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.
Foundation Model Assessment – Foundation models are popping up everywhere – do you need one for your proprietary image dataset? Get a clear perspective on whether you can benefit from a domain-specific foundation model.

Transcript:

[INTRO]

[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.

[INTERVIEW]

[00:00:34] HC: Today, I’m joined by guest Leo Grady, Founder and CEO of Jona, to talk about the microbiome. Leo, welcome to the show.

[00:00:41] LG: Thanks, Heather. It’s wonderful to be here.

[00:00:42] HC: Leo, could you share a bit about your background and how that led you to create Jona?

[00:00:46] LG: Yes, absolutely. I did my PhD in computer vision 20 years ago. I spent my whole career in the healthcare industry. I started at Siemens building AI in radiology. Then I was at a company called HeartFlow in the Bay Area building AI in cardiology. Then I was CEO of Paige.AI here in New York building AI in pathology. Then I spent some time on the BC side for a while, working with Jim Breyer, helping him invest in AI companies and healthcare. Then I started Jona about a year and a half ago.

[00:01:20] HC: What does Jona do?

[00:01:21] LG: Jona’s focused on the microbiome, which is all the organisms that live in your body and on your body, consist of this very complex ecosystem of bacteria, fungi, viruses, protists, archaea. What’s really remarkable to me about the microbiome is that it’s been linked to almost every aspect of human health, everything from GI disorders, autoimmune conditions, cancer, liver disease, obesity, longevity, but even Parkinson’s and Alzheimer’s, and depression, autism. All these things have been linked with the microbiome.

Yet despite this incredible science that has been linking the microbiome to all these different diseases and conditions, the literature is very complex. Every single month, there are more than 2,000 papers published on the microbiome. There’s been very limited translation of all of that science into clinical practice, despite the fact that there’s widespread interest among both patients and doctors in this area.

[00:02:32] HC: What role does machine learning play in this technology?

[00:02:34] LG: Well, the complexity of the microbiome, the fact that there are these trillions of organisms, this whole ecosystem that changes over time, and this complex literature, I really felt it was a perfect opportunity for machine learning because it’s a very data-rich environment. You have not only all these different species and strains. But also the DNA and different alterations in DNA can affect metabolic function in different ways. Because it’s such a data-rich environment and so complex, I don’t think any human being is going to be able to look at a microbiome test results or the DNA present in somebody’s microbiome and really understand what’s going on with their health.

Because of the complexity and because the microbiome has been shown to be so essential to underpinning human health, I believe that AI is really going to be the critical technology that allows us to decode the microbiome, really understand what is going on in somebody’s microbiome, and then help us engineer a new state of the microbiome to improve someone’s health.

[00:03:46] HC: You’ve mentioned microbiome data and the medical literature. What types of models do you apply to those, and how do you connect them?

[00:03:55] LG: We started with a product where we launched a cash pay test interpretation product. You can buy this test through [inaudible 00:04:08] medicine. You can buy it through direct to consumers. You can just go on our website and order it. It’s a stool sample. You send it back to the lab. You can do it at home. We do deep shotgun metagenomic sequencing. We identify the bacteria, the fungus, the viruses, everything that’s in your gut down to a strain level.

Then we built this large language model that reads all those papers. What it does is it matches the patterns found in your microbiome to every study that’s ever been published on the microbiome, linking your microbiome to perhaps different diseases, to different symptoms, to different allergies, to different capabilities to digest certain foods. Then the AI does a second thing and goes through the literature a second time and looks at all the potential interventions. It could be diet. It could be lifestyle. It could be supplements or whatnot.

Then we know that, for example, if we were looking at your microbiome, Heather, we could ask the question what happens if Heather goes vegan? What happens if Heather goes keto? What happens if Heather cuts out gluten? What happens if she takes metformin or some other supplement? Based on the literature, we can create a virtual microbiome of your gut. It’s like a digital twin of your microbiome and say, “What does the vegan Heather look like? What does the keto Heather look like?”

Then by looking at the patterns in your virtual microbiome, we can, again, match those against the entire literature. Help you not only understand what’s going on in your gut based on current science but also what you can do to change it to help you achieve your health goals.

[00:05:52] HC: This microbiome data, this is sequencing from the different strains that are detected in the stool sample. What does the sequencing data look like for those who aren’t familiar with it?

[00:06:03] LG: It really comes out of the raw sequencing machines, and these are FASTQ files. Then you run a bioinformatics pipeline to take all of that DNA and match it against databases where we can say this strain of DNA comes from E. coli or this strain comes from Akkermansia or Lactobacillus or whatever it is. Basically, we do the genetic sequencing. Then the bioinformatic step will link it to these different species and genuses and phyla. Then the AI takes the patterns of these different species and matches those up against all the studies in the literature.

[00:06:49] HC: What kinds of challenges do you encounter in working with this microbiome data and in building machine learning models using the species that you’re able to extract from it?

