# Turning AI Into the Infrastructure Advancing Human Health

## Abstract

Key takeaways Learn how digital health agents are reaching national scale and giving providers more time with patients. Explore how AI factories are accelerating target discovery and molecular design to shorten the path from research to clinic. Discover how robotic systems are scaling the reach of surgical care and medicines that must reach billions.

## AI Summary

- The presentation highlighted the integration of agentic AI and physical AI in healthcare, emphasizing how AI can assist in various aspects of patient care, from robotic-assisted surgery to mental health applications.
- The speaker discussed the challenges of a global shortage of healthcare professionals and how AI can help by automating administrative tasks and providing continuous patient support.
- A new platform, Isaac for Healthcare, was introduced to help train and deploy robotic systems in hospitals, including autonomous sample transport and robotic-assisted surgery.
- The BioNeMo Agent Toolkit was showcased for its role in accelerating drug discovery and lab automation, allowing scientists to design and test molecular candidates more efficiently.
- The RFD4 Proteina model, co-developed with the University of Washington Institute of Protein Design and the David Baker Lab, was announced as a significant advancement in protein design, capable of generating highly effective designs in a fraction of the time.
- The presentation concluded with a call to action for the audience to participate in the development and deployment of these AI tools to improve healthcare and scientific research.

## Transcript

**[00:00:04 – 00:00:10]** So, we keep you in suspense, as you know, but here we go.

**[00:00:10 – 00:00:17]** Sometimes, progress begins at home, with one steadier movement,

**[00:00:17 – 00:00:24]** a little more confidence, and care that meets you right where you are.

**[00:00:24 – 00:00:27]** It should listen in your language,

**[00:00:28 – 00:00:31]** understand what you need, and be able to connect you

**[00:00:31 – 00:00:34]** and your story to what's next.

**[00:00:34 – 00:00:39]** Don't worry, I'm going to be here the whole time, so your

**[00:00:39 – 00:00:42]** doctor can look up and be present, and support can move beyond

**[00:00:42 – 00:00:44]** the screen to precise action.

**[00:00:44 – 00:00:48]** The operating room itself can then help teams prepare

**[00:00:58 – 00:01:03]** While robotic instruments extend a surgeon's skill

**[00:01:04 – 00:01:07]** Supplies arrive where they're needed

**[00:01:10 – 00:01:16]** Laboratories turn samples into answers, and every connection

**[00:01:16 – 00:01:21]** builds a sharper picture that informs the next step, connecting

**[00:01:21 – 00:01:24]** digital health and physical A.I.

**[00:01:24 – 00:01:26]** To turn intelligence into action.

**[00:01:27 – 00:01:31]** Together, we're building the future of care, so that recovery

**[00:01:31 – 00:01:34]** can lead back to what matters.

**[00:01:34 – 00:01:41]** Another step, another adventure, another chance to move forward.

**[00:01:42 – 00:01:46]** And with that, let us all please welcome NVIDIA's Health

**[00:01:46 – 00:01:48]** Care and Life Sciences Vice President, Kimberley Powell.

**[00:01:48 – 00:01:53]** Thank you, Konstantin.

**[00:01:54 – 00:01:56]** Doesn't that give you the chills?

**[00:01:56 – 00:01:58]** It's pretty amazing.

**[00:01:58 – 00:02:00]** It makes you feel something deep inside.

**[00:02:00 – 00:02:05]** I have to say those micro duck robots that Jensen

**[00:02:05 – 00:02:08]** showed us earlier were incredibly entertaining.

**[00:02:08 – 00:02:11]** We're gonna be entertaining here, but there's gonna be an

**[00:02:11 – 00:02:15]** insurmountable amount of hope that all of you have helped design into

**[00:02:15 – 00:02:19]** what is gonna be this future of healthcare and future of medicine.

**[00:02:20 – 00:02:23]** And so, together, we can really start tackling what

**[00:02:23 – 00:02:27]** is a global challenge.

**[00:02:27 – 00:02:32]** We simply do not have enough healthcare professionals to

**[00:02:32 – 00:02:37]** serve the deep demand of healthcare services we need even today,

**[00:02:37 – 00:02:41]** never mind the severity of what's in front of us in the future.

**[00:02:41 – 00:02:45]** But let's talk about how incredibly exciting it is,

**[00:02:45 – 00:02:49]** and the momentum is super strong.

**[00:02:49 – 00:02:53]** Check out this survey done by Bessemer and Bain.

**[00:02:53 – 00:02:58]** It's showing here that healthcare industry is taking advantage

**[00:02:58 – 00:03:02]** of these incredibly fast innovation cycles.

**[00:03:02 – 00:03:07]** We went from chat that could kind of go questions and answers,

**[00:03:07 – 00:03:11]** and this was helpful AI, this was kind of giving you more

**[00:03:11 – 00:03:14]** information, maybe helping you learn a little bit.

**[00:03:14 – 00:03:19]** But it quickly moved into how can we leverage Modern approaches

**[00:03:19 – 00:03:24]** where all of a sudden AIs could use tools to go off and see, could we

**[00:03:24 – 00:03:26]** get some more workflow automation?

**[00:03:26 – 00:03:30]** And these are all helpful things, helpful things for the healthcare

**[00:03:30 – 00:03:32]** industry, but not quite enough.

**[00:03:32 – 00:03:35]** We are rapidly moving into this era where we

**[00:03:35 – 00:03:38]** are deploying agentic systems.

**[00:03:38 – 00:03:41]** This is AI that can do work.

**[00:03:41 – 00:03:44]** It can connect one step after the other.

**[00:03:44 – 00:03:48]** It's tackling problems all across the healthcare industry, from

**[00:03:49 – 00:03:53]** providers to payers, all the way through the life sciences industry.

**[00:03:53 – 00:03:58]** And this industry really needs us to do incredibly useful work.

**[00:03:58 – 00:04:01]** And that is what 2026 has brought us.

**[00:04:01 – 00:04:03]** It's no longer about a model.

**[00:04:03 – 00:04:05]** Jensen talked a lot about that today.

**[00:04:05 – 00:04:08]** It's about how we surround this model, the tools we give

**[00:04:08 – 00:04:12]** it, the knowledge we give it, the governance that we

**[00:04:12 – 00:04:16]** provide it, and the safety and the runtime so it can

**[00:04:16 – 00:04:19]** operate at the speed of healthcare.

**[00:04:19 – 00:04:25]** The other amazing aspect of this survey is the return on investment.

**[00:04:25 – 00:04:29]** If we think about digital systems of the past being

**[00:04:29 – 00:04:34]** deployed in healthcare, the return on that investment is typically

**[00:04:34 – 00:04:38]** measured on a two to three year, sometimes five or six year cycle.

**[00:04:38 – 00:04:41]** These were complicated systems.

**[00:04:41 – 00:04:45]** Because we're living in the world of agentic AI, and these

**[00:04:45 – 00:04:50]** agents can be accessible by voice, right on an everyday device

**[00:04:50 – 00:04:55]** like your phone, or on a tablet, or just in any computer that

**[00:04:55 – 00:05:01]** you use on every single day, the barrier to adopt is instantaneous.

**[00:05:01 – 00:05:05]** There's no, where do I click, what steps do I take?

**[00:05:05 – 00:05:09]** You are simply starting to operate in the most human

**[00:05:09 – 00:05:12]** way with a computer, which is via natural language.

**[00:05:13 – 00:05:17]** And because of that, the otherwise return on investment, which

**[00:05:17 – 00:05:21]** a lot of the healthcare industry thinks, it'll be two to three

**[00:05:21 – 00:05:27]** years, they're actually realizing that ROI within a year, and

**[00:05:27 – 00:05:31]** not only that ROI within a year, but it's three times the ROI.

