Scientist Stories: Christoph Lippert, Unleashing the Genetic Architecture of Heritable Traits
In this Data Science Seminar, Prof. Christoph Lippert presents his research on using machine learning, especially deep learning, together with statistical genetics to uncover the genetic architecture of complex, heritable traits. He begins with a concise overview of genome‑wide association studies (GWAS) and polygenic risk scores (PRS), illustrating how large‑scale GWAS meta‑analyses have identified thousands of SNP‑trait associations but individual effect sizes are small. Lippert explains how the INTERVEEEN consortium aggregates European biobank data (~2 million individuals) to compute PRS that can stratify disease risk (e.g., type‑2 diabetes) and guide public‑health screening. He then shifts to the challenge of high‑dimensional phenotypes such as MRI or retinal fundus images, where traditional GWAS would require compressing images to a single scalar. Lippert proposes a “transfer‑GWAS” pipeline: (1) use a pretrained convolutional neural network (CNN) to extract hidden‑layer embeddings from images, (2) optionally reduce dimensionality with PCA, and (3) perform GWAS on these embeddings using a kernel‑based two‑sample test (maximum‑mean‑discrepancy) or a linear model that incorporates covariates. He discusses two strategies for obtaining the CNN: (a) sample‑split training on the same cohort, which sacrifices test sample size, and (b) transfer learning from large external image datasets (ImageNet or a retinal‑specific Kaggle set). Empirical results on UK Biobank retinal images show many genome loci associated with image variation, some overlapping known eye‑disease genes, and a trade‑off between number of hits and effect strength depending on the pretraining source. Lippert acknowledges limited interpretability of image‑based associations and outlines ongoing work on explainable AI methods (collaboration with Wojciech Samek) and theoretical extensions (with Sonja Greven). The talk concludes with a description of rare‑variant set tests developed by his student Remo: (1) aggregate ultra‑rare variants into functional sets, (2) construct kernels weighted by neural‑network‑predicted functional annotations, (3) use a small‑permutation‑plus‑parametric‑fit approach to obtain accurate p‑values at GWAS scale. This method improves power and interpretability, exemplified by a discovered missense variant in PIEZO1 affecting red‑blood‑cell biomarkers. Overall, Lippert’s work demonstrates how deep learning can be fused with rigorous statistical testing to leverage massive, high‑dimensional biomedical data for genetic discovery and potential clinical translation.
Chapters
Good morning, everyone. Welcome to the Data Science Seminar. This event is jointly hosted by the Alice Live Heidelberg. A warm welcome to all the representatives. We welcome our guest speaker, Christoph Lippert. He's professor and chair of digital health and machine learning at the Hasso Plattner Institute for Digital Engineering and University Potsdam, and is an adjunct professor at the Icahn School of Medicine at Mount Sinai. Welcome to Brian Clark. He is group leader in the division of Professor Stegele. Thank you for taking over the introduction.
Yeah, thanks. I'm very happy today that we have Christoph joining us to give a talk on his work. As Doreen mentioned, Christoph is a professor and chair of digital health and machine learning at the Hasso Plattner Institute for Digital Engineering in the University of Potsdam. He's an adjunct professor at the Icahn School of Medicine, and his research has been focusing on developing novel methods that integrate machine learning and statistics to study the genetic architecture of heritable traits. So I have actually quite a personal interest in what he's going to be presenting today, I think. And Christoph especially likes to focus on interdisciplinary methods where you're combining genomics, imaging, and deep learning. To unravel complex relationships between genotype and phenotype. Christoph studied bioinformatics in Munich and did his PhD at the University of Tübingen and the Max Planck Institute in Tübingen, after which— well, before he joined the House of Platten Institute, he held positions at the Max Delbrück Center for Molecular Biology in Berlin, at Human Longevity Incorporated in Mountain View, and Microsoft Research in LA. Yes, and we're very curious to hear about some of the things he's been working on at these various positions. So Christoph, hand it over to you.
