Scientist Stories: Marylyn Ritchie, Integrating Genomics Data

All right. Good evening, everyone. As he said, I'm Marilyn Ritchie, and thank you very much to the organizers for inviting me. It's a pleasure to be here. I've never been to Denmark before, so this has been a lovely visit. So I'm going to talk about different methods and approaches for integrating genomics data. I spent a lot of time thinking about how to do this education session. When I teach at the university on this topic, I often go into a lot of details about the computational modeling tools that were evaluating. I looked at the timing of the session and opted to change it up, and I'm not giving my typical didactic lecture on this, but instead a more conceptual one. It's kind of late in the day. Depending on where you flew from, you might be as jet-lagged as I am, or maybe not quite so much. But I'm gonna touch on the different types of approaches that one might take, and also try to get you thinking about why you would wanna take different strategies. And I'm not going to propose that any one strategy is the right one. So first, to kind of set the stage for the talk, this is the only Manhattan plot you will see in this presentation, this cartoon. So what I would encourage everyone to do is to kind of get away from our typical kind of whole genome genetics type analyses where we're running GWAS and making Manhattan plots or We have an exome and we're running an exWAS and we make a Manhattan plot. We're going to put away the genome-wide one variant at a time approach, which I'll refer to as the hammer. No hammers in this talk. I want us to think about the other ways to look at data. There are a lot of different analysis tools that are out there, different strategies that we can use, not necessarily that are well accepted and published in kind of the human genetics and genomics literature. But a lot of other fields. They've been dealing with data at scale for decades. You know, people in astronomy have been dealing with terabytes of data for 20, 30 years. It's new for us, but they've been doing this for a long time. They're looking for complex models, big data. It's interesting. If you start digging into big data literature, genomics data is actually not very big. Just big for us. And it's becoming big as our technologies get faster and better. But I've been looking at how are other fields dealing with these kind of multi-scale data from multiple time points from lots of people. And many of you may have seen this. I gave this example at a talk at ASHG in an education session, not about data integration, about something else in genetics. But this example has stuck with me, and I've been using it quite a bit. I think it's important in this framework to think through. So how many in the room are familiar with the board game Clue? Or if you're from the UK, you probably call it Cluedo. Okay, so most, but not all. So I'll explain it briefly. So the way that the board game works, there's a board, has different rooms in the house where a crime was committed, a murder. There are suspects. And there are weapons. And so the whole point of the game is that each time you play, a new murder weapon, room, and murderer are selected, kind of put under the table, and then you have to try to figure out whodunit, where did they do it, and with what. I would encourage us, as we're doing studies of common disease or rare disease, to pretend that we're playing a game of Clue each time. That means that We don't know that— so just because in the last game it was Colonel Mustard in the kitchen with the lead pipe, which certainly when I was a kid and played this game, it seemed like it was always Colonel Mustard in the kitchen with a lead pipe. When you talk to other people, they'll say, "No, no, no, it was always Miss Scarlet, and it was in the study with the rope." So it's chance when you're playing that game, which set of cards comes together. But what I think might be happening in human disease, and it's pretty likely, is that it's different for each disorder. And so if your GWAS strategy worked and you found variants for this disorder that you were working on, that doesn't mean that it's going to work for the next disorder and the next disorder and the next disorder. It's a new game of Clue. We have to go back to basics and think about the disease process, look at the data in different ways, not assume that because it worked for one strategy it will work for— or that it did work for one strategy that it automatically will for everything else. So how do I conceptualize this? Well, who is guilty? Is it variation in the DNA or the RNA or the environment or is it protein? I would suspect that for some disorders it's mostly DNA variation. For others it's protein changes. Just because it's protein changes in one disease doesn't mean that it's protein for everything else. What are the weapons? So this is the how is it happening. We have so many papers in the literature that debate Is it common variants or is it rare variants? Is it epistasis? Clearly it's epistasis. Is it G by E or just the environment? It's pathway effects. It's network effects. It's all of the above. It's just which disorder are you looking at, what data do you have, and how are you modeling it? Sure, for some disorders— I've had arguments on Twitter about height. All of the data show that height is a polygenic additive with lots of common variants. You win. I've looked at the data. I agree. It's probably lots of additive common variants. Do I think that that means that for polycystic kidney disease, for schizophrenia, for