NVIDIA’s BioNeMo NIMs Now Power Protein Folding Inside Anthropic’s Claude Science

NVIDIA has published a detailed technical walkthrough showing how its BioNeMo protein-modeling tools can now run directly inside Claude Science, Anthropic’s beta application for scientific research. The result is a working pipeline in which an AI agent generates evolutionary sequence data, folds a protein two different ways, and compares the outcomes — all without a researcher manually stitching the steps together.

The tutorial, published August 31, 2026, is not a new product launch. It’s a demonstration of a capability NVIDIA and Anthropic first put in place two months earlier, when NVIDIA introduced its BioNeMo Agent Toolkit and named Anthropic among the companies building on it. What’s new here is a concrete, reproducible example of what that integration can actually do — and it’s worth a close look, both for what it shows and for what it doesn’t yet prove.

What Happened

NVIDIA’s developer blog walks through a full protein-structure workflow run entirely inside Claude Science. Using three NVIDIA BioNeMo NIM microservices — MSA-Search, OpenFold3, and Boltz-2 — the agent pulls protein sequences from the public database UniProt, builds evolutionary alignments, and then predicts structures for a protein called Seh1 both on its own and paired with a proposed partner protein, C1HCX1, from the fungus Paracoccidioides lutzii.

This builds on groundwork laid earlier in 2026. NVIDIA announced its BioNeMo Agent Toolkit on June 23, packaging more than a decade of its life-sciences models and libraries into tools that AI agents can call directly. In that announcement, NVIDIA named Anthropic and OpenAI among the frontier AI labs integrating the toolkit. A week later, on June 30, Anthropic launched Claude Science itself, confirming in its own product materials that the app connects to BioNeMo models including Evo 2, Boltz-2, and OpenFold3.

The August 31 post is best understood as the payoff of that groundwork: a real, working example rather than an announcement of something new.

What Is Claude Science, Exactly?

It’s easy to assume “Claude Science” is a specialized new AI model. It isn’t. According to Anthropic’s own FAQ, Claude Science “is a public beta app, not a model. It uses the same Claude models your plan includes.” What’s new is everything built around those models: scientific databases, lab-tool connections, and compute management, packaged into a single working environment.

The app launched in beta on June 30, 2026 for macOS and Linux, available to users on Anthropic’s Pro, Max, Team, and Enterprise plans (Team and Enterprise users need an administrator to turn it on). It’s designed to run wherever a researcher’s data already lives — a laptop, a lab’s Linux machine, or an HPC login node — coordinating literature review, data analysis, and computing jobs through a single conversational interface. According to Anthropic, the app also includes a background “reviewer” agent that checks citations and calculations for errors.

BioNeMo is one of several things Claude Science can plug into. The app also connects to more than 60 scientific databases and to outside partners like LatchBio and Helix for specialized data access.

How the Integration Actually Works

Getting BioNeMo running inside Claude Science isn’t a single toggle. Per NVIDIA’s tutorial, a user opens Claude Science’s Customize menu, selects Compute, then NVIDIA BioNeMo NIM, and connects. From there, they import BioNeMo’s agent skills from GitHub, add an NVIDIA API key, and set up an endpoint — either a locally hosted Docker container using their own GPU, or a connection to a remote GPU over SSH, an HPC cluster, or NVIDIA’s Modal-based cloud compute.

The tutorial’s own hardware requirements are a useful reality check: it calls for a workstation or cloud machine with an NVIDIA L40S or H100 GPU, and roughly 700 gigabytes of storage. Most of that is the sequence database used for building alignments — about 490 gigabytes for a partial version of the UniRef30 database — with the OpenFold3 and Boltz-2 model containers adding another 30 to 40 gigabytes.

Once the endpoints are running, the researcher doesn’t need to operate each tool manually. They describe the task in plain language, and Claude Science’s agent handles discovering the right BioNeMo skill, formatting the request, and running it.

The Demonstration: Folding a Protein With and Without Context

The tutorial’s central example is designed to answer a specific scientific question: does modeling a protein alongside a suspected partner change its predicted shape?

The workflow runs in three stages. First, the agent builds sequence alignments — both for each protein individually and for the pair together — using NVIDIA’s GPU-accelerated MSA-Search tool. Second, it feeds those alignments into OpenFold3 to predict the structure of Seh1 alone and paired with C1HCX1. Third, it repeats the same two predictions using a second, independently built model, Boltz-2, so the results can be cross-checked.

