Scientist Stories: Steven Gygi, Building Protein Interaction Network Maps

Hello everyone, and welcome to today's live webcast on Bio-Plex, a protein interaction network created from thousands of protein immunopurifications. My name is Kim Pham from LabRoots, and I'll be your moderator for today's event. Today's webcast is presented by LabRoots, the leading social media site for science professionals, and sponsored by Cell Signaling Technology. Cell Signaling Technology is a private, family-owned company founded by scientists and dedicated to providing the world's highest quality innovative research and diagnostic products to accelerate biological understanding and enable personalized medicine. Their employees operate worldwide from their US headquarters in Massachusetts and their office in the Netherlands, China, and Japan. For more information, please visit www.cellsignal.com. Before we get started, here are a few instructions. We want to hear from you during this interactive broadcast, so please ask questions or leave us a comment. Answers are also welcome. You can do this by hitting the green Q&A button at the lower left of the presentation window and typing in your comments and questions. We'll try to get to as many as we can, and we'll follow up with you if we don't have time for it today. Do you want a better look? You can enlarge the slide window by clicking on the screen icon in the lower right-hand corner of the slide window. If you can't hear or see this presentation properly, please let us know by clicking on the support button on the top right or use the Q&A button. We'll make sure to resolve any issues. This webinar has been approved for continuing education credits. Please click on the CE button at the bottom left corner and follow the process to receive your credits after the webcast. Now let's get right to today's presenter. We're proud to present Dr. Stephen Gigi. Dr. Gigi received his PhD from the University of Utah in the area of pharmacology and toxicology. He performed a postdoctoral fellowship with Rudy Ebersole at the University of Washington in 1996. He joined the faculty at Harvard Medical School in 2000, and he is currently a professor in the Department of Cell Biology. Dr. Gigi is a leading knowledge is to use mass spectrometry to answer fundamental questions in both normal and abnormal biology. He specializes in instrumentation advances for global cellular protein measurements. I'll now turn it over to Dr. Stephen Didi.

