In this session, Keith Murphy, founder of Organovo and CEO and founder of Viscient Biosciences, covered the recent developments in the field of 3D tissue bioprinting and human organ models on a chip or sometimes also called patient trials in a dish. He went into the different technologies that are being used and trialed, applications and markets that are going to be the first affected by these innovations, and also tissue examples that are currently possible to develop. He also provided a good overview of the regulatory hurdles that the field will have to go through with some examples of companies and organizations paving the way.


Presentation: Keith Murphy

  • I wanted to try and start with giving everyone a picture of some of the technology, big picture issues in the space, and walk through some of the ways we use these technologies.
  • Broadly speaking, I’m going to be speaking about 3D human cellular models. And there’s a lot of different ways one can interpret that – some people call it complex in vitro models, others call it human models in a dish. And then other people compare it to organs on a chip, which is actually a little bit different, because those are more of a 2.5D kind of thing. Then there’s also pure 2D which are essentially cells on a plate, the traditional monolayer culture as people are familiar with it, but you can get slightly more complex by having them in specific positions on a plate next to multiple different cell types. The 2.5D are going to be things with more complexity, like what Emulate is doing (a company out of Boston), they have an organ on a chip with a membrane that’s flexible with pressure on either side and your channels on either side, you put one cell type on one side, one cell type on the other, and that creates this kind of a 2.5D system because you get cell signaling across that membrane. But I’m gonna focus on the 3D stuff, because that’s what I know best.
  • What do you need for any of these more complex human models? What do you need to have to make things work?
  • Well, you need to start with cellular input – what you want to actually work with – and that can be many different things. Stem cells are a huge source of cells for these things – IPS stem cells. Human primary cells are also an important potential source. You can use any of those, but you have to think about what problem you’re trying to solve or what you’re trying to shed light on.
  • Then the fabrication platform is important. That’s 3D bioprinting, or a couple different options that I’m going to show later, including spheroids, organoids, things like that. Those obviously are the way you get somewhere. Where do you get to, what are the functional outcomes?
  • You need to be able to measure and see the benefit of what you’re actually working with. And also, you have to think about the fact that for some functional outcomes of disease, you might need a 3D model. So one of the reasons 3D models can be powerful is that if you think about 2D cells on a dish, when you’re trying to work with a disease like fibrosis, where collagen fibers twist between the cells in three dimensions and form a 3D fibrous network, that’s not something that happens on a 2D surface, and you certainly can’t see it on a 2D surface. So you have to step into 3D only to see if you’re doing anything in terms of fibrosis and changing it. So 3D can be important in terms of the functional outcome and what you can measure and how you’re going to measure it. But you have to think about all of that, and build assays to actually figure out what’s going on.
  • So the cell sources are very important. The good thing is that there’s a tremendous amount of strong work being done in all these areas and IPS cells really started to move quickly. I would say in general in 3D tissue generation and model use, there is a holdover from historical pharma’s desire to have highly reproducible results. One of the things pharma scientists like about working with animal models is that they can get similar results each time and have a positive outcome. So that’s one reason people strive for use of iPS cells, because if you take iPS cells, you can make up a bulk, and then work with the same bulk. iPS cells, in case someone’s not familiar, are a way to take a skin or other somatic cell from an individual and turn back the clock on it to allow it to become nearly omnipotent or potent stem cell, or pluripotent stem cell. Some people call it induced pluripotency. One of the challenges is that differentiation (to that state or pulling things back first to that state, and then differentiation into different states) is not as fully reproducible as people would like it to be. There are challenges with getting the same results every time, although that is the promise of iPS cells versus stem cells from humans or primary cells, which are starting with a donor source and therefore have inherent variability. All this stands in contrast to what I would call cell lines, which are the more historic standard (Caco-2, Henrietta Lacks’ cell line, as the traditional example). A lot of times what we’re comparing to when we build a 3D model is one of those former gold standards to compare to what people had beforehand.
  • One concept I think is important to understand is the cellular density of these models. So obviously, tissue in most cases, has a very dense set of cells in it. So in figure (d) you see the 100% cellular matrix that’s going to represent most of your soft tissues. Interestingly, if you’re talking about cartilage or material like that, that actually looks a lot more like figure (c) with various biomaterials forming the basis of it, and then some living codocytes that helped to replenish it and repair it. But most tissues and certainly the stuff I work with is all more like (d). So highly dense tissues in terms of the cell positions, and how many cells there are in a volume.
