From GEO Accession to Results

The Gene Expression Omnibus (GEO) is NCBI’s public archive for functional genomics data. It was launched in 2000 to hold microarray experiments² and now also holds a large share of the world’s published bulk RNA-seq studies^1,3. Most journals ask authors to deposit expression data there, so the dataset behind a paper you are reading is very often one accession number away.

GEO’s records nest inside each other. A Series (GSE) is one study: its title, summary, experimental design, contributors, linked publication and any files the submitters attached. A Series contains Samples (GSM): one record per biological sample or library, with its own annotations. Each Sample points to a Platform (GPL), which describes the technology: a specific microarray design, or for sequencing, an instrument model and organism. Related Series can be bundled under a SuperSeries, and for sequencing studies the raw reads live next door in the Sequence Read Archive (SRA).

Many people expect to download “the data” from GEO as one tidy table. For microarray studies that mostly works. For RNA-seq it usually doesn’t, and the reason comes from how GEO was originally built.

Every Series has a series matrix file: a text file with all sample annotations at the top and a data table underneath. For microarrays the table holds each sample’s processed intensity per probe, because GEO historically required those values on every Sample record. Sequencing studies are deposited differently: raw reads go to SRA, and processed data are attached as supplementary files in whatever format the submitter chose. So for most RNA-seq Series the series matrix carries the annotations but an empty data table. You can see it in GSE52778: its matrix file lists all 16 sample columns and zero data rows.

The supplementary files are the submitters’ choice, and the choice varies a lot. Some upload a clean raw-count matrix. Many upload only normalized values (FPKM, RPKM, TPM), differential-expression results, one file per sample bundled in a _RAW.tar archive, or Excel sheets with merged header rows. Count-based methods such as DESeq2, edgeR and limma-voom^6–8 need the raw integer counts, because they model the count noise themselves. FPKM/TPM tables have already removed the information those models need.

To fill that gap, NCBI now re-processes the raw SRA reads of many human RNA-seq Series through one uniform pipeline and offers the result as NCBI-generated raw counts: one gene-level count table per Series, aligned to the GRCh38 human reference, with GSM accessions as column names³. When this table exists it is usually the best starting point, because it is raw, complete and processed the same way for every study.

TransXplorer’s GEO import uses the Bioconductor package GEOquery⁴ to talk to GEO, and then applies its own rules to pick a matrix and build a metadata table. Knowing the order of those rules explains almost every result you will see, including the failures.

The rules, in detail

• Accepted input^GSE[0-9]+$ Spaces are trimmed and lower case is converted to upper case, so gse52778 works. An empty box gets “Please enter a GEO accession ID (e.g., GSE52778)”, and anything that is not a Series ID gets “Invalid GEO ID format. Please enter a Series ID like GSE52778.”
• Multi-platform serieslargest platform If the Series spans several platforms, the expression matrix comes from the platform with the most samples. Sample annotations are merged across all platforms (using only the columns they share), and when you confirm groups, any count column without matching metadata is dropped.
• NCBI-generated countsGRCh38.p13 file TransXplorer requests the human GRCh38 raw-count file for the accession and checks that what comes back is a real gzip table with at least two numeric sample columns. For non-human series this lookup simply finds nothing and the next source is tried.
• Supplementary filesname-based filter Runs only if the processing notes look like sequencing (words such as RNA-seq, counts, STAR, Salmon, featureCounts). Files whose names suggest DE results or normalized values (edger, deseq, diff, rpkm, fpkm, tpm, cpm, _vs_, logfc…) are skipped. Files named like count, raw_count, matrix or gene_expression are preferred, and the largest one is read. A generic .txt/.csv/.tsv file is only used when there are three or fewer candidates.
• Series matrix fallbackmax < 25 ⇒ microarray If the matrix table has values and the largest is below 25, the data are treated as log₂ microarray intensities and a note recommends limma. Otherwise they are labelled RNA-seq.
• Gene IDsversion suffix removed Trailing version numbers such as ENSG00000141510.17 → ENSG00000141510 are stripped from row names.
• MetadataGEO2R-style parsing Each characteristics line of the form key: value becomes a column named after the key. Administrative fields (contact, dates, protocols, file links) are dropped, as are columns where every sample has the same value. title and source_name_ch1 are always kept.

GSE52778 comes from Himes et al. (2014)⁵. Primary airway smooth-muscle cells from four donors were each left untreated or treated with albuterol, dexamethasone, or both, and profiled by RNA-seq. It is a good teaching study: the design is clean, it is paired by donor, and the submitters’ files show exactly the “FPKM only” problem from Section 3. We will compare dexamethasone vs untreated, the comparison the paper focuses on.

