How NVIDIA Turns Computing Into A New Lab Partner

Drug discovery is one of the hardest engineering problems on Earth, except it has not always been treated like engineering. Costs can hover around $2 billion per successful drug, timelines can run 10+ years, and too many patients still wait without a cure. We sit down with Rory Kelleher, who leads global business development for life sciences at NVIDIA, to talk about what changes when accelerated computing meets foundation models, generative AI, and agentic AI that can actually do work.

We break down how scientific agents differ from chatbots, and why tools matter as much as models. Rory explains NVIDIA’s BioNEMO Agent Toolkit and the idea of turning core life sciences capabilities into “agent skills” so biologists and chemists can run complex workflows through natural language. We talk protein design and protein binder design, co-folding, bioinformatics, target identification, and ADMET prediction for toxicity and safety, plus why this wave can “democratize” computational drug discovery for scientists who were never trained as programmers.

You’ll also hear a real example from Bristol Myers Squibb, where foundation models trained on proprietary sequences and compound libraries helped improve a sickle cell molecule profile until it reached first-in-human testing. We dig into what an “AI factory” looks like inside pharma, why teams want to run open models and local LLMs on secure infrastructure, and why scientific judgment becomes more important, not less, as agents increase throughput.

If you care about biotech, pharma R&D, life sciences AI, and the future of medicine, this conversation is for you. Subscribe, share the episode with a friend in research, and leave a review with the one workflow you want agents to tackle next.

Everyday AI: Your daily guide to grown with Generative AI
Can't keep up with AI? We've got you. Everyday AI helps you keep up and get ahead.

Listen on: Apple Podcasts   Spotify

Transcript

Copy Transcript Copied

Welcome And Why This Matters

SPEAKER_02 0:02

Hey everyone, really excited for this blockbuster chat today with NVIDIA on all things AI and biotech, life sciences, drug discovery, and more with you know a true innovator and leader in this space.

SPEAKER_00 0:19

I'm doing good, Evan. Thank you for hosting. Hey, Emma.

SPEAKER_02 0:23

Hi. Well done. And uh yeah, really excited uh for this chat. Yerman and I both have been following your news, of course, at NVIDIA. But this is a segment perhaps less covered by the mainstream media. So before that, maybe introduce yourself. Uh NVIDIA needs no introduction, but uh what does your team at NVIDIA actually do? What are you working on day to day?

Rory’s Role In Life Sciences

SPEAKER_00 0:48

Awesome. Yeah, once again, thanks for hosting. Uh just wonderful to be part of the part of the conversation. Love your podcast. Uh so my name is Rory Kelleher. I lead global business development for life sciences at NVIDIA. I've been here for coming up on a decade. My background has always been in enterprise technology, sales, and business development. And everything I've learned about biology and artificial intelligence and really the intersection of the two, I've learned over the past decade working with some of the most innovative and disruptive leaders in the space. Originally, that has been in what we refer to as the tech bioecosystem. And these are organizations who are reimagining what drug discovery looks like with computation, deep learning, and data at the core of what they do to try and turn the economics and the success metrics of what it takes to bring a drug to market on its head. And then increasingly over the past several years, uh, Large Pharma has begun to recognize that uh there is a more efficient, more effective way of being able to discover and develop drugs, and that is through increasingly using computational tools like foundation models and agents and uh accelerated computational tools. Um, and so our team engages with this entire ecosystem to uh understand the uh computational problems and bottlenecks and then leverage the NVIDIA computing platform from chips to systems to accelerated libraries to models to help the innovators do what their life's work is, which is uh deliver medicines to patients. Um, so that's that's what our team focuses on.

SPEAKER_01 2:24

That's

Why AI Now In Biotech

SPEAKER_01 2:25

fantastic. And you already started answering my next question, which is um why is life sciences such an exciting place to be working uh on AI projects right now? And maybe you can talk a little bit about your background before spending the last 10 years in this ecosystem, and if that informed kind of what you're doing now uh at all, um, and uh just dive into the details. Why, why now, why AI, why in this field in particular?

