Skills carry biological operating knowledge
Microsoft has added an open-source NVIDIA BioNeMo agent to Microsoft Discovery, allowing scientists to request protein binder designs in natural language and run the RFD4-Proteina workflow through Azure. The integration uses BioNeMo Agent Toolkit skills to teach the agent which endpoint to call, how to format molecular inputs and how to interpret outputs. Microsoft Discovery supplies orchestration, knowledge grounding and workspace controls, while NVIDIA’s model performs the GPU-intensive structure and sequence generation.
BioNeMo Agent Toolkit is a collection of structured instruction files for computational biology. Each skill captures model-specific operating knowledge that a general-purpose agent is unlikely to apply consistently from memory. The agent reasons about the scientist’s request and coordinates tools; the skill describes the valid workflow; an NVIDIA Inference Microservice performs computation. This separation makes domain expertise inspectable and reusable, though research teams still need to review the instructions as models, endpoints and scientific practice change.
Microsoft Discovery is designed for scientific research and development, combining agent orchestration, advanced reasoning, a graph-based knowledge foundation and high-performance computing. Its Bookshelf capability can convert proprietary documents into GraphRAG knowledge bases with citations. Enterprise deployments sit inside the customer’s Azure subscription and Entra tenant, using Entra ID, Azure role-based access, private endpoints, encryption and optional audit logging. Those controls provide a governed workspace but do not validate the scientific quality of a model-generated design.
A lightweight agent reaches managed GPU compute
The first contributed agent, nvidia-rfd4-proteina, targets protein binder design. RFD4-Proteina combines RFdiffusion and Proteina model lineages to generate a three-dimensional structure and amino acid sequence in one pass. It also supports motif scaffolding, enzyme design, small-molecule binders, partial diffusion and inverse folding. A scientist can provide a target structure and request candidates, after which the agent prepares the input, calls the model and returns structures, sequences and a summary.
Microsoft says RFD4-Proteina can generate hundreds of candidates in hours, with Discovery filtering and ranking a shortlist for researchers. That compression is potentially valuable where manual design cycles are slow, but rank is not experimental evidence. Protein behaviour depends on folding, expression, stability, specificity and biological context. Wet-lab validation remains essential, and teams should preserve rejected candidates and scoring criteria so a promising output can be traced back to the exact request, model version and filters.
The tool container is deliberately lightweight: a small Python image and script, without model weights or a CUDA stack. GPU computation can run through NVIDIA’s hosted API, Microsoft Foundry or a self-managed AKS deployment. This reduces the infrastructure burden inside the Discovery workspace and makes the pattern extensible to other models. It also means the agent’s security boundary spans several services, so identity, network and retention settings must be reviewed end to end.
Deployment choice changes the data boundary
Deployment choice materially changes data sovereignty. If the container and model run on AKS within the customer tenant, molecular data can remain inside that environment. Using NVIDIA’s hosted API sends requests to an external service. Microsoft states this distinction explicitly, which is useful for regulated research. Organisations should also examine support access, telemetry, regional availability and whether derived structures or prompts are retained, not simply whether the initial file crossed a tenant boundary.
The same skill-to-agent pattern is intended to support DiffDock for docking, Boltz-2 for structure and affinity prediction and GenMol for molecule generation. A modular catalogue can let scientists compose workflows without rebuilding every integration. Composition increases validation complexity, however, because uncertainty passes between models. A pipeline needs stage-specific acceptance criteria and should expose confidence, assumptions and transformation history rather than presenting the final shortlist as the output of one reliable system.
Installation uses the open-source Microsoft Discovery catalogue, a workspace secret for the NVIDIA API key and a locally built container. That transparency gives platform teams a starting point for code review and reproducibility. Production use still requires dependency pinning, image scanning, secret rotation, network restrictions and controls over who may launch expensive or sensitive jobs. Free credits are useful for testing but should not obscure the eventual compute and service costs of large candidate campaigns.
Generated candidates begin rather than finish discovery
The integration is a concrete example of how agent skills can bridge a general research environment and a specialised scientific model. Its promise is accessibility: protein biologists can describe an objective without becoming infrastructure engineers. Research organisations should establish an approval path for new skills and models, document their intended scientific use and prevent unreviewed community contributions from reaching sensitive projects. They should also retain negative experimental results so the agent does not repeatedly propose failed designs. The measure of success should be better experimental throughput and traceable hypotheses, not the number of structures generated. When workflow knowledge, computation and human validation remain distinct, agentic AI can accelerate exploration without pretending that a fluent interface has replaced scientific judgement.