Okay, thank you for this opportunity to speak today. I'm excited to try out this new format, which I haven't actually experienced before. I'm also grateful to Cell Signaling Technology for both inviting and sponsoring sponsoring me in this endeavor. I hope that it will become abundantly clear through this talk that protein-protein interactions can provide a powerful window into nearly any gene-specific biology. So here's the outline from my talk. It will be split into two parts. I will first go over the key concepts for about 15 minutes, and then I'll move on to an example of collecting a large protein-protein interaction dataset, the BioPlex dataset that we'll talk about. So first, we need a definition of proteomics, and so proteomics is really the systematic here analysis of proteins for many properties. This includes their expression, their activities, their structure, their modification state, and for today's lecture, for their interactions. And there are 5 key concepts that I'd like to go over to help us understand a little bit more about what is mass spectrometry-based proteomics. So the first concept is— that we should remember is what exactly is a mass spectrometer and what exactly does it measure? Well, a mass spectrometer is the smallest scale in the world. It can measure the molecular weights of atoms. For example, carbon is 12, nitrogen is 14. Let's take a case specific to proteomics though. So let's say we have 3 peptides and we want to know the sequence of those 3 peptides. Let's say one of these peptides is, say, 7 amino acids and one of them is 10 and the other is 12 amino acids. So we have 3 peptides that we're interested in and for those 3 peptides we want to know the sequence of those 3. Well, the mass spectrometer can help us with that. Infuse these into a mass spectrometer, and the mass spectrometer requires two principles to work. Everything needs to be in the gas phase, and everything needs to be charged. It can't work with neutrals. And so this electrospray ionization process here can accomplish both of those, and the peptides can then enter the gas phase in a charged form. And then they can be passed through this quadrupole mass analyzer here, which can allow for determining the molecular masses expressed as mass-to-charge ratios for those, and those are detected via an electron multiplier in this case. On the mass spectrometer. And so now we could know, if you look at the bottom here where there's a mass spectrum, on the x-axis there's a mass-to-charge ratio, m/z, and this is really where you're measuring the mass. And on the y-axis there's the abundance that's there. So this is the signal that the mass spectrometer is seeing. But how does knowing the How does knowing the mass of this number 12 here at 667, how does that help me to know that it's alanine, glycine, leucine, arginine? How can I learn the sequence based on just the mass? Well, to do this we need another step, another stage, and actually a different mass spectrometer, which is called a tandem mass spectrometer. So this is called multi-step analysis, multi-step analysis. In mass spectrometry. So this is a different situation here because we have a second stage of mass spectrometry which is used to collect the sequence information of the peptide. So a mass spectrometer is capable of collecting MS/MS spectra, and this is where the— in the experiment where the this 667 ion here is measured in the MS scan, and then a fraction of a second later, the mass spectrometer will turn over and try to collect an MS/MS spectrum. And let me explain what that is. An MS/MS spectrum is where that 667 peptide ion is physically isolated away from the other peptide ions, and then it is energized and collided here with, say, an inert gas in this dedicated collision cell here, and so it's fragmented. And then here in the third quadrupole, we read out the fragments that are coming through there. And so this mass spectrum here, this MS/MS spectrum here, it really represents all of the fragmentation information, the structural information from that 677 ion. And this really can tell us about the amino acid sequence here, because, for example, the mass difference between these two peaks here that might be 99 and that might be a valine, right? So you could start to learn something about the sequence information there. But the real key issue here is to do an MS/MS experiment, you want to fragment everything but only once. And so on average you take 10,000 molecules, but you're not breaking it down into all of the amino acids. You are breaking it on average only once, so you're creating two pieces. And when you measure the masses of those two pieces, you create a tandem mass spectrum. So peptides can be fragmented at many positions on the backbone, and we'd like to push out a poll here too just to see what you're thinking about and where you work. But here is most commonly the amide bond, which is the place on the ribosome where these amino acids are put together. That is the weakest bond, and that is where normally a peptide will fragment under these conditions. And so if you look at a peptide, it will have many places along the backbone where it could be cleaved. And for example, cleavage here, cleavage here would produce two pieces. It would create this piece here which contains the intact C-terminus and this piece here which contains the intact N-terminus. They have different masses. But if you fragmented it on average at all of these different positions, but on average only once per molecule, you would read out these many fragments and that would give you a tandem mass spectrum. And then you put it all together, it looks something like this where we can see that a tandem mass spectrum truly has a very rich supply of sequencing information contained within it. For example, cleavage is here between this, this serine and this glycine here that produces those two halves again as I was telling you, this piece here going down. Would have a mass of about 1172, and that would be this peak right here. So you can see it starts with a glycine here and ends with an arginine, and it has a mass of 1172. So you can see, as you look at this, you can see all these fragments and how they represent the sequence information. So to put this together a little bit more is the third concept. It's about database searching and protein ID, and the goal here is really to take this particular tandem mass spectrum to reassociate it back with the peptide to find the sequence of that peptide and then to assign that back to the protein. So we're just moving backwards towards where we started. To do that, you would think that you could just take the protein here and— take this protein here and then assign it over to the identified protein by just collecting the MS/MS spectrum of a protein. But typically, we don't deal with proteins that much at that level. This is a whole other field which is called top-down proteomics that's dedicated to just this. But for the majority of us, peptide sequence is much easier and it's more mature. To do this, we use kind of a circuitous route where we start at the protein and then we move to peptides here— trying to draw everything— we move to peptides here by digesting it with an enzyme, then through sequencing in the MS/MS in the mass spectrometer, we identify those peptides. And after identifying those peptides, we put them back together and associate them back to get to the protein. So there's this circuitous route that we take as we move to identify a protein by mass spectrometry. So how does this sequencing work? We assign the sequence match to an MS/MS spectrum. This is accomplished by several algorithms. The very first was called Sequest, and it is still the one my lab uses today. Sequest takes, for example, the entire protein database from yeast, and then it would digest that protein database in silico. So you create hundreds of thousands of peptides from those 6,000 proteins, and then only the peptides that had the mass that would— that could potentially match the peptide that I'm seeing would be considered. And then you would— Sequest would compare a theoretical tandem mass spectrum with an acquired one that I have here. And by correlating the score for this, we could learn and match the peptide back to what's in the database. So this is kind of how it works. It matches it back up using predicted spectra. And you can always go back and look at it and see how you did with that and inspect it and see how things look. And what you really notice when you do this is that there's really two ion series here, meaning we're sequencing it this way from the C-terminus, and these ions are the same thing, but they come from the N-terminus. So there's two different ion series, lots of information to help us to sequence a peptide by mass spectrometry. But the final thing maybe to remember is that we are trying to identify a protein here. When we digested the protein, we created many peptides. So we would expect to see many of these, more than one peptide identified by our database search. And indeed, in this example, there are 7 peptides that we can see here. And these 7 peptides all correspond back to the same protein. So this increases confidence in the correct identification. It really nails down the splice isoforms. You might find a peptide that crosses an intron-exon boundary, or you might find one that's specific to a particular exon. It also can provide a measure of abundance. If, if I identified by 7 peptides in this, under this condition for this protein, and the next time I identified by 70 under a different condition, The fact that I had 70 peptides matching would suggest, and you could rightly conclude, that there was more protein present in the other condition. So you can use it as a measure, as a quantitative measure as well, this number of peptides, and that's called spectral counting. You see that used quite a bit where the number of spectra is kind of a measure of how much the abundance of the particular protein was in the mixture. So the fourth concept is putting it all together. It's called shotgun sequencing. Shotgun sequencing is where peptides are loaded as mixtures onto reverse-phase columns with really small diameters. These columns are called