Part 3: From Ubiquitin to Calmodulin: Which Protein Should We Start With?

Real proteins: Where do we start?. Four cases to guide the research sequence. Nadoo Bio series cover.

Connecting Protein AI to Research · Three-Part Series

Part 1 · How BindClaw Works

Part 2 · Running RFD4-Proteina

Part 3 · Starting with Real Proteins — You are here

Access to a design model does not get a research project moving on its own. The model that generates candidates and the criteria used to judge them must be prepared separately, and we need to decide how to connect them. Choose the wrong first question, and even a good result may be hard to explain. A poor result may offer no clue about what to change. Using four proteins with publicly available structures, we will broaden the questions from structure comparison to function and state selectivity. These are plans for future research, not a set of results all obtained by running RFD4-Proteina. Structure prediction, sequence design, and evaluation tools need to be selected and connected according to each question.

Four stages featuring ubiquitin, HER2 and an antibody, PETase, and calmodulin

The molecular shapes in this diagram are illustrative. The ubiquitin and antibody shapes draw on public coordinates (1UBQ and 1IGT). The other two panels do not show the actual structures for those cases.

Ubiquitin: checking that the record holds together

Ubiquitin is a small protein that attaches to other proteins in the cell, signaling how they should be handled. It makes a useful first case because the exercise does not need to aim for a new discovery.

Comparing a predicted structure with the publicly available experimental structure 1UBQ for the same sequence reveals whether chains, amino acid positions, and result files remain correctly linked throughout the workflow. The comparison lets us check the numbering scheme and the connection between inputs and outputs. Reproducing a well-known structure should be interpreted separately from prediction performance on new proteins. [1]

HER2: comparing where a candidate makes contact

HER2 is a protein involved in signaling at the cell surface. The public structure 1N8Z contains the extracellular region of HER2 together with a Fab fragment of the antibody trastuzumab. Fab is the part of an antibody that recognizes its target.

We can use this structure as a reference to compare whether a candidate contacts the same region. Three claims need to be kept distinct: making contact at a similar site, actually binding strongly, and producing an effect in cells. Each requires different evidence. Comparing contact sites addresses only the first question. [2]

PETase: keeping shape and activity separate

Enzymes pose a different kind of question. PETase is an enzyme that helps break down the plastic PET. The public structure 5XJH offers a starting point for considering designs that preserve regions important to the reaction while changing the rest of the sequence.

A prediction that a candidate will retain its shape is different from a result showing that it breaks down PET faster. Whether enzyme performance has improved is judged by measuring degradation activity at matched temperatures and under matched test conditions. [3]

Calmodulin: evaluating two states together

Calmodulin has different shapes with and without calcium bound. The central question is whether we can evaluate both states together.

Running the same sequence twice does not amount to comparing two states. Inputs that represent the state, such as structures and ions, must actually reach the model for the comparison to hold. In the BindClaw study, 15 of the 18 final binder candidates bound to calcium-bound calmodulin. Of those, 13 showed no detectable binding to calcium-free calmodulin. This does not mean the binders attached to calcium itself; the result shows that binding differed with calmodulin's state. [4]

Keeping a consistent record across cases

For this kind of research in Nadoo Bio, researchers obtain structural data directly from the original source and upload it themselves. Recording both the source and the structure used makes it possible to revisit the reference when comparing results later. [5]

It helps to place a research specification between the agent and the execution environment. This document records the target and chain numbering scheme, the desired state and comparison state, controls, the hash of the evaluation code, the tools and weights used, the overall budget, and stopping conditions. The agent handles the questions and the reasons behind its choices. The execution environment handles tool contracts, execution conditions, and result provenance. The specification connects the two. [6]

Continuing the research when BindClaw is released

BindClaw's code is still listed as coming soon. When it becomes available, we plan to connect its process for translating research questions into computational tests with Nadoo Bio's execution and result records. [7]

Having access to generation is a starting point; it does not automatically make research faster. Evaluation criteria need to stay fixed, and the conditions that produced each candidate—and the reasons for revising it—need to be recorded. We must account for the cost of the entire search, including failed candidates, before we can say whether research has improved within the same budget. Ubiquitin asks about structure comparison and record continuity; HER2 about contact interfaces; PETase about enzyme activity; and calmodulin about binding in different states. Connecting the right tools and experiments to each question is the next task.

The day BindClaw's code is released, our first step will be to try connecting it to Nadoo Bio. We will share which research questions we can tackle first, along with the practical integration process and its results, in follow-up articles.

References

  1. RCSB PDB. 1UBQ ubiquitin structure. Public structural data, accessed 2026-11-13.

  2. RCSB PDB. 1N8Z HER2–trastuzumab Fab complex. Public structural data, accessed 2026-11-13.

  3. RCSB PDB. 5XJH PETase structure. Public structural data, accessed 2026-11-13.

  4. Cho, Y. et al. Adapting biological mechanisms for agentic protein design. Technical report, 2026-10-28.

  5. Nadoo Bio. Preparing research inputs. Product documentation, accessed 2026-11-13.

  6. Nadoo Bio. Agent operating procedures. Product documentation, accessed 2026-11-13.

  7. NVIDIA Digital Biology Research. Official BindClaw project. Official research overview, accessed 2026-11-13.