Scientist Stories: Anne Carpenter, Targeting diseases with morphological profiling

Hey everybody, welcome back. It's a real treat to be here at this venue to tell you about the research that's going on in my laboratory at the Broad Institute in Cambridge, Massachusetts. It's especially nice because I'm going to be telling you not just about the work that's going on in my own laboratory, but because we work on so many collaborative projects, you'll get a little glimpse of what a lot of different people are doing with images. So I hope you'll enjoy it. I first want to give a little bit of context. The images we work with are substantially smaller and lower resolution than many that we've seen so far in this conference. We work primarily on large-scale imaging experiments where we're testing lots of either genetic or chemical perturbations in multi-well plates. We use automated microscopes, and because we collaborate with so many people, these come from all kinds of different instrumentation, sometimes just conventional laboratory microscopes, and and sometimes those boxed instruments that collect a lot of images. Once we have the images, we run them through algorithms to identify regions of interest that are important to the biological question at hand, and then we do all kinds of— to extract all kinds of cell measurements, and then we do a lot of data exploration and machine learning. And a good bit of my group has transitioned to focusing primarily on data exploration as opposed to image processing in the past few years, so we'll talk about that as well. If you've heard of my lab, it might be because of CellProfiler. That's the open-source software project that we lead, and it's open source. It's used by thousands of biologists, and I'm not going to talk much about the software itself, but I will say that next week we have CellProfiler 3.0 coming out after a couple of years of development, and it has 3D capabilities. I'll mention that a bit. But the overall story that I want to tell today is a shift in how image analysis has been done over the past few years. And the— this shift involves increasing computational complexity, but also that increase in complexity has allowed us to attack biological problems that are increasingly complex. And for biologists to use cellular systems or biological systems that are more physiologically relevant and more complicated, and so the caliber of questions that we can answer is also increasing over time. And so I want to take you through this trajectory. I won't spend too much time on the first two I really want to focus on the latter, which has been the most recent work in my group. But first, let's talk about the— start at the simple end of the spectrum. Often it is the case that biologists say, hey, I'm going to be collecting these images, can you help me measure this thing that I want to measure? And so in such cases, your goal is to extract those features in an accurate and as automated a way as possible in order to score a particular phenotype. So a few examples of this are are things like counting GFP-labeled tuberculosis in a mouse macrophage system so that you can find antibiotics that don't just actually kill the bacteria right out, but actually allow the macrophages to survive the infection. And this could be working by interesting new mechanisms of action, which are sorely needed for antibiotics. So that was an example of a project with Deb Hung's lab. We— a much higher resolution project involved— this is one single nucleus of a cell— gamma irradiation in an oncology project aiming to identify genes in the DNA damage response pathway with Mike Yaffe's lab. So there the goal was, can you count up the speckles that are produced after irradiation and DNA damage is induced? Project with Fred Ausubel's lab over at MGH involves whole organisms. This is C. elegans. In this assay, C. elegans are the good guys. They are infected by Enterococcus faecalis just like humans are, and our goal here was to again to find anti-infectives, not antibiotics. These are compounds that allow the worm to survive but without directly killing the bacteria. And so these are targeting mechanisms like how does the bacterium get into the worm and how does the bacterium survive once it's inside the worm. And then lastly, one we're quite proud of because it's the first that's progressed to a clinical trial from my lab, involved a very simple DNA stain, the software identifying each nucleus and measuring the polyploidy of the nuclei. Here the goal was to treat AMR AML, which is a form of leukemia, not by killing the leukemic cells, but instead causing them to become polyploid and thus differentiate into a less harmful state to the patient. So these are all very simple examples. We're counting speckles, we're measuring fluorescence intensity, and so on. And so that's, that's that. The second stage I want to talk about is where the biologist knows what they're looking for, but they can't quite reduce it to a single particular feature, so it's not as easily described. And so in these cases, we use machine learning to to identify the phenotype of interest. And to do this, we made a tool. The important part of our contribution to these kinds of projects has been to make tools that put the power of machine learning into the hands of biologists. And so there's no one better to train a computer vision system than the biologist who is dedicated to a particular disease area and to a particular project and knows exactly what they're looking for and what the trade-offs are for accuracy and so on. So this tool uses cells that have been processed through CellProfiler or some other image analysis software where you've measured a lot of metrics about each individual cell. And the way it works from the biologist perspective is you just get a bunch of cells from your experiment and you drag and drop them into different categories. And it's a supervised machine learning system, so as you're dragging and dropping them into different categories, the software is learning what is it— what is it about the metrics that make this type of— these bipolar monoastral cells look different from the multipolar ones and different from prophase and monopolar and so on. So these are different stages of mitosis that were tested here. This general tool, I think what we're most proud of is that it requires very little tweaking from one complicated phenotype that one person might be interested to something that someone else might be interested in. The general pipeline