Cory Smith: Welcome to Galaxy Balance, a podcast exploring the intersection of cutting edge science, technology, and humanity's long-term future. Each episode features visionary thinkers working on transformative technologies, including synthetic biology, AI-driven science, longevity, and genome engineering. Today's guest on Galaxy Balance is Pranam Chatterjee, a bioengineer, computational biologist, and founder at the intersection of AI and synthetic biology. Pranam is an assistant professor at the University of Pennsylvania, where he leads the Programmable Biology Group, developing generative AI models to design novel peptides and protein biologics from sequencing to sequence data alone, opening new frontiers in therapeutic discovery and synthetic biology. Beyond academia, he's an entrepreneur and co-founded companies, translating computational bio design into real world applications, including Gameto. a venture exploring transformative approaches in reproduction and cellular technology. Pranam's work lies at the heart of AI-driven protein design, programmable biology, multi-objective generative modeling, and making him a perfect voice for a deep conversation about the future of life engineering technologies. Pranam, welcome to Galaxy Balance. Pranam Chatterjee: Thank you so much, Corey. Thanks for having me. Cory Smith: Excellent. So I love the whiteboard in the back. I love these brainstorming ideas and great to be featured in your office today. So I'd like to start with a bit about your background. What was your calling? What got you into science and where you are now? Pranam Chatterjee: Yes. thing. Yeah. Yeah, it's a very funny story, Cory. I don't think I actually started in science. My parents are both scientists, so I was always around it. But my first interest was actually in religion and philosophy. ⁓ I actually started as a religion major at Dartmouth College when I started my undergrad. And that stemmed from years of studying. Cory Smith: Yeah. Pranam Chatterjee: about different religions and different faiths. at a certain point, I studied it so much I even spent summers at Buddhist monasteries, went to Christian churches, went to Hindu temples just to learn. But at a certain point, I realized that the scientific method was also a really cool way of understanding the world too. And so that got me really intrigued. I went from Dartmouth, I transferred to MIT, and I spent all the rest of my time studying just doing computer science, molecular biology, and really pushing forward those fields, and then obviously staying on that. front and that train when I went to my PhD and my postdoc. yeah, it's been a very whirlwind journey. through that, I learned a lot about what types of problems I like to solve and how I like to solve them. That's what's brought me to what you just mentioned, the style of science we do, which is an integration of AI computation, but to design molecules that can do cool things. Cory Smith: I love the background and starting with religion. Not many scientists start with that and often people think of them as opposites, but in many ways I think they share a lot of features and the study of the awe, the cosmos, the beauty of life around us and a lot of previous scientists I think were religious, but more recently it's become much more of a secular field. How do you see the two different fields and potential overlap? Pranam Chatterjee: Yeah, exactly. Yeah. Thanks, everyone. Yeah, you know, at the end of the day, the science we do and the tools we develop are going to be hopefully being useful to people, right? And a lot of people believe very different things than we do as scientists. They have their own faiths, their own traditions. And for me, it's just important to learn them, to understand why people believe what they do, and such that when we do bring technology to them, we're able to have it be useful and be communicative to them. So I think that religion and philosophy just gives us a connect. to other people's faiths and traditions. We're the people bringing technologies and solutions to them as well. Cory Smith: Great. And do you have any key points in your mind of turning points in your career? Like what led you to the church lab or what really enticed you into this? Pranam Chatterjee: Yeah. Yeah, so you know, I always wanted to study, I always wanted to have applications in biology simply because, in biotechnology, simply because I felt that that was the area where you could solve problems for different key world problems. For example, medicine of course, but you can imagine agriculture, ⁓ know, ⁓ like bioenergy and food security, all of these can be solved with a biological solution, right? You can do gene editing. Cory Smith: Hmm. Pranam Chatterjee: of plants, can do, you know, can engineer microbes to produce more biofuels and bioethanol. And so for me, I was like, okay, I need to, I want to eventually engineer biology, but I need a good tool to do so, right? And so, you know, my first guess was maybe it was going to be material science or chemical engineering. But then I realized that computer science was probably going to be the future technology that we use to engineer biology. And so, studying pretty theoretical computer science as an undergrad, I said okay for grad school and postdoc, which is where I worked with George, I was like this is where, this is the time where we're gonna apply new algorithms to key problems. In George's lab it was reproductive biology, in my PhD lab at MIT and Joe Jacobson's lab it happened to be CRISPR. So yeah, the real turning point was I went from being a religion major, studied computer science, and then said, now is a time where I know enough about computer science that I can utilize it in a useful way. My PhD in postdoc was where that happened. Cory Smith: Great, and let's dig into the post-doc a little bit more. You mentioned studying reproductive technology. We've actually had Christian Cramion previously. So he's dug really deep into gameto, but perhaps you could describe a little bit more at the founding of gameto and that original idea behind it. Pranam Chatterjee: Not sure. Yeah, ⁓ nice. Yes. Yeah, it's an interesting story, right? So I was connected to George via my PhD advisor, Joe Jacobson, who's the inventor of the Kindle, which is such a weird story. He invented the Kindle. I went to Joe's lab. I was like, I'm going to edit. I want to do CRISPR because it seems like the most impactful technology. Joe was like, well, you're a computer scientist and you don't have any, ⁓ we don't have a wet lab. So we built up a wet lab. We did all of that work. And by the end of that time, I finishing up, George became one of my thesis advisors or thesis committee members and he asked me, hey Pranam, why don't you come do a postdoc in my lab? I would be really excited to have you and I was really excited to join his lab. But he was like, what do you want to work on? The same question that Joe asked me when I was doing my PhD. And I was like, you know, I think in stem cell design, stem cell engineering, one of the hardest questions is how do you make germ cells? We are able to make cardiomyocytes, neurons, pancreatic beta cells so easily, but how do you make the hardest cell type? And George was like, well, I have a great team at the lab that is also interested in similar questions. And Christian was a pretty new PhD student at that time who was really excited about