Thank you very much for the kind introduction, Brian. So let me share my screen. So I wanted to talk today about our work on genetics. And as Brian already said, we're interested in integrating imaging with genetics. And the tools that we use to do so typically involve machine learning, in particular today, nowadays, mostly deep learning methods with a statistical slant, I would say, to it. So let me give you an overview of the talk. I don't know whether we can cover everything in the 45 minutes, so I might actually skip some of these. But I want to first— I heard there are some students here also, so I want to give you a brief introduction what genome-wide association studies are, talk a little bit about the use of these studies for polygenic risk scores, And then on our work on developing statistical tests based on deep learning for genome-wide association studies of image or imaging traits. I want to present some work by my student Remo, who has been working on rare variant tests on sequencing data in the UK Biobank. And if time permits, on multimodal self-supervised learning that combines imaging and genetics. All right, so this is my group at the digital health cluster building in Potsdam, the Hasso-Plattner Institute. So I don't know, the Hasso-Plattner Institute is sort of a public-private partnership. We're completely funded by the Hasso Plattner Foundation. So Hasso Plattner is one of the SAP founders. He has given his wealth into a foundation that covers several not-for-profit endeavors, including museums. So he's lots of funding for art. But one of the largest items is the Hasso-Plattner Institute. We're around currently 25 professors, all in computer science, and 4 in this area between computer science and biomedicine. Well, all professors at the University of Potsdam, so we're basically actually a privately funded faculty at University of Potsdam. So that's kind of unique in the German system. So we are in a public university, we have our own bachelor-master programs in computer science. And yeah, but our money comes from the foundation. All right, so what are genome-wide association studies? So in genome-wide association studies, We have— we're usually interested in studying heritable traits, so we look at large populations of people, so typically in the tens or hundreds of thousands. Nowadays, large meta-analysis go into the millions of people, where we usually have a trait that can be simply measured by a single number, and on the other side we have the genomes of these people. Now, and what we're interested in is positions in the genome that explain differences, basically phenotypic heterogeneity. Now most of our genome is actually identical. There are only a few positions that vary, and this is typically represented in so-called single nucleotide polymorphism. So single nucleotides that are polymorphic, that are different between different groups of people. And yeah, so this is how an association would look like. Basically, you have the trait and you group the trait by people who have a C at one position and T at the other position, and you're basically looking for differences in the means using statistical models such as this linear regression model. And then you can basically do a simple statistical test, say Z-test or t-test on the regression coefficient. And your association test boils down to looking for effect size estimates that are significantly, statistically significantly different from zero. Now you can do that for the whole genome, and then what you will get is a plot like this where, because you're doing this for each of those single nucleotide polymorphisms that are spread all over the genomes separately. And for each of those, you're basically computing this p-value based on this effect size. And then you plot typically the log or the negative log of these p-values versus the position. And so what we see here is on the x-axis is the position in the genome, so we see the indices of the chromosome. On the y-axis, we see the negative log 10 p-values, and then you see that we get some positions that have several of these small p-values, and those are sort of associated loci. Why do we see many typically at the same spot? That's because neighboring positions on a genome are correlated by linkage disequilibrium. So that's basically by the way we inherit our DNA and the way crossover works in the inheritance of the chromosomes that causes this correlation. So this technique is not new and has been done for a long time and for many, many traits. And so for a long time, the NIH and the EMBL-EBI have been tracking lists of these associated traits. And I have been updating this plot here. The last time that they've updated is actually in 2018. So they don't do that anymore just because— so what you'll see is again, all the human chromosomes laid out and then each of those little dots is basically an association that has been found with on some of one of these studies with some group of traits. So these are tens of thousands of associations. As I said, not updated from 2018, so by now this is a multitude of what we're seeing here. So this is sort of on the one side quite successful because we see that a lot of our genome is associated to these heritable traits. However, So far, actually, relatively little has been done with these results because the individual effect sizes are largely relatively small and, you know, not actionable. So it doesn't tell you— if you know a single of these mutations, it typically doesn't help you very much on predicting the outcome. But there is now an interest in combining these association studies to polygenic scores, and this is something that we do as a member of the INTERVEEN consortium, where we pool sort of the largest European biobanks together. So we have data for genotype and phenotype data for close to 2 million, largely Europeans and some Americans, to combine these individual associations that have been found to polygenic scores. So that you can think of them as sort of taking for, say, for a single disease all associations and using the combined effect of all of these to predict the genetic risk of a person. So it's polygenic because we're using many of these individual SNPs to compute them. And