Anthropic links Claude Science to NVIDIA BioNeMo tools

Anthropic has integrated the NVIDIA BioNeMo Agent Toolkit into its Claude Science research workbench, bringing NVIDIA's life sciences tools into Anthropic's new public beta product for scientific research.

Claude Science lets researchers use natural language to direct software agents through scientific workflows. With the integration, those agents can use NVIDIA models, libraries and NIM microservices in the same environment, with access to NVIDIA compute resources deployed across different locations.

The deal targets a growing market for AI tools in drug discovery, genomics, medical imaging, molecular design and protein engineering. NVIDIA says 18 of the top 20 pharmaceutical companies already use BioNeMo, suggesting the software already has a broad foothold in large-scale life sciences research.

How it works

Under the setup, a scientist can describe a task such as analysing a genomic sequence, predicting a protein structure or designing a potential binder in plain language. Claude Science interprets the request and assigns the work to domain-specific agents built around established workflows in genomics, proteomics, single-cell analysis, cheminformatics and clinical research.

BioNeMo Agent Toolkit serves as the layer connecting those agents to NVIDIA's scientific software. It packages functions as callable skills and includes information on each tool's purpose and required inputs, allowing an agent to choose a suitable tool, prepare inputs and run the workflow.

Researchers therefore do not need to manually configure models, endpoints or software environments for each step. Instead, they can review outputs, refine questions and decide what to do next within a single interface.

NVIDIA highlighted several tools available through the toolkit, including Parabricks for genomic analysis, RAPIDS-singlecell for large-scale single-cell workflows, nvMolKit for cheminformatics tasks such as similarity search and conformer generation, and BioNeMo open models for biomolecular work.

It also pointed to BioNeMo NIM microservices, which package models as inference endpoints that can be called through an application programming interface. In practice, this gives AI agents a standard way to use scientific models in production research settings.

Drug discovery focus

One example described by the companies centres on designing inhibitors for common cancer targets. In that workflow, a researcher starts with a known cancer-linked antigen mutation and asks Claude to generate potential inhibitors, while the integrated NVIDIA tools handle prediction, optimisation and validation at scale.

The broader aim is to move life sciences research toward what both companies describe as an agentic model, in which software agents can reason through tasks, plan actions and call tools with limited manual intervention. In scientific research, that approach depends not just on language models but also on access to specialist computational tools that can handle large, complex datasets.

NVIDIA argues that this is where accelerated computing remains central. A scientific agent may need to fingerprint compound libraries, cluster likely hits, generate conformers, analyse genomic context and compare perturbation responses before suggesting the next experiment, and each step depends on the performance of the underlying software.

Open toolkit

BioNeMo Agent Toolkit is being offered as an open, harness-agnostic framework, meaning the same skills can be used across different agent systems and research platforms. That approach could help NVIDIA extend BioNeMo's reach beyond its own software ecosystem as labs and pharmaceutical groups test different AI platforms for scientific work.

For Anthropic, the integration adds domain-specific infrastructure to Claude Science at launch. Rather than asking researchers to connect specialist life sciences tools themselves, the workbench starts with preconfigured access to a set of established NVIDIA resources.

The announcement reflects a broader shift in scientific software, as developers try to combine conversational AI interfaces with workflow automation and specialised compute back ends. The commercial opportunity is significant, particularly in pharmaceutical research, where companies are looking for ways to shorten discovery timelines and reduce the cost of early-stage experimental work.

NVIDIA says Parabricks can cut genomic analysis from hours to minutes, while RAPIDS-singlecell reduces a 1.3 million-cell preprocessing and clustering workflow from 52 minutes to 25 seconds. It also says nvMolKit can speed up some cheminformatics operations by up to 3,000 times.

Claude Science is entering public beta, while the BioNeMo Agent Toolkit and its associated skills are already available through NVIDIA's developer resources and GitHub.