SPEAKER_00 2:55

Yeah, well, the the reality is that um there is still a massive number of diseases uh and patients with unmet needs, right, that that don't yet have a cure. And um I think it is the belief of many in this ecosystem that uh really we have uh a new tool in the toolbox for scientists, where uh artificial intelligence is the new instrument for how drugs will be discovered and developed. Um and frankly, this is this is needed not just because there's an incredible number of patients with uh, you know, without cures for their diseases, but also because the sheer cost and timescale of what it takes to bring a drug to market is uh is pretty astronomical, something around $2 billion per successful drug and takes of uh it takes upwards of 10 years. And so there are organizations that are trying to meaningfully compress those timelines and uh and the economics around that so that we're bringing the timelines from 10 years down to perhaps five years or less, and bringing the economics of what it takes to discover a drug uh to a fraction of what it stands today. And then uh in terms of my background and how I landed here, I mentioned my background is in enterprise technology, sales, and business development. You know, I I think um 10 years ago, my both my most progressive clients at the time were doing big data analytics in this thing called machine learning. And it was pretty early days. And I think what's what's happened over that 10-year time period is uh several waves of artificial intelligence or machine learning. We've gone from um predictive AI to generative AI, now to agentic AI. Uh, and really uh the combination of all of those different waves of artificial intelligence are starting to rear their heads in the life sciences space, um, where we're building foundation models that can understand biology or that can design molecules that have desired characteristics to be successful once introduced into a human body. Um, we are starting to see uh autonomous end-to-end systems that are starting to automate the scientific method from generating a hypothesis to designing an experiment to running that experiment in a robotic lab to analyzing the data set and there and therefore being able to drive what is the next experiment or question you want to do. And so it is it's just an incredible time to be at the intersection of artificial intelligence and life sciences. And there arguably isn't a more important, more important uh application of this technology than human health. And so it's an incredible, uh, incredible time to be participating in as he goes in the film.

SPEAKER_02 5:51

Indeed, what a time uh to be alive in general. You have a lot of solutions and products, uh, in particular, not just hardware, of course, that people might be familiar with, but tools and toolkits. Uh tell us about some of the relevant uh solutions for life sciences and biotech and give us a kind of plain in English overview of those tools.

BioNEMO Agent Toolkit Explained

SPEAKER_00 6:13

Yeah, well, one thing that maybe I'll shine a little bit of a light on is uh what we're we refer to as the Bionemo agent toolkit here in Nvidia. This is something that we launched in June. Uh, and it is important at this moment in time because of what are referred to in agentic artificial intelligence. And so if you take a step back and just look at what's possible today, large language models and agents can actually do work. They can do work for the first time. It's no longer just a chatbot that you can ask questions of, but you can give it a task or a goal, and it'll go out and access a bunch of different tools and information and actually come back with a finished product of work. Now, that work might be 80% of the way there, and a human might have to actually go do some additional work on top of it, but it has taken what could be hours upon hours worth of uh human effort and condensed that into uh now minutes. That's that's very exciting. Now, these agents become even more capable in doing work when you give them the right tools. And so BioNEMOAGENT Toolkit is a collection of models, libraries, um, and frameworks that we've turned into agent skills, right? Skills for the life sciences. And so if you think of what a biologist or a chemist might need to do in their day-to-day work, they might leverage computational tools such as co-folding or um tools for bioinformatics that might cluster cells. Uh, we've now taken all these accelerated tools, we've turned them into agent skills so that you can turn any general agent into a specialized agent that can actually do scientific tasks. And what this means is this can be a uh a peer to a scientist or an accelerant to the work that they actually do so that they can explore more ideas, they can get to answers more quickly, they can discard ideas that aren't aren't going places because they've run the experiments and learn that at such a faster rate. And so effectively what this means is we're going to start to see science start to be much more efficient, and we're going to be able to tackle more and larger, larger problems because we are able to give agents the right tools for them to actually do their

[Ad] Everyday AI: Your daily guide to grown with Generative AI

SPEAKER_00 8:29

work.

