Scientist Stories: Tiffany Amariuta, Causal cell types in human disease and complex traits

Thank you again so much for the invitation to present today at the MPG Primer. I'm very excited to talk about some of my work with the Price lab during my postdoctoral work and to hopefully give you some insight as to how the method works. And if you would like to use it, sort of go through a short tutorial. So today specifically, I'll be talking about our method to infer causal cell types driving human disease and complex traits. So the outline of my talk is that I'll start with a bit of background and then go into the method and simulations and finally show results on real traits. So starting with some background. So today I'd like to convince all of you that inferring causal tissues and cell types for human diseases is an important goal. And today I'll be using the words tissues and cell types pretty interchangeably. So let's imagine we have a disease variant that we identify from a GWAS. Most of the times this variant does not even lie within an exonic region of a gene. And so it might be very hard to link this variant to a relevant gene. But let's say it is nearby to a gene. We still have to ask the question, what is the effector tissue of the disease effect? This is not a very straightforward question, but is really, really critical to understanding the underlying genetic mechanism. So, for example, knowing the causal disease tissue is crucial to, one, describing the entirety of the biological mechanism of disease, two, selecting the cellular context in which to perform follow-up experiments, such as in an in vitro or in vivo setting, and then finally determining the physiological target of eventual pharmaceuticals. So disease-associated cell types can be identified by integrating GWAS with functional genomics data. So often in GWAS, we identify these peaks of association. And typically to understand the underlying functional mechanism of these genetic associations, we intersect them, their positions with different types of functional annotations, such as those pertaining to gene expression regulation. So this might involve annotations of chromatin accessibility, which is pertains to the 3D, 3-dimensional organization of DNA and relates to the either activation or suppression of gene regulation. And then maybe even more specifically transcription factor binding annotations. So for example, we might see that a particular variant resides within a motif sequence for a transcription factor. And that might lead us to hypothesize that the variant might increase or decrease the binding affinity of a particular transcription factor leading to some change on gene expression. However, again, the inference of which tissue or which cell type is the effector of this genetic variation is a very tricky question to ask. And this is because there is high co-regulation across tissues. And co-regulation is essentially the genetic regulation of gene expression. So here I'm showing a figure from the GTEx paper from 2020 demonstrating that there is a very high correlation of cis-regulatory effects across all GTEx tissues. And of course within modules of physiologically similar tissues such as brain tissues, we see even higher correlation, which is unsurprising. And so just if we were to compare, for example, the empirical correlation, which would be the correlation of the estimated eQTL effects, this is a very high correlation of 0.82. And then methods that have been devised to estimate rather the correlation of true effects sort of gives a more accurate estimate of what this would be, which is sort of set at 0.75. And this correlation structure will come in handy when we're developing our method and trying to devise a simulation framework to assess how we can sort of parse this information across tissues. So here I want to start by providing a cartoon demonstration of just how complicated it can be to understand what the co-regulation structure across tissues at a particular locus at a disease-associated variant. So let's imagine this red star indicates a variant found by a GWAS. Now let's say this variant co-localizes with a variant, or is in fact the same variant that has an eQTL effect with a gene in brain tissue. So just to, for background sake, an eQTL or expression quantitative trait locus is simply a SNP gene pair where there is a linear association between gene expression and genotype at that SNP. Now let's imagine due to an oracle, we know that this eQTL or this genetic effect on gene expression specifically in brain tissue is related or is causing the disease. Now let's say the exact same variant is also an eQTL for the same gene in heart tissue. However, again, because an oracle has told us this information, we know that this has no effect on disease. Now let's imagine a third tissue where instead there is an eQTL that is just in the same locus and so is in LD with the original GWAS variant. So I'm just sort of setting up a situation where without the true information from the oracle, it would be very, very hard to distinguish which tissue here is the relevant tissue for disease. And so these 2 examples of co-regulation that I've demonstrated is that one, the exact same variant can be an eQTL with a gene in 2 tissues, or variants in linkage can be eQTLs. For the same gene across tissues. So the problem that I'm formulating here is analogous to the need for