Using Claude Science with NVIDIA BioNeMo NIMs to run MSA-based protein folding and compare OpenFold3 and Boltz-2

Anthropic’s Claude Science was integrated with the NVIDIA BioNeMo Agent Toolkit and NVIDIA NIM microservices to run MSA-based (multiple-sequence alignment) protein structure prediction and to compare folding results across models. The example workflow requires a GPU-equipped machine and roughly 700 GB of storage. The test case compared Seh1 (UniProt C1GY11, 384 residues) modeled alone versus modeled with a predicted partner (C1HCX1, 976 residues) from the fungus Paracoccidioides lutzii.

Why this matters

Agentic AI can automate literature reading, hypothesis generation, model calls and experiment prioritization, speeding scientific cycles. Protein and complex prediction, however, requires domain tools and significant data; integrating these tools as agent-callable skills reduces friction and improves reproducibility.

System requirements and setup

• Hardware: local GPU such as NVIDIA L40S or NVIDIA H100.
• Storage: ~700 GB total. The msa-search NIM UniRef30-only database is ≈490 GB, and the Boltz-2 plus OpenFold3 containers total ~30–40 GB.
• Claude Science normally runs in a sandbox. Running BioNeMo NIM microservices requires compute endpoints exposing local/remote GPUs (Docker containers using the host GPU).

Setup highlights:

1. Import the BioNeMo Agent Toolkit skills from GitHub into Claude Science, store an NVIDIA API key and connect to local endpoints.
2. Download and launch msa-search, openfold3 and boltz2 NIM containers; for msa-search use the UniRef30-only profile to avoid the full 1.4 TB download.
3. Run smoke tests and approve endpoints in Claude Science.

Workflow overview

The workflow has three stages:

1. Generate MSAs (per-chain and species-paired A3M) using the GPU-accelerated msa-search NIM.
2. Predict structure with OpenFold3 for monomer and heteromer (mmCIF output, no templates).
3. Repeat the same predictions with Boltz-2 and evaluate each model independently.

The structural question addressed: how does Seh1’s predicted structure change when the proposed partner C1HCX1 is included?

Stage 1 — MSA generation

• The agent retrieved UniProt sequences and produced unpaired A3Ms for each chain and a paired A3M by greedy pairing related sequences from the same species. The search returned 202 sequences for each protein.
• Metadata were recorded (source, download time, accession, length, checksum, chain IDs: A = C1GY11, B = C1HCX1).
• Parameters used: Uniref30_2302-only, ColabFold search type, E-value 1e-4, up to 500 sequences. If paired alignment could not be produced, the workflow was to stop and report an error.

Stage 2 — OpenFold3 predictions

• Two OpenFold3 runs used the same Seh1 sequence: (1) Seh1 alone as chain A with its per-chain MSA, and (2) Seh1 (A) plus C1HCX1 (B) with both per-chain MSAs and the paired MSA. Outputs were mmCIF files; no templates, ligands or other constraints were used.
• OpenFold3 returned these confidence fields: confidence_score, complex_plddt_score, complex_pde_score, ptm_score, iptm_score (format, name, source were also present). runtime_metrics existed but was empty.
• Selected reported OpenFold3 values (with MSA):
  • Seh1 monomer: confidence_score 0.25, complex_plddt_score 82.3, complex_pde_score 0.58, ptm_score 0.82, iptm_score 0 (by construction for monomer).
  • Seh1 + C1HCX1: confidence_score 0.87, complex_plddt_score 78.2, complex_pde_score 1.07, ptm_score 0.79, iptm_score 0.86.
• Note: OpenFold3’s composite confidence_score weights interface predictions heavily; a well-folded monomer can still have a low composite score because it lacks an interface.

Stage 3 — Boltz-2 predictions

• The same two conditions were run with Boltz-2. Architectural difference: Boltz-2 does not accept a separate paired_msa field; it receives one MSA per chain and pairs internally (optionally concatenate_msas).
• Boltz-2 returned a richer set of confidence outputs including confidence_scores, ptm_scores, iptm_scores, protein_iptm_scores, complex_plddt_scores, complex_iplddt_scores, complex_pde_scores, complex_ipde_scores, per-chain and pairwise arrays, and full pae/pde matrices (up to 1360×1360 for the complex). runtime metadata was empty (metrics: {}).
• Selected reported Boltz-2 values (with MSA):
  • Seh1 monomer: confidence 0.79, complex_plddt 0.79 (0–1 scale), complex_pde 1.11, ptm 0.82.
  • Seh1 + C1HCX1: confidence 0.77, complex_plddt 0.75, complex_pde 1.33, ptm 0.76, iptm 0.82, protein_iptm 0.82.

The MSA is the load-bearing input

• The workflow compared MSA-enabled runs to single-sequence (no-MSA) runs. For the heteromer, interface pTM (iPTM) collapses without MSA in both models:
  • Heteromer iPTM (average over five samples, SD ≤ 0.006):
    • OpenFold3: with MSA 0.85 → without MSA 0.14 (drop 0.72).
    • Boltz-2: with MSA 0.82 → without MSA 0.19 (drop 0.63).
• Two robustness checks reinforced this conclusion:
  1. Increasing sampling (OpenFold3 diffusion_samples=5; Boltz-2: five samples with six recycling steps and 200 sampling steps) did not recover interface confidence in the no-MSA condition. Extra sampling does not substitute for evolutionary information.
  2. The two independent model families agreed within ~0.03 iPTM when given the alignment, indicating consistent cross-model signal.
• Monomer dependence: OpenFold3 required the alignment to fold the single chain well (pLDDT 82 → 36 without MSA). Boltz-2 folded the monomer more robustly from sequence alone (0.79 → 0.73) but still could not place the interface without MSA. Thus Stage 1 (MSA) is foundational for interface hypotheses.

Structural inspection

• For each model the Seh1 chain from monomer and heteromer predictions was superposed and the WD40 β-propeller region inspected.
• Both models reproduced the observation from Han et al. (AlphaFold Database expansion): the partner completes the Seh1 fold rather than remodeling it.
  • Core Cα-RMSD (monomer vs heteromer, Seh1 core): OpenFold3 0.68 Å (313/384 residues), Boltz-2 0.65 Å (307/384 residues).
  • The partner contributes additional β-strands that occupy the open edge of the propeller; the large full-chain RMSD arises from ~70 residues around the blade that shift as it closes.
  • Independently, both OpenFold3 and Boltz-2 position a cluster of C1HCX1 β-strands (approx. residues 305–391) against Seh1, consistent with the previously reported “three β-strands.”
  • The insertion is at the WD40 velcro-closure position, contacting Seh1’s N-terminal strand (residues 1–29) and C-terminal strands (~331–384).
• Caveat: these model predictions support a C1GY11–C1HCX1 interaction hypothesis but are not experimental proof of a biological association. Two independent models converging on the same local geometry is a strong, testable hypothesis; experimental validation is required to confirm binding.

Conclusions and next steps

• The NVIDIA BioNeMo Agent Toolkit made BioNeMo skills available to Claude Science agents, enabling reproducible MSA generation, NIM-based structure prediction and artifact preservation for review.
• MSA input is essential for confident interface prediction; additional sampling does not replace evolutionary information.
• Independent agreement between OpenFold3 and Boltz-2 on the local geometry provides a compelling hypothesis for experimental follow-up.

Explore the BioNeMo Agent Toolkit on GitHub to apply this workflow to other monomers and complexes.