Fix inverted span for minus-strand multi-exon GenBank locations

[Linxiushen]

What was wrong

GenBank._parse_location_node computes the aggregate span of a join/order location from the positional first and last parts:

location["start"] = located[0]["start"]
location["end"] = located[-1]["end"]

This assumes the parts are listed in ascending genomic order. They are not for minus-strand multi-exon features: GenBank/NCBI (and Bio.SeqIO.write) list the exons in transcription order, which is descending genomic order, e.g. a two-exon minus-strand CDS is written as complement(join(121..180,11..60)). For such records the code takes the larger coordinate as start and the smaller as end, producing an inverted span (start > end, i.e. a negative length). The start_partial/end_partial flags are taken from the same wrong positional parts.

This contradicts the method's own documented contract (join(1..100,200..300) -> {"start": 1, "end": 300}) and its stated goal of matching Biopython's CompoundLocation, whose .start/.end are always the min/max regardless of part order.

Trigger

Load any GenBank file with a minus-strand multi-exon feature:

CDS   complement(join(121..180,11..60))

datasets.load_dataset("genbank", data_files=...) then yields a location dict with "start": 121, "end": 60 — an inverted span (start > end, negative length). By Biopython's CompoundLocation semantics the aggregate span is min-start / max-end (here start=11, end=180 in the parser's 1-based convention), so the inverted output is a bug rather than a convention difference. This complement(join(...)) form is what Bio.SeqIO.write(..., "genbank") emits for a naturally-constructed two-exon strand=-1 feature, so any Biopython round-trip or real NCBI record of this shape is affected.

Fix

Aggregate the span over all parts: start = min(part starts), end = max(part ends), and take each partiality flag from whichever part actually carries the min-start / max-end coordinate. This matches Biopython's CompoundLocation.start/.end semantics. For already-correct ascending (plus-strand) inputs the result is byte-identical, so this is a pure correctness fix with no behavior change for the common case. parts (segment order) is left untouched.

Red / green evidence

Added test_genbank_minus_strand_multi_exon_join_span_is_min_max.

• Before the fix: the test fails with assert 121 == 11; end-to-end via load_dataset the row's location is {"parts": [[121, 180], [11, 60]], "start": 121, "end": 60} — an inverted span.
• After the fix: the test passes and the location is {"parts": [[121, 180], [11, 60]], "start": 11, "end": 180}, matching Biopython. The full tests/packaged_modules/test_genbank.py suite passes (71 tests, no regressions). ruff check and ruff format --check are clean.

Run: python -m pytest tests/packaged_modules/test_genbank.py -q

---

Prepared with AI assistance; verified locally by the submitter.

[lhoestq]

good catch ! cc @behroozazarkhalili for viz

related to #7951

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[behroozazarkhalili]

Thanks for the cc. Confirmed: the new test fails on main (start=121, end=60) and passes with this change, and the rest of test_genbank.py stays green. Min-start/max-end is the right aggregate. LGTM.

[Linxiushen]

I checked the existing hosted CI logs; I did not rerun the workflow. The Python 3.10 Ubuntu unit job reports 3510 passed and one collection error: Transformers 5.17.0 requires huggingface-hub<2.0, but the dependency-upgrade step installed 2.0.0. All 71 GenBank tests, including the new regression test, passed in that job.

The same dependency error appears in the same-day main Windows unit job and main Ubuntu integration job, with identical workflow and failing-test source. These are cross-matrix comparisons, not an exact-base rerun. GitHub annotations attribute the seven cancelled PR jobs to the failed unit matrix job; those cancellations still leave the full CI matrix incomplete.