Agentic AI is altering how analysis is finished. AI scientists can learn papers, suggest hypotheses, name fashions, and decide which experiments to prioritize subsequent. First proving their worth in software program engineering, coding brokers now write, check, and ship manufacturing code. Scientific analysis will be extra demanding and iterative. Researchers frequently consider proof, refine hypotheses, and run experiments that form subsequent choices. Even failed experiments can result in surprising insights.
Scientific issues additionally usually require domain-specific instruments, for instance, to fold a protein or characterize a molecule. Orchestrating and working these instruments will be difficult. Comparable packages might have vastly totally different atmosphere necessities or APIs. A general-purpose agent would possibly acknowledge {that a} process requires protein folding or molecular docking, however not which mannequin to run, easy methods to format the request, or which parameters matter.
NVIDIA BioNeMo Agent Toolkit closes that hole. The toolkit packages greater than a decade of NVIDIA BioNeMo life sciences fashions, libraries, and workflows into agent-callable abilities for biology, chemistry, genomics, and drug discovery.
Constructed to run with any agent framework, it permits complicated scientific workflows utilizing specialised area experience. On inner benchmarks, BioNeMo abilities elevate process correctness from 60% to 100% and roughly double token effectivity.
This put up makes use of Claude Science, Anthropic’s AI workbench for scientific analysis, with the NVIDIA BioNeMo Agent Toolkit and NVIDIA NIM to run protein construction prediction with multiple-sequence alignment (MSA) utilizing a number of folding fashions and examine outcomes.
Setup
NVIDIA and Anthropic labored collectively to combine the BioNeMo Agent Toolkit into Claude Science, enabling brokers to find, launch, and name BioNeMo NIM microservices straight.
Claude Science runs in varied configurations based mostly on the place your GPU is situated and your safety necessities. This tutorial focuses on working the platform on a machine with a GPU. Test the product documentation for GPU compatibility data. It’s additionally potential to run on a laptop computer and hook up with a distant GPU utilizing SSH, HPC, or cloud computing on Modal. For this tutorial, you want entry to a workstation or cloud machine with an NVIDIA L40S GPU or NVIDIA H100 GPU and Claude Science put in. The machine wants about 700 GB of storage for this workflow: the msa-search NIM’s UniRef30 database is ~490 GB (utilizing the UniRef30-only profile as a substitute of the total 1.4 TB set), and the Boltz-2 and OpenFold3 containers are 30–40 GB whole.
By default, Claude Science runs in a sandbox. Working BioNeMo NIM microservices requires compute endpoints that expose native or distant GPU sources. In Claude Science, choose Customise > Compute > NVIDIA BioNeMo NIM > Join. From there, import the BioNeMo Agent Toolkit abilities from GitHub, add an NVIDIA API key, and hook up with native endpoints—Docker containers that use the host GPU.
After importing the abilities, storing the API key, and configuring the native connection, begin a brand new mission and session. Then immediate Claude to create the required NIM microservice endpoints.
Obtain and launch the native BioNeMo NIM containers for the msa-search, openfold3 and boltz2 microservices. For msa-search, obtain solely the UniRef30 database so it downloads solely ~490 GB as a substitute of the total 1.4 TB. Run a smoke check on every to verify they’re wholesome.
When prompted, choose Approve endpoint for each fashions. Establishing these endpoints requires downloading containers for every microservice.
Steps
With the three NIM endpoints working, Claude Science orchestrates the structure-prediction workflow. To make the pipeline concrete, this tutorial examines the Seh1 protein and a predicted Mio-family companion from Paracoccidioides lutzii, a fungus that causes paracoccidioidomycosis. The instance is motivated by Determine 4e of Han, Tsenkov, Venanzi et al., “AlphaFold Database expands to proteome-scale quaternary buildings” (bioRxiv, DOI: 10.64898/2026.03.27.714458).
Seh1 and Mio-family proteins take part in conserved nutrient-sensing equipment, and buildings from different organisms recommend that Mio can full Seh1’s in any other case open β-propeller, making this pair a helpful check of MSA-supported complicated prediction.