ovarian cancer, for type 2 diabetes, that it's the same model? Probably not. It's probably something different. And so I think that we should look at the data in different ways. And the last one is where did that omic change occur? We mostly focus on blood. That's easy. Our patients and research participants are very willing to give blood. When you ask for some of these other tissues, they're a little bit less willing. And so we're going to have to kind of rely on blood for most of the time until we do have other tissues available. I think we're learning a lot from projects like the GTEx initiative, or the Genome Tissue Expression initiative, where we're learning which genes the expression is similar across different tissues. Which are not, so we'll know when blood is a good proxy or when skin is a good proxy and when we really do need to get the brain or kidney tissue that's going to be much more difficult to get. Okay, so to set the stage for this— the rest of the talk, we have a lot of variation that we have found with our hammer, right? Lots of GWAS hits. However, we know that biology is complicated. If you think about those GWAS hits, and many of you, including me, have published some of these lollipops on the plot. They explain very small amounts of the effect. And so we've been all debating and thinking and talking about where else do we find the heritability. It's hidden, it's missing, it's not actually there. The reality is we're trying to understand disease, and I don't want to get into the semantics of whether it's missing heritability or what. Biology is complicated. The diseases that we're studying, whether they're rare or common, are predominantly complicated. If you look at— this is one very simple KEGG pathway. Calcium signaling. And you can see, you know, single genes that interact with multiple other genes, with entire pathways. There are feedback loops. There are compensatory mechanisms. There are complementary and pseudogenes. It's complicated. I focused so much on the DNA variation and its complexity for a long time. For about the last 5 years, I've started thinking about the rest of molecular biology. I think Myself, just like many others, we were so focused in our own. So as Philip said, you can put an om on almost anything. Well, I kind of grew up in SNP data. So I was a genetics genomics person, not gene expression genomics, just the genome of DNA variation. And I think I had a lot of colleagues who started out doing microarrays. So they were gene expression. So they kept telling me, I don't know why you're looking at SNPs. All of the functions in gene expression, you really should be doing microarrays. And I had proteomic mass spec friends in grad school, and they were like, I don't know why you guys are wasting your time with DNA and RNA. It's all in the protein. You should be studying proteomics. We now talk to each other quite a bit because we're realizing disease is probably the relationship between DNA and RNA and protein, as well as all of the other omes that Philip talked about. So we published a review paper a little over 2 years ago on methods to integrate data. So we started to really dig through the literature and see, how is everybody else solving this problem? We wanted to be able to take data, and the test data that we worked with that I'll talk about is the Cancer Genome Atlas, or TCGA, where you have data from different cancer tissues. And within those data sets, they collected multiple omics data types. And so we started to think through, well, how would we put these data together? Kind of led to lots of literature searches to see how other people were doing it in our field, how people were doing it in other fields, and hence we wrote this paper to try to categorize how the methods can be classified. So this is similar to one of the slides that Philip showed, but this is trying to get us to kind of get out of our silos where we have the genome with lots of different data representations, the epigenome, the transcriptome, the proteome, the metabolome. Stick an omic on it and you can have a silo of data. We want to be able to put the data together in some sort of multi-omics way. So we took the methods and categorized them into kind of two broad categories, either multi-staged analysis or meta-dimensional analysis. And I'll explain what each of those are. And then I'll give you my biased perspective on which direction I think we should be going. Knowing that probably both directions will be important for different questions and disorders that you're looking at. So a multi-stage analysis is where you're taking one level of the central dogma to try to predict the next level and try to, in a hierarchical or stepwise way, go from DNA to RNA to protein. So we referred to things like eQTLs as being multi-stage. So it's SNPs that predict expression. Also looking at allele-specific expression. Also doing kind of overlap studies with functional units where you might take SNPs and then you see how do they overlap different types of sites in a very linear hierarchical way. And then you, based on those results, look for associations with some phenotype of interest. These examples are showing a binary or case-control type phenotype, but many of these methods are applicable to quantitative traits as well. As well. So the other strategy we refer to as meta-dimensional analysis. And let me first tell you why we called it that. We actually had quite a bit of back and forth with the reviewers of our paper. If you're in the audience, thank you for the heated debate. Thank you eventually for letting us