statistical fine mapping in GWAS. So often in GWAS, we get these peaks of association due to LD where the association strength is proportional to the total LD that the variant is being tagged by. And so statistical fine mapping allows us to estimate posterior probabilities on which variant is most likely responsible for the signal. Now, if we talk about this in the tissue space, what we essentially want to do is create a method that can identify the most likely causal tissue or putatively causal tissue among many tagging tissues. And of course, I want to use the term causal very carefully in this presentation because we are talking about statistical inference. And of course, any method that sees A particular dataset cannot make the inference that it is in fact a missing piece of data, such as a tissue or a SNP that is the causal factor. So here in this cartoon, I'm showing a piece of data from the Finucane et al. 2018 Nature Genetics paper, which is about using specifically expressed genes and the enrichment of SNP heritability around those genes to associate tissues with disease. So here essentially we see for BMI that there are many central nervous system tissues associated with BMI. Now essentially what I want to do with this method is to be able to pinpoint which single tissue or small number of tissues out of this peak is responsible for the association with BMI. So effectively what we're going to do is we're going to leverage the information that we get from transcriptome-wide association studies. And so with TWAS, effectively what we get is a way to associate genes with disease, and we can do this across tissues. So this is going to be a fundamental unit in our model. So I wanna begin by just briefly describing what a TWAS is and how exactly it does perform this polygenic colocalization of genetic effects on gene expression and genetic effects on disease. So the first step of a TWAS is to learn SNP gene weights from an eQTL model. So let's say we have a cohort such as GTEx. We will perform a multivariate regression of gene expression on cis SNPs, usually with some amount of regularization such as LASSO. We get these estimated effects, which again, I wanna stress are just a bit different from when you would do a typical eQTL analysis, which is a univariate regression of gene expression on the genotypes at a single SNP. In the second step of TWAS, we will take these estimated weights from the multivariate regression and supply a new genotype matrix, let's say from a reference panel. For example, in our method, we will use 1,000 genomes. Or in this cartoon example here, I'm actually talking about the GWAS cohort, but obviously for reasons of obtaining genotypes from GWAS cohorts, which is very difficult, this this exact implementation or demonstration of TWAS is not the one we will use. We will be using the summary stats-based version. But effectively what you would do is predict gene expression into this new cohort of individuals. And now we have a quantity that we refer to as genetically predicted gene expression. And so the last step is to compute a correlation of this genetically predicted gene expression which is an estimate of this genetic component of gene expression with disease status. And so as I alluded to, there is also a way to do this just using summary statistics, which effectively weights the SNP gene estimated effects by the Z-scores from the GWAS and takes into account the LD in the locus. So I wanna again establish some I guess like understanding for just how tricky this problem can be of co-regulation across tissues. So here I'm providing a very well-known example of the SORT1 gene, which we know regulates cholesterol specifically in the liver. So if you were to perform a TWAS on SORT1 across different tissues from GTEx, I show the results here. And the striking observation is that liver tissue is not the tissue with the largest z-score. And so here is just a really quick example of how it is not always glaringly obvious from a TWAS which is the most relevant tissue. All right, so more specifically and more quantitatively, what we found in our study is that TWAS association statistics are in fact proportional to the amount of tagged causal effects that arise due to co-regulation. And so here each dot represents a bin of genes and I've been them based on their co-regulation, their total co-regulation across genes and tissues. And I'll get into this a little bit more in detail in the next slides about just exactly how I compute or estimate co-regulation. But we can think of this for the time being as a quantity that's analogous to LD scores, but at the level of gene effects. So here what I'm demonstrating is that there is this relationship between the strength of the TWAS chi-squared and the co-regulation of the gene across tissues. And that is because we know and others have identified that TWAS statistics do indeed include direct causal effects of genetic effects on genes, on disease, as well as tagging effects of co-regulated genes in a locus and co-regulated tissues. So again, thinking back to that cartoon example of the brain, heart, and muscle tissues and how their eQTLs were either shared or in LD were examples of these co-regulated genes and tissues. All right. So now I'm going to transition into the method and simulation section. So I might just like take a quick break to ask if there are any questions on the background.

