Name: A More Potent Pipeline
Uploaded: 2026-06-09
Duration: 22 min

When it comes to drug discovery, most AI applications stop at prediction. A far more difficult problem is engineering the actual therapies—designing candidate molecules as well as the models to evaluate them. Lila Sciences and Nvidia are working together to build the "scientific superintelligence" to close that gap—with hope that it can accelerate the pace of discovery and speed drugs to market.

Geoffrey von Maltzahn, Chief Executive Officer, Lila Sciences 
Kimberly Powell, Vice President, Healthcare, NVIDIA 
Andrew Nusca, Editorial Director, Brainstorm, Fortune 

A More Potent Pipeline

Conferences

Um, I'm excited to have this conversation, so I'll kick it off by just putting it right in front of you. You both believe that science in this country is underinvested. Why? Oh, OK, I'll, I'll give a perspective here. One, there was a significant reduction by our current administration across the board of science. You heard what happened to funding in NIH, um, tens of billions of dollars eliminated at the exact moment where biology and healthcare is hitting its chat GPT moment. OK, if you think about um the the scientific method, if you actually say it out loud, what is it? You, you pose a question, you go gather context, you make observations, you reason over it, and then you take some action and that action provides more information and more context. What have I just described? Agents. They are exactly parallel, and so the scientific method is entering this new paradigm right now as we speak. And so we need to empower every scientist to have this tool and have it in abundance. And what you're going to hear a lot today is it's not only about um these scientific agents maybe being able to read biology or read science, it's actually can it do science. Agents is all about being able to do work. And so while science for the last many centuries has been human limited. We are now able to marry up both agentic science and labs and automation, physical AI together to create this closed loop on a massive accelerated scale. And so if we defund now while the rest of the world leans in, which Europe is leaning in, which a lot of the Asian countries are leading in, we will be left behind and we just cannot. Jeff, I'll, I'll pull on that final thread in that a lot of the reactions to changes in investments in science are based on the paradigms of human science in the past. And you know, expressions that it's unfortunate that we're, you know, underinvesting relative to what we've done historically. What Kimberly just landed is that the even bigger reason to be investing aggressively right now is that for the first time in human history, a dial called scientific intelligence. is in the picture and for the same reason that there were scaling laws of stuffing the internet into larger models with more compute that created higher intelligent AI systems, one can now conceptualize that scientific method that Kimberly was describing as a means by which to open up a new scaling law of scientific intelligence. And since scientific intelligence is the most important determinant of ROI in science, Um, it is going to become, it should be a national priority for the US and our allies to build the highest scientific intelligence on Earth. Falling below the scientific intelligence of one's adversary at a corporate level, at a sovereign level, is almost like an unimaginable competitive disadvantage. Yeah, I should ask the room how many, how many folks in here would call themselves scientists? Raise your hand. Alright. So for those who didn't raise their hands, you know, you've talked about in our conversations prior to this, um, science, um. Technology and AI in particular being a great unlock for science. Science is limited, is, is, is kind of how it has been characterized before. For those of us who are not scientists, what's so limited about science today that AI would be a great unlock. Uh, I'll take a stab at it, but, but Lila's building it. Uh, I'll take it from our perspective from being a technology company, um, as I described, what is the work that is now enabled by agents. Um, if you think about the scientists, where they start, they start in their own human observation, so that's their own human experience. So you're already right there limited by your experience and what you've observed in the world. So the next thing you have to do is you have to augment your own human experience with more context. And that means reading, let's take life and let's lean into life sciences to just put science in perspective. One of the most important areas of science is life sciences. And so we need to understand biology. Biology is something that the world doesn't completely understand now, that's why we still have so much unmet human disease. And so we have to now keep up, imagine a human trying to keep up with the number of papers that will be published during this talk, which is of the tens, and try to stay up on biomedical knowledge and all of that deep understanding. Um, that's again, human limited. And then they have to start asking, you know, the questions, the prompts. What is the scientific goal I'm trying to achieve? And so on and so on. And so it starts from the very first step of doing science that is human limited. And you take that all the way from what you can do sort of outside of the lab. The other really important paradigm to think about here is what we've experienced with agents and coding is incredible. You know, you're seeing 3 times the productivity of a software engineer, and that is wonderful. However, science is harder than that. Why? Because we don't have what software engineering has, which is all of this ability to verify the results. Did the code compile, did it pass the test? Did it do what you expected it to do? That's a rather simple way to evaluate and verify that your agent did what you expected it to do. When you're dealing with science, the only verifier is the lab and is the biology. Did it do what you expected it to do? And so that too has human limitation running the labs 24 hours a day, sometimes doing very exquisite laboratory work that is so, um, you know, deeply specialized, right? There's only so many scientists um that can actually go in and manufacture and work in labs. Let's say organoids and cell therapies and otherwise, right? We have other labs that do high content, um, you know, development of data, but there's a, there's all sorts. And so there's just limitations all along the way and this is what, what we are now in the moment to unlock a lot of that human limitation and connect the two together. In such a material way for the speed of science and so I, I love your perspective and maybe you should teach them about what you're building because