Nvidia introduces AI reference workflow for drug researchers

Among the customers Nvidia has won with its recently released catalog of pre-trained, customizable workflows are drug researchers who enable enterprise users to build their own artificial intelligence (AI) applications.

NVIDIA NIM^™ Agent Blueprints are reference workflows designed to help drug researchers and other customers build and deploy generative AI applications for uses such as virtual screening, information retrieval, and even customer service avatars.

The blueprint includes NVIDIA NIMs, or optimized cloud-native “microservices,” that enable developers to accelerate the deployment of generative AI models anywhere—be it via local workstations, on-premises data centers, cloud services, or GPU-accelerated workstations. (NIM stands for “NVIDIA Inference Microservices”).

By enabling biopharma companies to move from traditional screening with fixed databases to generative, AI-driven molecule design and pre-optimization, NIM Agent Blueprints are designed to help researchers develop better molecules faster. According to Nvidia, this represents a paradigm shift in the drug discovery process – particularly in converting “hit” compounds into “lead” compounds optimized for further development.

“Many of the pharmaceutical companies I visit have 60 million molecules in their library. And that is what they use as screening, these static 60 million molecules. But there are 10⁶⁰ potential molecules in the chemical space that could be a therapy,” said Kimberly Powell, Nvidia’s business development manager for healthcare GEN Edge.

“The paradigm shift here is in generative AI, and in particular the MolMIM system: It takes advantage of the generative effect and intelligently searches chemical space so that a molecule that the world may have never synthesized before can have the properties that you really care about,” Powell said.

According to Nvidia, a NIM Agent Blueprint called Generative Virtual Screening is helping drug developers deliver on AI's long-standing promise of reducing the time and cost of developing new therapies by accelerating virtual screening of small molecules using generative models.

Improving the “hits” through three AI models

The Blueprint identifies and improves virtual “hit” compounds – identified through screening as potentially biologically active – in a smarter and more efficient way. At the core of generative virtual screening are three key AI models:

• AlphaFold2, an AI model for protein folding developed by Google DeepMind. AlphaFold2 can predict 3D structures of proteins from amino acid sequences with atomic accuracy.
• DiffDock, the molecular docking model that predicts the binding structure of a small molecule ligand to a protein while optimizing several properties such as high solubility and low toxicity.
• MolMIM, a generative chemistry model that generates drug candidates optimized for user-defined properties. MolMIM can also design molecules optimized for binding to a specific protein target.

Each AI model is packaged in NIMs that integrate the microservices into a flexible, scalable, generative AI workflow. The blueprint pre-optimizes molecules for desired therapeutic properties using a generative AI approach.

“Virtual screening is still only one part of drug discovery. But we are working on models that cover the entire path from target discovery to lead identification. And we will develop blueprints as we progress through this drug discovery process,” Powell said.

“In computational drug discovery, generative AI will play a role from start to finish,” she explained. “Often, computational drug discovery has been considered in lead identification optimization, where we do a lot of simulation. But now, from target identification to lead optimization, we use a lot of computational methods.”

Other NVIDIA NIMs focused on drug discovery include:

• ESMFold, a “transformer” model – a neural network that learns context and therefore meaning by tracking relationships in sequential data – that can accurately predict protein structure based on a single amino acid sequence.
• Parabricks DeepVariant (the tool behind the Universal Variant Calling Microservice), a deep learning model designed to identify variants in sequencing datasets with short and long read lengths. Parabricks is designed to increase the speed of variant calling in genome analysis workflows by 50x compared to the original or “normal” DeepVariant implementation designed to run on central processing units or CPUs.

“We have up to four or five generally available NIMs and a whole host of other NIMs in preview for drug discovery and healthcare. We're going to be very active in announcing and delivering these applications. Every month you'll see a new rich offering of NIMs and blueprints,” Powell said.

Digital Human, PDF Data Extraction

In addition to the drug discovery blueprint, other NIM Agent Blueprints include a digital human workflow for use cases ranging from digital health to customer service, and a multimodal PDF data extraction workflow for Enterprise Retrieval-Augmented Generation (RAG) designed to generate more accurate answers from massive amounts of business data.

According to Nvidia, RAG can read images from any PDF file and provide insights based on what it sees.

“In healthcare, through biomedical research, everything we do with insurance companies, all the interactions between patients and doctors, there are PDF files everywhere that contain a huge amount of useful information. Now we can extract it and summarize it,” Powell said.

The digital human workflow, which can also be used in digital healthcare, uses customized avatars capable of automatic speech recognition – like Nvidia's interactive digital human named “James”. Human speech is converted into text, which enters a language model, which from there enters the RAG system, returns to speech synthesis and activates a user's avatar.

“When you go through this complete loop, you now have a representation of a real digital human that can understand, reason and respond, and also use the Audio2Face NIM,” Powell said. “You can meet James, and when a text-to-speech response comes, you actually show a different emotion on your face and can have a much more engaging conversation.”

The blueprints can be downloaded for free by developers and deployed using the NVIDIA AI Enterprise software platform.

Biopharmaceutical companies – including all of the top 20 in the space according to Nvidia – access and distribute NVIDIA NIM Agent Blueprints to enterprises worldwide through global systems integrators and technology solution providers such as Accenture, Deloitte, SoftServe and World Wide Technology (WWT). Cisco, Dell Technologies, Hewlett Packard Enterprise and Lenovo offer full NVIDIA-accelerated infrastructure and solutions to accelerate the deployment of NIM Agent Blueprints.

Accenture plans to adapt a NIM Agent Blueprint to the specific requirements of drug development programs. To this end, Accenture is collaborating with biopharmaceutical companies to optimize the molecule generation step within the MolMIM NIM.

Amazon Web Services' AWS HealthOmics – a service designed to help biopharmaceutical companies and health systems store, query, and analyze genomic, transcriptomic, and other omics data – provides all three NIMs that make up the blueprint, with the goal of streamlining the integration of AI into existing drug discovery workflows, Nvidia said.