Hi everyone, welcome to episode 103 of Cell & Gene the podcast. I'm your host Erin Harris, and in this episode I'm joined by Dr. Alexandra Snyder, Executive Vice President of R&D at Generate Biomedicines. If you've been following the evolving role of AI in biotech, you'll definitely want to listen in to this conversation.
Alex shares how Generate Biomedicines is using generative biology to redesign the way therapies are discovered and developed by leveraging a platform that integrates machine learning, structural biology, and high-throughput experimentation to accelerate and optimize drug creation. We take a pretty deep dive into their cutting-edge work on armored CAR-T therapies for solid tumors and how they're addressing challenges in tumor microenvironments, and why their approach allows for a broader and more efficient exploration of therapeutic hypotheses.
Alex also demystifies some of the hype around AI in drug development. She offers a really clear-eyed view of what it can do and what it can't do, and how Generate is working to bridge the gap between that promise and patient impact. All right, listeners. And so with all of that background, here is my conversation with Generate Biomedicines' Dr. Alex Snyder. Enjoy.
Hi, Alex. Welcome to Cell & Gene the podcast. Thanks for your time.
It's great to be here. Nice to see you, Erin.
Thanks. So, all right, tell us about Generate Biomedicines. Talk to us about the company, its mission, and then tell us a little bit about your role as EVP of R&D.
Thanks, Erin. So, Generate is a company that is trying to do what its name says. So we are really trying to reimagine how biologics are developed in order to address the greatest scientific and clinical challenges that face the field. So as you know, the field has advanced by leaps and bounds over the last 30 to 40 years, and so we're trying to identify how AI and machine learning, and how the integration of wet and dry lab techniques, can lead to yet another leap or bound. So our ultimate goal is to, as it says in our sort of in our vocation, is to instantly generate medicines. Now, of course, what does instantly mean? How many years will it take to get there? That's a very high aspiration. So right now, we're defining what are the steps to get there. How can we generate data, wet lab data, cryo-EM data, computational data, integrate those three things together to understand how to address biology more effectively and also more swiftly.
In terms of how we realize that vision, we're starting with biology that's quite well known, well described, so we could really test the platform. And so we've used techniques to develop potential medicines against known targets like TSLP, IL-13, and SARS for our first molecules. Moving on from there, we're moving kind of into bolder spaces. So we have a CAR effort that perhaps we'll talk more about today. We have another effort in cancer. We're actually trying to bind to the small molecule payload that's released by antibody-drug conjugates to try and protect people from the adverse effects caused by that free payload in the circulation. Moving beyond there into the future, we're looking at trying to get into cells, manipulate the biology within them, trying to target multiple things on different cell types. So we're really taking a running jump towards that leap by going through biology that's well known into areas that are less well known.
Excellent. That's so fascinating. And yes, we certainly will get into the CAR component of what you're doing. And I guess perhaps now is a great time to do that. You know, as you're approaching the design and development of armored CAR-T therapies to overcome immunosuppressive tumor microenvironment and solid tumors, talk us through how that's actually happening, and talk us through what that is and what that looks like.
Yeah, absolutely. So I think one of the challenges, one amongst many in the CAR space, has been how do you integrate multiple hypotheses into what needs to be one molecule that you're actually testing in clinic. And so the way that we've applied AI and machine learning to the CAR space is to actually manipulate different aspects of the CAR independently and concurrently and at high throughput to try and get to what we hypothesize to be the best CAR preclinically to then move into clinic. So more specifically, often when CARs are being made, a certain binder or a certain transmembrane domain or a certain intracellular signaling domain will be used, and then an armor may be added to something that's already known. And that's a fine method to use. What if, however, one could adjust each aspect of those concurrently? And so basically what we've done is to develop a suite of assays that can be performed at high throughput, looking at things like CAR proliferation, tumor cell killing, exhaustion markers on the CAR, what cytokines are released, how many CARs are expressed per cell, these sorts of parameters that we know matter to how a CAR performs in clinic, and assess them all at once. So by doing this at high throughput, we can say, okay, let's use one binder but change the transmembrane domain, let's adjust the binder and fix the transmembrane domain, let's adjust both at once. So we can actually test — those are examples, of course, you can adjust other things. By doing that and by measuring all the parameters I just mentioned together, you can really explore a much broader space of hypotheses and figure out which changes are actually making a difference to CAR function and performance and which are less important.
Now, to the question about armoring that you asked, I think the key challenge there is there are limits to what we can really do in vitro, and so I think we have pushed those limits as far as we can, but we also acknowledge that the ultimate test is how these things perform in humans. And so to that end, in addition to using our own platform and our own testing, we actually do a lot of what we call reverse translation: look at what's going on out there in the field and incorporate that into our own modeling. And so when, for example, you may have seen Carl June and his group's paper come out about the IL-18 armored CAR that seemed to be effective in the post-CAR failure setting, that is really informative in a way that no experiment done in a dish or a mouse could really be. So I guess what I'm getting at is we're very excited about what the platform can do to get us to what we think is the most potent, most potentially effective CAR. We also acknowledge that the ultimate test is what happens in patients.
Yeah, sure. That makes sense. And perhaps then talk us through the unique advantages associated with this. So, you know, what is generative biology offering in engineering these armored CAR cells compared to, say, traditional design approaches?
So I think it's this ability to measure all the parameters that you want and adjust them at the same time that's unique. So as you and listeners may be familiar with, typically these variables are changed in succession rather than concurrently. And so you kind of have to choose: do I want more expression, more CAR expression? Do I want to emphasize what cytokines are released? What this technology allows you to do is to say 'and' instead of 'or,' and to say 'altogether' instead of 'in succession.' So I want a high expression of my CARs per T cell, and I want high interferon-gamma, and I want high proliferation, and I don't want high exhaustion markers. So it's the ability to do this. Now, one aspect is doing it experimentally, right? The other aspect is how to integrate the data. So, you know, our conventional abilities to analyze and interpret such high-dimensionality data are somewhat limited. And when you use AI and ML to analyze the data, you can say, okay, here's where we were right, here's where we were wrong. What should we fix to address where we were wrong? So I think it's this ability to both generate, integrate, and then act on the data that is unique.