nanoscale here, nanoscale capillary columns, and they have very small diameters, maybe the width of a human hair, 75 microns in diameter, for example, but they're filled with the same material that you would see really almost in a normal HPLC column of bigger diameter. These are 5-micron glass beads usually, and they are derivatized, functionalized with say 18 carbons. So they're very greasy, they stick to peptides, and then as you increase the organic concentration that's going across the column, the peptides will begin to elute based on their hydrophobic properties. And as they elute, the mass spectrometer sees them, it detects first, and then it turns over and collects an MS/MS spectrum where it actually tries to sequence those. So as fast as the peptides are eluting, the mass spectrometer is trying to both detect them and collect the sequence information on those. And it just goes around in this big cycle doing this. And in a single hour, you can collect something like 10,000 tandem mass spectra across this gradient here. And so you could imagine that you could take hundreds of proteins and those could be digested and put those into this system here, and you could probably identify most of those proteins, if not all of those proteins, at the same time in the mass spectrometer. So it's pretty powerful. In fact, you could take it one step further and say, what if I had a HeLa cell? This has 10,000 expressed proteins in this cancer cell. If I took lysate from that and I added trypsin, I would create kind of an unholy mess of peptides, maybe a million peptides. Well, you could try to separate those peptides on these columns as well, and as they eluted, the mass spectrometer would still be sequencing as fast as it could peptides, and as long as peptides don't co-elute with the same mass at the same time, the mass spectrometer could still be able to identify those, and you could identify thousands of proteins that were present in that mixture at the same time. It's a very powerful technique. The final concept that we'll talk about for today is this affinity purification mass spectrometry. This is abbreviated as APMS. APMS here is where you have a tag protein. Here's the B here that I'm circling. This is your bait protein, and for your bait protein you have an antibody. Now that antibody is specific for the bait protein, but when you pull down on that bait protein using this antibody here, you not only get the bait, but you also collect the the proteins that are associated with it that are co-amino isolated at the same time. Now those proteins may be of interest to you because you want to identify those, and they could be either separated on a gel or more likely today you would just digest them directly and go into the mass spectrometer with those and be able to identify those through the protein identification technology that I was just talking about before. And so you can get there and then once you've identified those proteins, you can form a network there that kind of highlights what that bait protein here shown as A here was really potentially associated with in this mixture. Affinity purification is really a workhorse for identifying interaction partners in proteomics. That's the end of the introduction here. I'd like to move on now, and we'll have a little poll that might pop up again that just is asking what type of amino enrichment techniques you might use in mass spectrometry. In the next 20 minutes, I'd like to highlight our recent efforts to create a network of protein-protein interactions that encompasses essentially every protein in the human genome. So why would you want to do this? What's the value in protein-protein interactions? This is just an example that might help you. So tagging an unknown protein, in this case we've tagged this protein here called GatsSl3. GatsSl3 is a protein, if you look it up, of unknown function. There's nothing really known about it, but when we pull down on this protein, it came down with 5 other proteins. And these 5 proteins, what we noticed was that 3 of these, these ones I'm highlighting here, 3 of these proteins were actually part of another complex called the Gator2 complex. So this Gator2 complex is involved in sensing amino acids and relaying that information to mTOR pathways. Knowing that the Gator2 was pulled down with Gatsl3 here, we could implicate Gatsl3 potentially in the processes of Gator2. And if you look at the structure, the gene structure of Gator2, which was renamed Gatsl3, which was renamed to Castor1, what you see is it contains these act domains here. So there are two act domains, dual act domains here, and these act domains in bacteria are known to bind nucleic acids as well as amino acids, and the closest homolog in— to the human Akt1 is right here. It's this aspartate kinase that binds actually a lysine into this domain. So there's some evidence that these Akt domains can already bind amino acids, and so the hypothesis you would generate is that there's some amino acid that's being bound by Gatsl3, and that that is being then relayed to Gator2 and that that ends up working through that process with mTOR. It turned out to be true. This is the paper that was published here where it was figured out that, that Castor actually binds arginine, not lysine this time, very specifically and that it can sense the levels of arginine in the cell. And you can get all that hypothesis and those ideas directly from just, you know, some pulldown information. And so, you know, this is kind of one of the major reasons why you would want to do protein-protein interactions is to kind of generate these new hypotheses. Another reason you might want to generate or to identify proteins and interactors can be illustrated in this slide here, which is fairly complicated, but I think I can highlight that for you. So what I'm showing you here is a pathway from a signaling transduction pathway where you have a receptor, which is named R here, And that receptor is going to cause L and M here to disassociate, and then L is going to bind to some new partners. And at the end, all of this is going to end up that S is going to be— is a transcription factor. It's going to enter the nucleus and it's going to cause the translation of a new protein, which is going to end up in the outer membrane of mitochondria. So this entire pathway here from the receptor to the new protein, this pathway here, you might want to understand that. And if you had an antibody against L here that I've got right here circled, so if you had an antibody against L, you could pull it down. And when you pull down L, you can see the star graph around it. These are all the proteins that you would pull down with it. And if you knew the function of R or of, of lysine or, I mean, of K or J or some of these other proteins, you could learn a lot about what L is doing. And so that would be very interesting to you. But if you look at the structure of that star graph where you have L in the middle and these other interactors around it, that is not, does not resemble what the pathway really looks like. However, if you were to then do more reciprocal pull-downs where you pick M and T and U and N and P and pull down on all of those, and then you would create a network that looks like this. So by performing many immunoprecipitations, we end up with something that now begins to look a little bit more like the pathway. For example, L here is just 2 hops away from S, which is the transcription factor, and you can even see some of the regulatory, underlying regulatory pathways that are present there. And so this is a very, very strong argument, I think, for why one might want to collect an entire map of every human— of every— of the interactions of every human protein. And so that's what we set out to do about 3 years ago. The first paper was published last year, and this includes a network of thousands really of these APMS experiments. So take thousands of human genes, express those proteins, pull those down, and see what's interacting with that and see what we can learn. So this was the first paper from that. So we named it BioPlex. BioPlex stands for biophysical interactions from Orpheum-based complexes. Let me kind of explain a little bit more about what that means. So there's a collection of DNA called the Orpheum 8.1. So this is 13,000 clones, so 13,000 DNA sequences. So think of 130 96-well plates, and each one of these has, you know, 96 genes, DNA sequences on it. And so we wanted to take each one of those and to express those in a cell and pull those down and to identify the interaction partners from those. Our short-term goal was to do this for 2,500, which is about 10% of the human genome. And our long-term goal was to take 10,000 of these and to make a map for that. And I'll show you our progress. So this is the team of people that have been working on this. So there's more than a dozen people that were involved here. In fact, there's a cell culture team here, which really did a massive amount of work in growing those cells up. There's a biochemistry team. There's an instrument— there's an instrumentation person here, Joao Paulo, who looks a little bit tired but has been tireless, I guess, in these mass spectrometry experiments. The informatics and the data analysis was performed by these two people here, by Ramin Rad and by Ed Hutland. And Ed really drove putting together all the work for the first publication here. Lots of the data that I'm showing you here he put together. And of course, Wade Harper was a big part of this as well, as he's a co-PI on the grant along with Matt Soa, who we began this endeavor together. So it took a lot of people to do this, and let me just highlight for you the process that we go through for a single gene so that you can understand how we're going to do this for then thousands of genes. So for a single gene, we take, as I'm showing you here, there's this nice red spot right here, and that is, say, one clone. We take that out, and that gene has the structure that it's a gene, and then the C-terminus of that gene has an HA tag followed by a FLAG tag. So these are epitopes that can be used to pull down. So every gene will have an HA and a FLAG tag at the end of it. We then package this DNA into a virus