[00:06:58] LG: It’s really challenging. Actually, it was surprisingly challenging. Obviously, we need to be able to read all of these studies effectively. The first challenge was saying, okay, this is a paper about Parkinson’s disease. It was a study on 500 people that had Parkinson’s and 500 people who didn’t. The people who had Parkinson’s were higher in species A, B, and C and lower in species D, E, and F than the people who were healthy, right? Then maybe there were 50 different studies on Parkinson’s comparing healthy people against Parkinsonian people.

Then the first part is just reading all of that effectively and extracting the right content out of it and putting that into a knowledge base. Then the second part, which was also challenging, was now that we’ve extracted the information from, say, 50 different studies on Parkinson’s disease versus healthy people, each of these studies had different cohorts. You have different findings. Now, we’re looking at your microbiome, and we want to say Heather’s microbiome fits the profile of Parkinson’s disease according to these studies.

Then the second challenge was how do we actually combine these studies in a way that makes sense to really answer that question. We ended up having to do some pretty sophisticated modeling in order to be able to extract all that information.

[00:08:34] HC: That’s the patient sampling side of this. The other side is the medical literature. LLMs are quite the thing of the day right now, but are there differences in how you need to apply them to the medical literature? Are the challenges that come up there?

[00:08:47] LG: Yes. The approach to LLMs today, as you know, is you are training on these huge corpora to predict the next word or the next phrase. Then you can use this in a generative context to have a conversation with it or have it write a paper for you or something like that. That’s wonderful, and I think we’ve seen quite a bit of success with that approach to be able to ingest literature and answer questions about it, even being able to pass a variety of different exams, and so on.

A different level is required, though, when you are trying to make very specific interpretations of somebody’s medical data or somebody’s microbiome in our case because you need to extract the right information out of each paper. You need to be able to source every analysis that you make against the right literature. You have to extract that information in the right way. Any sort of generative usage of this kind of technology that’s prone to hallucination is just not going to cut it when you are working in a healthcare context or wellness context. There are a lot of challenges that forced us to really develop new kinds of techniques that are really suited to this problem. We can’t just rely on taking what’s out there today.

[00:10:13] HC: I imagine once you developed this new technology, both on the medical literature LLM side and in processing the patient microbiome data, it comes down to validating these models. How do you approach that in your use case?

[00:10:26] LG: Well, we have human oversight of all of the extraction piece. Basically, to take this Parkinson’s example again, if we apply the LLM to go and extract information out of the study on Parkinson’s, and it says “This was a study on 500 people and they fit this demographic profile, and we found these organisms were high and these organisms were low in people who have Parkinson’s." We have a human being that actually verifies that the LLM extracted that information correctly and can say, “Yes, that is correct. That can go in the knowledge base,” before we ever start applying that information.

The AI is doing that extraction. We have human oversight to make corrections to it. But once that paper has been extracted correctly, then we don’t need to look at it again. It’s a one-time review process on every study that that’s got pulled from the literature.

[00:11:29] HC: What about on the patient recommendation side, like for a particular patient to be sure that they are getting not just the correct but maybe the best recommendations for the data you have on them? Is there a validation step there?

[00:11:42] LG: All we’re doing is linking your microbiome with the literature, right? We’re not doing our own studies. We’re saying, “Hey, your microbiome fits the profile that was described in this study for diabetes or for abdominal pain or for brain fog,” or for whatever it is, right? So we’re not validating that that study is done correctly. I mean, this is a published peer-reviewed study that is there in the literature. We’re just linking your microbiome to that study. In theory, you could read all of these studies yourself, and you’d come to the same conclusion, but we have the AI do it for you.

[00:12:21] HC: I think part of that gets to my next question because for an end user to be able to trust the recommendations you provide, they probably need to have some explanation. The links to the original studies provided, is that the method of model explainability for the end user?

[00:12:37] LG: Yes, that’s right. If we say your microbiome has been linked with fatigue or with brain fog. Then when you see that linkage, you can click into it, and it’ll show you it’s because these organisms were high and these organisms were low. It’s because of these five papers, and then you link out to PubMed reading yourself.

[00:13:00] HC: Machine learning is advancing quite rapidly right now. There are new advancements hitting the headlines more frequently than ever before. Beyond the past year’s excitement over LLMs, are there other new advancements in AI that they are particularly enthusiastic about?

[00:13:14] LG: With respect to Jona or just in general?

[00:13:16] HC: Either.

[00:13:18] LG: Well, I’m excited about everything that we’re doing at Jona. As I mentioned, we’ve had to solve some really challenging technical problems. I think this ability to absorb the literature, to interpret biological information is really novel and very challenging to do in a way that is both explainable and meaningful and grounded in true science. A lot of my focus and our team’s focus over the last year and a half or so has been on how we develop this technology. I feel really excited about what we’ve built and where that’s going because medical data is getting more and more complex. Any sort of omic that is in the news, whether it’s the microbiome or genomics or proteomics or transcriptomics, the hallmark of all of these new modalities is that they throw off extraordinary amounts of data. I think we get beyond the ability of human beings to be able to look at it and make sense of any of it. Unless we can build the kind of AI technology that will allow us to not only measure that data, but then the AI can do the interpretation by linking to a known knowledge base, I think we’re going to have trouble actually translating any of that information into practice, both for patients and for doctors.

