Scientist Stories: Steven Gygi, Developing New Technologies for Mass Spectrometry-Based Proteomics

Steve is a professor of cell biology at Harvard Medical School, and as I think many of you know, one of the world experts on the use of mass spectrometry for proteomics. Thanks. I can probably do it from right here, I think. That's good. 2, 3, 4, 5, 6, 7, 8, 9, 10. Good. Okay, thank you for inviting me, Mike. Yeah, I'd never really been fully exposed to the MIT culture before today. This is my first talk here. And when I visited the bathroom downstairs, there's a huge chalkboard in the bathroom, wall to wall. And right by the final urinal there, there's this— you can see it's been erased, but it says Bob Weinberg, 1966. I hate the food. I will never return. To this building. So, you know, now I understand where some of these amazing ideas have really come from, you know, that have come out of this place. So it's just a little different across the river. We don't have chalkboards in our bathrooms. It's pretty exciting. So let's just make sure I'm going to go forward here. Okay. So my lab is mostly interested in in really measurement science. We're interested in being able to detect things that have never been detected on a scale that they haven't been measured before. And so we spend 80%, 90% of our time just thinking about measurement science and how we can improve that. And the other 10% of our time, we try to apply those technologies to important biological questions. I'll just go through a couple of background slides here to help us all understand. What the basic technology is that we're using. The term shotgun sequencing was coined by John Yates, and it really kind of talks about how we, how we collect all of the data that we do by mass spectrometry. So imagine that you have a gel here and you have a protein that's here. It's going to cure cancer, save children's lives. It's a really important protein. You want to know what it is. You can, you can excise that protein from the gel and then add a protease, say trypsin. That will create maybe 50 or 75 peptides. Those peptides then can be separated using really HPLC, but it's a very tiny version of HPLC. The diameter of the columns about the width of a human hair, very small, 75 microns in diameter. And the flow rates are very small as well. And then these peptides, as they elute into the mass spectrometer, it's going to switch back and forth between doing two things. First, it's going to measure the mass of those peptides. If it's 12 amino acids long, it's gonna measure something about 1,200. And, and that's the first thing it's gonna measure. But in a fraction of a second later, it's going to switch over to what's called MS/MS mode. Then it's not going to be measuring the mass of that peptide anymore. It's going to be measuring the mass of the fragments of that peptide. It will physically isolate that peptide, and then it will fragment it, and you'll get a fingerprint. All of the peaks in there correspond to structural information about that peptide. And so you do this on a rotating basis and you collect lots and lots of data. And the fragmentation spectra allows you to identify those peptides. And you also know something about the mass of those peptides as well. And so we can do this on a scale that's pretty amazing now. We don't usually work with a single gel band anymore. But if you had a protein complex like Wade just talked about, that might have 50 or 100 proteins. You could proteolyze that directly. And then in the mass spectrometer, you could measure all of those proteins directly. So you could identify hundreds of proteins there. And often we even look at total cell lysate. So you might have thousands of proteins. You make it more complicated by adding a protease. So now it's hundreds of thousands of peptides. Then those peptides can then be detected, analyzed and identified ultimately and assigned back to the protein. So it's a pretty amazing technology. And one of the key numbers, I guess, is that the newest mass spectrometers can collect about 1,000 of these MS/MS spectra per minute. So over the course of an hour, you might collect 60,000 of these sequencing attempts, assign those back to 4,000 or 5,000 proteins. So it's a pretty powerful technology. One of the other things that we make we make grand use of is our isotopes. And so J.J. Thomson is kind of a hero to all of us. He discovered the electron, which is a small thing, and he got a Nobel Prize for that. But then after that, he discovered isotopes, which is really cool. And then he went on to build the first mass spectrometer. So, you know, these isotopes that you see in here, they complicate our life, but at the same time, they make everything we do possible. And that's because if I had a peptide that had 50 carbons in it, 1% of the carbon is C-13 in nature. So that means I don't measure a single peak in the mass spectrometer. I measure something heterogeneous here where this would represent the population of that peptide that had no carbon-13s, just by chance. If you had a single carbon-13 somewhere in your peptide, you would be in this population. If you had 2 carbon-13s, you'd be there. We have this underlying principle that is in all of our data that we have to deal with, and it makes it more difficult. But it also allows us to do things that are wonderful. And that is we can artificially, as Wade showed in the previous slide, we can artificially add stable isotopes. Say you add a lysine that's enriched with 6 carbon-13s. It would move