that. So he connected us. But George also said, we don't really have funding for this project right now, but if you can figure out a way to help us raise money for it, we can definitely make it happen. And so for me, I was like, well, I didn't have a postdoc fellowship. I wasn't senior enough to get those fellowships. And so I was connected to an entrepreneur via George and others ⁓ named Martin Varsavsky. He's the founder of the Prelude Fertility Clinics and it's one of the largest fertility clinics in the world, set of clinics in the world. And he brought me down to Philadelphia and said, what do you have for me? And I was like, you know, I've developed this algorithm that can identify proteins that if we put them into stem cells, we can maybe convert them into oocytes. All I had was this silly little algorithm I had written to identify these proteins. And he was like, great, let's start a company around it and you can run it. And I was like, no, no, no, no. I don't want to start a company. I just want to be a postdoc in George's lab. And he was like, OK, that's fine. Why don't we put a little early money into your research so you can do the science. Cory Smith: Yeah Pranam Chatterjee: He donated money to the lab. used that early money to Christian and I to do some early work to prototype the idea of making us oocyte from a stem cell possible. And then he came back to me and Corey and said, well, okay, it seems like you're making this work. Let's now go and start a company. And I was like, no, I don't want to start a company. But we realized we needed money, right? And this is a, a trivial project to get it going. So what we ended up doing was we started Gamito. Cory Smith: Yeah. ⁓ Pranam Chatterjee: as a shell company just to raise money from venture capitalists to put, we got about $3 million as a seed round. Instead of starting the company, we actually put it as an SRA back into George's lab. Christian and I, I will give a lot of credit to Christian, really pushed hard on it. I was more on the computational side. We pushed and got. a lot of prototypes of the technology to work. We brought Merrick Smela, was a PhD student, Christian may have told you about. He helped us to develop other cell types in the ovary, like granulosa cells, theca cells. We worked together on this. And then we were like, well, now we have the technology. Gamito put the money in. We licensed the patents back to Gamito. And then Gamito, me and Dina Redenkovich, who is the CEO of Gamito, went out to look Altos and in California and we raised a 23 million dollar Series A and then from there the company really then took all the technology out. So it really showed you that a good partnership with George's lab and academic spin-out early allowed us to create now a technology that Christian and Dina have taken into phase three trials. We have many babies born from the technology. Yeah, and you know, me and Martine founded this company with just kind of a goal for me to Cory Smith: Yeah. Pranam Chatterjee: fund my postdoc in Georgia's lab. Cory Smith: Amazing story and just great work that they're doing. The first IPSC-derived phase three trial, huge, monumental for the field. I think great opportunity to become the first in the U.S. commercialized IPSC therapy. That'll be great for mainstreaming this. So that's a future that I see for it. And great story, because a lot of people, you then later went back to academia. So we'll dig back into that. But just wanted to tell you, we did get Merrick on the show as well. His was just released. So we talked about his Ovell company and ⁓ it seems that Pranam Chatterjee: Fingers crossed, Corey, fingers crossed. Yeah. Amazing. Cory Smith: of that goal of making the germ cells the support cells were actually what you got out of it first and had a much more clear regulatory path to get into the clinic. Pranam Chatterjee: Yeah. Yeah, and I think Merrick and Christian get so much credit for pushing the technology in so many directions. know, I came in with this algorithm that helped us to identify these TFs, but in reality, like, the fact that that strategy of utilizing IPSC-derived cells can work for so many applications, from OVAL, what Merrick's, the beautiful work Merrick's doing, to what Gamito's doing, shows you that we have a lot more to explore. And Gamito is really going forward and exploring many different directions for where these cells can be used, not just as IVF support cells, but also potentially as in vivo ⁓ tissue engineering applications. Cory Smith: Yeah, Christian shared the ⁓ ARPA-H application for hormone regulation for postmenopausal women. Incredible application and it's always great business decision when the same cells, same seed bank even, can be used for multiple clinical applications. Pranam Chatterjee: Yeah. Yeah, you know, I was telling, I told Dina when I started, like, ⁓ yeah, we're going to make these IPSC-derived oocytes. And Dina was like, you know, I think the first good, she's, mean, she's the person who's really made the company so successful, especially from a business side. ⁓ And she really told us that you have all this great technology, let's take a piece of it that has a very strong chance of application success and translation. And that's why Gamito was able to move so fast, is because we did a lot of that early work at the Harvard Lab, and then that translated so well in the business side of things. Cory Smith: Yeah, it's really such a great incubator and opportunity within George's lab. ⁓ yes, and I'll pose this to you. One person who saw my podcast said that I wouldn't know what to do. He's a fellow scientist. He says with de-extinction, with human reproduction, with basic technologies for reading and writing DNA, I would just spend the first six months looking around at what to do. How did you decide what to do out of so many opportunities? Pranam Chatterjee: Classic George story, Corey, as you can understand. Yes. Yeah, think that was the, you know, I'm a very, I've never been in a very traditional mentoring environment. I was in a immunology lab as an undergrad for a little bit, learned kind of classical good scientific practice. But remember in my PhD, my PI was the inventor of Yank, right? The Kindle technology, and he just said, what do you want to work on? And so when you're given that... open slate, you decide, you get to decide what you want to do and you get, and you realize, why don't I just work on the problem that is the coolest to me, right? Rather, regardless of how feasible it is, let me choose a problem that I just really think is cool and exciting. For PhD was CRISPR, like making a really ⁓ improved gene editing tool. For my postdoc, was like, what is the hardest biological question that could have a unique impact? ⁓ And for me, it was, you know, IPSC derived. oocytes was really exciting. So that open-ended science and then figuring out the technology is a fun way of approaching and choosing problems. I just happened to get fortunate with the problems I chose. Cory Smith: So if you brought that into your new lab, do you mentor that mindset of really independence and open ability to pick your own problem for your students? Pranam Chatterjee: Yeah, well you know it's funny because I am, I personally am very creative in the sense of the problems I want to solve. I have so many, like I want to solve all the big world problems. Now I'm not going to do that, but the- main thing is