Accession

GSE52778NCBI GEO · GPL11154

Samples

164 donors × 4 treatments

Comparison

Dex vs Untreated4 vs 4, paired by donor

Organism

H. sapiensprimary airway smooth muscle

1. 01
   Load

   Type the accession and click Fetch

   Open the Transcriptome Analysis tab. In the left sidebar, under 1. Data Input & Setup, below the file uploader, is a blue box titled Or fetch from GEO. Type GSE52778 and click Fetch.

   UI replica

   Or fetch from GEO

   GSE52778 ↓ Fetch

   Supports RNA-seq datasets with NCBI-processed counts. Microarray data auto-detected.

   The button switches to Fetching… and a progress bar walks through the sources from Section 4. For GSE52778 the second source succeeds: NCBI has generated a raw-count table, so TransXplorer never needs the FPKM file. A notification confirms “GSE52778 loaded successfully!” with the matrix size, a data type of RNA-seq counts, and the number of metadata attributes. The sidebar then shows GEO data loaded: GSE52778; its × link clears the data.

2. 02
   Assign groups

   Choose the grouping column, and don’t trust the default

   A dialog titled Select Sample Grouping Column opens. It previews the first five samples’ metadata, offers a Grouping column: dropdown, and below it shows “Preview: N groups detected” with the sample count per group.

   TransXplorer pre-selects a column using a simple rule: it prefers columns with 2–6 distinct, short values. For GSE52778 that rule picks ercc_mix:ch1, a technical flag recording which libraries received ERCC spike-in controls. It has two values, so it looks like a perfect two-group design. It is not your biology. (We checked this by running TransXplorer’s selection rule on the live GSE52778 metadata in October 2026.)

   UI replica · simplified

   ▦ Select Sample Grouping Column

   GSE52778 loaded: genes × 16 samples
   Select the column that defines your experimental groups.

   Pre-selected (wrong for this study)

   ercc_mix:ch1▾

   Preview: 2 groups detected

   Group	Samples
   -	9
   1	7

   Grouping column: (your choice)

   treatment▾

   Preview: 4 groups detected

   Group	Samples
   Albuterol	4
   Albuterol_Dexamethasone	4
   Dexamethasone	4
   Untreated	4

   Cancel✓ Confirm Groups

   Switch the dropdown to treatment (the parsed column; treatment:ch1 holds the same values), check that the preview shows four groups of four, and click Confirm Groups. TransXplorer matches each count column to its metadata row by GSM accession, assigns the groups, and confirms with “Groups confirmed from treatment: …”.

3. 03
   Annotate

   Select the organism

   The GEO import does not set the species for you. In the Select Organism section choose Human from Select Organism: *. The asterisk means it is required. The organism controls gene annotation and which pathway databases are offered later.

   NCBI-generated tables use numeric NCBI Gene (Entrez) IDs as row names. TransXplorer’s symbol-annotation step recognises numeric Entrez IDs alongside Ensembl IDs, so with the organism set, your result tables can carry gene symbols such as TSC22D3.

4. 04
   Design

   Account for the donors

   Once groups are confirmed, a Batch Effects & Covariates Detection panel appears in the sidebar. Because this data came from GEO, it shows a green box, “GEO metadata available from GSE52778”, with a Use GEO Metadata for Batch Detection button. You don’t need to upload a metadata file.

   Here the “batch” is biological: four donors, each contributing one sample to every group. Donor-to-donor differences are often larger than the drug effect, so they belong in the model. Select Manual Selection, click the GEO metadata button, and in the Configure Sample Metadata dialog set Group Column (Required): to treatment and Batch Column (Optional): to cell_line, then click ✓ Confirm Metadata Configuration. Alternatively, leave Automatic Detection on and let TransXplorer score the GEO columns for you.

   When a batch column is in use, the DE model becomes ~ batch + condition for all three packages. Because every donor appears in both groups, the donor and treatment effects are not confounded, so TransXplorer keeps the donor term. If a batch column were confounded with the groups, it would warn you and leave it out. For the reasoning behind this, see Understanding Batch Effects.

5. 05
   Compare

   Pick reference and treatment

   With four groups, Comparison Method: offers Specific Comparison or All Pairwise Comparisons. Keep Specific Comparison and set Reference Group (Control): to Untreated and Treatment Group (Case): to Dexamethasone. Positive log₂ fold changes will then mean “higher after dexamethasone”.