while each of those has a small effect, we see that on preliminary data from our consortium that if you compare, for example, the highest risk percentile versus the lowest risk percentile, or even the the sort of the average person, that in terms of risk of getting the disease at a certain age, there's a huge difference in number of years. So this is a plot for that predicts the sex-specific incidence for type 2 diabetes in Finland. And you see that this red line here depicts the official risk threshold for screening that is recommended by the Finnish government. And you see that people who live or are in this top first percentile actually should be screened at an age just above 20, versus if individuals that fall into the lowest percentile, they pretty much wouldn't require screening at all during their lifetime. So you see that by combining these scores into polygenic scores, we can actually predict meaningful outcome in terms of public health, from a public health perspective. And in the Intervene Consortium, we are not only computing these scores, but we are also working on trials that actually try to do interventional trials of communicating people their risk and to see whether preventative measures have an effect on their health. So these polygenic scores, they have multiple utilities. This is a figure that is taken from a review paper by Lewis and Vassos. So ideally, each person would get their risk estimates at birth because it basically— that's where genetics has sort of the largest influence because there hasn't been yet a lot of environmental exposure. That's actually where genetic risk would be most actionable, where if you identify the people that are at high risk, you can change lifestyle or ascertain early detection methods to these people. But they have value throughout the lifetime where as people get older, typically the genetic the value of the genetic component becomes smaller because your lifestyle will determine more of the variance. In effect, the way these polygenic scores are computed is pretty much always the same. So typically you have a database that contains these genome-wide association study results in the form of summary statistics. So these are the effect weights essentially, plus the confidence intervals of the effect weights. And now your goal is to predict the risk of individuals on some target dataset, which is not the same one as the effects have been estimated. So you are basically somehow using this target data to to build these weighted sums of these individual effects to combine them into a polygenic score. So, and yeah, there are various ways to do it, but in effect there are always these weighted sums that come out. And my student Remo, he has been computing a lot of these scores for a lot of traits on different biobanks that are part of the Interveen Consortium. And here you can see sort of how predictive they are for different traits. I won't go through all of this. An interesting thing is that— so each of those lines would be sort of a different existing method for polygenic score computation. So basically a different recipe to come up with this weighted sum. And you see that some perform generally quite good, and typically the best method would be, for example, this LD-PRED here or MEGA-PRS. So in case you have to do a study, these are good methods. There are also simpler ones that don't perform well, but what One thing that we observe is that for traits such as type 1 diabetes or arthritis, we see a huge variability in the way that these methods actually perform that cannot be explained by one method just being a tiny bit better than the other. And we actually hypothesize that this is largely not by differences in computing this weighted sum, but rather by differences in how you take into account the immune-related part of the genome, which is kind of hard to genotype and model. So it's more the preprocessing in this case than anything else. And these are exactly traits that have a large autoimmune component. You cannot only do that for binary traits, so disease or healthy, basically compute a risk, but you can also do that for continuous traits, of course, and then you're essentially computing a regression. So the archetype of heritable continuous traits in genetics is human height, and you see that This is a study from 2018, not by my group, but that was sort of using a lasso to predict, to do a polygenic prediction of human height in UK Biobank. You see that this is highly heritable and can be predicted quite accurately from the genome alone. And similarly, we can also look at continuous traits. On these different biobanks. Yeah, so this was the part on polygenic risk scores. Now I want to go more into sort of this methods that we are developing towards combining, you know, deep learning and statistics to develop deep learning-based statistical tests for for doing GWAS of imaging traits. So traditionally, GWAS have been done with single traits that can be summarized as single numbers. For example, the distance between head to toe of a person, human height, as I said, is this archetype of a highly heritable quantitative trait. And so you would collect the genetics, you would put a measure tape on people and then you're just essentially just computing a correlation between these two things, right? In large populations. So here you sort people by height and you would, for example, see an association between the chromosomal distribution of these people. So these are males and females. So you have an association between the presence of two X chromosomes leading to typically smaller height than X and Y. Now, summarizing, arguably summarizing the human phenotype into a single number is not very interesting. And as novel sensors improve, one can actually get much more insight in the human phenotype, for example, using MRIs, MRI, or also taking molecular measurements via various sequencing protocols. So, and then the analysis is not so easy anymore. So we need additional processing steps because we cannot just put all of these beating hearts, for example, represented by a single number and then compute a correlation coefficient to a SNP that causes more beating or more blood flow versus less or so. And in the UK Biobank, The