AI and organoids - Linda G Griffith, PhD

Okay, so my actual title didn't quite make it in because I was very, so I'm going to talk about order. We use AI all the time, machine learning, various kinds necessarily ai, a lot of machine learning things. But I want to talk about upgrading the organoids. And I know Samar Baz is here somewhere and I met him here at this meeting last year and it was amazing. And we've had him up to MIT. So this meeting really does catalyze also. So I want to talk about how we can start to build these lesion models of individual patients in the lab to start studying how they might respond to drugs. And this has worked from a lot of people in the lab. And I'm showing a picture here that has Steve Palmer in it because I'll show him at the end for a particular reason. We're working with him to try to actually get stuff moved into patients.

And so we talk about this all day and a lot of this, but what I want to call your attention to is that in women, the lesion in these patients, the local lesion microenvironment stands a huge spectrum of biophysical and biophysical features because you have them superficial, very highly vascularized. You have things invading into muscle or present in muscle, maybe invading, at least growing, going through the diaphragm, the bowel, et cetera. Often fibrotic, not always fibrotic. So there's many, many different local microenvironments that we need to think about. And in any patient it may be all these lesions causing symptoms or it may be a subset. So we've got a real mess. What we'd love to do is parse this mess and bring in vitro models up to a state that we could mimic features of these lesions to test what's going on. Alright, and again, thinking about the heterogeneity lesions, here's just a couple of papers, one from DR and another from the group in Australia showing different molecular phenotypes within different regions of lesions or in different lesions within the same patient.

So on the left, these are deep infiltrating lesions showing protease activity, MMP nine at the top in the center region of a dein infiltrated lesion where the cells are still highly columer and at the front where there's more what they call the invasive front in this paper. And then proliferation as assessed by KI 67, A marker of the cell cycle. Again, where you see the MP, you also see a lot of proliferation. And here peritoneal lesions is paper focused on that recently came out. I love this paper because it shows many different types of peritoneal lesions and it's showing the CD 10 positive. Those are characteristic of stromal fibroblasts and the endometrium and smooth muscle actin activated fibroblasts in different types of peritoneal lesions. So these are somewhat concerning if we're thinking of modeling and trying to address drug that would capture all these different phenotypes, but we can do it.

The other aspect of course is there's a sibling of endometriosis adenomyosis, and it's been discussed somewhat at this meeting, never enough probably to discuss, but it's when you have features of endometrial like tissue in the myometrium, and again, very similar pathology, but with some notable differences as I know there was at least one talk that went into that here. Something that my clinical collaborator, Keith Isaacson is hugely interested in. So one of the real, real interests of my collaborator, Keith Isaacson, is particularly adenomyosis and there's a lot of debate with all of these lesions. Are they actually invading or is it just proliferation and stuff pushing out? So we'd love to study this and build models for it. And of course we first look to what's going on in vivo, and I'm showing some images here for reference from a very beautiful paper by Animoto group where they use light sheet imaging.

So you take chunks of tissue, clear the lipid out, and then you can measure a piece of tissue that's half a centimeter, centimeter thick and real block of tissue and capture a lot of the features in the tissue. And in doing this, and this is a reference F-F-P-E-H and E section, they can map where the epithelial are. So this is showing the invasion of an endometrial gland into the myometrium. So this is the invasion here, and this is the presence of ectopic endometrial gland structures within the myometrium. So beautiful work, 3D reconstructions and they have schematic size here, how the endometrium has these mycelial, what they call my rhizome like structures and the glands. And then the hypothesis is that there's an invasion process that goes in. So something is invading. We don't know much about this invasion. We found we being Juan Neco, an amazing postdoc who's now an assistant professor of Tufts invasion in the Myometrium, also using light sheet imaging.

And that paper is in prep. It shows a lot of features of nerves and vasculature in addition to just the epithelial. And I hope he gets it published because it's an amazing set of images. So what we want to do in the lab right now, my role in this whole ecosystem of trying to unravel the mysteries is to build the kind of tools that would let us parse individual contributions to disease phenotype and then let us test mechanism and test potential therapeutics and do that in ways that might be specific for different patients. There's a lot of interest now and whether cancer driver genes may be somatic mutations giving rise to cancer driver genes may be important. We don't know if they are, if they're not, but in vitro models that can capture more of the in vivo physiology could be useful. So we think about recapitulating, biochemical and biophysical features of the lesion environment.

And I'm going to talk in particular about some synthetic extracellular matrix. Everything I've done here is in a standard organoid culture media, and I know Samir will have others. All the epithelial are actually from ectopic tissue, although we do make them from lesions. And most of the data I'm going to show except for some of the end is it happens to be just epithelia. Although we do an extensive amount of co-culture work. Okay, so the first thing that you do when you make organoids, you get a sample from the patient, you dissect out and make glands and stroma process to single cells and you culture it in something called matrigel with an extract of a tumor grown in mice. So it's a very messy reagent, but it's sort of this magic elixir that was identified first for gut organoids and now used for endometrium. And it's a very unsatisfying reagent because it comes from a tumor and has a lot of TGF beta.