[00:14:47] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[00:14:51] LG: Well, I think healthcare is an area that I’ve been passionate about for quite some time. I think the biggest challenges with AI and healthcare today are no longer technical, and they’re no longer regulatory. The fact is that with current AI technology and enough data, we can solve almost any AI problem that we want to now in healthcare. From a regulatory standpoint, when I started, we did not have many examples of FDA-approved AI systems.

Now, the work that I’ve been involved with at HeartFlow, at Paige. Paige was the first FDA approval for an AI system in pathology, still the only AI approval for an AI system I pathology. I feel like those areas have really been conquered in a lot of ways, and the FDA has been, I think, very forward-looking in laying out how these technologies will be regulated. But the main issue that I see in terms of translating AI into healthcare and into clinical practice is really payment. There’s no really established way to pay for these kinds of products. I think that is the number one barrier to clinical translation and adoption. It’s not just an issue of demonstrating value.

At HeartFlow, we had many clinical utility trials, multinational clinical trials, a nice recommendation. Still very, very difficult to get payment and to drive adoption of this sort of technology, and it’s unlike something like a drug or a device where people have established payment methods and processes. I think this is really the frontier in terms of AI for healthcare. I would really encourage anyone who is building a business in this area to really focus from the get-go on the business model and the perception of value and how these technologies are going to get paid for because that is going to be the number one issue that determines whether the technology gets adopted. [00:17:10] HC: Do you have any advice on how to handle that? Is it really case by case? Or are there some best practices on how you might handle the payment problem?

[00:17:20] LG: I just gave a talk on this recently to a whole variety of different AI companies and startups and academics. I think it really comes down to two things. One is the business model, right? You get a couple different business models, medical device, value-based care, selling it to pharma or life sciences. There’s cash pay. Then there’s also verticalization of the AI into a hardware, into a different kind of business. I think each one of those is challenging in its own way, and I wouldn’t say that any one of those business models has emerged as the winner.

I think the other area that people don’t talk about too much and don’t think about as much in terms of AI is the perception of value. By that, I mean if I develop a new candy bar, and I launch it out in the market, I don’t need to tell you anything about that candy bar. You can tell me that it’s going to cost under $10, right? If I build a new car and I launch that car, I don’t need to tell you anything about the car. But you know the price range is going to be somewhere in the order of, call it, 10,000 or 20,000 dollars to 80,000 dollars.

If I say I’ve built this new AI product, people don’t really know what payment range that they should expect for it, right? That’s because if I launch a candy bar, people are going to link it with other candy bars that they know on the market. If I link it with a car, you people are going to associate it with different – compare it to other cars in the market. But the perception around AI is really like an automated doctor or automated nurse or other healthcare professional. I think oftentimes because of that, people start trying to value this in terms of if I save the doctor five minutes, then that five minutes at the time is worth X. Then I can charge you half of X.

I think those sorts of ways of understanding and ascribing value to AI are really one of the key issues that has to be resolved. Furthermore, in terms of advice to somebody I would say this is how to frame the technology and the value of the technology so that people get comfortable with it and don’t think of it as like a robot doctor or robot nurse or something. I think it’s really key to getting it paid for.

[00:19:50] HC: I guess there’s a lot that still needs to happen for AI and healthcare to really catch on and for it to make business sense. There’s a lot that needs to be done with those business models and perception of value that you talked about there.

[00:20:03] LG: Yes. I think there really is because when you think of value in terms of human time, that is really problematic I think from the standpoint of trying to sell something. But the fact is that AI is technology, and people think of it as automating something that human does, but there are many different cases of AI where that’s not the case. One example is Subtle Medical, right? I don’t know if you know this company, but they use AI to speed up MRI reconstruction. This is not something that a doctor is doing. I mean, no one’s doing know 4A transforms on a patent paper or automating that. But by using AI in clever ways, you can add a lot of value. I think shifting the perception away from automating human tasks toward really the key value that gets provided by the introduction of the technology is the shift that needs to happen here.

[00:21:03] HC: Finally, where do you see the impact of Jona in three to five years?

[00:21:06] LG: Well, I think if we do our job right, then we will allow people to operationalize the current scientific literature and really start taking advantage of the microbiome, both to help them understand what’s going on in their own body and check and see if anything is looking off. But also then to be able to optimize their health, to be able to select the right diet choices, lifestyle choices, supplements, and so on to really help them improve their health.

As we think about going on a little bit longer, I think that having more data and reading more of the literature will allow us to start to translate beyond just operationalizing the literature to true clinical-grade diagnostics and new therapeutics as well.

[00:21:59] HC: This has been great, Leo. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[00:22:06] LG: You can connect with me on LinkedIn, Leo Grady. If you’re interested in Jona, our website is jona.health, www.jona.health. Yes, please come by. I love to hear from people.

[00:22:20] HC: Perfect. Thanks for joining me today. [00:22:22] LG: Well, thank you so much, Heather.

[00:22:23] HC: All right, everyone. Thanks for listening. I’m Heather Couture, and I hope you join me again next time for Impact AI.

[OUTRO]

[00:22:33] 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 and planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]