94. Gaming the System: NVIDIA's Vega Shah on Accelerating Biotech Breakthroughs

[00:00:04] Karl: All right, Dr Vega Shaw. Welcome to the grow everything podcast. We're so excited to have you here today.

[00:00:11] Vega: Thank you. I'm loving it already.

[00:00:14] Karl: Thanks so much. Vega, we're super excited to have you on the podcast.

[00:00:19] Karl: Number one, because I feel like we know you really well. I've interacted with you for years on Twitter. Number two, we have a very common, a good friend in Shelby Nussad of Compound VC who was on the podcast very early on. And three, we're super excited because we haven't had anybody from a company like NVIDIA on.

[00:00:38] Karl: we'd love to hear about your role. You are the lead for healthcare and life sciences. What are you looking for? What kinds of technologies or solutions are you helping to bring to market? And why, why is NVIDIA uniquely positioned to do that?

[00:00:53] Vega: Yeah. Oh, a lot of good questions. I kind of want to answer by first explaining, because I get this question a lot by first talking about why is it chipmaker? In healthcare and life sciences, what's up with that? Because get that question a lot. and think there's a good story behind So, for a long time, NVIDIA was well known for making GPUs. and if you are, a gamer, You know that the GPUs are great for making, the NPCs look really, really real. It makes their movements look real, their lip sync is great, their hair looks good. Pixar movies also use them. So, always known for that aspect.

[00:01:30] Vega: Someone at the helm, a leather jacket wearing person, had the foresight that, he was like, we're not a chip company, we're actually an accelerated compute company. that's truly what we do. Well, and with that in mind, they went out to build lots of verticals, within the company that serve different segments that can benefit from accelerated compute.

[00:01:54] Vega: And I don't even have to tell you this, but biology is a very data intensive field, we collect lots of data. It's complex data. and then the same applies for healthcare, right? On the clinical side, just different types of data. How great would it be if we could one accelerate compute, but also get insights using large language models using all of that data that's been collected in like the last couple of decades.

[00:02:16] Vega: And as technologies get better, for example, a sequencing gets better. We just collect more information, we're just gathering more and more of it. So they built the healthcare and life sciences. team within NVIDIA and they went and hired people that are like experts in that field. So we have a lot of PhD scientists that do research on large language models.

[00:02:36] Vega: , a lot of people that are also from that world that evangelize it. I would be an example of that. , and so my role is basically Thinking about ways in which, one life sciences, but also healthcare, how we can bring the products we have to that group. But also I think we've started, so we used to do in within NVIDIA, the healthcare life sciences built software.

[00:02:59] Vega: and they're all AI models that are designed to serve various segments. I would divide it into three broad categories. It's digital biology, digital health, and surgery. And digital biology is where we do drug discovery type work, and also genomics. digital health is where we provide LLMs that improve clinician patient interactions.

[00:03:19] Vega: and then digital surgery is where we actually have a product called hollow scan that sits within medical instruments that can give real time sort of image insights to doctors. So for example, if you have an MRI, A doctor can give voice commands and ask them to modify the MRI image and inspected in different ways.

[00:03:36] Vega: So you can get real time insights and make decisions as you're doing the procedure. and it's a really wide scope, and we want to widen that scope even more, because now we have all these generalized LLMs like Lama 3 that, are designed for multipurpose. Maybe we can use them for health care and for life sciences, My role actually is across the board on that, but even more importantly, like what I work on is our partnerships. So, NVIDIA has this great concept that we're not the experts on the science, so we partner with companies that are doing that and we're not the experts on the patients either. So we provide them the tools and we work really closely with them, but we don't actually create the software that goes direct to consumers.

[00:04:21] Vega: And I work with our partners that are cloud service providers. So we basically work with everybody. we're not competing. we work with AWS. We work with Azure, with Google Cloud, because a lot of our customers use, LLMs via those cloud service providers. so a lot of my work is how do we message and position with, cloud service providers or with partners.

[00:04:43] Erum: that's incredible. And I appreciate the example you gave when it came to the surgery part of it. I mean, that's something that we don't really talk about a lot, but it is super important it comes to at least health care and thinking of how to help individuals with complex diseases.

[00:04:57] Erum: But when it comes to complexity, there's a lot going on, Life sciences is complex. Healthcare is complex. A. I. Is complex. So gave us a surgery example. But what are some more immediate ways that? You're seeing today how A. I. Is revolutionizing health care and life sciences what are you seeing? That's really exciting across those three very complex spaces.

