Scientist Stories: Kathryn Lilley, Probing the Spatial Proteome

Hello everybody. Thank you very much for the invitation to speak here today about the work that we're doing in my lab in Cambridge. So I think amongst the speakers, I'm probably an outlier. Insomuch that my lab does not produce antibodies, it does not validate antibodies, and in fact, if you looked in my lab's freezers, we own very few of them. So with that out of the way, I'll talk to you about some methods that we've developed which we have used the antibodies to validate the methods. So it's obviously two-way traffic. So my talk is going to be pretty cell biological, so apologies in advance. So what we are interested in in my lab is the spatial proteome. So that is why proteins get to the right part of the cell and what they do when they're there. So you're all absolutely familiar with the way that cells are organized and higher organisms have got very complicated subcellular structures with many organelles, some membrane-bound, some not, and the proteins have got to get to the right place to carry out their function. They've got to meet their binding partners. They've got to meet their substrates if they're enzymes, and so on and so forth. And indeed, the cell is remarkably crowded. So I love the paintings of this guy, Dave Goodsell, who's from the Scripps Institute. And what he does is to take copy numbers of proteins and then draw out what he thinks the inside of a cell looks like. So this is the nucleus, the nuclear membrane, a nuclear pore, the outside, the cytosol, and you can see it's extremely crowded. So a protein gets to the right part of the cell based on a number of different features. It may have a localization signal, so it may have something that looks a bit like this, sets of basic amino acids which are nuclear localization signals, which means that these proteins can associate with import machinery and get into the nucleus. A protein may end up in the right place or in fact multiple places depending with whom it associates. So this PRAC protease can be present in more than one place, and what dictates where it is is simply which isoform of its binding partner P38 it binds to. If it's binding to this isoform, it's a nuclear protein. If it binds to this isoform, it's cytoplasmic because the binding masks its nuclear localization signal. Proteins might end up in different places depending on their isoform status. And in my lab, we've shown that this protein, leucine aminopeptidase 3, has got an alternative splice form, one form of which puts a different exon at the front end, which allows that protein to be trafficked into the mitochondria. Where the short form of the protein is cytosolic. Also post-translational modifications, so a protein can be in completely different parts of the cell depending on its modification state— status. So a pro-apoptotic protein, Bid, is cytosolic, but if it's cleaved by a caspase, its N-terminus is then myristoylated and it ends up associated with the mitochondrion. And also protein structures. So the sort of poster child for differential localization to do with protein structure is the protein aconitase. Depending on whether it's binding RNA or not will depend whether it's a metabolic enzyme in the mitochondria or a cytoplasmic RNA-binding protein. And more and more, there is evidence to suggest that proteins end up where they do simply because of where they are translated. So translating a protein in situ is energetically favorable. It's more easy to shunt RNA around the cell than it is protein. It allows efficient assembly of protein complexes. If something's produced in the same place as its binding partners, it's more likely to find them. Therefore, it's stopping off target effects. And importantly, mistranslation, or being translated at the wrong part of the cell, has been shown to give rise to a number of different diseases. So I hope I persuade you that 50% of the proteome is actually in more than one place. So why this is is open to speculation. Some of it is to do with the fact that proteins, because of their function, traffic around different parts of the cell, but some of it is also to do with increasing number of evidence that proteins have to moonlight. So what I mean by that is that proteins, depending where they are, who they're binding to, and what their structure is, may have a completely different function. And this goes back to the C-value paradox, which has been around for a while now, which basically is saying that as we've evolved, the size of our genome has not evolved to keep up with the amount of complexity which is involved in higher organisms. Sadly, there's no straightforward way to assess this multifunctionality. So if you've got a series of proteins and you think that they may be metabolic enzymes if they're in one place, and structural proteins if they're in another part of the cell, it's very difficult to be able to sample that. So my lab is basically, it's a technology development lab, and what we're trying to do is to answer a few important questions and fine-tuning and developing technology to allow us to do this. So we're interested in the spatial difference between the transcriptome and the proteome, and also where translation takes place. So we want to know