[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]
[0:00:34] HC: Today, I'm joined by guest, Guru Banavar, founding CTO and Head of Discovery AI at Viome, to talk about the human microbiome. Guru, welcome to the show.
[0:00:43] GB: It's great to be here, Heather. Thank you.
[0:00:45] HC: Guru, could you share a bit about your background and how that led you to create Viome?
[0:00:49] GB: Absolutely. I have been in the field of AI since – I would say maybe the late eighties. So long-term commitment to the field, and I've seen the ups and downs of the field, the winters, and the springs, and the resurgence, and so on. After my Ph.D., I joined IBM Research in the mid-nineties, and had a lot of very interesting analytics-oriented projects, also distributed systems-oriented projects. But I eventually found myself to be the VP, the founding VP of the AI group at IBM Research. This was in the sort of the late 2000s, early 2010s, when there was a pretty interesting turn of events at IBM, where we decided that we would invest in creating the next generation of AI systems, which many people now know as Watson.
Of course, the project that got Watson into the limelight was the Jeopardy Watson System, which at the time, of course, there was a large group of people who were very much involved in that project. And 2011, there was the breakthrough Jeopardy live show where the machine beat the world champions on a very difficult and natural language processing base system. So that was kind of the launch of a very important, I would say, business initiative at IBM, which I ended up leading. There was a ton of interesting verticals, but a ton of horizontal technologies that was built as part of that project and follow on, which I was part of.
One of the verticals in that set of activities was in the area of life sciences and healthcare. It turns out, that computational biology, which was sort of a long-running, long-term field ended up sort of merging into AI, if you will, and applied specifically to healthcare life sciences. Over the first few years of my tenure as an AI leader at IBM Watson, I literally fell in love with the field of computational biology, and AI put together. At some point in time, I decided that the impact that I wanted to make in the field of computational biology, life sciences, and healthcare could be done only if I joined a few of my friends from the broader community, and started a new company, which we did in at the end of 2016, beginning of 2017, we started Viome. That's how I got to Viome. I jumped ship from IBM.
I said, this is the right way to solve the really complex and probably one of the defining problems of the 21st century, which is chronic disease. So I decided that we need to build Viome from the ground up as an AI company. Because biology, at the end of the day, is super complicated, and it is very much an information science, as much as it is a biological science or life science. It is an information science. We need data and AI as a critical part of the infrastructure that is needed to solve the big problems of biology, life science, chronic diseases, and so on. That was my motivation for jumping ship. Over the last six or seven years, we have made a ton of progress, and we have, I believe, created sort of a new paradigm for addressing the big problems of healthcare and life sciences.
[0:04:51] HC: What does Viome do and why is this so important for chronic disease?
[0:04:55] GB: Yes. First of all, chronic diseases are the majority of the diseases that we face as humanity today. In the whole world, every other person you might meet has some kind of chronic disease, whether it's a metabolic disease or a gastrointestinal or digestive issue, or it's an immune system issue, cardiovascular, mental health, all of these chronic diseases. Acute problems are very well set up to be handled by the traditional healthcare system. But this chronic disease, which is maybe 70%, 80% of all of the diseases that we face is not handled very well by the traditional health care system.
We need a different way of thinking about it. The traditional healthcare system is focused on infectious diseases from the past century, antibiotics, and all of those things. Then, of course, acute problems like trauma, for example, you need surgery, you need other kinds of very immediate solutions to problems. Those are the kinds of things that traditional healthcare is set up. But in order to address chronic disease, you need to look at the very long term, you need to look at decades, not even years. You need to look at decades of somebody's life.
The modern lifestyle has brought us to a point where there's an overabundance of a lot of things. There's a lot of food. There's an extra amount of calories. Probably, the standard American diet is really bad at helping us maintain our health. Then, there is a lot of stress, there's very little sleep. All kinds of issues are kind of conspiring to create a set of chronic diseases like I've mentioned before, that can only be addressed if you take a long-term preventative lifestyle-oriented approach. And you dig deep into the molecular biology of individuals. It's not going to be enough if you just look at the symptoms. Because many of these chronic diseases don't have symptoms externally, you need to look deep inside, and you need to see what's going on at the molecular level.
That's one of the starting points for Viome. When we say molecular insights, which is what we provide at Viome, we are specifically looking at RNA, meaning gene expression, not DNA. DNA is just your genome, and all the genes that you were born with, and you probably die with. You have the same set of genes that you have, and there may be some slight modifications, but there are a lot of repair mechanisms that are constantly fixing it. But RNA, which is maybe a few tenths of the percent of your genome that gets expressed is changing constantly, depending on your lifestyle.
When you eat a meal, when you get stressed, when you travel, when you do exercise, each one of those activities, in fact, creates a different set of genes that get expressed. What we decided at Viome was, that we would measure and we would analyze these expressed genes, which are RNA molecules. It's not easy to measure and analyze RNA molecules for a number of reasons. One of them is that RNA is fundamentally not a long-term stable molecule. You need to capture it, preserve it, and you need to analyze it, keeping in mind that there are going to be all of these issues in the collection phase, which we've already solved. We can preserve an RNA-based sample for 28 days, and we've published papers about this.
Once you sequence that data, you end up getting a very different type of data set, compared to a general genomic DNA data set, even when you look at metagenomics, you get the full set of genomes of all of the organism, not just the human organism, but also all of the microbial organisms. That's an important element of being human. You have a lot of microbes living inside of you, and on top of you, in your gut, in your mouth, on your skin. Everywhere in your body, you have microbes living on you and inside you. We need to be able to capture the activities, meaning the gene expression of the microbes themselves. Not only the human gene expression, but also the microbial gene expression. Because the microbes and the humans are constantly sort of interacting with each other to metabolize a lot of the things that are going on in our body. Not just the food.
When you eat your food, the first set of organisms that process the food — are your microbes, your microbiome in your gut. And then, the metabolites and the proteins that come out of that whole process are somehow metabolized and digested by the human system, and then it gets into your bloodstream, and gets into the rest of your body, and so on. So understanding the whole molecular biology of RNA is a very big part of what we do in Viome.