[00:05:20] Vega: Yeah, I want to give you, two distinct examples, So let's talk about life sciences first. the main product that we're like really excited about right now is BioNemo. BioNemo is our collection of LLMs that addresses like drug discovery. And mainly The initial portion of the R& D pipeline when you're looking for novel candidates.

[00:05:42] Vega: the way we address it is that BioNemo supports large language models. there are almost always open source and, we GPU optimize them. So we take models that have already been built, for example, like AlphaFold and we optimize it. So it. works in a compute efficient way on NVIDIA's GPUs, So you can obviously go to GitHub and set up the tech stack yourself, but we found that is actually the blocker for biotech companies, The effort and the know how needed to set up that entire tech stack, if you want to use AlphaFold in a production system is actually heavy lift especially startups.

[00:06:18] Vega: So we want to lower that barrier by giving people these models and making sure they get the right amount of compute, making sure they know we benchmark them and how we tested them for accuracy. So we basically. And we also make everything open source, what we've built AI models ourselves, that we also publish on and make it open source.

[00:06:40] Vega: we have a solutions architecture team that like literally helps with the setup of it. my best way of putting it is we help companies do it correctly because our expertise is the hardware and the software. and then we let them go off and do the research and predict the proteins and look for small molecules.

[00:06:56] Vega: so that's by Nemo. The other distinct example I want to give you is on the digital health side. We're really excited about that as well in a very different way. we with company called Hippocratic AI. That's building, agents that can interface with patients, and give them critical information before procedures.

[00:07:14] Vega: So, a great example is if you're about to get an operation done, you will need both advice and feedback in the beginning on, what to eat, and what not to do, to prepare for your surgery. Now, often this is done by healthcare workers, and literally, they have to call someone.

[00:07:30] Vega: And one, they have to find the time to call someone and they have a very short amount of time to give critical information to the person, not to mention, there's always things like language barriers or accents and interfacing that is difficult to navigate. so Hippocratic, the company that we're working with is using our large language models to create this complex agent that can talk to people.

[00:07:53] Vega: It has more time. It can really spend time answering questions. It can understand different accents. I think not yet, but we'll eventually also speak different languages and basically prepare people. So it's filling a gap that really. Exists, but there's no person that can do it right now, our work there is help them build that architecture because they're using multiple elements, Like, audio to face they're using, a couple of, translating models that, translate language. so I see the impact in two different directions in that in life sciences is where it very much accelerated compute for drug discovery and finding new candidates and moving drugs faster.

[00:08:33] Vega: And then on the clinical side, how do we improve, patient outcomes by just. getting them more information and attention?

[00:08:39] Karl: I'm super curious. I mean, I'm thinking about in the life sciences, in genomics, we don't have, digital models of cells at this point, maybe you might be able to have a model of even a mitochondria, but I even think like really basic life hasn't been modeled yet.

[00:08:55] Karl: And, I have so many questions in terms of what you guys are thinking about from a data point of view, because life sciences is probably one of the top 3 most data intensive, , fields. I know finance is another 1 and health care is the 3rd how do you think about that Tsunami of data being generated and how your language models are able to deal with that. And how do they improve over time? I'm sorry. It's like 3 questions in 1.

[00:09:22] Vega: that's a good question. So, for genomics, we do have a product called Parabrix, which does exactly what you just said. It does pre processing of NGS data, and effectively helps GPU accelerate it, So if you are using a GPU to do that processing, it will do it much faster than a CPU.

[00:09:40] Vega: speed is an element of why pre-processing is difficult, Speed and compute necessary. So that's what we're addressing. Again, I'll go back to this many times where the accelerated compute company, we don't have an army of genomicists thinking about the model. so we do.

[00:09:57] Vega: Assist that process in some ways. but then we also researchers that interact in the genomics world to do collaborative research, it's not something we're not thinking about, but we also don't think of that as like our strong suit. I agree as a scientist from my previous life.

[00:10:13] Vega: There aren't good ways to model a whole cell think the problem is multi layered, it's the data, we have a lot of data, but what we've collected is sometimes unusable because it's not in a standard format or it's missing information. but then there's also just a lot we don't fully know.

[00:10:31] Vega: we are still adding to the body of knowledge as we go and it's a complex problem.

[00:10:37] Erum: You talked about, GPUs, talking about microchips and devices. Have you worked with companies like Illumina you're talking about accelerating compute. So what about miniaturization and. Cost you said speed i'm trying to understand can I take my phone and be able to sample?

[00:10:55] Erum: some water and be able to do a whole analysis I mean, is that where we want to get to in like nvidia's vision?

[00:11:01] Vega: Not our vision, because we're always the person in the background helping the vision, but we do work with sequencing instrument providers, several of them, again, Parabricks is probably the most popular product there. it could lend to the application that you just talked about, where you do preprocessing faster for NGS data, but we.

