Key Takeaways
- Without MSA, protein interface confidence collapses from 0.85/0.82 to 0.14/0.19 in OpenFold3 and Boltz-2.
- Both models independently place the C1HCX1 β-strands only when evolutionary alignment is provided.
- The BioNeMo-Claude Science workflow records every input, metric, and error for reproducible hypothesis generation.
Table of Contents
The Alignment That Decides Interface Confidence
Two independent protein folding models reached nearly identical confidence scores for a predicted protein interface when supplied with evolutionary alignment data. Remove that alignment, and both models lost the interface almost entirely.
The multiple-sequence alignment, or MSA, is the evolutionary context that tells the model how residues have varied across related proteins. In this test, the MSA proved to be the load-bearing input for interface prediction.
NVIDIA’s developer blog published the workflow on August 31, 2026, showing how the BioNeMo Agent Toolkit combines with Claude Science and NIM microservices to run multiple-sequence alignment and structure prediction. The demonstration centered on the Seh1 protein and a proposed partner from Paracoccidioides lutzii.
According to NVIDIA’s developer blog, the result was not a subtle performance shift. Interface confidence dropped from 0.85 and 0.82 with MSA input to 0.14 and 0.19 without it for OpenFold3 and Boltz-2, respectively.
Inside the BioNeMo-Claude Science Orchestration
The workflow packages NVIDIA’s life sciences models, libraries, and workflows into agent-callable skills for biology, chemistry, genomics, and drug discovery. Claude Science agents can discover, launch, and call BioNeMo NIM microservices directly from a configured compute endpoint.
The setup requires an L40S or H100 GPU and roughly 700 GB of storage. The largest dependency is the msa-search NIM’s UniRef30 database at about 490 GB when using the UniRef30-only profile rather than the full 1.4 TB set.
Three NIM endpoints carry the workload: msa-search, openfold3, and boltz2. After importing BioNeMo Agent Toolkit skills from GitHub and connecting local Docker containers to the host GPU, Claude Science can create, test, and approve each microservice endpoint.
The structural question was specific: how does Seh1 change when modeled alone versus with the proposed C1HCX1 partner? The agent first generated unpaired MSAs for each chain, then built a paired alignment by matching related versions found in the same species.
MSA Search returned 202 sequences for each protein. The agent recorded accession, length, checksum, chain ID, and source data before passing the alignments to the folding models.
OpenFold3 received per-chain MSA and paired MSA inputs. Boltz-2 received per-chain A3M files because it pairs MSAs internally and has no separate paired MSA field.
- Hardware: NVIDIA L40S or H100 GPU with Claude Science installed.
- Storage: Approximately 700 GB for the msa-search database and model containers.
- Microservices: msa-search, openfold3, and boltz2 NIM containers.
- Inputs: UniProt sequences for C1GY11 and C1HCX1, with 202 MSA hits per protein.
Why Both Models Collapse Without MSA
The benchmark measured interface confidence using iPTM, the metric that reports predicted contact between two chains. With MSA input, OpenFold3 reached 0.85 and Boltz-2 reached 0.82 for the Seh1–C1HCX1 heteromer.
Without MSA input, those numbers dropped to 0.14 and 0.19. The gap was not a single-run anomaly; five samples clustered tightly with a standard deviation at or below 0.006.
Extra sampling did not rescue the interface. Running OpenFold3 with five diffusion samples and Boltz-2 with five samples, six recycling steps, and 200 sampling steps left the no-MSA interface collapsed within 0.01 iPTM of the single-sample value.
The monomer results reveal different model dependencies. OpenFold3’s pLDDT fell from 82 to 36 without MSA, while Boltz-2 held monomer pLDDT at 0.73 from sequence alone but still could not place the interface.
- OpenFold3 with MSA: 0.85 iPTM; without MSA: 0.14.
- Boltz-2 with MSA: 0.82 iPTM; without MSA: 0.19.
- Sensitivity: Five samples per condition, standard deviation at or below 0.006.
- Sampling robustness: Additional samples and recycling steps did not substitute for MSA input.
A Reproducible Hypothesis Engine for Autonomous Discovery
After the confidence analysis, the agent superposed Seh1 from monomer and heteromer predictions. Both model families produced core Cα-RMSD values below 0.7 angstroms, with the major change confined to the WD40 β-propeller edge.
