Resources for Computer Vision Teams:

LinkedIn – Connect with Heather.
Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.
Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.

Transcript:

[INTRODUCTION]

[0:00:03] HC: Welcome to Impact AI. Brought to you by Pixel Scientia Labs. I’m your host, Heather Couture. On this podcast, I interview innovators and entrepreneurs about building a mission-driven, machine-learning-powered company. If you like what you hear, please subscribe to my newsletter to be notified about new episodes. Plus, follow the latest research in computer vision for people in planetary health. You can sign up at pixelscientia.com/newsletter.

[EPISODE]

[0:00:34] HC: Today, I’m joined by guest Andrei Georgescu, CEO and co-founder of Vivodyne, to talk about drug discovery. Andrei, welcome to the show.

[0:00:42] AG: Hey, thanks so much for having me, Heather.

[0:00:45] HC: Andrei, could you share a bit about your background and how that led you to create Vivodyne?

[0:00:49] AG: Sure. Sure. So, my background focused really heavily first on the high-throughput integration of microfluidics for lab-on-and-ship applications. We were looking at, could we do biosensing of different pathogens? Can we use the high-throughput chemistry possible with microfluidics to produce all your nucleotides of many different sequences? Those interests shifted very much to exploring the depth of human physiology by growing human tissues within microfluidics during my Ph.D at Penn with my mentor, Dan Huh, who I then spun off Vivodyne with.

But the scope within the work at Penn was, can we take this conceptually novel approach to recapitulating the complexity of human biology in microfluidics and can we scale it up to produce data at large scale? The interest really stemmed from this feeling that, scientifically and at the edge of what is known in biology, we have this great ability to go down the rabbit holes of complexity in a very narrow fashion, right? We can pick our favorite receptors and proteins and transmembrane proteins. We can do cryo-EM and we can understand their structures and binding sites and how these structures change when they bind certain ligands. We can pattern antibodies across a sheet of glass and look at their affinity to binding some compound. And we can go very deep on the individual function of a certain receptor or a protein. But the integrative knowledge of understanding what happens if I modulate this thing? What happens to everything else in a cell or in a tissue or in a person? That is a very difficult level of understanding for us to have in biology currently. It’s very, very difficult to make headway on that, that kind of integrative understanding.

My kind of deep interest became, how can we scale up a process where we can learn with the effects of these modulations and this dial turning really are on the scale of very many human tissue substrates, where we can learn at the complex level of physiology how very small changes to signaling within cells and within tissues will manifest as a change in the phenotype and the function of that tissue.

The kind of lowest-hanging area where this interest came together is in the development of new drugs, to kind of figure out how to fix bugs in our own biology by applying this understanding to learning the mechanisms we can leverage and exploit to make those corrections. So, yes, my journey began from making microfluidics at a large scale to making tissues at a large scale to making data at a large scale and then learning from that large-scale data to produce this integrative understanding.

[0:03:26] HC: So, what does Vivodyne do? And why is it important for the drug discovery process?

[0:03:31] AG: Sure. So, Vivodyne grows human tissues at a very large scale so that we can understand human physiology and we can test directly on it in order to discover and develop better drugs that are both safer and more efficacious and that enter clinical trials with much more certainty in what exactly we are manipulating when we dose a certain target or a receptor or apply a certain therapy.

[0:03:53] HC: What role does machine learning play in this?

[0:03:57] AG: Yes, the important role of machine learning is if you consider broadly how data is processed, let’s say even in a lower throughput way, currently in fields where the like output modes of data are very complex. So, we look, for example, at studies in animals, right? We have many different histology sections of tissues from mice that have been subjected to different conditions. You look at those sections and there’s no very simple thing to quantify.

In other words, I can’t look at the total brightness of the image or I can’t very easily look at the amount of H&E staining, for example, and say, “Oh, this one’s got more staining. It must mean there’s more disease.” Instead, we have to rely on, or have had to rely previously on very, very deeply experienced pathologists to look at these cells and try to make sense of what is going on, either in the context of a new disease or the reaction to a certain therapy, or a combination thereof.

What is very difficult and different here from a much more simple, like the unidimensional kind of readout, like the intensity of a signal, is that there is so much context that that pathologist needs to use to understand what something means in a certain part of that image. If you have a purple spot within a pink cell, it is a nucleus, but if you have a purple spot outside of a cell, it might be a sign of developing fibrosis, let’s say, or a dark pattern. Or if a cell is very large on the surface of an epithelia in that section, it is an epithelial cell and it is healthy. Or if you have a very large cell in the stroma of that tissue, it could be cancerous. In reality, the variations are actually even more subtle than that, right? There’s more specific context that has to be used.

