 align:middle line:84% The interesting thing they hypothesized 01:00:33.830 --> 01:00:38.480 align:middle line:90% is that evolution optimizes. 01:00:38.480 --> 01:00:42.110 align:middle line:84% In other words, the animal is going to live on the edge. 01:00:42.110 --> 01:00:48.230 align:middle line:84% Produce maximum energy with a certain amount of oxygen input. 01:00:48.230 --> 01:00:49.850 align:middle line:90% And actually they measured it. 01:00:49.850 --> 01:00:53.142 align:middle line:84% They took the metabolic network of E. coli-- 01:00:53.142 --> 01:00:54.350 align:middle line:90% that's fairly well described. 01:00:54.350 --> 01:00:57.425 align:middle line:84% You have essentially the entire E. coli network around 01:00:57.425 --> 01:00:59.550 align:middle line:84% and you can download it and you can play with this. 01:00:59.550 --> 01:01:01.370 align:middle line:90% And that's pretty well known. 01:01:01.370 --> 01:01:03.770 align:middle line:84% And that's simple, just linear equations. 01:01:03.770 --> 01:01:06.710 align:middle line:84% You solve it and then you have an idea 01:01:06.710 --> 01:01:09.302 align:middle line:84% what are the allowed parameter combinations. 01:01:09.302 --> 01:01:10.760 align:middle line:84% These are very simple combinations. 01:01:10.760 --> 01:01:12.800 align:middle line:84% So this is like just stoichiometric numbers, 01:01:12.800 --> 01:01:15.150 align:middle line:84% how many things go in and come out. 01:01:15.150 --> 01:01:19.370 align:middle line:84% And then you simply measure that if you measure the oxygen 01:01:19.370 --> 01:01:22.670 align:middle line:84% input or the energy source for growth, 01:01:22.670 --> 01:01:24.620 align:middle line:90% how fast is the mass produced? 01:01:24.620 --> 01:01:26.000 align:middle line:90% How fast will the animal grows? 01:01:26.000 --> 01:01:29.400 align:middle line:84% They show that it is always living on the edge-- 01:01:29.400 --> 01:01:30.900 align:middle line:90% so it's optimizing. 01:01:30.900 --> 01:01:33.170 align:middle line:84% And then you are switching from one energy source 01:01:33.170 --> 01:01:34.100 align:middle line:90% to another one. 01:01:34.100 --> 01:01:35.990 align:middle line:84% It's going to find the other edge. 01:01:35.990 --> 01:01:38.885 align:middle line:84% And if you do the experiment 10 different times, 01:01:38.885 --> 01:01:41.510 align:middle line:84% the animal is going to take very different routes to get there. 01:01:41.510 --> 01:01:47.780 align:middle line:84% If you keep taking time series points during the development, 01:01:47.780 --> 01:01:48.930 align:middle line:90% it's finding a new edge. 01:01:48.930 --> 01:01:52.130 align:middle line:84% But it's always arriving to the new edge 01:01:52.130 --> 01:01:57.500 align:middle line:84% and it's going to find that new stable optimal growth 01:01:57.500 --> 01:01:58.920 align:middle line:90% condition. 01:01:58.920 --> 01:02:01.610 align:middle line:84% So this is actually very interesting 01:02:01.610 --> 01:02:04.130 align:middle line:84% because this is an entirely non-dynamic description. 01:02:04.130 --> 01:02:05.870 align:middle line:84% It's a very simplified description. 01:02:05.870 --> 01:02:07.610 align:middle line:84% But it teaches you something very 01:02:07.610 --> 01:02:11.030 align:middle line:84% important about the biology of E. coli, 01:02:11.030 --> 01:02:12.420 align:middle line:90% and it can be very useful. 01:02:12.420 --> 01:02:15.800 align:middle line:84% For example, for chemical engineers 01:02:15.800 --> 01:02:18.690 align:middle line:84% who are working in fermentation, this is very important. 