   The other defaults are sensible starting points: DEG Analysis Package: edgeR with TMM normalization, an adjusted p-value (FDR) threshold with quick presets (0.01 / 0.05 / 0.1), and Min Log2FC Cutoff: −2 / Max Log2FC Cutoff: 2. With only four donors per group, an |log₂FC| cutoff of 1 gives a broader, more typical gene list. Why the three DE packages usually agree is covered in What is Differential Expression?.

6. 06
   Run

   Run QC, then differential expression

   Click Run Exploratory QC first. It normalizes the counts and computes PCA/UMAP and batch diagnostics, and the results open in the Exploratory Analysis results tab. On the PCA, look for samples clustering by donor as well as by treatment; that is the pairing at work. Then click Run DEG Analysis.

   If you click Run DEG Analysis before QC, an Exploratory Analysis Required dialog offers Run Exploratory Analysis First. QC only needs to run once per dataset. After that you can run several comparisons, for example Albuterol vs Untreated, without repeating it.

7. 07
   Read

   Volcano plot and gene tables

   Open DEG Results & Visualisation. The All DEGs, Upregulated Genes and Downregulated Genes tabs each have a download button (Download All DEGs and so on). In Volcano Plot, click Generate Volcano Plot to draw an interactive plot you can hover, zoom, and export with Download Plot.

   Sanity check against the paper. Himes et al. reported 316 differentially expressed genes for dexamethasone vs untreated (Benjamini–Hochberg adjusted p < 0.05). These included well-known glucocorticoid-responsive genes, DUSP1, KLF15, PER1 and TSC22D3, plus CRISPLD2, which they showed dexamethasone induces at both mRNA and protein level⁵. Look these up in your table or hover for them on the volcano. Your exact counts and p-values will differ: the authors used their own hg19 alignment and a different DE method, while NCBI’s table is aligned to GRCh38. The direction and rough ranking of these genes is what should agree.

   If the classic genes are missing, check three things before blaming the biology: the grouping column (step 2), the reference/treatment direction (step 5), and that the organism is set so symbols resolve (step 3).
8. 08
   Interpret

   Pathway enrichment

   Open the Pathway Enrichment results tab. The organism is carried over from your DEG run (“Auto-detected from DEG Analysis”). Under Gene Set to Analyze: keep Both Up & Down-regulated (Recommended) or split by direction. In Select Pathway Databases: the default for human is GO Biological Process 2023; you can add KEGG, Reactome, WikiPathways, MSigDB Hallmark and others. Then click Run Pathway Enrichment. Results appear as an Enrichment Table, an Enrichment Plot and a Gene-Pathway Network.

   That is over-representation analysis (ORA) on your significant gene list. For a threshold-free view, the top-level Enrichment Analysis tab has a GSEA mode that accepts a pre-ranked list. Export Download All DEGs and use gene symbol plus log₂FC as the ranking. When to prefer which is the subject of GSEA vs ORA.

GEO holds hundreds of thousands of Series deposited over more than two decades in very different formats, so no importer handles them all. The map below starts from what you actually see after clicking Fetch and leads to a fix. The numbered cards underneath give the detail for each branch.