It becomes very difficult to pull meaning from these images. And it becomes also very hard to ask these subject matter experts, like, what exactly are you looking at? What is the recipe to making the conclusion that you’ve made? Because there are so many aspects of different parts of content in that image to which they are attending when they are making that conclusion that it’s almost intuitive and not a kind of logical sequence, right? The conclusion they make.

So, much of the same way, if we look at the data that we generate from the kind of perturbation and testing on these human tissues, we end up with very large 3D phenotypic images with hundreds of thousands of cells in a tissue context and these very subtle changes in that phenotype that result from us dosing them with drugs or trying this challenge or myriad and often very subtle. So, it becomes very difficult to apply the very basic approaches to intensity quantification or the counting of objects in these images as biased approaches to understanding what is in them. And we have to take this more complex approach, much like a pathologist would take on slides of tissue, except at a much higher scale, of tens of thousands of these samples in a single experiment that we run.

So, the role of machine learning is to learn from a vast wealth of data what features in these images, both on phenotype and transcriptome and secretome, what are the features that most specifically indicate the, let’s say, regression or development of a disease in an otherwise healthy tissue. Kind of to summarize, we use machine learning and AI as a mechanism to understand the complexity of very deep data and to very efficiently apply that complexity and infer from what we’ve learned across the very large breadth of data that we collect.

[0:07:27] HC: So, the input to these models are the high-resolution 3D images that you mentioned? Or are there other types of data? What’s the output from the models?

[0:07:35] AG: Yes. So, the interesting thing is, there are many different modes of data that we work with. One is image data. We take a couple of different approaches to working with these. We have image data where we have single cell resolution but it’s a scope of large tissues. Hundreds of thousands of cells, kind of matchstick size, let’s say. The equivalent of large biopsies that you could take from people, except we grow these tissues at high throughput with these big robotic platforms.

So, as one input, we have these 3D images. As a second input, we can do transcriptomics and single-cell sequencing and ATAC and site seek and all of these genomics measurements on these same tissues and see how they respond from a signaling perspective in these perturbations. Then third, we can look at the secretome, right? What are they releasing and how are they signaling both to other cells within that tissue and externally, downstream within a person?

There’s two things that we can do with these different modes of data. The first is we can try to come to some insight or some conclusion from that data directly. The most simple example is, this is a very biased approach, we’d biased in the sense that we know what we are looking for. If a person has learned that staining a sample to look at the deposition of collagen one might be a great marker to assess how much fibrosis has developed and so we stain for collagen one and we see where fibroblast and stromal cells are producing collagen and we then quantify those things. Or we have a vascular phenotype that is induced by some vascular pathology in response to some drug and we perform segmentation and geometric analysis of that vascular structure as a result of that staining. We use segmentation models to mask out the vasculatures and approach them. We have bone marrow tissue and we’re looking at the production of progenitor cells and we count the cells that are positive for some signal.

These are approaches that use these biased inputs in the sense that a person has said, “These are the readouts that I’m looking for. I know that these things matter. They’re at least involved in some process,” and what are the results? The second kind of mode or approach to interrogating this data is an unbiased one where you are trying to understand given some input or given some measurement. For example, I know that in clinical trials, I have this very large kind of database of whether some drug has produced toxicity in some tissue. I take those same drugs that I have this nice kind of input lookup table of, and I test them on our tissues, and we get some phenotype out of those tissues, and I learn to predict what the clinical outcome would have been from the output of testing on our tissues, right?

So, given some 3D image of these tissues, can we predict or can we infer whether that patient would have an increase in some biomarker at some week of treatment? And what that produces is an unbiased model where you’re relying on deep learning to do the thing that it’s very, very good at, which is understand what features in these images to use in which to attend to when producing that inference. So, the outputs that we create are both biased in that a person can say, “We want this piece of data, we want that piece of data, we want to quantify what exactly it is,” or they can be unbiased in that they are linking some input data to some output prediction and being unbiased in its training on which features it picks up to do that.

[0:10:48] HC: Of those two types of models, have you found one to be more helpful than the other?

[0:10:52] AG: It depends on context, right? So, I think what it also plays out a little bit is the value of both in explainability, because you think in this unbiased mode, you are reliant on having a lot of data to learn these features. Then, oftentimes, okay, well, you’re getting really good prediction from one input to an output, but you want to have a sense of explainability, like what exactly, which features are driving those predictions. Both to kind of increase our understanding of biology, maybe there’s something there that we hadn’t seen, right? By having all this data and in training this inferential model, we have found this new thing that it not pays attention to.

