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]

[00:00:03] 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]

[00:00:34] HC: Today, I’m joined by guest, Ashley Zehnder, co-founder and CEO of Fauna Bio, to talk about drug discovery. Ashley, welcome to the show.

[00:00:43] AZ: Yeah. Thanks so much for having me, Heather. Looking forward to it.

[00:00:44] HC: Ashley, could you share a bit about your background and how that led you to create Fauna Bio?

[00:00:49] AZ: Yeah. Of course. Myself and actually both of my co-founders have somewhat non-traditional backgrounds for the company. I guess it’s biotech at large. Not for the company we’re building. I came out of the veterinary sphere. My background is in veterinary medicine. I did an internship and a residency focused on exotic animal medicine. And so, for those that are not familiar with exotic animals really are considered within veterinary medicine, really anything that people keep is pets that are not dogs and cats. Birds, mammals, reptiles. Not so much zoo animals. But I really spent a lot of my clinical time working with these very different animals that had very different ways that they presented with different kinds of diseases. Everything from cancer to heart failure. Kind of everything in between.

I became very interested in, why is it that different species present in different ways? And why do certain species get different kinds of diseases at a higher prevalence than others? What’s the underlying biology that really relates to that difference? I became interested in that from an academic standpoint and applied for a cancer biology PhD at Stanford, which was focused on supporting veterinarians who wanted to do PhDs.

And it was really there that I started to appreciate this very large knowledge gap that existed between people that had come up through, say, medical background, human medical background and were doing research for drug discovery or other types of translational research. Really only thought about doing discovery for humans by looking at either human data directly or really working with mouse and rat models where they try to induce a human-type disease and had very little understanding or appreciation for the diversity of disease traits that we do see across a number of different other species that we can learn from.

My original intention was to go do a PhD at Stanford, learn as much about cancer biology, and go back and be a clinician who also taught students in residence in an academic institution. And realized that there was actually a much bigger problem that needed to be solved, which is this knowledge gap.

I was fortunate enough to then start a postdoc at Stanford also with Carlos Bustamante, who was the new Chair of Biomedical Data Science. He was really working with postdocs. Looking at very large medical data sets. I was looking at clinical data and comparing clinical data between humans and animals. And actually published a few ML papers out of that work actually which are relevant to this podcast.

But then I met my two co-founders, Katie and Linda, who are both working on very different aspects of different animals and how you use them for translation. Our CTO, Linda Goodman, her PhD was in human drug discovery and human genetics. Looking at risk factors for heart disease. And realized it was actually pretty hard to figure out what genes were driving disease by just thinking about humans in isolation and not thinking about humans in the context of their hundreds of millions of years of evolution as mammals. And what other genes have we kept that are consistent across all of those hundreds and millions of years? And those genes that are highly conserved, or as we call them constrained and that they don’t mutate very often, are actually usually doing something important. That’s why nature has selected them. She really became interested in looking at those highly-conserved genes in disease. What are they doing?

And then my co-founder, my CSO’s background is also in human genetics. But she rotated in the lab of Dr. Sandy Martin at UC Denver who was really one of the four founders of looking at hibernation genomics and really trying to understand how species that hibernate throughout the winter – which is fascinating and we’ll probably get into a little bit of that – How they protect themselves from damage that happens during that time period. And it turns out they’re using just highly conserved mammalian genes. The same ones that you and I have, but in really remarkable ways. And so, both of them started out working really in human medicine but really gravitated toward looking at other animals for models of how we can really improve human health.

[00:04:35] HC: And so, what does Fauna Bio do?

[00:04:37] AZ: Fauna Bio at its heart is really a novel drug and therapeutic discovery company. We use AI and genomics as a way to really identify what are going to be the most impactful targets for new therapeutic programs across a broad number of diseases. We can use that data to make our own drugs that we’re planning on putting in the clinic as part of our own pipeline. But then we also can do discovery work for large pharma that are working in really difficult disease areas.

One of the more recent examples of that is the work that we’re doing with Eli Lilly to look for new therapeutic ideas around obesity. Folks, and probably if you’re not under a rock, are aware of how much immuno-GLP agonists and new therapies for obesity have been in the news. A lot of that science was really done decades ago. And a lot of that research was done really focused on diabetes. And then people realized that those drugs could be used for obesity.

And so, there hasn’t been a lot of effort put into really new therapeutic strategies for obesity for decades. And so, it’s a really high bar to actually get obesity drugs approved because there’s a lot of concern about safety. And so, what is interesting for a company like Lilly is what Fauna Does, is really bring wholly new data sources to the problem of drug discovery.