The headline finding is about how much that alignment step matters. With the alignment data included, both models found strong evidence of an interaction between the two proteins — a confidence score called interface pTM (iPTM) of 0.85 for OpenFold3 and 0.82 for Boltz-2. Without it, those same scores collapsed to 0.14 and 0.19. NVIDIA reports the result held up consistently across five repeated samples per condition, and that increasing the sampling budget didn’t change the outcome — suggesting the alignment data itself, not simply more computation, is what drives the result.

Source: NVIDIA developer blog, August 31, 2026. Figures come from a single company-reported case study (five samples per condition) and have not been independently replicated.

Structurally, both models independently predicted that the same handful of C1HCX1 strands would attach to the same region of Seh1’s folded structure — a detail that echoes a related, previously published observation in an external scientific paper on protein-structure databases that NVIDIA cites as the inspiration for the experiment. NVIDIA is careful to note that this consistency across two different, independently designed AI models is what makes the result interesting, not proof that the interaction is confirmed. As the company puts it, the two models converging “creates a compelling hypothesis rather than proof of binding.” Confirming that biology would require laboratory validation, which hasn’t been reported.

Performance Claims: What’s Verified and What Isn’t

It’s worth separating two different kinds of evidence in this story.

The specific numbers from the Seh1 case study — the iPTM scores, the structural comparisons — come directly from NVIDIA’s own single demonstration. They’re detailed and specific, but they represent one worked example, not an independently repeated or peer-reviewed study.

Separately, when NVIDIA first announced the BioNeMo Agent Toolkit in June, it cited internal benchmarks claiming that BioNeMo’s skills raised AI agents’ task-completion accuracy from 60 percent to 100 percent, and roughly doubled how efficiently they used computing tokens. That, too, is a company-reported figure; NVIDIA hasn’t published the underlying benchmark methodology in public detail.

There is at least one piece of genuinely independent evidence. An independent technical benchmark published by K-Dense tested BioNeMo’s skills using Claude Code (a different Anthropic product, not Claude Science) across several Claude models. It found that giving an agent access to BioNeMo’s skills substantially improved its ability to find the correct technical steps for a task — for example, correctly identifying an API’s exact endpoint path five times out of five attempts with the skill available, compared to two out of five without it.

One caveat deserves particular attention. NVIDIA’s own GitHub repository for the BioNeMo Agent Toolkit — separate from its polished blog post — notes that the hosted OpenFold3 endpoint, along with the combined MSA-Search-to-OpenFold3 workflow, has experienced timeouts in testing and “should not yet be used for pass/fail or model comparison.” That’s a reliability flag from NVIDIA’s own engineering documentation, sitting somewhat in tension with the confident, detailed OpenFold3 results presented in the tutorial itself.

Still in Beta, With No General-Availability Date

Both halves of this integration remain works in progress. Claude Science is explicitly labeled a beta product, and Anthropic says it will “continue to refine the platform” based on user feedback. The BioNeMo connection inherits that same status. NVIDIA’s underlying BioNeMo Agent Toolkit — the open-source code and skill library on GitHub — is generally available and free to use, but the specific pathway that lets it run inside Claude Science has no announced date for moving out of beta.

Not an Anthropic Exclusive

It’s tempting to read this as an NVIDIA-Anthropic partnership story, but the evidence doesn’t support that framing. NVIDIA’s own June announcement named OpenAI, alongside Anthropic, as a frontier lab integrating the BioNeMo Agent Toolkit. Dozens of other companies — from drug discovery software makers like Schrödinger and Dassault Systèmes to lab automation firms like Thermo Fisher and Tecan — are also listed as building on the same toolkit.

The wider competitive landscape is moving quickly on multiple fronts. Google DeepMind, whose AlphaFold system won the 2024 Nobel Prize in Chemistry for solving the protein structure problem, launched its own “Science Skills” bundle at Google I/O in June 2026, connecting tools like AlphaFold Database and AlphaGenome to its Gemini-based agent platforms. Commercial-grade access to DeepMind’s newest folding model, AlphaFold 3, largely flows through Isomorphic Labs, a DeepMind spinout with its own pharmaceutical partnerships. Notably, a technology news report from late July 2026 said DeepMind had reorganized the team originally responsible for AlphaFold, moving researchers onto Gemini-focused work — though the AlphaFold database and models themselves reportedly remain available. That report comes from a single outlet and has not been independently corroborated elsewhere in this reporting.