  • And this is actually an area that’s very important to understand, because people use different techniques, and they don’t always work with the most dense cells. So sometimes people start with gel cell mixes that look almost like (b) – like 10% volume of cells in something. Things like organoids started with that in the early days, basically asking: Can we just grow cells in 3D culture, starting with the surrounding gel material and see what happens? Will the cells behave better just because they are in 3D? They did get some good results, and it showed that they can grow, take over and be 3D even if you start with that. And you can get interesting and good early results, so people still use them.
  • In the context of solving some deeper biology questions, nowadays you start as close to 100% cellular as you can, because in reality, in the soft tissue of interest, you want the cells to be interacting with one another. So you’re going from a 2D model where cells interact with plastic, and that dominates their shape and their behavior because that material interaction is very different and foreign to them, to a 100% cellular environment, where if you compare their native environment in the body with what they end up with in the dish, there are cells all the way around them. And that enables them to act the most normal because they’re getting the mechanical signals from those other cells that are correct, they’re getting the cell signaling, and the biochemical signaling is going to be very comparable.
  • And so you’re solving a different problem if you start to work that way. Historically, in terms of animals, we’re engineering a disease model in an animal. Now we ask: Well, how do we actually just find a way to replicate the conditions in the body, so the cells will act exactly as they do in the body? Where the biology will be inherent, versus a petri dish where they’re interacting with plastic and that alone strips them down a bad path for many cell types.
  • I’m gonna use some terms that are not settled fully by folks. Organoids is a term that can mean many things, and organoids and spheroids can be used interchangeably. So it’s a little bit hard to say that there’s any final answer on any of this in terms of the terminology, but I tend to think of organoids as taking a gel material and working with those lower level of cell densities where you’re surrounding it, usually with something like a Matrigel, and often the choice of the biomaterial that you use can make a big difference. So you have to do a lot of bioengineering work that goes into the choice of materials. But Matrigel is basically an animal derived extracellular matrix. It’s a bit unpredictable in terms of what signaling factors are in there because it’s coming from a living source. But essentially, in an organized setup, you’re going to get a lower density in terms of the number of cells the cell density can then grow and you can get something that’s much denser over time.
  • There’s great work being done on this. There’s a guy named Hans Cleavers out of the Netherlands, University of Utrecht, who is leading the way in this. He actually became a member of the Board of Directors at Roche, because he does such great work. He’s working mainly on guts, but has expanded that work to IPS sourced intestinal models. This work is really going forward and you see valuable comparisons to human biology when you move things into these models. And in the case of the gut, that’s going to be seeing the formation of villi and the cellular lining correctly.
  • Now another way to work is to start with spheroids. The distinguishing factor is that they are going to be formed by a hanging drop method, where you put a small drop of cell suspension on a cover slip and hang it upside down, allow that to serve as a sort of a concave bowl, where the spheres, the cells, will actually aggregate at the bottom and form a solid sphere of cells. Spheroids tend to be solid – 100% cellular. More recently, companies like Corning Biosciences have introduced ultra low attachment multiwell plates, where you put your cell suspension in, and then the cells self assemble at the bottom of that into a spheroidal formation. You can get some really good results out of these.
  • And then the challenge, what you can reach for at the next level, is that you don’t have a lot of ability to create structure or morphology inside of these. Typically, tissue has not fully mixed sets of cells – they form in regions – and there are structures inside the tissue next to one another. And that’s what 3d bioprinting can give you.
  • So 3D bioprinting is a way to take the same materials – gels and cells – in the form of a bio ink that you can place on a surface in high resolution and layer by layer build up a structure, where you can position the cell types in specific positions. You can make things quite large with this, certainly up to a millimeter in thickness. If you combine it with biomaterial scaffolding, you can potentially go much thicker, but gravity starts to become the biggest issue.
  • In the sense of the technologies I’ve worked with in the Organovo space bio printing platform, we’ve tended to work in the 300 micron to one millimeter thickness range, because that’s supportable with our techniques, and with a high density cell suspension in gel – so the bioink is 90+ % cellular. And the whole point is to actually make the gel go away as quickly as possible and allow the cells to survive and adhere together as a tissue.
  • And that’s the platform real quick to show you that.