how proteins end up in the right place, how that changes and what the dependence of that is based on their post-transcriptional and post-translational status, how the location changes upon perturbation, if proteins adopt different structures in different parts of the cell, what role RNA plays in all of this, not only getting itself translated in different parts of the cell, but also acting as structural binding partners and scaffolds. And what is the relationship between the protein and its corresponding transcript? So if you look where the transcript is, you look where the protein is, do they correlate in any way? And also, if you do have two distinct pools of protein, the same protein but in different locations, are they interchangeable?, or are they translated as completely separate pools and there's no crosstalk between them whatsoever? So these are fairly ambitious questions and aims, and we're certainly nowhere near answering many of these questions. But what we're in the process of doing at least is to develop technologies to be able to at least address some of these. So I want to start first with experiments that show us where a protein is. So this is to get, as Matthias was talking about earlier, an atlas of the proteins in a cell, a street map of the proteins. So I don't need to labor the point at all with this audience that if you want to know where your protein is, you can get one of your lovely antibodies, fluorescently tagged, do immunohistochemistry, show where it is. You could also, if you wanted to, perhaps create a fusion protein and use— with GFP and use microscopy methods to show where that is. So these are very important and very valuable methods, but they're limiting in many ways. Tagging, for example, may direct your protein to the wrong part of the cell. If you're using an antibody, then you want to be sure of its specificity, and it's pretty low throughput. So if you want to look at many proteins on a large scale and across many different perturbations or time points, it's going to be difficult to do that. Another set of methods which are gaining popularity are to use proximity tagging methods. So here your question is going to be a little different. Rather than the question being, here's my protein, I'm fascinated by GAPDH, where is it? This question is, I'm fascinated by this particular subcellular niche, and I want to know all the proteins that are present in it. So here you can direct enzymes that will produce activated biotin when they get to the part of the cell that you're interested in. That activated biotin will then biotinylate any protein nearby, and then those proteins can be extracted using streptavidin, and then mass spectrometry will tell you what the proteins are that are present. In some cases as well, and these are relatively rare cases, you can carry out organelle enrichment to a point where your preparation is pretty homogeneous, and then again the mass spectrometers will give you a parts list of what's present. Now, there aren't that many subcellular niches that can be purified to homogeneity, and with any amount of contamination, your parts list is not necessarily going to distinguish between a genuine resident of whatever the niche is that you're interested in all the contamination. So the other issue with these methods where you're looking at a compartment at a time is that it will tell you what's there, but it won't tell you what the proteins are that you identify are capable of. So if those proteins can be in more than one place, you won't know what that other place is. So the methods that we've been working with for many years now are not entirely novel in their design, but were first muted by this guy here. This is Christian de Duve. He was a Belgian cell biologist. He got the Nobel Prize for the discovery of the peroxisome and the lysosome, and he argued that there was no point trying to carry out purification of organelles and subfractionation of different niches because you were never going to get anything pure. So he argued if you carried out biochemical fractionation of cells, then what you could do is to look at the distribution of proteins throughout these fractions, and if you knew that you had a marker for a particular location, then any other protein that was showing the same distribution, you would assign to the same location. So his methods for doing this was to use equilibrium density centrifugation, where you crack open your cells and you separate out the subcellular niches, the organelles, based on their buoyant density. So the idea here is that the method is holistic and you're looking at as much of the cell as you possibly can in a single experiment. And again, to reinforce the correlation profiling of this, if you take your fractions that have been fractionated based on some physical properties such as density or size of the organelle, then if you look at the abundance of your marker proteins that you know are mitochondrial or you know are from the endoplasmic reticulum, and you work out what their correlation profile is, so their abundance through your fractions, then any protein which is showing a similar correlation, you would assign to the same organelle. So using these methods more than 10 years ago now, we developed a method called