Skip to content

NCBI — ELink gds→sra returns no links for series with SRA data

Status: draft, not filed Archive: NCBI Entrez Re-checked: 2026-10-15 — the census was not re-run. On a fresh geo-sample-1000, a different population, ELink returns nothing for 128 of the 669 series whose SOFT file names experiments (19.1%, against 17.5% in the census), and the six series on the pathology page still reproduce in its test. Evidence: pathologies/elink-gds-sra-missing-links.md Where to file: info@ncbi.nlm.nih.gov, or the NCBI Help Desk

Regenerate the affected list before filing. The table below is a census of reprocessed-gse, so that is the corpus to re-run — into a fresh cache: survey run resumes by default, so against the main cache it would skip every series and replay the dated answers, and --no-resume there would overwrite them.

CACHE=/tmp/elink-recheck.sqlite
for route in "gse->experiment:soft_family" "gse->experiment:elink_gds_sra"; do
  fetch survey run --route "$route" --corpus reprocessed-gse --cache "$CACHE"
done
fetch survey compare --route "gse->experiment:soft_family" \
                     --route "gse->experiment:elink_gds_sra" \
                     --corpus reprocessed-gse --cache "$CACHE"

The census is several hours of E-utilities traffic. --corpus geo-sample-1000 is the quick check that the rate has not moved; it is a different population, so its rate is not the headline below.

Draft report

Subject: ELink gds→sra returns no links for GEO series that have SRA data

We resolve GEO series to their sequencing data programmatically, and have found that elink.fcgi?dbfrom=gds&db=sra reports no SRA neighbours for a substantial fraction of GEO series that demonstrably have SRA experiments.

This is measured by census rather than by sample. We queried all 13,045 GEO series that have been through our reprocessing pipeline — series that are known to carry sequencing data — by two routes: ELink, and the !Sample_relation = SRA: lines in each series' own SOFT family file.

ELink gds→sra SOFT family file
Series resolved to experiments 10,752 (82.4%) 13,022 (99.8%)
Series reported as having none 2,293 (17.6%) 16 (0.1%)
Unique experiments returned 178,879 308,639

For 2,280 series, ELink returned nothing while the SOFT file named experiments. ELink returns about 58% of the experiments GEO's own records point to.

Taking GSE27679 as a small example:

  • esearch db=gds resolves it to UID 200027679.
  • elink.fcgi?dbfrom=gds&db=sra&id=200027679&cmd=neighbor returns no link records. Passing linkname=gds_sra explicitly makes no difference, and cmd=neighbor_history behaves the same way.
  • The SOFT family file at ftp.ncbi.nlm.nih.gov/geo/series/GSE27nnn/GSE27679/soft/GSE27679_family.soft.gz names nine SRA experiments, beginning SRX046652.
  • esearch db=sra&term=SRX046652[ACCN] returns count 1, so the experiment is in SRA.
  • ENA lists seven runs for SRX046652 (SRR121517–SRR121523).

The affected series are not confined to any era: GSE106544 (2,396 samples, all with SRX accessions in its SOFT file) returns no ELink neighbours, and so do series from 2011 through 2022. We specifically tested and ruled out an age effect — ELink succeeds for GSE21739, older than every affected series in our first sample, and for GSE56066, newer than all of them.

In every case we checked, the records exist on both sides and GEO's own metadata states the relationship; the Entrez link table does not express it.

The response is an ordinary HTTP 200 with no error field, so callers cannot distinguish "this series has no SRA data" from "the link is missing". In our census ELink produced zero outright failures — its incompleteness is entirely silent. That is why we are asking rather than quietly working around it.

Three questions:

  1. Is the gds↔sra link table expected to be complete, and is there a known reason these series are absent from it?
  2. Is there an Entrez query that recovers the relationship? Searching db=sra for GSE27679[All Fields] also returns 0, so we have not found one.
  3. Would a list of the 2,280 affected accessions be useful to you? We can supply it, and regenerate it on request.

Going the other way, we found three series where ELink and the SOFT file return experiment sets of the same size with zero overlap (GSE114373, GSE150508, GSE164404). In each case the SOFT-named experiments exist in db=sra but carry no runs, while the ELink-named ones do — consistent with a replaced submission where the link table was updated and the GEO record was not. If the two are maintained by different processes, that may bear on question 1.