**[00:05:31 – 00:05:36]** So every dollar spent, there's$ 3 of value coming in.

**[00:05:36 – 00:05:37]** This is extraordinary.

**[00:05:37 – 00:05:41]** The health care industry has never seen these kind of results before.

**[00:05:41 – 00:05:44]** And that's what brings us to such an incredible time.

**[00:05:44 – 00:05:48]** Now, just to remember, what is AI?

**[00:05:48 – 00:05:53]** It's not a model, it's now about agents, agents helping us do

**[00:05:53 – 00:05:58]** useful work, extend the time that humans have to do the incredible

**[00:05:58 – 00:06:00]** work that they'd love to do.

**[00:06:00 – 00:06:04]** And so we take this model, of course that's one of the

**[00:06:04 – 00:06:07]** elements of the system, but we provide it now a harness.

**[00:06:07 – 00:06:10]** I think of it as the body, the harness, and then we give

**[00:06:10 – 00:06:14]** it arms and legs and eyes and all of these things so

**[00:06:14 – 00:06:18]** it can be essentially a companion to us, a co-worker with us,

**[00:06:18 – 00:06:20]** a thought partner, a lab assistant.

**[00:06:20 – 00:06:24]** And we're working on so much technology because there is

**[00:06:24 – 00:06:28]** the need that the specialization, the exquisite work that's

**[00:06:28 – 00:06:33]** done in the healthcare industry, requires a lot of specialization.

**[00:06:34 – 00:06:39]** And so we work on how do we extend this platform in ways that can

**[00:06:39 – 00:06:46]** unlock this insane potential of, let's call it, two decades at least

**[00:06:46 – 00:06:48]** of digitizing healthcare data.

**[00:06:48 – 00:06:52]** Europe is extremely rich with digital data.

**[00:06:52 – 00:06:54]** We are in the United States as well.

**[00:06:54 – 00:06:59]** But you need to be able to give access to that data in

**[00:06:59 – 00:07:03]** a way and make sense of that data, use it as a knowledge foundation

**[00:07:03 – 00:07:08]** so you can ground these otherwise generalist AI with the deep

**[00:07:08 – 00:07:11]** understanding of what transpires in the healthcare industry.

**[00:07:11 – 00:07:16]** So Jensen talked about the NVIDIA Agent Knowledge Foundation.

**[00:07:16 – 00:07:21]** These are tools That can sit within the data platforms that are right

**[00:07:21 – 00:07:25]** now being kind of re-architected and rebuilt to become AI ready.

**[00:07:25 – 00:07:29]** We must make this otherwise dark and not useful healthcare

**[00:07:29 – 00:07:31]** data; we need to make it AI ready.

**[00:07:31 – 00:07:34]** We need to empower our agents with it.

**[00:07:34 – 00:07:38]** But we need to do it with security and with safety in mind.

**[00:07:38 – 00:07:42]** And so, also Jensen introduced the incredible platform we've

**[00:07:42 – 00:07:43]** built with OpenShell.

**[00:07:43 – 00:07:49]** This is a secure runtime and sandbox for agents to operate,

**[00:07:49 – 00:07:52]** where, just like when you hire a new employee, or maybe

**[00:07:52 – 00:07:57]** you have a child you're trying to expose to new things, you sandbox

**[00:07:57 – 00:07:59]** their ability to access things.

**[00:07:59 – 00:08:01]** I have teenage daughters.

**[00:08:01 – 00:08:05]** I don't love cell phones or social media, so I sandbox them

**[00:08:05 – 00:08:08]** on what they can and can't access.

**[00:08:08 – 00:08:11]** This is a no different phenomenon here, right?

**[00:08:11 – 00:08:14]** So you may start your agents with just a small amount of

**[00:08:14 – 00:08:19]** very restricted access to start doing that work, and as trust goes,

**[00:08:19 – 00:08:21]** you can start to open that up.

**[00:08:21 – 00:08:26]** Or you can design specialized agents who talk to each

**[00:08:26 – 00:08:28]** other and pass along.

**[00:08:28 – 00:08:33]** There's a very common term, standard operating procedures, that

**[00:08:33 – 00:08:35]** is all throughout this industry.

**[00:08:35 – 00:08:39]** Well, imagine that each operating procedure could be its own

**[00:08:39 – 00:08:41]** very deeply specialized agent.

**[00:08:41 – 00:08:45]** And to do that, we think a lot about our strategy

**[00:08:45 – 00:08:47]** around open models.

**[00:08:47 – 00:08:51]** NVIDIA has been investing for the last many years in

**[00:08:51 – 00:08:57]** creating an open model ecosystem At the frontier or the frontier, if

**[00:08:57 – 00:09:02]** you heard Jensen, that nuance where we are building open source models

**[00:09:02 – 00:09:07]** that are at the frontier such that you can take and customize them

**[00:09:07 – 00:09:11]** for the deep specialization that we need for this industry, and then,

**[00:09:11 – 00:09:15]** in other cases, we're completely creating the frontier, and

**[00:09:15 – 00:09:20]** these are going to be incredible tools for the healthcare industry.

**[00:09:20 – 00:09:23]** Let's just take a review of some of them that I'm gonna show you today.

**[00:09:23 – 00:09:29]** The opening video said two phrases, agentic AI meets physical AI.

**[00:09:30 – 00:09:33]** Healthcare is an extremely physical industry.

**[00:09:33 – 00:09:34]** At the end of the day, I'm here, right?

**[00:09:34 – 00:09:38]** I'm not an avatar, I'm not a pixel floating in the air.

**[00:09:38 – 00:09:42]** We need to do things in the physical world to achieve what we'd

**[00:09:42 – 00:09:44]** like to do to live healthier lives.

**[00:09:44 – 00:09:50]** So, Nemotron is a family of open language models and vision language

**[00:09:51 – 00:09:56]** models that allow for us to take what is otherwise generalist.

**[00:09:56 – 00:10:01]** So we're going to have a lot of examples today about how, with

**[00:10:01 – 00:10:06]** a small amount of post-training, we can get exquisite performance

**[00:10:06 – 00:10:12]** and essentially surpass the frontier on specialized tasks.

**[00:10:12 – 00:10:17]** There's other places where, again, I will say we are the frontier.

**[00:10:17 – 00:10:22]** In the areas of Cosmos, these world foundation models, these

**[00:10:22 – 00:10:27]** world and action models with GRUT, two things that we need to

**[00:10:27 – 00:10:31]** start establishing the conditions, whether that's generating synthetic

**[00:10:31 – 00:10:35]** data, generating synthetic worlds, and then training, essentially,

**[00:10:35 – 00:10:38]** the brain of the robot, with all the policies of how they're

**[00:10:38 – 00:10:42]** going to operate in the real world, those models with Groot and

**[00:10:42 – 00:10:45]** Cosmos are completely the frontier.

**[00:10:45 – 00:10:47]** And then we also have a wonderful family, and we're going to get

**[00:10:47 – 00:10:52]** it towards the end, which is really around biology foundation models.

**[00:10:52 – 00:10:57]** In drug discovery, it's a very arduous process.

**[00:10:57 – 00:11:00]** It's a very long process, but we are now starting to be

**[00:11:00 – 00:11:04]** able to model the world of drugs in a computer and do things at scales

**[00:11:04 – 00:11:07]** that we could never do before.

**[00:11:07 – 00:11:12]** But let's think about how agents are truly starting to

**[00:11:12 – 00:11:14]** join the workforce in healthcare.

**[00:11:14 – 00:11:17]** And as you know, we're all patients at the end of the

**[00:11:17 – 00:11:22]** day, and many of you are innovators along these steps of the journey,

**[00:11:22 – 00:11:26]** perhaps all along them, or perhaps in discrete areas.

**[00:11:26 – 00:11:32]** But you'll see that, and for there to be a true transformation,

**[00:11:32 – 00:11:38]** we want AI to be infused at every single step of this process.

**[00:11:38 – 00:11:42]** And in fact, it is, and there's a ton of that innovation happening

**[00:11:42 – 00:11:46]** here all across Europe.