I want to show my students, hey, if you keep your mind open and you say, I'm going to be the one that solves climate change or cures cancer or whatever big problem you can think about, you're very capable of doing so. So the answer is yes. But we make sure that the tools we use to solve them, Corey, are the ones that my students are good at and what I am good at. So while you want to solve a problem, you still have to be self-aware and know that these are the things I'm good at technically. And can I utilize my algorithm work, my protein engineering work, whatever, to solve it? So it's like kind of like starting big and then honing a path to the solution that you're able to traverse. And that's what I try to teach my students. Cory Smith: Excellent. And so there must have been a moment where Gamito was really launching off and they wanted you to just go into that full time. How would that decision made that you would pursue the academic route? Pranam Chatterjee: Yeah. I think it was very clear and I told the team very early that I was, you know, I actually by the end, by the time Gamita was raising, I'd already gotten my faculty job at Duke. And so I was like, this is what I want to do. I was very clear to Martin as well when we started that. And so I think it was very clear that I wasn't going to be the SCSO or director of science. And so it was very Cory Smith: Hmm. Pranam Chatterjee: exciting that Christian really wanted to take that role and he was very interested in leading a startup and being in industry. So it made sense. I was like, well Christian, you're the guy. You actually did so much of the exciting work at our lab, at George's lab, so you're the perfect person to take over. I'm gonna excuse myself. I'll say stay on the scientific advisory board, but I'm gonna go and do what I, you the next set of problems that I'm. Cory Smith: Yeah. Pranam Chatterjee: personally wanting to ⁓ solve, which happened to be coinciding with this exciting new revolution of AI in biology. Cory Smith: Excellent. And on that point of iPSC-derived cells, the algorithm you made to pick proteins, do you still work on that in your new lab? Pranam Chatterjee: Yeah, so we've come back to it, Corey. So the first set of work when I came to my lab at Duke and then we moved the lab to Penn was to design peptides and molecules that can hit undruggable targets. We had really cool methods to design these molecules. We're able to hit disordered proteins, proteins causing cancer, all of these targets. But now, guess what we've come back to? Now the big problem is You may have heard in the AI biospace is this idea of a virtual cell. ⁓ It's the term that's being thrown around. But what is that virtual cell? It is essentially what we did with Gameto, which was... generate, create perturbations, in our case, we're transcription factors, that you would put into cells and you would convert them into a new cell state, which in our case was iPSCs to ovarian cells, right? So in reality, Gameta was the first virtual cell company. The only difference now is we're designing, using AI molecules that will induce the cell states that we want. And whether that is ovarian cells, oocytes, cardiomyocytes, whatever, these are the problems that now my lab is very excited to work on. Cory Smith: And so any cell types in particular, or you're still working on a virtual cell map of essentially all cell types. Pranam Chatterjee: I think we're not working on all cells. I still am very excited about ovarian cells. I simply think that that is an understudied area. Gameto is really pushing on it. But clearly, there's a lot of heterogeneity there, too. And from just not just the support cells and the oocytes, but other cells, like stromal cells, ⁓ hormone-producing cells, I think we have a huge opportunity there to create tools and therapeutics that affect half the population, right, the female population, and help them in their reproductive journeys. yes, the cell types that we're trying to create now ⁓ utilizing molecules that our AI models are designing are primarily ovarian cells, endometrial cells, know, cells of... pretty much the cells of the reproductive system. However, we have collaborations for neurodegenerative disease as well as cancer as well. Can you convert a cancer cell back to a non-cancer cell? So it doesn't just have to be a ⁓ forward stem cell to something. It can be cancer cell to non-cancer cell, disease cell back to healthy cell, plant cell that doesn't fix nitrogen to a plant cell that fixes nitrogen. The idea is to be able to convert cell states that you want. This is the dream of the field, but we have the molecules and the experience with gameto on doing it successfully. Cory Smith: And you mentioned molecules, so Gameto primarily did it with transcription factors. Are you using small molecules or completely different proteins or least evolved transcription factors? Or what's the strategy you can share? Pranam Chatterjee: Correct. So we as you. As you remember, for the last few years, my lab has become the AI for peptide lab. ⁓ Why peptides? You know, the funny story goes is that while we were doing Gameto, we were developing this technology that utilized peptides to attach to enzymes such that the peptides bound to a protein and the enzymes modified it. So this is the second company I started called Ubiquitix, ⁓ which started about the same time as Gameto, where we were basically designing molecules to edit. protein. you know CRISPR is ⁓ CRISPR is like a guide RNA and a Cas9 binds, edits the DNA. We were developing peptides attached to like a kinase or a ubiquitin ligase, bind, edit the protein. So we were trying to make a CRISPR, we've been making a CRISPR system for proteins and peptides just happen to be this short guide RNA like very clonable molecules. And so I was like, well, I'm just going to get really good at making peptides. And what that's evolved into is this ability to make peptides that can target various proteins, but also peptides that can convert. cell states, T-peptides can bind to a lot of proteins, they can modify them, they can block them, and that can evolve how a cell state looks from time point one to time point two. So yes, I think the answer is our molecule of choice is peptides. They're small, they're deliverable, GLP-1s clearly have worked very well in the clinic, so let's make a more drug-like molecule. Transcription factors are a little big, and that's why they work so well in cell engineering contexts, like amidos context because you can do this ex vivo and then go into the cell or use them as IVF support cells but peptides are much more drug-like and they give us a lot more opportunity for targeting. Cory Smith: Excellent. And if you could get a little bit more into those tools that you're using to design these peptides. I know you've created tools like Peptune and PEPMLM. Could you describe some of the basic foundational tools you've developed and their applications? Pranam Chatterjee: Yeah. Yeah, it's great. I will say I'm a slightly controversial figure in the protein design, maybe molecule design space. Our thesis has been, do we need the structure of proteins to be able to design them? That's been the Nobel Prize was won via for AlphaFold, for RF Diffusion, for all these great tools at the Baker Lab and DeepMind, and everyone else has been developing. The thesis I always made was, Do we really need the structure of a protein to be