NVIDIA BioNeMo and Anthropic Claude Science: How AI Agents Are Speeding Up Scientific Discovery

Scientific research has always moved at the pace of its slowest step — usually the part where a researcher has to leave their actual question behind and go configure software, wrangle compute resources, or format data for yet another specialized tool. Anthropic and NVIDIA just took a serious swing at fixing that problem.

Anthropic has released the public beta of Claude Science, an AI workbench purpose-built for scientific research, and it now integrates directly with the NVIDIA BioNeMo Agent Toolkit. Together, they let scientists describe a research task in plain English and have Claude orchestrate the entire computational workflow behind it — from genomic analysis to protein structure prediction to molecule design — without the researcher ever touching a config file.

Here’s what the integration actually does, why it matters, and how it could reshape the day-to-day work of computational biologists, chemists, and genomics researchers.

What Is Claude Science?

Claude Science is Anthropic’s AI workbench for scientific research. Instead of clicking through dashboards or writing custom scripts, researchers converse with Claude directly, describing what they want to accomplish. Claude then interprets that request and coordinates the necessary steps to get it done — end to end.

Want to analyze a genomic sequence? Predict a protein structure? Design a molecule that binds to a specific target? You describe the task, and Claude Science handles the orchestration, calling on specialized, domain-trained agents to execute each step.

What the NVIDIA BioNeMo Agent Toolkit Brings to the Table

NVIDIA has built one of the most extensive GPU-accelerated computing stacks around — spanning hardware, software frameworks, scientific models, microservices, and domain-specific tools for life sciences. The BioNeMo Agent Toolkit packages all of that into callable, well-documented “skills” that an AI agent can use.

Rather than Claude having to guess how to run a piece of scientific software, the toolkit tells it exactly what each tool does, what data it needs, and how to call it. That’s what lets Claude Science pick the right tool for the job, format the data correctly, run the computation on NVIDIA’s infrastructure, and hand back a usable result — all without human micromanagement.

This isn’t a niche integration, either. NVIDIA BioNeMo is already running in production at 18 of the top 20 global pharmaceutical companies, so the toolkit is plugging Claude into infrastructure that’s already deeply embedded in the industry.

From Natural Language to Lab-Ready Output

The specialized agents inside this system are trained on established protocols across genomics, proteomics, single-cell analysis, cheminformatics, and clinical research. That domain knowledge is what allows Claude to map a researcher’s plain-language request to the correct computational step, rather than requiring the scientist to know which specific model or pipeline to invoke.

A good example is cancer drug discovery. A researcher can start by identifying a known cancer-causing mutation, then simply ask Claude to design candidate inhibitors targeting it. From there, Claude Science and the BioNeMo toolkit work together — using NVIDIA NIM microservices — to run high-throughput inhibitor prediction, optimization, and validation as one continuous pipeline, instead of a series of disconnected manual steps.

Faster Genomic and Single-Cell Analysis

Speed is the whole point of this integration, and the numbers involved are substantial.

• Genomic analysis through NVIDIA Parabricks drops from hours to minutes, letting the agent factor genomic context into decisions in near real time.
• RAPIDS-singlecell, built by the scverse open-source community, compresses a 1.3-million-cell preprocessing and clustering job from 52 minutes down to roughly 25 seconds — turning single-cell analysis into a live part of the agent’s reasoning rather than an overnight batch job.
• nvMolKit accelerates cheminformatics tasks like similarity search and conformer generation by up to 3,000x, letting agents iterate across huge chemical spaces quickly.

The toolkit also gives agents access to advanced open biomolecular models — including Evo 2, Boltz-2, and OpenFold3 — so there’s a purpose-built model available for each stage of a research workflow, whether that’s sequence analysis, structure prediction, or molecule generation.

Production-Ready Deployment via NIM Microservices

For teams that need to move beyond experimentation, NVIDIA packages its biomolecular models as BioNeMo NIM microservices — containerized, pre-tuned inference endpoints designed for production use. Claude Science’s agents interact with these through a single stable API, meaning a research team can go from prototype to production deployment without re-engineering the underlying infrastructure.

Importantly, the BioNeMo Agent Toolkit is designed to be open and agent-framework-agnostic. The same scientific skills can, in principle, work across different AI agent platforms — not just Claude — which gives research organizations flexibility rather than locking them into one ecosystem.

What This Means for Researchers

The bigger idea here is tightening the loop between human reasoning and computational execution. Instead of a scientist submitting a job and waiting hours (or days) for results, the goal is a workflow where they can ask a question, get a fast answer, refine the query, and keep iterating — while staying focused on the actual science instead of the plumbing underneath it.

Anthropic is currently in the public beta phase for Claude Science and is actively gathering feedback from researchers on what additional integrations and domain specialists they’d like to see next. Engineering teams interested in the underlying tools can access the BioNeMo Agent Toolkit through NVIDIA’s developer resources and GitHub repositories.

Frequently Asked Questions

1. What is Claude Science?

Claude Science is Anthropic’s AI workbench built specifically for scientific research. It lets scientists use natural language to direct Claude through complex, end-to-end research workflows instead of manually configuring software and models.

2. What is the NVIDIA BioNeMo Agent Toolkit?

It’s a toolkit that packages NVIDIA’s GPU-accelerated life sciences models, libraries, and microservices into callable “skills” that AI agents like Claude can use to execute scientific computations.

3. How does Claude Science use the BioNeMo Agent Toolkit?

Claude Science connects natively to the toolkit, which tells Claude’s agents what each computational tool does and what data it needs. This lets Claude select the right tool, format inputs correctly, run the job on NVIDIA compute, and return results.

4. What kinds of research tasks can Claude Science handle?

It supports tasks across genomics, proteomics, single-cell analysis, cheminformatics, and clinical research — including things like analyzing genomic sequences, predicting protein structures, and designing candidate molecules.

5. How much faster is genomic analysis with this integration?

Genomic analysis run through NVIDIA Parabricks can drop from hours to just minutes, letting agents incorporate genomic context into their decisions almost in real time.

6. What is RAPIDS-singlecell and why does it matter here?

It’s a single-cell analysis tool from the scverse community that compresses a 1.3-million-cell preprocessing and clustering workflow from 52 minutes to about 25 seconds, making single-cell analysis fast enough to be part of an agent’s live reasoning process.

7. What models are available through the BioNeMo toolkit?

Researchers get access to advanced open biomolecular models including Evo 2, Boltz-2, and OpenFold3, each suited to different stages of a research pipeline.

8. What are BioNeMo NIM microservices?

They’re containerized, pre-optimized inference endpoints for NVIDIA’s biomolecular models, designed for stable, production-grade deployment. Agents access them through a single API.

9. Is the BioNeMo Agent Toolkit limited to Claude?

No. The toolkit is designed to be open and agent-framework-agnostic, meaning its scientific skills can work across different AI agent platforms, not just Claude Science.

10. Is Claude Science available to everyone right now?

Claude Science is currently in public beta. Anthropic is actively collecting feedback from researchers to guide further integrations and domain-specific agent development.