Part 1: Define the Test Before the Design: How BindClaw Works

BindClaw: Define the test first. Set the research question before designing. Nadoo Bio series cover.

Connecting Protein AI to Research · Three-Part Series

Part 1 · How BindClaw Works — You are here

Part 2 · Running RFD4-Proteina

Part 3 · Starting with Real Proteins

A protein's sequence is the order of its amino acids. Protein AI can predict a three-dimensional shape from that sequence and propose new sequences to fit a desired structure. Researchers want to know what comes next: will this candidate bind at the required site? Will it work only in a particular state?

Designing a shape that binds a target is different from achieving a desired effect. An inhibitor has to compete with the target's usual binding partner. A design intended to bring two proteins together has to hold on to both. NVIDIA's BindClaw starts from this distinction. [1]

The generative model and the layer directing it

BindClaw is a research project on agents for binder design. Binders are proteins designed to attach to a target; antibodies also fall into this category. One distinction matters here: a generative model that produces structures and an agent that decides when and under which conditions to run that model operate at different layers. BindClaw handles the latter.

Four stages: define the goal, build computational tests, design and iterate, and measure experimentally

The molecular shapes in the diagram are illustrative ribbon representations drawn from public coordinates. They are not computational results for a particular design, nor do they imply that a candidate passing a computational test will produce the same effect in practice.

The workflow has four stages. First, define the goal. Once the desired function has been extracted from a paper or an instruction, Assay Builder translates it into a computational test. Intake extracts the target and design constraints, Target ID resolves the target, and Strategist proposes a mechanism of action and the partners or states that need to be evaluated alongside it. A mechanism of action is a hypothesis about the interactions that produce the desired effect.

Design and iteration follow. A generative model produces candidates, and the computational results guide changes to the contact interface or generation settings before another run. The evaluation criteria remain unchanged. The final stage is experimental measurement. [1]

A computational assay resembles a test specification

Suppose we want to recognize a protein with calcium bound. We also need to run the condition without calcium to make a comparison. The question is whether the candidate scores well only in the desired state. This kind of computational test is called a computational assay. It resembles a test specification in software development, but it does not capture all of biology. Passing the test does not guarantee function.

This also explains why changing evaluation criteria midway undermines a comparison. If the criteria need to change, they should be versioned, and earlier candidates should be reevaluated under the same criteria before we can claim an improvement.

Turning research questions from papers into five tests

BindClaw reads the desired effect from a paper and translates the contacts, states, and molecular arrangements needed for that effect into computational criteria. Panels A–E of the official figure show how different research questions become different tests. [1]

The diagrams below explain the research questions and computational criteria schematically. The experiments shown at the end are follow-up steps for testing computational candidates.

A. How should a candidate engage two receptors to mimic IL-2 and IL-15?

Schematic: compute contacts and arrangement for an IL-2/IL-15 mimic binding two receptors, then test the cell response.

IL-2 and IL-15 are proteins that send signals to immune cells. A candidate designed to mimic them needs to contact both the beta and gamma components of the receptors that receive those signals. The computational test checks whether there is sufficient contact with both and whether the two receptors come together in the intended arrangement.

B. Which state of Hsp70 should the candidate recognize?

Schematic: compare ATP and ADP states of DnaK computationally, then measure binding and function.

Hsp70 helps other proteins maintain their shape. This example uses DnaK, a member of the Hsp70 family, and places the same candidate in the ATP and ADP states associated with its operating cycle to compare them. Although the starting paper aimed at activation, the test here examines computational evidence that the candidate would preferentially recognize the ATP state.

C. Could engaging two sites help bring a receptor into the cell?

Schematic: check a design bridging D11 and D6, compare a disrupted-link control, then test cellular uptake.

Endocytosis is the process by which a cell encloses material from outside and brings it inside. In this example, two binders are joined into one design so that each contacts a different site on a receptor. The comparison checks not only whether both bind at the intended sites, but also whether they contact unintended sites and whether contact disappears in a control candidate with one side's binding function disrupted.

D. How should a small molecule be held to switch on fluorescence?

Schematic: assess the DFHBI pocket pose and burial, then measure fluorescence from selected candidates.

DFHBI is a small molecule that can fluoresce when bound inside a protein. In redesigning the existing barrel-shaped protein mFAP1, the test checks the position and orientation of DFHBI in its binding site and whether it retains its flat shape. A separately generated binder is also presented, but it is not the same barrel-shaped design. A pass in this figure concerns molecular arrangement; it is not a measurement of fluorescence brightness.

E. Can a candidate selectively recognize digoxigenin?

Schematic: compare digoxigenin with related molecules computationally, then measure binding selectivity.

Digoxigenin is the small molecule chosen as the recognition target in this example. The test examines the candidate's computational binding score and contact geometry for digoxigenin, then compares them with results for three other molecules. Rather than looking only at a score for the target molecule, it asks whether the candidate can distinguish it from the others. [2]

F. Keep the criteria, change the generation conditions

Schematic: inspect weak receptor predictions, revise generation settings, and reassess with the same fixed criteria.

Panel F shows how the test results from A informed the next round of design. When the first candidates showed weak predicted contacts with the two receptors, the agent added conditions to encourage helical shapes, narrowed the length range, and generated more candidates. The ipTM values shown for the two receptors rose from 0.07 and 0.19 in the first round to 0.86 and 0.90 in the next. ipTM is a computational metric of confidence in the relative arrangement of proteins. These values are not experimental success rates or results showing immune-cell activation. They record that changing the generation conditions produced candidates with higher structural confidence. [2]

Changing the model can change the answer

One observation from the design stage stands out. EGFR is a cell-surface protein that receives growth signals and is a common target in cancer research. When it was used as the target, Proteina-Complexa and RFD4-Proteina showed opposite preferences for binding sites. In other words, which binding site was favored depended on the generative model.

This example shows what needs to stay fixed and what needs to be compared when we ask the same question of different models. Without setting evaluation criteria first, we cannot distinguish a difference caused by changing the model from one caused by improving the design. [1]

What the calmodulin campaign showed

Calmodulin is a protein that relays calcium signals in cells. It has different shapes with and without calcium bound. Through a design process involving human participation, the BindClaw team found binders that preferred the calcium-bound state.

Of the 18 final candidates tested experimentally, 15 bound to calcium-bound calmodulin. Of those 15, 13 showed no detectable binding to calcium-free calmodulin. The results demonstrated that designs could distinguish between the two states.

The denominator matters when reading these numbers. The 18 sequences were selected from 1,553 evaluated designs. This does not mean that most of the initially generated candidates worked. Nor does undetectable binding prove that there was no binding at all: it means binding was not detected under the measurement conditions. The report also acknowledges that the weights used for the final ranking could not be reconstructed. That information is needed to reproduce the performance. [2]

The public page describes the method step by step, but the code is not yet available to download. Its Code section still says “Coming soon,” so we cannot currently check BindClaw by running it ourselves. [1]

We therefore started with what we could run. The next article covers the scope of our RFD4-Proteina runs in Nadoo Bio, their measured runtimes and costs, and the terms of use. RFD4-Proteina is one of the generative models BindClaw uses at the design stage.

References

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

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