able to design a good molecule? PEP MLM was one of those models that really tested that hypothesis. What if we just gave the model the sequence of the target protein, as well as in PEPTUNE, the functional constraints we want for that molecule? I want it to bind. I want it to not be toxic. I don't want it to be immunogenic, et cetera, et cetera, et cetera. And then what we realized, Corey, was that the problem was that If you look at the structure of a protein, you're really only looking at one state of that protein. And usually, where do we do structural biology? We do it in cryo-EM, x-ray, crystallography. We look at it in a very static state. And so because we started to train language models, like PEP MLM, on protein-peptide interactions, we realized that all the information that a language model has about the protein and peptide incorporate structural constraints functionally. all that. We don't need to look at the structure to make a geometric fit to the molecule. PapMLM, we made peptides that were able to Huntington, viral phosphoproteins, etc., all via a simple language modeling task, which was here's a target, here's the peptide, hide the peptide sequence, regenerate it. Basically, your classic language modeling task that chat, GPT, Claude, Gemini would utilize, we just applied it for peptide design. And as I told you, the molecules worked. But we realized, too, like I said, it's not just about binding. It's also about getting the therapeutic properties in the molecule that will allow it to succeed in a clinical trial. And because we just use sequence, we are able to incorporate those functional constraints. So, peptune, MOGDFM, are you ready? These are models that are coming out of our lab that say, I want to design a perfect peptide that binds to, ⁓ you know, Huntington protein, but also is permeable, non-toxic, non-immunogenic, you know, has good therapeutic properties, generates. These are models that are coming out of our lab that say, I want to design a perfect peptide that binds to, you know, Huntington protein, but also is permeable, non-toxic, non-immunogenic, you know, has good therapeutic properties, generates. And that's what we get. We get a cyclic short GLP-1-like, if you want to say it, peptide that we can dose vivo get therapeutic effect that we want. This is the future of therapeutic design. And that's what we get. We get a cyclic short GLP-1-like, if you want to say it, peptide that we can dose in vivo and get the therapeutic effect that we want. This is the future of therapeutic design. It's not about binding. It's not about virtual cells necessarily. What it really is about is getting a good molecule and making sure it has a good chance of passing phase one, two, three clinical trials. Cory Smith: Excellent description, Pranam. And what I'm thinking of is vibe coding has clearly changed software design. And what I'm thinking of how this applies to vibe coding biology. Essentially, you've described at these higher levels. We want it non-immunogenic, GLP-1-like delivery, binds this protein. How in the future can you simply say that and then automate the rest? Clearly it requires very specialized labs now. Pranam Chatterjee: Yes. Yeah. Absolutely. You're bringing a really good point here, too. I think the question I've started to think about is where does natural language interface with this too, right? Because live coding is where you add natural language, but then you have ⁓ code that comes out of it. Biology is the same way. What if I could just describe this is what I want and then I generate? What we've learned is that not yet, I think it's going to happen, where natural language will be the best way of incorporating these priors and these constraints into design. The models that we're developing right now are more domain specific. We're making models that literally the generative model makes a SMILES string of acyclic peptide. And we have predictors. We call this the peptiverse. We have predictors that predict toxicity, solubility, permeability, binding. They guide the generative model to sample peptides that have those. properties and are Pareto optimal across them. So we had a recent study just a few weeks ago where we saw what if I instead of utilizing these labeled predictors of therapeutic properties, I instead use quen or llama open source. Of course, we can eventually use you know, claw gemini GBT. And we said, what if I instead of utilizing hard predictors to guide my generator, I utilized language models to guide it. And what we found was if you had good label data of properties and you trained a good predictor of those and you use that predictor to generate molecules, the molecules are better. But if I use language and I did not have, say I don't even have a predictor for protease stability, which is why GLP-1s took 20 years to get to the clinic, I don't have a predictor of protease stability because there's not enough data on that. I'm just going to Quen to... Cory Smith: Yeah. Pranam Chatterjee: say, hey, I want this peptide to be protease stable. And what we found was that can help, but not enough yet that would make a meaningful effect of how good the molecule would be. So one day we'll get there. But I think right now we're still at a point where domain-specific generative models, sequence models, Cory Smith: Mm-hmm. Pranam Chatterjee: guided by function predictors really make molecules that are ideal and we have animals in our lab. We're translating these theoretical formulations into in vivo and early results from our lab seem to show that this works. Cory Smith: Excellent. And taking a step back for George's picture of big picture long term, I'm imagining a world where even this is for, you said virtual cell, but virtual organisms. Like the work that they're doing at Colossal of bring back the mammoth that you have sequins, but another, like let's say, have you read How to Make a Dragon? Barck Noffler. Yeah, excellent idea of like. Pranam Chatterjee: Yeah. Yeah. Yes, of course. Yeah, that's a great book for a symbioguy, right? Cory Smith: Yeah, yeah, mean, great application there for the audience. He basically describes how we could use stem cell biology and genome engineering to bring about a real dragon and what are the difficulties, what are the challenges, and how one day this might be possible. And I'm imagining another vibe-coded organism level of create a unicorn, create a dragon, and I could lay out steps and potentially execute it for you. Pranam Chatterjee: Yeah. Yeah. And I agree, I think those are very possible because of the tools. Remember, it's not just about the software, it's about the hardware too. To do programmable biology or in this case, vibe-coded biology, I love that term. You should definitely, like, this should be the term we are thinking about in the future, of vibe-coding biology. I think what we need are both advancements in the AI, which is fine. Of course, we're, know, labs like mine and obviously the big labs like Anthropic, DeepMind are doing this. We need those advances, but we're never going to really get to good biological applications if the hardware does not advance at the same time. And we are seeing, obviously, with... technologies like gamete technology and all the stuff that came out of George's lab on stem cell engineering, gene editing of course, right, also out of George's lab and others. My whole PhD was on CRISPR Cas9 engineering. And now I think