This tutorial makes use of a two-protein system from Paracoccidioides lutzii: the nuclear-pore protein Seh1 (C1GY11) and an uncharacterized proposed companion (C1HCX1). The instance asks the structural query: How does the anticipated construction of Seh1 differ when it’s modeled alone versus with the proposed companion?
To analyze this query, Claude Science builds evolutionary context for the goal sequences utilizing MSA. It then passes these alignments to 2 impartial folding fashions, preserving the inputs and outputs so the single-chain and two-chain predictions will be examined inside every mannequin.
The workflow has three phases:
Generate MSAs for single-chain and species-paired sequences with the GPU-accelerated MSA Search NIM.
Predict construction for a single chain and a two-chain complicated with the OpenFold3 NIM.
Repeat the predictions with the Boltz-2 NIM and consider every mannequin independently.
Stage 1: Generate MSAs
The agent retrieves each protein sequences from UniProt and creates two kinds of alignment. It first creates an unpaired alignment for every protein to offer details about every chain’s construction. It then creates a paired alignment by matching associated variations of the 2 proteins present in the identical species. If adjustments in a single protein constantly correspond to adjustments within the different, that sample may help the fashions predict the place the proteins would possibly work together.
On this run, the search returned 202 sequences for every protein. The agent recorded the supply and sequence for Seh1 (384 residues) and C1HCX1 (976 residues), together with checksums for verification. It didn’t trim both sequence or present structural templates, ligands, or different constraints.
Use this immediate:
I’m evaluating the Seh1 protein C1GY11 by itself and with C1HCX1. Load the msa-search-nim talent and obtain each present FASTA sequences from UniProt. Save the place every sequence got here from, whenever you downloaded it, its accession, size, checksum, and chain ID: use A for C1GY11 and B for C1HCX1.
Create an A3M alignment for every protein with the MSA Search NIM. Use Uniref30_2302 solely, the ColabFold search kind, an E-value of 0.0001, and as much as 500 sequences. Then create a paired A3M for the 2 proteins, in A-then-B order, utilizing grasping pairing.
Save the request, response, A3M information, endpoint, metrics, and any errors. If you happen to can’t make the paired alignment, cease and inform me; don’t quietly change it with a special sort of alignment.
Stage 2: Predict construction with OpenFold3
OpenFold3 takes a molecule record with per-chain msa (unpaired) and, for complexes, paired_msa. The agent runs two circumstances with the identical Seh1 sequence: the monomer (chain A with its MSA) and the heteromer (chains A and B with their MSAs plus the paired MSA). Use mmCIF output, no templates, and retailer each returned pattern and the arrogance fields the service truly exposes.
Use this immediate:
Load the openfold3-nim talent. Use the saved alignments to make two OpenFold3 predictions:
1. C1GY11 by itself as chain A.
2. C1GY11 as chain A with C1HCX1 as chain B.
Use the paired alignment for the two-protein prediction. Return mmCIF information. Don’t use templates, ligands, pockets, or different constraints.
Save the request, response, buildings, runtime particulars, and each confidence rating that the service returns. These scores present mannequin confidence solely; don’t say they show the proteins work together. If paired MSA enter is unavailable, save the error and cease as a substitute of continuous with an unpaired prediction.
The response uncovered confidence_score, complex_plddt_score, complex_pde_score, ptm_score, iptm_score (plus format, title, supply); runtime_metrics was current however empty.
The monomer’s iptm_score is 0 by building, and OpenFold3’s composite confidence_score weights the interface closely — which is why a well-folded monomer (pTM 0.82, pLDDT 82) nonetheless scores low on the composite. Learn confidence_score inside OpenFold3, not throughout fashions.
Stage 3: Repeat the predictions with Boltz-2
The identical two circumstances run by Boltz-2 with the identical chain IDs and no-template coverage. One architectural distinction issues: Boltz-2 has no separate paired_msa area. It takes one MSA per chain and pairs internally (non-obligatory top-level concatenate_msas), so every Boltz-2 chain acquired its per-chain A3M.
Use this immediate:
Load the boltz2-nim talent and repeat the identical two predictions with Boltz-2: C1GY11 alone as chain A, then C1GY11 plus C1HCX1 as chains A and B. Use the saved per-chain alignments within the format supported by the talent.
Don’t activate affinity prediction or use templates, ligands, pockets, or contact constraints.