So, explainability to understand new things, but also explainability to validate that we are not paying attention to some artifact that might have introduced bias into our samples and biased in the sense that, maybe we received, maybe cells that were isolated from certain patients that were in a clinical trial compared to a control healthy group from a different sample site, have some variability in how well they stain for some target, or if the experiments were done sequentially and there was variability in the lot of some reagent that we use to gather that data, maybe those things are other differences.

So, using biased methods to provide explainability to these unbiased methods is oftentimes they are both equally useful, but together they are even more useful than they would be as the sum of the parts, right? You get like the product of the parts here.

[0:12:18] HC: Yes. It’s great to hear about those two pieces working together because deep learning, as much as it is powerful, it is very black box. This mechanism of interpretability that you talked about is really important.

[0:12:31] AG: I think it’s interesting that many of the tools that we are using to, in a way, at least conceptually, but maybe not an implementation, but at least conceptually mimic the way that human brains think, allow us at least a little bit of window into understanding that explainability.

So, we can do these really interesting things. If you have an attention network, let’s say that is active in a vision transform or something, we can look at what it is attending to in the input to understand in a contextual basis when is some information important compared to other cases. We can also invoke or force constrictions in the model. For example, given some input, there has to be an attention network upfront that limits the real network to only seeing a single tile of that image or a single subset of that image. You train it to pick the most informative part of an image creating the prediction that’ll infer, and in so doing, you can quite literally focus in on what it is paying most attention to, or like, what is the most important part of this image that will create the most amount of predictability in a model if it had to choose only one area to perform this inference on, right? You’re co-training. I can only pick one little piece of this image to predict on while also training that downstream model that will do the prediction, which is kind of in an effect, an attention network of its own.

So, it’s interesting the things that we can do, given enough data to understand what these inferential models look at and how we can learn what they’re looking at without it being just some black box that we magically assume is looking at the right things.

[0:14:01] HC: I imagine quantity of data is key to getting some of these insights. How do you go about gathering data to train these models? I think you’d mention something about growing tissue?

[0:14:11] AG: Yes. So, when we think of language models, let’s say at large, we are just now reaching a point where we’ve kind of begun to saturate in this approach of just scale the model, scale the amount of data on which it is training, and the model will become smarter. We are just now kind of hitting the limits on that and I say we, as in like the AI field at large. We have grown to be training our language models on essentially the content of the entire Internet, which is a great thing to have.

It’s like you imagine the centuries of human effort in producing that training data just so this LLM could learn it all and have some amount of insight come out the other end if you prompted the sample from that training set. That does not exist in biotech, sadly. We don’t have centuries of openly available and very deep clinical trial data on which we can trade in from which we can learn.

There’s many reasons for that. One is there are privacy issues for the outcomes of those patients in clinical trials. The second is the invasiveness with which you can take data from those patients. If in a clinical trial, you can take a blood draw and you can look at a biomarker or two and have a good sense of progression of an underlying disease, well, that’s great. You don’t have to take an invasive biopsy. You don’t have to take a large sample. It becomes easier to monitor more patients.

The downside is we now end up with segregated data where depending on the clinical trial, we’re looking at different markers, the patients are being tested for different periods, even like the analysis of that data, the volume of blood, the time of day they’re coming in for that, all of these things begin to diverge. So, you end up with extremely segmented and inconsistent datasets and even that is in a perfect world, right? You imagine all the noise between different hospitals taking similar data, for example.

Then third, most of the time, it is very expensive for this data to be produced. So, if I am a pharma company and I’m taking a $100 million bet on a clinical trial and it doesn’t go the way that I expect, the only value that I really have left out of that is the data that I have produced. Although altruism is great and all, I’ve spent $100 million producing some data and I can’t so easily just release it to the world and say, “Hey, all the other pharmas that are also trying to make revenue. Here’s some free data.”

We have this issue of inconsistency in existing human data, inaccessibility of that data, and variability in what type of data is collected. There is no amazing central source of data on which the amazingly predictive clinically facing models of what happens to this patient if I give them this a drug. There’s no centralized data on which potential models of that type can train. That’s just because of the difficulty of obtaining that data consistently.