the third pillar, Corey, is simplicity on the hardware. I think stem cell engineering, CRISPR. Cory Smith: Yeah. Pranam Chatterjee: CAR-T, mRNA, all of this are great tools. They help us to do such sophisticated biology. But my hope is that we can also simplify the actual actuators that do the manipulation. This is why I love peptides, small molecules, these, like, dosing a small molecule into a system or into a patient or into an animal, it's so easy, right? It just goes in, it goes, does its thing, and it's gone. There's no, like, big, thing that has to actuate and you have to deliver a very specific way. So my hope is that... we can both push simplicity on the output side, right? The last molecule. But the complexity comes from the design, the AI, the, you know, all the pieces that we're currently developing on the tech side. So I feel like that's the future. Simple outputs, complex design parameters. And create a molecule that I can dose and be like, here's a unicorn. I just put in a small molecule and made that happen. That I think would be the dream, but you know, we're gonna have to really work hard on the hardware first. Cory Smith: Yeah, you hit on a good point, Pranam, that the hardware for current experimental workflows is very complex. Talk about what a human can do, very complex workflows, but talk about programming it into a liquid handler, it can do very simple things, but then to change the experimental design or improve it often requires a total redesign of the hardware, or at least how it's configured. So I love this idea of simplifying the execution, like complex design, complex inputs, but with a standardized format. Pranam Chatterjee: Yeah. Yeah. Yeah. Cory Smith: So Ari, is your lab working on any automation? Pranam Chatterjee: Yes. Yeah, so we are not an automation lab. However, we are starting to see the benefits. We have a peptide synthesizer that is fully automated, does its thing. It's just that simplification has really revolutionized our lab's workflows. Now the model can design a peptide, whatever it needs to do, know, bind, do all these things. Cory Smith: Hmm Pranam Chatterjee: press a button, essentially. There's a little bit of setup, but the whole robotic system handles the design. And you get this molecule, and just that is by itself has been powerful. But there are lot of great labs that I work with, collaborate with, like Emma Corey's lab. And she's a Kevin Esfalt student who's, so she's like a second generation George student, if you can imagine. And then obviously Mo Khalil and others are really developing robotics-based systems for biology. Cory Smith: Hmm. Yes. Yep. Pranam Chatterjee: and then now you're starting to see, I can make the perfect molecule, right, ideally. And they have the tools to really trans-, to make it, do the functional experiments, etc. So even if it's not a fully vibe-coded biology, as you said, those innovations simultaneously will merge such that now... we're going to really have a unified workflow that creates the molecules and the tools that we need to program biology. So I see it happening very soon. However, I do think we're still needing to develop the hardware. We're not perfect on the algorithms. We're not perfect on the robotic systems that make complex experiments easy. But we've got to get there. Cory Smith: Yeah, we live in a very exciting time where these opportunities, these exponential, increasing technologies are converging, as you say. And what advice do you give to your students to keep relevant, to keep up, and to prepare yourself for where the world is gonna be? Pranam Chatterjee: I love this question. You're asking the most important questions and the one that keeps me up at night. At ⁓ the end of the day, I'm a professor, a faculty member, advisor. My job, as much as it is to develop good science and do good research, is to make sure that my students are successful. I want to first say I'm very lucky to have the best students, I think, that I could possibly have. My students have won the Rhodes Scholar, one of my students, Sophia, very well known now in the AI world. She's an undergrad, considered the prodigy of our field, I would say. She's so smart and she's won a lot of awards. And so for me, I think about their futures and be like, where do you set yourself apart such that people will still want to give you the opportunity to solve problems that maybe... other problems have been automated or solved away. And it really comes down to the two main things. One, on the AI side, I think the folks that are going to be very good at pushing new AI theory are going to be still in demand. Because as all of the automation of coding, training models, I mean, I'm sure if you don't know it already, I could teach you how to train a PyTorch ML model in like 10 minutes and you'd be like training your own ML model. It's just, it's not as difficult as it used to be, but it's really the folks who are very good at the math that I think are still having a meaningful impact in where AI is moving. So for my AI, I have two sides to my lab. Half my lab are pure AI scientists, the other half are pure experimentalists. For the AI scientists, I really push them to be good at core algorithm math development, like things where you have to be creative. not execution, but more just like, have a great idea, I'm going to show that this works. That's the place I want them to innovate, and I think that'll keep them relevant for a long period of time. On the experimental side, I think it's exactly where you just mentioned, which was interface with robotics, interface with AI tools. My experimentalists, they're not computer scientists in any sense of the word. They probably can write very simple programs, but they speak with AI folks. every day because half my lab are AI folks. They're translating the AI outputs. They're utilizing robotic systems. They're playing, they're creating a workflow that is modern. And I think that is, these two skill sets and these two types of scientists are going to really survive this. this big transformation that we're seeing. And so tried it. I'm lucky that I'm trained in both sides of these things, and I'm able to teach them. But most of my students either operate in one or the other. Cory Smith: Yeah. Excellent. And you had a good record of translational and therapeutic impact. What are some new diseases or targets that you are excited about applying? You mentioned Huntington binders, but just what's top of your mind on that field? Pranam Chatterjee: Yeah, of course. mean, know, just quick story on the Huntington stuff. When I was a PhD student, we didn't really have a lot of NIH or NSF grant funding. the first person, that first organization that gave me some money to do our CRISPR gene editing work was the Huntington Disease Foundation, CHDI. And so we are really excited about keeping translating our molecules for that disease. But I will say our disease focuses, as an academic lab, focuses on mostly rare diseases, right? These are the diseases that a venture capitalist is not going to find ⁓ very easily because there's no market for a three person rare disease. We work on Alexander disease, is a ⁓ story that we've, a disease where we've been focused and featured a lot of because we're so passionate about, you know, coming up with a solution for this rare neurodegenerative disease, but also rare cancers like pediatric sarcomas ⁓ as well as viral diseases as