  • With all these tools, what are the things you can actually do with them? You’ve got these models you’re building and the different tissues. The things that we tend to focus on all fall into these categories on the slide. Preclinical safety, so toxicity testing, you use normal tissues for that. And then efficacy and disease modeling, taking normal tissues, and you can create a disease tissue using the normal tissue. For creating a disease, that would be akin to taking an animal and inducing disease in the animal. So if you’re trying to create fibrosis, you can induce that chemically in the model. And then in a diseased tissue setting, you can also think about starting with disease patient cells. So primary cells that are diseased, or in the case of a genetic disease, IPS, or primary cells that come from an individual with that genetic deficiency or that genetic condition – it should actually express your disease appropriately if you start with those cells. And that can include IPS, obviously. And then we’ve also looked at therapeutic opportunities. I won’t talk about transplant tissues today, although that’s a rich area of bioprinting work for sure. Also drugs – once you have the disease tissue, then you’re moving towards building an actual new therapy that doesn’t actually have to be the term drug in the FDA terminology, so this actually applies to cell therapies and other things as well. So any new therapy can be tested on these models as long as they’re appropriate for the testing you’re trying to do.
  • This is an example of how to build a 3D tissue, and what you have to consider. This is essentially the structure of a liver tissue, you put four different cell types into this structure we’re building here – hepatocytes and stellate cells, which together are the bulk of the overall liver, have to touch each other at interfaces, and that’s important for the function of the stellate cells. And so we create regions of these cells, not so much that we’re reproducing the exact liver biology at the microarchitecture level, but that we actually get these cells interacting in space more or less like they do in the body. As things improve, people can get better and better and actually start to more accurately reproduce tissues. People at Rice University or in companies like Volumetric Bio are working on technologies to do that better and better.
  • But essentially, if you look at what’s been done with the original Organovo liver platform, you have a region inside the tissue, where these two things are interacting in space, and that’s what gives it some of the characteristic biology of the liver. In the 24 plate tissue array you can see a tissue basically three millimeters across and half a millimeter thick, and has about 1 million total cells in it. And we tend to play with the ratios of cells, but the first place to start is to ask: What is the ratio of cells in the native tissue that you’re trying to reproduce? In this case, it is 60%, plus hepatocytes.
  • To show what this has built, you can look at the various characteristics of liver tissue. For example E-Cad+ is indicative of tight junction formation between cells, it’s exactly what you’re never going to see in a dish, in a 2d model. These cells are all interacting in a cuboidal way, which is another way to say they’re happy and doing what they should. That’s when they express E-Cad+ and that’s what you see happening throughout this tissue, a really important aspect of the biology underneath. And because we include stellate cells, in the upper left you see actual formation of lumen. So the vascular structures, small vascular structures, that are basically proto capillaries to form the red CD31, are surrounding those hollow structures that form and would eventually be able to carry blood, but mostly just serve as passive diffusion routes that make the issue able to be a little larger.
  • And what can you show if you achieve normal biology for something like this? This is a drug that I’m showing here that was not seen as toxic and went into human trials, was still not seen as toxic in humans, because they did a small number of humans in the initial trials. But by the time it got released into millions of people, it was toxic, and had to be pulled from the market. So unforeseen liver toxicity, basically a classic case of missed toxicity not seen in animals as things move forward, because you can test a lot of animals, and then a small number of humans, and you might miss things. So what this is basically showing is that, this toxicity would have been clearly seen in these models had they existed at the time. This is really the kind of thing you want to show.
  • And we did a whole set of these, with our ExVive Liver tissue to show that we could be much better at predicting toxicity than 2D hepatocytes or rat models. And part of that was showing that for a set of these classic case drugs that were missed and got on the market, that they clearly showed as toxic in our model when compared to human results.
  • Regulatory issues I wanted to speak to here. There’s a lot of work ongoing to move these things forward and use them for toxicology. There are companies like Insphero that make spheroids and work very hard on using those for tox modeling. And then there are academic and industry consortia that are geared towards trying to find the best solutions and build sort of standards. The standards don’t exist yet, but they’re working towards that.
  • EMA has been very busy and supported a lot of work, including with Insphero, which is overseas, and the EPA in the US has been very active on this, just for predicting the toxicity of industrial chemicals better. They’re very focused on that, they need higher throughput, cheaper solutions. But nonetheless, they have a strong interest in being better at predicting before they approve industrial chemicals. And so they’re very keen on this and putting together platforms and better ways to do it. And then the FDA has what I would call a “Predictive Toxicology Roadmap”, that includes these kinds of models that are working towards their goals. One thing to think about is that this is really one leg of a stool. There’s a lot of predictive work being done in silico, and a lot of other things that they try to use to be more predictive in toxicology. So what I think of these models as is actually avoiding the “garbage in garbage out” problems of in silico models, because that’s where they have a big problem right now. You still have a black box in terms of the human where you can’t section and stain as easily as you can using these small models. And so you get a very loose understanding of what’s going on in a person in clinical trials. Whereas in the human model, you can really look at things. And so I think that’s going to feed even better, predictive toxicology overall through in silico use of the data and better prediction because of that.