One last part I would say is, the business model that we set up for Viome is also tremendously new. It is a direct-to-consumer business model. That's how we started. And we've collected a lot of data from our customers. Our customers get an app, and they send us samples from their homes, whether it's a stool sample, a blood sample, a saliva sample, send us samples from home. They get a kit at home and they purchase, a solution from us, which we are an e-commerce company, basically.
When they send us the sample, we process it, we send them back molecular insight, we send them back recommendations for food, for lifestyle, for supplements, and everything else. That process of getting data from customers also helps us do the R&D, and the product innovation that we need to create. So we have this massive data set, we have a product and R&D innovation flywheel. That business model has also now reached a level of maturity that we have almost 600,000 samples from more than 100 countries now around the world. That's a massive data set.
That data set helps us discover lots of interesting and very insightful new biological phenomena, biomarkers, targets, pathways, and so on, and so forth, which I'm happy to talk about when we look at this really massive data set, along with their phenotypes. Meaning, all of the symptoms, the lifestyle, the demographics, their diets, their medications, we get all of that data. So that's the metadata, and the molecular data is the base data, and we are able to combine the two to create a ton of insights that can be used for delivering value to customers through food recommendations, and supplement recommendations, and other things.
[0:11:58] HC: How do you use machine learning here? How do you turn RNA data into insights for your customers?
[0:12:03] GB: Heather, I am one of those AI people who believes that you first focus on the problem, and you bring all of the tools you need in order to solve the problem. AI to me is not just one thing, like the latest buzzword, like generative AI, or something like that. For me, AI is an ML, is a set of tools, and you take the right tool for the right problem. In the case of the problem that we have focused on, which is to address chronic disease. It turns out that you really need to focus on the full range of AI and machine learning toolsets. We use knowledge-driven or model-driven techniques. We also use data-driven techniques.
For example, we use a ton of domain knowledge captured as ontologies, and then we process it through reasoning algorithms of different types, and optimization algorithms. That's one part of the toolset. We also use sort of the classic supervised, and unsupervised types of machine learning for our molecular data. Because remember, we have the phenotypes, which is the metadata for our molecular data. So we can use supervised learning for differentiating between a disease cohort versus a healthy cohort for any disease, right? We have dozens of diseases for which we have analyzed the differences in the molecular data between a disease cohort and a healthy cohort.
We also use a lot of unsupervised techniques, clustering, and various types of analysis techniques for using unsupervised learning has helped us to learn about molecular or biological pathways. Now, of course, in the last few years, there's been a ton of progress with new architectures that have helped the world of generative AI take off. The transformer architecture, which of course, there's one set of basic ideas, but we also use potentially variants of the transformer architecture to apply that to molecular data. Applying something like transformers, not just the language, which we also do, by the way. We can learn from the written-down knowledge of the field of biology using a large language model, but we are also now exploring how to use our massive molecular data set to train transformer-based models.
That whole array, so everything from, I said, knowledge-driven, model-driven, data-driven generative AI. That entire set of AI and machine learning tools are applied in this context of solving the problem and providing value to customers.
[0:14:49] HC: What kind of challenges do you encounter in working with this RNA data and the large data set that you've collected from customers? In particular, what challenges do you encounter in training models based off of it?
[0:15:01] GB: Yes. First of all, I want to make sure that everyone understands that molecular data is super high-dimensional data. You can imagine millions of features in a given sample, in order to understand why that's the case. When we collect a sample, and when we get the output of a sequencer, after the RNA molecules have been detected, the first set of raw data that we get are known as reads. That could be in the tens of millions of reads, which then need to be bioinformatically interpreted or compressed if you want to think of it like that. Into all of the transcripts, meaning the genes that are expressed within the human organism, as well as the microbial organisms.
Of course, we also detect the genomes of these organisms. Meaning that we need to – one of the algorithms that we use is called alignment in bioinformatics. You take every one of these reads, which could be a few 100 nucleotides long, and align them to either genes or genomes that are maintained in large catalogs. That's a massively complicated algorithm that requires a ton of technological innovation, including performance, cost, and so forth on the cloud data pipelines, and so forth. We do all of that.
Then, we try to interpret even after getting the genes and genomes, we are looking at millions of features in a single sample. And let alone, the number of samples we have, which is millions of samples themselves, right? Then, we look at how to interpret the pathways from those expressed genes, and that requires a level of domain expertise and domain knowledge. So you understand what kinds of molecules and enzymes, for example, are interacting with what other kinds of molecules to produce a certain metabolite. For example, if you want to see how short-chain fatty acids are produced in your gut, through the gut microbiome, starting from, let's say, complex carbohydrates, fibers that are – it goes through maybe about 10 or 15 steps in the biochemistry of these complex carbohydrates before it gets turned into short-chain fatty acids, which are metabolized, that are very good for anti-inflammatory properties in your gut.
Understanding those pathways requires both domain knowledge as well as multiple types of machine learning in order to understand the pathways. Then finally, you look at the phenotypes, you look at people's symptoms, people's lifestyles, and medications, and so forth, and you're able to then figure out what are the pathways that are active in people who have metabolic disease, or who have autoimmune diseases, or mental health diseases. For example, serotonin, or GABA, or one of those important molecules in the body. Those are the kinds of problems that you want to solve in order to address chronic disease.
If you just look at sort of a hardcore machine learning lens, there are a number of unique issues. For example, when you look at gene expression data, when you look at large, very high dimensional molecular data, especially with the microbiome, there's a problem called compositionality, which is that, you don't want to look at absolute values, absolute levels of expression when you're quantifying the expression of genes, for example. But instead, you want to look at the relative activity across a single sample, meaning all the features across a single sample, versus across the multiple samples. When you look at all of those data, let's say analysis problems, it is actually a pretty difficult dataset.
Then finally, I want to say we do a lot of very classic scientific methods. When you're looking at a group of people with a disease versus a group of people without that disease, you have to be super careful because you're applying the fundamental scientific method, and you have a hypothesis, or sometimes you don't have a hypothesis, right? Both cases are true. You could just do data-driven experiments, where you want to control for confounders of different types between the two groups. When you control for confounders and – ideally, of course, we would do a randomized control trial, even within the data set that we do. But even when you're not able to do randomized control trials, you do observational retrospective analysis, you want to control for confounders. So we apply all of those techniques from classic epidemiology and causal inference in order to understand what are the drivers of the molecular pathways for a given disease.