Now, to get to clinical trials and drug development, pharmacists test on a lot of animals, right? We currently – let’s exclude the ambitions of our company, Vivodyne, from this. We do not have really any better way of predicting the complexity of a person than to test on a complex mammal. Let’s say like a mouse. But the problem there is you have very low throughput. You don’t have the throughput of cells in a well plate to test on mice. You have a couple dozen mice, let’s say, in a certain study of a drug. You run into this problem of scale of that data and then you have to also kind of apply some of those constraining variability driving differences in how you look at and examine these animals, right?

In one mouse, we might section tissue A and B and look and stain it for this marker. In another set of animals, we might look for a different panel, or you might do a blood draw, or we might only weigh them. So, there’s also inconsistency there and low throughput. There’s no holy grail of data that we can use to train even the most like low-hanging fruit of models that we think, “Wow, wouldn’t it be predictive if only we had this data.” To understand how human physiology works, even in very kind of basic, such a thing exists, very basic contexts.

If we go even further back to the things that are currently possible to scale, you have cells in well plates and cells in the many thousands of well plates of a two-well plate. We can treat all those wells containing the same cells with a different challenge. We can give them a different drug. We can give them – we can knock out a different gene. We can knock out a different gene in each cell within those wells and do sequencing. We can do all these very deep and scalable things to cells in well plates.

But the problem is we are excising so much complexity of human physiology. There are so many interactions between different types of cells that allow those cells to break symmetry and to respond to the conditions of their microenvironment in that tissue and to respond to systemic differences in physiological response that we are excising completely when we have a single-cell type that is just growing on a sheet of plastic.

So, we finally get to a point where we have some throughput, but the depth of that data is suddenly ridiculously shallow. For all the scale is worth, we now have such a shallow basis from which to project into this much more complicated output space of human physiology. We’re like trading realism for throughput as we go from humans to animals to cells in a dish, and although throughput goes up, we’re really losing the realism that makes it important.

So, the purpose of Vivodyne is let’s change that premise. If this reasoning that well, there will never be a way to have both is a product of, I think, a very simple premise, which is we only have cells in a dish or we only have animals and then clinical trials thereafter. Our mission is, can we have the best of both worlds? Can we have our cake and eat it too? We then think of how can we gain deep samples of human physiology at a very large scale? Maybe, well, because people exist already, how about we just take biopsies of different tissue types within clinical trial volunteers, let’s say, or patients in clinical trials? But we run into that same issue of invasiveness, right? So maybe in some oncology clinical trials, maybe we can get some biopsies, but they are not hundreds of thousands of them. And for other tissues in that person, they probably would not really like to have biopsies of all their different organs taken just in this like, “Oh, we’re just making data and that’s it.” So, it’s very hard for systemic biopsies across many tissues to be gathered in a way that is useful and scalable. If we take one step back from taking a biopsy, the question then becomes, what if we grew that equivalent? What if we grew a biopsy of a liver? What if we grew a biopsy of bone marrow, of the lymph node, of the lungs and airway, of all these different tissues? What if we grew biopsies of diseased tissue and cancer tissues? Can we make those tissues that we grow as close as possible to the real thing? Because if we have an operation that allows us to grow these very realistic tissues at a large scale, then we can kind of get the benefit of the depth of human physiology, but if the operation is scalable, also the throughput of data that we need.

So, the way that Vivodyne does this is we grow human tissues and we rely on the self-assembly of those tissues to grow them over the course of about a week or two into their mature state. But because these tissues are growing by self-assembly, we can scale the process to many tens of thousands, if not hundreds of thousands, different tissues or the same tissues with different or the same tissue is exposed to the same challenges, but from different cell donors and people, and we can grow this very large base of deep and wide data that we can use as a basis for this training that is not otherwise available.

Yes, to address this glaring lack of trainable data, we create that data by growing it at scale by being able to test outcomes kind of just in time. So, we have a new hypothesis of a new pathway is discovered or new mediator of a pathway is discovered. We want to know what happens if we shut it down or what happens if we activate it and amplify it. We can generate that data just in time and add it to a set. This allows us to create the depth and the scale of data that’s required to kind of push through this bottleneck of data and availability in biotech. So, data becomes wildly, wildly important. It becomes the fundamental constraint of biotech AI. Our company works very hard to create this process to produce that data.

[0:22:11] HC: That sounds like an incredibly powerful and innovative way to solve this data problem. I definitely haven’t heard that solution before.

[0:22:19] AG: Thank you. Yes. We end up taking credit for a lot of the logic that is employed by cells like every day in our own bodies, right? We remodel their environment. They keep our tissues perfused with blood. They keep them from becoming too stiff or too soft. So, we are just exploiting the things that these cells know how to do already and then gloriously take credit for all of it.