LOP-IT, which is the localization of organelle proteins using isotope tagging. The idea is that you take your cells or your organism, whatever it is, you crack open the cells, and these are millions of cells at a time, so from the get-go, this is not a single-cell method. But you take your fractionated cells and then use a dense— a centrifugation method. So the first one I'll talk about is equilibrium density centrifugation to carry out subcellular fractionation. So the idea of equilibrium density centrifugation is that your organelle will end up in the tube at a position which is consistent with its buoyant density. So the heavier organelles will end up at the bottom, the lighter ones at the top of your tube., and then you fractionate. This is a Western blot, so you do use antibodies. So then what you do is to work out whether your fractionation has worked at all by using an antibody in a Western blot to show that you have got some sort of pattern that your antibody to your marker protein is showing, and that that pattern is not the same for every different organelle marker.. And you can carry out side fractionations, for example, if you want to know proteins that are associated with chromatin, and that could be a separate preparation. But at the end of all this, you're measuring all your fractions within the same experiment and then putting the data back together again and trying to make sense of it. So this is the mass spectrometry method that we tend to use. It's not the only one we could use, but this is what's known as an isobaric tagging method. The idea being that the separate fractions that you get from your centrifugation tubes, you convert your proteins to peptides, and then for each set of peptides you label with a specific chemical tag. This tag binds to all amino groups, so the N-terminus of the peptide will bind, as will any lysine residue which happens to be in your peptide. So all the peptides from your first fraction get labeled with one tag, All the peptides from your second fraction that you've got down your tube are labeled with the second tag, and so on and so forth. And we now have 11 different tags at our disposal. You merge everything together, carry out mass spectrometry, and we're using different rounds of mass spectrometry. And if a peptide's present in all your fractions, then you will see a mass which corresponds to that peptide. But then you carry out tandem mass spectrometry, And you fragment your peptide. You get useful ions which will tell you what the sequence of the peptide is and therefore what protein it's been derived from. But the tags drop off to give you the abundance of that peptide and therefore the protein it came from in your samples. So this is what an MS/MS spectrum looks like. We've got some lovely ions here which various search engines will tell us what the sequence is. But down the bottom end of our mass range, we have the tags, and they recapitulate the correlation profiles of our proteins through the fractions that we've got from our subcellular fractionation. So this is now our workflow. We carry out whatever sorts of fractionation we want up here, but ultimately we've got our samples, our proteins that have been fractionated,, and we want to know what the correlation profiles of the proteins are through our fractions, and then what we can do is to use methodologies to visualize what we get. So we've got very, very complicated data. We're now able to identify and give an address to around 10,000 proteins per experiment, and for each one of them we want to have a look at its correlation profile, and then how it matches the markers correlation profiles. And ultimately what we're going to do is to visualize our data using principal components analysis, which will cluster together proteins that are behaving in the same way. And we know what our clusters are because hopefully our clusters contain our protein markers. So this is the sort of thing that we get. This is a principal component plot of mouse embryonic stem cells. Anything in red here are either marker proteins for the mitochondria or proteins that we've classified to the mitochondria using machine learning tools. If I've showed you this in 3 dimensions, above it is floating the peroxisomes, so we've got peroxisomal markers and proteins that we can classify to the peroxisome. Down here, because of the method that we've used, we've subdivided the nucleus. Into proteins here which are associated with chromatin, so DNA binding proteins, transcription factors, et cetera, and proteins here which are nucleoplasmic, also containing the nucleolus. We can visualize and resolve from one another components of the secretory pathway from the endoplasmic reticulum out to the plasma membrane, and we've also got a nice cluster which is consistent with cytosolic proteins. Because of what— how we're fractionating the cells. We've also got information on large protein complexes such as the proteasome, and here these two clusters are the two different ribosomal subunits. We can see 14 discrete locations, and actually with additional markers we can see more than that. But importantly, what we're seeing is the steady state location of proteins in a large number of cells. So Mattia showed very nicely this morning that a