Those are the spectrum of problems. It's not just a tech problem. This is a science and tech problem, because you need to understand both the scientific part, the technological part, and the eventual application that needs to be interpretable, and that needs to be explained in the beginning. We are explaining it to common users. These are direct to consumer type of applications we are talking about. We need to be able to translate all of this complex technology and science into understandable consumer-level information. That's itself a challenge. All of these are the types of challenges that we have in understanding molecular biology and connecting that to chronic disease. And then making something useful for our customers such as, here are the foods that you should eat more of, here are the foods that you should not eat. I can take a minute to explain what that means, but that's the set of challenges we have.
[0:20:59] HC: Those are a number of areas of expertise that really are important. Like you said, it's not just the machine learning. All of this has to come together in order to train useful models and make something that's useful for the end user. How does your team plan and develop a new machine-learning product or feature? In particular, what are some of the early steps you take in that process?
[0:21:21] GB: Yes. First of all, I want to say that we have an entire stack of molecular technology stack, so to speak, from the bottom to top. At the very bottom of the stack, we have laboratory technologies, where we figure out how to process the samples and turn the physical molecules into digital information, the reads, the sequences, and the genes and genomes that can be then digitally processed. That is a fundamental data generation layer that is absolutely critical.
Then the next level up is the molecular analysis layer, which includes the bioinformatics I talked about earlier. That requires a lot of very complicated scaling, accuracy issues, and so on. But it also includes the pathway analysis, the biomarker detection, the target identification, and those kinds of techniques in that second layer. In the third layer, we have a number of different models, we have engines, we have ontologies, and we have even clinical research drivers. Like we could recruit people who come to Viome and consent to join one of our clinical studies or clinical trials that are going on. All of these layers are important to keep in mind when I talk about how we plan out our machine-learning projects, how we do the early phases, and so on.
We usually start with the health or life science challenge that we're trying to address as an example. If, for example, we see that we need to learn how to address sleep disorders in our customer population. When people joined Viome, we asked them a number of questions. For example, we asked them, what was your goal? What was your reason for joining Viome? Then, we asked things like, what are your top issues? We see that a lot of people have metabolic issues. Meaning, weight gain, or some other issue now with the metabolism. But a number of people in the, I would say, in the top five, you also see something like sleep issues. Many people are not able to sleep well, whether it's insomnia, or sleep apnea, whatever reason it is. People don't have a great night's sleep.
If we see that in our consumer questionnaires, then we decide that, yes, we need to focus on that problem. Now, when you focus on the sleep problem, there is a lot of existing data that you can analyze, because we have maybe tens of thousands of people who have already told us that they have sleep issues. One of the first things we can do is, literally isolate the group of people that have told us that they have sleep problems, and then compare that group of people with people who've told us that they don't have sleep issues. Meaning they sleep really well at night.
One of the first steps we would want to do in that case is to understand the differences in the gene expression, in the gut through the stool samples that we get in the blood, which is the blood samples that we get. It's a finger-prick blood sample. So people have sent us their human transcripts and their microbial transcripts through those two samples. We can understand the difference between people who have sleep problems versus people who don't have sleep problems, right? That is a good starting point for us. In order to appreciate how difficult just that kind of a simple analysis is, again, think about the high dimensional data set we're talking about.
Now, you have these two groups and you have to figure out what are the critical molecular pathways that are different between these two groups. In order to do that, you need to apply a number of different tools that I mentioned earlier. A simple one would be all of the data curation and cleaning work that you need to do for the molecular data itself. Then, you start focusing on what kinds of pathways are expressed in the people with sleep issues, and the people without sleep issues. In order to figure out what pathways are expressed, there's a number of standardized annotation database. For example, the Kyoto Encyclopedia of Genes and Genomes, KEGG database has a lot of domain data about different pathways that are already identified by a lot of the scientific community.
We basically go through the catalog, and we figure out which of those pathways are expressed in the people with the sleep problems, the people without sleep problems, right. Then, you get to a maybe a reasonably low dimensional data set that you can meaningfully analyze, and interpret, and say, "Okay. You know what? These are the issues that are problematic in the people with sleep issues." We have a clinical research team, we have a biomedical team, and we also get to them, and we start asking questions about what are the ingredients in food, and medications that could be impacting those pathways, that's a little bit downstream from the machine learning work. But the machine learning work has to produce interpretable information, such that our clinical research and our clinical nutrition team is able to understand and figure out what kinds of molecular substrates in nutrition could be impacting these pathways.
That's sort of the set of things that that we think about when we want to solve a problem. So it's very much problem-driven. That's kind of how I want to summarize the whole thing. But it requires a ton of not just horizontal AI machine learning technologies, but it requires a significant amount of domain knowledge in order to be able to narrow the problem to things that matter. We're not at a at a place where we can simply have emergent AI-driven insights, like we have in the case of maybe some of the generative technologies. We're not there yet. We will probably get there in the next few years. Maybe in the next five years, maybe later, I don't know, but we're not there yet.
Until that happens, there is a lot of scientific knowledge that exists in the domain that needs to be leveraged in order to be able to narrow this problem into something that we can solve today for our customers.
[0:27:48] HC: That domain knowledge is essential, and I see that for applications across healthcare, and even climate change, and agriculture, and different spaces like that, when you're into a niche, and trying to solve a very impactful problem. That domain knowledge becomes very critical.
[0:28:05] GB: Agreed. Agreed.
[0:28:06] HC: Are there any specific technological advancements that made it possible to build your technology now? Wouldn't it have been feasible maybe even a few years ago?
[0:28:14] GB: Yes. I would say that I'm going to interpret the word technology very broadly here. Okay? I'm not just going to talk about digital technologies. I'm also going to talk about laboratory technologies. We essentially invented and scaled a few laboratory technologies, and we also invented and scaled a few digital technologies in order to create the molecular technology stack that I talked about earlier on.