**[00:11:46 – 00:11:49]** We're working with Sword Health, I think you saw them in the

**[00:11:49 – 00:11:53]** beginning, where they have AI companions that really

**[00:11:53 – 00:11:57]** help you stay connected to the health services you need

**[00:11:57 – 00:11:59]** in between appointments.

**[00:11:59 – 00:12:02]** That could be for physical therapy; that could be for

**[00:12:02 – 00:12:04]** mental health applications.

**[00:12:04 – 00:12:07]** And that's an incredible thing for a patient to have access

**[00:12:07 – 00:12:11]** to healthcare services when your doctors can't be readily

**[00:12:11 – 00:12:15]** available but can still continue on your healing journey.

**[00:12:15 – 00:12:20]** One advanced is helping to essentially help patients

**[00:12:20 – 00:12:24]** navigate the system, triage patients, schedule patients,

**[00:12:24 – 00:12:26]** all across the UK.

**[00:12:26 – 00:12:29]** Heidi is giving healthcare professionals, again, this

**[00:12:29 – 00:12:35]** very natural language ability to take language, transcribe it,

**[00:12:35 – 00:12:39]** marry it up to incredible evidence about the patient's history or

**[00:12:39 – 00:12:44]** medical evidence so that they can enhance clinical decision support.

**[00:12:44 – 00:12:46]** And so your doctor can look up, right?

**[00:12:46 – 00:12:49]** They're not typing anymore in the room when you're with them.

**[00:12:49 – 00:12:51]** And that's an incredible transformation.

**[00:12:51 – 00:12:56]** And I would say on behalf of both entities in the room,

**[00:12:56 – 00:12:59]** the doctor wants to look you in the eye and the patient

**[00:12:59 – 00:13:01]** wants to be looked in the eye.

**[00:13:01 – 00:13:04]** And this is what being able to use voice and language models

**[00:13:04 – 00:13:07]** and agentic systems is enabling.

**[00:13:07 – 00:13:13]** Hologen is incredibly taking all of the modalities of healthcare

**[00:13:13 – 00:13:19]** data, all these insane signals that we can and have been capturing, but

**[00:13:19 – 00:13:20]** we sort of look at them discreetly.

**[00:13:20 – 00:13:24]** What if, .. what if, .. what if .. we could make

**[00:13:24 – 00:13:26]** a large medicine model?

**[00:13:26 – 00:13:28]** That's their term; they coined it.

**[00:13:28 – 00:13:30]** And why would you do that? Because this is now a model

**[00:13:30 – 00:13:33]** that's really speaking all the signals of what we capture

**[00:13:34 – 00:13:35]** throughout the patient journey.

**[00:13:35 – 00:13:38]** So we can start predicting outcomes for the patient,

**[00:13:38 – 00:13:42]** helpful to the providers around the world, gonna be extremely

**[00:13:42 – 00:13:46]** helpful for clinical trials in the life sciences industry.

**[00:13:46 – 00:13:51]** And Dawn Health, just recently announcing with Merck, how

**[00:13:51 – 00:13:56]** do you help, again, connect patients who are on these therapies

**[00:13:56 – 00:14:02]** to a good experience, being able to understand how are their symptoms

**[00:14:02 – 00:14:05]** and keeping them on regimen and all of that deep connection?

**[00:14:05 – 00:14:07]** Wouldn't we all love that?

**[00:14:07 – 00:14:10]** There's some incredible medicines coming out, but we want the

**[00:14:10 – 00:14:13]** support while we're taking them to know, is this normal or not?

**[00:14:13 – 00:14:16]** And so Dawn Health is helping do that with the

**[00:14:16 – 00:14:17]** pharmaceutical industry.

**[00:14:17 – 00:14:20]** So you can see we started at the beginning to the end

**[00:14:20 – 00:14:25]** and everywhere in between, and this is the phenomenal future that

**[00:14:25 – 00:14:27]** we're all starting to experience.

**[00:14:27 – 00:14:31]** And it is, as Constantine put it, it is lighting up Europe.

**[00:14:31 – 00:14:35]** You know, there's half a billion people that have access now to

**[00:14:35 – 00:14:40]** Daedalus and how they're infusing AI throughout their entire system.

**[00:14:40 – 00:14:44]** The Avelios is an AI-native electronic health record

**[00:14:44 – 00:14:47]** system, just thinking completely differently about how we

**[00:14:47 – 00:14:51]** could infuse AI throughout the whole process.

**[00:14:51 – 00:14:55]** And CoreT, creating all of the medical AI and clinical

**[00:14:55 – 00:14:56]** AI that can be built into, they're built into hundreds

**[00:14:56 – 00:14:58]** of applications now, because it's just a matter of time before we get

**[00:14:58 – 00:15:01]** to the point where we're going to be able to do that; every single,

**[00:15:01 – 00:15:06]** again, part of the patient journey and the healthcare process, we

**[00:15:06 – 00:15:11]** can utilize the ability for voice and language models and connecting

**[00:15:11 – 00:15:16]** it with data and infusing that into incredibly useful work.

**[00:15:16 – 00:15:17]** So super, super exciting.

**[00:15:17 – 00:15:21]** Now, we need to make sure that all of that awesome work

**[00:15:21 – 00:15:26]** we do in the digital world can also find its way to driving

**[00:15:26 – 00:15:28]** solutions in the physical world.

**[00:15:29 – 00:15:31]** As I said, the World Health Organization projects that

**[00:15:32 – 00:15:35]** we're going to be something like 10 million healthcare

**[00:15:35 – 00:15:39]** workers short in a few short years.

**[00:15:40 – 00:15:43]** So if we're going to meet that demand, we have to meet

**[00:15:43 – 00:15:44]** it in the physical world.

**[00:15:44 – 00:15:49]** It is the case that a lot of our healthcare professionals spend

**[00:15:49 – 00:15:54]** 30, sometimes upwards of 50 percent of their time doing administrative

**[00:15:54 – 00:15:57]** type work, non-clinical work.

**[00:15:57 – 00:16:02]** And so we need to find a way that we could imagine robots, you're

**[00:16:02 – 00:16:05]** going to see a lot of them walking around here, could be delivering

**[00:16:05 – 00:16:08]** things throughout the hospital so we're not detracting nurses

**[00:16:08 – 00:16:10]** from the patients that need them.

**[00:16:10 – 00:16:15]** We, of course, want to see robots and systems that are living

**[00:16:15 – 00:16:18]** inside patient rooms, providing information and experiences.

**[00:16:18 – 00:16:22]** And absolutely, we have folks who are pioneers in

**[00:16:22 – 00:16:26]** the robotic-assisted surgery space, and that is going to

**[00:16:26 – 00:16:30]** be a place of absolute innovation.

**[00:16:30 – 00:16:34]** And so we need to say, how do we teach a machine to act in this

**[00:16:34 – 00:16:37]** real world as it keeps changing?

**[00:16:37 – 00:16:42]** There's a lot of software and infrastructure that we

**[00:16:42 – 00:16:44]** have to make available.

**[00:16:44 – 00:16:49]** Over the last couple of years, NVIDIA has been taking the

**[00:16:49 – 00:16:53]** incredible breakthroughs that we've helped pioneer in the

**[00:16:53 – 00:16:55]** automotive industry.

**[00:16:55 – 00:16:58]** Being able to drive a car in the real world has a lot

**[00:16:58 – 00:17:05]** of technical parallels with being able to autonomously drive machines

**[00:17:05 – 00:17:07]** inside of a surgical suite.

**[00:17:07 – 00:17:11]** From a technological platform perspective, incredibly parallel.

**[00:17:11 – 00:17:15]** So we've been working on a platform we call Isaac for Healthcare.

**[00:17:15 – 00:17:22]** That really helps us take and build on the incredible libraries

**[00:17:22 – 00:17:27]** across NVIDIA, domain adapt them for the workflows of physical AI.