[00:11:20] Vega: Entirely rely on them to tell us what the vision is. We don't try to shape the vision of a genomics company because they are the experts on that.

[00:11:28] Erum: So here on grab everything we talk about sometimes life sciences with I mean biology is everything We say from our friends Kristen Ellis of lower carbon capital when she came on, and I always say this everywhere we go. I'm sure Carl's tired of me saying it, but what she said is that the Venn diagram between climate and biology is a circle.

[00:11:46] Erum: It does the same thing, And we're biology, biology is everywhere. And you're talking about life sciences here. Have you encountered any companies that are looking at climate biotech, any other solutions? I know you're specifically looking at health care and life sciences, but just in terms of NVIDIA, how have they supported the vision of sustainability and regenerative companies?

[00:12:08] Vega: Yeah, yeah, yeah. Good question. So. the breadth of the entire company's product also goes into material sciences, predictive modeling and material sciences is 1 of the fields that's working heavily in climate change. They're basically working on carbon capture technologies, Materials that can make it more efficient or speed up carbon capture.

[00:12:27] Vega: they wouldn't come to us. we're not that department, but we do get companies that want to work with us on that problem. And then NVIDIA has supported weather predictive modeling, which can be expanded out to, global climate change tracking is the best way to time we show up, it's always like, here's our hardware and here's our software. How will you use it? And we will imagine with you how you want to use it. So we often rely on people that are obsessed about these problems to come to us and tell us, here's what we need. Can you build it?

[00:12:59] Vega: so yeah, there's carbon capture companies, definitely some that NVIDIA has talked to. at the Santa Clara headquarters, we have a, demo room where we show all the technologies that we've supported, and we do have a weather predictive modeling, demo over there, which is really impressive because it can predict what the day will be like and show you the pressure patterns for the next several weeks and things like that.

[00:13:21] Vega: Yeah,

[00:13:21] Karl: That's amazing. and what it makes me think of is I saw yesterday an Autodesk video where they talk about really being able to be much more comprehensive when it comes to any kind of construction in terms of the impact that construction has from a carbon point of view, which I'd never heard them talk about before.

[00:13:41] Karl: And I wouldn't be surprised if all that is just running on Nvidia hardware.

[00:13:44] Vega: I hope so.

[00:13:46] Karl: So you're also, part of the inventors team. and we're curious about what that role looks like and what trends are you seeing in tech bio and or biotech startups? how do you support them? What are you most excited about?

[00:14:00] Vega: Yeah. Oh, that portion is a small part of my role, but also one that's very dear to me because I've been advisor before. I love the biotech community and I'm constantly trying to learn about new startups. So I really enjoy that work. I'll talk about a few of the companies that's publicly have received investment from nvidia.

[00:14:22] Vega: 1 really popular example is recursion, they were together on the stage at, J. B. Morgan last year and, announced to the world that we're investing in recursion. They also use our bio NEMO platform for some of their predictive modeling. and they, use their own. custom model called phenom beta on by any the other company I will mention this is more recent is to rate therapeutics.

[00:14:47] Vega: we've invested in them and they're doing small molecule discovery. I think the commonalities across the board. So, 1, we look at a wide range of companies, for investment, the commonalities tend to be like, they're using AI, for discovery. and often they have a combination of platform and, validating it in lab as well, Cause that's where the power is. so terrain, for example, is doing something called reverse engineering of small molecules using their own AI model where you can, Describe the parameters you want for a small molecule, and then it will generate that for you, which is incredibly powerful. Because in the past, what we did is search for the needle in the haystack, we would look at large databases, small molecules searching for certain parameters. Now we can go the other way and actually just create them. Very impressive work. We just announced, they just published a paper, it's a preprint, I think, and they, talk about this reverse engineering technology that they've created.

[00:15:40] Vega: So, yeah, I think we're looking at a lot of companies that are doing that style of drug discovery.

[00:15:45] Erum: That's super interesting. because we just started working with a company that's doing something very similar. DeNovo small molecule generation. is that similar to what Terrace doing?

[00:15:57] Vega: they might agree with me, but I think they call it reverse engineering.

[00:16:00] Vega: that's the, so,

[00:16:02] Vega: yeah,

[00:16:02] Vega: I don't know.

[00:16:02] Erum: yeah, makes sense, Because, you're looking for specific binding affinities, specific solubility, there's so many parameters, you know, what you need. in order to design it, I mean, that's brilliant messaging. Congratulations to them.

