Re-using public single-cell data without inheriting someone else’s mistakes
A practical route from a GEO accession to an analysis you can trust — what to check in the metadata, when raw counts matter, and the traps in combining datasets.
Do this on your own data: the first thing to do with a matrix pulled off GEO is find out what shape it is actually in. Drop it into the free QC report before you build anything on top of it: omics.hstgenomics.com/qc-check. An example of what a finished re-analysis looks like: omics.hstgenomics.com/sample.
Want it done rather than explained? We do analysis only — send a count matrix, an h5ad or Seurat object, Cell Ranger output or a public accession and get publication-ready results with the code that produced them, at a fixed price agreed before any work starts: omics.hstgenomics.com/quote.
Public data is the cheapest experiment you will ever run. It is also the easiest place to build a conclusion on someone else's undocumented decisions. This is the sequence that keeps that from happening.
1. Work out what you are actually downloading
A GEO series may contain raw FASTQ, a filtered count matrix, a normalised matrix, or an author-processed object. These are not interchangeable. A "filtered" matrix has already had barcodes removed by criteria you cannot see; a normalised matrix cannot be used for pseudobulk differential expression, because you need raw counts. If only processed data is available and your question requires counts, say so in your methods rather than transforming something back into a shape it never had.
2. Read the metadata like a sceptic
Check three things before any analysis: how many biological subjects are actually represented (a series with 20 samples may be four subjects with five timepoints); which chemistry and platform version was used, since v2 and v3 10x chemistry differ substantially in sensitivity; and whether the samples were processed in one batch or over months. Any of the three can be the largest source of variance in your re-analysis.
3. Re-run QC yourself
Do not trust the authors' filtering, even when it is documented. Re-compute per-cell metrics and look at the distributions. Published datasets routinely contain a sample with 30% mitochondrial reads that was kept because it did not matter for the original question — and does matter for yours.
4. Decide honestly whether datasets can be combined
Combining two studies of "the same tissue" is where most public-data projects go wrong. Different dissociation protocols change cell type proportions; different chemistries change detection; different tissue sites are simply different tissue. Integration will make the UMAPs overlap regardless — that is what integration does — so overlapping is not evidence that the combination was valid.
A workable test: integrate, then check whether a cell type you know is present in both datasets keeps a consistent marker profile, and whether the study of origin still predicts cluster membership after correction. If one study contributes essentially all cells of a cluster, treat that cluster as a study artefact until proven otherwise.
5. Keep the provenance
Record the accession, the file names, the download date and the checksums. Public datasets are revised more often than people expect, and "we downloaded GSE" with no date is not reproducible. This matters at revision time, when a reviewer asks why your cell counts differ from the original publication's.
6. Give credit precisely
Cite the original data paper, not just the accession, and follow whatever data-use terms apply — controlled-access human data has real conditions, and using it outside them is a problem no result is worth.
What re-analysis is genuinely good for
Validating a finding in an independent cohort, testing a hypothesis before spending money on an experiment, building a reference for label transfer, and answering the reviewer question "does this hold in published data?". Those are all strong uses. What it is not good for is manufacturing a new headline result from a dataset whose design was never intended to answer your question — reviewers can see the difference.