  • And then overall, everyone’s moving away from animal research. Europe’s been pushing faster than the US. But I think you’re seeing a trend. Personally, I think with things that exist today, it’s almost unethical to put something into humans without having tested them. Because even if it’s only 5% of the time, you’re going to see some things that are surprising in these models, and you should be investigating them. That’s not the case today, the FDA still doesn’t have any sort of requirements. And I think the real way to push towards that is actually engaging at a legislative level to push for that more and be more demanding. That’s something I don’t do day to day, but I think it’s something that will happen long term. I just don’t work in that space right now.
  • We segwayed a little bit into drug discovery. And one of the things that gave me confidence is what I mentioned about methotrexate. One of Organovo’s partners is the University of North Carolina Chapel Hill, where they are working just on exposing the standard liver model I showed you to methotrexate and seeing really wonderful and characteristic patterns of fibrosis forming in there. So there’s a clinical sample image down here on the lower right, and the bridging fibrosis, these long tendrils across space with specific nodes in some areas, that’s very characteristic of this kind of methotrexate induced fibrosis. So methotrexate is a drug used for rheumatoid arthritis. It’s used chronically by a lot of people, and a fair number of them get liver fibrosis over time. And this model was able to really replicate that in a high fidelity way. That gave me confidence that it could actually be used in this area to do drug discovery, because we’re seeing such good biology, both in the tox setting, but specifically in that replication of that methotrexate induced disease.
  • That’s actually where Viscient Biosciences was formed to go after liver work in disease modeling. I see the biggest potential in crossing that gap from where you have a lead drug and you’re picking one of 6 to 10 drugs at that point, and you’re going to put it into animal tox studies. But then eventually, that’s the same drug that’s going to go into human clinical trials. At that stage pharma doesn’t have a lot of information about what the best drug is. In fact, they’ve already made some decisions based on needing to make it good for the animal as well, just so they can continue to study it. If your drug is perfect for humans, but didn’t work in animals, you couldn’t actually use it under the old models because you don’t have any way to see it as you advance it. And so there’s a lot of missed opportunity in those 6 to 10 molecules you’re choosing from. That’s a big part of the problem with the translation into success in the clinic, because we only have, I think the last review that just came out by the industry group Bio, said that the success rate for clinical trials and drugs in the past 10 years was 8%. So if you put a drug into human clinical trials, 8% of those drugs eventually got marketed. That’s a huge failure rate, the biggest driver is the animal to human species gap. It’s not the only problem, but it’s one of the biggest drivers. And that’s where these models can be really useful, you’re going to see better prediction of the eventual result.
  • And how do you do that? You can actually use it the same way as you use animal models today. You’re going to have a disease model and a mouse that’s engineered with the disease, or a knockout mouse. And you’re going to somehow cure that by blocking the gene, and that mouse then kind of validates: “Hey, we have a gene of interest to go after with a drug.” Then you’re going to go into cells in a dish in the case of traditional screening and find a way to just knock down that target and show that you’re knocking down that gene. And then pick the best chemistries that do that, go back into animals and show that the chemistry you picked does that in the animals, and then you’re off to the races. You can do the same thing in parallel with any bio printed or similar tissue, you just go through the same way and you’re hopefully seeing a better result, because you’re using human cells and the human context. One caveat is that in the stage of screening, you can’t always use bio printed models – they’re not high throughput enough. And so I’m showing spheroids here, because that’s something people have used to do some of the screening work, you get up a little bit high throughput, just like the cells are a step-in to get higher throughput for the animals. Same thing here, same concepts, you’re trading off a little complexity for throughput, and then coming back and getting the final answer in your original model.
  • Here’s an example of a high fidelity bioprinted model of liver fibrosis. This fibrosis is NASH, which is non alcoholic steatohepatitis – alcoholics always get fatty liver over time, and now we see it in patients who are not alcoholic, but are obese. And as they get deposits of fat in their liver, their liver starts to build collagen, maybe to offset the softness that the liver is getting at that point. And we’ve actually been able to reproduce this disease using the cells from patients that have it. When their cells become available for research use, we can build the same type of liver model I showed you with some tweaks that personalize it and make the disease come out.