There's none in the chat right now, but I was wondering which— I guess, are there particular examples other than the one you've chosen that you think have of biology that's been best elucidated with these kind of methods?

So are you referring specifically to like findings from TWAS that have been—

Exactly.

Okay. So the method that we're presenting today is kind of like the first one to sort of try to partition this space. So yeah, I think when I get into the results on real traits, we'll be able to see like, oh, there are actually some examples of potentially counterintuitive identification of tissues that might be involved or simply pinpointing of tissues and the identification of other tissues as just being tagging. Perfect.

So it needed further innovation from this initial setup.

Yes, exactly. So before we were dealing with just association statistics and there was really no quantitative way to attribute the signal to either causal or tagging components. And again, putatively causal is what I'm going to I'm gonna refer to this term as.

Thanks so much.

Yep. All right, so moving on to methods and simulations. So as I promised, I'm going to go a bit more into detail about just exactly how I am estimating co-regulation because now we've identified that this is an important quantity, right, that drives up our TWAS chi-square statistics. And if we are actually going to try to determine which tissue is leading to the chi-square statistic profile that we have, we need to accurately estimate co-regulation. Just like we would need to accurately estimate LD scores if we were doing the same thing with GWAS statistics. So here I'm proposing that these coregulation scores essentially can be estimated using gene expression prediction models and a reference panel. So 1,000 genomes, for example. So the first step again is the same first step as TWAS. We are estimating these SNP-gene estimated effects in a multivariate regression. And now we are predicting gene expression into a reference panel. So here we're using 1,000 genomes. And now the way in which we sum up effectively the coregulation score across genes and tissues works on a gene-tissue pair basis. So the focal gene-tissue pair, GT, exists in a particular locus. And of course, there will be other genes in that locus that could potentially share eQTLs. So we want to make sure that we include information about the genetically predicted gene expression of those genes as well and all of the other tissues that exist in our model. And just for the sake of understanding this example, all other tissues, you could think about all tissues in the GTEx dataset. So here what we do is we use the genetically predicted gene expression. And actually, I would call this w-hat because it's estimated rather than being the true genetic component of gene expression. But effectively what we want to do is take the correlation of these values, W-hat, across different genes and tissues with the focal gene G tissue T pair. So we would sum up what the correlation of the sentinel pair is with each gene in the locus with each other tissue T prime. And this also includes the focal gene. And we do include some bias correction that we can talk about later when we're in the simulation section. Alright. So here is effectively the model or the expectation of the TWAS chi-square statistics that we have derived from the principles of effectively genetic effects on, or like starting with the linear model of additive genetic effects on phenotype, we have derived this TWAS chi-squared model as a function of these co-regulation scores that I talked about on the previous slide. And this quantity H squared G E T prime, which is going to be the thing that we are trying to estimate with our method. And this is going to be effectively our causal parameter that allows us to distinguish the contribution across different tissues. So TCSC is the method, tissue co-regulation score regression, and it estimates tissue-specific contributions to disease. So again, just more in more detail, I'll go through these quantities. So the chi-square statistic on the left is represented as a vector of gene tissue associations computed from TWAS. For a specific disease. N is the GWAS sample size. And now in this summation over all tissues T prime. So again, for example, think about having all tissues in the GTEx database. We'll take a sum of the coregulation scores that we've computed on the previous slide. So this notation refers to the score, the total coregulation of gene G in tissue T. So that's the focal pair. With every other tissue T prime. And then we— this gets weighted by what the sort of causal contribution of tissue T prime is. And because this quantity is a— like, it's an aggregate quantity across the entire tissue, the Per gene quantity would be this quantity divided by the total number of genes in tissue T prime. And here specifically we're referring to the number of cis-heritable genes, which means which genes actually had this eQTL model that in the previous slide was actually able to predict gene expression to some degree. And I can talk a little bit more about the details if there's questions about that. All right, so just to give more intuition, our method specifically determines that a tissue T prime causally contributes to disease if genes with high co-regulation to this tissue T prime have higher TWAS chi-square statistics, whereas genes with lower co-regulation to the tissue T prime would have lower TWAS chi-square statistics. So this is sort of the pattern that we're looking for. In our dataset. And we can actually generalize our model above to partition the TWAS chi-squared statistics across pairs of diseases. So we can identify tissue-specific contributions to the covariance of two diseases or two traits by regressing products of TWAS z-scores on coregulation scores. So the only thing that would change is the expectation becomes an expectation of products of Z-scores. And then this estimated term over here becomes an estimated covariance instead of an estimated heritability. All right, so on the next slide, I just want to again provide some visualization to really illustrate what is happening in our regression in TCSC. So like I described on the previous slide, we have this vector of TWAS chi-squared statistics. So this