protein may be in one place at one part of the cell cycle and another place at another. We're seeing the average because we're working with millions of cells simultaneously. And these little gray crosses are proteins for which we're not classifying. So these are proteins that are either part of some structure that we don't give a name to, but more likely they're proteins of mixed location, so proteins that can be in more than one place. And this is around 50%. So we've had to develop some pretty hefty machine learning tools to be able to deal with these data, not only to look how good our methods are and compare them with other people's workflows, because clearly we want the best resolution that we can possibly get between different organelles. We've also just come up with a Bayesian mixture modeling method to give the probabilities of proteins that are in more than one place to give some idea of the mostly nuclear or mostly cytosolic, and also using orthogonal data sets such as immunohistochemistry data to add confidence to our assignments of proteins to different subcellular locations. We can start to look at biology in action. So we look at proteins that are in more than one place. There's a lovely cluster of importins and exportins between the nucleus and the cytosol. Whose job it is to traffic between one place to another. We can also look at the steady state location of members of signaling pathways, and of course it's comforting to see that your receptors are in the plasma membrane, your terminal transcription factors are on the chromatin, and then you've got the intermediate members of the signaling pathways that are consistent with recycling vesicles. We've also got— can see proteins in complex. We don't do anything to cross-link to keep our proteins together. But you can also see where you have very stable complexes that members of the components of the complex sit on top of one another. Down here, for example, in yellow, we've got components of EIF3, which is involved in translation. And we can see proteins as well, as I said earlier, that are in different places depending on their alternative splicing status. So this is back to LAP3, the long isoform which associates in the mitochondria, and the short isoform which is in an intermediate position between the plasma membrane and the cytosol. So we need to be able to capture dynamic events. So just briefly, this isn't published, but if we treat cells in different ways, and these were a monocytic cell line which we treated with lipopolysaccharide,, we can look at to see what's changing with sets of proteins in terms of their location. So there are certain different features that you would like to see. So for example, sets of proteins that appear because you've treated the cells in some way, so these are now expressed given the perturbation. Proteins that don't change, so proteins, even those that are in more than one place, that remain in more than one place, but also proteins that are moving. So this example I'm showing here is protein kinase C-alpha, which relocates from the cytoplasm to the plasma membrane, and the transcription factor STAT6, which moves from the cytoplasm into the nucleus. So the methods are consistent and reproducible enough to pick up changes of proteins moving from A to B. Now, we've over the years had to prove and show that our methods are valid and that they that make sense and that what we're not doing, particularly in terms of proteins that are in more than one place, is simply picking this up as an artifact of the methods that we're using to fractionate the cells. So many years ago, we talked with Matthias's colleague, Emma Lundberg, and her team about if we were going to use a new cell line, what would be the preferred cell line from them so that we could compare our data with a large amount of beautiful images that Matthias alluded to this morning. So Emma's favorite cell line at the time was U2OS. So what she did was to supply us with this cell line and also the conditions that they'd used to grow them in their large-scale imaging study. So we got our equivalent LOP-IT map. This was work by Katerina Galadaki. And then had the job of using our machine learning tools to classify where we thought proteins were in the cell. And then came the mouth— the mouth in our hearts, hearts in our mouth moment, which was to compare our classifications with Emma and her team's classifications. So these are completely orthogonal methods. She's got single cells and they're intact. We've got millions of cells that we hacked to pieces and hope that the proteins stay in the right place. Interestingly, she also thinks from their data that half the proteins are in more than one place. So what we showed was astonishing, and I don't think at the outset that we ever imagined that there would be such great amounts of agreement. So for certain subcellular locations, the agreement is eye-watering, and that particularly is for the endoplasmic reticulum and the mitochondria. There are some places where we don't agree.. But that, to some extent, is to do with nomenclature. So Emma calls things vesicles where we might have some idea because of sets of proteins which are clustering together what those vesicles may be. She was also able to see subcellular