[00:16:14] Erum: we'll give it a messaging lab stamp of approval, but, that's incredible. So that. Is 1 word that I think When it comes to generative AI and, people say it a lot and I'm sure the term is again, I'm bringing this back to messaging lab is turn.

[00:16:27] Erum: It's use a lot. but ultimately, a lot of this is simulation, Simulating would you agree with that? Do you have differences between the terms generative AI and simulation? What do you think about that?

[00:16:39] Vega: that's a good question. Yeah. And I think I've heard three different terms, prediction, simulation and generation.

[00:16:45] Vega: it's hard to place them really rely on the model builders to tell me, and Toray definitely calls it a generative AI, platform but honestly it's hard to tell you because i don't think the early models of alpha fold would not have been called generative they were more predictive

[00:17:00] Vega: that's answer for me i think it's like a fuzzy nomenclature

[00:17:04] Karl: yeah. I mean, we've talked about it in terms of when do you start saying, simulate versus generative, it does start to get complicated very quick. if you don't know what people are talking about,

[00:17:14] Vega: think it also goes back to that this field is so new that we don't have words to describe a lot of when we don't have standard bodies either we don't have A consortium that's nonprofit that's setting the standards for how to use generative AI, especially for things like drug discovery.

[00:17:32] Vega: I mean, the FDA is attempting to do that by setting up a committee, but we're early days. And so I think, A lot of these things will have to be defined, otherwise you just have to rely on, the company or the model makers describing it. Yeah, it's not just, the newness of AI. It's also our understanding of what its capabilities are continues to evolve. And then at the same time, we barely understand the complexity of biology. I had said we don't have a digital model of a cell.

[00:17:59] Karl: And to me, what's fascinating about that is we don't even know how many processes that are going on in a cell at any given time. And there are so many of them. So I keep thinking. That's going to impact the way we think about things like a GPU, because to be able to create those kinds of models, which we're going to want, and we're just talking about single cells.

[00:18:17] Karl: We also want to be able to model, organisms, and then also get to the point where we can model, biome consortia. So all of those things are going to require different kinds of language models and, and we don't have the definitions for those things yet.

[00:18:30] Vega: totally. Yeah. one of the things you said reminded me of the work that I did when I was in grad school, which was discovering new species of bacteria and then constructing de novo their genomes. You'd think the technical side was hard, doing the de novo assembly of a genome of a bacteria, but actually the hard part was.

[00:18:48] Vega: convincing everybody that this is in fact a new species, Because you have to go look at the entire world's database and see if anyone had found it, do a bunch of alignment against large databases. And even then, yeah, we're 99 percent sure this is new, right? There's so many things we don't know about the living world, including, microbiome and that to me is very exciting because I think there's a lot of potential to find new enzymes, new drugs there.

[00:19:14] Vega: Yeah.

[00:19:15] Erum: Oh, my God. You're speaking with your language. This is something that Carl and I have been talking about a lot with some other, biologists and, people that are in the industry of personal care and beauty of food and looking at. And again, we're talking about when you talk about grow everything, beyond just using a coli, you we discovered one, one thousandth of the Bacteria or microbes that exist on earth and we're living and breathing, discovery and seeing what's out there and just fascinating what types of functionality these microbes can imbue on a process, I don't think it's a fish that Excretes glass deep underwater because of the pressure, you know, just like, wait, what? that's crazy. What kind of biology? I mean, I'm sure pressure has a lot to do with it. It's just fascinating to see, what's out there and can we simulate, generate, predict.

[00:20:03] Erum: Which, microbes are there and like what they can do and then how do we incorporate them into like an application that could be helpful to us Totally. Totally. another cool thing that now I'm taking you on a total tangent, but

[00:20:18] Erum: Please, we love

[00:20:19] Erum: tangents.

[00:20:19] Vega: read this last night. I was reading this pre print that just came out where they found an entire novel protein superfamily in the gut microbiome, called oglans. And I love that.

[00:20:30] Vega: We're just walking around with novel proteins inside us. That just blew my mind. I had lay down and think about,

[00:20:37] Karl: Yeah, I saw that on your Twitter feed. and I had seen it somewhere else. And I was like, I got to go back and read that prepend because it's so crazy what we know is so small compared to what we have to discover when it comes to the complexity of biology.

[00:20:51] Karl: so, being able to just appreciate that and then. To think about how we model it and this problem that you were saying where people are convincing people that you had a new species. Wait, you mean, there's just not a database. You could interrogate these genomes against, you know, where is that?

[00:21:07] Erum: Yeah. what are these novel proteins do? any information on that? Any teaser?

[00:21:13] Vega: I'm trying to remember. I read it very late, last night, but I think they were found in these viroid entities within the gut microbiome. I have to go back and look.