the mass of that peptide over to here. You would still have all the isotopes associated with that. But you would have 6 guaranteed stable isotopes that would be in here. And then by comparing these 2 peaks, you could have a binary comparison of what's in there. And so a few years ago, we were able to do proteome-wide comparisons at a binary level, right? So you could say with a treated or non-treated or cancer or non-cancer, you could look at protein expression levels on pretty much a comprehensive scale. But what if you wanted to do this with more samples than that? What if you don't want to just look at two things? So, so now multiplexing is emerging as really probably the key new technology, I think, in our field. And for example, if you had 6 samples, let's say you had 3 treated samples and 3 that weren't treated, and you wanted to analyze those at the same time, you could do that using something called these isobaric tags. There's either iTRAQ reagents or TMT. Are the two kind of brand names that, that they come by. But they're very, very special in that they use isotopes, so they have all of that that I was telling you about before, but they also have a principle that allows them to not be distinguishable when, when the mass spectrometer operates in that first mode where it measures the mass. So when I measure the mass of the peptide, I don't see 6 peaks, even though I have 6 different labels. I see one peak and they stack up on top of each other. So that's pretty cool because you could imagine you could have 24 of these or 96 of these and it would be no less complicated at that level than it was before. So that's the future, right? Putting 96 of these together. But later on when the mass spectrometer switches over to sequencing, they do, they do fragment differently. And so they separate into, say, these are 1:1:1 ratios here, but you could You could imagine that you could then quantify all 6 of them in a single scan. And so this is the way that multiplexing happens, is that you can label things, mix them together, and it's no more complicated than it was before, which is wonderful. And then you can quantify them in a subsequent fragmentation spectrum, you know, that you have there. So, for example, if you cared about the brain and the liver, you might take 3 mice and you might take out the brain and the liver, and then you might want to look at what's different at the protein level on a very large level between that after you had first proteolyzed each one of those, made them into peptides. You would then label those with one of the reagents at their means. Then you would mix them together. So now you have one really complicated mixture of peptides. And then you would usually— we fractionate that into a few peptide fractions because maybe 1 hour or 2 hours isn't enough time on the mass spec. It's a shotgun technique. We want to spend 24 hours maybe doing the sequencing. So after we fractionate it, one run on the mass spec might look like this, and the mass spec switching back and forth. Back and forth. It's sequencing and it's identifying peptides. And at the end, you might have one spectrum might look like this, where this is from a protein called cytochrome b5, and of course this is a liver-specific protein. And so you can see in the liver it's much higher for these 3 labels than it is for those 3 labels. And so you get the idea that this is how it works. We've since then extended the labels from 6 to 10, and we used a pretty neat trick. If If you have a carbon-12 atom and you move it to a carbon-13 atom, that's a neutron, right? That has a mass of 1. If you take a nitrogen-14 and you move it to a nitrogen-15, that's also a neutron, has a mass of 1. But the masses are not the same. Those neutrons don't have the same masses because the binding energy in the nucleus is different. And because of E mc², they're slightly different. And so in the mass spectrometer, they're resolved at the third decimal place. They're about 6 milli-mass units different. A carbon-12 to a 13 compared to a nitrogen-14 to a 15. That means if I use a nitrogen isotope instead of a carbon isotope, I actually can increase— simply by using this nitrogen here, I can actually increase the number of labels. So now we have 10 labels because we had a nitrogen that we could play with. And if somebody wanted to add another nitrogen, they could have more labels and more labels. And so you can see how you could play games with this trick to increase the number of labels much greater. But this was a nice trick to get to 10, which is where we stand right now. So I'll apply some of these ideas now to showing you where we're at with the multiplexing. And I'll try to stick with colorectal cancer for most of these experiments. So we like to study colorectal cancer because the cell models are just really good. And of course, there's always the Wnt signaling alterations, but at the same time, there's layers on top of that. With lots of the other classical kind of oncogenes or tumor suppressors, the KRAS, the BRAF, the others are there. And there are cell lines that are— that also then have removed, say, the KRAS oncogene so that you have the same cell line with and without that. So you can do a lot of nice comparisons. They're really nice models that aren't really available in many other cancer types. And also they come in two forms. So you really have this kind of hypermutated tumor where you have perhaps 10 times more mutations, or you have these kind of non-hypermutated versions. So, so there's