One of our core intellectual property elements is the meta-transcriptomic laboratory technology, which essentially, isolates, detects, and processes what we call the informative RNA molecules in any given sample. That required a number of sort of biochemistry-level technology breakthroughs. And eventually, the processing breakthroughs involve things like minimizing what we call sample-to-sample crosstalk, figuring out how to remove the non-informative molecules in a sample, and also improving the accuracy of this whole methodology way beyond. I'm talking about 10,000 times better than other existing technologies out there in the world of sequencing and the previous generation.
Our generation which is, I maintain our lab technology is the absolute latest generation and 99% or more of the entire healthcare life sciences industry is still in the previous generation or two generations ago. In our latest generation, which we have many patents on, we have figured out how to scale that to the point where it can be available as a consumer product. If you just go to a research center to get the same laboratory processing done, it'll cost you 10 times more. It will be maybe thousands of times less accurate. That's the first set of breakthroughs that we have to make in the laboratory technology space. On the digital side, just getting the bioinformatics to work on these massive dimensional data sets, and do the alignment, and do it at a cost point that is not going to break the bank took a lot of work.
We've come up with algorithmic improvements, and I'd call them some know-how and how to pull together a stack and a pipeline of data and algorithms together, that enables a very optimized, efficient way to scale this to the point where we've now reduced the cost and the time taken by an order of magnitude over the last few years. Now, there's also some new recommendation-type technologies that uses molecular data and food ontologies, that together create the kinds of both precision food and precision supplements that we are able to manufacture for each customer of Viome.
When a customer comes to Viome, we send them precision supplements that is made for each individual. It's not made for a big group of people, it's for you, and you alone, based on all the pathways, and all of the molecular patterns we see in your gut, and your mouth, in your blood, and so on. All of those things required a suite of algorithmic improvements, and scaling, and bringing down the cost to the point where this can actually be sold at a consumer product level. Those are the kinds of things that I would bring to the forum.
[0:31:57] HC: Is there any advice you could offer to other leaders of AI-powered startups?
[0:32:00] GB: Yes. In general, if I kind of step back from all of the work we've done, my first piece of advice would be focused on the problem, right? I mean, don't just pick up a solution that exists out there, like a technology, and then make that a hammer and go looking for nails, and everything might start looking like nails and you start applying and you figure out that it doesn't really work. In our case, what we ended up doing was we focused on a problem that was massive, and that the entire world was taking, in my opinion, the wrong approach. The basic big health care system was broken, because they were focused on acute and infectious type diseases, not on chronic disease, right? That's a big, huge problem right there.
It impacts such a large part of the global population that that is one of the biggest problems of the 21st century, so to speak. Start with the big problems. When you figure out what kinds of applications can address that big problem, you look at the tools that you have in mind, and you pull out the tools that are necessary to solve that problem. And of course, along the way, you'd have to innovate quite a lot. I mean, there's a lot of tech hype, there's a lot of AI hype that happens. A lot of people kind of try to sell every new hyped-up technology as the solution to everything. I've learned that that is definitely not the case. I would advise other leaders of AI-powered startups to be very careful about how you pick your solution toolset, based upon the problem that you want to solve. That's kind of my biggest advice.
Now, if you look at a healthcare and life science space, there's a lot of other issues, which is, a lot of the science, a lot of the data that's out there is really poor quality. The studies are not very well designed, they're usually very low powered, and there's a lot of hypotheses that are going around with very little additional validation, but people kind of somehow propagate it. They become almost myths, and it just becomes something so widespread that it's difficult to get out of it.
I would question the fundamentals if possible. If you're in a science-based domain, I would question some of the fundamentals, and try to address those fundamentals using the latest available technologies, and the available data and science. Those are my two things. Focus on the problem, question the basics.
[0:34:35] HC: Finally, where do you see the impact of Viome in three to five years?
[0:34:39] GB: Okay. I would love for us to be able to create a new model of health care to handle the chronic disease epidemic in a few years. Okay. What that means to me is that we need to be able to address these longer-time horizon issues, which can only be done at home. These are lifestyle factors. You cannot ask somebody to keep going to a hospital, or a clinic for chronic diseases because this happens every single day in your life, when you make the decision of what to eat, where to go, how stressed, how much you sleep, how much you exercise. Those are the things that make a difference to your chronic health, from chronic disease. I think we need to learn chronic health.
That starts at home, and it has to be prevented. It's not the usual system thing today, of solving the problem when the heart attack happens, and you extend somebody's life by another 10 years, but they have such a poor quality of life for those last 10 years. What if you could look ahead a few decades before cardiovascular disease starts, and you figure out how to prevent it, and how to manage it in such a way that you have a health span that is the same as your lifespan, meaning that you're healthy, and you have a great quality of life until the time that you die. The time that you die is going to be extended by quite a bit, compared to what you normally would have if you hadn't done this new model of healthcare.
Finally, I think this new model of healthcare is continuous. This is not just going to be like a one-shot thing, it becomes part of life, it becomes a coach, a friend, angel on your shoulder, so to speak, that is constantly with you, with the technologies we have who's helping you, the technology, the AI is helping you live your life in such a way that you can prevent chronic disease. You can have the best life has to offer, you have a high quality of life for the longest period of time possible. That's where I would see the impact of Viome. If that happens in the next three to five years, I'd be the happiest guy on the planet.
[0:36:52] HC: This has been great, Guru. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?
[0:37:00] GB: Well, check me out on LinkedIn. My last name is Banavar, B-A-N-A-V-A-R, and first name is Guruduth, G-U-R-U-D-U-T-H. That should be easy to find. Also, of course, if you check out our company website, V-I-O-M-E.com, you can see the solutions we have available today. We also have a number of blogs where we put out all the science and the technology we develop, so you can check that out as well. Finally, of course, if you want to follow or tweet to me or to us at Viome, feel free to do so and we will be happy to connect with you and respond. Thank you very much.
[0:37:41] HC: Perfect. I'll link to those in the show notes. Thank you for joining me today.
[0:37:44] GB: Thanks, Heather.
[0:37:45] HC: All right everyone. Thanks for listening. I'm Heather Couture, and I hope you join me again next time for Impact AI.