[00:21:22] Karl: it's funny because I'm looking at the pre print right now. These are an unrecognized class of by roid like elements that they call obelisks that they identified in the human gut meta transcripts. So that's just the beginning of this thing. So it's pretty insane.

[00:21:41] Vega: when I posted it, someone very matter of factly said, yeah, there's tons of undiscovered proteins inside the human gut microbe. And we haven't cultured many of those species of bacteria. And I was like, oh, my God,

[00:21:51] Karl: Yeah, yeah, yeah, yeah, we've mentioned it on the podcast and we've had, one of the co founders on cultivarium, focused research organization based out of Boston, and one of the things that they were really working on was this ability to culture I want to say gut microbiome, or whatever, bacteria that live in the gut microbiome, because we still don't know how to do that.

[00:22:11] Karl: And we still have to characterize all those microbes that are in there. And as important as they are again, this is 1 of those modeling problems that we're going to have to get to, because how you eat the way you live all impacts on microbiome. And we all have a unique 1 and then we've also had Stephanie color of Persephone.

[00:22:29] Karl: And she told us that there is this extinction event happening in the gut microbiome of infants where they're missing essential microbia, because they're either not being delivered vaginally or because the mother doesn't have them. And so that is going to have huge health implications.

[00:22:45] Karl: So there's a lot there.

[00:22:47] Vega: oh my gosh. Yeah, and I, can tell you a bit about my, history of working as a scientist. When I first finished my postdocs, I did my postdoctoral at Joint Genome Institute. And then, my first job outta postdoc was at a, health and gut microbiome company called January ai. And that's exactly what they were doing.

[00:23:04] Vega: They were providing a service where customers would, have their gut microbiome sequence, but. One of the goals was to track and maybe, monitor A1C impacts. So if you're like pre diabetic. And so we would give them both an insulin monitor, even people that aren't diabetic, but are like pre diabetic, and then also track their gut microbiome.

[00:23:25] Vega: And I tested some of it as I was with the scientists, like trying to figure out what happens. Well, there were so many interesting things, that came out of it. Your relative diversity changes so much, when you travel, one big one was going to Japan, because you're consuming a lot of seaweed.

[00:23:42] Vega: there's an entire group of microbes that are basically, recruited into your gut microbiome from eating seaweed. And so your gut microbe changes so much when you travel to certain countries. and then my next job actually was high throughput isolation of microbes. this company called isolate bio and they had microarrays that could isolate.

[00:24:02] Vega: and we tried gut microbiome. I think we had most success with soil and water, but we certainly tested gut microbiome a lot. Cause that's a great application. And we were working with UC Davis on this project for, milk. So mother's milk, how that is a source of microbes for infants. is all in my past.

[00:24:21] Vega: This is before I wanted full software.

[00:24:23] Erum: I mean, I'm sure it requires a lot of GPUs to model a lot of in

[00:24:29] Erum: any type of biomes,

[00:24:30] Vega: absolutely. Not just model, but image analysis uses a lot of GPUs. so coming back to Nvidia, we do have, a large language model tools that do image analysis specifically. So to, fill in gaps and images to do some level of prediction based on the images you already have.

[00:24:48] Erum: But images of cells or microbes?

[00:24:51] Vega: yes. We have done

[00:24:52] Erum: Oh, oh, interesting. You should look up Moni. that's what it's called. And then we talked about how the scan, which is the digital, surgery one.

[00:25:02] Erum: And are these models being used together? we just started working his company and, enough to be dangerous or I'm completely misunderstanding what binding does. But in terms of the modeling, is that using smiles data? So

[00:25:14] Erum: it's like the string. Okay.

[00:25:16] Erum: Okay,

[00:25:16] Vega: by an email is even broader where it supports multiple models, It's a framework for you to use your model. On. can use the model you created yourself or an open source model. And one way we facilitate is we GPU optimize the model and we give it in a container.

[00:25:33] Vega: and you can run that container anywhere you want. You can run it on AWS. You can run it on premise. we're calling it a NIM, which is NVIDIA inference microservice. So bio NEMO is implemented using a NIM and, You could be using multiple models using BioNemo in conjunction, So you can be using many in tandem, including AlphaFold or, recently we put out two models that I want to talk about.

[00:25:58] Vega: DiffDoc, which, predicts, how proteins and ligands, bind with each other. And MoleMim, which is, actually also like inverse generation of small molecules using parameters.

[00:26:10] Karl: And, so much of this stuff that we're talking about is complicated science, topics. And in general, we talk to other scientists, but we're always thinking, from the messaging lab point of view, also with some of our clients is how do we translate that stuff into things that are more user friendly depending on who the end user is.