TGF beta is of course involved in signaling between stroma and epithelia and responsive progesterone. It varies from lots a lot. It doesn't support co cultures and on and on and on for why it's a love hate relationship that everyone who works with it has. So what we did over about a decade actually and inspired in part by Kevin Osteen and Steve Palmer is completely replaced matrigel with a synthetic matrix, meaning every component of this is made in a lab with no animal components ever entering in it. Everything is done by som phase peptide synthesis or polymer synthesis. And it's very simple. We have a precursor that has some peptides that interact with cells. We cross-link it with a peptide, you mix these together with cells and get a beautiful clear gel. And it took us about 10 years to make this because the devil's in the details.

There's a lot of engineering design I will skip except I do want to show for those who care about cell biology, some facets. The key features are we've got these tethers of a synthetic polymer polyethylene oxide that interact via ligands for RINs on the surface of cells. When the cells are first put into this matrix, they need some signals that say it's okay, you have an adhesion, you have an anchor. So then the cells will produce their own extracellular matrix. I'm showing here fibrin, which would be characteristic of a stromal cell, and they will assemble that. And what the matrix that we named does is it captures that secreted matrix. And so we have a number of different adhesion ligands. The most common one we use in the lab has two. One for RGD binding ligands that has an extra site and one for collagen binding immigrants.

And then there's unpublished work we have with other ligands for endothelia and neurons. I can talk to people about at dinner tonight if they want. And it's very important to have the right concentration. Integrin clustering is really important. And then we have these matrix finding peptides. So we published an enormous amount on this and on the principles of the design, and again, I can describe to you, but in Juan Echos postdoc work, he showed that you could support co-culture of endometrial epi, organoids and stroma and take them through a simulated menstrual cycle by adding progesterone and estrogen. And then actually at the end you could add rizone and block estrogen signaling and simulate menses. And so features morphologic features biochemical and other features in these co cultures. The red are stromal cells. This is an acton stain for the apical side of the epithelial. So the morphological features you can see here, and there were a lot of characterizations in this particular paper regarding production of various endometrial specific molecules in inflammation and in a non-disease kind of state.

So what we've started to been doing lately is trying to get a lot larger and start to look at differences between patients with certain gynecology disorders and donors who are not diseased in that way. They may have fibroids or something or controls controlled donors. And we do use a lot of AI for this. So one thing we're doing now, and I'm not going to show you an enormous amount of data, just to give you a sense, is we grow these in droplets. So there's just a little droplet sitting in a culture dish, and we can follow the growth over here. Seven days of culture going from little tiny, we start with individual cells and then even by day one we see some organoids. And all the blue here are identified by an image analysis system as an organoid. Okay, so we can follow the size of the organoids and other features because we're capturing an actual 3D image we scan through.

And then this is comparing two types of gels we want to compare. I mentioned we want to mimic features of a soft microenvironment, so a biophysical that may be comparable to the endometrium, but we know that fibrotic lesions or lesions and muscle have a very stiff microenvironment. So we can ask questions. Are there differences in the growth of these lesions like these organoids in different biophysical microenvironments? So here I'm just showing how we would grow these, and this is unpublished data from a couple of postdocs in the lab. And if we compare a controlled donor to an endometriosis owner, and this is just an end of one right now, this is using automated image analysis to characterize the average number of organoids. We can characterize other features that actually have similar curves in those other metrics like size and so on. And this is looking at two different experiments done at different times.

We can see that for the controlled donor there seems to be a faster growth in stiff matrices, whereas for the endometriosis donor they grow faster, but they're comparable in both matrices. So there may be some endogenous just I'm on all the time signal and we don't know that yet. We need a lot more donors. But this is giving you an idea of using automated image analysis. And we can further go in and we can do staining for things that we can benchmark against InVivo. For example, KI 67, as I showed you at the beginning of lesions, a marker of proliferation we can stain for this proliferation marker and for dpi. And then we can go in and again, up here was the endometriosis patient, the organoid growth, and here's the KI 67 for this particular donor. And it's not significant for KI 67 and you wouldn't expect it to be because these look the same there.

Okay? So these are the types of things we can do and compare for example, the endometriosis and control and compare against matrigel. So we're now starting to get these use, I guess you could call this ai. It is AI because it's image analysis to get these kind of features out of our donut. The other facet of this is to start observing the emergence of these. What happens to these organoids when they are encountering an interface? One hypothesis about the growth of some of these lesion or the invasion is this is an actual mechanical fracturing or there's some kind of invasion into tissue planes that are in the muscle. And so we've also been characterizing the biophysical environment for generating what looked like endometrial gland lumen structures in these hydrogels. So if you look at a droplet and you look at day seven, you can see that there's this green haze, and that's because organoids have fused and grown along the surface to make a gland lumen structure.