structures that we had no idea that we had any sort of resolution in our data. For example, when we took her nucleolar proteins, or proteins that she thought were nucleolar from her imaging, we could then see within our nucleoplasm cluster that we'd actually got a subcluster, and these nucleolar proteins were clustering very nicely together in our data. And it's also interesting to see that where some of the antibodies Emma was less confident with and some of the antibodies she was more confident with, that our agreement scaled with that confidence. So the more confident that she was in the targets of the antibody and their reliability,, the more there was with agreement with our data. When we have a look at overlap with proteins that we think are in more than one place, then this also was— we were able to see the same sorts of structure in detail in EMA's dataset. And importantly, when we then merged our data with their data, we could carry out different types of classification. And to cut a long story short, using the F1 macros, we were able to see that we improved classification and reproducibility by combining these two very different datasets together. So since this time, we've increased our LOP-IT repertoire. These are three different workflows. In fact, we've got five different workflows that we now use in the lab. They're there for different reasons. The one that I've described using density centrifugation— sorry, equilibrium density centrifugation gives us the maximum resolution. We have a method which is based solely on differential centrifugation. It's easier, doesn't require as many— as much sample size, but we take a hit on resolution. And we have a new method where the idea is to try and maintain RNA integrity. So not only can we get a map of where the proteins are in the cell, but also a map of where the transcripts are in the cell, again to see the commonality of transcript and proteins that correspond to one another in the same place and those that don't. The code names in the lab for these, mostly to do with their ease of use, is Rolls-Royce, Ford Focus, and Skateboard. And they require different amounts of cells, different amounts of times. The rate limiting step for all of them is mass spectrometry. Spectrometry, and the rate-limiting step after you've done that is the data analysis and applying the machine learning tools. But just to show you the difference between our LOP-IT-DC workflow, which we are just about to submit, and our HyperLOP-IT workflow, by the time we've carried out the two different types of centrifugation, we're still using the same downstream quantitation methods and mass spectrometry., and importantly, the same types of machine learning. And these for the same cells, this again is U2OS. You can see that we've got better resolution with HyperLOP-IT, but if you look at the overall structures, there's still a lot of commonality between the two. What is important to show is that where you think that you know biology because the textbooks, Albert and Stryer and Leninger tell you where metabolism, for example, takes place, this is not necessarily the case. So when we look, for example, at glycolysis and gluconeogenesis, which I was always told was cytosolic, those enzymes do come out as being cytosolic, but a proportion of them are in more than one place. And when you have a look what they are, it's intriguing to see that these are the proteins which control flux through these pathways. So it may be, who knows, that one of the ways of actually controlling, sort of a crude way of controlling flux through pathways, is to simply move the important rate-limiting enzymes out of the way. So just finally, and I'm doing this because he said he wanted to know what the acronym meant, we want to capture the RNA as well. And we want to capture the RNA binding proteins which we are, which are very important at locating RNA to the right part of the cell. RNA is a very delicate molecule. It's never allowed out on its own. It's coordinated by proteins, so can we fish out the RNA binding proteome? So the way that it's done at the moment is to cross-link with UV protein to RNA and then pull out the RNA based on its poly-A tail. Now, not all RNA has got a poly-A tail, poly-A tail, and this also requires far too many cells, this many if we were going to try and combine this method with LOP-IT. So we've got this many cells down to half of one of these with our new method. The method's called OOPS, or Orthogonal Organic Phase Separation. We tried to call it something more sensible, but once it was called that, it stuck. And it's so easy it's untrue, so if nothing else, I will share this easy method with you. So the idea is that if any of you have worked with RNA, you know that the way that you purify it from your cells is to use the trizol reagent. So the trizol reagent, when combined with chloroform, will partition. RNA goes into the aqueous, protein goes into the organic. If you cross-link them together, then they end up at the interface. And they end up not only at the interface, but if you take them through multiple rounds of this trizol chloroform extraction, which is super simple, that you end up with very pure RNA-bound protein. So the very interface that the protocols say do not