01:02:18.690 --> 01:02:20.750 align:middle line:84% You can really optimize the production 01:02:20.750 --> 01:02:22.340 align:middle line:84% of certain type of proteins if you 01:02:22.340 --> 01:02:26.240 align:middle line:84% understand the metabolic network underneath. 01:02:26.240 --> 01:02:28.160 align:middle line:84% So one more thing, it's just the end. 01:02:28.160 --> 01:02:29.958 align:middle line:90% Again, this is not entirely new. 01:02:29.958 --> 01:02:32.250 align:middle line:84% This is a very old paper-- it's not actually very old-- 01:02:32.250 --> 01:02:34.820 align:middle line:90% 1994. 01:02:34.820 --> 01:02:38.750 align:middle line:84% Then people describe like the coagulation pathway that 01:02:38.750 --> 01:02:40.080 align:middle line:90% has been known for a long time. 01:02:40.080 --> 01:02:41.830 align:middle line:84% And you can actually describe pretty well. 01:02:41.830 --> 01:02:44.630 align:middle line:84% So of course, modeling has been around in biology as well. 01:02:44.630 --> 01:02:47.270 align:middle line:84% It's just, we didn't have that much measurement. 01:02:47.270 --> 01:02:49.700 align:middle line:84% And there was certainly no public or private funding 01:02:49.700 --> 01:02:51.212 align:middle line:90% to do all the experiments. 01:02:51.212 --> 01:02:52.670 align:middle line:84% And so, this is what we are having. 01:02:52.670 --> 01:02:55.210 align:middle line:84% So traditionally, biology try to predict things 01:02:55.210 --> 01:02:56.210 align:middle line:90% and then do experiments. 01:02:56.210 --> 01:02:58.070 align:middle line:84% And people did everything in their head 01:02:58.070 --> 01:03:01.015 align:middle line:90% and based on their intuition. 01:03:01.015 --> 01:03:02.390 align:middle line:84% And what we are trying to do here 01:03:02.390 --> 01:03:04.430 align:middle line:84% is speed up or make it more efficient-- 01:03:04.430 --> 01:03:07.723 align:middle line:84% this process you have to experiment. 01:03:07.723 --> 01:03:09.890 align:middle line:84% You have the prediction, you do modeling in between. 01:03:09.890 --> 01:03:11.690 align:middle line:84% That's going-- in silico, you have 01:03:11.690 --> 01:03:13.100 align:middle line:90% lots of alternative solutions. 01:03:13.100 --> 01:03:16.160 align:middle line:84% You throw out all the obviously wrong solutions, 01:03:16.160 --> 01:03:19.130 align:middle line:84% and you just experiment to test the ones that 01:03:19.130 --> 01:03:20.870 align:middle line:90% are supported by your model. 01:03:20.870 --> 01:03:26.420 align:middle line:84% And that can speed up designing of new therapeutic approaches. 01:03:26.420 --> 01:03:28.620 align:middle line:84% For example, combinatorial treatment, 01:03:28.620 --> 01:03:30.750 align:middle line:84% which is very difficult if you think about it-- 01:03:30.750 --> 01:03:36.120 align:middle line:84% a combinatorial explosion of all potential treatments. 01:03:36.120 --> 01:03:39.305 align:middle line:90% Any questions? 01:03:39.305 --> 01:03:46.150 align:middle line:84% AUDIENCE: Do you see any similarities with-- 01:03:46.150 --> 01:03:48.020 align:middle line:84% I just like [INAUDIBLE] that came out 01:03:48.020 --> 01:03:49.520 align:middle line:84% of the paper [INAUDIBLE] in terms 01:03:49.520 --> 01:03:54.610 align:middle line:84% of the physics of kinetic gas theory 01:03:54.610 --> 01:03:56.534 align:middle line:84% and just the dynamic [INAUDIBLE].. 01:03:56.534 --> 01:03:59.080 align:middle line:90% 01:03:59.080 --> 01:04:03.280 align:middle line:84% Some of the models that people are building. 01:04:03.280 --> 01:04:06.868 align:middle line:90% Or is that [INAUDIBLE]? 01:04:06.868 --> 01:04:09.410 align:middle line:84% ZOLTAN SZALLASI: I haven't seen any experimental verification 01:04:09.410 --> 01:04:11.245 align:middle line:90% one way or another. 