phenotyping is actually done using many of these sensors. So there is MRI, detailed MRI measurements are taking pretty much for the whole body and also various molecular measurements, but we are currently mostly interested in the MRI part. So the challenge here in doing basically finding genetic changes in variability in these traits is that we have quite complex phenotypic data. So we have, say, 50,000 MRIs of the brain or of the heart. And on the other side, we have the genomes and we want to somehow associate this. Traditionally, we would have summarized this into a single number or a single binary trait, but what we would like to do is to not have this sort of strong compression, but rather use the data directly. And the idea here is to use deep neural networks to do so. So this is the goal. On one side, we have all your metric images. So you don't have just a single image, but many of these slices. And what you want is to do your GWAS. So of course, yeah, a simple way to use machine learning would be to actually say, okay, well, let's predict a single number from these images. So say predict the size of some organ in the brain, the ventricles or so, and then do the standard GWAS. Now the problem with this is that A, you need to train this neural network, so you first need to annotate all of these images and basically need to know what you're looking for. So what we are really interested in is rather just look at all the variability that is in the images. And the idea that we had is that by not looking at an output layer, basically by prediction for a single trait, but rather at a hidden layer of a neural network, basically by doing statistics there, we get basically a feature extractor that gets all kinds of interesting information that is contained in the image out, and then we can do statistics on that. And it turns out that this naive idea actually is quite a good idea and performs also quite well. So I had a student, Matthias, working on this problem. He is a mathematician. He looked at it as a more from a theoretical side. And what he found is that if under some circumstances, basically you can view this neural network or this hidden layer projection of the neural network as a kernel of the image and then apply existing kernel tests to this and that this would actually represent something that would resemble a non-parametric statistical test. So if you basically do a two-sample test, so you have the— you group all the individuals by their genotype and look at the distances on the projection layer of the mean of these two sets of individuals, then that gives you a statistical test, a two-sample test that is nonparametric. So there's the setting. We take images from two distributions, so Xi and Yi, so the people with a C and the people with a T. And the null hypothesis would be that these hidden layer projections come from the same distributions. And if we just perform a simple two-sample test on these encodings, then Matthias was able to prove that the test would be nonparametric in the sense that if you let the sample size go to infinite, you would find that your p-values would go to zero independent of what the differences in these traits are. So if there are differences in distribution, you don't have to specify what they are. You just need a large enough sample size, you will see differences at some point for some some n. Now the problem is, how do we get these neural networks? So the devil is in the detail. So the problem is, if we use the dataset that we're doing the test on to train the neural network, we're actually using the data twice, which gives us an invalid test statistic. So there are two ways to mitigate that. One is to do a sample split. That's something that has been traditionally done. So on one part of the data, you train a neural network. On the other part of the data, you do this hidden layer projection and you do your test. The nice thing about that is that it doesn't need an external dataset, and the problem is that it reduces the available sample size for testing. So there's obviously some form of trade-off where if we make the training dataset larger, the testing set becomes smaller, and we're trading off the quality of the trained neural network versus then power on this trained neural network. The other idea that was proposed originally in this paper by Matthias, and it's actually quite cool because it's so simple, is to use a neural network that has been trained on a similar but different dataset, for example, ImageNet, and transfer the learned encodings to our dataset. So we just use this ImageNet-trained model to get our hidden layer projection, and then we would perform the testing. On this dataset. And this is actually what— so if this different dataset is similar enough, then all these theoretical properties of this nonparametric test would hold. Based on this two-sample method, we saw that this is sort of something good to do. It's actually the test statistic for this two-sample test was really simple. It was basically just a difference in the mean, just the MMD statistic, maximum mean discrepancy. It's a standard kernel test. We then modified this to basically use a linear model on the same projection to do a GWAS that also enables us not only to do a two-sample test, basically look for difference between two groups, but also incorporate information about covariates into our test. So the idea is completely simple. As we have our neural network, we project our images and then we take our normal way to do GWAS on this hidden layer projection. And yeah, this is sort of the basis of this transfer GWAS paper, which was a follow-up then by Matthias where he applied this to retinal fundus images in the UK Biobank. So in the UK Biobank, we have retinal fundus images, I think of 75,000 people from the— no, we have 150,000, total of 150,000 images. Might be that several people have been sampled multiple times, but the 150,000 images are also left and right eye. So the exact sample size is roughly half of that. Here is actually the information. So we have 64,000 subjects, so we have done repeated measurements of several of these. The method works as follows. So you have You first train your neural network, which can either