touch this interface actually has got a large number of RNA binding proteins bound to RNA anyway. But after cross-linking, it becomes a very, very enriched fraction. So I won't go through all this other than to say that when we do mass spectrometry on the proteins that's pulled out, it makes some sort of sense, but also adds to the list of proteins that we know bind RNA. We can have a look at the binding domains. This is the message I want to get over, I suppose, is that there's an awful lot of RNA-binding proteins that seem to be able to bind RNA without having a canonical RNA-binding sequence. We, by combining different proteases with ways of fishing out peptides that contain RNA after you've chopped with RNAs, you still get sort of a nub of RNA present, which is going to be phosphorylated. We can actually work out the binding sites of where the RNA is on some of these unusual proteins. Many of these binding sites seem to map with domains of proteins which are targets for drugs. So I think that is an important message. Some of them make perfect sense. Some of them are surprising. When we did— because the method is so simple and we don't need very many cells to be able to do this, we were able to map RNA binding proteins through the cell cycle, at least through cell cycle arrest and release. And what we showed was that the good old glycolytic enzymes, which I've banged on about already, are RNA binding proteins. Some of them were known to be, very few of them shown to be in vivo. We've shown that basically all of them are, and because the method's so simple, we've gone right back through evolution looking at Arabidopsis and Drosophila and yeast all the way to E. coli, and E. coli, they bind RNA as well. So this feature has been around a long time. During cell cycle arrest, they step back. When you then take the cells and un-arrest them, they become RNA-binding proteins again. And because we can map the structures, we can see, for example, GAPDH, where the RNA is binding. And the fact that this is mapping to its coenzyme binding site and its Rossmann fold. And we can now, because we don't need very much material, we can now combine it with the LOP-IT method to show the subcellular location of these RNA binding proteins, which seem to be in many, many, many different compartments of the cell and not just confined to the cytosol and the nucleus. So many messages I hope I've got over. Importantly, half the proteome seems to be in more than one place. The EAPS method is a stupid name, but at least it's quick and flexible and allows you to retrieve RNA-binding proteins from very small sample sizes, which means there's a lot of cell biology which can now be studied and unearthed. And I think these are the messages that I put in a jet-lagged state on the train yesterday. The subcellular targets of antibodies, a validation would be to compare with what we find using a very, very orthogonal method. 50% of proteins are of no fixed abode, and so if there are multiple signals in IHC, then that makes some sort of sense. Therapeutic targets, if your protein is a moonlighter, then you might think that your therapeutic target is going to knock out one feature of that protein, but actually, It may be knocking out activity that you don't know about and is essential to the cell. And I think we've shown, but I haven't shown you in any detail, that RNA binding capacity is much greater than imagined in proteins and also in some cases maps to sites for which drug targets have been designed against. So here's my lab. There's a lot of them. If they're in red, they're informaticians. If they're in blue, they're not. We have lots of collaborators here, there, and everywhere, and most of the work that I presented today has been funded by the Wellcome Trust.

Thank you very much for your attention. [APPLAUSE] It's completely elegant and fantastic work. The briefest of brief questions.

Please.

I wonder, how does your fractionation workflow change on the principal component analysis, for example, you would add insertion? So are you looking at lipid association of proteins, because that can really change, that can shift your fraction to the opposite side of the post.

We don't use detergent when we fractionate the cells because if we did, the organelle integrity might be impacted. And if that were the case, then that would impact the behavior of the organelles in our differential centrifugation methods. What we have done previously, certainly in our early— so our reason for being was to map membrane proteins in the Golgi apparatus in plants. Plants, that's how we started out 15 years ago. It's quite a difficult thing to do because plants are horrible, very difficult to crack the cells open. And there, because we were only interested in membrane proteins, what we did once we'd fractionated was to use carbonate washing to fish off any peripherally associated proteins. So what we ended up with then were very meaty proteins with transmembrane spanning regions. At that point we had to use a lot of detergent to try and solubilize them. But we certainly don't use them at the point of fractionating the cells, because if we did, we'd lose all the different densities and sizes of the organelles, which are the very things that we're separating upon.