SAN JOSE, CA—Just as users train artificial intelligence (AI) models to search for a druggable target or predict the structure of a protein, so Nvidia founder and CEO Jensen Huang says he trained himself to deliver this year’s keynote address on Monday at the company’s NVIDIA GTC 2024 conference.
“I hope this will turn out as well as I had it in my head,” Huang quipped to the audience of employees, developers, partners, investors and other industry watchers who filled the SAP Center to its roughly 19,000-seat capacity.
Judging from the loud applause when Huang finished his address about two hours later, he did.
Nvidia is the Silicon Valley-based microprocessing giant that has expanded its market-leading footprint in AI chips to industries that include the life sciences. Huang wowed the crowd by rattling off a series of corporate tech and commercial announcements.
While Nvidia received much attention by unveiling its next-generation NVIDIA Blackwell AI computing platform — designed to run real-time generative AI on trillion-parameter large language models at up to 25x less cost and energy consumption than its predecessor — many of the company’s announcements focused on drug discovery and genomics.
One was an expansion of NVIDIA BioNeMo™’s generative AI platform for drug discovery with foundation models capable of analyzing DNA sequences, predicting how proteins will change shape in response to a drug molecule, and determining a cell’s function based on its RNA.
BioNeMo is designed to simplify and accelerate the training of models on proprietary data, scaling up the deployment of models for drug discovery applications.
Among new foundation models available in BioNeMo is the platform’s first genomics model, DNABERT, which has been trained on DNA sequences from the human reference genome, Hg38.p13. DNABERT is designed to predict the function of specific regions of the genome, as well as analyze the effects of gene mutations and variants, by generating a dense representation of a genome sequence by identifying contextually similar sequences in the human genome.
For each nucleotide in the input sequence, DNABERT computes embeddings that can be used for a variety of predictive tasks. DNABERT can be applied for both commercial and non-commercial use.
Nvidia said it will make additional models for accelerating protein structure prediction, generative chemistry, and molecular docking prediction available via BioNeMo.
Among them is scBERT, which has been trained on data from single-cell RNA sequencing. scBERT is designed to let users apply the sequencing data toward tasks that include predicting the effects of gene knockouts and identifying cell types such as neurons, blood cells or muscle cells.
Another new model, EquiDock, will be among BioNeMo models capable of predicting the 3D structure of an interaction between two proteins. EquiDock is designed to better enable drug discoverers to tell if a molecule will be effective.
After explaining how over the years, AI has learned to understand text—starting with the simple word “cat”—followed by text, images, video, and speech, Huang introduced Nvidia’s life-sciences announcements by posing, then answering, a not-so-rhetorical question: “What else can you understand that you’ve digitized?”
Generative Revolution
“It turns out that we’ve digitized a lot of things: Proteins and genes and brainwaves. Anything you can digitize, so long as there’s structure, we can probably learn some patterns from it. And if we can learn the patterns from it, we can understand its meaning. If we can understand its meaning, we might be able to generate it as well. And so therefore, the generative revolution is here,” Huang declared.
As important as the new inference models Nvidia announced was its new technology for enabling users to access them. The models will be available to users through a new technology unveiled as NVIDIA Inference Microservices or NIMs.
NIMs are optimized cloud-native “microservices” designed to let developers accelerate deployment of generative AI models anywhere—whether through local workstations, on-premises data centers, cloud services, or GPU-accelerated workstations.
Nvidia says NIMs will also expand the pool of available developers of AI models as anywhere from 10-100 times more enterprise application developers gain the ability to transform their companies by applying AI.
Developers, in turn, will be able to simplify AI model development and packaging using industry-standard application programming interfaces (APIs) that enable developers to update their AI applications quickly, in many cases with as little as three lines of code, according to Nvidia.
Nvidia’s NIM catalog features 25 models across “healthcare” fields ranging from drug discovery to imaging, medtech, and digital health.
