On August 31, 2026, NVIDIA published a technical blog post demonstrating how the BioNeMo Agent Toolkit works inside Anthropic's Claude Science workbench to orchestrate protein structure prediction using multiple-sequence alignment and multiple folding models. The tutorial walks through a concrete workflow: retrieving sequences for the Seh1 protein (C1GY11) and a predicted Mio-family partner (C1HCX1) from Paracoccidioides lutzii, generating paired and unpaired alignments with the MSA Search NIM, and then predicting both monomer and heteromer structures with OpenFold3 and Boltz-2 NIM microservices.

The benchmark results show that evolutionary alignment is the load-bearing input for interface prediction. With MSA input, both OpenFold3 and Boltz-2 achieved high interface confidence — iPTM scores of 0.85 and 0.82 respectively for the Seh1–C1HCX1 heteromer. Without MSA, those scores collapsed to 0.14 and 0.19. Structural superposition indicated the Seh1 fold is completed rather than remodeled when the partner is present, with both models independently placing the same C1HCX1 β-strands at the WD40 velcro-closure position, reproducing observations from Han, Tsenkov, Venanzi et al. (bioRxiv, DOI: 10.64898/2026.03.27.714458).

What's new

  • BioNeMo Agent Toolkit packages over a decade of BioNeMo life sciences models, libraries, and workflows into agent-callable skills for biology, chemistry, genomics, and drug discovery. On internal benchmarks, BioNeMo skills raise task correctness from 60% to 100% and roughly double token efficiency.
  • Claude Science integration enables agents to discover, launch, and call BioNeMo NIM microservices directly. The tutorial requires a workstation or cloud machine with an NVIDIA L40S or H100 GPU and about 700 GB of storage (the msa-search NIM's UniRef30 database is ~490 GB using the UniRef30-only profile).
  • Three NIM microservices are orchestrated: msa-search (GPU-accelerated MMseqs2 over UniRef30), OpenFold3, and Boltz-2. Each runs as a Docker container exposing a local GPU endpoint.
  • Two folding models compared on the same Seh1 monomer and Seh1–C1HCX1 heteromer predictions. OpenFold3 uses a separate paired_msa field; Boltz-2 pairs internally via optional concatenate_msas.
  • Confidence metrics reported as returned by each NIM: OpenFold3 exposes confidence_score, complex_plddt_score, complex_pde_score, ptm_score, iptm_score; Boltz-2 returns a richer set including full PAE matrices (up to 1360×1360 for the complex).

Why it matters

The integration moves agentic AI from software engineering into computational biology, where workflows demand domain-specific tool orchestration. A general-purpose agent may recognize that protein folding is needed but not which model to run, how to format requests, or which parameters matter. The BioNeMo Agent Toolkit closes that gap by encoding model-specific knowledge into skills that any agent framework can call. The benchmark also reinforces a practical lesson for structure prediction: paired MSA input is not optional for complex interfaces — without it, interface confidence collapses even when monomer folds remain plausible.

Our take

The tutorial is a reproducible, end-to-end example of an AI scientist executing a multi-step structural biology workflow — retrieving sequences, generating evolutionary context, running two independent folders, and comparing confidence outputs. The hardware requirement (L40S or H100 plus ~700 GB storage) and the UniRef30 database download mean this is not a lightweight demo; it targets research teams with GPU infrastructure. The real signal is that NVIDIA has packaged its BioNeMo stack as composable, agent-callable skills rather than monolithic pipelines, making it portable across agent frameworks (Claude Science, Baseten, and others).

Sources