**[00:17:28 – 00:17:32]** So, Apian is pioneering, and they're advancing autonomous

**[00:17:32 – 00:17:35]** sample transport, if you saw that robot going down the

**[00:17:35 – 00:17:39]** stairs and delivering samples for lab teams, really building

**[00:17:39 – 00:17:44]** digital twins using NVIDIA Omniverse and Isaac for Healthcare.

**[00:17:44 – 00:17:46]** And you can begin with data and simulation.

**[00:17:47 – 00:17:50]** You start with some real world data, and then you can use

**[00:17:50 – 00:17:54]** world models like Cosmos to enhance that digital data.

**[00:17:54 – 00:17:56]** Oops.

**[00:17:56 – 00:17:57]** I guess it was already on there.

**[00:17:57 – 00:17:58]** Can I go back?

**[00:17:58 – 00:18:02]** And enhance that digital data to create a whole bunch of

**[00:18:02 – 00:18:05]** different scenarios and a whole bunch of different corner cases

**[00:18:05 – 00:18:08]** that you wouldn't want to have to first experience in the real world.

**[00:18:08 – 00:18:10]** You could do it in simulation first.

**[00:18:10 – 00:18:13]** It becomes an enhanced training data set, and it goes in to

**[00:18:13 – 00:18:16]** do reinforcement learning and post-training, so that

**[00:18:16 – 00:18:20]** when you do get to the real world, you have that ability with a much,

**[00:18:20 – 00:18:26]** much more robust model to operate in incredibly complex environments.

**[00:18:26 – 00:18:30]** And so they use Isaac and Groot to train the policies after

**[00:18:30 – 00:18:32]** they created these digital twins.

**[00:18:32 – 00:18:35]** And then you have a real-time computer that's inside the

**[00:18:35 – 00:18:40]** robot, able to perceive, sense, reason, and then act.

**[00:18:40 – 00:18:42]** It's a similar situation.

**[00:18:42 – 00:18:45]** So if we can solve it for healthcare, hospital

**[00:18:46 – 00:18:49]** automation-type activities, we can start to translate that in.

**[00:18:49 – 00:18:54]** However, we're working on a lot more domain-specific tools.

**[00:18:54 – 00:18:59]** So as part of Isaac for Healthcare, we've introduced medical physics

**[00:18:59 – 00:19:02]** simulation, because the body, once you go inside of it,

**[00:19:02 – 00:19:06]** physics really gets in the way, and things like you start to deform,

**[00:19:06 – 00:19:11]** and when tools hit different anatomy, you experience different

**[00:19:11 – 00:19:14]** things, and the robots are going to have to learn to deeply adapt

**[00:19:14 – 00:19:17]** to that in incredible real-time.

**[00:19:17 – 00:19:20]** And so, similar situation, once you have those simulated

**[00:19:20 – 00:19:24]** environments, you could essentially build a training data set

**[00:19:24 – 00:19:28]** and then simulate that in a simulator here and post-train

**[00:19:28 – 00:19:32]** these policy models so that they have much more capabilities

**[00:19:32 – 00:19:34]** as they go into deployment.

**[00:19:34 – 00:19:38]** So we're working with an incredible ecosystem here, not only in

**[00:19:38 – 00:19:44]** hospital automation, but also in clinical and medical robots.

**[00:19:45 – 00:19:48]** And so deploying robots, there's a third part of this here,

**[00:19:48 – 00:19:51]** so we talked about data, we talked about simulation, is how do

**[00:19:51 – 00:19:53]** you actually deploy these things?

**[00:19:53 – 00:19:56]** And the world of medical devices has typically, and frankly

**[00:19:56 – 00:20:01]** robots of all kinds, have typically architected their systems.

**[00:20:01 – 00:20:06]** If you go and work with our friends who are typically OEMs of medical

**[00:20:06 – 00:20:10]** systems, they have different BUs, whether they're a CTBU or

**[00:20:10 – 00:20:15]** ultrasound business unit or an MRI, and each one of them tends to go

**[00:20:15 – 00:20:17]** off and redesign that top system.

**[00:20:17 – 00:20:20]** They pick their own CPUs, their FPGA combination, their

**[00:20:20 – 00:20:23]** GPU combination, and then they have to get to the hard

**[00:20:23 – 00:20:26]** part, is how do I architect it, and how do I write the

**[00:20:26 – 00:20:30]** software to run that in a latency budget that would actually allow

**[00:20:30 – 00:20:33]** for us to deploy real-time systems?

**[00:20:33 – 00:20:36]** And so here's an example, kind of a blooper, one of

**[00:20:36 – 00:20:42]** our amazing product managers, where he took and he ran, he trained

**[00:20:42 – 00:20:44]** a robot policy in simulation.

**[00:20:44 – 00:20:47]** And in simulation, it actually worked fine.

**[00:20:47 – 00:20:49]** This is called gravity compensation.

**[00:20:49 – 00:20:52]** You saw that the micro ducts, they had incredible gravity

**[00:20:52 – 00:20:54]** compensation and could get back up.

**[00:20:54 – 00:20:57]** If you don't know how to compensate for gravity, you know, look

**[00:20:57 – 00:20:58]** at what's happening.

**[00:20:58 – 00:21:01]** And so why is this happening?

**[00:21:01 – 00:21:03]** In simulation, everything looks perfect.

**[00:21:03 – 00:21:06]** When you go try to deploy it onto that real hardware

**[00:21:06 – 00:21:09]** system, you're finding all this latency because you're moving

**[00:21:09 – 00:21:14]** data, you're going from a sensor to an action, and in between, you're

**[00:21:14 – 00:21:17]** moving things with an unoptimal path across that otherwise

**[00:21:17 – 00:21:22]** what I would call old-school, off-the-shelf architecture.

**[00:21:23 – 00:21:26]** Jensen uses a word called extreme co-design, and we

**[00:21:26 – 00:21:29]** do that in our data centers, and you saw what we introduced

**[00:21:29 – 00:21:35]** with our AI factories this morning; it's no different for edge compute.

**[00:21:35 – 00:21:38]** We have to think about it like it's a data center living

**[00:21:38 – 00:21:43]** in the robot because you have to be extremely efficient of sensing

**[00:21:43 – 00:21:46]** something in the real world and taking action in the real world.

**[00:21:46 – 00:21:48]** Any latency in between there, you're going to have things

**[00:21:48 – 00:21:53]** like falling down, and so, with the ability of a real-time

**[00:21:53 – 00:21:56]** runtime, which we've been working on, and the extreme

**[00:21:56 – 00:22:00]** co-design of our IGX, where you have a completely integrated

**[00:22:00 – 00:22:06]** IGX store with all of our Holoscan sensor bridge, the CPU, the

**[00:22:06 – 00:22:10]** GPU, and the high-speed networking, and a standardized networking

**[00:22:10 – 00:22:13]** NVIDIA, NVIDIA Holoscan, AI, factories, to get information in

**[00:22:13 – 00:22:17]** and out, you have a much different story, a much different experience.

**[00:22:17 – 00:22:23]** And so what we've been able to do is take extreme co-design,

**[00:22:23 – 00:22:28]** build a real-time runtime that we call NVIDIA Holoscan,

**[00:22:28 – 00:22:32]** and today we are announcing Holoscan 5, which takes advantage

**[00:22:32 – 00:22:35]** of that extreme co-design.

**[00:22:35 – 00:22:39]** So, we look at the IGX through our architecture, and we've created

**[00:22:39 – 00:22:47]** this real-time runtime that allows you to go from sensor, Processing

**[00:22:47 – 00:22:52]** into all the AI inferencing that we need into the actuation.

**[00:22:52 – 00:22:58]** And so Holoscan Sensor Bridge is an incredible innovation

**[00:22:58 – 00:23:03]** to essentially make that IO insanely fast and standardized, so

**[00:23:03 – 00:23:06]** you're not going through all these different interfaces, so everything

**[00:23:06 – 00:23:09]** can kind of speak one language and at very, very high rates.

**[00:23:09 – 00:23:13]** And so this unified runtime, Holoscan traction has been

**[00:23:13 – 00:23:16]** incredible, almost 2 million downloads, and it's, a lot

**[00:23:16 – 00:23:21]** of that has even just come in the last, you know, five to six months.

