Why was Multiple-Sequence Alignment (MSA) so critical in the protein structure prediction benchmark?

The benchmark demonstrated that MSA, which provides evolutionary context by aligning related protein sequences, is the 'load-bearing input' for interface prediction. When modeling the Seh1 protein complex, both OpenFold3 and Boltz-2 achieved high interface confidence scores (iPTM of 0.85 and 0.82) with MSA input. Without it, those scores collapsed to 0.14 and 0.19, proving MSA is essential for accurately predicting how the two proteins interact.

Run NVIDIA BioNeMo NIM Microservices for Protein Structure Prediction in Claude Science

NVIDIA and Anthropic Enable AI Agents to Run Complex Biology Workflows

NVIDIA and Anthropic have integrated the NVIDIA BioNeMo Agent Toolkit with the Claude Science AI workbench, enabling AI agents to directly discover and call specialized microservices for complex life sciences research. The collaboration showcases a concrete workflow for agent-driven protein structure prediction, a process that traditionally requires significant domain expertise. By packaging sophisticated bioinformatics models into agent-callable skills, this integration addresses a critical gap between general-purpose AI agents and the specialized tools needed for scientific discovery.

The companies demonstrated a workflow where an AI agent within Claude Science orchestrates multiple NVIDIA NIM microservices to predict the structure of a protein complex from the fungus Paracoccidioides lutzii. This multi-stage process involved generating multiple-sequence alignments (MSA) and then passing them to two independent folding models, OpenFold3 and Boltz-2, to compare results. The benchmark underscored the necessity of domain-specific inputs, revealing that interface prediction confidence collapsed without MSA data, dropping from iPTM scores above 0.8 to below 0.2.

Platform: Anthropic Claude Science AI workbench for agent orchestration.
Toolkit: NVIDIA BioNeMo Agent Toolkit provides skills for biology, chemistry, and genomics.
Microservices: NVIDIA NIM endpoints for MSA Search, OpenFold3, and Boltz-2.
Hardware Requirements: A workstation or cloud instance with an NVIDIA L40S or H100 GPU.
Storage Requirements: Approximately 700 GB for databases and model containers.

This integration signals a move toward making agentic AI more practical for real-world scientific applications. By abstracting the complexity of running specialized models into containerized NIM microservices, NVIDIA is building an ecosystem that lowers the barrier for automated research. This model, where a generalist agent can orchestrate highly specific computational experiments, could significantly accelerate discovery cycles in fields like drug design and genomics by automating the iterative, and often tedious, process of evaluating hypotheses and analyzing model outputs.

The integration of NVIDIA BioNeMo NIMs into Claude Science is less about a single technical achievement and more about establishing a blueprint for the future of scientific research. It positions NVIDIA's full stack—from silicon to NIM microservices—as the essential infrastructure for AI agents performing specialized, high-value work in the life sciences, effectively building a deep competitive moat in this emerging market.

>> Verify Original Transmission at NVIDIA