all these kind of how do you classify these tumors, as well as, as what's really going on with the layering of these additional, these additional oncogenes. So just as an example here, down here are 7 different cell lines. These are tumors, these are colorectal cancer cell lines, and here's just a few of the, from the sequencing projects, a few of the mutations. They're colored based on what type of mutation they are. But just look and count up here. If you look at the COLO205, there's 1, 2, 3. If you look at the HCT16, there's about 10. And the same with the HCT15s. The HT29 is, again, not a very hypermutated, so it has fewer. It has only 4. And so is the HT55. So, again, you get this idea, again, that these tumors come in two different types, even just from that kind of an analysis that you can do just by looking at it and trying to understand what's really going on with these. So as a first step, we took— we had an 8-plex at that time. We hadn't synthesized the other two reagents. And so we said, let's look at the protein expression profiling on a global level across 8 colon cancer cell lines. They all come from the same source. So you really shouldn't have a lot of tissue-specific effects. They should be mostly due to either aneuploidy effects, gene dosage, or perhaps the mutational state, you know, of these cells. And so we pulled out these 8 and we went through the process of proteolyzing them, labeling them, mixing them together and into the mass spectrometer. And looking at the data. And when we were done, we went back and did this in biological triplicate. So we had now 3 different 8-plexes, right, and the overlap of all of those was 7,200 proteins. We quantified each of those in each of those replicates. And so if you think about that for a minute, it's actually really astounding because I have 24 measurements then, because I have 8 cell lines times 3 biological replicates for each protein. So I have 7,200 measurements times 24. So there's really 172,000 protein measurements in this dataset. And underneath of that is the peptide measurements. There's about 10 peptide measurements per protein. So there's probably north of 1 million different measurements in this dataset that when you put it all together, you know, you have to make sense out of as well. So again, these experiments are pretty astounding, you know, where they're headed and what they can do. And this is just an example. Rather than show you a correlation of how the replicates look, just so you can understand it, I just pulled one out to make it easier to look at. It's awful big on the screen. But this is, this is villain 1. And what you can kind of see here is that the 3 replicates really cluster together very nicely here. And these replicates weren't, weren't like I took 3 dishes right next to each other and those are my biological replicates. We did the first replicate, waited a few months, and then went back and thawed the cells, remade them, did them again. So they— not only do they have a lot of variability in many proteins, but they also are very reproducible in where they grow back to. So, you know, that's just something to think about there. But you can see that at least 4 of these cell lines have very low levels of this particular protein, and then 3 or 4 of these are definitely expressing those. And even from here to here, there's a 4-fold difference, you know, between those for that. And so this is the type of data that exists across the 7,200 proteins. Here's another protein involved in mismatch repair. You can see in the LOVO which is a hypermutated cell line, this is likely missing or very low, and different levels in other cells. So, again, the reproducibility was one of the first issues, and just the wide variability, too, where these cells are just remarkably variable in their expression levels, but at the same time they're consistent. And you can do other things with this. You can just ask questions and say, show me everything that's unique to the HT-29 cell line. That's what I care about, is HT-29 cells. And you might see that, for example, here's a gamma-synuclein is pretty much exclusive here. And you might think, what if I knock that down? Would that kill that particular cell line? And it's interesting just to look across the 8 cell lines and say, what are the completely unique proteins just to one cell line? Many, many of these are secreted proteins or they're plasma membrane proteins. So it is— they are kind of marking themselves as different, you know, as different cells. As they go through that. And so you get, you know, just a list of things that are like that. And you can do— play all kinds of games like that. So I won't have time to show you too many things about what we did with this data, because I want to get to the signaling part of this. But one of the things you can do is cluster the data, obviously. The 5 red dots here correspond to the hypermutated cancer cell lines that we had, and the blue dots correspond to the non-hypermutated. So they kind of cluster based on that. You can You can kind of see that. But also, you can just ask questions at the breakpoints here and say, what are the categories of proteins that are really influencing this particular branch here on the tree? And for this one, for example, it's signaling. And so if you look over here at the cluster close— more closely, I'm putting a line here between the 5 cell lines over here that correspond to the hypermutated and the 3 cell lines over here that are the non-hypermutated. And you can