call it this as well. So meta-dimensional is not a word, and that's what they argued. This is not a word in the dictionary, and we were aware of that. We were making up a word to make a point. It's not just multivariate. They wanted us to call it multidimensional because it's multiple dimensions of data that we're putting together. But it's not just multiple dimensions. It's also multivariate. So we want to be able to build multivariate models of multiple dimensions of data. So it's multi-multi. And we kind of looked in Greek and decided that that was meta. And so we called it meta-dimensional. And eventually we won the reviewer over, and the paper was allowed to call it meta-dimensional analysis. The idea is this. It's not just one layer predicting the next layer. It's how do the layers come together in some kind of combined way to explain or predict a phenotype. And even within this strategy, we broke them down into three areas. So one is called concatenation-based integration. And in some ways, this is the simplest to think about. This is where you would take each of the datasets, so perhaps a SNP dataset matrix, a gene expression matrix, and a microRNA data matrix, and you would concatenate them into a mega data matrix. So you'd have patients down the rows and the genomic variables in the columns, and, you know, the first however many would be SNPs, the next would be your expression, and then your microRNAs. And it would make a very large data matrix. Many features to search through. You would use some sort of machine learning algorithm to do that analysis and then look for association with phenotype. And I will show you an example of one of those strategies today. The next would be a transformation-based integration. This is where you would take each of the data types and map them individually to some other space, so do some sort of transformation of the data, and then combine them. I think that the work that Philip talked about I would project is a transformation-based integration. So they analyzed each of the sets. They did those correlation analyses of each— the transcriptome and the proteome and the metabolome— and then put them together in some sort of pathway analysis. So I would say what he presented is a transformation-based integration. The last one is model-based integration. That's where you do some sort of machine learning or computational modeling of each data data types separately. And then you take the resulting models and put them together into another machine learning algorithm to see, can you take the components of these models and put them together? I think in the presentation today I'll be covering at least the first two. I don't think I ended up having time to fit the third one in. But I'll try to remember to tell you if I do. If I did, and I'll explain that that's what that one is. So why is this type of integration important? Or why do I think it's important? I'm going to give you a partial real biology example, and it's partially hypothetical. This example is in the paper, so if you go to the paper, you can read more of the details. I'll try to give the Cliff Notes now. So what I have shown here are three different candidate pathways for breast cancer. So this is a schematic of the estrogen metabolism pathway. These are two DNA repair genes, and this is a schematic of cell cycle. Those are all true biological truths. This pathway also— one other kind of true thing before I tell you my hypothetical is that metabolism of estradiol down this right side of the pathway is believed to be more carcinogenic than the left side of the pathway. This is work that was published by Fritz Parrell at Vanderbilt University back when I was a graduate student. He's a biochemist and he had lots of evidence to say that kind of the byproducts down this side of the pathway were more carcinogenic. So now for the hypothetical. So let's say that you have copy number variants in cytochrome P450 1A1 that lead to less expression of this particular gene. And then you have common variants in cytochrome P450 1B1 that leads to more function of that gene. So right from the start, more estradiol is gonna go down this bad side of the pathway. Then you also are unfortunate enough to have rare variants in catechol-O-methyltransferase and glutathione S-transferases, which are two points in the pathway where those estrogen byproducts are deactivated and taken out of commission. If you have rare variants in both of these that lead to less inactivation of that estradiol, byproduct that will lead to DNA damage. So let's say after that you also have some sort of methylation change in XRCC1 and gene expression change in XRCC3 that lead to less DNA damage repair. So you have more DNA damage happening because it came down the right side of the pathway and did not get inactivated, and then less of it is repaired because you have inactive DNA damage repair. Then to top it off, you have some sort of protein change in cyclin B that leads to upregulation of cell cycle. So what would happen? That means you have more DNA-damaged cells not being repaired, replicating over and over. I mean, that's what cancer is, right? It's our own cells that have some sort of damage that then replicate. So as I said, this is hypothetical, but it seems plausible that it could be variation that is both common SNPs— or CNVs, copy number variants, common SNPs, rare variants, methylation changes, gene expression, and some sort of protein change that lead to something like breast cancer. If you did not model all of these components together, I'm not sure that you would find