[OUTRO]
[0:37:55] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe, and share with a friend. 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]
[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]
[0:00:34] HC: Today, I'm joined by guest, Guru Banavar, founding CTO and Head of Discovery AI at Viome, to talk about the human microbiome. Guru, welcome to the show.
[0:00:43] GB: It's great to be here, Heather. Thank you.
[0:00:45] HC: Guru, could you share a bit about your background and how that led you to create Viome?
[0:00:49] GB: Absolutely. I have been in the field of AI since – I would say maybe the late eighties. So long-term commitment to the field, and I've seen the ups and downs of the field, the winters, and the springs, and the resurgence, and so on. After my Ph.D., I joined IBM Research in the mid-nineties, and had a lot of very interesting analytics-oriented projects, also distributed systems-oriented projects. But I eventually found myself to be the VP, the founding VP of the AI group at IBM Research. This was in the sort of the late 2000s, early 2010s, when there was a pretty interesting turn of events at IBM, where we decided that we would invest in creating the next generation of AI systems, which many people now know as Watson.
Of course, the project that got Watson into the limelight was the Jeopardy Watson System, which at the time, of course, there was a large group of people who were very much involved in that project. And 2011, there was the breakthrough Jeopardy live show where the machine beat the world champions on a very difficult and natural language processing base system. So that was kind of the launch of a very important, I would say, business initiative at IBM, which I ended up leading. There was a ton of interesting verticals, but a ton of horizontal technologies that was built as part of that project and follow on, which I was part of.
One of the verticals in that set of activities was in the area of life sciences and healthcare. It turns out, that computational biology, which was sort of a long-running, long-term field ended up sort of merging into AI, if you will, and applied specifically to healthcare life sciences. Over the first few years of my tenure as an AI leader at IBM Watson, I literally fell in love with the field of computational biology, and AI put together. At some point in time, I decided that the impact that I wanted to make in the field of computational biology, life sciences, and healthcare could be done only if I joined a few of my friends from the broader community, and started a new company, which we did in at the end of 2016, beginning of 2017, we started Viome. That's how I got to Viome. I jumped ship from IBM.
I said, this is the right way to solve the really complex and probably one of the defining problems of the 21st century, which is chronic disease. So I decided that we need to build Viome from the ground up as an AI company. Because biology, at the end of the day, is super complicated, and it is very much an information science, as much as it is a biological science or life science. It is an information science. We need data and AI as a critical part of the infrastructure that is needed to solve the big problems of biology, life science, chronic diseases, and so on. That was my motivation for jumping ship. Over the last six or seven years, we have made a ton of progress, and we have, I believe, created sort of a new paradigm for addressing the big problems of healthcare and life sciences.
[0:04:51] HC: What does Viome do and why is this so important for chronic disease?
[0:04:55] GB: Yes. First of all, chronic diseases are the majority of the diseases that we face as humanity today. In the whole world, every other person you might meet has some kind of chronic disease, whether it's a metabolic disease or a gastrointestinal or digestive issue, or it's an immune system issue, cardiovascular, mental health, all of these chronic diseases. Acute problems are very well set up to be handled by the traditional healthcare system. But this chronic disease, which is maybe 70%, 80% of all of the diseases that we face is not handled very well by the traditional health care system.
We need a different way of thinking about it. The traditional healthcare system is focused on infectious diseases from the past century, antibiotics, and all of those things. Then, of course, acute problems like trauma, for example, you need surgery, you need other kinds of very immediate solutions to problems. Those are the kinds of things that traditional healthcare is set up. But in order to address chronic disease, you need to look at the very long term, you need to look at decades, not even years. You need to look at decades of somebody's life.
The modern lifestyle has brought us to a point where there's an overabundance of a lot of things. There's a lot of food. There's an extra amount of calories. Probably, the standard American diet is really bad at helping us maintain our health. Then, there is a lot of stress, there's very little sleep. All kinds of issues are kind of conspiring to create a set of chronic diseases like I've mentioned before, that can only be addressed if you take a long-term preventative lifestyle-oriented approach. And you dig deep into the molecular biology of individuals. It's not going to be enough if you just look at the symptoms. Because many of these chronic diseases don't have symptoms externally, you need to look deep inside, and you need to see what's going on at the molecular level.
That's one of the starting points for Viome. When we say molecular insights, which is what we provide at Viome, we are specifically looking at RNA, meaning gene expression, not DNA. DNA is just your genome, and all the genes that you were born with, and you probably die with. You have the same set of genes that you have, and there may be some slight modifications, but there are a lot of repair mechanisms that are constantly fixing it. But RNA, which is maybe a few tenths of the percent of your genome that gets expressed is changing constantly, depending on your lifestyle.
When you eat a meal, when you get stressed, when you travel, when you do exercise, each one of those activities, in fact, creates a different set of genes that get expressed. What we decided at Viome was, that we would measure and we would analyze these expressed genes, which are RNA molecules. It's not easy to measure and analyze RNA molecules for a number of reasons. One of them is that RNA is fundamentally not a long-term stable molecule. You need to capture it, preserve it, and you need to analyze it, keeping in mind that there are going to be all of these issues in the collection phase, which we've already solved. We can preserve an RNA-based sample for 28 days, and we've published papers about this.
Once you sequence that data, you end up getting a very different type of data set, compared to a general genomic DNA data set, even when you look at metagenomics, you get the full set of genomes of all of the organism, not just the human organism, but also all of the microbial organisms. That's an important element of being human. You have a lot of microbes living inside of you, and on top of you, in your gut, in your mouth, on your skin. Everywhere in your body, you have microbes living on you and inside you. We need to be able to capture the activities, meaning the gene expression of the microbes themselves. Not only the human gene expression, but also the microbial gene expression. Because the microbes and the humans are constantly sort of interacting with each other to metabolize a lot of the things that are going on in our body. Not just the food.
When you eat your food, the first set of organisms that process the food — are your microbes, your microbiome in your gut. And then, the metabolites and the proteins that come out of that whole process are somehow metabolized and digested by the human system, and then it gets into your bloodstream, and gets into the rest of your body, and so on. So understanding the whole molecular biology of RNA is a very big part of what we do in Viome.