**[00:23:21 – 00:23:24]** Holoscan was originated out of the healthcare business unit

**[00:23:25 – 00:23:27]** because it's a really hard problem.

**[00:23:27 – 00:23:29]** We go for the hardest of problems, why?

**[00:23:29 – 00:23:31]** Because it's gonna have applicability elsewhere.

**[00:23:31 – 00:23:36]** And now Holoscan is being deployed in all sorts of industries, high

**[00:23:36 – 00:23:40]** performance computing industries, into general robotics, into

**[00:23:40 – 00:23:47]** manufacturing, as well as being adopted into medical applications.

**[00:23:47 – 00:23:51]** And so the Holoscan ecosystem is extremely rich, as I said,

**[00:23:51 – 00:23:55]** into industrial robotics, into medical robotics.

**[00:23:55 – 00:23:59]** But you see that the sensor and being thoughtful about

**[00:23:59 – 00:24:04]** how sensors and sensors interact with that computer is really,

**[00:24:04 – 00:24:05]** really important.

**[00:24:05 – 00:24:08]** So the sensor bridge and all the actuators, that's essentially

**[00:24:08 – 00:24:11]** the eyes and then moving an arm.

**[00:24:11 – 00:24:14]** The system designers, a lot of these systems need to go

**[00:24:14 – 00:24:16]** into regulated industries.

**[00:24:16 – 00:24:19]** They might need to live inside of an OR, they might need to be tested

**[00:24:19 – 00:24:20]** for different kind of things.

**[00:24:20 – 00:24:25]** And so we have a whole ecosystem of system and design partners who

**[00:24:25 – 00:24:29]** take the otherwise infrastructure underneath, build a very thoughtful

**[00:24:29 – 00:24:33]** design that can be industrialized for your industry, and then

**[00:24:33 – 00:24:35]** connect it to your sensors.

**[00:24:35 – 00:24:36]** And so this ecosystem is incredible, and

**[00:24:36 – 00:24:38]** we're so proud of it.

**[00:24:38 – 00:24:42]** And so please go check out Holoscan 5.

**[00:24:42 – 00:24:43]** You saw it this morning.

**[00:24:43 – 00:24:47]** We are incredibly excited to announce the work that we're

**[00:24:47 – 00:24:51]** doing with Siemens Healthineers to really look at the next generation

**[00:24:51 – 00:24:54]** of adaptive radiotherapy.

**[00:24:54 – 00:24:57]** As we spoke about, when you think about the work that has to happen

**[00:24:57 – 00:25:03]** inside of these really, really complex but insanely high-value

**[00:25:03 – 00:25:09]** and life-saving pathways like radiation therapy, it's a very hard

**[00:25:09 – 00:25:14]** process, and we need to connect the agentic and digital world.

**[00:25:15 – 00:25:18]** You know, the first thing we have to do is we take imaging data,

**[00:25:18 – 00:25:21]** and we have to deeply understand it and make a plan for being able

**[00:25:21 – 00:25:24]** to administer radiation to a tumor.

**[00:25:24 – 00:25:25]** And that is very human-driven today.

**[00:25:25 – 00:25:29]** It's very manual, contouring and doing the planning and

**[00:25:29 – 00:25:32]** simulating the radiation dosage.

**[00:25:32 – 00:25:35]** All of that exquisite planning has to come in.

**[00:25:35 – 00:25:38]** And so, Siemens has decided, why wouldn't we turn that

**[00:25:38 – 00:25:40]** into an agentic workflow?

**[00:25:40 – 00:25:46]** Let our planners be 50% more effective, more efficient, I

**[00:25:46 – 00:25:49]** should say, in being able to build up these plans so that they can get

**[00:25:49 – 00:25:54]** these treatments to patients faster and work in natural language.

**[00:25:54 – 00:25:57]** We all want to not have to know what buttons to click.

**[00:25:57 – 00:25:59]** Work in natural language.

**[00:25:59 – 00:26:00]** Make the plan.

**[00:26:00 – 00:26:04]** Take that planning, pull it into simulation, because it

**[00:26:04 – 00:26:07]** is the case, as we said, physics is at play with our human body.

**[00:26:07 – 00:26:08]** It's changing all the time.

**[00:26:08 – 00:26:11]** It's changing as we're radiating it.

**[00:26:11 – 00:26:14]** So how can we simulate not only what is changing in the

**[00:26:14 – 00:26:19]** human body, so patient simulation with our medical physics,

**[00:26:19 – 00:26:23]** but also simulate how is the medical device actually operating?

**[00:26:23 – 00:26:25]** Image-guided radiation therapy is a robot.

**[00:26:25 – 00:26:27]** It has been for a long time.

**[00:26:27 – 00:26:29]** You know, Siemens has been a robotics company

**[00:26:29 – 00:26:31]** for a very long time.

**[00:26:31 – 00:26:34]** Now they can think about how we enhance that simulation

**[00:26:34 – 00:26:39]** to then drive what the future goal is, is to have true real-time

**[00:26:39 – 00:26:41]** adaptive radiation therapy.

**[00:26:41 – 00:26:45]** And with that, we're co-designing also the real-time run computer

**[00:26:45 – 00:26:47]** for these systems to go be run at.

**[00:26:47 – 00:26:50]** So it's an incredible, incredible opportunity to work with such

**[00:26:50 – 00:26:53]** a sophisticated group.

**[00:26:53 – 00:26:56]** I have family members who have been treated with this technology,

**[00:26:56 – 00:27:00]** I'm indebted to it, and I'm delighted to take it into its next

**[00:27:00 – 00:27:02]** generation of adaptive therapy.

**[00:27:02 – 00:27:07]** And this is truly the blueprint of what all medical devices are

**[00:27:07 – 00:27:09]** going to look like in the future.

**[00:27:09 – 00:27:13]** You're going to have all the digital information to set these up

**[00:27:13 – 00:27:15]** or protocol these machines, right?

**[00:27:15 – 00:27:18]** You're going to have the interface that might be voice generated

**[00:27:18 – 00:27:22]** to help patients even, you know, autonomously navigate getting their

**[00:27:22 – 00:27:24]** own medical studies being done.

**[00:27:24 – 00:27:27]** And then you're going to also want to have that real-time

**[00:27:27 – 00:27:32]** sensing to control that truly makes all medical devices robotic.

**[00:27:32 – 00:27:37]** And so we're really jumping on and taking advantage of

**[00:27:37 – 00:27:39]** an amazing effort.

**[00:27:39 – 00:27:44]** If you guys haven't checked it out, OpenAge is a huge

**[00:27:44 – 00:27:47]** community-driven data set.

**[00:27:47 – 00:27:52]** Open and H for Healthcare, for healthcare-specific data sets.

**[00:27:52 – 00:27:57]** We started it in robotics, in clinical robotics, where we have

**[00:27:57 – 00:27:58]** many, many different embodiments.

**[00:27:58 – 00:28:04]** We have the richest, you know, over 750 hours, it's the largest

**[00:28:04 – 00:28:05]** data set in the world.

**[00:28:05 – 00:28:09]** And so today we're announcing Open H for Hospitals to really

**[00:28:09 – 00:28:13]** help find a lot of different and diverse hospital environments.

**[00:28:13 – 00:28:17]** So we can start to learn by real-world data about What's

**[00:28:17 – 00:28:18]** going on in those environments?

**[00:28:18 – 00:28:21]** How does it work? What are people even doing?

**[00:28:21 – 00:28:24]** That becomes the basis of us then being able to take

**[00:28:24 – 00:28:28]** advantage of things like Cosmos and synthetically generate

**[00:28:28 – 00:28:32]** a whole bunch of more environments, many more tasks, different

**[00:28:32 – 00:28:36]** types of people, different Volumes of people, all of that

**[00:28:36 – 00:28:38]** can essentially come into play.

**[00:28:38 – 00:28:41]** And so we're working with a wonderful set of collaborators

**[00:28:41 – 00:28:45]** to really introduce this next generation, or I would say

**[00:28:45 – 00:28:48]** next evolution of the OpenAge data sets, this one particularly for

**[00:28:48 – 00:28:50]** hospitals and hospital automation.