this effect, because it could be that it's kind of the combined effect of all of these things happening simultaneously that leads to the cancer cells replicating. As I said, it's a hypothetical example. Unfortunately, we don't have a lot of strong biological examples to model things based on. So we've been just kind of thinking about them to try to develop strategies to analyze them. So now I'm going to give you a few examples of where we've applied these meta-dimensional strategies. And it looks like they are working. So first, this is a breast cancer data set from TCGA. We were looking at censored survival data and we were looking at multi-omics data. So the— we did a preprocessing step to allow us to do the survival analysis. I won't go into the details, it's in the paper. The input data that was available was copy number, methylation, gene expression, and protein expression. We had two different outcomes, both survival days, which we did the kind of survival analysis, and then survival status, which was just a yes or no. Did they survive out to a certain time point? So a case-control type outcome. The modeling strategy that we used is called grammatical evolution neural networks. It's in a software package called ATHENA, which is the Analysis Tool for Heritable and Environmental Network Associations. It's a tool that my lab has been working on, oh, for 14 years, I think. Really long time. So it's using a multi-layered neural network strategy. The grammatical evolution part is a computer science strategy that evolves computer programs. So one of the problems with neural networks generally is that you have to set them up at the start. And I'll show you a neural network in a few slides so that you'll see what I'm talking about. But it's lots of nodes and you have to tell the neural network how many input nodes do I want, how many layers or hidden layers should I have, and how are they connected to each other. So you typically start the network with that, and then what neural networks will do is this process to modify the weights, or how much each variable contributes to the model. But typically you start by saying, this is the model, now fit the model. We have no idea which variables should go into the model. We don't know how they're connected. We don't know how big it should be. And so this grammatical evolution strategy will evolve that. It will generate millions of neural network models on a big computing cluster. And then through processes like crossover and mutation, kind of take bits and pieces of one and cross it over with another. So two neural networks get together and then they have a neural network baby and you get to see what the cross product of those neural networks looks like. So it's using theories of evolution and survival of the fittest to create computer programs. There's a lot of literature on this. We've published a lot of papers. I'm happy to send them if anybody's interested. I'm not going to go into more detail on that for now. But it does allow you to do this kind of multi-layered network strategy. So what did we do with the data? Well, first we tried each data type on its own, and the fitness here is 1 minus the mean absolute difference. So higher is better. Each one alone did not perform as well as putting all of the data types together. So each data type alone kind of had okay performance on modeling survival, but they definitely did better when we combined them all together. Gene expression was the closest to the integrative strategy, so there was the most signal coming from gene expression. We definitely got above and beyond that by combining other data types. And the details are in the paper along with the model. Oh, that was concatenation-based integration. So we just took the datasets and made one super dataset, and that was what we fed to the neural network. So next we took a knowledge-driven approach to to look for interactions in ovarian cancer. Again, this is TCGA. Here we were using only one data type, gene expression. This was a proof of concept, kind of adopting sort of some of the strategy that Philip talked about with the pathway type analysis, but we tried to kind of go one kind of sidestep. We mapped the gene expression data to different knowledge sources before doing the modeling. So we took the raw gene expression variables, mapped them to pathways, gene ontology groupings, and protein families. And then we fed those in as the input. So we basically— if, for example, a given pathway had 4 gene expression variables, we did some very simple math. We summed them together and divided by 4. It was overly simple, and I was shocked that it worked. But then we allowed for the analysis on the pathways. The gene ontologies, the protein families, or all 4— or all 3 combined. And this was just to see, can you do some sort of knowledge mapping of the data and then combine the variables? Part of the reason for that is to try to work toward a space where maybe one day we could combine data where the different omics are collected on different people. That's a huge kind of wish list item that you could take TCGA data, and GTEx data and ICGC data and analyze it all together. And right now that's difficult because you might have SNPs and copy number variation on one tissue, and then you have methylation and metabolomics on another set. And so there's not a really good way to put the data together. So we started thinking, if you could map the data to some common ground, then could you combine the data across datasets? And so this was kind of a pilot to see, even with one set of samples, 3 samples