colabfold_complex_e2e

ColabFold Complex End-to-End Reference

This directory is a small runnable reference that starts from curated ColabFold-like complex fixture outputs and produces the AFDB integration artifacts the toolkit is expected to emit.

Selected fixtures:

• Homodimer: AF-0000000066074510 -> Q46806 / Q46806
• Heterodimer: AF-0000000300000101 -> A0ABS2QMZ4 / A0ABS2QMF5

Input source:

• Raw fixture source files live under tests/fixtures/colabfold_real_examples/.
• This example copies the selected fixture PDB and score JSON files into input/ using normalized AFDB-style names: *-model_v1.pdb and *-meta_v1.json.

For the monomer reference example, see examples/colabfold_monomer_e2e/.

Regenerate

Run the repo-owned helper from the repository root:

.venv/bin/python scripts/generate_colabfold_e2e_example.py \
  --duckdb examples/uniprot_example_subset.duckdb \
  --output-dir examples/colabfold_complex_e2e \
  --example-id AF-0000000066074510 \
  --example-id AF-0000000300000101 \
  --python-bin .venv/bin/python \
  --dssp-algorithm mkdssp \
  --bcif-backend biotite

That helper executes the same script sequence used to populate this directory and records the exact subprocess commands in config/commands.txt.

Flow

The generated files follow the old Nextflow end-to-end order:

1. Normalize raw ColabFold fixture inputs into input/.
2. Validate PDB/score consistency via validation_results.tsv.
3. Convert ColabFold scores to AFDB confidence + PAE JSONs in scores/.
4. Write per-model manifests in chain_manifests/ and model_manifests/, then merge them into merged_manifests/.
5. Export per-model metadata JSONs in model_jsons/ and batch them into model_batches/.
6. Export per-chain metadata JSONs in chain_jsons/ and batch them into chain_batches/.
7. Enrich the complex model and chain metadata with computed iPSAE/interface summary metrics from ipsae/ipsae_summary.csv.
8. Export ModelCIF generator input JSONs in modelcif_input/.
9. Generate ModelCIF files in modelcif/.
10. Run DSSP and write enriched mmCIF files in dssp/.
11. Generate enriched PDB files in modelpdb/.
12. Attempt BCIF generation and write results to bcif/.

Complex Metrics Source

• uniprot/templates/colabfold_example_modelcif_metadata.json is static scaffolding for ModelCIF metadata categories, provider text, and software parameter definitions.
• That template must not contain fake computed iPSAE values.
• The computed complexPredictionAccuracy_* values in the committed complex JSON outputs come from the iPSAE summary CSV and are injected into:
  • model_jsons/ and model_batches/
  • chain_jsons/ and chain_batches/
  • ModelCIF global QA metrics during ModelCIF generation
• The lightweight complex e2e example currently carries only the iPSAE-derived metric set. The separate clash/interface-analysis fields (complexPredictionAccuracy_N_clash_backbone, complexPredictionAccuracy_N_clash_heavyAtom, numberOfInteractions) are not part of this committed example unless a later stage adds them explicitly.

Tooling Caveats

• Reproducing the metadata JSONs and ModelCIF inputs requires the committed DuckDB subset examples/uniprot_example_subset.duckdb, which contains only the UniProt accessions used by the monomer and complex examples.
• mkdssp was available locally and was used for the DSSP step.
• Mol* cif2bcif was not on PATH, so the example uses the explicit biotite backend instead.
• The committed .bcif files were generated directly from the DSSP-enriched CIF files in dssp/. The per-model status is recorded in run_summary.json.
• pydssp is not installed in this environment, so it was not used.
• The heterodimer fixture metadata is exported as fragment chains using local chain ranges from the curated fixture set. This example does not reconstruct authoritative UniProt residue offsets beyond what is committed in the fixture corpus.
• Those local fragment ranges are preserved through the generated ModelCIF entity/reference alignment metadata and the heterodimer complex now gets an AFDB-style fallback complex name (Complex of .../...) in the JSON and PDB outputs.
• The committed complex metadata schemas require the computed iPSAE blocks on both per-record files and batch files, so removing those fields from a complex model_jsons/*.json, model_batches/*.json, chain_jsons/*.json, or chain_batches/*.json file will fail validation.
• Legacy PDB DBREF records remain a format caveat for long UniProt accessions. The mmCIF/ModelCIF and JSON artifacts are the authoritative metadata outputs for the heterodimer example.