We look at species that can massively change their metabolism. Change their level of hunger or satiety. Rebuild skeletal muscle even though they haven’t been eating. And these are all disease these traits or protectant traits that are very relevant for humans that are on some of the currently approved GLP-1 agonists like Wegovy and Zepbound are names that people would have heard of.

From that perspective, it’s helpful for big pharma who have oftentimes spent decades working with human genomics data to think about those diseases in a different way. And Fauna really brings that data to the problem of drug discovery. And that’s important because we’ve had the human genome for close to 30 years now. A little over 20 years. And there’s still a lot of complex diseases that need new therapies. And we really need new data to think about those problems.

[00:06:34] HC: And what role does machine learning play in this technology?

[00:06:37] AZ: A lot actually. As you can imagine, when we start to intersect not only genomics data, which we integrate. Looking at actually the genetic bases with these different species, we do a lot of discovery from one species called the 13-lined ground squirrel. But it’s really a prototypical hibernator and represents a number of different species that hibernate.

But to understand what’s special about those species, we have to compare those to now up to 452 mammals, including humans, and look at what genetic bases have been conserved over those hundreds of millions of years of evolution I mentioned earlier. Obviously, we got those genomics data, which is a huge data set. Those take supercomputers to analyze and annotate, which we collaborate with the Broad Institute to do that.

But then what Fauna works on a lot more and intimately, or more familiar with inside the company, is looking at how the gene expression changes. Whether genes are going up or down. Whether there’s more genetic data and protein being expressed at times where animals are protected from damage.

And so, one just quick example of that, ground squirrels, when they hibernate take their body temperature down to almost freezing. They stay there for a couple of weeks at a time. Their bodies are kind of 90% depleted of oxygen. Their heart rate is five beats per minute. Normally, it’s 250. Their metabolism is about 1% of normal. And then over the course of just an hour, they’ll rapidly warm back up to normal body temperature. And that process is actually pretty damaging. It’s not that dissimilar to what happens to a human after a heart attack or a stroke.

You could imagine, if you had a heart attack or a stroke every couple of weeks over the winter, your heart and your brain probably wouldn’t be functioning so well when you woke up. But these animals do it every year. And that’s part of their biology. We can then start to look at how genes are changing when we know that they are repairing the damage that does occur during those time periods.

We have to then intersect that expression data, is what it’s called when you look at the RNA protein expression, against what we see in human disease. Then we’ll take data sets from, say, humans with heart failure, or neonatal mouse data where they can still regenerate parts of their hearts. And we’ll intersect those and we’ll look for enrichment of this opposite but protected pattern that we see in the ground squirrels.

And so, we use machine learning really to identify what are the most impactful targets as we start to overlap this data? We have now 35 external mostly human-centered data sets into a very large knowledge graph. We have a graph neural network that runs on top of that knowledge graph to help us with target predictions. And that is really born out of the fact that this is just too many data sets for someone to work with without using machine learning and AI to be able to figure out what are the most important target concepts for us to work on.

[00:09:20] HC: These machine learning models that help you draw connections, are they supervised models that give a certain input and it needs to predict a particular target?

[00:09:31] AZ: Yeah. We’re really trying to improve our ability to predict a gene’s role in a human disease. And so, as you mentioned, it’s about drawing those connections within the knowledge graph. Between a gene that may not have been associated with a human disease before and building evidence for that connection. And so, that’s really what a lot of the graph neural network ML is doing.

[00:09:47] HC: Is this gene expression data you’re working with? Or is there full sequencing data as well? And for those who aren’t familiar, what do these data types that you’re working with look like?

[00:09:56] AZ: Yeah. Sure. Both. The discovery is largely driven by expression data. And so, this is largely RNA sequencing. When we have – kind of going back to our basic biology, if we have your genome, has the literal base pair sequence that kind of makes you you. Those base pair sequences make concrete packets called genes that can be mutated in different diseases.

But sometimes, in a disease process, more of those genetic data get transcribed into what’s called RNA, which is a kind of intermediate step to then making proteins, which are kind of essentially the machines that really function throughout the body. And so, we look at the level of RNA in different tissues at different time points because that really tells us how much the body is really using a particular gene at any one point in time.

And so, we use a lot of RNA sequencing data. Historically, we were doing a lot of what’s called bulk RNA sequencing where you just kind of look at tissue expression for a whole tissue. Now, really, our workforce is becoming single-cell RNA sequencing, which the name uh suggests, is looking at expression of RNA from one specific cell. And so, you get much more fine-grained detail around what cell types are participating in a disease process.