One last part I would say is, the business model that we set up for Viome is also tremendously new. It is a direct-to-consumer business model. That's how we started. And we've collected a lot of data from our customers. Our customers get an app, and they send us samples from their homes, whether it's a stool sample, a blood sample, a saliva sample, send us samples from home. They get a kit at home and they purchase, a solution from us, which we are an e-commerce company, basically.
When they send us the sample, we process it, we send them back molecular insight, we send them back recommendations for food, for lifestyle, for supplements, and everything else. That process of getting data from customers also helps us do the R&D, and the product innovation that we need to create. So we have this massive data set, we have a product and R&D innovation flywheel. That business model has also now reached a level of maturity that we have almost 600,000 samples from more than 100 countries now around the world. That's a massive data set.
That data set helps us discover lots of interesting and very insightful new biological phenomena, biomarkers, targets, pathways, and so on, and so forth, which I'm happy to talk about when we look at this really massive data set, along with their phenotypes. Meaning, all of the symptoms, the lifestyle, the demographics, their diets, their medications, we get all of that data. So that's the metadata, and the molecular data is the base data, and we are able to combine the two to create a ton of insights that can be used for delivering value to customers through food recommendations, and supplement recommendations, and other things.
[0:11:58] HC: How do you use machine learning here? How do you turn RNA data into insights for your customers?
[0:12:03] GB: Heather, I am one of those AI people who believes that you first focus on the problem, and you bring all of the tools you need in order to solve the problem. AI to me is not just one thing, like the latest buzzword, like generative AI, or something like that. For me, AI is an ML, is a set of tools, and you take the right tool for the right problem. In the case of the problem that we have focused on, which is to address chronic disease. It turns out that you really need to focus on the full range of AI and machine learning toolsets. We use knowledge-driven or model-driven techniques. We also use data-driven techniques.
For example, we use a ton of domain knowledge captured as ontologies, and then we process it through reasoning algorithms of different types, and optimization algorithms. That's one part of the toolset. We also use sort of the classic supervised, and unsupervised types of machine learning for our molecular data. Because remember, we have the phenotypes, which is the metadata for our molecular data. So we can use supervised learning for differentiating between a disease cohort versus a healthy cohort for any disease, right? We have dozens of diseases for which we have analyzed the differences in the molecular data between a disease cohort and a healthy cohort.
We also use a lot of unsupervised techniques, clustering, and various types of analysis techniques for using unsupervised learning has helped us to learn about molecular or biological pathways. Now, of course, in the last few years, there's been a ton of progress with new architectures that have helped the world of generative AI take off. The transformer architecture, which of course, there's one set of basic ideas, but we also use potentially variants of the transformer architecture to apply that to molecular data. Applying something like transformers, not just the language, which we also do, by the way. We can learn from the written-down knowledge of the field of biology using a large language model, but we are also now exploring how to use our massive molecular data set to train transformer-based models.
That whole array, so everything from, I said, knowledge-driven, model-driven, data-driven generative AI. That entire set of AI and machine learning tools are applied in this context of solving the problem and providing value to customers.
[0:14:49] HC: What kind of challenges do you encounter in working with this RNA data and the large data set that you've collected from customers? In particular, what challenges do you encounter in training models based off of it?
[0:15:01] GB: Yes. First of all, I want to make sure that everyone understands that molecular data is super high-dimensional data. You can imagine millions of features in a given sample, in order to understand why that's the case. When we collect a sample, and when we get the output of a sequencer, after the RNA molecules have been detected, the first set of raw data that we get are known as reads. That could be in the tens of millions of reads, which then need to be bioinformatically interpreted or compressed if you want to think of it like that. Into all of the transcripts, meaning the genes that are expressed within the human organism, as well as the microbial organisms.
Of course, we also detect the genomes of these organisms. Meaning that we need to – one of the algorithms that we use is called alignment in bioinformatics. You take every one of these reads, which could be a few 100 nucleotides long, and align them to either genes or genomes that are maintained in large catalogs. That's a massively complicated algorithm that requires a ton of technological innovation, including performance, cost, and so forth on the cloud data pipelines, and so forth. We do all of that.
Then, we try to interpret even after getting the genes and genomes, we are looking at millions of features in a single sample. And let alone, the number of samples we have, which is millions of samples themselves, right? Then, we look at how to interpret the pathways from those expressed genes, and that requires a level of domain expertise and domain knowledge. So you understand what kinds of molecules and enzymes, for example, are interacting with what other kinds of molecules to produce a certain metabolite. For example, if you want to see how short-chain fatty acids are produced in your gut, through the gut microbiome, starting from, let's say, complex carbohydrates, fibers that are – it goes through maybe about 10 or 15 steps in the biochemistry of these complex carbohydrates before it gets turned into short-chain fatty acids, which are metabolized, that are very good for anti-inflammatory properties in your gut.
Understanding those pathways requires both domain knowledge as well as multiple types of machine learning in order to understand the pathways. Then finally, you look at the phenotypes, you look at people's symptoms, people's lifestyles, and medications, and so forth, and you're able to then figure out what are the pathways that are active in people who have metabolic disease, or who have autoimmune diseases, or mental health diseases. For example, serotonin, or GABA, or one of those important molecules in the body. Those are the kinds of problems that you want to solve in order to address chronic disease.
If you just look at sort of a hardcore machine learning lens, there are a number of unique issues. For example, when you look at gene expression data, when you look at large, very high dimensional molecular data, especially with the microbiome, there's a problem called compositionality, which is that, you don't want to look at absolute values, absolute levels of expression when you're quantifying the expression of genes, for example. But instead, you want to look at the relative activity across a single sample, meaning all the features across a single sample, versus across the multiple samples. When you look at all of those data, let's say analysis problems, it is actually a pretty difficult dataset.
Then finally, I want to say we do a lot of very classic scientific methods. When you're looking at a group of people with a disease versus a group of people without that disease, you have to be super careful because you're applying the fundamental scientific method, and you have a hypothesis, or sometimes you don't have a hypothesis, right? Both cases are true. You could just do data-driven experiments, where you want to control for confounders of different types between the two groups. When you control for confounders and – ideally, of course, we would do a randomized control trial, even within the data set that we do. But even when you're not able to do randomized control trials, you do observational retrospective analysis, you want to control for confounders. So we apply all of those techniques from classic epidemiology and causal inference in order to understand what are the drivers of the molecular pathways for a given disease.