SPEAKER_01 9:08

This is so interesting. So, Rory, like uh you, I worked in different industries throughout the years, and

(Cont.) BioNEMO Agent Toolkit Explained

SPEAKER_01 9:15

of course, nothing is

Practical Workflows Scientists Run Today

SPEAKER_01 9:16

as meaningful and as exciting to me personally as uh working in healthcare and life sciences, uh, what could be um you know more important than human health that enables everything else as possible? So um give us some specific uh real-world examples. What can scientists actually do with it today? Um something can be done faster that took forever before, or are some things possible that are just were not possible at all without these tools?

SPEAKER_00 9:49

Yeah, I think this combination of agents and um accelerated computational tools for the life sciences is starting to democratize uh computational approaches to everyday biologists and everyday chemists. And so you can take a uh a biologist who was never trained in computer science and might not know much about using the latest protein design models or co-folding models or models for target identification to understand the causal biology behind the disease. And they can now interact with an agentic system that has access to these tools using natural language. And so they can say, you know, given this target, help design protein binders that might bind to what is ultimately causing this disease. And that the agentic system will then go orchestrate the workflows, spin up the computational uh tools that might be AI models, um, go run the experiments in silico, analyze the data, present the data, and then allow the scientists to figure out where to go next, right? And that that can happen across uh examples like protein binder design or uh small molecule property prediction and admet prediction for toxicity and safety. And so there's a whole host of specific use cases that are now approachable to people who were not computational. The example I'll provide is um the same thing that happened in coding, where everyday uh call it business users now are able to code in natural language using tools like Codex or Cloud Code. Everyday scientists can now access computational tools and orchestrate end-to-end workflows in ways that would have taken teams uh much longer and much larger to be able to actually deliver on. So that's that's what exciting is exciting to us is these these computational tools are just so much more approachable now.

SPEAKER_02 11:54

And this is seems very much a today's story, not a future story. I I see Ristolmeyer uh Squibb announced they're expanding their work with NVIDIA, including your agent

Bristol Myers Squibb Real Results

SPEAKER_02 12:05

toolkit. Tell us what you can about that relationship and how they're using, you know, new medicine, new, you know, AI to leverage new medicines and discoveries to get to patients faster.

SPEAKER_00 12:19

Well, we've been working with the team at Bristol Meyer Squib for greater than three years. They first purchased uh what we call an AI factory, which is um a computational system that can turn uh raw enterprise data into enterprise intelligence. Um, and over the past three-year time horizon, they were able to, through collaboration with our team and leveraging the system, prove to themselves that there is a better mousetra for how they discover and develop drugs. And so what they were using this system for is to take their proprietary sequences and cohorts and compound libraries, train foundation models that were generative to be able to uh design new molecules that had better characteristics and more likely to actually succeed in the clinic. One of the examples that the chief scientific officer, the head of RD, Robert Plange of BMS gave is a uh a sickle cell therapy where they had designed a molecule and they had basically hit a ceiling. They weren't seeing the profile or the characteristics improve to a point where they felt comfortable moving it forward into clinical trials. They turned it over to the machine learning team that had been building these protein design foundation models. They were able to improve the uh the profile of that molecule to the point where today that molecule is first inhuman. It is actually being tested in humans because it now has the profile thanks to the machine learning models, but they believe it both addresses the drug, uh, the causal uh disease uh drivers behind the drug and is going to be safe in patients. Um and uh that wasn't possible without these computational approaches before. So now they are doubling down with an investment in Nvidia's latest architecture uh called VeraRubin, and they're using it to pour some gasoline on the fire and turning their enterprise proprietary data into proprietary intelligence. And they're going to be using it uh to run open models on local data sets as well. So those those open models could be from the BioNemoagent Toolkit, where we have models for protein binder design or admet prediction or target identification. But they'll also be able to run local LLMs on their own hardware, hosted in their own environment. And they've effectively uh secured their own future in bringing in what is a new form of labor, like computation plus these agents is a new form of labor that they can now assign work to, right? And so that's what they're investing in is a platform that allows them to accelerate the work that they do and become a lot more efficient in uh in the development of drugs.

unknown 15:05

Wow.