is across all gene tissue pairs, specifically the pairs where the gene was identified as cis-heritable in the tissue using GCTA. So this vector can be very long in practice. This can be hundreds of thousands of quantities. The next big term is this coregulation score matrix. So this is really setting up the multivariate On the columns, we have all of the tissues in our dataset, such as GTEx. And on the rows, it is the corresponding to all of the gene tissue pairs that are in this vector on the left. And each cell in the matrix is the coregulation score of that particular gene tissue pair with that particular tissue, which we refer to as T prime. And so T prime, we call it T prime just so that it is distinguishable from the tissue's T that corresponds to this vector on the left. And then the effect sizes that we are estimating, that we're, you know, our causal parameters that we're interested in using to tease apart tissues, there will be as many as there are columns. So again, for each tissue in a dataset such as GTEx, we will estimate this H squared, this should say H squared GE. And this is what we learn in the regression. So hopefully this just gives a little bit more of a visualization as to what is going on in the regression. So typically we would estimate, let's say 50 quantities of these based on the size of GTEx. Okay, so now moving into simulation results, we've actually used the co-regulation structure that I showed in the beginning of the presentation across GTEx tissues to give us a basis for the covariance that you would expect across tissues in terms of genetic effects on gene expression. And so we've tried our best to make these simulations as realistic as possible. They are based on real genes and real LD between variants modeled from 1,000 genomes. So the The, I'm here showing 3 very important quantities to evaluate the performance of TCSE: power, calibration or type 1 error, and the bias on the estimate of the H2GET parameter in both the causal and tagging tissues. So for these simulations, we've used the Mancuso Lab t-WAS simulator as our primary way of generating data. And we've used 10 tissues where in these primary simulations, one is causal. So effectively for power, TCSC is moderately powerful and you can see at average GTEx tissue sample size, which hovers around 300, this power is modest and around 30%. So just kind of demonstrates that it is not a really simple task to tease apart tissues when there is such a high amount of co-regulation. This power would obviously increase as tissues become more independent from one another. And so in our simulation framework, it's actually a very difficult task because we don't really model specific— specifically regulated genes across tissues. That are also true components of the disease architecture. And so all of the eQTLs that are in our simulations are indeed shared or correlated to some degree across tissues. So we are really testing the ability of this strategy to disentangle causal from tagging effects. For Type I error, we're happy to see that it does hover around 5%. It is definitely significantly greater than that, but not really and it's not very much a cause for concern. And later I will show comparisons to other methods that aim to identify tissues that are associated with human complex traits and polygenic diseases that have extremely large type I errors. So this method does offer an improved solution to what is currently out there. And then lastly, for bias, we've simulated a trait where 10 of the heritability of the genetic effects on gene expression that are related to the trait is 10%. And so we're quite accurate here with our estimates across different eQTL sample sizes. And the estimate for non-causal tissues, we hoped to hover around 0, but in fact it is significantly non-zero. But in In the long run, it doesn't seem like it affects much the Type I error because it still is quite well calibrated. Just giving more insight into simulations, talking about power specifically, even though the power of tCSC we demonstrated to be quite modest in primary simulations, there are indeed parameters that you can toggle that help you understand what would increase power in a real setting and what would decrease power. So the two parameter settings that I want to highlight here, of which there are many parameters to play with, and that can be checked out in our preprint. But essentially, the number of tagging tissues has a large effect on power. So for the most part, adding tagging tissues makes it more difficult to distinguish which is indeed the causal tissue. And that's because they all have this really high amount of co-regulation modeled off of empirical results from GTEx. So this is not surprising, right? If you had a model with, let's say, only 2 tagging tissues, it would be much easier to identify the true causal tissue. Likewise, in terms of the number of tagging genes, if there are fewer tagging genes in the causal tissue, it is easier to estimate or to identify which is the causal tissue because all of the T the GWAS chi-square statistics that we have for the causal tissue are coming from a causal gene, or at least I'm showing the continuous effect of that ratio increasing. So with the purple bars, we have the largest number of true causal genes in the causal tissue with the smallest number of tagging genes. And now another thing I want to mention that we identified in simulations that is really important when considering applying tCSE to real trait analysis and how to process this data, for example, from GTEx. We've identified that the type I error is near 5% and approximately well calibrated only if the tissues have equal sample size. When running the eQTL multivariate regression. And that is because having larger sample sizes at tagging tissues will make the estimated eQTL effects more accurate and will have an effect on the model such that the model believes that there's more causal signal that can be attributed to those tissues. So it's not a surprising result, but is an important consideration for processing data in real life when applying this method. So in practice, what we do is we try to downsample tissues to have the same number across a lot of tissues without having to throw away a lot of tissues that just have really small sample size