well. think it's these diseases where it makes sense to translate the molecules you design academically, where you can run a trial at Penn Medicine or at CHOP here in Philadelphia rather than spinning out of company and translating it. However, companies are very powerful too, where for diseases that are much larger, like for example, ⁓ liver diseases that we work on at Ubiquitix, larger cancers that we work on with my new company, Recognition Bio and others, that's great for the companies to translate. But academically, we're really focused on rare neurodegenerative, rare pediatric cancers, bioremediation applications too, right? These are maybe not the sexiest things to work on. from a company perspective, but are really exciting to work on academically where the pressure to find the market fit is smaller. Cory Smith: And the new administration has a focus on rare diseases. So has that influenced your target or companies that can come up with a perhaps more streamlined path? Because I know from the venture side, the payoff side, rare diseases aren't worth it from a capitalist pure point of view. They might have certain niche applications, but like you mentioned, GLP-1, they're the real winner where you have essentially the whole world, most of the world. Pranam Chatterjee: Exactly. We have to think really carefully there. The good thing is that the amount of money you need to do with AI, the amount of money we need to translate and make a new molecule and have it work is much lower than it used to be because the molecules coming out of our models are pretty good. It ⁓ doesn't require ⁓ a large pull to make that happen. So I think that when we choose diseases, we... want to have diseases where we have a close connection to ⁓ the patient families, to the biologists that have been studying that disease for a long time, and really figure out what is the right way to solve it, and are there regulatory pathways that make it easy for us to take this quickly into a patient and see if it works. And if that happens, like for example for Alexander disease, or for ⁓ pediatric sarcoma, like Ewing sarcoma, Then we get this feedback much, much quicker, ⁓ and we're able to develop the molecules better. So yes, we choose a lot of the diseases where we have a technology fit, a regulatory fit, and an academic fit. And if it all comes together, we have an area that we can work on. Cory Smith: Excellent. I love your experience, Pranam, on drug development and regulatory and how AI is going to interface. I know there was a town hall last year at the FDA, and there were lots of open discussions about a fear that China can get into the clinic so much faster. And there, of course, were questions of what safety might they be compromising. So what are your thoughts on how we can be more agile at the US while maintaining ⁓ competitive safety and ethical standards? Pranam Chatterjee: Yeah. Thanks Yeah, it's tough. mean, I think that first of all, I really think China is doing an amazing job at taking good risks and hopefully pushing their technology forward. We have to do the same. You know, at end of the day, we are both competing on the technology side, but also on a talent side as well. ⁓ And, you know, we really need to make our science a lot more collaborative and open to hopefully scientists from around the world. That's number one. But I think on number two, the U.S. has to realize that we do have a lot of the best technologies being developed here. And we need to have paths to try them out as fast as possible. Sorry for the sound in the background. And so we, so as a result, we have to make sure that the scientists at the early stages and are connected to doctors and are connected to regulatory officials saying, don't, if you're gonna work on this disease, make sure that what you're thinking about has a straightforward path to ⁓ approvals and has a straightforward path to conducting the early safety trials. If we don't do that, then what's gonna happen is we're gonna spend all this time and effort, creating a technology, and then once we get to the FDA, they're like, yeah, this is not going to fly. This is back to my point of make things simple, right? Like, create technologies that are easy to translate, that you will not face regulatory hurdles. Like, if we can make a small molecule and a peptide, do that. Don't go and do a crazy weird ⁓ cell and gene editing therapy if you can avoid it. If we need... Cory Smith: Cow. Pranam Chatterjee: gene editing and cell therapy as the only way to solve it. For example, some genetic diseases, you just need to insert a gene. Like, you've to insert the gene. Then take that step. So I think it has to be smartness on our end on picking modalities that make sense for the disease and are easy to translate, but also a willingness on the regulatory side to say, hey, some diseases, some ⁓ indications just need complex therapeutics. And we have to make sure that those have straightforward, streamlined approval processes that help us to translate. If not, we're just going to fall behind as a country. Cory Smith: I love it, Bernan. Back to your simplicity. Simplicity when available and complexity when needed and no other alternative. Pranam Chatterjee: Yes. Exactly. Yeah. Otherwise we're just really going to make it hard on ourselves. Right? Like, I see folks who are developing these complex, like, know, they're going to deliver five proteins ⁓ using a genetic circuit and inducible upon...amazing. I love the technology. It's so cool. But every piece in that circuit is another piece. Cory Smith: Yeah. Pranam Chatterjee: that the FDA or regulatory bodies or patients are going to have to tolerate. And if I could do the same thing, but I could give dose of a small molecule that has great bio-distribution and high specificity, or a peptide that just does all of that, which one would you take? I think everyone would be, take this. So it's just that type of mentality. Cory Smith: Yeah. And it's great to put those circuits in biology, but everything mutates. So if it's many cells that require it, there's going to be some attrition from mutation. Pranam Chatterjee: Yeah. But like I said too, we're getting so good at the design as well. Who knew a few years ago we were going to be able to just out of the box design a peptide that does these 10 things? Or a small molecule that is able to bind, have low tox, non-immune. We're able to insert functionality into our designs. This is what AI is giving us the opportunity to do. So maybe we do not need to settle for a semi-okay small molecule because it's easy. We can actually get what we get with a circuit, get what we get with ⁓ complex biologic. and put it into a more deliverable simple therapeutic, all the complexity just happened in development. That's all. Cory Smith: Hmm, and AI is only as good as the training sets or the label sets as you've described. What efforts do you think are needed to fully collect that data set? There's so much diversity of life on earth and proteins, the oceans largely untapped and unsequenced. Where do you think are the next frontiers to gather that training set to improve this? Pranam Chatterjee: You know, I used to say that, I love this question because I used to say that, ⁓ know, everyone's like, ⁓ it's all about the data, it's all about the data, garbage in, garbage out. And I, there's a fundamental disagreement I have with that point because I know that good AI