here, and we infect that into 293 cells. Now, in the 293 cells, we create a stable cell line. Once we have a stable cell line which is expressing the tagged form of that particular gene, we grow up 6 plates. Those 6 plates then are harvested and we do an immunoprecipitation experiment using an antibody against the HA tag, and then we precipitate whatever proteins come down from that, we precipitate those, we add trypsin to digest those, and then we analyze those in duplicate on a mass spectrometer, and this is just an example of that. And so this is the process that we go through for a single gene to identify the interacting proteins. Let me show you what it looks like for the project here. So every single month, we take out about 6 plates shown here. So 6 plates is about 600 baits, 600 clones, and those 600 clones during that month are going to be packaged into DNA, they're going to be made into stable cell lines, they're going to be purified, purified by immunoprecipitation after they've been grown out and shot on the mass spectrometer. And so this all goes through this, this particular process. Now to do this, we're analyzing everything in duplicates. We use 70-minute runs, 40-minute gradients, very quick. We use a mass spectrometer called an Orbitrap mass spectrometer, and we have 3 of these. Quadrupole exactive, QE mass spectrometers that collect both high-resolution MS and high-resolution MS/MS information. And we invert the run order and use different instruments in order to avoid any carryover between the replicates. These instruments run 24 hours a day and they're dedicated to this project, so it's a pretty mass spectrometry-heavy project. So here's the first network that we put together, and this included about 2,500 different experiments. Now, only 2,275 here are shown, and that's because about 300 of these did not have an interactor. So you had to have a single interactor to make it into here, so they just interact either with themselves or with, or, you know, or with nothing. So there were 2,200 proteins, and the entire network is made up of 7,600 genes in total. So this is what it looks like, and that's what we got for the first one. And the important thing here was that there was just a ton of new information here where only about 86% of what we identified was not known before. So the real challenge in these types of experiments is is demonstrated here. So if you have a protein here, and let's say this protein is CDK1, if I pull down on CDK1, it interacts with the proteins that I'm showing you here, all of these different proteins that I'll highlight around here are interactors. But what I really see in the mass spectrometer is this. The mass spectrometer is so sensitive that it can actually see that it can actually see about 500 proteins, because it, there's some background that's in there that's identified, and it's only a single-step purification. And so the real key is how do you go from pulling down 500 proteins, some of which are background, some of which are interactors, how do we go from doing that all the way over to identifying these 16 interactors that are there?. And so these can really be divided. In order to do this, we kind of divide the data into 3 parts. We say that there are bona fide interactors here, there are nonspecific background proteins, which makes up about 95% of what's there, and then there's one other category which had never been considered, I don't think, in protein-protein interactions, and that is that there are actually false positives in there. And it may not seem like a big deal because it's only 1%, the false positives, false positives. So that means that there's only 5 out of these 495 here, only about 5 of these proteins are wrong. But because they're false positives, meaning that they're actually wrong identifications, they can match to any protein. In fact, they can match randomly. And when they match to any protein, they may match to a very rare protein that I've never seen before. And that's dangerous because that would look like an interactor. It would look like it was enriched because it hadn't been seen before, so it was a specific interactor. So, by modeling that in our, in the way that we identify these interactors, we've been able to really remove that problem from the dataset where we don't have these random false positives don't end up in there, and the background can be modeled out because we have thousands and thousands of pull-downs all with the same antibody, so we know the background really well. So, that's kind of how we did it. And what I'll tell you more about that is that we had 9 different parameters, 9 different parameters here. So each one of these parameters could try to tell the difference between those 3 categories. And you'll see one of them up here, you know, might be interesting is the Z-score, of course, is pretty important. So a Z-score would be how enriched among the 2,500 pull-downs that protein that I identified in that one pulldown, how enriched was that over all the other pulldowns that I had ever seen? And what you can see is that, you know, is that these other 9 parameters as well are used, and we use something called a naive Bayes classifier, which we trained using all of the other plates except the one we were using to do the analysis. We take one plate out, we use all the other plates, all of the data from the other plates as training data, and then we create this classifier which spits out probabilities of whether you're an interactor or false positive or your background. And so that's how we created— we moved from here over to here to get this map. So here's the network again. This network again is represented by almost 2,600 different APMS experiments, and it contains almost 23,000, 24,000 interactions. So what can we learn from these maps? And if you just look at them here, here's 3 examples, and what I'm showing you here is everything that we got for this bait. So XRCC2, it pulls down these 3 proteins, 1, 2, 3, So these 3 proteins, that's all it pulls down, and these are all involved in the BCDX2 complex, so it makes sense. And these are responding to DNA damage. And of course over here you see this EIF4E, which pulls down the entire complex. Everything is known. And the SEPT1 protein over here pulls down all of these other SEPT proteins. It pulls down SEPT1 to SEPT11. So, you know, pretty amazing. Pretty amazing coverage on some of these complexes and the way that it works. That was just some good evidence that it really works well. So how do you benchmark how well you're doing? So there's a gold standard in the field called the Corum database. So the Corum database is a carefully valid— contains carefully validated complexes that are often from low throughput experiments. It is about as close as we may come to a gold standard. Here's one quorum complex which is called the Arp23 complex, so this is involved in actin nucleation, and there are 7 known members. You can see the 2 blue, the 2 highlighted in blue here are the ones that were baits, and of course they pull down the other 5 members as part of that. So you get the entire complex is returned here based on just having these these two baits. So in this case, we have 100% coverage across this quorum complex. Here are 6 other examples, I mean, I guess 5 other examples of quorum complexes, and you can see the blue are baits again, and that we were able to pull down basically the entire complex many times, except for this, this one down here where we didn't do as well. But overall, you do pretty well for how you pull down these complexes. And if you look at it in general across all of the complexes, what you see is that if I draw the line here, what you see is that for about 80% coverage, so we found 8 out of 10 members of these complexes more than half of the time. So that's a lot, and we found 100% of these complexes a third of the time. So, you know, there's really tremendous coverage on these well-known complexes that were acquired at low throughput that are annotated in this Quorum database.. And when we don't do well, what you notice is that we often, like right here you can see this additional, the blue ones are the ones that were baits and the reds are the ones that are interactors, but the purple one here was one that was done after that, so it wasn't done in the first 2,500 but it was done at say number 3,000 or something. And you can see that it begins to add in a little bit more of what's happening with this complex. We fully believe that as we go forward, all these complexes are going to be filled in based on just including more and more baits. Once you have this large network, really gigantic network, what can you do with it? I guess we all went to high school, so we all understand what it means to be in a clique. The definition of a clique is that everything interacts with everything else. And the first, the best example maybe of a clique is something like this where you have a 3-member clique. So this is where all 3 members are interacting with each other as you can see. This is a 3-clique. And there is an algorithm called clique percolation. And clique percolation would take that same clique and it would try to add additional cliques onto the sides of that one. As shown here. And by doing that, you can increase the size of that clique. And so this is really a way to kind of cluster or to find communities in a network. So that's what it's trying to do. It's trying to find clusters, find communities within a network. And if you apply this to the Bioplex network, you allow it to go through this clique percolation algorithm, what we found is that there was about 350 different communities. So these 7,000 proteins in this network, they can be filtered back into about 354 communities. Now some of these communities are very large and contain, you know, many, many proteins, but not many. This is way up here. So this is one that had, say, the splicing proteins or RNA binding proteins up here with more than 100 members, a few of those complexes. But the vast majority are down here, and these have less than 10 members. And so you can see that most of these are just small complexes of proteins that are interacting all with each other and that they form communities. And what can