[0:22:38] HC: So, how do you go about validating your machine-learning models? Is it as simple as just growing more tissue or is there more to it?

[0:22:46] AG: Where that is almost like unfairly easy in our cases, we can predict a certain outcome. Then, because of the very low marginal effort required to like literally just like test that within our tissues, we can produce that data as an outcome and use it as the training cost, as the loss function, in other words, right? We can compare to known data. We can train it on that. We can produce some inference and say, “Okay, this unknown compound or this unknown pathway. What happens if we modulate that?” We can predict some outcome and we can validate it against tissues that we test with that drug that have actual data for it.

That’s a kind of training accelerant and an active learning approach to increase the accuracy of our models. It also lets us understand what parts of the training data like landscape are too sparse or completely vacant and where the areas from which sampling produces like really out-of-distribution outcomes, we can fill it in with new data that we generate and the kind of comparison data that we use to ultimately validate these predictions, human outcome data from clinical trials.

But the proxy of having the depth of human data in our tissues allows us to overcome what’s otherwise pretty shallow data from clinical trials like a single biomarker or like self-reported side effects and so on.

[0:24:00] HC: Great to hear those extra steps there. In developing these models and in validating them, it sounds like there’s a great deal of knowledge that goes into it. It’s not just machine learning, it’s knowledge of biology and of tissue and of all how this comes together. How does this knowledge from the domain experts get integrated with machine learning developers? How do they work together in order to create the powerful models that you need?

[0:24:23] AG: Yes. I think one of the great philosophical and conceptual challenges of our company kind of elevated above the sciences, we have to be really good at biology. We have to be really good at hardware because we’re growing these tissues with robotics in order to scale them and we have to be really good at the kind of machine learning side, so we know what we’re even getting with the data.

The hardest part of like architecting the company has been how we center those domains of expertise around their common overlap, which is kind of like tech bio at scale with the software aspect. So, on the biology side, we try to hire for folks that have had a lot of time or that have had a lot of training in exploring biology from an engineering kind of perspective. You think of have they been involved either in research or in industry work where their role was to understand how some system of biology works and to either reverse engineer it or to kind of like use it to some end, right? Sometimes in drug development, other times in tissue engineering, or in bioengineering, or in diagnostics development and so on.

Within the robotics side, the kind of key thing there is having enough exposure to biology that the context of what we’re doing makes sense. If someone has only looked at – if they’re a mechanical engineer, for example, and they’ve worked at the nanoscale, and they’ve worked on doping a new type of more faster-switching transistor, they’re down there at the nanoscale, it might not be so conducive to interfacing with biology. But if they’ve worked in systems where you are doing microfluidics-based diagnostics, or you are building robotic systems for brain implants and robotic surgery, and all of these kinds of aspects that lend themselves well to an interface with biology, then we have kind of wide bandwidth across those people.

Then lastly, in machine learning, the kind of core areas of domain expertise that we look for are in folks that have worked specifically in models employing multimodal data inputs across kind of depth in those modes. So, a kind of simple example is in like contrastive learning and you’re looking at models like a clip, they have a caption, they have an image, and you learn the associations between those. In our case, we have to do that, but it becomes even harder because it’s an image, it’s a transcriptome, it’s a secretome.

So, folks that have had this experience and have developed this intuition for how do we structure an embedding like this, like joint representation and in some latent space. Or in some projection into which an image of a tissue and a transcriptome of a tissue will take on the same embedding or the same vector in that space. An understanding of those things that comes intuitively makes the rest of what we’re doing almost like second nature. So, you’re looking at a caption that’s had a picture of three dogs and making sure that you’re labeling the right image with three dogs applies itself unexpectedly well to an image where you’re looking at a single-cell sequencing dataset and a phenomics 3D image and saying, “To which image does this transcriptome apply?”

We look for those specific expertise sets in order to structure a team that can work really well with these interfaces and just have really broad bandwidth across those interfaces to communicate with each other. That becomes the fundamental mechanism we use to make sure that we can kind of grow and scale efficiently and that these different teams are not all building different things without context for what the others are doing.

[0:27:55] HC: Is there any advice she could offer to other leaders of AI-powered startups?

[0:27:59] AG: Oh, boy. I hesitate to give advice. I can say only my own kind of learnings and that is really, I have found as a technical leader of the company that the easiest way to kind of leverage and to bet on technical depth as a way to make our company kind of more likely to be successful or more differentiated from the things that other people are doing, is to bet really hard on that kind of technical axes, right?