and try to meet in the middle. So in practice, what we use is a cutoff of 320 individuals, which is, like I mentioned earlier, around the average sample size of GTEx. So as promised, I wanted to show a comparison of TCSC to other tissue trait association methods. So on the right, I have a list of a few that have been published over the last few years. In the interest of time, I will just quickly mention what these methods are. So, RolyPoly, RTCColoc, LDSC-Seq, and CoConet. I want to mention the two, or like the one defining feature that differentiates two classes of these methods. One is methods that utilize eQTL data. So TCSE utilizes eQTL data and RTCColoc does as well. The other three methods utilize gene, like total gene expression rather than genetic effects on gene expression. So they will instead utilize either specifically expressed genes or the co-expression networks of genes and then incorporate the GWAS data for the genetic effects on disease that are nearby those relevant genes for different tissues. And so on the left, what I'm showing is the ROC curve for the different methods. And we were very happy to see that tCSC has a superior AUC compared to the other methods. So in the next slide, I'm actually showing, I think the reason that's driving the particularly lower AUROC of the other methods is the very high false positive rate. So I wanna return to the statement I made where yes, I identified that the type I error of TCSE sometimes does rise above 5%, but it is very well calibrated compared to other methods that try to identify trait-associated tissues. That being said, these methods never claim to try to identify putatively causal tissues. So it is a bit of a different test because these are— those are association tests rather than trying to apply causal inference strategies or partition association signal in a multivariate model. But in essence, I think we can attribute the high performance of tCSC to its specificity and ability to really identify by tagging tissues from causal tissues. So now I want to move into the last section, which is results on real traits, which will actually culminate with a short tutorial of how TCSE works and how you could go about downloading it from the GitHub and running it on new data. So what we've done in terms of real trait analysis is apply TCSE to gene expression data from 48 GTEx tissues. And 78 GWAS summary statistics with a very large sample size. So in total, TCSC was able to identify 21 causal tissue disease pairs at 5% FDR. We note that TCSC was also able to increase the specificity of known tissue disease associations. So what I mean by that is in GTEx, there are categories of tissues such as adipose tissues, and there will be like 2 specific ones within that category. Or artery tissues. And then there will be, for example, 3 tissues within that category. And while previous methods or the previous understanding, even in the literature of some tissue-specific biology, had implicated these groups of tissues, TCSC is now able to pinpoint one within the umbrella. And then lastly, TCSC finds 17 causal tissue trait covariance pairs at 5% FDR. So this is the cross-trait analysis. So here's what a typical output of tCSC looks like. So here on the Y-axis, I'm showing traits, and on the X-axis, I'm showing different tissues. And the quantity shown here is the H squared G-E-T that I've been talking about in the past, divided by the common SNP heritability. And the reason why we divide that quantity by the common SNP heritability is just to make it a bit more comparable across traits. So here, full boxes indicate an FDR of 5%, and then smaller boxes represent an FDR between 5% and 10%. So on the next slide, I'll get into some examples. But the ones with full boxes are the 21 tissue trait pairs that I talked about on the previous slide. So for a summary of our results, before I get into that, I first do want to say that our results were able to replicate quite well. So for 10 traits, we were able to collect independent GWAS summary statistics and found that the direction of effect was consistent. And for a large proportion of tissue trait pairs, we even reidentified a signal at 5% FDR, which was really heartening to see. So some example tissue trait pairs that I want to point out for different reasons of specificity or identifying known associations or identifying plausibly novel associations. The first is finding that adipose subcutaneous might be a putatively causal tissue for HDL. And this is because previously, according to the literature, we knew that adipose was associated with HDL, but there is visceral adipose and there's subcutaneous adipose. And it does seem to be consistent with a new association— or not, not so new, but an association of adiponectin, which is associated specifically with this type of adipose, and also with HDL levels, while the tissue itself had not previously been associated with HDL. In terms of a known association, aorta artery was found to be the putatively causal tissue for diastolic blood pressure. And this is good because just by definition, DBP is indeed the measure of how much pressure is exerted on the aorta when the heart is relaxed. TCSE also found that aorta artery was a putatively causal tissue for glaucoma. I'm labeling this as a plausible, potentially novel association. This is something that hasn't been talked about before. It might be consistent with the fact that high blood pressure is a known risk factor for glaucoma, but it might also be a result of having the true causal tissue missing in our dataset such as an eye tissue. And so on the next slide, I'm going to dive a little bit more into why the aorta artery might come up with our method as associated with glaucoma. And then finally, heart left ventricle was found as the putatively causal tissue for platelet count. And we think this is quite consistent with the fact that platelets do cause blood clots in response to damaged blood vessels. And so indeed, if, the heart is pumping out blood, it might have an effect on serum levels of platelets. Returning back to this aorta artery glaucoma example, I wanted to dig into where GTEx sampled the aorta artery tissue from and then relate that to the vasculature around the eye. Here we can see from GTEx that they sampled from all these different