Save the request, response, buildings, runtime particulars, PAE when obtainable, and the arrogance values returned by Boltz-2. If a price will not be returned, report that as a substitute of estimating it. Repeat the predictions with out MSA utilizing OpenFold3 and Boltz-2. Evaluation and analyze all the outcomes, then present a abstract.
Boltz-2 generated full PAE outputs (write_full_pae=true).
Boltz-2 returned a richer confidence set — confidence_scores, ptm_scores, iptm_scores, protein_iptm_scores, complex_plddt_scores, complex_iplddt_scores, complex_pde_scores, complex_ipde_scores, plus per-chain and pairwise arrays and full pae/pde matrices (as much as 1360×1360 for the complicated).
The one requested area the service left empty was runtime metadata (metrics: {}).
The MSA is the load-bearing enter
As a result of the workflow additionally ran single-sequence (no-MSA) predictions, the worth of Stage 1 will be measured straight. For the heteromer, the interface pTM (iPTM), the metric that reviews on the anticipated contact between the 2 chains, collapses with out an MSA in each fashions:
With MSA enter, iPTM reaches 0.85 for OpenFold3 and 0.82 for Boltz-2. With out MSA enter, it falls to 0.14 and 0.19, respectively. The 5 samples cluster tightly (SD ≤ 0.006), indicating that the distinction is constant throughout samples.
Two robustness checks help this outcome. First, working every situation with a extra beneficiant sampling funds (OpenFold3 diffusion_samples=5; Boltz-2 5 samples with six recycling steps and 200 sampling steps), the image stays unchanged. The with-MSA interface remained excessive, and the no-MSA interface remained collapsed (similar to the single-sample values inside 0.01 iPTM). Additional sampling doesn’t substitute for evolutionary enter.
Second, the 2 mannequin households (with totally different architectures and, for Boltz-2, a special MSA-pairing mechanism) land inside 0.03 iPTM of one another when given the alignment, which is reassuring cross-model settlement.
The monomer outcomes present the sort of MSA dependence every mannequin has. OpenFold3 wants the alignment even to fold the one chain (pLDDT 82 → 36 with out it); Boltz-2 folds the monomer pretty nicely from sequence alone (0.79 → 0.73) however nonetheless can’t place the interface with out an MSA. In both case, Stage 1 represents extra than simply an non-obligatory pre-processing step—it serves because the very basis of the interface speculation.
Examine the construction the workflow produced
The agent then examines whether or not the proposed companion adjustments the anticipated Seh1 construction. For every mannequin, it superposes the Seh1 chain from the monomer and heteromer and examines the WD40 β-propeller area.
[
The monomer and heteromer views use the same camera position after alignment of the Seh1 cores, showing where the partner strands occupy the open propeller edge.
The structural analysis reproduces the source paper’s Figure 4e observation in both independent models:
The Seh1 fold is completed, not remodeled. Superposing Seh1 between monomer and heteromer gives a core Cα-RMSD of 0.68 Å (OpenFold3, 313/384 residues) and 0.65 Å (Boltz-2, 307/384). The partner adds to the existing propeller; the large full-chain RMSD comes only from the ~70 residues around the blade that shift as it closes.
Both models place the same partner strands. Independently, OpenFold3 and Boltz-2 position a cluster of C1HCX1 β-strands from roughly residues 305–391 against Seh1 — consistent with the paper’s “three β-strands.”
The insertion is at the WD40 velcro-closure position. The partner contacts Seh1’s N-terminal strand (residues 1–29) and C-terminal strands (~331–384), the edge where a WD40 propeller closes its final blade.
It’s important to note that these models predict a C1GY11–C1HCX1 interaction. However, it’s not yet verified experimentally as a true biological association. In this case, two independent models converged on the same local geometry, creating a compelling hypothesis rather than proof of binding. The real value of this workflow lies in providing a reproducible, evidence-backed way to form and inspect these structural hypotheses—leaving the final word to experimental validation.
Get started
BioNeMo Agent Toolkit enables Claude Science agents to generate MSAs, run NIM-based structure predictions, and preserve artifacts for review. Explore the toolkit on GitHub to apply this workflow to other monomers and complexes.