[00:26:27] Karl: How do you think about that,

[00:26:28] Vega: Yeah. Yeah. so that is one of the challenges when NVIDIA went looking like, what should we build? One of the challenges was the absence of GUIs. so we do include that as well with BioNemo. And you guys can go try it out today. If you go to build. NVIDIA. com, you can try out our BioNemo models.

[00:26:46] Vega: and you can put in your own, Molecular data file or we have examples on there that you can test out. So, the ability to have an interface. We find is very important. We also work with companies that build their own interface as well, and they, have more knowledge of what those scientists want.

[00:27:06] Vega: 310 is a good example, where they've built, GUI that is very nice, very user friendly, and it can do that predictive modeling as well.

[00:27:14] Karl: think it's worth it just for our audience, the people who don't know what a GUI is, it's a graphic user interface, it's so you do not have to look at lines of code when you're looking at your computer.

[00:27:26] Vega: exactly. Yeah. and I come from the world of dotmatics and bench sling, which are two well known, data induction platforms, And in my last jobs, I was a product manager at dotmatics for, integration with the ELN and, GraphPad Prism, which everyone loves GraphPad Prism for analytics.

[00:27:45] Vega: One of the main things was what user interface do we give scientists so that they can one easily and in very few clicks get to what they need to do, So it's surprising what causes a lot of delays in Processing of data. It's a lot of copy pasting. It's a lot of clicking.

[00:28:02] Vega: and it's often because, the U. I. Isn't necessarily built by someone that ever worked in a lab, think about why those software are like that. It's because a lot of UX Is missing the context of what is the day to day of a scientist? so I had a lot of fun in that job because I could add that.

[00:28:18] Vega: It's like, yes, that won't work. It's too much clicking and I have gloves on. I can't do that.

[00:28:25] Karl: Yeah, I'm thinking a lot about the workflows what does that workflow look like? How do you set it up in a way that. Makes it easy for people, but I think, again, so much of this stuff is so new and you have the scientists who are doing the work, but they don't necessarily know how to develop those gooeys that are going to make it easier for people to just move the parts around so that you can actually design this up front.

[00:28:46] Karl: so I think it's very exciting, to see how that's going to evolve.

[00:28:49] Vega: Yeah, absolutely. my opinion. Take it with a grain of salt. Different people will tell you different things. I don't think that biotech companies should try to build their own. GUI or their own software internally. the analytical part or building a large language model, totally.

[00:29:05] Vega: But trying to build a GUI as a bad tech company feels like you will do it poorly, it will scale poorly. and then the one person that does it, if they ever quit, going to not be.

[00:29:15] Karl: Yeah Totally.

[00:29:19] Erum: that's a good point. That's a good point. But at the same time, there's a lot of tech bio companies and when we're talking about drug development, at least, a lot of them, they're like, oh We're gonna get a platform for big pharma to use But then big pharma is five customers or let's just say that, maybe it's 20 customers and then they have to charge a certain amount of money in order to Be valuable because a lot of Tech bio companies are venture backed, Unless, they're a bunch of smart people that, they're not venture backed. So they need to make millions of dollars in revenue in order to, do that. so some of the companies that we've spoken with, they started out as tech bio, then they're like, all right, we're just going to create assets and then sell them to big pharma and take it all the way because that's the proof point to be able to do that.

[00:29:58] Erum: But have you seen, I any, tech bio companies that are platform based that are doing it right where, like you were saying, like a pharma company, they're not building it in house. Like, all right, we're going to work with this company,

[00:30:08] Vega: Oh, good question. two thoughts I completely agree. the platform plays really hard. The platform play minus an asset is really hard because the asset is where most of the investment and money is around, but also your platform, does it work or not? We're only going to know if there's assets coming out of it,

[00:30:26] Vega: So there's this inevitable pattern of you start building a platform. And often it's companies started by founders that more tech people than biologists. And they come in, they're like, yeah, we'll build a platform. Of course, this is generalizable, we can build something.

[00:30:40] Vega: But it, Often tends to either go via partnerships to the asset play, even Schrodinger, this very big platform company, which in my mind is actually like a small molecule like Schrodinger is a totally a success story, but they also have labs they also do experimental work to validate some of the work they do.

[00:30:57] Vega: So, yes, platform plays hard. I think. The tech file companies that I see that are doing it right Again, I'm such a biased person to talk about it. But recurrence hitting all the important points, high throughput data collection, large language models, and then also assets, And it's the trifecta of wins in my mind.

