ELink gds→sra returns no links for series that demonstrably have SRA data¶

Status: open, unreported Found: 2026-09-14 Archive: NCBI (Entrez) Severity: silent data loss

Measured on geo-sample-1000, a seeded stratified draw from all 275,238 GEO series. Of 150 series surveyed, 24 carry SRA experiments according to their own SOFT family file. ELink gds→sra returns zero links for 6 of those 24 (25%).

Symptom¶

elink.fcgi?dbfrom=gds&db=sra returns a linksets array with no linksetdbs for a GEO series whose SOFT family file explicitly names SRA experiments. There is no error, no warning, and no HTTP failure — the response is a well-formed 200 that says the series has no SRA data.

Anything resolving GEO → SRA through ELink therefore reports these series as having no sequencing data at all, rather than failing in a way anyone would notice.

Affected accessions¶

From the 2026-09-14 survey. Each series' SOFT file names the experiment in the third column, and that experiment exists in NCBI's own db=sra.

GEO series	Experiments in SOFT	Example experiment	In `db=sra`?	ENA runs
GSE27679	9	SRX046652	yes (count=1)	7
GSE29158	20	SRX014928	yes (count=1)	1
GSE34359	2	SRX111797	yes (count=1)	2
GSE35884	4	SRX117991	yes (count=1)	1
GSE49732	4	SRX466511	yes (count=1)	1
GSE49780	6	SRX467063	yes (count=1)	1

Reproducer¶

# 1. The series resolves to a gds UID.
curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=gds&term=GSE27679%5BACCN%5D+AND+GSE%5BETYP%5D&retmode=json" \
  | jq -r '.esearchresult.idlist[0]'
# -> 200027679

# 2. ELink reports no SRA neighbours. Expected: 9 UIDs. Actual: none.
curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/elink.fcgi?dbfrom=gds&db=sra&id=200027679&retmode=json&cmd=neighbor" \
  | jq '.linksets[0].linksetdbs'
# -> null

# 3. But the SOFT family file names the experiments.
curl -s "https://ftp.ncbi.nlm.nih.gov/geo/series/GSE27nnn/GSE27679/soft/GSE27679_family.soft.gz" \
  | gunzip | grep -oE 'SRX[0-9]+' | sort -u | head -3
# -> SRX046652, SRX046653, SRX046654

# 4. And that experiment is in SRA, with runs available from ENA.
curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=sra&term=SRX046652%5BACCN%5D&retmode=json" \
  | jq -r '.esearchresult.count'
# -> 1
curl -s "https://www.ebi.ac.uk/ena/portal/api/filereport?accession=SRX046652&result=read_run&fields=run_accession&format=tsv&limit=0"
# -> SRR121517 ... SRR121523

Passing an explicit linkname=gds_sra makes no difference, and searching db=sra for GSE27679[All Fields] also returns 0 — so there is no alternative Entrez query that recovers the link either.

What it is not¶

Not an age effect. The first hypothesis was that the link table was never backfilled for older submissions. It does not hold: ELink succeeds on GSE21739, which is older than all six failures, and on GSE56066, which is newer than all of them.

That hypothesis only arose because the run that produced it was biased. A 150-series sample was taken by slicing a sorted corpus, which returns the lowest accession numbers rather than a sample — it covered only GSE ≤ 59,184, the oldest fifth of GEO, so every failure in it was necessarily an old accession. The census settled the question properly: affected series span 2011 to 2022, and include GSE106544, a 2,396-sample series that returns nothing. The slicing bug is fixed, and the benchmarks page records it.

Not a client bug. The links are absent with and without linkname, under cmd=neighbor and cmd=neighbor_history.

Not private or withdrawn data. Every experiment resolves in db=sra and every one has downloadable runs in ENA.

Workaround¶

Read the SOFT family file. !Sample_relation = SRA: carries the experiment accessions for all six series. This is what fetch10xmeta already does, which is why the problem has not bitten the reprocessing pipeline.

Note the failure is not symmetric — see SOFT omits the SRA relation. SOFT and ELink fail on disjoint populations, so neither is a safe sole source. The resolver takes the union and records which route produced each accession.

Upstream report¶

Not yet filed. Draft body in upstream/ncbi-elink-gds-sra.md.

Regression test: tests/pathologies/test_elink_gds_sra_missing_links.py. It asserts the links are still missing. If it fails, NCBI has populated them — update this page and re-rank the route.