that tell you about, about this graph? So here, for example, what you can see is that I'm highlighting this box right here. So this is panel I, as we call it here, and it has 6 different communities that I'm highlighting. There's an orange community, a blue community, light blue community, a green, purple. And if you look down here, you can see it's blown up with the actual protein-protein interactions that are shown within it. And what you see is pretty amazing. What you see is that this network partitioning that's done by this clique percolation, that it really reflects the real organization that's inside the cell, the real architecture. For example, you have the regulatory and the core particle for the proteasome, you have these other proteins that are associated with the proteasome that are nearby, and so you can really get this kind of architecture, this structure that can be pulled out based on these communities and how they're organized with each other. Here's just another example of how that is. If you look here at this number 3, which is right here, It's blown out over here, and then you can see it over here. So this has 3 communities, fairly large communities. This largest community here, the purple one here, this is a— this community here is a large phosphatase community, gigantic. And there's also a kinetochore and an RNA polymerase community here. So what this suggests is that there are common regulatory factors or there are common mechanisms that are involved in regulating these, that probably the phosphatase is is highly involved in regulating these two other communities that can be pulled out together. It's pretty cool. One of the other things you can do with these communities is you can begin to do kind of gene set enrichment and just kind of looking at each community and seeing what's present at higher than you would expect, what different categories are enriched in there. And so if you, you know, if you look here, right here you can see that there's something called cluster 72, that in cluster 72 there's a protein right here in the center called C11orf74. So C11orf74 is a protein of unknown function that nothing's known about, but everything that it interacts with in the network, its whole community, they all are red, and that means that they're all part of intracellular transport. And so you can imply then that likely this protein as well, its function is going to have something to do with cilia. And so this is what you can do, and it's really pretty astounding actually because if you look over here to the cluster 11B and cluster 11A, remarkably what you see is that this is the mediator complex, all of these, but the mediator complex its architecture is actually into two pieces that exist in two different groups in the nucleus. And you see that perfectly manifested here in the self-organized in these clusters. And so you can really use this to help you understand even some kind of structural information about what's going on in the cell. Here's another example that I like here. And let me see if I can highlight it for you. These 3 right here show up right here. So these 3 clusters, we call them cluster 6A, 6B, and 6C. So there are 3 different colors. I don't know if you can see it or not. There's like a purple, a red, and a green. But each of these clusters does the same function. They are histone acetyltransferases. That means they take an acetyl group and they place it onto a lysine. But there's known to be 3 separate complexes. That do this, and that these three have a different function, but they are kind of related to each other in some ways. And you can kind of get that picture from the map here. What you can see is that they segregate completely based on which of the three complexes they're in, as well as some of the interactions that they have between each other. So it's really, really pretty great. So this network also, this enables us to predict do some predictions. And one of the things you could try to do is to predict the, the subcellular localization. So this is what we've done here, and this is the example that I'm showing you on the slide here, is if you look here in the center, you'll see PSMA1. So this is PSMA1. If I start on the network on that big hairball and I look at everything that's one hop away, that's right next to it, it'll be on this circle that I'm going to draw right here. So everything that's one hop away will be on that circle. So there's maybe 15 proteins that are directly associated with PSMA1 in the network. Two hops away in the network is everything on this outer ring. And so what I'm showing you here then in the colors of the balls that are, that are for these interactors is whether it's a cytoplasmic, a nuclear, or both, you know, for these, for the prediction on UniProt. You just look it up on UniProt, it tells you what the localization is. And so we can make predictions based on these for the one in the middle. So for PSMA1, if you come over here on this matrix, you can see that it's predicted to be both cytosolic and nuclear, right here. Cytosolic and nuclear because they're— in fact, every member of the proteasome, as you go down here, every member of the proteasome is predicted to be cytoplasmic and nuclear except right here. So this protein is called PSME1, PSME3. PSME3 is a nuclear-specific, nuclear-specific, um, subunit of the proteasome, and it doesn't show— it shows up only as a nuclear protein here. So again, you can use this approach then to predict for thousands and thousands of proteins, we can predict their localization based on this approach. It's pretty cool. Here's another example where we predict mitochondrial localization. Here it is right here, mitochondrial localization for all of these members of this of these complexes as this is. So just another example there. One of the main reasons that we did this was to try to add function for the first time to proteins that no one's ever studied before. There are thousands and thousands of proteins, thousands of genes, which there's nothing known about them. They just basically have a name. And many of them are just called something like C16ORF82. So they call them C-ORFs because they don't have a function and they're just named that way. Here I'm showing you 271 different C-ORFs, and what you can see here is that there's the first piece of information. The red ball here is showing you that C-ORF, and it's binding to some other protein. For the first time, we have some information about it and what it's doing. It's kind of beautiful, and it kind of shows maybe a little bit of function about many of these proteins for the first time. We can also predict their localization. So for this, in this case, what we were able to do here is show that, look, like all right here, all of these C-ORFs really, I guess, would be predicted to be plasma membrane. So we know something else about these unknown proteins, right? There's some new mitochondrial ones here. There are some proteasome, there's one, I mean, one that's the peroxisome here. And one to the lysosome even, right? So there are new members at new localizations in the cell, and we can get that directly from the network. So there's a lot that you can do with this network to predict things. And you can even take those C-ORFs, as I'm showing you here, like here's one here in the middle, and you can see that, well, let's look at this one up here first. So this is the catenin complex here, catenin A and catenin B. B, and these two were known to interact with each other before, and they sever microtubules, and it was never known that there was another subunit here. So this actually turns out to be a third subunit that can be another regulatory subunit for this complex. And so you get that for these CORFs, you can add some function to those and find out what they do. So finally, there's a viewer that you can download from the website. If you look at the Bio-Plex website, there's a viewer., and in that viewer you can just take like here's Vav1 for example, you can look at Vav1 and find out what was interacting with it, there's often lots more information there that you can try to pick out your favorite protein and try to understand it. So in summary, really we showed that most of the complexes are faithfully depicted, and that most of these interactions that we found had never been found before, so this was pretty amazing. We had 86% were new here, really thousands and thousands of new interactions. And that we could subdivide the network into hundreds of clusters, these communities, and these were united by shared function and localization. Also, I didn't show you this for time purposes, but you could think about doing this. You have the domains on these proteins as well, and you could ask questions to see if any two domains are enriched or are associated more than you would expect by chance, right? So we did that, and for thousands of domains now, we can predict that they're associated with each other for some reason, that they have something to do with each other. And so we've done that domain analysis as well, which is something new. We can also predict subcellular localization, as I showed you. And finally, all these new ideas about novel proteins, this is really the most exciting thing for me was to be able to just see so many unknown proteins for which we now have our first hint about what they're doing and what their function is. So I guess that's the end of this part here before we take some questions. And I would just like to really thank the two sponsors for this project. So this was dual funded. It was funded by the National Human Genome Research Institute, which is part of NIH, And it was also funded by Biogen. So we had a pharmaceutical company as well as government funding to do this because it really took millions of dollars to be able to do this. And our goal is still to finish the entire human genome as much as we can. And on the website, I think there's almost 6,000 or more of these experiments now, whereas the publication that I showed you only covers 2,500. So there's a lot more data there. To be looked at. I'd also just like to thank Cell Signaling Technology for sponsoring this event here and for helping us out with getting this webinar today.

