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.

[INTRODUCTION]

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

[INTERVIEW]

[0:00:34] HC: Today, I’m joined by guest, Hanadie Yousef, CEO and Co-Founder of Juvena Therapeutics, to talk about protein-based therapeutics. Hanadie, welcome to the show.

[0:00:43] HY: Thank you so much for having me, Heather.

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

[0:00:48] HY: Yeah, sure. I am a stem cell and neurobiologist by training. I like to say, with an accrued expertise in the biology of aging and mechanisms underlying tissue degeneration, accrued over a decade through my PhD work and postdoc research at UC Berkeley and Stanford collectively, where I was really focused on elucidating changes in protein signaling that occur during the aging process, either systemically in blood circulation, or locally in tissue micro environments, and showing how those changes lead to what’s known as a loss of tissue homeostasis, and the onset of numerous debilitating symptoms and phenotypes from the loss of, for example, our own body’s stem cell function and ability to get the appropriate cues to repair and regenerate tissues due to the increase in systemic, low-level inflammation, and really showed how those changes lead to debilitating degenerative diseases.

Importantly, we were able to show that we could actually target many of these mechanisms and protein signaling pathways to actually reverse the process and promote tissue rejuvenation. It was really during this time period in these formative years that I had this light bulb moment, in that we were looking at different proteins and pathways to target for promoting tissue rejuvenation, we actually began looking at proteins secreted by stem cells, which are, in many ways, the most regenerative source for tissue regeneration. Being that stem cells such as human embryonic stem cells are responsible for developing every tissue in the body and even in an entire human being.

We were able to show that we could actually isolate these stem cell-secreted proteins and use them to promote things, like enhanced neurogenesis, the survival of human cortical neurons in the face of a beta and promotion of muscle regeneration and reduction in scar tissue formation in aged animal models, or in aged human cells, or disease cells in vitro. It was many of these ideas that really led to the idea around Juvena, which is to build a robust screening platform that enables us to mine stem cell secretomes, meaning proteins secreted by stem cells, in order to match different secreted proteins with therapeutic potential to cells and receptors in which they can have regenerative, or disease-modifying effects followed by extensively validating them and then translating them into engineered protein-based therapies for specific degenerative diseases.

I had the great fortune then of meeting my co-founder, who’s a proteomics and machine learning expert, and then working to co-found Juvena in 2017, get funding in 2018 and really launch full-time operations by mid-2018.

[0:03:49] HC: Tell me more about what Juvena does and why it’s important for healthcare. There’s a lot of technical aspects of what you do. Maybe some of it you can broaden out for this audience.

[0:04:00] HY: Yeah, sure. At Juvena Therapeutics, we are really focused on discovering and developing tissue restorative biologics, which are protein-based therapeutics that really aim to restore homeostatic protein signaling in order to promote improved tissue function, ultimately for the treatment of multiple age-related and chronic diseases, but we have focused so far on really muscle wasting and metabolic diseases.

We’re doing this through an AI-enabled platform that essentially enables us to mine proteins secreted by stem cells to identify, map individual proteins that have therapeutic benefits, followed by validating them. Then once we really show that they work in both human and vitro disease models and then in animal models, we go on to protein engineer them for more drug-like properties in order to ultimately develop them as drugs for human diseases.

This is so important for healthcare because many diseases associated with aging today really have no FDA-approved therapies, or cures. As we get older, unfortunately, all of us, we start to lose our mobility, our health, our vitality. We start to get diseases related to decline in various tissues, such as heart disease, muscle wasting, and neurodegenerative disease. Juvena is part of really, a new approach to really leveraging the biology of aging and underlying mechanisms associated with why our tissues decline in function, in order to target this biology so that we can treat a broad swath of diseases, restore health, and really get to the heart of what’s causing a lot of these chronic conditions.

I think this has dramatic implications for our ability to really switch from today’s healthcare system, which I like to call more of a sick care system, where we’re really treating symptoms, not always going and treating really the heart of the cause of diseases, to a true healthcare system where we can oftentimes be preventative in nature and target not only halting the progression of disease, but actually restoring tissue function, restoring health, really reversing in many ways the aging process so that we can live those later decades of life to the fullest with our health, with our mobility, with our cognitive capabilities still intact, ultimately, to the very end is our goal.