For example, in the heart example I gave you before, we can look at data from the heart. But we can say, “Okay, what are the genes doing in a heart cell? Versus what are the genes doing in maybe a fibrotic cell that may be in the same tissue or an inflammatory cell that’s in charge of looking at inflammation?” And so, you can ask those more detailed questions of that data if you have single-cell RNA sequencing. That’s the workhorse.

We do this all on the background though of the whole genomics. What we are really focused on at Fauna is translating the insights from these species into drugs to help people at the end of the day. And so, that doesn’t help us if the genes that we’re working on only exist in, say, the 13-lined ground squirrel. It’s really cool for the 13-lined ground squirrel. It doesn’t really help us.

Against the background of all the work that we do is that very large alignment of 452 mammals that really cover every major group of mammals on the planet. And so, we look to see, are the pathways that we’re identifying, how similar are those pathways in ground squirrels, and bats, and horses, and humans, and marine mammals, and elephants? Are they just highly conserved pathways that most mammals have?

And if we’re seeing changes in those highly conserved pathways, they’re much more likely to do something important in people as they do in the species we’re doing discovery in. And so, that level of conservation is important for us to make sure that our work will translate into impact for people.

[00:12:29] HC: How do you go about collating this data set from the 400 animals that you mentioned there?

[00:12:35] AZ: Yeah. Sure. We do collaborate very closely with the Zoonomia Project at the Broad Institute. Linda, our CTO, did her postdoc with some of the main investigators with the Zoonomia effort. That’s actually a consortium of labs that are generating data from many, many, many different species. There are actually a really nice set of publications for those that are interested that came out in Science in April of last year that really explained the data sets, have the primary data, talk about the implications. It was a whole issue of Science dedicated to Zoonomia. That’s a very nice resource for people that are interested.

But we collaborate with the Broad. And they do a lot of those alignments. Because I said that takes a supercomputer about a year to really line up all of those genomes. And that, again, forms the background upon which we then layer the Fauna data that we generate looking at how those highly conserved genes are changing in disease-relevant settings in the species that we work with. It’s like kind of the foundation upon which we build the rest of the Fauna platform is that work from the Broad.

[00:13:32] HC: What kinds of challenges do you encounter in working with this genomic data? And in particular, training machine learning models based off of it?

[00:13:40] AZ: Yeah. And I think this comes up a little bit in one of your later questions is thinking about validation, right? How do we know that it’s working? I think with genomics data, there’s obviously variability and quality. And that’s something that is getting better actually over time. One of the things that really enables the work that we do at Fauna is the improvement in the quality of sequence data that is able to be generated for a much lower cost. It’s just a heck of a lot cheaper to generate a whole new genome for a species than it was when they first sequenced the human genome, which cost a couple of billion dollars if I remember off the top of my head. Now it’s down to tens of thousands of dollars if you really need to generate a whole new genome. And generating sequence data for an individual, animal, or person can be cheaper than that.

It’s much cheaper to generate higher-quality genomics data than it used to be. And, similarly, much cheaper to generate the kinds of expression data. RNA sequencing, for sure. Proteomics is still a little bit more complicated. But it’s certainly easier than it has been in the past to generate very high-quality single-cell RNA sequencing. We’re doing a lot of that for the work that we’re doing with Lilly. Tons of data sets we’re putting on the platform that we’re doing for drug discovery.

The challenges on the technical side are getting much easier. I think the challenges on the interpretation side are still there. Is a gene in a certain location in one species the same gene in a different location in another species? That annotation and understanding. What does a sequence mean in terms of an ultimate genetic product? That could be a little bit more complicated. And there’s some nuance to that. But, fortunately, my CTO and her colleagues, that’s what they did their PhDs in. And then my CSO as well. They both have deep expertise in that field. And how do you annotate, is what it’s called, the genomics data?

We fortunately have domain expertise in that. But that can be tricky if you don’t know what you’re doing. And then really figuring out what are the right data sets to query. I think when you’re dealing with biology as complicated as hibernation science as a discovery platform, again, these are animals that massively change their entire physiology every couple of weeks and in a very short – like I said, an hour time course is very short. Capturing the right data at the right time from the right tissues is something that Fauna is particularly good at. And we’re using more and more automated types of machine learning to map those time point signatures to human disease. It’s less sort of domain expertise-driven.