Among drug discovery-focused NIMs are DiffDock, the molecular docking model designed to predict the binding structure of a small molecule ligand to a protein; ESMFold, a “Transformer” model—a neural network that learns context and thus meaning by tracking relationships in sequential data—which can accurately predict protein structure based on a single amino acid sequence; and AlphaFold2, an AI model for protein folding developed by Google DeepMind. AlphaFold2 can predict (3D) structures of proteins from amino acid sequences with atomic-level accuracy.
Another NIM, MolMIM, is a generative chemistry model that generates drug candidates optimized for properties defined by users. MolMIM can also design molecules that are optimized to bind to a specific protein target. And another NIM, Universal DeepVariant, is designed to deliver 50x speed improvement for variant calling in genomic analysis workflows compared to the original or “vanilla” DeepVariant implementation designed to run on central processing units or CPUs.
For drug discoverers that require imaging, the VISTA 3D microservice can be used for accelerating creation of 3D segmentation models.
100+ BioNeMo Users
According to Nvidia, more than 100 companies are using BioNeMo. Among them is Cadence Design Systems, a software developer based in San Jose.
On Monday, Cadence joined Nvidia to announce a generative AI collaboration intended to dramatically accelerate drug discovery. Cadence will combine its cloud-native Orion® molecular design platform with BioNeMo and Nvidia microservices in a partnership intended to expand Cadence’s therapeutic design capabilities, namely AI-guided simulation of molecules and lead optimization, as well as shorten time to results.
Orion allows researchers at pharmaceutical companies to generate, search, and model data libraries containing hundreds of billions of compounds
“Our pharmaceutical and biotechnology customers require access to accelerated resources for molecular simulation. By leveraging BioNeMo microservices, researchers can generate molecules that are optimized according to scientists’ specific needs,” Anthony Nicholls, corporate vice president at Cadence, said in a statement.
BioNeMo users include numerous drug developers:
- Astellas Pharma is using BioNeMo to accelerate molecular simulations and large language models for drug discovery—as well as using the Tokyo-1 AI supercomputer toward that work.
- Iambic has agreed to contribute its NeuralPLexer model as a BioNeMo cloud API for noncommercial use, by helping researchers predict how a protein’s 3D structure changes in response to a drug molecule.
- Insilico Medicine has integrated BioNeMo in its AI-accelerated drug discovery workflow, which has grown into a pipeline of more than 30 therapeutic assets—six of them in clinical stages. Nvidia nurtured Insilico during its startup days by inviting it into its NVIDIA Inception program for startups.
- Recursion uses BioNeMo to offer its Phenom-Beta AI transformer model, designed to extract insights from cellular microscopy images to help researchers better understand cell function.
- Terray Therapeutics is developing a multi-target structural binding model with help from BioNeMo. Terray is also training generative AI models for small molecule design on Nvidia DGX Cloud.
Nvidia says developers can access NIM microservices through its NVIDIA AI Enterprise 5.0 software platform, by using NVIDIA-Certified Systems on premises, as well as through public cloud platforms such as Amazon Web Services (AWS), Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure.
Nvidia announced additional life sciences-focused collaborations as expansions of existing partnerships with Amazon Web Services (AWS), Microsoft, and other companies.
BioNeMo foundation models will soon be accessible on AWS HealthOmics, a service designed to help life sciences organizations and healthcare providers store, query, and analyze genomic, transcriptomic, and other “omics” data. AWS will use the NVIDIA Blackwell GPU platform.
Microsoft said it will combine its Microsoft Azure cloud platform with the NVIDIA DGX Cloud and NVIDIA Clara suite of computing platforms, software, and services, with the aim of helping life sciences organizations and healthcare providers accelerate clinical research and drug discovery, broaden access to precision medicines, and enhance medical image-based diagnostic technology.
GEN’s’ attendance at NVDIA GTC 2024 was funded by Nvidia.