see that the signaling branch over here, these are the proteins that came back as significant from that signaling category. And you can see that it's kind of different here. For the 3 replicates here, they're not really— there are proteins in signaling that are always changed, but it's not the same proteins. I don't see one protein that goes across and really kind of unifies that that's what's changed in those cells. And so this is kind of a theme that we're seeing too, that you really can't have a protein-centric view at all in these datasets. It has to be either pathway or even network-centric. Or you're not going to understand what's going on, because this truly tells you, you know, those kinds. And same with cytoskeleton, which is what really changes over here. You see genes in cytoskeleton in different cell lines— each of these is a different cell line— but you don't really get the sense that there's any really unifying theme across there other than the category, you know, itself. So this is just one of the ideas that we had. So I'd like to kind of end here with a longer example here of of phosphotyrosine and what's going on there. So it turns out that the DLD1 cell line is actually available both as the way it was isolated from the patient that has a KRAS mutation, as well as in another form where the KRAS mutant gene has been removed. And so they have a wild-type copy, probably from their mom— their dad's probably the bad one— but they have a wild-type copy still there of KRAS, but the mutant version has been removed. And so if you take these two different cell lines, which are isogenic except for, you know, that particular— either having the oncogene or not— and you then want to look at how they're wired or how the signaling is happening, you might want to look at EGF signaling. It's going to turn on a lot of interesting things for you. You could add EGF and you could run a time course here. You could add EGF to the other cell line and run a time course there. And if we used our multiplexing platform We would need to modify it a little bit in order to allow us to measure phosphotyrosine differences. And so to do this, after we digested it, we didn't label immediately each of these 8 samples. We first took them over a titanium dioxide column in order to purify the phosphopeptides. Once we had the phosphopeptides, we labeled those, save money. And then we put those together that were labeled into a single— so now they are a single pot of phosphopeptides, some of which are phosphotyrosine. And we purified the phosphotyrosine using the cell signaling antibody to do that. And so after pulling out just the tyrosine phosphorylated peptides, we put those into the mass spectrometer in a single run. So all the data I'll show you from this is about 3 hours on the mass spectrometer. It's just an example of one so we can kind of orient ourselves where we're at. Here's the scheme again. If you look at the wild type here, there's not much— on this EGFR1172, there's not much tyrosine phosphorylation at that site. But immediately at 5 minutes, it goes up like it should. And then it comes down 10 minutes and 30 minutes. But if you look at the KRAS mutant in these DLD1 cells— so these have the oncogene— they start very, very high. So they're already phosphorylated at that site. And it actually goes the other direction. When it sees EGF, which means you must be activating a phosphatase or turning off a kinase or both in order to have it return to normal, you know, to levels down here. So this is the kind of thing we started to see almost immediately in this, is how they're just wired a little bit differently there. And it gets a little hard to look at because they're so different. There's around 200 phosphotyrosines that I've got on this list here. Here's what they look like in a heat map., but it, you know, normally you would expect to see it look like I said, you'd expect it to go to the 5 minutes and then back down. And it kind of does that. A lot of the yellow in here at 5 minutes, it's going up and coming down. So that's kind of the major thing that's happening. But a lot of tyrosines just don't do anything or they go down. So it is really kind of a mixed bag of things that happen there. But the other side is also kind of a mixture. I just kind of show this heat map to let you know how dynamic it really is. It isn't some obvious, heatmap that really helps you to understand what's going on. But maybe you can understand this better if you use, maybe reduce the dimensions a little bit with, say, principal component analysis. So we had 200 phosphotyrosines that I was measuring there. And that's 200 dimensions of data. If I just look at what's happening in general in another axis in 2 dimensions, it would look like this.