[00:31:16] Karl: they were just in the news yesterday because they're about to release data on one of their assets but the initial assets that we're going to see data from recursion are actually not assets that they develop using ai They're either things they unlicensed or things that so I think it's going to be interesting and hopefully they'll be successful with those Because we're all looking for the successful ai design molecules that will have a big impact on treating some disease

[00:31:39] Erum: Yeah. hopefully it's our client, Carl.

[00:31:41] Karl: Well, we do have a client that's doing that in the molecular blue space, they're doing very well.

[00:31:46] Vega: So this is something that's been on my mind for a while. have there been IND filings of a, computer aided? Drug discovered and this is something that I don't even know how to track this necessarily because it's not something we call out, but it'd be really interesting to see if.

[00:32:01] Vega: An AI based drug gets through IND filing.

[00:32:04] Karl: I'm going to answer your question and I'm going to caveat it by saying, probably read the paper at the beginning of this year, there is a science paper that has tracked this I might be totally wrong I read it very quickly. I think that there's a couple of things that are.

[00:32:17] Karl: Close to the finish line, but I don't think anything's in the clinic beyond maybe a phase 1 or phase 2 again, I'm totally caveating this by saying, I don't know. I don't remember. I might be wrong, but let me find the paper and I'll send it to, um, Vega, because I think it's super interesting.

[00:32:35] Karl: And I think it is something all of us should be tracking. I mean, what we do know is that we've seen some really nice investments in companies that are using AI. I you mentioned recursion, I mentioned that, they're about to. release data on their assets. we work with a molecular glue company that is doing really well on the investment side and is getting some things also into the clinic and all their stuff Is ai optimized and designed because all the molecular glues are based on Not our client many of are based on thalidomide and that's where the molecular glue basis comes from um Our client doesn't use that as the foundation for their model.

[00:33:11] Karl: But, there's been some really nice investment in that space. I'm super excited. I think we should all be tracking this.

[00:33:17] Vega: the FDA is also, I think, in tandem working on creating one better policies on how to like grapple with AI based technology. Not just in the drug discovery, but throughout the R& D pipeline, How do people use AI for even planning clinical studies and things like that?

[00:33:35] Vega: so I think it'll get more refined, how we think about the use of AI and taking drugs to market.

[00:33:41] Karl: I feel like it was not too long ago. I 1 of the areas we've done a lot of focus in this year has been on by manufacturing and I remember a year ago or so reading, the impact of. AI in drug development. there's so many aspects of drug development that are so complicated.

[00:33:58] Karl: I mean, everything from yes, discovering drugs and, safety profiles before you even put them in patients, but things on the clinical trial side There's a huge amount of things that need to be optimized there And that is starting to be added or or augmented with ai Optimized with ai and then also on the biomanufacturing side, So all of these places have the opportunity for people to apply these technologies and optimize them and change them Yeah,

[00:34:26] Erum: the company in silico, they have an A. I. Design drug that received F. D. A. I. N. D. Approval back in April. I don't know how many tests they've done before that, and I don't know if it's just pure A. I. But they have worked with F. D. A. So they might be working on those frameworks that you're talking about Vega, but I'll share this link in the show notes.

[00:34:45] Erum: and, I think that's very exciting if they can, Accelerate the pace of drug development because are so confident in our simulations of drugs. I'm bringing that word to you. I'm bringing that word. I'm bringing it to the industry. Simulate, NVIDIA. I'm so impressed with, I think lot of people are going to listen in to our episode because they're just so curious, but NVIDIA has been in all over the news. one of the highest market value companies in the world, Apple, Microsoft, and NVIDIA, just crazy to see, I think 80 billion in revenue, which is incredible, Who's the biggest client?

[00:35:23] Erum: where are they getting the most money from right now? Is it from Apple and Microsoft? I mean, and then also who's the biggest healthcare and life sciences client

[00:35:32] Vega: Good question. we are a company that. Our technology only shows up with partners, It works when we partner with lots of companies. And as you know, Healthcare and Life Sciences is like incredibly distributed.

[00:35:43] Vega: So I couldn't name one big customer. It would be many, many, many customers. Especially the life sciences portion, because it's more new. We want to work with everyone, What we want is to drive adoption. and we're thinking about the developers that work on this. We're not necessarily thinking about revenue.

[00:36:01] Vega: And it is within the company's ethos, before Rev. So that is actually something that everyone says, we obsess about the end user, we obsess about the developers that are going to use these technologies will make their life easier. and revenue will happen.

[00:36:17] Vega: And in their hiring practices. even though I read this wild number, like 75 percent of NVIDIA employees are millionaires. I'm not one of them, by the way.

[00:36:27] Erum: Not yet.

