Two diagrams. Open the SVGs directly — they are plain text and render in GitHub.
1. Pipeline — pipeline_DAG.svg
Three input modes (genome, protein, pangenome) converging on one shared spine:
00c pharokka → 01 phold → 02 FoldSeek → 03 compare → 04 curate → 05 output
↑ → 06 phynteny → 07 integrate
Step 02 forks: 02d searches local FoldSeek databases, 02w uses the public
search.foldseek.com API. Both produce identically shaped best_hit.csv /
top3.csv, so everything downstream is unaffected by which one ran.
02w is experimental. The web API accepts structures only, so it needs an
ESMFold folding step first (scripts/lib/esmfold_cache.py), and it is subject to
server-side rate limits and timeouts. Use 02d for anything you intend to
publish.
The purple lane marks FoldFirstAskLater (FFAL), Hannelore Longin's structure-first annotation pipeline. FFAL can replace steps 01–02 entirely: it produces per-CDS structural hits, and an adapter reshapes them into what step 03 expects, so phageFACTor's curation runs on FFAL's evidence.
Status: working, not yet merged. Integration is in progress with the FFAL
authors; the branch that carries it is dev-tanais on their repository. Nothing
in this repo depends on FFAL — the lane is optional and the adapter lives outside
this repo.
2. Scoring and curation — scoring_curation_DAG.svg
Where a FoldSeek hit is filtered, scored, ranked and finally named. The
superscript numbers refer to the word-lists catalogued in
code_lists.md. Prose version of the same logic:
curation_gates.md.
Two things the diagram is meant to make obvious:
- The informativeness gate removes most hits, before any scoring happens.
- Trusted calls are never re-annotated. Only pharokka-hypothetical genes enter curation at all; a sequence-homology call is passed through untouched.