is the same. You just sit down in front of the computer for anywhere from an hour to maybe an entire day if it's a super difficult phenotype. And at the end of the day, you have a system that can score your phenotype as as well— usually as well as you can. And so some projects that we've tackled using this have allowed us to interrogate some more complex cellular systems. One example is this co-culture. You don't see the stroma in this picture. We're looking at GFP-labeled hematopoietic stem cells, and in particular leukemic ones. And what we're looking for here is these cobblestone regions. So this is where the cells have sort of flattened out and lost their round appearance. What this allowed us to do is to identify drugs that can potentially treat leukemia without damaging hematopoietic stem cells, which is crucial for a real therapeutic. Secondly, a collaboration with Sangeeta Bhatia's lab at MIT. She's very interested in engineering human livers. They're hard to come by. The waiting lists are incredible for them. So it's very difficult to get a liver for patients, but it's also difficult to get hepatocytes for experiments because they're so rare. And so she did a screen in order to identify chemicals that would stimulate the proliferation of these hepatocytes in culture. That will allow her to have material to do experiments, but also eventually, hopefully, to have enough material to make human livers as replacements. And so in this experiment, the goal was to differentiate these starred cells, which are the human hepatocytes, versus the rest of the cells, which are fibroblasts. And in this case, the difference was just that the cells were a bit smaller, a bit rounder, a bit blurrier. But it required machine learning to really distinguish them nicely. Another whole organism project with the Ruvkin Lab at MGH involves fat metabolism. It's hard to study metabolism in cell culture because we don't know all the components that are involved. And so here, looking in a whole organism with obvious therapeutic potential, they were interested in identifying genes whose knockdown would alter fat metabolism. So Carolina Walby in the group, who many of you may know has her own lab now at Uppsala University, She developed some very nice algorithms to disentangle the worms, and then after you've done that, it's very simple to line— make a lineup of these worms, and then we use machine learning to look at the different distributions of fat along the body. And then lastly, in a glioblastoma project, we were looking for genes that modulate the transition of this particular cell type from a more neurosphere phenotype to a more neuron-looking phenotype. So that's it. That's what we've been doing for machine learning. The things we've been working on in this domain as well are three-dimensional analysis. And if you download CellProfiler next week, you can test out this capability. This has been a really great collaboration with the Allen Institute for Cell Science, which is collecting a lot of large-scale 3D time-lapse imagery. And I encourage you to check out their Cell Explorer. It's a really great resource, and also a lot of pretty pictures to look at that they've captured. Another area that we're working on in image processing involves deep learning. And again, our goal is to take advanced techniques and put them into biologists' hands. So we don't have any user-friendly— we don't have a user-friendly tool for this particular application, but it's on its way. Where the idea is it's very difficult for experts even to agree on the staging of malaria infection in red blood cells. So as this parasite is infecting, the cells take on very, very subtle changes. And so we were able to train a deep learning algorithm that could match at least the accuracy of experts. And here's a t-SNE plot where the individual data points are the red blood cells that are infected at different stages. So we're actually now using the deep learning model to teach us, since the experts are pretty confused about the various stages, the deep learning model is teaching us what happens in the transition of this parasite over time. Okay, so we've gone through these first two examples, and what I'm excited to tell you most about today involves profiling. So in profiling, our goal is to characterize samples even if we don't know what phenotypes we're looking for in advance. And this one I really hope that you'll pay attention to because I think it's probably applicable to— the strategy that we're taking here may be quite applicable to, I would imagine, very many of the researchers in this room, because the goal is really to just allow computational techniques to augment your ability to examine images. So you might ask, well, if we don't know what phenotype we're looking for in a given experiment, what on earth do you stain for? And so it'll become a little clearer over time, but first let me explain an assay that we tend to use. It's called— it's called cell painting, and our goal was just to stuff as many stains as possible into a single super cheap assay. So we wanted to run this at scale on hundreds of thousands of samples, so it couldn't involve antibodies or expensive reagents, nor fancy microscopes either. So we run it on a conventional high-throughput microscope, and it stains these different organelles and has been pretty robust and convenient. So that's the cell painting assay we use. And what we do, if we want to test a set of perturbations— there might be tens of thousands of them— and we treat the cells with that perturbation, we fix and stain the cells with the cell painting assay and extract as many features as possible. So instead of extracting the one thing that the biologist biologist tells us that they want to look for. We measure everything we possibly can, and it's usually more than 1,000 features, and we usually have a few thousand cells for each sample. And once we've done that, we can measure similarities. So the whole rest of my talk comes down to— can we— it comes down to the intuition of a microscopist, which is my background. Can we look at a cell population that's quite heterogeneous, and can we say this cell population looks like that one, and this one doesn't look like that one, and these cluster together and those don't. That's fundamentally the process that we're taking here, which you can do by eye to some extent, but you certainly don't want to do it at this scale. And secondly, just the human visual system can only see certain things in images. And so we're