it's incredible. Tell everybody what Lila Sciences does and while you're on the subject, what exactly is scientific superintelligence, which is a term that your company likes to use. So Lila's mission is scientific superintelligence to solve humankind's greatest challenges. That is of course a very lofty one, yeah, OK, uh-huh, um, and scientific superintelligence sounds like it's a distant part of the future, and, and of course it is. It's also an accurate statement of the past now. So many of you will have heard of Alpha Fold, a Nobel Prize winning contribution to science from DeepMind. That is super intelligent on the task of predicting how genes across the kingdoms of life fold into the engines of life that we call proteins. And um our founding team of Lila had been pioneers of applying generative AI to disparate domains of science for 3 to 6 years before Chachi PT emerged, and we'd concluded that the models we'd created had become super intelligent for antibody design, for peptide design, for designing organic molecules and even inorganic materials. And since the only thing that distinguishes human science from the rest of the animal kingdom, science, and $10 trillion of annual revenue from approximately 0 is intelligence, if intelligence goes up anywhere in science, that's the most important thing in the history of science. And as Kimberly was just saying, like, when you spend your life specializing in science for decades, you can lose track of what a flaw it is that as a species we've had to shatter the world into thousands of little subspecialties in order to be able to comprehend it. And um A wonderful piece from 2019 by last year's Turing Award winner, which is like the Nobel Prize in Computing, called The Bitter Lesson of AI, which I recommend all of you look, uh, look up, is a two-page blog post that succinctly observed something that's been even more prominent in the years since and applies to science as well. Which is that in 70 years of AI research leading up to 2019, every single time AI researchers took an approach that was specialized to the human way of doing things, Those were destroyed by the most general and scalable approach. So we concluded that the most general version of AI systems being able to run the scientific method that Kimberly described and thereby enable a new kind of manufacturing process for intelligence, like think vast buildings filled with scientific instruments running the scientific method around the clock across materials, chemistry and life sciences. With all those hypotheses and conclusions flowing into a single kind of scientific mind. would be the future of something like 100% of. We set that in motion 3 years ago and have been honestly startled by how quickly we've gained conviction that what I just described is true. It's moving very quickly. Can For the uneducated here, can AI discover a drug right now? Uh, emphatically, but it depends on how you describe the, like where the challenge starts. So everything that Lyla's been doing has been living in, um, brilliant scientists inside of Lila stating, um, that we should go after some grand challenge. I'll give you a few examples, but we've, we've been applying a common system across material science, chemistry, and life. So one of the challenges we've been after is discovering new catalysts that can split water into hydrogen and oxygen. The commercial standards today use precious metals, specifically iridium and ruthenium. Those are rare enough that I imagine some of you haven't previously heard of those elements. And um A year ago, Lila's agents were making reasonable suggestions, like one of our um uh senior team members was a professor at Caltech. This has been among his domains of his life's work. And a year ago he'd say, my my lab never tested this, but this is an interpolation between what my lab was doing and this other lab. A few months ago, he started describing that a third of the agent's suggestions didn't make any sense at all. There was no reason to believe that that element or that crystal structure had any role in that reaction. And those are now humankind's highest performing catalysts for splitting water into hydrogen and oxygen, and it's quite moving to hear him talk about it in that he feels like he's seen a scientific discovery from far off in the future. We're now seeing this across designing antibodies, designing mRNAs, designing small molecules, and I think to many people this clawed code-esque moment of when a new intelligence gets injected into science and changes it forever thereafter. It may not feel imminent, but it is right around the corner. So I can't help but think so all these things we're talking about, um, heavily heavily regulated, right? We need standards in science we we have these things drug discovery for example very long process, very expensive, very regulated um some of these other things we're talking about, same deal. If we inject AI into the mix, even patents, by the way, even scientific patents, if we inject AI into the mix, are our regulatory frameworks ready for that today? I mean, we're, there's a lot of adaptation that's gonna go on, um, but, but let's be clear on, on, you know, the subject matter of science and it being regulated for. Human consumption, I mean, I think the opportunity here is upstream of that, right? So there's, there's regulatory measures that have to be, um, clearly put in place whenever you're putting humans at risk, OK, but if we kind of put that somewhat to the side because the FDA is working really hard, you know, thousands now of, um, AI medical devices that are being approved, you know, they're, they're working really hard so that. Things that have gone through the necessary um gates are going to reach patient we you know we want these things to reach patient whether that's a new surgical robot, a new computer vision algorithm that can detect something like pancreatic cancer or otherwise, right? But if we come back to upstream of that and the ability to do science, one thing that I think is just really, really important here is um. A lot of the work that we do together is a huge investment of Nvidia's in an area we call open models, open AI, not open AI, the company, open artificial intelligence models that we do and we invest in to create a generalized somewhat capability, but in areas we can uniquely contribute. So physical AI world foundation models that Lila uses for. Their lab, um, biology foundation models that they can use to do that antibody model development, um, language, multimodal language models that can, you know, read and parse through all of the medical literature and understand that. And then they, we, we create them as open models so that amazing companies like Jeff can take them, adapt them for the core capabilities that they're trying to