Thank you very much. Okay, that was a terrific presentation, Dr. Gee. Thank you so much for bringing that information to us. Before we get started with the live Q&A session, we'd like to get some feedback from the audience. If you could please give us your thoughts on the poll that's appearing on your screen right now, we'd really appreciate your feedback. And as a quick reminder on how to reach us today, questions can be sent via the green Q&A button at the lower left of the presentation window, and we'll try to get to as many questions as we can. And it looks like we have some great questions already. The first question is, what are the ramifications of using just a single cell line, such as HEK292T, for Bio-Plex?

Yeah, that's a good question. I mean, our goal was to make a single pass through a human cell line. Now, we realize that many of the proteins that we expressed were never meant to be in a 293 cell. So we took olfactory receptors, we took albumin, hemoglobin, things like that. We expressed those in this 293 cell and tried to do their interactors. It was completely unbiased. And so by using a single cell type, we had some advantages that everything was the same and was unbiased, but there's some disadvantages that there may be, the interactors may not be expressed in that cell or other reasons. So that's definitely a caveat to this, and the hope would be to go in and do other cell lines as well as to do other conditions, maybe only during mitosis is some complex formed and we wouldn't discover it using this as well, so I think using a single cell line is a, was kind of a necessary evil here, but I think it still worked out pretty well for us.