SPEAKER_01 15:07

So interesting. So you just talked about uh Bristol Meyer Squip collaboration with um NVIDIA. So this

Open Platform For Startups Too

SPEAKER_01 15:14

is as big pharma as a big pharma can get. So I want to go back to you talking about the idea of using this technology to democratize um discoveries and innovation in the field. So is this technology and maybe a bio uh Nemo agent toolkit something that small biotechs or even startups can actually pick up and start using? Or is it really built more for uh bigger players in the farm industry?

SPEAKER_00 15:43

This is an open platform that the entire ecosystem can leverage and build on top of. And so um, you know, we make these models available on Hugging Face, we make this code available on uh on GitHub, and anyone can go and grab it and build on top of it. Um, you know, we've got a number of different partners who just point their agentic system to the GitHub and say, integrate all these tools, and then help me understand what I might be able to accelerate based off the type of work that I'm doing today. Um, and so that's what's exciting for us is like we try and build platforms to enable an entire ecosystem, whether you be a large pharma, whether you be a tools developer, like for molecular dynamics tools or for bioinformatics or cheminformatics tools that ultimately are helping the biotech ecosystem uh advance their programs forward, or you're you know a small three-person biotech who's just getting started. Uh, you know, because we've built this open platform, it's easy for folks to leverage. And because NVIDIA's computing platform is ubiquitous, it's available in every cloud through every OEM. It's even available on uh on local machines, whether it be like a DGX Spark or uh an RTX laptop. Uh, anyone has now got these incredible systems at their fingertips.

SPEAKER_02 17:05

That's amazing. I love that. Uh, particularly uh someone who tinkers and toys with AI and LLMs on my own hardware. Uh it's really exciting, but there are real-world challenges in using and deploying and monitoring and managing AI technology. What are some of the challenges you've observed in AI for drug discovery? Uh, some of the obstacles, you know, roadblocks, and

Biggest Roadblocks And Human Judgment

SPEAKER_02 17:29

how do you get them past those challenges typically?

SPEAKER_00 17:33

Yeah, I think um, as with anything, even with knowledge work, there's an element of like human taste and intuition that is still required to drive value out of these tools, right? And so, you know, a lot of people ask like, oh, uh, is uh are coding agents replacing software engineers? Or would scientific agents replace scientists? And I actually think it's going to be the inverse. I think scientific taste and scientific domain expertise, knowing which questions to ask, knowing where to drive these systems is becoming more important than ever. And if you look at in the coding world, uh, we actually have a demand for more coders than ever before because it takes an engineering mindset to be able to think about how to use these systems and drive, you know, 10 times more throughput than a team was capable of doing before, right? And I think the same is going to happen in science in science as well, um, where uh Ashley McGargee, the the um CEO of Genentech, said that she's found because of these agentic systems, they need more scientists than they've had because they have more ideas that they can now pursue. As we talked about at the top of the at the top of the call here, there's still an immense number of diseases that have no cures. And how wonderful would it be if we can make one scientist as effective as 10 or 50 as 100? And like that is that is the incredibly exciting stage that we're in right now, is not that we're going to be reducing the number of scientists that the world needs. It's that we've made every scientist more efficient and effective so that they can go pursue more problems and ultimately start to bend the curve of disease on human health.

SPEAKER_01 19:20

Yeah, so in healthcare, of course, uh the conversation sometimes goes to AI, replacing doctors. But no, it's the doctors with AI will become better doctors because they'll be able to focus more on patients. Um the empathy can come through again if AI takes some of the you know mundane tasks off of their hands. And I love your um phrasing of uh this technology can create a new form of labor as opposed to reducing the labor pool. So as you look ahead, what should people in the wider life sciences, pharma, biotech ecosystem keep an eye out uh

What To Watch Next In Models

SPEAKER_01 19:56

for next?