be done on a very large dataset of just images that are scraped from the web. That would be based on ImageNet. You train a convolutional neural network, or similarly, you could also do this based on another dataset of retinal fundus images. So this would— he actually compared this to using a Kaggle dataset of roughly n 100,000 individuals where the task in this challenge was to do prediction of diabetic retinopathy. But both of these, and in this case you would predict what class the image belongs to, so this, the class label here would be dog, here I think this is I don't know what's on this image, but here's another dog. But it's basically something that clearly is completely unrelated to the biomedical domain here. But still, we observed that these representations were nicely transferable to our original fundus images on the UK Biobank. So this would be the process. We take our study samples from the UK Biobank, we take this trained convolutional neural network, we remove the output part, so we stop at a hidden layer projection, and then we have sort of our features extracted from the images. Now this layer is These are quite a lot of features, and now you run into the problem of multiple hypothesis testing as you're combining them to a single test, because in the end, these representations are not sufficiently disentangled in the sense that each feature would correspond to a different feature in the images. Distributed representations, right? So these represent just some variation in the image. So we do another feature reduction step using PCA to something that is manageable, and then we test— do basically a multivariate test of the resulting principal components projection. And then we get our Manhattan plot similar to the one that you saw before by testing on these principal components. I said we do two types of pre-training. One is using ImageNet. The nice thing about that is that we don't have to do it ourselves. We can just download a neural network from the web. And if we look at the first 10 principal components of this hidden layer representations, we can actually see what these features correspond to and that we actually observe— sorry, someone was coming into my office— we observed that some of these dimensions correspond to brightness and so on. So these are just general image features. Some correspond to the quality of the images, but also correspond to eye diseases if you look at sort of extreme values of these principal components. The other type of pre-training was based on this Kaggle competition on diabetic retinopathy prediction. And here you see even more features that correspond to eye diseases. These are clearly more clear visible here. If you look at the results in terms of loci that we find is that we find many places in the genome, many loci that have prior been associated to different eye-related traits, but also non-eye-related traits. One observation when we observe the— compare the results on this ImageNet pre-trained model and the IPACS pre-trained model is that on the ImageNet pre-trained model, we find more associations. However, the ones that we find are not as strong. And on the IPACS model, we find fewer associations, but the ones that we find are stronger. And this can be— the interpretation of this is that the IPEX model, because it is predicting, is basically specifying on predicting a single IDZ, it extracts much more sort of task-specific features from these images compared to this ImageNet pre-trained model which just extracts any image features that are useful for classification of natural images. But both ways this approach works quite well and we were quite happy with the results. Now, a downside of this is of course that we get associations with any kind of image variation. So they do lack interpretability in some sense. So we don't know what the individual features correspond to. Of course, we can look at the association list directly. In this case, we actually took some samples that— or some of the SNPs— and then actually trained the neural network the other way around to predict the genotype from the image. And then we can look at sort of differences in the images that best predict sort of the different genotypes in order to explain. But this is not yet sufficient for us really to— I don't think this is completely there in terms of the explainability. So there's a lot to do. And that's something that we're still working on. Since January, we are a member of the DFG research unit on fusing deep learning and statistics towards understanding structured biomedical data, where one of the projects that we are involved is to develop explainability methods for these statistical tests that work on images. And in a second project, we are extending the methods and theory behind these statistical tests. So one, this theory project is together with Sonja Greven, who is a statistician at Humboldt University of Berlin, and the explainability project is together with Wojciech Samek from the Heinrich-Hertz-Institut, who is an expert in explainable AI. So let me summarize this part. So we develop powerful statistical methods for genetic associations. This is active work. And the goal is to make use of complex high-dimensional data. That we have already achieved. And we aim to generate also more interpretable results and You also get some more results on the theory side. Now for the last part of this talk, I want to briefly talk about the work of Remo on rare variant tests for GWAS. So these tests also use neural networks, but they don't use them on the phenotype data. But rather on the genetic side. So far we talked about these single nucleotide polymorphisms, which typically the ones that we look at in GWAS are frequent. As we are more and more moving towards sequencing large populations of individuals, and in fact all of the UK Biobank, which is pretty much the most important dataset that is available for these genetic studies, are undergoing whole genome sequencing. And already for all of them, I think for all of them, the exome, whole exome sequencing data is already available. The problem is, As we look at sequencing data, we find more and more extremely low frequent variants where we are in the regime that even though we