segments in red. Now, of course, these segments are very far away from any vasculature that would be around the eye. But if we look at this schematic on the right, we see that there are all these blood vessels that are immediately connected to the aorta, and it's very possible that these tissues and other cell types share a lot of similar properties. Again, because of Maybe the true causal tissue is missing and this is simply the tissue that is the most similar, the most co-regulated with the true causal tissue. This is probably why tCSE was able to pinpoint this one over other artery tissues such as coronary artery or tibial artery. Now returning to comparisons of other methods. Now in the real trait scenario, whereas before we were comparing in simulations. Here we're just demonstrating that other methods are indeed less specific in implicating tissue trait pairs. And so this is consistent with the high false positive rates that we identified in simulations for these other methods and the lower overall area under the curve. So here we're showing the results from RTC-COLOC and the LDSC-SEG method. And each pair of tissues here are highly co-regulated, and a black line separates these pairs of co-regulated tissues. So on the right, you can see that t-CSC is quite distinguishing for every pair. It is only indicating one as the most likely causal tissue for a trait. And so here with LDSC-seq, we see that there's still some residual signal for co-regulated tissues. But for the most part, outside the pair of co-regulated tissues, it's quite good at distinguishing. Whereas RTCColo probably has the biggest problem with specificity, finding lots of ubiquitous signal across very different physio— very physiologically different tissues. And so, yeah, again, I just want to call back to simulations that we've kind of expected this result in real trait analysis. Due to what we saw with the poor specificity of these other methods and simulations. All right. And then another thing I just want to point out is that, again, it's another consideration for real trait analysis. So what happens when our model assumptions get violated? And model assumptions aren't something I've specifically talked about in this presentation yet, and probably for the sake of time. But a key assumption is that we assume that the gene trait or gene disease effect sizes are IID, independent and identically distributed across genes and tissues. So even though genes and tissues are co-regulated, the effects that they have on disease we expect to be independent. And so here we tested some scenarios where we violate these assumptions just to see how TCSE would perform. So in this particular example, we've duplicated the causal tissue in the model and added some noise to avoid colinearity issues. But effectively, what we see is that if you were to have two very, very— two tissues with the same effect on disease at the gene level, the ability of TCSC to pinpoint those two causal tissues now versus the background of the tagging tissues is still quite good. The type I error still hovers around 5%. And the estimate of heritability, because now these two tissues are sharing the same signal, it's halved. So before we were looking for a quantity of 10%. Now because we have a duplicate tissue, we're looking for each to be estimated at 5%. And while it's noisy, this is approximately what TCSE is finding. So we're actually quite happy with how TCSE is performing in the event of model assumptions being violated. And just talking a little bit more about when we would expect these model assumptions to be violated. In tissues, we're working with GTEx tissues in these empirical examples of applying TCSE. And tissues are quite different from one another. And of course, with modern RNA sequencing in single cells, the differences between cell types and cell states is getting ever increasingly smaller. So of course, you can imagine a scenario moving from tissues all the way to populations of single cells that you could have neighboring clusters of single cells that probably have shared gene-level effects on disease. And so this example in playing around with the model assumption violations is actually quite important because that would represent what would very realistically happen if we were to apply TCSE to single-cell data, which is on the pipeline. So the last real trait example that I wanted to demonstrate is the cross-trait tCSC results. So here we've partitioned the products of z-scores for TWAS for BMI and red blood cell count across coregulation scores across tissues. And interestingly, while no tissue was associated or whole blood was associated with red blood cell count in the single trait analysis, But the tissues I'm showing here, pancreas and the brain substantia nigra, were not associated with either BMI or RBC count in the single trait analysis. As you can see with the pink and green bars, these are not significantly greater than zero. So these did not pop up in the single trait analysis. These are those pi t results that I showed before. But the blue bar is the estimated— the cross-trait result, which is the covariance of the trait in a particular tissue divided by the genome-wide covariance. So it's kind of similar to how with piT we've divided by the SNP heritability. So interestingly, what we see is that gene trait effects in the pancreas seem to be in opposite direction for BMI and red blood cell count, evidenced by this negative covariance estimate. Significantly negative, I'll say. And then in the brain, the covariance is significantly positive. So this is really interesting. Perhaps the genes that do indeed have a true causal effect on BMI are consistent with their direction of effect, as are the genes with causal effects on red blood cell count. So this sort of just proposes a way that one could partition or ask which tissues are underlying the genetic covariance between two traits or two diseases. So here's just the full overview of our cross-trait TCSE results. And of course, for the sake of time, I've only presented one example. But there are other things that we can delve into when we look at, look at all of these results. And I do think that the most interesting examples are where for a pair of traits, there will be two tissues implicated in the cross-trait analysis