Those are the spectrum of problems. It's not just a tech problem. This is a science and tech problem, because you need to understand both the scientific part, the technological part, and the eventual application that needs to be interpretable, and that needs to be explained in the beginning. We are explaining it to common users. These are direct to consumer type of applications we are talking about. We need to be able to translate all of this complex technology and science into understandable consumer-level information. That's itself a challenge. All of these are the types of challenges that we have in understanding molecular biology and connecting that to chronic disease. And then making something useful for our customers such as, here are the foods that you should eat more of, here are the foods that you should not eat. I can take a minute to explain what that means, but that's the set of challenges we have.
[0:20:59] HC: Those are a number of areas of expertise that really are important. Like you said, it's not just the machine learning. All of this has to come together in order to train useful models and make something that's useful for the end user. How does your team plan and develop a new machine-learning product or feature? In particular, what are some of the early steps you take in that process?
[0:21:21] GB: Yes. First of all, I want to say that we have an entire stack of molecular technology stack, so to speak, from the bottom to top. At the very bottom of the stack, we have laboratory technologies, where we figure out how to process the samples and turn the physical molecules into digital information, the reads, the sequences, and the genes and genomes that can be then digitally processed. That is a fundamental data generation layer that is absolutely critical.
Then the next level up is the molecular analysis layer, which includes the bioinformatics I talked about earlier. That requires a lot of very complicated scaling, accuracy issues, and so on. But it also includes the pathway analysis, the biomarker detection, the target identification, and those kinds of techniques in that second layer. In the third layer, we have a number of different models, we have engines, we have ontologies, and we have even clinical research drivers. Like we could recruit people who come to Viome and consent to join one of our clinical studies or clinical trials that are going on. All of these layers are important to keep in mind when I talk about how we plan out our machine-learning projects, how we do the early phases, and so on.
We usually start with the health or life science challenge that we're trying to address as an example. If, for example, we see that we need to learn how to address sleep disorders in our customer population. When people joined Viome, we asked them a number of questions. For example, we asked them, what was your goal? What was your reason for joining Viome? Then, we asked things like, what are your top issues? We see that a lot of people have metabolic issues. Meaning, weight gain, or some other issue now with the metabolism. But a number of people in the, I would say, in the top five, you also see something like sleep issues. Many people are not able to sleep well, whether it's insomnia, or sleep apnea, whatever reason it is. People don't have a great night's sleep.
If we see that in our consumer questionnaires, then we decide that, yes, we need to focus on that problem. Now, when you focus on the sleep problem, there is a lot of existing data that you can analyze, because we have maybe tens of thousands of people who have already told us that they have sleep issues. One of the first things we can do is, literally isolate the group of people that have told us that they have sleep problems, and then compare that group of people with people who've told us that they don't have sleep issues. Meaning they sleep really well at night.
One of the first steps we would want to do in that case is to understand the differences in the gene expression, in the gut through the stool samples that we get in the blood, which is the blood samples that we get. It's a finger-prick blood sample. So people have sent us their human transcripts and their microbial transcripts through those two samples. We can understand the difference between people who have sleep problems versus people who don't have sleep problems, right? That is a good starting point for us. In order to appreciate how difficult just that kind of a simple analysis is, again, think about the high dimensional data set we're talking about.
Now, you have these two groups and you have to figure out what are the critical molecular pathways that are different between these two groups. In order to do that, you need to apply a number of different tools that I mentioned earlier. A simple one would be all of the data curation and cleaning work that you need to do for the molecular data itself. Then, you start focusing on what kinds of pathways are expressed in the people with sleep issues, and the people without sleep issues. In order to figure out what pathways are expressed, there's a number of standardized annotation database. For example, the Kyoto Encyclopedia of Genes and Genomes, KEGG database has a lot of domain data about different pathways that are already identified by a lot of the scientific community.
We basically go through the catalog, and we figure out which of those pathways are expressed in the people with the sleep problems, the people without sleep problems, right. Then, you get to a maybe a reasonably low dimensional data set that you can meaningfully analyze, and interpret, and say, "Okay. You know what? These are the issues that are problematic in the people with sleep issues." We have a clinical research team, we have a biomedical team, and we also get to them, and we start asking questions about what are the ingredients in food, and medications that could be impacting those pathways, that's a little bit downstream from the machine learning work. But the machine learning work has to produce interpretable information, such that our clinical research and our clinical nutrition team is able to understand and figure out what kinds of molecular substrates in nutrition could be impacting these pathways.
That's sort of the set of things that that we think about when we want to solve a problem. So it's very much problem-driven. That's kind of how I want to summarize the whole thing. But it requires a ton of not just horizontal AI machine learning technologies, but it requires a significant amount of domain knowledge in order to be able to narrow the problem to things that matter. We're not at a at a place where we can simply have emergent AI-driven insights, like we have in the case of maybe some of the generative technologies. We're not there yet. We will probably get there in the next few years. Maybe in the next five years, maybe later, I don't know, but we're not there yet.
Until that happens, there is a lot of scientific knowledge that exists in the domain that needs to be leveraged in order to be able to narrow this problem into something that we can solve today for our customers.
[0:27:48] HC: That domain knowledge is essential, and I see that for applications across healthcare, and even climate change, and agriculture, and different spaces like that, when you're into a niche, and trying to solve a very impactful problem. That domain knowledge becomes very critical.
[0:28:05] GB: Agreed. Agreed.
[0:28:06] HC: Are there any specific technological advancements that made it possible to build your technology now? Wouldn't it have been feasible maybe even a few years ago?
[0:28:14] GB: Yes. I would say that I'm going to interpret the word technology very broadly here. Okay? I'm not just going to talk about digital technologies. I'm also going to talk about laboratory technologies. We essentially invented and scaled a few laboratory technologies, and we also invented and scaled a few digital technologies in order to create the molecular technology stack that I talked about earlier on.
