Anthropic’s Claude Science integrates NVIDIA BioNeMo Toolkit

Anthropic has integrated NVIDIA’s BioNeMo Agent Toolkit into the public beta of Claude Science. The integration allows researchers to invoke NVIDIA-accelerated models and NIM microservices directly from the Claude environment to run life-science research workflows.

Claude Science is Anthropic’s AI workbench for scientific research. In the public beta, scientists can submit plain-language requests that the platform converts into computational steps executed across connected services and hardware.

The BioNeMo Agent Toolkit exposes high-performance compute and preconfigured scientific tools as callable skills within Claude. Domain-specialized agents receive metadata about each tool’s purpose and required inputs, select the correct function, format data, call the deployed NVIDIA service and return results for review. This removes the need for researchers to manually configure models, network endpoints or software environments.

Researchers can ask Claude to perform tasks such as analyzing a genomic sequence, predicting a protein structure or designing molecular binders. For those requests, agents can orchestrate sequences of operations that include fingerprinting compound libraries, clustering candidate molecules, generating conformers, scoring structures and comparing genomic or perturbation data.

The integration brings specific models and libraries into Claude workflows, including Evo 2, Boltz-2 and OpenFold3, and software such as Parabricks, RAPIDS-singlecell and nvMolKit. NVIDIA packages many of these models as BioNeMo NIM microservices, which are containerized inference endpoints that present a single API for production deployments. NVIDIA reports that 18 of the top 20 global pharmaceutical companies run BioNeMo in production environments.

NVIDIA reports performance improvements for several components: Parabricks can reduce genomic processing from hours to minutes; RAPIDS-singlecell cuts a 1.3 million-cell preprocessing and clustering job from about 52 minutes to 25 seconds; nvMolKit can speed tasks like similarity search and conformer generation by up to 3,000 times. Faster runtimes allow agents to include heavy computations within iterative reasoning loops instead of relegating them to offline batches.

The BioNeMo Agent Toolkit is open and harness-agnostic so the same callable skills can be reused across different agent frameworks and enterprise research platforms. Engineering teams can access the toolkit and example scientific skills through NVIDIA developer resources and GitHub. Anthropic is soliciting feedback from researchers during the public beta phase.

The integration adds NVIDIA-accelerated models and production deployment options to Claude Science for tasks in drug design, genomics and single-cell analysis.