Just for example, some people do today who live to become centenarians, right? There are rare individuals, but some people who will enter those later decades of life and really still have their health intact, and oftentimes, have the great fortune of maybe just passing away in their sleep versus due to various debilitating symptoms, or heart attacks, strokes, cancer. That’s ultimately the goal of Juvena, to really enable people to age with dignity, to continue to contribute to society, and to really maintain their health until the very end.

[0:07:16] HC: How do you use machine learning to tackle this?

[0:07:18] HY: Truthfully, I’d say, a variety of ways. At Juvena, we’ve integrated different machine learning and deep learning techniques through every cycle of our screening process, from the in silico predictions of secreted proteins that have therapeutic benefit through to the human cell model screening of these proteins for their ability to alter phenotypes and be disease modifying through to animal model testing, which we call in vivo target validation, and then all the way through to then identifying the top leads and engineering them into better versions of themselves with more drug-like properties.

For example, Juvena leverages simple machine learning classifier models to decode and deconvolute proteins that we predict are secreted, what we call non-canonically secreted proteins. We leverage deep learning models to predict phenotypic effects from chemical, or genetic perturbations of cells. We also leverage transformer models to really infer changes in protein sequences. Once we know a protein, we know its sequence, we can actually infer changes that can enhance the stability of the protein, or improve its potency, its functionality, and then we’re also leveraging now generative AI to even predict novel protein sequences that can bind particular drug targets with improved functions.

We’re also leveraging deep learning models that can, for example, help us better quantify protein expression in cells from bright-field images, or more robustly quantify behavioral changes in animal models treated with particular protein drug candidates. Ultimately, it’s leveraged at every stage of the process from in silico prediction, screening through to the actual engineering and drug development.

[0:09:21] HC: What type of data do you work with in order to do that? For those in the audience who are less familiar with some of those data types, what do they look like?

[0:09:29] HY: Great question. At Juvena Therapeutics, we leverage multiple data modalities from – one of the first steps of our platform is in actually leveraging quantitative proteomics, specifically mass spec-based differential quantitative proteomics, in order to look at and identify proteins that are enriched in stem cell secretomes, meaning proteins secreted by stem cells, compared to non-regenerative secretomes, such as the proteins secreted by either senescent cells, or fibroblasts, in order to really build a proprietary library of proteins that are enriched for their regenerative, or immunomodulatory capabilities, meaning their disease-modifying effects.

We then also leverage, in addition to quantitative proteomics, multiomics from single cell, or book RNA sequencing through to genetic data, as well as chemical data, biochemical and biophysical properties of the proteins, through to also high content imaging and robotics-enabled imaging that generates multiple images, which we leverage to better – basically, train our models for the various use cases that we’re using them for. Ultimately, it’s really a multi-modal platform with multiple data modalities that really feed into our ability to build predictive models and identify new drug candidates.

[0:11:02] HC: How do you gather this data? Is it something that you need to annotate in order to use it in your machine-learning models?

[0:11:07] HY: Yes. There always has to be, in our case, in many ways, a human in the loop, in order to assess the quality of the data. There often is annotations that are leveraged in order to build these training models.

[0:11:23] HC: Where do you get the data from? Is it collected in a lab, in your facility, or is it through collaborations, or through some other means?

[0:11:31] HY: At Juvena Therapeutics, we are a fully integrated end-to-end platform company, where we generate a lot of the data in-house from our own wet lab, where we have primary cell models that we’ve established, different preclinical mouse disease models. We will often work sometimes with seros. We will process the samples, do the experiments, but then, for example, use a proteomics mass spec facility to generate the spectral data, and then we will annotate it, analyze it, train models on it.

Oftentimes, with some of the microscopy and image generation, as well as also, we have our own sequencer. Much of that data is generated in-house. But we do take advantage of different seros that might have the instrumentation that we don’t have in-house to generate some of the raw data that we then analyze, annotate, and utilize.

[0:12:27] HC: What kinds of challenges do you encounter in working with all these different types of data?

[0:12:31] HY: I’d say, some of the challenges that we encounter really in having enough data, right? I think in drug discovery and development, one of the things, unlike for example, large language models, ChatGPT, which were trained on trillions of bits of data. When it comes to wet lab data generation, sometimes you’re really limited by just the quantity of data that you can generate. Given the limitations of training data, you might overfit a model, or not have it be as accurate as it can be. That’s what I’d say is some of the challenges, is really in just the amount of data that we can generate to really build and train models that could be then broadly applicable, that we can generalize across different types of data sets.

