program.md

proteinCLAW autoresearch

This file is for autoresearch mode only. For a manual pipeline run, use the direct workflow in README.md.

The root loop runs the phosphopeptide design pipeline and lets an agent tune the experiment surface between runs.

Setup

Before starting the loop:

1. Read README.md for the shared setup steps.
2. Initialize submodules with git submodule update --init --recursive.
3. Run pixi run -e protein-dev protein-bootstrap.
4. Start Prefect with pixi run -e protein-dev prefect-start.
5. Validate the environment with pixi run -e protein-dev python prepare.py.
6. Confirm that results.tsv exists with the expected header row.

Experiment Surface

The root train.py file is the only place the autoresearch loop should edit. It now exposes two editable surfaces:

• build_plan() for run-level parameters
• the inlined pipeline.run(...) graph inside run()

The baseline there starts from:

• 10 RFD3 backbones
• 5 MPNN sequences per backbone

The loop should compare runs using the printed summary:

• higher pass_count
• lower best_peptide_ca_rmsd
• higher best_phosphosite_hbonds
• higher best_po4_fraction_buried

Run each experiment with:

pixi run -e protein-dev python train.py > run.log 2>&1

Logging

Append each attempt to results.tsv as tab-separated values using the summary fields from train.py.

Suggested statuses:

• keep when the score improves
• discard when the score is equal or worse
• crash when the run fails before printing a summary

Loop

The basic loop is:

1. inspect the current branch and commit
2. edit train.py
3. run the pipeline
4. read the summary
5. record the result
6. keep or discard the change based on the score

Do not use this file for direct pipeline runs. Use the README workflow instead.