NVIDIA BioNeMo Agent Toolkit Brings Accelerated AI to Life Sciences Researchers in Claude Science
NVIDIA Anthony Costa ● Covered by 3 sources
Anthropic's new Claude Science workbench now plugs into NVIDIA's BioNeMo Agent Toolkit, letting AI agents run real lab workflows on GPU-accelerated science tools. That means researchers can ask in plain English for a protein design or genomic analysis and get it done in minutes instead of days.
Based on reporting by NVIDIA, Anthony Costa — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
NVIDIA has spent over a decade building an accelerated computing stack for life sciences — chips, libraries, models, microservices, the whole pipeline. This week that stack got a new front door. Anthropic launched Claude Science, a conversational research workbench, and wired it directly into NVIDIA's BioNeMo Agent Toolkit, so scientists can describe a task in natural language and have it executed on real GPU-accelerated infrastructure without touching a config file.
The mechanics are fairly elegant. BioNeMo Agent Toolkit packages NVIDIA's scientific capabilities — genomics tools, protein folding models, cheminformatics engines — as callable skills that Claude Science's domain-specialized agents can pick up and use. A researcher might type something like "design inhibitors for this cancer-causing mutation," and the system routes that through models like Evo 2, Boltz-2 and OpenFold3, executing on NVIDIA compute wherever it's deployed, then hands back results for the scientist to inspect and refine. It's an iterative loop: ask, compute, review, ask again, without ever leaving the conversation.
What makes this more than a chatbot wrapper is the speed underneath it. NVIDIA Parabricks turns genomic analysis that used to take hours into a job that finishes in minutes. RAPIDS-singlecell, built with the scverse team, shrinks a 1.3-million-cell clustering workflow from 52 minutes down to 25 seconds. And nvMolKit claims up to a 3,000x speedup on cheminformatics tasks like conformer generation and similarity search. Those numbers matter because agentic workflows only work if each tool call returns fast enough to stay inside a reasoning loop — nobody wants an AI agent sitting idle for 50 minutes between steps.
The toolkit is deliberately framework-agnostic, meaning the same scientific skills can plug into other agent platforms beyond Claude Science, and it's already available through NVIDIA's developer channels and GitHub. NVIDIA notes that 18 of the top 20 pharmaceutical companies already use BioNeMo in some form, which gives this integration a built-in audience rather than a cold start. Claude Science itself enters public beta today, with Anthropic actively asking researchers what other domain specialists they want added next.
My take — AI-written commentary, not fact-checked reporting
This is the clearest sign yet that the AI agent hype is finally colliding with infrastructure that can back it up — an agent is useless if the tools it calls take an hour to respond, and NVIDIA just spent years quietly building the plumbing for exactly this moment. I'd rather see this kind of vertical integration in science than in another chatbot feature war, because drug discovery timelines actually matter to real people, not just quarterly earnings calls.
Read more about this at: NVIDIA