**[00:28:50 – 00:28:54]** And the cool thing is, it's right at the time that with Siemens,

**[00:28:54 – 00:29:00]** also NVIDIA, where with the RoboCup for the first time, they actually

**[00:29:00 – 00:29:02]** have a healthcare challenge.

**[00:29:03 – 00:29:04]** And that speaks to me.

**[00:29:04 – 00:29:07]** I remember the days, you know, years ago when the NeurIPS

**[00:29:07 – 00:29:10]** conference was starting to build in biology and building

**[00:29:10 – 00:29:14]** chemistry, and you get really excited that, like, okay, all

**[00:29:14 – 00:29:18]** these brilliant computer scientists and AI researchers are going

**[00:29:18 – 00:29:20]** to Turn their sites to healthcare.

**[00:29:20 – 00:29:21]** This is fantastic.

**[00:29:21 – 00:29:25]** This is going to be, you know, healthcare becoming a destination

**[00:29:25 – 00:29:28]** for all of the roboticists in the world, and this is

**[00:29:28 – 00:29:29]** a wonderful thing.

**[00:29:29 – 00:29:32]** So I would say we're going to be, you know, helping to

**[00:29:32 – 00:29:35]** publish recipes; we're publishing these data sets, so you can

**[00:29:35 – 00:29:39]** really take advantage of all of our open source tools and data

**[00:29:39 – 00:29:42]** and participate in this challenge.

**[00:29:42 – 00:29:47]** Okay, we're going to move on to the next sector here, where

**[00:29:47 – 00:29:51]** AI, not only is it happening in the walls of the hospital and extending

**[00:29:51 – 00:29:55]** beyond the hospital, but the other part of this is treatments.

**[00:29:55 – 00:30:00]** And we need to take what is right now $2 billion, 12 years,

**[00:30:00 – 00:30:02]** you pick your number, it's really long, failure rate's

**[00:30:03 – 00:30:08]** too high, too expensive, and take this opportunity to connect

**[00:30:08 – 00:30:16]** Digital and physical AI so that AI can become a new lab companion.

**[00:30:16 – 00:30:19]** And so it's truly joining this scientific method.

**[00:30:19 – 00:30:21]** What is the scientific method?

**[00:30:21 – 00:30:26]** You hypothesize, you go off, you do a ton of research, you

**[00:30:26 – 00:30:29]** have to synthesize a whole bunch of information, think about what

**[00:30:30 – 00:30:33]** experiments you want to drive, go make something, go test something,

**[00:30:33 – 00:30:34]** and go around the loop again.

**[00:30:34 – 00:30:37]** Okay, this is the scientific method.

**[00:30:37 – 00:30:41]** And so you can take everything that we've built for agentic AI and

**[00:30:41 – 00:30:45]** apply it for a lot of that upfront, you know, science hypothesis

**[00:30:45 – 00:30:47]** generation and doing your research.

**[00:30:47 – 00:30:55]** You then want to connect that AI to what I call a digital dry lab.

**[00:30:55 – 00:30:58]** These are some of the domain-specific, industry-specific

**[00:30:58 – 00:31:03]** tools that, you know, 50 years ago are instruments on a lab

**[00:31:03 – 00:31:08]** bench, but are now being able to start being represented in models.

**[00:31:08 – 00:31:10]** AlphaFold taught us that.

**[00:31:10 – 00:31:13]** And so NVIDIA has the BioNeMo Agent Toolkit that we've been working

**[00:31:13 – 00:31:17]** on for a long time with all these domain-specific tools that we can

**[00:31:17 – 00:31:23]** now give agents, to start exploring vast amounts of chemical and

**[00:31:23 – 00:31:28]** protein space and refine designs in the dry lab before they ever go off

**[00:31:28 – 00:31:30]** and go and get into the wet lab.

**[00:31:30 – 00:31:34]** And then even when you get into the wet lab, you want that wet

**[00:31:34 – 00:31:35]** lab to essentially be autonomous.

**[00:31:35 – 00:31:37]** You don't want it to become the bottleneck.

**[00:31:37 – 00:31:40]** You don't want there to be, you know, if somebody's not

**[00:31:40 – 00:31:43]** there to pipette something or carry a sample from one cell to

**[00:31:43 – 00:31:48]** another, to slow down the learning process of the scientific method.

**[00:31:48 – 00:31:50]** And so that's what's really exciting here, is connecting

**[00:31:50 – 00:31:53]** these two together.

**[00:31:53 – 00:31:57]** So we also are announcing an incredible advancement in

**[00:31:57 – 00:31:59]** work that we're doing with Danaher.

**[00:31:59 – 00:32:04]** That news went out this week, of really advancing autonomous science

**[00:32:04 – 00:32:07]** and thinking about it end to end.

**[00:32:07 – 00:32:12]** And so, with Danaher, they've got their VECO and GeneData

**[00:32:12 – 00:32:16]** platform that is now completely agentically driven.

**[00:32:16 – 00:32:19]** They're integrating in the BioNeMo agent toolkit so that

**[00:32:19 – 00:32:24]** you can put these molecular designs, workflows together at

**[00:32:24 – 00:32:27]** the digital workbench, if you will.

**[00:32:27 – 00:32:29]** And then also looking at how do we make and test these

**[00:32:29 – 00:32:34]** more efficiently across their very, very vast portfolio of instruments.

**[00:32:34 – 00:32:37]** And there is where you can take edge computers like the

**[00:32:37 – 00:32:42]** DGX Spark, and you can even make that individual instrument

**[00:32:42 – 00:32:44]** intelligent in nature.

**[00:32:44 – 00:32:49]** An ability to protocol it and QC it and make it more efficient.

**[00:32:49 – 00:32:52]** And so we're working on the end-to-end process to really

**[00:32:52 – 00:32:56]** imagine all of the so many labs out there that exist already

**[00:32:56 – 00:33:03]** today and, you know, let it be, you know, scientist-led design, but

**[00:33:03 – 00:33:08]** allow for a closed loop scientific method to truly transpire.

**[00:33:08 – 00:33:12]** So every single experiment that's done within any scientific

**[00:33:12 – 00:33:17]** organization is the scientific data flywheel back into the

**[00:33:18 – 00:33:21]** agentic system, and you turn that crank as fast as you can go.

**[00:33:21 – 00:33:26]** And so science soon here should go as fast as compute.

**[00:33:26 – 00:33:29]** And wouldn't that be incredible to unlock things?

**[00:33:29 – 00:33:32]** So it's an amazing partnership that is underway with ourselves

**[00:33:32 – 00:33:35]** and Danaher, and they're releasing and working on

**[00:33:35 – 00:33:38]** a ton of innovation in this area.

**[00:33:38 – 00:33:43]** Lila Scientific, also they're building the super

**[00:33:43 – 00:33:47]** intelligence really for labs.

**[00:33:47 – 00:33:51]** Again, connecting both the digital world and the physical world.

**[00:33:51 – 00:33:54]** And this is the point I wanted to describe to you.

**[00:33:54 – 00:34:00]** Science has, in large part, by necessity, been quite narrow

**[00:34:00 – 00:34:01]** and been siloed, right?

**[00:34:01 – 00:34:03]** It's deeply specialized.

**[00:34:03 – 00:34:08]** And so we have the opportunity here to take those specializations

**[00:34:08 – 00:34:11]** and give Lila the power to be at the frontier.

**[00:34:11 – 00:34:14]** So for a lot of the exquisite work that they're doing, because they

**[00:34:14 – 00:34:20]** have designed a lab that can work across materials and chemistry and

**[00:34:20 – 00:34:26]** proteins, to specify the tasks that they really need using the NVIDIA

**[00:34:26 – 00:34:31]** NEMOTRON family and surpass the frontier and then have a completely

**[00:34:31 – 00:34:34]** different cost structure for being able to run that at scale.

**[00:34:34 – 00:34:37]** And so here you can see just evidence of that, for essentially

**[00:34:37 – 00:34:40]** doing the design of synthesis recipes, you know, here's

**[00:34:40 – 00:34:44]** GPT-Astra and with a small amount of post-training, this

**[00:34:44 – 00:34:49]** was published publicly, about$ 2,000 worth of post-training,

**[00:34:49 – 00:34:53]** being able to essentially, as Jetson said, they are now the

**[00:34:53 – 00:34:55]** frontier for this particular task.