If you’re a technical founder, do something that is incredibly, incredibly hard because the ability for you to do that thing will grant you much more leverage than creating what is otherwise a much more simple and generic business and trying to scale on the principle of efficiency to market or cost of goods or something that is much easier for others to do. You end up building a much deeper moat if you try the very hard thing and you’ll be surprised that interest of partners in working on that very hard thing, both in the way of folks that want to work with you to join the company, investors that want to work with you to accelerate the company, and partners commercially that value what you’re doing and I want to engage with you there.

So, relatively shallow advice, but hopefully, at least confirmatory, it’s like it is okay to do something that is very hard. And in fact, it grants you much more leverage to do that thing and it makes the process much easier than saying, “Well, no way you’re going to buy this. This is too much like sci-fi. Let’s make a marginal improvement to something and maybe more investors will be interested. Or maybe more customers will be interested.”

[0:29:29] HC: Finally, where do you see the impact of Vivodyne in three to five years?

[0:29:34] AG: When we think of drug development and how it goes, it doesn’t go so well in clinical trials. Failure rate is kind of low to mid-90%, so 93% to 94% failure rate. You kind of think, why is the failure rate that high?

You can skim off about 10% of failures because of downstream like market impact. There’s no patients, maybe there are not a sufficient number of patients that might take an otherwise safe and effective drug, or insurance doesn’t want to cover it, and the biopharma doesn’t do so well. There’s management problems at the pharma, but that leaves about 90% of failed drugs, which is about 88 or so percent of 88%, 85% of all the drugs that are taken into clinical trials that fail in about equal proportion because of safety and efficacy concerns.

The interesting kind of catch there is all of the drugs that have entered clinical trials must have demonstrated to the FDA to an incredibly high standard that they’re probably going to be very safe and effective in people. We’ve done that in animals. So, you think of every drug for all these different indications, there’s like one and a half thousand INDs that are submitted each year to allow a drug to enter clinical trials. Each of those one-and-a-half thousand INDs have done something pretty amazing in animals. Those animals are not any simpler or less complex than we are.

Organ systems of mammals are all there in the animal. The signaling is equally deep. The only difference between us and animals like mice is there are small, small changes that are kind of scattered everywhere in our genomes that just change the statistical chances that some interactions happen or just change the epitope of a certain binding domain, and that propagates across the entire integrated product of that genome and how it manifests into the world in producing this huge difference in a mouse, which is this tiny little thing in us. We can work these miracles and animals, but why can’t we do it in people?

It is really just the fact that the failure rate of that 93%, 94% is because we’re testing it for the first time on people, and these differences between humans and animals have manifested so that the thing that we’ve optimized on the animals by virtue of having refined on them and tested the thing multiple times has just been not the right thing that works on people. Maybe if we had the opportunity to refine and optimize that drug on human tissues before clinical trials that we would have the same success rates and the same ability to fine-tune those drugs as we currently do on animals where we can beat all sorts of different types of solid tumors.

I mean, you look at how frequently in the press you get these reports of cures to different types of cancers and regressions of Alzheimer’s and extensions to lifespan and all of these things that we seem to be able to do really easily in animals but we just can’t seem to do well in people or at least very efficiently in people. You think, well, maybe if we can do in people what we do in mice which is test on them, and in this case, in an ethical way by lab growing human tissues of substrates, that maybe the amount of successful drugs so we can accelerate into clinical trials and into lives of people that have all sorts of really nasty diseases can be accelerated, right?

So, we hope that over the course of the next half decade and decade and beyond that, we will be able to produce drugs. We will enter this world of plenty where the development of new drugs against diseases becomes a far more successful, reliable, and predictive process, and we’re able to make much safer and much more effective drugs just by virtue of being able to optimize that therapeutic on human tissues before giving it to people for the first time in clinic.

[0:33:09] HC: This has been great, Andrei. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[0:33:15] AG: They can find out more about Vivodyne at vivodyne.com. I’m along for the ride.

[0:33:19] HC: Perfect. Thanks for joining me today.

[0:33:21] AG: Yes. Thank you so much, Heather. I really appreciate you having me.

[0:33:23] HC: All right, everyone. Thanks for listening. I’m Heather Couture, and I hope you join me again next time for Impact AI. [OUTRO]

[0:33:33] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe and share it with a friend, and if you’d like to learn more about computer vision applications for people in planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]