trying to maximize the capabilities of the computer and leverage them as much as possible. So all the projects I'm going to describe involve this basic concept. Can we treat cells with something and then see which samples look similar to each other? Many of these technologies are very useful to the drug discovery process, so I'll describe them kind of along this timeline. You might be aware that generally in the drug discovery process you develop some assay which you believe— if I find a drug that has the right response in this assay, I believe that's going to end up being an effective drug for the disease in humans. And then you screen a set of small molecules known as a library. Once you have a hit, something that looks If it's positive in the assay, then you develop it chemically to develop something that's a therapeutic candidate, and then eventually clinical trials and out to a drug. So how can we use imagery in this morphological profiling strategy that I've described in this process? The first— I'll just talk about a couple projects in depth. One is, can we find disease-specific signatures and then screen drugs to reverse the signature? Let's make it concrete with an example. And this example is actually not from my lab. This is from the University of Utah, but I was so excited about them taking this approach using CellProfiler software and their own assay that I now serve on their board. So the basic concept here was this lab had been studying a hereditary disorder known as CCM, and they knew that it was caused by a single gene. So you lose function of the single gene, you get the disease. And so they could replicate that in a cell model by knocking down the gene, CCM2, and it doesn't take image analysis to tell you something happened, right? So we know this gene is causative, we knock it down, we get some phenotype, and then they ask, can we screen a bunch of drugs and find any that reverse the phenotype? And what I love about this experiment is that They chose two sets of drugs. They chose a set of drugs that was selected by experts as being closest to healthy looking after treatment with the drug, and then they allowed the computer to select a set of drugs. And it turns out that the drugs chosen by automated analysis outperformed those chosen by experts who had been studying this disease for quite a while. And so even in a case that's blazingly obvious, this can be a good approach. And certainly for cases that are more subtle where we can't see a difference, it might be even more so. So they've really gone to town on this in the context of the company, and they now have looked at a huge number of monogenic loss-of-function genes so that they have a couple hundred rare disease models that bigger pharma companies can come and screen their drugs against. So that's definitely very high throughput and very highly parallelized. In my own lab, we have taken that basic idea, but in a much more restricted way. We're looking at patient samples where patients have various mental illnesses. And our goal is to follow the same kind of approach except using patient cell lines instead of genetic mutations. So the idea is to take cells from patients that have a particular mental illness and compare them to those that are well-matched but don't have a mental illness and look for any phenotypic difference. So I don't have results to show today, but we have some inklings. It's a very challenging experiment. It's statistically going to be underpowered no matter what you do, but we have some evidence that we can see differences, especially for major depression and schizophrenia. So that's the basic idea. But I would encourage you to think about this approach instead of— if you don't, you might have patient cell lines that would be interesting to just measure everything you can possibly measure. You don't have to use the cell painting assay. You can use whatever stains you have available to you to see what differences you can find. Or you may have a gene you're interested in or a compound you're interested in. Can you treat cells with those, with a perturbation of that gene or with a compound, and just be open-minded about what phenotypes you look for. It can be just a nice approach, even in basic research. The second example I want to describe involves looking at gene and allele functions by grouping together similar genes and alleles. So in this project, we were overexpressing genes in U2OS cells, and following the same procedure, we overexpressed the gene We stain it with the cell painting assay, and then we look at how similar do the profiles look. And we tested a couple of hundred genes that spanned a lot of diverse pathways. Our goal was to kind of test out the phenotypic space here, and we were really pleased with the results. First of all, I was shocked that half of the genes we tested yielded a phenotypic signature in this assay. So keep in mind, this is a single cell line at a single time point with just 6 stains imaged in 5 channels. It's a very simple assay, and yet half of the genes did something to the cells, something detectable to the cells. So we were a little nervous. Well, maybe that's high, but maybe they all do the same thing to the cells. Maybe the cells are just unhappy, and so 48% look exactly alike, and then 2% maybe have some interesting phenotype. But that wasn't the case. We saw a really nice diversity of phenotypes, as you can see there. What we saw were a lot of So when we had two clones of the same gene, two independent clones, they clustered right next to each other. Where we had different pathway members, here's BRAF and RAF1 and KRAS are all right next to each other. We had a few mutations in the experiments, so constitutively active mutations of genes fell near their wild-type counterparts. And even though in this experiment we intended it to be a pilot, we didn't expect to discover anything new, we noticed that The Hippo pathway was one of the nicest clusters that was formed here, and we saw that a gene on the opposite side had an opposite profile to this— to the Hippo pathway, and that was in the NF-kappaB pathway. We had some collaborators who were interested in this pathway and agreed to follow up on this and verify that it really was true, that there is this negative transcriptional crosstalk between these two pathways that was undiscovered, even among these relatively well-known genes that we had tested. So that was a surprise to me that an opposite profile would actually be a biologically meaningful thing, but