models can take in garbage, but still produce very clean outputs. We've seen it with the natural language models. Any data is usually good data. Reddit data is good, you know, ⁓ like junk, like forums or, or like newspaper, like editorial, whatever, like good and bad data is good. I think. But what the framework that new modern AI methods are taking is this. Here's a generative model that has been trained on billions of unlabeled data. I'm talking about just text or just molecules. don't molecules don't do specific things. They're just molecules. That's fine. Now we have specialized predictors, right? Predicting toxicity, predicting. ⁓ thermo stability, predicting whatever you want for your molecule, that needs good data. But the good data for those models does not have to be billion data points. That just has to be thousands, hundreds of thousands, maybe just like 500 data points. I can train a good classifier of toxicity with a thousand good data points of toxicity. I don't need a billion data points. Of course, if I had a billion data points, it would be better. So the main goal now is to, for whatever functionalities that we want in our next generation of technologies, we need to have good label data. hopefully at the patient level, but I'll take cellular level, environmental level, such that I can train these predictors. I have generators of whatever I want to make. These now guide the generator to make the molecule I want. So because we've decoupled this problem in modern AI ⁓ methods, if you can imagine, this gives us flexibility. We can go for high quality data, and we can use it to train smaller predictors. that guide big generators to make the things that we want. So yes, where do we get that data? This is now, I've told my experimentalist, hey, you don't need to get a billion data points. I just need a thousand good ones. It can adjust the style of experiments, the locations where we get the data, much better than where I'm like, I gotta have a million data points. They're like, all right, mRNA display, like some high throughput data collection method. We now have a lot more flexibility. Cory Smith: Yeah, excellent description there. ⁓ And stepping one layer back for AI, something that's really popular hitting the internet. I'm not sure how much you know about it, but it's OpenClaw running local AI agents. was originally Clawbot. It was too close to Clawbot, so they changed their name. Are you implementing any of these AI agent scientists in your research or personal life or in your research lab? Pranam Chatterjee: Yeah. Yeah. I think it's something where we have to look to ourselves and say, are we at a point where our time, our capability can be really enhanced with the AI tools? Of course, day to day, I'm sure you and I and everyone, we utilize the classic out of the box GPT-CLAWD, maybe an open source model for our work. But I think agentic AI is an area that is going to have a meaningful impact. We've seen with labs like James Zou's lab, other labs in the US where they're really utilizing AI scientists, future house, right, as companies that are really innovating on that front. For me, I still think that we are, you know, personally I think that we, good utilization of AI tools integrated with domain knowledge at ⁓ the scientist's level is the most productive way of conducting science, so it's an integration. But I think agentic AI gives me the ability to de-risk a lot of early ideas and workflows very quickly. And so for me personally, not yet the lab, in my day-to-day work, I am utilizing agentic AI to automate ⁓ workflows for running the lab, helping the companies, know, logistically. scientifically, think it's going to require, like we talked about earlier, an integration with the hardware capabilities, the experimental capabilities, the robotic capabilities that we have in the lab. Once we have that integration built up, I mean, I hate to say it. love, know, students are the best part of my job. But you can definitely imagine that agents are going to play a big role as scientists to innovate new technologies. Which tools are we going to utilize? Is it going to be open source or closed source? I think it's going to be time will tell, but we're seeing great progress in that area. Cory Smith: Excellent. And as a leader in the AI field, what do you think are your main ethical responsibilities, especially with your background in religion? What do you feel to be a good steward of this technology that you think about? Pranam Chatterjee: Yeah, you know I'm in the weird spot where I'm, and there's a growing number of folks, but it's still a small community of AI on the bio front, right? So I think of course in regular AI, in natural language processing, ⁓ you know, big LLMs, I think there's a whole community of AI safety there. That's not, definitely not my expertise, but my expertise, you know, I'm just where the time frame that I happen to... to really innovate in the field is to pioneer this area of AI bio. And really, you can imagine the scariest thing that could happen in our world is that an AI tool or an AI agent could design a very pathogenic virus. These types of things keep you up at night because that could really be the end of everything. So I think it's just the safeguards that we put into our generative models. where we can make a lot of things. I can make a peptide that does really good therapeutic applications, but someone else could use the same generative model to design a very toxic toxin at the same time. So how do we ensure that our generative models are constrained enough such that we are making good things, but we prevent ourselves from going down the rabbit hole of designing harmful, syn-bio, know, molecular tools. And for us, it's educating the next generation of scientists in the lab on like where these, how do you use these tools effectively, but to training them in ways that are constricting exploration of dangerous latent spaces, right, of these models. We have that ability to train those models. Right now, training those really good models is still very theoretically complex, still hard. There's ⁓ expertise needed for that. So we have to put our expertise in shifting model outputs to be following ethical safety guidelines. In biology, it's very important. Cory Smith: I think just because you can do it doesn't mean you should do it. That line from Jurassic Park, I'll share one story from the church lab. I applied to the DARPA Safe Genes Grant, which was very interested in gene drives, containment. We won that award and went through it. But during the presenter, someone from George's lab, I won't say who, said, why don't we just create the worst virus we could imagine and then think of the safety protocols for that? And George, of course, says, I don't think that's a good idea. I think you can focus on the safety independent of that, but. Pranam Chatterjee: yeah. Yeah, exactly, yeah. Cory Smith: Clearly people who have brilliant minds who think of an engineering challenge can do very dangerous things with some of these technologies. Pranam Chatterjee: Yeah. Absolutely. I think that's the fear, right? Is that it's there. It's there. The technology is there. I mean, right now, there are genomic language models that can design new DNA sequences. There are protein models and whatever. And the main thing is that if someone comes in and prompts that model correctly or conditions the output the way they