Microsoft’s life-sciences AI work is generally built around its Azure AI Foundry platform, an approach industry analysts describe as lower-effort for customers already committed to Azure infrastructure, and higher-effort for those who aren’t.

In short: this is one entry in a broader, fast-moving race among AI labs and cloud providers to make agentic tools useful for biology, not a two-company exclusive.

What It Costs to Run

Neither NVIDIA nor Anthropic has published a specific price tag for running this workflow beyond the hardware requirements already noted above.

For general context, NVIDIA’s NIM microservices are free for prototyping through NVIDIA’s developer program, subject to rate limits of roughly 40 requests per minute according to third-party trackers. Production use is expected to require an NVIDIA AI Enterprise license, which third-party aggregators put at around $4,500 per GPU per year, or roughly $1 per GPU-hour on cloud infrastructure. These are general licensing figures, not a cost estimate for this specific workflow, and financial or contractual terms between NVIDIA and Anthropic have not been publicly disclosed by either company.

Limitations and Open Questions

A handful of caveats are worth keeping in mind. The biological interaction predicted in NVIDIA’s demonstration hasn’t been experimentally confirmed; it remains a computational hypothesis. The software on both sides of the integration is in beta, meaning reliability and features are still likely to change. Confidence scores from different folding models aren’t directly comparable to each other — OpenFold3 and Boltz-2 calculate their overall confidence figures differently, something NVIDIA flags explicitly in its own documentation. And it’s not yet clear how widely researchers are actually using this specific BioNeMo-in-Claude-Science pathway, as opposed to Claude Science more broadly or the BioNeMo toolkit through other AI agents.

Where This Leaves Things

NVIDIA’s case study demonstrates a real, working pipeline and a measurable effect from including evolutionary alignment data — but it remains one vendor-reported example, not independently replicated science, and the predicted Seh1–C1HCX1 interaction is still a hypothesis, not a confirmed finding.

Worth watching next: whether this integration moves out of beta, whether outside researchers publish results using it, and how competing efforts from Google DeepMind, Microsoft, and OpenAI’s own BioNeMo integration develop from here.

Frequently Asked Questions

Is Claude Science a new Claude AI model? No. Anthropic describes it as a beta application, not a model — it uses the same underlying Claude models included in a user’s existing plan. What’s new is the surrounding toolset: database connections, lab compute management, and now, BioNeMo integration.

What is the NVIDIA BioNeMo Agent Toolkit? It’s an open-source collection of AI-agent-callable tools, announced by NVIDIA in June 2026, that packages more than a decade of NVIDIA’s life-sciences models and software into skills that AI agents — not just Claude Science — can use for tasks like protein folding, molecular docking, and genomic analysis.

Which NVIDIA BioNeMo models work with Claude Science? Anthropic’s own product materials name Evo 2, Boltz-2, and OpenFold3 specifically. NVIDIA’s August 31 tutorial demonstrates three tools in combination: MSA-Search, OpenFold3, and Boltz-2. The broader BioNeMo toolkit includes additional models, such as OpenFold2, ProteinMPNN, RFdiffusion, and DiffDock, usable by AI agents generally.

Do I need a specific GPU to run BioNeMo NIMs in Claude Science? NVIDIA’s tutorial specifies an L40S or H100 GPU and about 700 gigabytes of storage for the full pipeline, largely due to the size of the sequence databases used for alignment. Researchers without local access to that hardware can connect to a remote GPU over SSH, an HPC cluster, or cloud compute.

Is this integration available to everyone right now? Claude Science is in public beta on macOS and Linux for Anthropic’s Pro, Max, Team, and Enterprise plans, with the BioNeMo connection available as part of that beta. Neither company has announced a date for general availability.

Is this integration exclusive to Anthropic? No. NVIDIA’s own announcement named OpenAI, alongside Anthropic, as a frontier AI lab integrating the BioNeMo Agent Toolkit, and dozens of other software and life-sciences companies are also listed as adopters of the same open toolkit.

Has the protein interaction from NVIDIA’s demonstration been confirmed in a lab? No. NVIDIA describes the predicted interaction between Seh1 and its proposed partner protein as a hypothesis supported by two independent AI models agreeing with each other — not as an experimentally validated result.