we've seen that in a number of these different pathways where you have negative regulators of a pathway give you one profile, and you'll see the opposite profile going in the opposite direction. What's really exciting about overexpressing genes and testing their function is that it's applicable to personalized medicine. So today, if you get your tumor sequenced, Your doctor will say, okay, good news, we figured out what your mutation is, you have H574Q. But sadly, 90% of the time, roughly, they have no idea what to do with that information. So it's a case where the technology has sort of outpaced our ability to actually do anything. And so we like the idea of, will it be possible to take, for example, every Every mutation that's ever been seen in BRAF, can we overexpress each one? Can we map them relative to each other and figure out the functional impact of each mutation? So the wild-type BRAF is down here in this experiment. In the cell line, it clusters with the empty kind of negative controls. So it doesn't really have a phenotype. But when you have all these various mutations, you see various responses. And so we're really excited about the potential for this. To allow us to rapidly kind of map genes against each other. So if you're interested in this concept, it's also been done with mRNA data, and this paper came out just last year on this concept. But we're really excited about the prospect of using imagery just because of its inexpensive nature. Okay, I'll just talk about a few other examples but not go into tremendous detail. Things we can do using images as this unbiased data source: we can predict the outcome of a complex assay. Use— by correlating its readout to cell painting profiles. This was done successfully in the context of a pharmaceutical company, Janssen, over in Belgium, and in collaboration with a number of different academic and industrial collaborators. But just really beautiful work that I encourage you to take a look at this bioarchive, where they were able to predict the results in an oncology and a central nervous system assay using images from a totally unrelated imaging assay they just had lying around. And that allowed them to, instead of screening a huge library, they were able to computationally just look up what should be hits in this remaining library and just test those and found very good success rates, 60-fold and 240-fold enrichment in the hits that they were looking for. So the idea of not even having to screen things anymore because we can use images to predict the results of biological assays. It's pretty exciting. So I showed you where we can match genes versus genes. Can we also find compounds that match gene profiles, or opposed— opposite of gene profiles, so that we don't even need to make disease-specific assays anymore? So the concept here is you take this query gene, you look up a library of well-characterized compounds, and you can perhaps immediately find an inhibitor of a gene that you care about just by matching up the profiles. We have some evidence that this works. We have some stunning anecdotes, and we have some solid statistics, but we, we won't believe the results until we've actually discovered some new compounds that really are inhibitors of things. So we're in the process of doing the wet work testing for those projects. Another example application, we could create smaller, more efficient compound libraries by identifying compounds that are dissimilar from each other. So if you're thinking of testing a bunch of compounds, you don't want to test 50 of them that kind of do the same thing. And so what we do is we would run the library through— and this has been done and published— you run the library through a single imaging assay and allow that to let that select individual, just a few examples from each of the different clusters. We can also assess the function of newly synthesized compounds. So Stuart Schreiber's lab is at the Broad, and they synthesize an awful lot of compounds, and he now has a rule that if you're a chemist and you join his lab, you have to learn how to do cell culture, and you have to run your new synthetic compounds through the cell painting assay when, when they join the lab. And so every week as you create new compounds, you run them through the standard chemistry to kind of figure out what you have, but they also run it through the cell painting assay in order to look at the performance and to see how the performance of the compound changes in a very broad way as a function of the different chemical groups that they've been playing around with. So they found that very helpful to increase the cycle. Usually it takes That loop takes months for a chemist to get any information back about what has their compound done, or they don't know what assay to run to figure out what their compound does, and that speeds up the process. And then another example: can we determine what a compound does by matching it to known chemicals or to genetic profiles? And this is very well proven in the field. I won't go through any papers or details, but generally speaking, if you have a lot of compounds and you know that a that some share a mechanism of action, you'll see their profiles will cluster. And so if you have an unknown compound, a query compound, you can look it up against a library of existing chemical perturbations. But what's really exciting is to look it up against genetic perturbations as well, because that can give you a clue as to how your compound is working, which is a pretty valuable step along the pathway to getting a drug approved by regulatory agencies. So I've just skimmed through some different examples. I hope you'll find those concepts interesting. If you would like to read more about the applications, there's a review article here. If you want to learn about the best practices, the computational pipeline that tends to be used in the field, 20 labs in this field got together and wrote a paper, and it's really comprehensive about what's recommended at each stage, so you can take a look at that. There's a more popular article, and then there's a couple of conferences that are— that cover this kind of work. So I hope you've appreciated the trajectory here. I hope that you're inspired a bit to push push your image analysis beyond just the things you want to measure and maybe go a bit beyond that to, to see what might be lurking in your images that you haven't gone looking for. And with that, I would be happy to answer questions. That is how you time a talk.