and 1 data type, does this work? And we were very pleased to see that it did. So these are 2 different fitness metrics, balanced accuracy and area under the curve. Again, higher is better. So we looked at each— so we did gene expression analysis alone, so the raw data, and that's what's shown in this row. And then we tried to map the data to pathways, and these were KEGG pathways, gene ontology groupings, protein families. And then we combined the 3. This did not include gene expression in raw form. It was just these other three. And in both balanced accuracy and area under the curve, the integrative strategy performed much better than any of the individual strategies. The closest was definitely the KEGG pathways, but the integrated strategy looked to work much better. So I would say this is a transformation-based integration. We transformed the data to something else and then did the modeling. OK, so the third real data example. This is a paper that came out in JAMIA right on New Year's Eve of 2016. And this was actually discussed by Russ Altman. He does a year in review at a bioinformatics— or the Translational Biomedical Informatics Conference on the West Coast. It's in San Francisco every year. And he selects papers that— are of note in different topic areas. And this paper was the one that he selected for data integration strategies. So in this one, we were looking at meta-dimensional models. So it's multi-omics again in ovarian cancer, but we wanted to use that knowledge-driven strategy, so mapping the data first before we integrate. So here we're using genome, epigenome, and transcriptome data. There is proteome data available, but we chose not to use it for this particular study. This schematic, it's from the paper just to show the concept, but the basic idea is that we map each data type to pathways, and then we put the pathways in as the input for the modeling. Again, it was survival analysis. Here we did early versus late, low-grade versus high-grade, and then short versus long-term survival. We also used the neural network strategy. For this as well. So I said we use the copy number data, methylation, and gene expression, and have these three different outcomes. If you look within each one— so I have them highlighted, they might be a little hard to read— but we have the single data analysis, the knowledge-driven analysis, and then the integrated. And in all cases, the integrated strategy worked better than the single data type on its own or the mapped data type on its own. It seems— and so in this example, we only mapped to pathways. So we didn't do the pathway gene ontology. Because we saw the KEGG pathways looked to be the best, we tried that strategy. Kind of now what we're trying is, can you map them to multiple things? We're trying to kind of add one variable at a time to the method so that if it doesn't work, we can try to figure out which aspect of it didn't work. So in all three, the integrated strategy worked well. As well. And this is an example of one of the network models. So this is what a neural network looks like. So the different colors are indicating the different data types. So the green is gene expression. The, well, blue or gray, I guess depending which screen you're looking at, is the methylation. And then the yellow are copy number variants. And so the way that you would interpret one of these models— so people always say that neural networks are not interpretable, so I'm going to show you that I think they're very interpretable. So this is a gene expression variable multiplied by a weight of 0.46. Methylation has a weight. Each one of these has a weight that they get multiplied by. These up here come together in a multiplication function. So you multiply this one times this one times this one times this one, and that is what goes in here. Then you have these two gene expressions that are— one is subtracted from the other here. Then that mathematical product and that mathematical product This one is subtracted from that one with a weight attached to each one. So I could write this out in one equation of math. And I didn't because it's the end of the day and nobody wants to see math. But that's how simple these models are. This is a two-layer network that has input variables from three of the data types. So a little bit more education material. So we are confident that these models look to be working. There are tons of simulation studies and they work well in simulation, but the beauty and the flaw of simulations is that you make up what you want it to find and then you see if it can find it. And if it doesn't find it, you tweak it until it finds it. It's not very satisfying. Are the models that we're finding in TCGA real? I'm not sure. We now need to see how do you extend those or try to replicate those in other datasets. So that's something that we're starting to look into. But thinking about network models. So if you are not familiar with networks, but you wanna do more networks. So this particular paper, the gear was about epistasis, which is not at all what we're talking about. But the way that they explain networks in this paper is so crystal clear, I think that your grandmother could understand a network if she read this paper. Full disclosure, this is open peer review, and I was one of the reviewers of this paper. And so I did really like the paper. But I'm gonna sum it up in one image, and I am assuming if you've never done networks you will understand it. The way that they explain it is that you could have Kevin Bacon nodes or Marlon Brando nodes. A Kevin Bacon node, who's heard of 6 