01:04:11.245 --> 01:04:12.620 align:middle line:84% Most of these approaches actually 01:04:12.620 --> 01:04:13.700 align:middle line:90% were done by physicists. 01:04:13.700 --> 01:04:17.660 align:middle line:84% Like people describe large Boolean networks 01:04:17.660 --> 01:04:19.280 align:middle line:90% based on stained glasses. 01:04:19.280 --> 01:04:21.250 align:middle line:84% And then they said, well, it's going 01:04:21.250 --> 01:04:22.505 align:middle line:90% to behave something like this. 01:04:22.505 --> 01:04:26.010 align:middle line:90% 01:04:26.010 --> 01:04:28.640 align:middle line:84% If you can formulate experimental testable 01:04:28.640 --> 01:04:32.000 align:middle line:84% hypothesis based on that, and then you do the experiments 01:04:32.000 --> 01:04:34.340 align:middle line:84% and you see this is the way it works or doesn't work, 01:04:34.340 --> 01:04:35.190 align:middle line:90% it's great. 01:04:35.190 --> 01:04:38.510 align:middle line:84% The problem is that there are some theoretical approaches 01:04:38.510 --> 01:04:41.750 align:middle line:84% that physicists work out that just cannot be translated 01:04:41.750 --> 01:04:43.850 align:middle line:84% into anything biologically testable. 01:04:43.850 --> 01:04:45.440 align:middle line:90% That's my only problem. 01:04:45.440 --> 01:04:47.040 align:middle line:90% And those are not that-- 01:04:47.040 --> 01:04:49.040 align:middle line:84% of course, those are not that obvious how to do. 01:04:49.040 --> 01:04:53.690 align:middle line:90% 01:04:53.690 --> 01:05:04.267 align:middle line:90% AUDIENCE: [INAUDIBLE] 01:05:04.267 --> 01:05:05.850 align:middle line:84% ZOLTAN SZALLASI: Possible to use what? 01:05:05.850 --> 01:05:07.150 align:middle line:84% AUDIENCE: Gene expression data [INAUDIBLE].. 01:05:07.150 --> 01:05:08.942 align:middle line:84% ZOLTAN SZALLASI: What gene expression data? 01:05:08.942 --> 01:05:10.590 align:middle line:90% AUDIENCE: [INAUDIBLE] 01:05:10.590 --> 01:05:12.923 align:middle line:84% ZOLTAN SZALLASI: Yeah, that's a good question obviously. 01:05:12.923 --> 01:05:16.590 align:middle line:90% And that's-- I mean if-- 01:05:16.590 --> 01:05:19.470 align:middle line:84% if you were here when I talked about microarray measurements 01:05:19.470 --> 01:05:24.270 align:middle line:84% or noise, those are pretty noisy measurements. 01:05:24.270 --> 01:05:26.927 align:middle line:90% So you can do it. 01:05:26.927 --> 01:05:29.010 align:middle line:84% If you want to do parameter fitting or microarray, 01:05:29.010 --> 01:05:30.450 align:middle line:90% I wouldn't do it. 01:05:30.450 --> 01:05:33.690 align:middle line:84% Microarray is compressing and if you 01:05:33.690 --> 01:05:36.330 align:middle line:84% can guess pretty well what's going up or down right now, 01:05:36.330 --> 01:05:39.360 align:middle line:84% that's already a nice achievement. 01:05:39.360 --> 01:05:41.220 align:middle line:90% One can play with this and-- 01:05:41.220 --> 01:05:44.970 align:middle line:84% well, microarray is improving and we are certainly 01:05:44.970 --> 01:05:45.780 align:middle line:90% taking our share. 01:05:45.780 --> 01:05:48.960 align:middle line:84% Showing, for example, that based on some chips, 01:05:48.960 --> 01:05:52.522 align:middle line:84% like half of the [INAUDIBLE] probes are wrong. 01:05:52.522 --> 01:05:54.480 align:middle line:84% They are just not what they are supposed to be. 01:05:54.480 --> 01:05:56.897 align:middle line:84% And if you throw those out, you have very good [INAUDIBLE] 01:05:56.897 --> 01:05:57.640 align:middle line:90% measurements. 01:05:57.640 --> 01:06:00.595 align:middle line:84% So one can improve this type of measurements. 