[0:13:20] HC: How do you validate your models, especially considering the challenges in gathering sufficient data?

[0:13:26] HY: It really depends on what the model is being used for. For example, for our models that are predicting phenotypic responses of proteins, so essentially, predicting therapeutic proteins, we will actually then recombinantly produce those proteins and test them to see if they actually are inducing that phenotypic response, so a lot of in vivo validation of the predicted proteins. We also benchmark our model’s performance by doing a head-to-head comparison of performance metrics to a base case, or the current state-of-the-art model, if available. We also train models using different architectures and training data processing to determine optimal parameters. And we perform cross-validation testing.

[0:14:11] HC: Are there any specific technological advancements that made it possible to do this now, when it wouldn’t have been feasible even a few years ago?

[0:14:18] HY: Absolutely. For us, advances in quantitative proteomics really opened the door for us to be able to mine thousands of potential secreted proteins and link them to over 70,000 known disease indications. It would take a massive amount of wet lab screening and experimental testing to really get there. Historically, there’s really been technical limitations to being able to actually map and decode the therapeutic potential of secreted proteins. Advances in quantitative proteomics have enabled us to take even a nanogram quantity of starting material, meaning the proteins that we can feasibly actually isolate and purify from cultured stem cells and be able to really deconvolute those proteins and actually identify what proteins are enriched in these different cell secretomes.

There’s also been really, really rapid and continuous advances in sequencing technology from single-cell RNA sequencing to other types of multi-omics datasets that have really enabled us much more quickly, cheaply, and robustly to generate the type of multi-omics data that is enabling us to really take this systemic approach to utilizing this data to identify different trends and markers that can be used to identify, essentially, new drug candidates and new drug targets. Those have been, and then, of course, in all sorts of robotics-enabled screening microscopy, I’d say we’re really going through a revolution, I’d say, in biotechnology, that is just enabling much more cost-effective data generation than we saw even just a few years ago. This is opening so many doors to being able to leverage all that data to basically, adopt different AI technologies that can help us shed more insights into different patterns that are coming out of the data and also different potential drug targets and drug candidates that just a human-based hypothesis-driven approach might miss, or just not really be powered enough to really get some of those insights in such a rapid fashion.

[0:16:46] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[0:16:50] HY: I would say, just be really cautious about overfitting models and trying to generalize without appropriate training data for accuracy. Sometimes it actually might be more appropriate to use more traditional biostatistical modeling, or standard techniques for your purpose. AI isn’t the solution to everything. Oftentimes, you do still want to have that human in the loop and really test the accuracy of these models just based on the limitations and data inputs and training data that goes into it.

[0:17:27] HC: Finally, where do you see the impact of Juvena in three to five years?

[0:17:31] HY: Say, in three to five years, Juvena will hopefully have its first drug through clinical validation well on its way towards commercialization, which is excitingly a muscle regenerative candidate that can act to improve not only muscle regeneration but muscle metabolism, muscle function that’s broadly applicable across age-related muscle wasting through dystrophies. We also are developing a whole pipeline of protein-based therapeutics that are coming out of our platform.

I really see in three to five years, Juvena having a validated protein drug discovery platform that’s actually enabling us to build a pipeline of tissue-restorative biologics that can really change the way we age in a way in which we’re targeting many diseases of aging that can revitalize our health and really allow us to really look forward in many ways to those later decades of life.

[0:18:30] HC: This has been great, Hanadie. I appreciate your insights today. Where can people find out more about you online?

[0:18:36] HY: Sure. You can find out more about us at our website, juvenatherapeutics.com, as well as our Juvena Therapeutics LinkedIn page.

[0:18:44] HC: Perfect. I will link to both of those in the show notes. Thanks for joining me today.

[0:18:49] HY: My pleasure, Heather. Great questions and look forward to hearing this and many other interviews that you have had.

[0:18:57] HC: All right, everyone. Thanks for listening. I’m Heather Couture, and I hope you join me again next time for Impact AI.

[END OF INTERVIEW]

[0:19:06] 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]