SPEAKER_00 19:58

Yeah, I think I think one thing is that um to the point you just made, you're you you're not going to be replaced by AI. You're going to be replaced by somebody who uses AI. Therefore, everyone should be engaging this technology. Tinker with it, play with it, understand it, see what's possible, see how it might augment the work that you're doing today or give you the path capacity to do far more than you've been doing today. Um, and I think what we're going to see is that uh foundation models that once again can be orchestrated by agents or or large large language models are becoming more and more capable. And so I'll give you a few examples. A year ago, some of the leading molecular design foundation models uh for uh for antibodies would yield, say, a 1% hit rate. In one year's time, we're now seeing a 30 to 40% hit rate in one year's time, right? And so if that trend continues, we're going to see extremely capable models that can design molecules with desired characteristics. Now, building a molecule is only one stage of drug discovery. Another is really understanding what is driving the disease. And so we've got many partners in this ecosystem who are setting up industrial scale biological data generation to be able to train foundation models for uh target identification and validation. And this is with um, you know, single-cell biology and uh digital pathology and all these different biological data modalities. And if we can better predict what is actually causing the disease, then we have a better likelihood of being successful in the clinic because we can actually design trials with people who are likely to respond to the molecule that we've actually developed. And so I think what people should look out for is that um we're gonna start to see these technologies and the combination of these technologies begin to compound and build on top of it. And it's gonna follow a similar path to how we now see um chip design take place. You know, 40 years ago, the way that people used to design chips was on drafting board. Now you can design a full chip end-to-end in silico, and when you push a button, you know on day one that chip is going to work. And I think that's the exciting future that we have in drug discovery and medicine is that simulation and agentic systems and foundation models are going to get us to a point eventually where designing drugs becomes an engineering discipline, not just a uh a guessing game, right? And so that's that's what's super exciting, Irma.

SPEAKER_02 22:48

Yeah, really amazing to hear. We're of course uh in Boston, which we like to think of as ground zero for biotech and and life sciences innovation.

Events Ahead And Final Thanks

SPEAKER_02 22:59

Um, and there's a lot going on around town this fall. What are you up to? Any any insight into your travel and customer events? And I'm sure you'd be on the road uh quite a bit here and beyond.

SPEAKER_00 23:09

Yeah, we've got a we've got a few exciting things coming up. We've got um GTC Berlin that's taking place uh in mid-October. And so we'll be engaging with the European ecosystem, as you know, uh, for the top 10 leading pharmaceutical companies, call Europe home. And there's an incredible ecosystem of um leading researchers that are building companies uh in the drug discovery and life sciences space. And so we'll be engaging with that ecosystem. Uh JP Morgan is not too far off around the corner. That is the you know, the the mecca, if you will, of the entire healthcare ecosystem. Uh, and then we'll we'll likely also be showing up in health because um, you know, digital health is an extremely fast-growing segment that we're engaging with. Think of these as agents who are transforming the physician and patient interaction or the patient interaction with the healthcare system, or helping to um, you know, address the uh claims denial problem that many people have experienced over time. And so, like this is a whole ecosystem that too is going to be transformed because of agentic AI and accelerated tools. And so that's just a quick glimpse into some of the milestones that we see coming up over the over the next few months here.

SPEAKER_02 24:28

Well, I can't wait to watch it all unfold. And thank you so much for sharing your time and insight into the incredible innovation happening at NVIDIA and beyond. Appreciate your time.

SPEAKER_00 24:39

Thank you, Evan.

SPEAKER_02 24:40

Thank you very much.

SPEAKER_01 24:41

And thank you for everyone watching and listening and sharing and participating in these conversations. We'll see you.

SPEAKER_02 24:47

Yes, definitely reach out for us to come on the show. Also check out our TV show at techimpact.tv, now on Fox Business and Bloomberg Television Monthly. Thanks, everyone. Thanks, Rory.

SPEAKER_00 24:57

Cheers, guys. Thank you.