So we're now exploiting this to make features that look like endometrium. If we image, you can see the beautiful luminal epithelium characterized by apical actin. And if we image further down, these cells have deposited their own basement membrane characterized by laminate and have a beautiful polarity. So if we now go in and compare what happens in a soft and stiff matrix to these organoids in the interior and the top of the matrix, these are cells in the interior in a soft matrix and they will emerge out. But in a stiff matrix, you can see they take on that hugely, hugely invasive phenotype. And in fact, if we now go in and characterize what happens to the endometrial organoids growing in a soft matrix, they have this pulsing structure likely due to the hydrostatic pressure generated by fluid transporters that then are regulating the hydrostatic pressure balancing with growth.

But now if we go in and look at cells in a stiff matrix, you could see that they have a very different morphology. It looks more like a lesion, sort of this correlated feature. These are proliferating and now you see them invade out. So we can start to use these kinds of systems to ask questions about invasions. So I'm giving you the initial, but I am someone, you always worry that what you're seeing with the stiffness is due to a particular feature of a matrix that's simply correlating with stiffness and not the true molecular structure. So we asked, this is actually my Halloween pumpkin. I really, really worry in everything that we're inferring causation from correlation. So stiffness is one feature. So what we've done recently is go in and make a completely orthogonal way to create gels that have the ability to support organoids and stroma, but also are cross-linked with hyaluronic instead of a peptide cross-linked with hyaluronic acid.

Hyaluronic acid is a charged polysaccharide that's present in the endometrium naturally and engages a receptor called CD 44. And so we can make a gel that's very beautiful, and we find that indeed we can support. So over on the left, here's our peg peptide gels. Again, cells are very beautifully polarized in a stiff gel. They take on this invasive phenotype, but this is relatively new data. If we look at them in a hyaluronic acid hydrogel, and this is at day 14, soft matrices, they look fairly similar, but in the stiff matrices, they're not as correlated, although they do have some morphological changes. So starting to really map out our molecular phenotypes. One very interesting feature though in the peg ha gels, if we look at our in vivo light sheet microscopy images of adenomyosis lesions. So this is an interior of an adenomyosis lesion, and what you see are these little tiny gland like structures.

This scale bar is a thousand micro current. And over here we're seeing little tiny structures also in the interior of these lesions in organoids, in the peg hydrogel. So what I've tried to show you today is we're building up a compendium of tools still in the process of characterizing them, but optimistic, we can capture features not only a biochemical but biophysical space. We get influences on morphology capture features that we see in vivo of morphology. There's donor variation in response to the stiffness. We can get some invasion phenotype, but we want to further characterize this in orthogonal systems. There's a lot that we're doing with media studies, co cultures, and putting organoids from lesions in these matrices and trying to then apply these to specific types of, for example, organoids with different cancer driver mutations, perhaps in collaboration with Samir. I'll just show a couple of slides at the end.

Those are all static cultures in a Petri dish. But of course, we would love to model these nascent little lesions interacting with blood vessels because this is likely what we're going to be targeting with drugs. How do we really observe what's going on in a lesion when it's really almost in situ? So for that, we're adapting technology developed by my fabulous colleague Roger Cam. We have a joint student, Ellen K, this is NIH supported work. And he developed a way in a microfluidic device that has a fluid channel here, a fluid channel here, and a central region pinned off by these posts which appear here in black. If you inject into that central region of fibrin gel precursor that has vascular cells, they will self-assemble and form per fusible vessels over four days. And what you're seeing here are monocytes that are going through from one side to the other because these vessels are completely fusible.

And this is actually a commercial chip. You could go do this at home if you want it. It's actually quite easy to do. What we've done is we've adapted our synthetic hydrogel to support formation of these vessels. So here in green are vessels, and this is unfortunately not maybe the best. We don't have a video here, but what we've done in this experiment is injected along with the vascular cells, we've injected organoids and stroma, and what we see are three proto lesions developing, and this is day five. So we're now in the process of starting to use this to evaluate how some drugs may interact. Oh, it wasn't the final one, sorry. I was going to mention that we're interacting with Steve. What we're doing now, we have a project with Steve Palmer who's here, and I have wanted for 10 years to work together to evaluate June kinase inhibitors ever since we identified June kinase as a potential target independently of his own speculation that it would be a fabulous target for endometriosis.

So we're now just starting to work together, but overall, I think the field needs a lot of boost to get more interest, not just in endometriosis, but in the overall processes of menstruation. So later this year, we're going to start a moonshot for menstruation science, and you'll get emails if you sign up on our email website. And we parallel that with breast cancer because breast cancer 40 years ago was completely not talked about. If we start talking about menstruation, it'll bleed over. Well, sorry, the bad word. It will spill over into all these other facets of gynecology. So with that, I'll end my talk and say thanks to my many collaborators on this work, including Dave Trumper on the devices and my husband Doug Berger on the systems biology, and to the new collaborators, Steve Trump, and of course Keith Isaacson and his partners for the clinical collaboration.