[DimaMolod]

Second of three behaviour-preserving PRs that split alphafold3_backend.py; stacked on #658. This one takes input preparation out of the backend. No flag, output or file name changes.

What moves

AlphaFold3Backend.prepare_input was 1,230 lines holding 28 nested functions. It is now about 40 lines:

• alphafold3_inputs.py (new).
  • The pure helpers become module functions: chain construction, monomer normalisation, region slicing of JSON chains, template mmCIF fixes, chain IDs, and the job name (_object_name_fragment, _build_output_job_name).
  • The helpers that read prepare_input's state become methods of a FoldInputBuilder: the chains so far, the chain IDs in use, which chains carry MSAs translated from AF2 features, the debug records, and the caches.
  • prepare_input makes a fresh builder per call, so the caches of feature metadata and template mmCIF checks still last one preparation, never the process. _supported_chain_kwargs loses its cache: it runs a few times per fold, and caching it would have pulled the pure chain helpers into the builder.
• alphafold3_debug.py (new). The translated-MSA debug writer and the final-MSA A3M writer that was nested in predict_structure.

The moved code changes only in indentation, apart from closure state becoming builder attributes. That rewiring is spelled out and count-checked by the script that did the cut.

Commits

1. The move. Tests patch and call the names where the moved code now looks them up. The string naming rules are tested in output_paths, where they live since refactor: move AF3 output writing and resume out of the backend #658.

2. Builder tests:

   • chain IDs roll over from Z to AA;
   • chain IDs given in JSON and generated ones never collide;
   • two preparations in one process share no chain IDs and no cached feature metadata;
   • A3M region slicing keeps only the insertions inside the region;
   • JSON chain regions slice sequence, residue IDs, PTMs and MSA together.

   Making the metadata cache process-wide fails the shared-state test, as it should.

3. PreparedFold. prepare_input returns PreparedFold(fold_input, output_dir, resolve_msa_overlaps) instead of single-key {fold_input: (output_dir, flag)} dicts. predict drops its length and legacy-shape checks. In the tests, unpacking becomes attribute access; no assertion changes.

Verification

• CPU equivalence harness, after the move and again after PreparedFold. It covers 15 inputs: monomer, homo- and hetero-oligomers, chopped regions, RNA, DNA, ligand, PTMs, JSON inputs, multi-seed, and two folds per call. Each runs prepare_input and then process_fold_input(model_runner=None) with debug MSAs and templates on. Every _data.json, debug artifact, output dir and resolve_msa_overlaps flag is byte-identical to main.
• Unit + integration, both release images: the same results as main plus the 6 new tests. AF2 image: 937 passed, 26 skipped. AF3 image: 1008 passed, 15 skipped. The same pre-existing environmental failures as on main (no pytest-xdist, no network).
• GPU, full AF3 suite with RUN_EXPENSIVE_GPU_TESTS=1: 39/39 passed, including TestAlphaFold3MetadataEndToEnd (real feature generation through inference and ModelCIF), resume, fast kernels and the MMseqs/regression checks.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-13T18:22:55.132940Z	03c7eb7	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.