  • And we get exquisite patterns of the disease. So the right side here is a clinical sample, the left is one of our bio printed samples. And the characteristics of this disease are fat deposits forming in the liver over time, and then the fibrosis, the blue collagen formation. So we get really clear patterns of disease.
  • We also go and look at the gene expression level to see if these match up. And we have high fidelity comparison between the gene expression of the human liver disease versus healthy and what’s upregulated in orange or downregulated in green.
  • Same thing for the healthy versus disease bioprinted model on the top row there, what’s upregulated as orange and what’s downregulated in green. You see these gene sets that are important in this disease, really high fidelity of the comparison here – only a couple misses in gene sets. And so that allowed us then to say: “Okay, we’re taking this to pharma.”
  • And also we test the final step (like methotrexate and matching that) – if you take other drugs and put them into this disease model, could you match what was done in the clinic with these, and that’s been achieved now as well. These are all targets for NASH that had been taken to the clinic, and had phase 3 outcomes, or in one case phase 2, because that was a failure. And taking the animal model prediction on the left. So almost all of these obviously were predicted to be positive based on the animal model, they would have an effect of down regulating the disease. So the right side of these are being predicted positively. Otherwise, you wouldn’t have taken it to human studies and phase 3 studies. But they almost all failed. There’s one exception to that, but they almost all failed. So that lower right quadrant on the, on the left is your animal model saying it would work, but it did not work. On the right our model said those would not work, and they didn’t work in the clinic eventually. But we also correctly predict the one drug that’s in development that does work – FXR, which intercepts beta cholic acid. And we see other compounds that clearly do work in our model, do have positive phase 3 data, although they’re still working towards approval. 
  • And interestingly, that was not necessarily advanced by animal model results, they had a lot of human data, by the time they started a clinical program with that drug because it was developed in a different indication. So it was driven by the fact that they had really, really interesting human data, which is, of course, the most predictive you can get. They were developing it in a different liver disease.
  • Okay, and then where do we go with that? Well, then we validate targets ourselves. So we show that if we block any of these genes, we actually reduce the amount of collagen in those models. And that becomes a validated target – the same way you would block them in animals or do a knockout and block. And then if you see disease go down with the gene dit, that becomes a validated target that you can go after. And that’s what we’ve done with these models to see the same thing. So Viscient right now is focused on moving forward to find drugs to these specific validated targets, and also partnering on these targets with pharma companies.
  • Okay, so regulatory issues on this drug development side. One of the big challenges here is that there’s not really any precedent yet, even though we’ve had successes. So there’s been one big success, but there’s no sort of standard way to do this. There’s a company called Vertex Pharmaceuticals that has built its entire cystic fibrosis franchise on 3D cell culture models starting way back in 2001 when they got going, but that didn’t become a template for the industry. Actually even Vertex’s scientists can’t reproduce new models, they haven’t focused on that in 20 years. So there’s not really a clear path for anything, but the FDA is extremely scientific and collaborative. And they said: “Hey, if your data are supportive, we’re going to support you. But you may need to do certain things, you can’t just tell us that animal models don’t work, we need some data to show that – you can’t just assert it. Let’s see what you can get to and what are the roadblocks.” And then you may need to show them probably that even if blocking your target in animals wouldn’t result in the animal having a disease cured,you still have to show that you knocked that gene down, that drug can get through the animal’s gut, actually get into the bloodstream, block that target and knock down the protein level. Even if that’s not going to result in a cure in the animal. They want to see that some of those things can happen with a drug.
  • Okay, so regulatory issues on this drug development side. One of the big challenges here is that there’s not really any precedent yet, even though we’ve had successes. So there’s been one big success, but there’s no sort of standard way to do this. There’s a company called Vertex Pharmaceuticals that has built its entire cystic fibrosis franchise on 3D cell culture models starting way back in 2001 when they got going, but that didn’t become a template for the industry. Actually even Vertex’s scientists can’t reproduce new models, they haven’t focused on that in 20 years. So there’s not really a clear path for anything, but the FDA is extremely scientific and collaborative. And they said: “Hey, if your data are supportive, we’re going to support you. But you may need to do certain things, you can’t just tell us that animal models don’t work, we need some data to show that – you can’t just assert it. Let’s see what you can get to and what are the roadblocks.” And then you may need to show them probably that even if blocking your target in animals wouldn’t result in the animal having a disease cured,you still have to show that you knocked that gene down, that drug can get through the animal’s gut, actually get into the bloodstream, block that target and knock down the protein level. Even if that’s not going to result in a cure in the animal. They want to see that some of those things can happen with a drug.