where one has a positive contribution to the covariance and one has a negative contribution. And it's interesting because we can compare this to what is the genome-wide direction of covariance or genetic correlation and think about how that relates to each specific tissue's contribution. All right, so I want to wrap up the presentation with talking about how one could use TCSE and get started quickly. So we have a GitHub up and running, and hopefully it has all of the information that you would need to one, reperform all of the analysis the analyses that I've presented here and that we've done in the paper, but also how you can arrive with your new set of GWAS summary statistics and run TCSE quickly using the coregulation scores that I've already computed. Because if you do want to apply TCSE to GTEx data, there's really no need to process it all again, especially now that we've downsampled tissues so that they have equal sample size and meet those criteria for considerations that we've identified are important in simulations. So here's just a list of the data that is provided on the TCSE GitHub repository. So first, obviously, there's the link to clone the— well, first you would clone the TCSE repo. Then an important step of TCSE, right, is running TWAS and running these multivariate eQTL models. And so to do that, we use Fusion from Sasha Gusev's lab. And so the path to clone that repo is there. The path to download the 1,000 Genomes PLINK files is there, as well as a path to download the results from the multivariate eQTL regression. So these are the GTEx gene expression weight files that I've used in the method. So you can download that, which again pertains to downsampled GTEx tissues. So this would be the preferable one to use if you were applying tCSA. And then finally, coregulation scores, which need not be recomputed if you're using the same gene expression data that I am using. Alright, so here's just sort of a list to get of like the quick start. So if a user has a new set of GWAS summary statistics, for example, that weren't analyzed in our preprint and they wanted to know what is the most likely causal tissue for this trait, what they can do is arrive with their new set of GWAS summary statistics. We recommend that this, the disease or trait does have a high SNP heritability that's significantly greater than 0 and preferably a z-score greater than 6. This was a criteria that we used for the traits that we analyzed in our paper. You just wanna make sure that there's enough of a genetic component that when you try to partition it across tissues, you have power. Next on the repo, we have a quick start tutorial. So it advises how to format those summary statistics. How to run TWAS on those summary statistics using the already computed gene expression prediction models. And then lastly, running the t-CST regression, which is very fast and only takes approximately 2 minutes per trait and can be parallelized across traits if you're running multiple traits. So there's also instructions if you'd like to start with new eQTL data. So if you have, um, like a cohort other than GTEx, and you have genotype and gene expression data, there are instructions for using FUSION essentially to recompute the gene expression prediction models and co-regulation scores. And then finally, 3 types of analysis can be performed as are outlined in our preprint. So this primary analysis where we've tried to use as many tissues as possible, so So out of the 49 GTEx tissues, we use 39. So we've either downsampled tissues to have equal sample size at 320 individuals, or we've meta-analyzed very highly co-regulated tissues that had small sample size in order to get them up to 320 as an effective sample size. So that's why this number is less than 49. And then in the secondary analysis, we simply took the 23 tissues for which we had exactly 320 individuals. And then finally, a tertiary analysis where we just focused on brain tissue, where the motivation behind that was for a brain-related trait, perhaps it's not necessary to include non-brain tissues if you know that the causal tissue is somewhere in the brain. Because we know from simulations, power does decrease for the more tagging tissues that are added. So here's an example. If you are to follow all the steps on the GitHub, hopefully you're able to produce a plot like this. So for example, this is using ADHD summary, GWAS summary statistics. And this is just an example from the brain-specific analysis, which plots a little bit nicer because there's fewer tissues. So here it's a nice example because TCSE is indeed implicating a single tissue. The hippocampus. And you can see the pi T prime results for all of the other tissues that are not significantly different than zero. And just quantitatively speaking, TCSE estimates that 13% of ADHD heritability can be attributed to the genetic component of gene expression in the hippocampus. All right. So you can also download just a table of all of the results across all of the 78 traits that we analyzed and those 39 GTEx tissues. So just in conclusion, I've talked today a lot about how co-regulation scores and TWAS statistics can be used to infer putatively causal tissues underlying disease heritability in the single trait analysis or disease covariance in the cross-trait analysis. TCSE was able to identify some new biologically plausible tissue disease pairs such as the aorta artery and glaucoma. But we also, you know, provide some reason for why the TCSE might identify this tissue if the true causal tissue is indeed missing. And then lastly, TCSE might be more informative when applied to different types of eQTL data. So GTEx is steady state eQTL data. So perhaps if we were to apply this to dynamic eQTL datasets, case-control eQTL datasets, or single-cell eQTL datasets, we might have more biologically compelling applications of tCSE. So we're looking forward to those future developments. And so with that, I just want to acknowledge all of the great mentorship I've received from the Price Lab, including from Alkes and colleagues. You can check out our manuscript, which we published last summer on bioRxiv. And yeah, so this is a link to my lab website and just a nice picture of the scenery around UCSD. So with that, I will be happy to take any questions.

