Part 2: Candidate Generation Is Ready: Running RFD4-Proteina in Practice

RFD4-Proteina: A hands-on run. Candidate generation: time, cost and terms. Nadoo Bio series cover.

Connecting Protein AI to Research · Three-Part Series

Part 1 · How BindClaw Works

Part 2 · Running RFD4-Proteina — You are here

Part 3 · Starting with Real Proteins

In protein design, the generative model produces candidates. Deciding how often to run it, under which conditions, and how to evaluate the results happens outside that model. An agent such as BindClaw handles this outer layer by translating a research goal into computational tests and iterating on designs. RFD4-Proteina works inside it, generating candidate sequences and atomic structures together. Developed jointly by the University of Washington's Institute for Protein Design and NVIDIA, the model generates compressed numerical representations of molecular types and shapes alongside coordinates. [1]

What generation-only means

Four stages: prepare an annotated CIF, generate under the specified conditions, receive candidate files, and evaluate separately

This diagram illustrates the inputs and outputs when running the model in Nadoo Bio. It is not the output of a particular run, and generated results do not include quality scores.

Nadoo Bio offers this model for generation only. Supply one annotated CIF file containing the design conditions, and it generates up to four candidates. It returns a structure file for each candidate, a table listing the files, and a record of the settings used for the run. CIF is a standard format for molecular atomic coordinates. An annotated CIF also specifies which parts to preserve and which to redesign.

The results contain no quality scores. Choosing among candidates requires refolding—predicting how a designed sequence folds—and docking, which calculates how it fits against the target. Experimental testing comes at the end. Completing generation means that candidates have been produced; it does not establish that they are useful. [2]

Measuring runtime and cost

On an NVIDIA L4 GPU, generating four candidates took 171 seconds for a 100-residue input and 301 seconds for a 250-residue input. A residue refers to an individual amino acid in a protein. At the published rate of $0.000829 per second, these runs work out to roughly $0.14–$0.25 each. [3]

The length increased by a factor of 2.5, while runtime increased by about 1.76. These two measurements alone are not enough to predict runtimes for other lengths or input configurations.

From a tool list to an execution plan

MCP is a way for AI agents to check a tool's input requirements and request work. Retrieving the tool list through this connection returns tools including Boltz-2 for structure prediction and ProteinMPNN for sequence design. RFD4-Proteina is also on that list.

Inputs come next. The agent looks up the files and settings each tool requires, checks them against the material it has, and creates an execution plan. That plan records the inputs, the settings to run, and the estimated cost. The job is submitted once the researcher opens, reviews, and approves the plan.

That approval is an important point in the research record. It fixes what will be run and under which conditions, providing a reference when tracing back from the result files later. [4]

Three questions about the terms of use

First, what do the terms cover? BSD 3-Clause applies to RFD4-Proteina's code and model weights. Weights are the numerical values a model acquires through training.

Second, what is allowed? BSD 3-Clause is a permissive license that allows commercial use, modification, and redistribution.

Third, what must be retained or respected? Redistribution must preserve the copyright notice, license conditions, and disclaimer. The names of the copyright holders or contributors must not be used to endorse or promote products without prior written permission. [5] [6]

The model card explains that separate terms apply to the optional OpenFold3 refolding files. This component is used for refolding and calculating metrics; it is not needed to generate candidates. Anyone also using refolding-based evaluation needs to check that component's terms separately. [6]

Now that we can generate candidates, the next questions follow naturally: which protein should we start with, and what would count as success? The final article lays out four examples and an order for exploring them.

References

  1. RFD4-Proteina project. Official documentation. Official model resources, accessed 2026-11-13.

  2. Nadoo Bio. RFD4-Proteina tool overview. Public product page, accessed 2026-11-13.

  3. Nadoo Bio. Pricing. Public product page, accessed 2026-11-13.

  4. Nadoo Bio. Checking execution support for each tool. Product documentation, accessed 2026-11-13.

  5. RosettaCommons. RFD4-Proteina code LICENSE (commit 6805869). BSD 3-Clause license text, accessed 2026-11-13.

  6. NVIDIA. RFD4-Proteina model card (revision a06e2509). Documentation from the weights distribution repository, accessed 2026-11-13.