But at least when we started the company, a lot of it was based on our understanding of the hibernation literature and the physiology that had been done. That’s less so now. It’s just complicated biology. And so, I think the tricky part of working with genomics data is not always the genomics data. It’s actually what we call the phenotype data. It is what is happening in the animal at the time point when you took your data, right? Is it a sick animal? Is it a healthy animal? Is it a fat animal? Is it a skinny animal? Is it a warm animal? Is it a cold animal? And having really good descriptions of that phenotypic data, much more complicated. Because people describe things in different ways. There’s not a lot of standardization depending on where you’re getting the data from. And it’s just more complex than literally reading the base sequence from a sequencing run.

Actually, it’s kind of the shadow problem of genomics data is not the genomics data. It’s the phenotype. And, actually, we looked at acquiring a number of other biobanks when we started Fauna. We acquired the one that our CSO built during her PhD because we understood the phenotype very well. Because she built the data set.

We looked at acquiring a few other biobanks. When we started the company, we went and talked to a bunch of investigators and tried to understand how they were describing the data. And it wasn’t very consistent. They weren’t using consistent time points. They weren’t using consistent measurements to understand what was the state of the animal when we took the data.

And without that, the data itself is not very meaningful. And so, what we switched to now is working directly with academic labs to design the studies. Work with those investigators to generate the data. Understand exactly how they were generated. And that data is “clean enough” then to go into the platform to use for discovery. That’s actually more of the problem, is understanding what was the state of the creature from which you took the genomics data. Then the genomics data itself sometimes.

[00:17:43] HC: Let’s talk about the validation piece. If your models seem to perform well on the data that you have, how do you know that they’ll work well on other human data that’s not in your data set?

[00:17:53] AZ: Yeah. For sure. There’s a couple of layers to that. One of those is really more purely computational validation, which is looking at how well does our data predict existing and well-known drug targets? Can we find well-known drug targets? And does our platform rank them highly? And then what are the other kinds of targets that are ranked similarly that are maybe not as well known? And so, really looking at what has been done before? What has been successful? Mapping against that. And then really just generating an AUC of our predictive capability for those known and successful drug targets. But, obviously, that is still computational. But trying to find gold standard examples to compare our predictive capabilities to.

The other one is sort of less sexy. It’s doing experiments on the bench, right? A lot of what Fauna does, we’re about half-half computational research team. And there’s team members who do both. But the lab in Emeryville, California is really our in vitro science team. And experts there. Have done their PhD and postdoc work in modeling human diseases, human cell models. We’ll take genes and compounds that we can find with the platform. Test them in human cell models and see if we see an effect.

Sometimes we need to go straight into animal models of disease. And so, we have an in vivo team that works up actually in Oshkosh, Wisconsin. We have a collaborative research agreement there with the university. And so, we’ll test changing a particular disease or testing with a compound and seeing if it works.

I think there’s sort of two levels of that validation. There’s the computational, what genes have made good drug targets before. Can we find them? And do they get enriched with our approach? And they do. Strongly, actually. And then can we find new targets? And then we have to sort of test them in real-life experiments. And I think that combination of data is quite powerful.

[00:19:33] HC: You’ve mentioned a couple of instances where the domain expertise of understanding the data, the phenotypes is very important for getting your data in the right format. Probably for training models, all that type of thing. How do your machine learning developers collaborate with these domain experts to develop this technology and doing it right?

[00:19:53] AZ: Yeah. There, we try to work with labs that are really at the cutting edge of how we use these data sets directly for medical applications that are really thinking about translation. One example of that is the Zitnik Lab at Harvard. We’ve got a consultant there who’s doing her PhD really and exactly what we do, which is building large knowledge graphs for target discovery.

And so, making sure that we are setting up the data sets in the right way. Using the right annotations. Using the right terminology and the right curation. Curating the data in the right way so that the data is clean. Removing stuff that is not clean and making sure we’re using the right statistics to validate the nodes in the different models. It’s really working with the academics who are developing a lot of those methods at the time. We work very closely with those labs.

[00:20:37] HC: Are there any specific technological advancements that made it possible to build this technology now when it wouldn’t have been feasible a few years ago?

[00:20:44] AZ: Yeah. I think the one that comes to mind is actually the one that I mentioned a few minutes ago was really this drop in sequencing cost. We are generating a lot more higher quality data now than even when we started Fauna. And, certainly, those two or three years before we started Fauna were all of us were finishing up our PhDs. It was feasibly intractable frankly to start from scratch with a new species, generate a genome that you needed, generate sequencing data without spending potentially millions of dollars to do all of that. Now those technologies have advanced to the point where the sequence quality is much better and it’s just much cheaper and more efficient. There’s contracting companies that we send out to do sequencing that get us high-quality data very quickly and for a cost that it’s very affordable for a startup.