Thank you so much, Tiffany, for an exceptional talk, and congratulations on starting your lab.

Thank you so much, Sarah.

Please, if anyone has questions, feel free to type in the Q&A or raise your hand and we'll unmute you. I had a question about how you would apply this to single-cell data given, as you were saying, the loss of power. Do you think, would you start by looking at a single cell type in each tissue or starting with a single tissue and looking at cell types or just go for all of them?

Yeah. So we've started to work on this in my group and it's really interesting because there's so many avenues for statistical innovation with analyzing single-cell eQTL data in this manner. For example, even identifying single-cell eQTLs, that is also ever-evolving. And there's— one could apply linear models. One could apply Poisson models. One could apply univariate models. One could apply multivariate models. So there's many different ways to do this. And also, the thinking of— of the simplest way would be to pseudo-bulk single-cell data and treat TCSC just like we employed it in GTEx. But we're thinking of more interesting ways that really leverage the continuity of the single-cell data and sort of more play around with sort of like the interactions at the genetic level with different cell types. And so yeah, there are many ways to do this. And yes, power is concern. But at the same time, there are really large single-cell datasets out there in terms of not just the number of cells, but the number of donors exceeding the number of donors and thus genetic variation available to partition that we had in GTEx. So there's actually like some potential for an even greater increase in power also because of the disease relevance of the the cells that are assayed. So again, just like saying the steady-state gene expression in GTEx of these bulk tissues probably washes out the signal of any specific disease cell type that is a smaller proportion of the total tissue, and maybe using single-cell data will actually give us better power to see those genetic effects.

I will jump in with a question too. That was such a cool talk, Tiffany. Thank you so much. I wanted to follow up on this point that you raised about how— let me see if I can get it right— the sharing of cell types across tissues can sometimes enable slash result in you're detecting an association to a tissue that shares a cell type but isn't necessarily the tissue that is disease relevant. And I'm wondering from looking at the data that you have whether you have any general impressions of which tissues are A, undersampled of course, and B, should be prioritized because they come up so often potentially through this cell type sharing.

Yeah. Yeah. So I think like, for example, tissues that come up really often as being associated with traits, the tissue that comes up the most often by far is whole blood. That could be because there are one, we have a lot of GWAS for blood cell traits and these traits are also very accurately measured. So there's probably more power in the GWAS compared to other like case control traits, for example. But that being said, there's also like so much heterogeneity in whole blood, you have like 1% that are lymphocytes and actually related, very important for immune function. And so, yeah, this dataset is not perfect if you're trying to identify associations with immune diseases because finding whole blood is not interesting, but also whole blood is such a diluted version of what cell types are actually relevant to those diseases. So I hope that answers your question. And if not, I'm happy to—

No, you did that. That's super cool. Thank you. And I mean, I guess the really cheerful thing is that as you add more tissue types, your specificity to the data, your specificity necessarily increases, which is very cool.

Yeah, yeah, exactly. It remains pretty constant.

Yeah.

Great.

Well, thank you again for presenting to us today. I hope you enjoy being back in Boston for the day. And I hope the city shows off a little bit and can compete with this lovely view that you have.

Thank you so much, Sarah and Diane, for inviting me. Um, it was a pleasure to give this talk this morning.