NVIDIA brings BioNeMo microservices to Claude Science

NVIDIA has partnered with Anthropic to integrate the NVIDIA BioNeMo Agent Toolkit into Claude Science, a scientific research workbench. This integration allows AI agents to discover, launch, and call BioNeMo NIM microservices directly. To run this setup locally, researchers need an NVIDIA L40S or H100 GPU and approximately 700 GB of storage. This storage accommodates the 490 GB UniRef30 database for the msa-search NIM, alongside 30 to 40 GB for the Boltz-2 and OpenFold3 containers. On internal benchmarks, utilizing these BioNeMo skills raised task correctness from 60% to 100% while doubling token efficiency.

The system orchestrates a complex workflow to predict protein structures, demonstrated using the nuclear-pore protein Seh1 (C1GY11, 384 residues) and its proposed partner C1HCX1 (976 residues) from the fungus Paracoccidioides lutzii. The agent retrieved sequences from UniProt, generated multiple-sequence alignments (MSAs) using the GPU-accelerated MSA Search NIM with an E-value of 0.0001 for up to 500 sequences, and fed them into OpenFold3 and Boltz-2. The search returned 202 sequences for each protein.

The benchmarks highlighted the critical role of evolutionary alignment. With MSA input, OpenFold3 achieved an interface confidence (iPTM) of 0.85, and Boltz-2 reached 0.82. Without MSA, these scores collapsed to 0.14 and 0.19, respectively. For the monomer, OpenFold3's pLDDT dropped from 82 to 36 without MSA, while Boltz-2 dropped from 0.79 to 0.73. When analyzing the structural outputs, both models independently placed the C1HCX1 beta-strands (residues 305 to 391) at the WD40 velcro-closure position of Seh1, contacting its N-terminal (residues 1 to 29) and C-terminal (residues 331 to 384) regions. The core C-alpha RMSD between the monomer and heteromer was 0.68 angstroms for OpenFold3 and 0.65 angstroms for Boltz-2.

For life sciences practitioners, this integration simplifies the orchestration of disparate molecular biology tools that often have conflicting environment requirements. Instead of manually formatting requests and managing APIs, researchers can use Claude Science to automatically run multi-model pipelines. This agentic approach provides a highly reproducible, evidence-backed method to generate and inspect structural hypotheses before committing to expensive wet-lab validation.