**[00:34:55 – 00:34:58]** And that's incredibly powerful for the work that is ahead.

**[00:34:58 – 00:35:01]** So let's take a minute, I know you might have seen this if you were

**[00:35:01 – 00:35:05]** at Jetson's keynote, but I want to have you see how Lila is truly

**[00:35:05 – 00:35:13]** connecting the digital and agentic world to a fully autonomous lab.

**[00:35:14 – 00:35:17]** Scientists can imagine far more experiments than

**[00:35:17 – 00:35:19]** they can carry out.

**[00:35:19 – 00:35:24]** Lyla Sciences is connecting AI to an autonomous laboratory,

**[00:35:24 – 00:35:28]** accelerated by NVIDIA.

**[00:35:28 – 00:35:32]** Here, a scientist's goal is to design a therapy

**[00:35:32 – 00:35:35]** that uses a patient's own immune cells to recognize

**[00:35:35 – 00:35:38]** and fight B-cell lymphoma.

**[00:35:39 – 00:35:41]** First, scientific reasoning.

**[00:35:41 – 00:35:44]** Lila Iris takes the scientist's goal and

**[00:35:44 – 00:35:47]** builds an experimental plan.

**[00:35:48 – 00:35:53]** IRIS is built on NVIDIA Nemotron and NVIDIA Agent Toolkit;

**[00:35:53 – 00:35:57]** then the agents go to work in the computational dry lab

**[00:35:57 – 00:36:02]** using domain specific models built with NVIDIA BioNeMo.

**[00:36:02 – 00:36:05]** They can run millions of virtual experiments,

**[00:36:05 – 00:36:10]** narrowing the candidates before testing them in the lab.

**[00:36:11 – 00:36:16]** Those candidates move into the AI science factory.

**[00:36:16 – 00:36:19]** Autonomous robots make and test them.

**[00:36:19 – 00:36:23]** NVIDIA Isaac and Omniverse help develop and simulate

**[00:36:23 – 00:36:25]** the robotic systems.

**[00:36:25 – 00:36:28]** Jetson and Thor provide the compute.

**[00:36:28 – 00:36:32]** The experiments measure how the candidates perform.

**[00:36:32 – 00:36:36]** The experimental results go back to Iris and inform the

**[00:36:36 – 00:36:38]** next designs and experiments.

**[00:36:39 – 00:36:42]** Then the loop begins again.

**[00:36:42 – 00:36:47]** Each cycle gives scientists more evidence to build on,

**[00:36:47 – 00:36:52]** because behind every experiment is the hope of a new treatment.

**[00:36:52 – 00:36:54]** Incredible.

**[00:36:54 – 00:36:57]** Incredible. Future of science right there, and they've

**[00:36:57 – 00:36:58]** already built it.

**[00:36:59 – 00:37:02]** So it's an incredible opportunity.

**[00:37:02 – 00:37:07]** NVIDIA has put together also what we call the BioNeMo Agent for

**[00:37:07 – 00:37:10]** Life Sciences Developer Example.

**[00:37:11 – 00:37:14]** We essentially put it out, it's an open-source GitHub

**[00:37:14 – 00:37:17]** that gives you a reference of how would you put this all together.

**[00:37:17 – 00:37:22]** Taking advantage of NVIDIA, Agent Toolkit, Nemotron, the

**[00:37:22 – 00:37:25]** BioNeMo Agent Toolkit tools, and putting it together in

**[00:37:25 – 00:37:29]** a workflow so that you could build your own agentic systems.

**[00:37:29 – 00:37:34]** We've published this. You can experience it on build.com.

**[00:37:34 – 00:37:39]** You can also get it on GitHub and our brev and put in your

**[00:37:39 – 00:37:41]** own molecules and kind of test it.

**[00:37:41 – 00:37:45]** And so it's the way that we are going to build these scientific

**[00:37:45 – 00:37:48]** agents, is we are going to have many specialized models.

**[00:37:48 – 00:37:53]** These are multi-model systems that are going to be continuously

**[00:37:53 – 00:37:56]** kind of coming into the next frontier of models.

**[00:37:56 – 00:37:59]** We've got to be ready for the next model to come.

**[00:37:59 – 00:38:03]** But this is the architecture of the modern software system,

**[00:38:04 – 00:38:07]** and so we're working with incredible partners who are

**[00:38:07 – 00:38:11]** doing just this and essentially building scientific agents

**[00:38:11 – 00:38:12]** with the BioNeMo Agent Toolkit.

**[00:38:13 – 00:38:16]** Bolts, obviously for molecular design platform where they've

**[00:38:16 – 00:38:19]** built their own foundation models, they now are turning

**[00:38:19 – 00:38:23]** that into an agentic system so that they can turn that molecular design

**[00:38:23 – 00:38:25]** on steroids and go really fast.

**[00:38:25 – 00:38:30]** BioElevate as an AI scientist platform, and BenchLink, where

**[00:38:30 – 00:38:34]** so many, tens of thousands of companies around the world

**[00:38:34 – 00:38:37]** use already electronic lab notebooks and giving them the

**[00:38:37 – 00:38:41]** agentic work style and connecting it to the experimental data

**[00:38:41 – 00:38:43]** in real time as it's coming in.

**[00:38:43 – 00:38:47]** And then Therna for doing RNA medicines discovery.

**[00:38:47 – 00:38:50]** All of these different applications now have the ability to, one,

**[00:38:50 – 00:38:55]** become agentic, and two, use these domain-specific tools so that

**[00:38:55 – 00:38:59]** the data flywheel and scientific flywheel can turn really fast.

**[00:38:59 – 00:39:03]** Now, I want to introduce you to what is a new

**[00:39:03 – 00:39:05]** model at the frontier.

**[00:39:05 – 00:39:08]** We're super excited to be announcing today...

**[00:39:08 – 00:39:13]** The RFD4 Proteina model that has been co-developed with

**[00:39:13 – 00:39:17]** the University of Washington Institute of Protein Design

**[00:39:17 – 00:39:21]** and the David Baker Lab, Nobel Laureate, and NVIDIA researchers.

**[00:39:21 – 00:39:24]** This model is incredible.

**[00:39:24 – 00:39:27]** It's a new model that has reasoning, which we started to have

**[00:39:27 – 00:39:33]** in our last Proteina, and it has this incredible programmability.

**[00:39:33 – 00:39:36]** I'm going to explain to you really what that means.

**[00:39:36 – 00:39:41]** In one way, you have this shared latent space that connects

**[00:39:41 – 00:39:44]** protein sequences and their structure, and it allows the

**[00:39:44 – 00:39:49]** model to design the two together, and then with the ability to

**[00:39:49 – 00:39:53]** sort of program it with the inputs to the model so that it can do

**[00:39:54 – 00:39:56]** these incredibly refined designs.

**[00:39:56 – 00:40:01]** The model is state-of-the-art, as proven by experimental

**[00:40:01 – 00:40:05]** verifiers that were gracious to spend the time with us on this.

**[00:40:05 – 00:40:09]** This model essentially replaces, call it, dozens and dozens

**[00:40:09 – 00:40:13]** of other specialized models into one model for all of

**[00:40:13 – 00:40:15]** these extremely complicated tasks.

**[00:40:15 – 00:40:19]** And they're doing things that we couldn't design proteins

**[00:40:19 – 00:40:20]** to do in the past.

**[00:40:20 – 00:40:24]** So, let me show you a little bit about how this is a new

**[00:40:24 – 00:40:27]** performance frontier in a couple of ways for protein design.

**[00:40:27 – 00:40:30]** One is, for the amount of compute, how many designs

**[00:40:30 – 00:40:34]** can you make, is off to the right, and then for that amount

**[00:40:34 – 00:40:36]** of compute, what is your hit rate?

**[00:40:36 – 00:40:40]** What are the, what are the, how many of them would you actually

**[00:40:40 – 00:40:41]** think about putting into the lab?

**[00:40:41 – 00:40:43]** Because that is the expensive part.

**[00:40:43 – 00:40:46]** And so, before investigating the, you know, lab work, we,

**[00:40:46 – 00:40:50]** we ask whether the protein is predicted to fold, right?