The Seh1 fold is completed, not remodeled.
OpenFold3 and Boltz-2 independently placed the same cluster of C1HCX1 β-strands near the WD40 velcro-closure position. That placement is consistent with the observation reported in Han, Tsenkov, Venanzi et al.
The two models reproduced the reference paper’s Figure 4e observation through independent architectural routes. That convergence strengthens the structural hypothesis without turning it into experimental validation.
The workflow does not claim a proven biological interaction. It offers a reproducible method for forming and inspecting structural hypotheses before experimental testing.
On internal benchmarks, BioNeMo skills raise task correctness from 60% to 100% and roughly double token efficiency. The Claude Science example illustrates the tactical reason: the agent preserved the right input types instead of quietly substituting an unpaired alignment.
The bigger signal for AI teams is not that two models predicted a protein interface. It is that an agent preserved the input types, sampling conditions, confidence fields, and failure modes across two independent architectures without silently swapping data.
For scientific agent developers, this shifts the benchmark from raw answer accuracy to process integrity. The workflow records the request, response, A3M files, endpoints, metrics, and errors at each stage, creating an auditable trail from sequence to structural claim.
That auditability matters because the interface prediction remains a hypothesis, not a verified interaction. Two independent models converging on the same local geometry is strong evidence, but the final word still belongs to experimental validation.
The commercial implication is clear: agentic scientific computing is beginning to look like software engineering’s coding-agent moment. The value is moving from model access to orchestration, validation, and reproducibility layers.
For platform builders, the workflow’s storage and compute requirements also set a real barrier. A roughly 490 GB database, plus model containers and an L40S or H100 GPU, is not a lightweight API call; it is an infrastructure decision.
The decisive variable in agentic protein structure prediction is no longer the folding model; it is whether the agent treats evolutionary alignment as non-negotiable. For teams building auditable AI-driven scientific workflows, programmatic SEO and AI automation is how Andres SEO Expert approaches it — talk to the team here.
Frequently Asked Questions
What is the role of multiple sequence alignment (MSA) in protein structure prediction?
MSA provides evolutionary context that tells folding models how residues have varied across related proteins. In the article, removing the MSA caused interface confidence to drop from 0.85/0.82 to 0.14/0.19 for OpenFold3 and Boltz-2, respectively, showing that MSA is the load-bearing input for interface prediction.
How does the BioNeMo-Claude Science workflow orchestrate protein structure prediction?
The workflow packages NVIDIA’s life sciences models into agent-callable skills. Claude Science agents discover, launch, and call BioNeMo NIM microservices (msa-search, openfold3, boltz2) from a configured compute endpoint. The agent generates unpaired MSAs, builds paired alignments, and passes the correct input formats to each folding model.
What hardware and storage are required to run the BioNeMo NIM microservices?
The setup requires an NVIDIA L40S or H100 GPU and approximately 700 GB of storage. The largest dependency is the msa-search NIM’s UniRef30 database at about 490 GB (or up to 1.4 TB for the full set). Local Docker containers are connected to the host GPU.
How did OpenFold3 and Boltz-2 perform with and without MSA input?
With MSA input, OpenFold3 reached 0.85 iPTM and Boltz-2 reached 0.82 iPTM for the Seh1-C1HCX1 heteromer. Without MSA, those scores dropped to 0.14 and 0.19, respectively. Five samples clustered tightly with standard deviation at or below 0.006, confirming the effect was not a single-run anomaly.
Can additional sampling or recycling steps substitute for missing MSA input?
No. Extra sampling did not rescue the interface. Running OpenFold3 with five diffusion samples and Boltz-2 with five samples, six recycling steps, and 200 sampling steps left the no-MSA interface collapsed within 0.01 iPTM of the single-sample value, demonstrating that MSA is irreplaceable.
Why is process integrity important in AI-driven scientific workflows?
Process integrity ensures that agents preserve input types, sampling conditions, confidence fields, and failure modes without silently swapping data. The workflow records requests, responses, A3M files, endpoints, metrics, and errors at each stage, creating an auditable trail from sequence to structural claim. This shifts the benchmark from raw answer accuracy to reproducible, verifiable scientific reasoning.