[00:36:28] Vega: But they hire people that are obsessed about their Field of choice.

[00:36:33] Vega: So if I go talk to someone in the robotics team, it'll be this person that's dedicated their whole fricking life to robotics, they're not even like looking at the stock every day or trying to see how much they're worth. They're just obsessed about the problem and they truly hire people like that.

[00:36:47] Vega: So even though there's so many wealthy people that work at the company, when I go to the office, people are driving rundown cars, like they're old, Prius from 10 years ago, it's a very interesting place because everyone imagines NVIDIA employees I'm sure some have, but people aren't going out buying yachts.

[00:37:04] Vega: They're still obsessed about the problems. And I really love that about NVIDIA. And speaking to who are the biggest customers? I think The way we, provide these tools, these LLMs to, end users is via cloud service providers, we work really closely with AWS, Azure, Google, Oracle, and they tend to be, the largest consumers of GPU, which, uh, I mean, it's kind of common sense, you

[00:37:29] Erum: Yeah.

[00:37:30] Karl: Yeah. My 1st introduction to, Nvidia was, well, 2 things. 1 is, my sons were all picking stocks a long time ago. And my 1 son was like this is going to be a big company. my two youngest sons built computers so the GPUs was a big gaming computers.

[00:37:47] Karl: whatever they put in, that was going to be the biggest expense was the GPU. they spent a lot of time schooling me on, why it had to be this. Nvidia GPU for the computers they were building.

[00:37:59] Vega: Yeah, I think gamers are still part of the ecosystem and we still build great stuff for gamers. like playing games, but I never built gaming stations. So it was never something that I thought about. I think for me, NVIDIA became a thing that maybe two years ago and I realized they're in like Life Sciences and I got really excited.

[00:38:19] Vega: It was like, Oh my gosh, it wasn't even the stock. I think my dad two years ago texted me and he's like, you should probably buy NVIDIA stock. That was one line. It was like maybe buy NVIDIA stock.

[00:38:30] Erum: great fatherly advice there.

[00:38:32] Karl: There you go.

[00:38:33] Erum: that's incredible. I was just going to say we're touching on a topic that Carl and I had talked about, and I definitely foresee this happening in the future is the gamification of biology. And like, we're talking about how do you simulate a cell?

[00:38:47] Erum: How do you create a game that can not only But also help, um, create a realistic model of a cell or just biology, I don't know if you've seen anything like that, or you could talk offline about how we could do this

[00:39:01] Vega: Totally. Well, has already been a thing, hackathons have done protein folding experiments. and protein folding has been a big gaming, experiment for a while. 100 percent agree we should gamify biology. We're in, truly the, age of biology. for two reasons.

[00:39:17] Vega: It's the infinite problem. There's so many things to be solved. It can keep us occupied. Till our species gets extinct, right,

[00:39:25] Erum: Yeah.

[00:39:26] Vega: right to our end. We can keep discovering things in biology. So if anyone's looking for an infinite problem, this is the right field. You'll never get bored. I also think we should gamify healthcare.

[00:39:36] Vega: We should gamify taking care of ourselves because we're already so addicted to all our devices. I hope AI agents and various LLMs are used to how to get. Better nutrition, how to sleep better, how to have work life balance. I need that too.

[00:39:53] Erum: yeah, I agree. think it's a great way to wrap up our conversation. I mean, it's been complete joy speaking with you. when Carl mentioned that, you you're going to come on the pot. I was so excited because my first experience with NVIDIA is I watched the demo day that happened 2 years ago.

[00:40:09] Erum: and I saw that they were getting to life sciences and I was like, Carl, this is huge that NVIDIA is getting into it. And I learned about the bio NEMO platform a little bit there, and I was like, this is incredible. And then, of course, as everyone gets keeps hearing about NVIDIA, this big company, what's going on?

[00:40:22] Erum: so thanks much for shedding light on what's happening at NVIDIA, all the amazing things you guys are supporting. You're supporting all the vision of these great companies that are helping us explore more, when it comes to biology.

[00:40:33] Vega: Thank you, Carl. Thank you. Um, this was a wonderful conversation. I'm so glad we talked about things I haven't thought about in five or six years. So it was great. I loved

[00:40:42] Karl: Thank you. Yeah, no, I'm really glad. Thank you so much.

[00:40:46] Erum: We'll certainly have you back on for sure. cause I know you're a bit newer to NVIDIA. So whenever one year anniversary, and you're a little more comfortable with telling us some inside scoop. Ooh,

[00:40:58] Vega: then too. So yeah, definitely would love to talk about that.

[00:41:01] Erum: awesome. you so much.

[00:41:04] Vega: Thank you.