have hundreds of thousands of individuals, we lack statistical power. Here you see a plot of, an example plot of the the exome data of the UK Biobank, where on the x-axis you see the minor allele frequency, or the logarithm of the minor allele frequency, and on the y-axis you see the frequency. And you see that for very small minor allele frequencies, you see hundreds of thousands of individuals, or hundreds of thousands of these variants that have No, it's actually millions of these variants that have these small minor allele frequency. And actually going back to my time at Human Longevity Inc., where we sequence a lot of whole genomes, you actually, the number of variants that one finds that largely happen, say in only one individual actually goes into the hundreds of millions, if not even approaching a billion. So obviously we don't have statistical power to study all of these. So we need some way to aggregate the information from many rare variants. And this is something that has been done using, for example, these variance components models where what we do is instead of testing, basically doing a linear regression based on a single SNP, we find some strategy to group these single variants into sets of variants and then ask the question whether taken together they explain any variability in our phenotype of interest. So we're replacing many tests by a single test on these individually very low frequent variants. And basically this corresponds to, if assuming that all of these effect sizes follow a normal distribution, And you integrate over this normal distribution that corresponds to actually writing down a model that has a covariance matrix that is built from the set of variants that you're interested in. And this test then boils down to asking whether if we scale this covariance matrix with a single scalar, whether the scalar is 0. Then this covariance matrix would explain no variance in the phenotype or is larger than zero, which would mean that this group is associated. So there's an idea that has been proposed in the GWAS field for also a long time ago. There's a paper from Harvard in 2011 It's called SCAT, Sequence Kernel Association Testing. And yeah, you can also do population structure correction if you want to in these models. And in the past, I've worked a lot also on these set tests. Now the problem is computing these tests is not so easy. So what actually does Scott paper does is it uses an asymptotic approximation to the test that makes it quite fast, but we find actually that this doesn't have a lot of power, so it's extremely conservative. The gold standard would be to do a permutation test where you would, for example, permute the vector of all the phenotypes and and in order to generate an artificial null distribution. The problem is that while it's powerful, it's prohibitively slow. So back in the days, we had the idea that we could do a small number of permutations and then fit a function to this permuted distribution in order to extrapolate into the tail of a very small values. And we found that this actually works quite well empirically and is computationally as efficient as these asymptotic approximations. Now, we have done sort of an update on this where we improved these parametric fits and small number of permutations parametric fits to scale to current GWAS. So this is work by Remo in the paper that I just outlined to tackle dataset sizes in the millions. And we added a small proof that this method that we found works empirically quite well also controls the family-wise error rate in a correct way. Another thing that Remo did here is to extend the way we compute the kernels by basically including functional annotations into the kernel. So if you If you look at this figure here, what we're interested in is actually testing this covariance matrix that is built from individual SNPs jointly. Now, there are actually many ways how you could construct such a covariance matrix from many SNPs, and the idea here was to use functional annotations to build a weighted covariance matrix. Where variants that have some function would get a higher weight compared to variants that likely have no function. That should remove basically the dilutive effect of all of these variants that are likely not likely to have a function and increase power. On the other side, it's also more interpretable because we're reducing the sets of variants only to the ones that we have prior information on. And he used regulatory genomic datasets to train neural networks to basically predict the functional— the function of these variants, and then incorporated these into the tests. And due to the lack of time, I won't go on in this detail, but in a sense, These are convolutional neural networks, fairly standard methods in functional genomics where you predict for each stretch of sequence the effect of this sequence in an experiment. Yeah, so why is this useful? Well, the first thing is motivation is to gain statistical power because we have these many, many variants, which results in also many ways to group these variants. So that results in a large multiple hypothesis testing problem. But we also find that it improves interpretability because it focuses on groups of variants that likely have a function, and this function tells us, you know, some, or gives us some hint of causality of how these variants work. This is an overview again of the methods, but I think I want to come to conclusion because we're running already out of time. So the results of the study I won't discuss, but he found, as I say, so he applied this largely to blood biomarkers and found quite interesting results. For example, he found a missense variant in Piezo1 that actually likely causes stomatocytosis. This increased in fact red blood cell turnover and resulted in lower levels of the blood biomarker that he looked at. Yeah, so I won't explain the large part, but rather want to conclude this talk And say thank you to the people who actually did the work again, and I'm happy to take your questions.
More from Joshua
Scientist Stories: Hans Waterham, The value of phenotypic assays in peroxisomal disorders
Scientist Stories: Joel Zlotogora, Genetic carrier screening programs for reproductive purposes