— and you could see that you might start with the wild type, you might start over here on average for these 200, and they generally just shoot across the x-axis and end up over here at 5 minutes, then they go back to the 10 and they end up at the 30. They don't end up back at the wild type, but they end up kind of getting back towards the wild type, and that's kind of a typical trajectory. With the mutant form, it does the same thing at the 5-minute time point, it generally does go up, although I showed you the example of one that doesn't, and then it tends to move, you know, back, but in a much more blunted— so this is kind of the way you can perhaps look at some of the data, things like that. You can also do something where you can take all of those 200 phosphotyrosines, and I can plot them back as their weights, their loadings, back onto this graph here. And so if you walk around and look at these, you can get the full view of everything that's going on there. For example, up here I have this ephrin B6, and it doesn't have any change in the first 4 time points. But it's not— but it is different in its level in the KRAS mutant, but it's not dependent on EGF for what happens. If I look at another one over here, I see this Yes1 kinase. This is the activation loop, I think, on Yes1. It's not affected at all by EGF, but it comes down when you add EGF in the KRAS mutant. And over here you see this channel here that is tyrosine phosphorylated, but very late at 30 minutes, and it is— it also is tyrosine phosphorylated here, but at a reduced rate in the mutant. So you can walk around and see blunted peaks, or you can see some that have, like on the MAP kinase, you can see it going up and down but not doing anything in the mutant when KRAS is hyperactive. And so you can really get an idea about everything that's going on here. So this is a kind of a great way to kind of get an idea about what are the possibilities in these spectra. So we went back, and they also— there's also another cell line called HCT116 cells, and they also had a KRAS active mutant when it was isolated from the patient. But a cell line's been created such that it doesn't have the KRAS again. And so we had another cell line that we could compare what we learned from the DLD1 cells, compare that back to the HCT now, and ask questions. And it's quite different, because remember what I showed you for this one, where this goes up and down, but this one just comes down. In the mutant form. Over here you can see that it definitely goes up and down again in the wild type, but it also just goes up and down in the mutant. And so KRAS isn't wired— it's not wired in the same way in the DLD1 cells, such that you don't see— you don't see this, you know, you don't see this in those cells. And the kinase ES1, again, which John Blenis told me he thinks is probably responsible for this particular site, because it's decreasing here and this is decreasing. It's not changing in either of the cases, you know, for its phosphotyrosine. So, you know, we have a bigger dataset there. The PCA looks very different for this one. It doesn't go across. If you remember, it went across the x-axis last time in the way that it worked. In this case, it just kind of rolls up to 5 minutes. It kind of rolls up to 5 minutes and then back down to 10 and 60 minutes over here, and the same with the mutant. So, you know, we're just starting to get a handle on layering of these particular oncogenes on top of, you know, the same cell lines and seeing what happens to those. And of course, you can look at the loadings for these, and they look very different. You know, they're just not— they're just not as exciting, I guess, as the DLD1. For whatever reason, there's really only 4 things that happen. You either see things that are there in the mutant and not there in the wild type, but they're not responsive to EGF, or you see things that go the other direction and Like clock, you know, here, vimentin goes up, clock goes down. You see the EGF signaling here is maybe a little bit blunted in, say, you know, both cases. And that's what you see when you go around here. So it's just a different setup, a different wiring, but you kind of get an idea about what's happening in these cells compared to the DLD1 cells. So that's kind of where I wanted to take you today on a little journey through where we're thinking about and what we're doing. And it is kind of a fairy tale ending. I like this slide because it reminds me that didn't always turn out the way it should, and maybe Snow White's not that happy right now. She got what she wanted, she married the prince, but he's watching Polo on the TV, she's got at least 4 kids, which I guess is a good thing, and, you know, so, but we have to make sure, I think, again, that in proteomics, as we get these fantastic new technologies, that we treat it like a fairy tale as well, and that we make sure that it ends right, and that we, you know, we get where we want to get with these, because The technologies are evolving so quickly that it's hard to even keep up, right, let alone afford them. But it's hard to keep up with these. And so I've shown you that multiplexing is pretty mature now. I mean, actually, we can use it at almost a global scale, you know, with what we're trying to do. I showed you some examples of 6-, 8-, and 10-plex that we can do. And I tried to show you that phosphorylation is going to play a major role, I think, in proteomics and in signaling in the future. And really, the complicated experimental designs are what we're really excited about now. We might have more later, but with 10 you can already think about some pretty complicated experimental designs that you might want to do. And with that, I just thank the members of the lab who did all the work. I didn't really do any of this work. Most of the phosphotyrosine work that I talked to you about was kind of spearheaded— the figures that you saw in the statistics were done by Ed Hutland in the lab. Mark Jedrachowski grew all the cell lines, and David as well helped him and did some of the resistical analysis. Robert Everly and Graham McAllister are the instrumentation guys that made this happen, and Robert's really done a lot on the phosphotyrosine pulldowns as well. So thank you for your time.