Again, building on two foundations, I guess. One is that broad sequence data I mentioned before. The other foundation is really those improvements and sequencing technology. In terms of ML technology, obviously, improvements in graph neural network AI. Certainly, it’s been helpful for this knowledge graph that we’re building. We are starting to integrate more and more types of generative AI in the platform largely from the perspective of making the platform more accessible and intuitive for folks who don’t code, right? The folks on the Fauna team that are not building the platform, how can they use it more effectively? How can they query it in a way that allows them to get more value out of the platform? And that allows us to scale the impact of that platform across more users of the organization.

And so, those types of technical advancements really have to do with how do we ingest data? We’re experimenting with ingesting Fathom recordings from internal team meetings to be able to have the AI learn from those technical discussions and be able to use that in answers to responses – responses to questions.

And so, we’ve got somebody in-house who’s really just focused on integrating the newest and best types of AI and thinking about how to essentially lower the barrier entry for folks to access and get insights from our own data. And so, that’s something that obviously is changing minute by minute it feels like sometimes. But trying to make sure we’re picking up the best of that and applying it on the platform.

[00:22:49] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[00:22:53] AZ: Yeah. I do founder calls increasingly frequently over the past year or so who are starting companies that are kind of in the AI drug discovery space with different approaches. One of the things that I try to impress upon founders is really understanding your differentiation.

Particularly, I get a lot of founders who want to talk about how to partner with pharma and want to understand sort of how those partnerships come together. Pharmas see a ton of AI platform companies. And a lot of them are using really largely public domain data and thinking about it in a different way, which may be fine for internal discovery. You may find something really cool, which is great. But in terms of selling the platform approach, a lot of big pharma are becoming very savvy in terms of their use of AI.

They have whole departments focused on AI drug discovery. There are large consortiums focused on AI and drug discovery. There’s becoming increasingly less that those big pharma can’t do themselves actually quite well. There may be a few exceptions. But most of the large pharma have bought into that this is an important part of drug discovery and are building those departments to do that.

You cannot really just differentiate on the fact that we are a really strong AI team. So are they. What is the differentiation? What is the data set that you’re generating that they can’t generate? Or is there a way you can think about that data in a very different way? You really do have to have a very strong story around differentiation and why that differentiation impacts a particular part of drug discovery. It might be clinical trial optimization. It might be looking at safety. Or it may be looking at ways to repurpose drugs that have failed in phase four. It may be early types of discovery.

There’s many points along that drug discovery path where AI companies can differentiate. But that story has to be really clear. Because, otherwise, it’s very hard to kind of get out of the signal-to-noise that is the AI kind of discovery landscape in biopharma.

[00:24:44] HC: And, finally, where do you see the impact of Fauna Bio in three to five years?

[00:24:49] AZ: It’s a great question. We really want to put our own drugs in the clinic. Increasingly, companies are recognizing that the way to have an impact even as a genomics or discovery platform technology is to own your own programs and see them into the clinic and seeing an impact on patients.

And so, aside from really any type of platform metrics that we might generate in terms of how many targets we find or how big our data sets are, the real metric that matters is, are we making drugs that meaningfully change disease for potentially, in some cases, millions of people around the world for some of the diseases that we’re looking at. And so, having that impact. Getting drugs into the clinic. Finding some wholly new strategies for obesity, which some market analysts are calling a hundred-billion-dollar market. Having a company like Fauna bring those new ideas to a company like Lilly, which is now the biggest Pharma company in the world, really shows the power of the approach that we are taking, which is quite differentiated from most other approaches.

We really want to see those impacts on patients. And, hopefully, we’ll get there within the next few years. We’ll start to put drugs in the clinic. Timelines are always variable. But our internal programs are moving along quite nicely. We are very hopeful in that regard.

[00:25:56] HC: This has been great, Ashley. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[00:26:03] AZ: Yeah. We actually just relaunched our website. It’s a good time to go to the website. We actually have a lot of blog posts on the website from each of our founders. Have our founder’s story there in terms of how we think about the company. Some new press around our Lilly partnership and some nice links out to references around that. And, actually, some nice links to the Zoonomia Consortium that I was mentioning before. And so, the background on comparative genomics and why it’s important. Happy reply to people on LinkedIn. My email is ashley@faunabio. It’s very easy to hack. But happy for people to follow up that have questions.

[00:26:37] HC: Perfect. Thanks for joining me today.

[00:26:39] AZ: Yeah. Of course. Happy to do it.

[00:26:41] HC: All right, everyone, thanks for listening. I’m Heather Couture. And I hope we join me again next time for Impact AI.

[OUTRO]

[00:26:50] 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 and planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]