BioNeMo Agent Toolkit skills now power a Microsoft Discovery agent, nvidia-rfd4-proteina, that designs protein binders from a plain-language request. This post covers the science behind RFD4-Proteina, the architecture that connects the two platforms, and the steps to add the agent to your own Discovery workspace.
Drug discovery has always been a race against time. Finding a protein that can precisely grip a disease target, blocking it, degrading it, or flagging it for immune destruction. It has historically taken years of laboratory trial and error. AI is changing that timeline. With the NVIDIA BioNeMo Agent Toolkit now available on Microsoft Discovery, scientists can run state-of-the-art Protein/BioFM design workflows through a natural language interface, entirely within their Azure environment.
What Is BioNeMo Agent Toolkit?
The NVIDIA BioNeMo Agent Toolkit is an open-source collection of skills and structured instruction files that give AI agents deep domain expertise in computational biology. Each skill encodes what a BioNeMo expert knows about a specific model: which endpoint to call, how to format inputs, what the outputs mean scientifically, and how to chain models into validated pipelines.
Skills are installable by any AI agent platform: npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit.
The agent reasons and coordinates the workflow. The skills supply domain knowledge and guide how tools are used. The NIM (NVIDIA Inference Microservice) does the GPU computation.
What Is Microsoft Discovery?
Microsoft Discovery is an extensible Azure platform that brings together agentic orchestration, advanced reasoning, a graph-based knowledge foundation, and high-performance computing for scientific R&D. At its core is the Discovery Engine, a cognitive orchestrator that mimics the scientific method, reasoning across proprietary research data and external scientific literature to generate and validate hypotheses, storing outcomes in a long-term knowledge graph. Scientists interact through natural language; Discovery handles the compute.
For enterprise deployments, Microsoft Discovery deploys into your own Azure subscription and Entra tenant. The Bookshelf feature lets teams convert proprietary documents into Knowledge Bases using GraphRAG (a Microsoft Research technique), which agents use as grounding context with citations. Security controls include Entra ID authentication, Azure RBAC, private endpoints, encryption at rest, and opt-in audit logging.
RFD4-Proteina: The Science
RFD4-Proteina is a single model that unifies the RFdiffusion and Proteina AI model lineages, developed in collaboration with the Institute for Protein Design at the University of Washington. It generates a protein’s 3D structure and amino acid sequence in a single model pass, replacing the separate backbone design and sequence design steps that prior-generation pipelines required.
Supported design tasks
Why it matters for drug discovery
A cancer-driving protein like KRAS or HER2 drives the abnormal growth of a diseased cell. A precisely designed protein binder can block that protein’s signal, degrade it, or flag the cell for immune destruction. The challenge is designing a binder with the right 3D shape to interact with a specific part of the target protein.
RFD4-Proteina generates hundreds of candidate designs in hours. The Discovery agent filters them, ranks them, and delivers a shortlist to the scientist that is ready for wet-lab validation.
The Connection: Bringing BioNeMo Skills into Discovery Agents
The natural fit between these two platforms is the agent pattern they share. BioNeMo skills provide model-specific guidance. Discovery agents are system prompts that tell an agent how to use a tool container. The translation is direct. We built the first BioNeMo Discovery agent using RFD4-Proteina as the demonstration.
The tool container is eight lines of Dockerfile — python:3.11-slim plus a single Python script copied from the BioNeMo Agent Toolkit. No model weights. No CUDA stack. Builds in under a minute.
Why This Architecture Works
For scientists: no infrastructure expertise required. A protein biologist can design binders by describing the target in plain language. Microsoft Discovery handles the rest.
For IT and platform teams: the tool container is a lightweight Python image (python:3.11-slim, no model weights, CPU-only). GPU compute runs on a managed service: either the NVIDIA hosted API or Microsoft Foundry, where the entire compute stack is managed for you. AKS is also an option for teams that prefer self-hosted deployments.
For data sovereignty: molecular data stays within the customer’s Azure tenant when the container and model run on AKS inside that tenant. When the NVIDIA hosted API is used instead, requests go to that external service, so teams with strict residency requirements should choose a self-hosted deployment.
For extensibility: the same pattern — BioNeMo skill → Discovery agent.yaml → lightweight tool container → model running on AKS or Microsoft Foundry — applies to every BioNeMo model. DiffDock for molecular docking, Boltz-2 for binding affinity prediction, and GenMol for de novo molecule generation can each become a new Discovery agent by following the same contribution template.
How to Add the Agent to Your Local Discovery Workspace
The agent is contributed to the open-source microsoft/discovery catalog.
Requirements
- Microsoft Discovery workspace (desktop app or Azure)
- NVIDIA API key from build.nvidia.com/ipd/rfd4-proteina (free credits available)
- Docker (for local container build)
Steps
- In Discovery, run
/add-example-agents - Point to the community catalog:
https://github.com/microsoft/discovery - Select
nvidia-rfd4-proteina - Set your
NVIDIA_API_KEYin workspace secrets - Build the tool container:
cd agents/nvidia-rfd4-proteina/tools/nvidia-rfd4-proteina
docker build -t nvidia-rfd4-proteina:1.0.0 .
- Start designing: “Design protein binders for my target structure in
target.cif.”
What’s Next
This agent is the first of many. It provides a starting point for connecting additional BioNeMo models and workflows to Microsoft Discovery, with BioNeMo Agent Toolkit skills guiding their use. The agents below show where this pattern goes next:
Together these agents give scientists access to NVIDIA’s full computational biology stack through the same natural language interface they already use for every other scientific workflow.
Get Involved
- BioNeMo Agent Toolkit (open source): github.com/NVIDIA-BioNeMo/bionemo-agent-toolkit
- Microsoft Discovery app: github.com/microsoft/discovery
- RFD4-Proteina (free credits): build.nvidia.com/ipd/rfd4-proteina
- Contribute an agent: follow the authoring guide at
docs/authoring-guides/agent-authoring-guide.mdin the Discovery repo
NVIDIA BioNeMo is NVIDIA’s platform for AI-driven drug discovery and molecular biology. The BioNeMo Agent Toolkit is open source under Apache 2.0.
|
Aniket leads Global Business Development for NVIDIA’s Ecosystem Cloud partners in Healthcare and Life Sciences (HCLS). Previously, he was AWS’s lead for the Advanced Computing/HPC strategy for HCLS customers. Prior to AWS, Aniket worked in various technical roles for over a decade at DNAnexus, Qiagen, Knome, Pacific Biosciences, and Novartis. He holds a master’s degree in Bioinformatics from Northeastern University. |
|
|
Pouneh Partowkia is a business development and strategic partnerships leader at NVIDIA with over 20 years of experience driving enterprise technology adoption and growth. She works with customers and ecosystem partners across industries to accelerate AI adoption, with a focus on strategic partnerships, accelerated computing, and AI-driven innovation. |
|
|
Vinay Maddi is a CSE & GTM professional at NVIDIA, focused on accelerating enterprise AI adoption through strategic partnerships with Microsoft. His work spans Healthcare and Life Sciences, Manufacturing, and Automotive, with an emphasis on agentic AI, accelerated computing, and scientific discovery. |