Great. Your next question is from Christopher at the Providence Cancer Center, and he asks, what about protein polymorphism? Does the software consider known alleles and SNPs for protein coding genes?

Yeah, I mean, that's a great question. You certainly can consider those. Typically we use what we call a reference genome. And this is like for mouse, it's the B6 black, right, so it's not a different type of mouse, it's a B6 black mouse, so we don't, or for humans it's just the reference genome, we don't really consider the polymorphisms, we could, in the past it hasn't made that big of a difference because, you know, there are multiple peptides per protein and so you're definitely going to miss the one that had the polymorphism, but the peptide that had the polymorphism, but you're not going to miss the protein most likely. So the answer is that the software can. If you want to include that as a database, you could. But for now, we don't.

We just use the reference genome. OK, great. Next, we have a question from Daisel at the University of Cambridge who asks, why are you sure not pulling down more than one protein with the target?

Of an antibody? Let me read that. Why are you not pulling down more than one protein?

I guess, how are you sure that you're not pulling more than one protein?

Oh yeah, how am I sure that I'm not pulling down more than one protein as a target? So that's a great question as well. So in this particular case, what we're using is we use the same tag for all 2,500. So think of it that we have the gene, and at the end of the gene we have a tag and that tag is called the HA tag. There's an antibody, we have only one antibody that recognizes that HA tag. So every pulldown is done with the same antibody. Now that antibody may recognize that tag, it may also recognize something else, but overall we know the background really well for what that antibody recognizes because we use that same antibody for everything. So it's not that we're using native antibodies here. We're not going in and pulling down using an antibody that was made specifically to pull that down. We're using an antibody against a tag which is expressed.

So it's the same antibody for everything. Okay, great. Your next question's a little bit long, so feel free to read along with me. It comes from Michael at Harvard Medical School who asks, since it seems like you're measuring a static condition in one condition, Would you always expect that you would always be in agreement with quorum? Wouldn't quorum try to capture everything across conditions?

Right, so that's a great, that's a great question again too, and so I wouldn't expect to capture everything, one, because 293 cells may not have that complex, we may not have the conditions right as the questioner says there, and so I wouldn't expect it to always agree, but we were kind of surprised by how much it did agree, and of course quorum complexes are much more of the kind of universal complexes like I showed you, the ARB2/3 complex, the APC complex, some of these other complexes, the proteasome, the ribosome complexes that are very much well-defined.

Great. Next we have a question from Maria at Seville University who asks, how can I know theoretically if two proteins interact in the plasmatic membrane with a stimulus?

So I guess the question there is asking if I need to have the stimulus present in order to cause that interaction to occur. So again, you could certainly do the pull— you could tag and do the pull-downs in the presence of that stimulus, and it would allow you to— let me go back to the question here that fell off my screen here. But you certainly could do the pull-down in the presence of the stimulus. And see if they interacted. Now, dealing with plasma membrane proteins, that's a very much another difficult situation there, and it may not work. It may be that you don't have the right detergent to keep the membrane proteins happy, or it may be that that interaction is not stable enough without the exact membrane that you need. In that case, you might have to use something like cross-linking. You could move to chemical cross-linking where you might stabilize through covalent cross-linking those two together and then find them. There's another idea called BioID, which is where somebody, instead of putting an HA tag on the end of the protein, you could put what's called a biotin ligase type thing called Bira, and then you could— that would biotinylate the nearest neighbors instead of pulling them down. You would use biotin to pull them down. So there are ways that you could look at this, and especially if that person wanted to email me, I could provide more information about what I would do in that case.

That was fantastic. Next we have a question from Atsushi at Johns Hopkins Medical School, who asks, have you tried another cell line such as neuronal or U-cell?

We have done some of this work, not a lot, maybe 100 or so, in a cell line called an SY5Y cell line, which is kind of thought to be more of a neuronal type cell line. We've done some in Jurkat cells, which are, you know, obviously derived from immune cells, but we are currently, our next phase of this project is to create a second map in another cell line, and the second cell line that we've chosen is the HCT116 cell line, so this is a colon cancer cell line, which is, you know, one of the issues with 293 cells, of course, is that they have, I don't know, 80 chromosomes or something. So they're not 100% normal, but an ACT116 cell is a bit more normal, I would guess. It's a diploid cell, it's got 46 chromosomes, it looks better. So we'll see how the map, as we create the map again a second time in an ACT116 cell, how that compares. But I don't have, we don't have plans to do it in a neuronal cell, they don't transfect that well, or in an immune cell yet.