Have you given any thoughts to how these models could be uniquely applicable to study and development of longevity therapeutics? Like parabiosis, epigenetic reprogramming, etc? Is anybody using your systems or systems like these to understand how these novel techniques work?

  • There is a lot that can be done, but just a few initial thoughts. For example removing things that are building up in the space between the cells – that cannot really be replicated in different models. Or stem cell population and reservoir, how to study stem cell differentiation in true biology context, unless you have something 3D it’s a pretty big challenge. Or if you think about aging, then that can be partly described as depletion of stem cell availability, so you could study how to maintain stem cell reservoirs, have them multiply in situ better in the 3D model than you could otherwise. So there are highly relevant applications but I haven’t seen anyone doing it yet.


I presume a lot of the value of human trials is that the drug is getting exposed to a wide variety of genomes and phenotypes — thus if, say, 2% of people have an adverse reaction for whatever reason, you discover this. I presume you’d want to do something similar in organoids where you have a wide variety of genomes available (gender, race,…), and run them in parallel to get this data early. Is that possible?

  • If you start with IPS cells or primary cells, you certainly should be able to do “mini clinical trials in a dish. At a stage where 6-10 drugs are candidates to move through animal safety testing into trials, that’s the primary way that the models can help with this cost benefit analysis. Use cells of 20 patients to understand outcomes for 100 patients. From a stats perspective, you could predict 30% failures, it might even turn into biomarkers, and then you can restrict the clinical trials only for the people that his would work in, so not only the clinical trial would get smaller by 30%, but you also get 30% of zeros from the results, so you could potentially get much more efficient with the trials.


Have you ever tried to include immune cells in your 3D models, e.g. tissue resident macrophages, etc? What happens when you put those in?

  • The gaps in the gene expression profiling with NASH liver are all immune related, so that is basically what you are talking about. We see a clear gap in that.


Is it difficult to do?

  • It’s not especially difficult actually, it’s more about the trade off choices you have to make and focusing on the more major things first.


How far are we from a whole (or most of a) body on a chip? Is there a roadmap to that?

  • There are attempts to create pathways to that. There are some really cool things that are being done to integrate multiple cell types – I think it’s UCF where they have done some work to show that if you expose drugs that are known to be metabolized in the liver but then cardiotoxic, but the cardiotoxicity was unforeseen because it’s the first pass metabolism that turns it into the cardiotoxic metabolite, they made chips that allowed them to expose to the liver and then to the cardiac cells. And they saw that they were able to see the toxicity. So there are integration efforts even from just academia, but also in the more top down way DARPA has created challenges to integrate 10 tissues for example and those are ongoing. It’s hard to do that because we are nowhere near perfect in any individual tissues, so to get to 10 is a big leap, almost too daunting for some of the participants in the challenges, but you don’t get anywhere if you don’t shoot for the moon. 



Seminar summary by Bolek Kerous.