degrees of separation from Kevin Bacon? Lots of people. So almost anyone, if you just start talking to people, you will find someone in your circle of, kind of within 6 connections, that knows or has met Kevin Bacon. Why? Because he is in tons and tons of movies, TVs, commercials, ads. He's everywhere. Rarely is Kevin Bacon the leading man. He's always the supporting character. He pairs up with lots of people, but he's never the driver of, the movie or the show. Whereas Marlon Brando is the leading actor, period. Never is he the supporting actor. He's always the leading actor. There are hub nodes and spoke nodes. Marlon Brando, he, he's it. He has his own small network and he's the only thing you need to get function from that network. Kevin Bacon is connected to tons of things, so he's a hub in a tons of spokes. So an effect on Kevin Bacon is gonna affect lots of other things. An effect on Marlon Brando, it's gonna be a big one, but it's only gonna affect a few small things. They also talk about Twitter and how, kind of the difference between Lady Gaga's kind of billions of followers, so she's another hub in a spoke network, and then somebody else that I don't recall right now. But it's a great example of learning about networks. Okay. A few thoughts to leave you with. So one of the things that any kind of naive viewpoint in coming into this type of analysis would think about, well, can we just evaluate all the combinations of variables? Because one of the shortcomings of machine learning is that most of the methods are stochastic. What does that mean? That means that we don't evaluate all of the models. We kind of start with a set of models and we tweak them and then we see if we find the best solution. But they don't— we don't evaluate every single combination of variables in our dataset. Why don't we do that? Because it's— we don't have enough time in the world to do that. So I'm going to show you just a SNP example. So pretend we only had SNP data, which here we're talking about multiomics. So you have all the SNP data and the copy number variant, the gene expression, the metabolomic, et cetera. So let's say you were going with— and this is an old slide. I will update in 2 clicks. But this has been a problem since the early days when we said, hey, wouldn't it be cool if we could have 500,000 SNPs and do a project like the HapMap? That would be awesome if we could have that. For perspective for the students, this is what people talked about when I was a grad student. My thesis project was on 25 SNPs, and that was a big data set. And so there was this pie-in-the-sky idea that like, oh my gosh, one day we could have a HapMap. And look at 500,000 SNPs. Wouldn't that be amazing? That would be amazing. It will never happen. And it happened soon after that conversation. But I was working on a method to exhaustively look at combinations of SNPs. And so my mentor said, well, if and when we get this HapMap data, would you be able to do those analyses? I think so. And he's like, OK, can you go check how long it will take? So 500,000 SNPs, then if you do all pairs of SNPs, it's 10 to the 11th. 3-way combinations, it's 2 times 10 to the 16th, and so on. So 5 SNPs at a time from 500,000 is 10 to the 26th models. I'm like, oh, that's probably doable, right? He was like, well, I don't know. How fast does your algorithm run? Well, it takes about 1 second per model. And so it was going to take 10 to the 18th years to run the HapMap back in 500,000 SNP world. Don't we wish we only had 500,000 SNPs? Now we have at least 5 million. And if you've imputed to the HRC, you've got 17 million good ones. That is 10 to the 20th days. This is assuming that we do 1 million tests per second now, because our computers have gotten a lot faster since 1999. For a frame of reference, because that's just a ridiculous amount of time, the Big Bang theory was 10 to the 12th days ago. So we are talking longer than the universe existed to do this. So if there's a computer scientist in here, I know GPUs, quantum computing, we are getting faster. As we're getting faster, we're making you more data. So we can't exhaustively look at everything. We have to be smarter. We either need to filter the data with statistics or biology or something. There's no way that we can look at everything. The types of algorithms that we can use. So there are two kind of classes of methods. So one is hill climbing and one is kind of stochastic. Hill climbing methods, you start with a model. And statisticians like this approach because you start with kind of a model that you know. And so a hypothesis. We think that these variables are important. And then we gradually add to the model and subtract from the model. And kind of like forward and backward selection type approaches. This is what back propagation does in neural networks. You add and you subtract variables or you change the order of the variables and you just take kind of stepwise approaches. Stochastic, you just throw tons of models out and then you kind of randomly make changes to them or maybe with some sort of method to your madness make changes to them. You try to cover a lot of the search space much more randomly or stochastically. So I would argue that the hill-climbing methods aren't going to work in this multi-omics space. And the reason is very simple. What we're working with is a simple fitness landscape. So this is an image of Mount Fuji. Normally when I teach this in a class, I have everybody close their eyes, but based on the time of day, I'm not asking anyone to close their eyes. So pretend that your eyes were closed