One of our core intellectual property elements is the meta-transcriptomic laboratory technology, which essentially, isolates, detects, and processes what we call the informative RNA molecules in any given sample. That required a number of sort of biochemistry-level technology breakthroughs. And eventually, the processing breakthroughs involve things like minimizing what we call sample-to-sample crosstalk, figuring out how to remove the non-informative molecules in a sample, and also improving the accuracy of this whole methodology way beyond. I'm talking about 10,000 times better than other existing technologies out there in the world of sequencing and the previous generation.
Our generation which is, I maintain our lab technology is the absolute latest generation and 99% or more of the entire healthcare life sciences industry is still in the previous generation or two generations ago. In our latest generation, which we have many patents on, we have figured out how to scale that to the point where it can be available as a consumer product. If you just go to a research center to get the same laboratory processing done, it'll cost you 10 times more. It will be maybe thousands of times less accurate. That's the first set of breakthroughs that we have to make in the laboratory technology space. On the digital side, just getting the bioinformatics to work on these massive dimensional data sets, and do the alignment, and do it at a cost point that is not going to break the bank took a lot of work.
We've come up with algorithmic improvements, and I'd call them some know-how and how to pull together a stack and a pipeline of data and algorithms together, that enables a very optimized, efficient way to scale this to the point where we've now reduced the cost and the time taken by an order of magnitude over the last few years. Now, there's also some new recommendation-type technologies that uses molecular data and food ontologies, that together create the kinds of both precision food and precision supplements that we are able to manufacture for each customer of Viome.
When a customer comes to Viome, we send them precision supplements that is made for each individual. It's not made for a big group of people, it's for you, and you alone, based on all the pathways, and all of the molecular patterns we see in your gut, and your mouth, in your blood, and so on. All of those things required a suite of algorithmic improvements, and scaling, and bringing down the cost to the point where this can actually be sold at a consumer product level. Those are the kinds of things that I would bring to the forum.
[0:31:57] HC: Is there any advice you could offer to other leaders of AI-powered startups?
[0:32:00] GB: Yes. In general, if I kind of step back from all of the work we've done, my first piece of advice would be focused on the problem, right? I mean, don't just pick up a solution that exists out there, like a technology, and then make that a hammer and go looking for nails, and everything might start looking like nails and you start applying and you figure out that it doesn't really work. In our case, what we ended up doing was we focused on a problem that was massive, and that the entire world was taking, in my opinion, the wrong approach. The basic big health care system was broken, because they were focused on acute and infectious type diseases, not on chronic disease, right? That's a big, huge problem right there.
It impacts such a large part of the global population that that is one of the biggest problems of the 21st century, so to speak. Start with the big problems. When you figure out what kinds of applications can address that big problem, you look at the tools that you have in mind, and you pull out the tools that are necessary to solve that problem. And of course, along the way, you'd have to innovate quite a lot. I mean, there's a lot of tech hype, there's a lot of AI hype that happens. A lot of people kind of try to sell every new hyped-up technology as the solution to everything. I've learned that that is definitely not the case. I would advise other leaders of AI-powered startups to be very careful about how you pick your solution toolset, based upon the problem that you want to solve. That's kind of my biggest advice.
Now, if you look at a healthcare and life science space, there's a lot of other issues, which is, a lot of the science, a lot of the data that's out there is really poor quality. The studies are not very well designed, they're usually very low powered, and there's a lot of hypotheses that are going around with very little additional validation, but people kind of somehow propagate it. They become almost myths, and it just becomes something so widespread that it's difficult to get out of it.
I would question the fundamentals if possible. If you're in a science-based domain, I would question some of the fundamentals, and try to address those fundamentals using the latest available technologies, and the available data and science. Those are my two things. Focus on the problem, question the basics.
[0:34:35] HC: Finally, where do you see the impact of Viome in three to five years?
[0:34:39] GB: Okay. I would love for us to be able to create a new model of health care to handle the chronic disease epidemic in a few years. Okay. What that means to me is that we need to be able to address these longer-time horizon issues, which can only be done at home. These are lifestyle factors. You cannot ask somebody to keep going to a hospital, or a clinic for chronic diseases because this happens every single day in your life, when you make the decision of what to eat, where to go, how stressed, how much you sleep, how much you exercise. Those are the things that make a difference to your chronic health, from chronic disease. I think we need to learn chronic health.
That starts at home, and it has to be prevented. It's not the usual system thing today, of solving the problem when the heart attack happens, and you extend somebody's life by another 10 years, but they have such a poor quality of life for those last 10 years. What if you could look ahead a few decades before cardiovascular disease starts, and you figure out how to prevent it, and how to manage it in such a way that you have a health span that is the same as your lifespan, meaning that you're healthy, and you have a great quality of life until the time that you die. The time that you die is going to be extended by quite a bit, compared to what you normally would have if you hadn't done this new model of healthcare.
Finally, I think this new model of healthcare is continuous. This is not just going to be like a one-shot thing, it becomes part of life, it becomes a coach, a friend, angel on your shoulder, so to speak, that is constantly with you, with the technologies we have who's helping you, the technology, the AI is helping you live your life in such a way that you can prevent chronic disease. You can have the best life has to offer, you have a high quality of life for the longest period of time possible. That's where I would see the impact of Viome. If that happens in the next three to five years, I'd be the happiest guy on the planet.
[0:36:52] HC: This has been great, Guru. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?
[0:37:00] GB: Well, check me out on LinkedIn. My last name is Banavar, B-A-N-A-V-A-R, and first name is Guruduth, G-U-R-U-D-U-T-H. That should be easy to find. Also, of course, if you check out our company website, V-I-O-M-E.com, you can see the solutions we have available today. We also have a number of blogs where we put out all the science and the technology we develop, so you can check that out as well. Finally, of course, if you want to follow or tweet to me or to us at Viome, feel free to do so and we will be happy to connect with you and respond. Thank you very much.
[0:37:41] HC: Perfect. I'll link to those in the show notes. Thank you for joining me today.
[0:37:44] GB: Thanks, Heather.
[0:37:45] HC: All right everyone. Thanks for listening. I'm Heather Couture, and I hope you join me again next time for Impact AI.
[OUTRO]
[0:37:55] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe, and share with a friend. 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]