want, it's so hard because these tools are democratized, right? Not even our generative tools we put about on Hugging Face, they're open source, anyone can download them. So you're really at the mercy of the goodness of people, right? And their ability to translate their technologies in a safe and ethical way. communication is important, but honestly, we have to also be reactive and defensive, where there will be bad actors in AI bio and biotechnology. We just have to be as easily as it is for me to make a therapeutic for ⁓ a drug. Right now, we can make good molecules. We should be also ready to make anti-toxins and anti-virals very quickly, ⁓ such that if that happens, we're ready. It's a sad way to operate, a very reactionary way, but I think we're going to hit a future where that's just something we have to do. Cory Smith: Yeah, think, yeah. And important for raising everyone up. George brought this up in the podcast, that increased income inequality or disparity for people, that's gonna allow them or promote the environment that they will want to or need to use these tools for misuse. And that. Pranam Chatterjee: For misuse or for, I for advance, I don't think we have to be so negative in the sense of, it's for advancement, but advancement and misuse kind of have, can overlap in a way, right? Remember, people are just really trying to ⁓ be successful and do what they wanna do, but if that doesn't overlap with an ethical mindset and I'm gonna avoid misuse, we're going to have that. So both education but also, defensiveness and reactionary ⁓ preparedness is going to be important. Cory Smith: Excellent. And what do you think about artificial general intelligence? You're really within the AI biospace, but this is a big discussion. And ⁓ what are your thoughts on when it might come about and the risks or benefits of that? Pranam Chatterjee: I think the saddest thing for me is that as an AI person, part of it is like, I don't know if I like it. I like that we're developing all of these great tools that are gonna help us make good therapeutics, enable folks to be very productive. We're getting robotics systems, self-driving cars, all everything. But there's certain things that I feel make us very human. For example, everyone's like, I'm gonna solve Navier-Stokes with Gemini and you're like, ⁓ that is so sad because you that is like the most wonderfully human thing we do. Like what is the best thing you can do? Like the coolest capacity we have, our brains have, you know, folks that can solve these very difficult theorems, prove out whoever solves, you know, proves Navier-Stokes is going to be, it's going to be such a great win for humanity. So, you know, for me, AGI is a point where I think AI is already able to do things that, if we said they can do it, get a perfect score in the Putnam competition, we'd be like, holy crap, that's one of the hardest things we do as humans, and now they're already able to do so. Our bar is so high that we may get to a point where every difficult problem that humans want to solve can easily be solved by the tool that we've built. Is that a future that we want? Maybe, right? Maybe it'll be very beneficial and powerful. But for me, I think that there's still a lot of core questions that human solving will be a really important moment for us. So I think AGI is kind of here, Corey, if I had to be honest with you. think we are already, AI is able to do things that humans can do very well, but it'll continually evolve and there will be a point where AI tools are just better at everything than we humans are capable of. then... you know, we have to reevaluate our own utility. Cory Smith: Yeah, think it'll be easy. We're gonna watch it along the way, and I agree, I don't think a world where it does everything will be a pleasant one. Maybe it will, but I think we need to have some purpose from what I've seen of with ourselves and with ourself. If you lose purpose, you lose that mission in life. We have to find a new mission if AI can suddenly do most things, most challenging things better than us. Pranam Chatterjee: Yes. Right. We need. Yeah. But I don't think we need to be so pessimistic. I think we adapt as humans. We find our purposes in different areas. Honestly, I think art, culture, know, music, even though we have AI designed music and art, no one likes it, right? Like who likes AI designed music and art? So I think that we're going to maybe come back to maybe the more human parts of us and really prioritize those aspects, even if AI can do it. So that'll be so fun to see. As AI achieves its maximum capability, Cory Smith: Yeah. Yeah. Yeah. Pranam Chatterjee: What pieces of humanity do we want to keep? Is it the arts and culture? Is it some science? Is it math? Whatever. I think that's something that we'll figure out as a people. And I'm very excited to be a part of that. And I'm excited that we're developing the algorithms that maybe could incorporate those wishes as we progress forward. Cory Smith: Excellent. One thing I'd to close with is science fiction. Big passion of mine and a lot of scientists. I wanted to ask if science fiction has influenced you and how? Pranam Chatterjee: Yeah, I mean, of course. I think when I started my whole career ⁓ in science, I really wanted this ability to change, right? To be able to say, hey, this is the biological system or the system we have. I don't like it. I want to fix it. It causes diseases. I want to fix it. It makes the environment bad. I want to fix it. And I think that reading a lot of science fiction really gave you the dream, right? Early, right? When you were reading... ⁓ you know, even like books like 1984 or others that were really imagining a future that was very technologically advanced. You're like, what are the advances I want to see? And hey, it hasn't happened yet, so let me be a part of it. So for me, science fiction was like reading science fiction didn't really push me in like random directions. It really grounded me in saying, what part of the future can I play a role in? And for me, it was in biology, right? Because like I said, I think biology is a great opportunity. We're not the ones making natural language chatbots or new ads or drones or anything. We're making therapeutics. We're making environmental solutions. We're making food security solutions. That's the future that science fiction inspired me to pursue. yeah, anyway, think it's great that we now are able to realize those things that we were reading about 20, 15, 20 years ago. Cory Smith: Excellent, and do you have any specific book recommendations or authors that influenced you? Pranam Chatterjee: Yeah, mean, you know, personally, I'm so in the now right now that I kind of just spend all my time reading textbooks and papers. personally, think 1984 was my favorite book, really, of all time, because it really inspired me. So I'm sure everyone's read it. But if you want to read it again, definitely take a chance at it. Cory Smith: Great recommendation, it stands up time, very relevant for today as well, so great reread. Well, Pranab, thank you so much, I really appreciated you taking the time with me, great discussions, great to learn about what your lab is doing, and ⁓ definitely gonna follow you closely and look forward to those next publications. Pranam Chatterjee: There you go. Thank you, Corey. Thank you so much for having me. Hopefully we'll talk again very soon. See you. Cory Smith: Yep, bye.