**[00:40:50 – 00:40:53]** And so you can use alpha-fold 3, for example, to make sure

**[00:40:53 – 00:40:57]** this, this, this, um, structure is folding, um, and has the,

**[00:40:57 – 00:41:01]** the structural and, and confidence criteria that you want.

**[00:41:01 – 00:41:04]** And so higher means greater share, meets that criteria;

**[00:41:04 – 00:41:08]** and farther to the right means that, we have faster generation,

**[00:41:08 – 00:41:11]** and that's what I mean by a new performance frontier.

**[00:41:11 – 00:41:13]** And that was for protein-protein interactions.

**[00:41:13 – 00:41:16]** Now for small molecule binders, you know, this

**[00:41:16 – 00:41:18]** is a distinct capability.

**[00:41:18 – 00:41:22]** Protein and small molecules are adapting together during

**[00:41:22 – 00:41:26]** this design process, which is extremely exciting.

**[00:41:26 – 00:41:29]** And this is essentially five times the design.

**[00:41:29 – 00:41:31]** Five times the designs.

**[00:41:31 – 00:41:34]** And for proteins, evolution, it just so happened that we

**[00:41:35 – 00:41:37]** found something that worked and we kept going.

**[00:41:37 – 00:41:40]** It wasn't perfect. Here we can keep looking, keep looking, keep looking

**[00:41:40 – 00:41:43]** until it's absolutely perfect.

**[00:41:43 – 00:41:47]** And we're going to be able to do that now with test time compute.

**[00:41:47 – 00:41:51]** And so, it's incredible to think that these protein models

**[00:41:51 – 00:41:54]** and these biology foundation models are really getting

**[00:41:54 – 00:41:56]** into this scale deployment phase.

**[00:41:56 – 00:42:01]** Compute really does mean better protein models, and so we

**[00:42:01 – 00:42:05]** have an awesome amount of partners who are working on this.

**[00:42:05 – 00:42:10]** Of course, you can go get your new DSX PodFactory, if you like.

**[00:42:10 – 00:42:15]** But we also have this available today in all sorts of ways.

**[00:42:15 – 00:42:20]** So with OpenAI ChachiBT Codex, there's a new custom

**[00:42:20 – 00:42:25]** RD4 Proteina molecule viewer in UX that you can use.

**[00:42:25 – 00:42:29]** Anthropic with our BioNeMo Agent Toolkit in Claude Sciences,

**[00:42:29 – 00:42:31]** it's available there.

**[00:42:31 – 00:42:33]** And then it's available on all of these scaler partners,

**[00:42:33 – 00:42:35]** so Nebius' Token Factory.

**[00:42:35 – 00:42:39]** So if you wanted to go do a design program and spin up a thousand

**[00:42:39 – 00:42:42]** GPUs, it is ready to go for you.

**[00:42:42 – 00:42:44]** Now, I have one special thing before we close.

**[00:42:45 – 00:42:46]** I know we're at time here.

**[00:42:46 – 00:42:50]** I want us to hear from the Nobel laureate himself, David

**[00:42:50 – 00:42:53]** Baker, and his team at University of Washington, and some of

**[00:42:54 – 00:42:59]** our researchers at NVIDIA about how incredible this model is.

**[00:43:00 – 00:43:03]** Proteins in nature solve all the problems that arose

**[00:43:03 – 00:43:06]** during evolution.

**[00:43:06 – 00:43:10]** They digest our food, they mediate the currents going

**[00:43:10 – 00:43:15]** through our brain when we think, they power our motion.

**[00:43:15 – 00:43:17]** They're almost like magical things and the promise of

**[00:43:17 – 00:43:21]** being able to design new proteins is that we can now solve these

**[00:43:21 – 00:43:24]** problems that weren't relevant during evolution but are really

**[00:43:24 – 00:43:27]** important for humans today.

**[00:43:29 – 00:43:32]** These are challenges nature never faced, things that nature never

**[00:43:32 – 00:43:35]** had the time or chance to develop against, such as degrading man-made

**[00:43:35 – 00:43:39]** pesticides, curing Alzheimer's disease, finding this new

**[00:43:39 – 00:43:45]** cancer target, fixing nitrogen in a more environmentally friendly way.

**[00:43:46 – 00:43:48]** This project began about a year ago with a very small group

**[00:43:48 – 00:43:52]** of researchers from both NVIDIA Research and the Baker Lab here

**[00:43:52 – 00:43:54]** at the University of Washington.

**[00:43:54 – 00:43:57]** The collaboration with NVIDIA has been great in fundamental

**[00:43:57 – 00:44:01]** level algorithm optimization, the incredible expertise.

**[00:44:01 – 00:44:04]** It showed me what kind of innovation is possible at

**[00:44:04 – 00:44:09]** the right place and the right time with the right resources.

**[00:44:12 – 00:44:17]** RFD4 Proteina unites concepts from the RF Diffusion series

**[00:44:17 – 00:44:21]** of models with concepts from the Proteina Complexo series

**[00:44:21 – 00:44:24]** of models developed at NVIDIA to make something that's more

**[00:44:24 – 00:44:28]** powerful than any previous model.

**[00:44:30 – 00:44:32]** When we first saw the results of the computer,

**[00:44:32 – 00:44:34]** we couldn't believe them.

**[00:44:34 – 00:44:35]** They were actually too good to be true.

**[00:44:35 – 00:44:38]** We wanted to put the model to a real test.

**[00:44:38 – 00:44:42]** It takes so long to go from that idea or proof of concept

**[00:44:42 – 00:44:46]** to like a vial of a drug that can be injected into a patient.

**[00:44:46 – 00:44:50]** It's like a 10 to 12 year process, and it's like a billion

**[00:44:50 – 00:44:52]** dollars in between.

**[00:44:52 – 00:44:57]** With the introduction of RFD4 proteina, we see these molecules

**[00:44:57 – 00:44:59]** that we can take directly off the computer.

**[00:44:59 – 00:45:03]** And not only do they bind to the target that we intend,

**[00:45:03 – 00:45:08]** there's this 10 to 1,000 fold increase in the tightness with

**[00:45:08 – 00:45:10]** which these designs are binding.

**[00:45:10 – 00:45:14]** It's an incredible advancement in terms of using these

**[00:45:14 – 00:45:19]** computational tools to speed up the process of drug design.

**[00:45:21 – 00:45:24]** I work on making enzymes that break down pesticides.

**[00:45:24 – 00:45:26]** Enzymes are great because they do very selective

**[00:45:26 – 00:45:30]** chemistry, so ultimately it's cleaner for the world.

**[00:45:31 – 00:45:33]** I've been working on this problem for a year.

**[00:45:33 – 00:45:36]** With the new model, RF Diffusion 4 Proteina,

**[00:45:36 – 00:45:38]** we did it in just four days.

**[00:45:38 – 00:45:41]** And not only that, but the best one that we found was

**[00:45:41 – 00:45:44]** over a thousand times better than the previous enzymes

**[00:45:44 – 00:45:47]** I had developed over the past year.

**[00:45:47 – 00:45:52]** The way chemistry is done is going to be changed forever.

**[00:45:52 – 00:45:55]** I'm very excited to see all the great things that people

**[00:45:55 – 00:46:00]** do with RFD4 proteina and how it makes the world a better place.

**[00:46:02 – 00:46:03]** It's incredible, right?

**[00:46:03 – 00:46:06]** Yeah, I know.

**[00:46:07 – 00:46:10]** See, I got chills for you guys, too, right?

**[00:46:10 – 00:46:11]** Chills, chills.

**[00:46:11 – 00:46:14]** What you can do in four days in what would otherwise take

**[00:46:14 – 00:46:18]** you a year and somehow a thousand times more effective.

**[00:46:18 – 00:46:21]** Again, to all of you in the room, we are in this together.

**[00:46:21 – 00:46:24]** I see so many familiar faces.

**[00:46:24 – 00:46:25]** Thank you for everything you do.

**[00:46:25 – 00:46:29]** We are delighted to be a part of the incredible mission

**[00:46:29 – 00:46:33]** statement you've set for yourself and have an incredible GTC

**[00:46:33 – 00:46:37]** and go get used an RFD4 protean and design some proteins.

**[00:46:37 – 00:46:38]** All right, thank you very much.