Great. We have lots of great questions coming in. And the next one, it comes from Rupert, who says, "Interactome is established from EGK293. Would future interactome studies also include other cell types that are tissue-specific or disease-specific, such as cancer cells?" Right, that's a great question.

And again, like I said, we're going to be— we are making the map again right now in HCT116, which is the cancer cell line, so that will be will be interesting. We have plans to make it in several other cell lines, maybe in an IPS cell, you know, or stem cell. So, there are plans to do this, and of course, environment is important, or conditions are important too, so it would be great to be able to make a map under different stimulations of the cell as well, not just different cell types.

Interesting. Next, we have a question from— 2 things. DuPont, who asked, can weak transient interactions be identified by this method?

Yeah, I mean, that's a pretty good question. I would define a weak or transient interaction as something like a kinase binding to its substrate, where it might be 5 micromolar, might be the binding constant there. So I don't think it can detect those types of things, but I do think it's not really detecting only strong binders. So again, I think the answer to the question is it's kind of in the middle. It doesn't detect weak, but it doesn't only detect strong. I think things that are fairly transient, it can detect, but no, it's not at the stage where it's detecting substrates for a kinase or an enzyme reaction or something like that, where it's a much weaker binding that's happening. These would have to be fairly stable interactions.

They have to survive the washing. Great. Next, we have a question from Rose at the NIAID at NIH who asks, could the non-interactors be interacting with carbohydrates or lipids?

Yeah, that's a great question. So there's 300 proteins that we put through here that didn't act like they were interacting with anything. And it could be. It could be the protein's job is to— if it's an enzyme and it doesn't work with anything else, it doesn't have to interact with anything else. There's no rule that says you have to do that. No, I think that's great, and you know, some of these proteins, again, remember we tagged them at the C-terminus, and if they needed their C-terminus for function, that might be another reason why they wouldn't interact with whatever they were doing, right, because they weren't functional anymore. So there's a couple of reasons why they might not be there, but certainly the one stated by the questioner is right.

Very cool.

Next we have a question from Omar at Muta University who asks, "Is protein interaction different from protein aggregates?" Yeah, I mean, I think, you know, obviously protein aggregates are usually precipitated, they're much more difficult to solubilize, and a protein interaction is just how things are normally happening, the normal physiology in the cell. You have, you know, proteins interacting with each other and working together to do their normal physiological processes. So protein aggregates is more of a kind of a, more towards the disease states.

Okay, next we have a question from Chow Feng at the U of T who asks, can you give more details about how to get rid of false positives?

Right, so for the false positives, the way we did that is it's a bit more technical, so it might be hard to understand, but you have to model those false positives. It's only 1%, but the way we modeled the false positives was is that we had done the search and we had the proteins in the forward direction. We also had the proteins in a reverse direction when we did that, and so it was pretty obvious to us that we had a good training data for things that were completely false positive. Anytime we detected and it passed all of our filters, anytime a protein that was completely bogus that was called a reverse protein ended up in our network, right, and so there were like 10 or 20 of these that made it through. Those proteins are false positives by definition because they're made from fake proteins. So we would use those to train the model and find— to show how those got into there is the same way that the other ones that we don't know which ones they— we don't know about those. So again, we used a known amount of bogus information to help us understand and to be able to predict what those false positives were like and what they were doing.

Very interesting. Next, we have a question from Kirk at the UAMS, who asks, is this method compatible with TMT labeling to compare interactions across multiple conditions?

Yeah, it's a great question. It might be, because the problem you have is that you have different expression levels of each of these proteins. So if I took 10 different baits, and I use TMT, which can measure 10 different interactomes at once, the issue would be that if one of the baits was expressed at 50 times higher level than any of the others, it might be more difficult to detect the interactors for that one, because for the one that was 50 times less. So, you could do it if the baits were all expressed at similar levels, say within 2 to 5 fold of each other in the cell, I think it would work. Work. Otherwise, I think you would rather stick with different conditions but the same bait would probably be a better way to use TMT.

Very interesting. And then we have a question from Sun at King's College of London who asks, have you cross-checked interactions with pathway?

Yeah, so we have done that as well, so we look over different, different, you know, all the different KEGG pathways and all the different websites and different things, we've been overlaying these data. And other people have actually picked up our network and overlaid it on top of networks and pathways as well. I really think it's going to be the best network, certainly the most accurate network that we have to date. So I definitely think you can use it for networks and pathways.

Okay, that was great. The questions were great and the answers were Excellent. I think we have a lot more questions than we have time for today. But if we didn't get to your question, we'll follow up with an email. And I would like to once again thank Dr. Stephen Gigi for his presentation. Do you have any final comments for us?

No, just thank you so much for listening. And I hope you'll use the network.

It's a great thing. Thanks. Fantastic. I would also like to thank our— again, our speaker today for his informative presentation, as well as our sponsor, Cell Signal Technology, for making it possible to bring this presentation to you. Today's webcast will be available for on-demand viewing through September of 2016. You will receive an email from LabRoots alerting you when it's available for on-demand and posted on labroots.com. Please feel free to forward this announcement to any colleagues who weren't able to join us for today's webcast. Thank you for logging in and participating in our broadcast today, and we'll see you next time.