and that— so that you could imagine having a blindfold on. That's the point. So if you were a blindfolded mountain climber that got dropped out of a helicopter onto this hill, First, I would say be very careful. That's a pretty dangerous activity. But if you landed safely, if you had a blindfold on and you were trying to find the peak of the mountain, that was your goal or your objective, if you took a step forward and you felt yourself going up, you would know that you were going in the right direction. If you took a step forward and you went down, you would know it was the wrong direction. And so even with a blindfold on, you could kind of keep turning and stepping until you get to a point that every direction in the circle that you step, you go down. That means you have found the peak of that mountain. So if what we're working with in biology looks like this, you could land on the mountain and take steps kind of up and down until you get to a peak where anywhere you step you would go down. And if you took your blindfold off, you would see, lo and behold, I've reached the peak of the mountain. And this is what a hill-climbing algorithm will do. It won't work well if the landscape looks like Waimea Canyon. Or any other large canyon. This one's in Hawaii, which is my favorite place in the world, and that's why I show this one. There are lots of peaks. Well, how do we translate this to common or rare disease? It is likely that there are multiple combinations of genes and proteins that explain the disease. It's probably not one model. It's heterogeneous. It's probably, you know, in this set of patients it'll be this set of genes and proteins. In this set of genes it'll be the environment plus this methylation plus this gene expression. If we don't use methods that allow us to search a lot more of the space, I worry that we'll miss things. So lastly, I'm sure you're thinking that, gosh, there are so many limitations to what I've talked about, because there are. So context is important. As I mentioned early on, blood is the tissue of choice because it's easy. We know that the effects are tissue and time dependent. When we are collecting data, We don't typically do broad unbiased sample collection of all tissues and time points because it's extremely expensive. We collect the tissues only in cases. We rarely have good control samples of tissues. It's usually at one time point, and as Philip talked about, we're definitely learning that you need these variables at multiple time points because they change, they're dynamic. So that's an issue. The data that we have are incomplete. Even as we get better technologies, we're filling in more of the gaps of the genome and the transcriptome and the methylome, but we are not yet at the point that we have all of the data points. The biological knowledge that we're working with is incomplete. So if you are doing some sort of pathway analysis, I get critiques all the time that KEGG pathways, they're not right. Like there are wrong things in them. I know that. And they're incomplete. They're not completely filled in yet. And the network topology of those pathways is wrong or incomplete. So as we learn more biology, we find out that, you know, we thought this node was connected to this node, but it turns out this is connected to this, which is connected to that. And so that connection goes away. We know all that. But I would argue or ask the question, do we need to know all of these things in order to make inferences about biology and actually move the field forward? Obviously, I think the answer is no, because we're doing that. So I'm gonna give you a very simple example. So based on this image, does anyone know what they're looking at? You can just shout it out if you think you know what it is. I hear whispers. All right, how about if I turn the image? Does that help at all? The sea. You got it. From that very lossy data, so very compressed, compressed, you can hardly see the data points. He can tell that was the C. Now we couldn't tell here, but once we turned it, he could tell it was the C. So as we add more data points, things will become more clear. We will start to get things into view and get the full picture of what we're trying to look at. But from a very, very limited, with lots of missing data view, he was able to tell it was the C. That's what I would argue is happening with our analyses. We're not getting it all right. We're not finding every gene and protein that's important. We're probably missing some of the connections. But we're probably finding some things that are really there. And it's just a matter of more data and more time, and then we'll get to the point of having all of it. And this is my happy place. So when people say you're in a stressful situation, you're about to get your blood drawn, and they say just close your eyes and go to your happy place. This is mine. I take a picture of it every year at the conference that I go to. So to close, I just want to acknowledge the members of my team. This is my lab. It's a group of biologists and statisticians and computer scientists. It takes kind of thinking about these things in many different ways and challenging each other quite a bit to come up with some of the wacky things that we're doing. Some of them work, some of them don't, but they stick with me and try it anyway, even when I say, "This probably won't work, but please try it." 'Cause I think that even all of those failures is allowing us to learn more and move the field forward. So with that, I'll close, and if there's time, I'll take questions. Thank you very much.