What happened
NVIDIA says it integrated the BioNeMo Agent Toolkit with Anthropic’s Claude Science so agents can discover and call local or remote BioNeMo NIM microservices. Its technical post demonstrates a workflow that retrieves protein sequences, generates evolutionary alignments, runs OpenFold3 and Boltz-2 predictions, and preserves the resulting artifacts for review.
NVIDIA’s Aug. 31 technical post describes BioNeMo Agent Toolkit as a package of life-sciences models, libraries, and workflows that can be exposed as agent-callable skills. The company says the toolkit covers biology, chemistry, genomics, and drug discovery, and is designed to work with different agent frameworks.
In the demonstrated setup, Claude Science orchestrates three NVIDIA NIM microservices: MSA Search for evolutionary alignments, OpenFold3 for structure prediction, and Boltz-2 for a second set of predictions. NVIDIA says internal benchmarks raised task correctness from 60% to 100% and roughly doubled token efficiency; those figures are company-reported and are not independently tested in the post. The tutorial requires Claude Science, an NVIDIA API key, and access to a workstation or cloud machine with an NVIDIA L40S or H100 GPU. NVIDIA estimates about 700 GB of storage for the workflow, including roughly 490 GB for a UniRef30-only database and 30–40 GB for the Boltz-2 and OpenFold3 containers. The post also describes alternatives in which Claude Science runs on a laptop while connecting to a remote GPU through SSH, HPC infrastructure, or cloud computing.
The tutorial uses local Docker containers connected to the host GPU and requires health checks before the endpoints are approved. For its biological example, the workflow compares the Seh1 protein C1GY11 from Paracoccidioides lutzii by itself and with a proposed partner, C1HCX1. NVIDIA reports that the agent retrieved sequences of 384 and 976 residues, respectively, and found 202 sequences for each protein during MSA searches. It created unpaired alignments for the individual chains and a species-paired alignment for the two-chain case. The workflow then ran monomer and heteromer predictions with no templates, ligands, pockets, or other structural constraints, using mmCIF output and retaining requests, responses, structures, metrics, and errors.
The post reports that OpenFold3 produced an interface pTM, or iPTM, of 0.86 for the heteromer with MSA input, while Boltz-2 produced 0.82. In additional five-sample comparisons, the reported MSA-versus-no-MSA values were 0.85 versus 0.14 for OpenFold3 and 0.82 versus 0.19 for Boltz-2. NVIDIA says the samples clustered tightly, with standard deviations no greater than 0.006, and that more generous sampling budgets did not materially change the pattern. The two model families use different architectures and handle MSA pairing differently, but their MSA-supported interface scores were within 0.03 in the comparison described.
Source details: developer.nvidia.com ↗
Why it matters
The example shows how an AI research agent can connect specialized biological models into a reproducible workflow rather than merely summarize scientific literature. It also reports that evolutionary alignment was decisive for the predicted protein interface in this test, while making clear that model agreement is not proof of a biological interaction.
The practical significance is the orchestration layer. Protein folding and complex prediction require specialized inputs, model-specific formats, and parameters that a general-purpose agent may not know how to select or apply. In NVIDIA’s account, Claude Science identifies the needed services, prepares the sequences and alignments, starts the endpoints, submits the predictions, and records the artifacts. That makes the AI agent an operator of a scientific pipeline, not simply a conversational interface to a single model.
The results also illustrate why input quality can matter as much as model selection. In NVIDIA’s test, removing the MSA caused the predicted heteromer interface confidence to fall sharply in both OpenFold3 and Boltz-2. OpenFold3’s reported monomer pLDDT fell from 82.3 with MSA to 36 without it, while Boltz-2’s corresponding value moved from 0.79 to 0.73 on its 0–1 scale. The comparison suggests that evolutionary information was especially important for placing the contact between the two chains, although these are confidence outputs from prediction services rather than measurements of a real biological complex.
The structural comparison adds a more specific hypothesis. NVIDIA says both models placed a group of C1HCX1 beta strands near the open edge of Seh1’s WD40 beta propeller, at a position associated with closing the final blade. Superposing the Seh1 cores reportedly produced C-alpha RMSDs of 0.68 angstrom for OpenFold3 and 0.65 angstrom for Boltz-2 over most of the chain. The company interprets this as the partner completing an existing fold rather than substantially remodeling it. Agreement between independent computational models makes the proposed geometry more useful as a target for investigation, but it does not establish that the proteins bind in living cells or perform the proposed function together.
What to watch next
The main open questions are whether the workflow generalizes beyond one protein pair, how much it costs and how long it takes to run, and whether laboratory experiments confirm the predicted association. NVIDIA reports internal toolkit benchmarks and model confidence scores, but the post does not provide independent replication, runtime measurements, or experimental validation.
The clearest limitation is biological validation. NVIDIA explicitly says the workflow predicts a C1GY11–C1HCX1 interaction and that the result has not been experimentally verified. The example is motivated by a prior structural observation involving Seh1 and a Mio-family partner, but this post itself reports a computational analysis.
Follow-up work would need laboratory evidence, such as biochemical or cellular tests, to determine whether the predicted association exists and whether it has a biological role. Generalization is also unresolved. The demonstration centers on one protein pair from one fungal species, one UniRef30 database profile, and specified search and sampling settings. The reported confidence changes show that MSAs were load-bearing in this case, but they do not establish that every protein complex benefits to the same degree or that a high interface score reliably predicts binding across different proteins.
NVIDIA’s internal task-correctness and token-efficiency figures likewise lack the benchmark definition, test set, comparison baseline, and independent evaluation needed to assess their broader meaning. Operational details will determine whether this is usable outside a demonstration. The post requires substantial local storage, compatible NVIDIA GPUs, container downloads, an API key, and manual approval of endpoints.
It says runtime metrics were present but empty for the reported runs, so readers cannot infer latency, energy use, cloud cost, or throughput. Boltz-2 and OpenFold3 also report confidence values on different scales and expose different fields, which limits direct numerical comparison. The next meaningful evidence would be reproducible runs by outside researchers, published resource measurements, broader protein-complex tests, and experimental confirmation of the structural hypothesis.