01:06:00.595 --> 01:06:02.970 align:middle line:84% But right now if you just take any microarray measurement 01:06:02.970 --> 01:06:05.310 align:middle line:84% from literature or yourself, and you 01:06:05.310 --> 01:06:07.020 align:middle line:84% want to do a parameter fitting on that, 01:06:07.020 --> 01:06:09.360 align:middle line:84% that's essentially noise propagation. 01:06:09.360 --> 01:06:12.750 align:middle line:84% If your question is, if I do a modeling-- 01:06:12.750 --> 01:06:15.150 align:middle line:84% I have a fairly good description of the model-- 01:06:15.150 --> 01:06:16.860 align:middle line:90% very large network. 01:06:16.860 --> 01:06:19.170 align:middle line:84% And my microarray measurement can give you, 01:06:19.170 --> 01:06:22.470 align:middle line:84% with a 95% confidence, who goes up or down, 01:06:22.470 --> 01:06:25.110 align:middle line:84% and my prediction is just going to be from the model-- 01:06:25.110 --> 01:06:27.330 align:middle line:84% for many proteins, who goes up or down. 01:06:27.330 --> 01:06:30.270 align:middle line:84% And you match it in a probabilistic fashion 01:06:30.270 --> 01:06:31.740 align:middle line:84% to the outcome of the microarray. 01:06:31.740 --> 01:06:33.300 align:middle line:90% In that case, yes. 01:06:33.300 --> 01:06:36.757 align:middle line:84% And that is being done and that's a useful approach 01:06:36.757 --> 01:06:38.340 align:middle line:84% because you have a very large network. 01:06:38.340 --> 01:06:39.870 align:middle line:84% And you're going to say that I want 01:06:39.870 --> 01:06:42.420 align:middle line:84% to predict how most or many of the parameters 01:06:42.420 --> 01:06:43.930 align:middle line:90% are going to change. 01:06:43.930 --> 01:06:46.710 align:middle line:84% And if I can then tell you who goes up or down, even 01:06:46.710 --> 01:06:49.800 align:middle line:84% without giving you very good quality numbers, 01:06:49.800 --> 01:06:51.580 align:middle line:84% that's going to be something very useful. 01:06:51.580 --> 01:06:52.830 align:middle line:90% So that's something actually-- 01:06:52.830 --> 01:06:57.120 align:middle line:84% I must have had an extra slide somewhere, which is-- 01:06:57.120 --> 01:06:58.650 align:middle line:90% yeah, actually I do. 01:06:58.650 --> 01:06:59.810 align:middle line:90% That was it-- huh. 01:06:59.810 --> 01:07:00.690 align:middle line:90% That's it. 01:07:00.690 --> 01:07:03.130 align:middle line:84% Kind of predicted what I wanted to. 01:07:03.130 --> 01:07:05.220 align:middle line:84% So there is a low throughput data 01:07:05.220 --> 01:07:07.710 align:middle line:84% where you have very good quality models. 01:07:07.710 --> 01:07:10.050 align:middle line:84% And there are other ways of going that high throughput 01:07:10.050 --> 01:07:12.210 align:middle line:84% data that are not that high quality 01:07:12.210 --> 01:07:14.280 align:middle line:84% and you can do probabilistic modeling. 01:07:14.280 --> 01:07:16.830 align:middle line:84% You can create-- you have one network, 01:07:16.830 --> 01:07:18.810 align:middle line:84% and you create lots of alternative networks 01:07:18.810 --> 01:07:21.790 align:middle line:84% with individual parameter changes. 01:07:21.790 --> 01:07:25.920 align:middle line:84% And then you just keep pruning that tree-- which model 01:07:25.920 --> 01:07:29.670 align:middle line:84% is giving you the best fit to that low quality microarray 01:07:29.670 --> 01:07:30.460 align:middle line:90% measurement. 01:07:30.460 --> 01:07:32.670 align:middle line:84% So in that case, yes, you can use it. 01:07:32.670 --> 01:07:35.690 align:middle line:84% For parameter fitting, I wouldn't do it. 01:07:35.690 --> 01:07:40.000 align:middle line:90%