- Published on
Run NVIDIA BioNeMo NIM Microservices for Protein Structure Prediction in Claude Science
- Authors
- Name
- DeepHorus
NVIDIA and Anthropic have integrated the NVIDIA BioNeMo Agent Toolkit into Claude Science, Anthropic's AI workbench for scientific research, enabling AI agents to discover, launch and call NVIDIA NIM microservices for protein structure prediction. The integration allows researchers to run multiple-sequence alignment (MSA) based folding workflows using several models and compare their results directly within Claude Science. The announcement underscores a broader shift toward agentic AI in science, where AI scientists read papers, propose hypotheses, call models and decide which experiments to prioritize next. Coding agents proved this pattern first by writing, testing and shipping production code, and scientific research now demands a similar but more iterative approach.
Scientific problems frequently require domain-specific tools, such as folding a protein or characterizing a molecule, but orchestrating these tools is difficult because similar packages may have vastly different environment requirements or APIs. A general-purpose agent may know that a task calls for protein folding, yet not which model to run, how to format the request, or which parameters matter. The BioNeMo Agent Toolkit closes that gap by packaging more than a decade of NVIDIA BioNeMo life sciences models, libraries and workflows into agent-callable skills spanning biology, chemistry, genomics and drug discovery. Built to run with any agent framework, the toolkit enables complex scientific workflows using specialized domain expertise.
On internal benchmarks, BioNeMo skills raise task correctness from 60% to 100% and roughly double token efficiency, highlighting measurable gains for research teams. The setup requires a machine with an NVIDIA L40S GPU or NVIDIA H100 GPU with Claude Science installed, though users can also run on a laptop and connect to a remote GPU via SSH, HPC, or cloud computing on Modal. Storage demands are substantial: roughly 700 GB, including the msa-search NIM's UniRef30 database at about 490 GB when using the UniRef30-only profile instead of the full 1.4 TB set, plus 30-40 GB of Boltz-2 and OpenFold3 containers.
By default, Claude Science runs in a sandbox, but running BioNeMo NIM microservices requires compute endpoints exposing local or remote GPU resources. Users navigate to Customize > Compute > NVIDIA BioNeMo NIM > Connect, import the BioNeMo Agent Toolkit skills from GitHub, add an NVIDIA API key and connect to local Docker endpoints that use the host GPU. After importing skills and configuring connections, users start a new project and prompt Claude to create the required NIM microservice endpoints. Containers for msa-search, openfold3 and boltz2 are downloaded and launched, then smoke-tested to confirm health before both models are approved.
With the three NIM endpoints running, Claude Science orchestrates the structure-prediction workflow. The tutorial demonstrates the pipeline on the Seh1 protein and a predicted Mio-family partner from Paracoccidioides lutzii, a fungus that causes paracoccidioidomycosis. The example is motivated by Figure 4e of Han, Tsenkov, Venanzi et al., "AlphaFold Database expands to proteome-scale quaternary structure predictions," grounding the demo in published research.
For researchers, the workflow means domain tools are callable by agents rather than manually configured, reducing environment friction and improving consistency across runs. Comparing results from multiple folding models also strengthens confidence in predicted structures before wet-lab validation. Looking ahead, the collaboration between NVIDIA and Anthropic suggests that agent-driven, model-comparative pipelines could become standard practice in structural biology and drug discovery.