create. And then build in all, they're the experts, they're the experts in doing science. I don't claim to be, but I've created the conditions where we can massively accelerate their ability to build this super intelligence, and then they will build in the necessary guardrails and because the open model ecosystem exists and it's open tools and open guard railing. You create the conditions that we can create a much, you know, safer and faster and more specific level of of of intelligence for a lot of different domains. And so, you know, in that regard, we're working really hard to democratize this capability. Because if it's only closed models that exist, there's, there's not a lot of ability to engage and create the conditions that all of these applications across life sciences and materials science and all of these um amazing and important industries can thrive in the age of AI. Uh, it just wouldn't be possible. And so, so that's how we think about it. How do we create the building blocks that you can build in the necessary safety considerations and specialization so that these applications are behaving the way that they're expected to behave and put them in the hands of those subject matter experts. Now I want, I want to come to the audience in a second for a question, but I have one more for you, Jeff. Um, thank you for that, Kimberly. Um, so we've been talking about the building blocks and the foundation and just, and we've been talking about life sciences as an enormous category of things. Break it down for us. What are these new capabilities going to help us get to faster? Specifically, are we looking at solving certain types of cancers? Are we looking at certain classes of drugs that we've had difficulty cracking? What do you think is most promising? Because of these new capabilities. Yeah, what's, what's important and beautiful about this era of science is that the, the root node underneath all the root nodes that you just described is intelligence. Again, going back to the slope that is observable in the rest of the animal kingdom, science versus humans. Any extrapolation from that means that higher intelligence and science. It is going to radically improve humankind's ability to be able to do what each of us hopes a medicine will do when we step into a healthcare system, which is prevent the disease from ever happening or reverse and cure a disease that's already underway. And in fact, if you look at the last century of, of medical progress and you didn't know the order things took place, you might struggle to arrange them. In that 100 years ago humankind discovered the first antibiotics, those cure diseases every minute around the world. In close proximity to that humankind discovered vaccines. Those prevent some of the highest areas of In all of human history, morbidity and mortality from arising. Since those, the number of times we've been able to use the word cure or to prevent the category of diseases from ever happening has been um very rare. And um it's because as you wander into the labor of love that is science, it is really difficult for the human mind to conceptualize any, any fraction of the combinatorial spaces at play, like what Kimberly just described agents doing. To design an antibody or to design a small molecule is they are stepping into a combinatorial space that is atoms of the universe vast. And you know, try as we do, the human mind is not very good at navigating those sorts of terrains. And so for the same reasons that Having a model that has read everything on the internet. Is an exceptional resource for humankind, um, agents that have the ability to be able to earn the equivalent of billions or trillions of PhDs by running the scientific method around the clock and have benefited from concurrent. A cycles of the scientific method at civilization scale, right, it's going to create a kind of tool that's never existed to do what every scientist dreams of, which is sail to the very farthest edges of human knowledge and Go further. Amazing. Alright, let's go right here. Oh my goodness, we only have time for one. I will go right back here in the back, yes, with the white, your name and who you're with. Good to see you. Can you guys hear me? OK, there we go. Thanks with Nanaware from New York. Uh, I want to touch on Kimberly, your point because I totally agree with the open model need, especially when we think about identifying molecules and proteins. I totally get that. I think that's going to speed up our identification. Immensely, but when we think about actually creating the therapeutic or in my case, we create digital diagnostics and cardio metabolics. At some point we have to fix our risk management when we go to. Right? And we can do predetermined change of control, but we might be able to send updates every year, uh, once a year on post-market surveillance, for example. So how do you think about how AI, maybe this comes down to the GPU level, but I'm curious, Jeff, your thoughts as well. Traceability and design controls have to go back to the root somehow when we think about safety and efficacy, and I'm, I don't, I might be out of my depth, but I don't see how AI. How, how do we go backwards to the root when we're thinking about inputs for traceability and design controls in that regard, because then we could literally be updating FDA. On a constant basis. Yeah, I mean, briefly from a technology perspective, there's a couple ways we think about trying to tackle this problem. In, in one effort, it's how can we build, if you will, digital twin models, let's say the virtual cell, multi-scale, um, very large capable models that can represent biology and then simulate. More in in computer so that you have more, you know, efficacy, more ability to present that as simulation information to the FDA. This is, we believe is a near term future that must happen, uh, and so that's, that's one incredible thing. But I think what we can't dismiss is I'm not, we're not saying do everything in silicon. There is still a lab. There must still be a lab, but I think what you're going to be putting into the lab and that that verifier, that validation that we should speed that up much, much more quickly, narrow it down and have that one turn. I don't know that we need to trace back to what piece of information on the internet was in that model that got us there. So I think where you put the traceability. Back to should be something that was observed in the lab for the, for the, you know, and then hopefully we build this digital twin and these digital model futures that have enough fidelity, which we're just not there yet, but do eventually have enough fidelity that the FDA will also start to accept some of those measures. We must leave it there, but thank you so much, Kimberly. Jeff, thank you very much. Thank you.