← All use cases

Discover protein folding models

Convert protein-folding hypotheses into benchmarked experiment cycles.

In Goal mode, ZeroTwo researches new architectural changes to AlphaFold2 and implements them, aiming at better protein folding results.

Advanced
Long-running

Starter prompt

/goal Test whether adding higher-order geometric state alongside the pair representation improves a small AlphaFold2-style model. Start from the baseline in this repo, write PLAN.md with the hypothesis and the ablations first, and then iterate against validation lDDT-Ca on the benchmark. Log every run in EXPERIMENTS.md, compare only against matched baselines, and tell me plainly when a gain is within noise.

Explore a protein-folding architecture hypothesis

Give a Work or Code task one bounded idea, a working baseline and a benchmark, then let Goal mode implement, track and iterate. The example question here is whether explicit higher-order topological states, such as face and tetrahedral terms next to the pair tensor, help an AlphaFold2-style model.

Define a bounded experiment

Start small. Pick one hypothesis, a minimal baseline you can train quickly, and a benchmark with a single validation metric. Have the agent cover each change with unit tests and microbenchmarks before it spends time on full training runs.

Run the search with Goal mode

Write the goal as a falsifiable hypothesis with a plan of explicit ablations, then let the task hill-climb validation lDDT-Cα across turns. Goal state is stored on the server, and you can check it with /goal, /goal pause and /goal resume. A goal on a cloud task keeps going after you close the window. A goal on a local task is picked up only by the desktop app that started it, so it waits, and no time counts, while that computer is offline. Ask the agent to keep PLAN.md and EXPERIMENTS.md so every run can be audited afterwards.

Training needs real compute. A cloud task runs in a managed sandbox and keeps going when you close the app. A local desktop task uses your own machine and its GPU, if you have one, but it only runs while the desktop app is open and the computer is online.

Review the result

An agent can speed up implementation and debugging while you steer the science, but it does not solve folding for you. Treat a gain as real only after matched comparisons and replicates, and see Follow a goal for how to set the stopping condition.

Resources

The public benchmark plan and NanoFold competition repositories describe one benchmark for this kind of experiment.

Related use cases