Inspiration

Designing a protein binder involves much more than generating a sequence. Researchers need to understand which candidates to investigate, why a model favours them, and what remains uncertain before committing time and resources to experiments. Yet there remains a clear lack of integrated platforms that bring these steps together in a fast, transparent, and evidence-driven workflow.

Therefore, we built MyProtein to unify this process in one research workspace. and grounded in traceable evidence. Instead of presenting an unexplained score, it connects protein design and candidate screening with traceable model metrics and structural evidence—enabling faster, better-informed decisions.

What it does

MyProtein let a user define a target protein and submit a binder-design task. It uses Tamarind Bio's BoltzProt integration to generate candidate binders and predicted complex structures.

The workspace then presents the returned candidates in a ranked evidence table. For each candidate, users can inspect model confidence values, see whether it passes the configured screening thresholds, and understand the reasons behind that outcome. Users can select candidates for a follow-up Boltz-2 complex prediction and inspect the associated structural evidence.

In one real run against human IL-2, the workflow returned 50 candidates. Six met our current model-score thresholds. These results are computational predictions: passing a threshold does not establish experimental binding, specificity, biological activity or suitability for a wet-lab study.

How we built it

We built a Python backend and a browser-based frontend around a target-specific research workflow. The backend submits and tracks Tamarind jobs, retrieves results, checks candidate and structure data, and stores task records locally.

Our screening logic uses explicit rules and evidence rather than asking a language model to invent a conclusion. The interface connects each candidate's outcome to its model metrics and provides access to predicted structure files and interface-contact information.

Challenges we ran into

A major challenge was distinguishing what the models actually returned from what we hoped to conclude. A confidence score is not a measured binding affinity, and a predicted contact is not proof of biological function. We therefore made missing evidence and the limits of each assessment visible.

We also had to handle long-running external jobs, account limits, incomplete results and structure-rendering delays. Keeping task state and candidate provenance attached to the correct target was essential for a trustworthy demo.

What we learned

We learned that the most useful part of an AI research tool is often not a single model call, but the system around it: clear inputs, traceable outputs, honest uncertainty and a way for researchers to inspect the evidence.

What's next for MyProtein

We want to improve interactive structure visualization, add stronger site-specific and off-target checks, and make it easier to compare independent predictions before a researcher chooses candidates for experimental validation.

Built With

  • bio`?`boltzprot`?`boltz-2`
  • python?`javascript`?`html`?`css`?`tamarind
Share this project:

Updates

Submission history