Ron, I'm inspired, but I'm chagrined a little bit. Everything you showed is flat in your culture wells. And it's wonderful for what you showed, but have you pushed this to real 3D volumes, not 3D in a cell, if that's what I think you're talking about, and multiple different cells in tissues, not a single line that's just been perturbed? I realize that's not for screening purposes, but for interpretive purposes.

No, absolutely. So the basic— the concept— the profiling concept you can apply to any matrix of data, whether that data whether the data came from 2D or 3D doesn't matter. So that is certainly applicable to everything. One question is how do you get nice data out of 3D culture systems, not just 3D images of 2D culture systems. And I think the— in the high-throughput world, we're definitely a bit behind. Certainly people in this room have done a great job extracting beautiful measurements out of true 3D, even time-lapse imagery. That could be fed into a profiling pipeline just like this. But in the high-throughput world, it's a little bit behind that. Things have to be kind of 10x more robust to function in our world. So we can't really stand too much tweaking. So yeah, it is a little bit of a problem.

So to follow up on that, from that cell paint reversion assay you showed, which is really cool, How many of the drugs that you tested on 2D then fell apart when they took it further?

2D versus 3D, we haven't done any such experiments. Yeah. Yeah. So we have not done any comparisons. And I've seen only a very few quantitative comparisons between any given 2D assay versus a potential 3D assay. But it's a hot topic in the field. Is it worth it to do— to make organoids and to make liver spheroids or cerebral organoids and so on. That's definitely a very open question. It's a lot of extra work if you don't need to do it, so people want to know the answer for sure.

So at one point you mentioned that when you were doing this general profiling, you found that when you're overexpressing some proteins, they had the opposite phenotype of the other, like the Hippo pathway way with NF-κB. What's the opposite phenotype?

Yeah, that's a great question. I didn't expect this to work at all, and so let's take a super simple case, you know, overexpressing a gene makes the nuclei get bigger. Oh, look, an inhibitor of that pathway makes the nuclei get smaller. Okay, that's maybe logical, but we're looking at, you know, a 1,500-dimensional feature vector, and the cells are changing. I mean, you know, looking through the microscope, cells are changing in all kinds of ways, and So just bumping a pathway this way doesn't necessarily imply that it should look, you know, the opposite. What does opposite even mean in this phenotypic space? And so I can't— I don't know that I'm going to provide a satisfying answer to your question because the answer is I have no idea. I just know that it seems to work. I just know that the—

What was opposite?

I'm sorry, what was opposite?

What was opposite that you were looking at?

How did you define opposite?

Something in the— Opposite, it's just anti-correlated profiles. Is how we define opposite. And so we could see it in certain cases. So if in some cases overexpressing a gene caused the cells to take on a very asymmetrical appearance, and then if you looked at the opposite phenotype, you could tell that the mitochondria were really evenly dispersed around the nucleus and were very symmetrical looking, and the cells were kind of more roundish and flattish instead of elongated. So in some cases we could make a very nice biological— we could understand biologically what was meant by opposite. But I wouldn't expect that to happen all the time. In fact, I would hope that we wouldn't always be able to interpret it by eye, because I would hope that the computer is picking up more than we can. But it's definitely a lot more satisfying when we can understand what's happening in any given case.