1. “Invalid GEO ID format”, “not found”, or a network error Only Series IDs are accepted. A GSM (one sample), GPL (platform) or GDS (curated DataSet) ID is rejected before anything is downloaded. “GEO dataset … not found” usually means a typo or a Series that is still private. “Network error: Could not connect to GEO” is usually temporary: NCBI’s servers are sometimes busy, so wait a minute and try again. FixCopy the GSE from the paper’s data-availability statement. For a GDS, open it on GEO and use its reference Series.
2. “No Count Matrix Found for GSE…” The dialog says how many samples have metadata and shows the first few rows, but no source produced a usable matrix. Typical causes: NCBI hasn’t generated counts (or the series is not human), and the supplementary files are a per-sample _RAW.tar archive, an Excel workbook, a normalized-only table (FPKM, RPKM, TPM, CPM), or a file whose columns cover fewer than half the samples. The dialog itself suggests three ways forward: upload your own count matrix with this metadata, process the raw FASTQ files from SRA in the FASTQ Processing tab, or use uniformly processed resources such as recount3 or DEE2^9,10. FixIf a raw-count file exists in an awkward format, download it, tidy it into genes × samples, and follow Bringing your own data. If only reads exist, see FASTQ to results and the FASTQ Processing tab.
3. Microarray series When the matrix comes from the series-matrix table and its maximum value is below 25, TransXplorer labels it “Microarray (log2-transformed values)” and recommends limma⁸. Two caveats. First, the platform annotation table is not downloaded, so rows are the array’s probe IDs (for example 1007_s_at), and gene-symbol mapping and enrichment may not work. Second, older arrays deposited on a linear (non-log) scale exceed 25 and are labelled “RNA-seq counts” even though they are not counts. FixFor a serious microarray reanalysis, map probes to genes with the platform’s annotation first, then upload the gene-level table. Set Normalization Method: deliberately; already-normalized data may need No Normalization.
4. Normalized-only supplements (FPKM, RPKM, TPM) Supplementary files whose names contain fpkm, rpkm, tpm or cpm are skipped, so a series that offers only those ends at “No Count Matrix Found” instead of loading them as counts. A normalized table with a generic file name, or normalized values in the series matrix, can still get through, and the success notification will still say “RNA-seq counts”. Count models (DESeq2, edgeR, limma-voom) assume raw integer counts, and normalized values break their noise model. FixOpen the source file from the GEO page and look at a few values. Raw counts are whole numbers; FPKM/TPM have decimals. If you see decimals, look for NCBI-generated counts or re-quantify from SRA.
5. “No matching sample IDs between count data and metadata” Groups are attached by matching count-matrix column names to GSM accessions. NCBI-generated tables always use GSM IDs, but a submitter file might label columns Ctrl_1, KO_2… In that case Confirm Groups cannot link them and stops with this error. If only some columns match, the unmatched ones are dropped. FixDownload the file, rename its columns to the matching GSM accessions (the sample titles on the GEO page show the mapping), and upload it with a metadata file whose Sample column holds the same IDs.
6. SuperSeries, multiple platforms and mixed organisms A SuperSeries bundles several Series, often different assays (RNA-seq plus ChIP-seq, bulk plus single-cell) or species. GEO keeps platforms separate, and a platform is tied to one organism, so a human+mouse study has at least two. When a fetch returns more than one platform, TransXplorer uses the matrix from the platform with the most samples, which may not be the one you want. FixOn the SuperSeries page, find the list of SubSeries and fetch the specific GSE for your assay and species. Then set Select Organism: * to match.
7. The pre-selected grouping column is wrong The default favours columns with 2–6 short distinct values, so spike-in flags, sex, lanes or batch dates can win, as ercc_mix did for GSE52778. The preview turns red with “need at least 2 groups” if a column has only one value. FixRead the GEO page’s Overall design first (Section 2), then choose the column that matches it. If no single column defines your groups (for example, genotype × time), build a small metadata file with a Sample column of GSM IDs and a combined group column, upload it under Upload Batch Metadata File:, and pick that column as the Group Column.

1. Barrett T, Wilhite SE, Ledoux P, et al. NCBI GEO: archive for functional genomics data sets—update. Nucleic Acids Res. 2013;41(D1):D991–D995.
   doi:10.1093/nar/gks1193
2. Edgar R, Domrachev M, Lash AE. Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res. 2002;30(1):207–210.
   doi:10.1093/nar/30.1.207
3. Clough E, Barrett T, Wilhite SE, et al. NCBI GEO: archive for gene expression and epigenomics data sets: 23-year update. Nucleic Acids Res. 2024;52(D1):D138–D144.
   doi:10.1093/nar/gkad965
4. Davis S, Meltzer PS. GEOquery: a bridge between the Gene Expression Omnibus (GEO) and BioConductor. Bioinformatics. 2007;23(14):1846–1847.
   doi:10.1093/bioinformatics/btm254
5. Himes BE, Jiang X, Wagner P, et al. RNA-Seq transcriptome profiling identifies CRISPLD2 as a glucocorticoid responsive gene that modulates cytokine function in airway smooth muscle cells. PLoS One. 2014;9(6):e99625. (GEO: GSE52778)
   doi:10.1371/journal.pone.0099625
6. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15(12):550.
   doi:10.1186/s13059-014-0550-8
7. Robinson MD, McCarthy DJ, Smyth GK. edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics. 2010;26(1):139–140.
   doi:10.1093/bioinformatics/btp616
8. Ritchie ME, Phipson B, Wu D, et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 2015;43(7):e47.
   doi:10.1093/nar/gkv007
9. Wilks C, Zheng SC, Chen FY, et al. recount3: summaries and queries for large-scale RNA-seq expression and splicing. Genome Biol. 2021;22(1):323.
   doi:10.1186/s13059-021-02533-6
10. Ziemann M, Kaspi A, El-Osta A. Digital expression explorer 2: a repository of uniformly processed RNA sequencing data. GigaScience. 2019;8(4):giz022.
    doi:10.1093/gigascience/giz022
11. Lachmann A, Torre D, Keenan AB, et al. Massive mining of publicly available RNA-seq data from human and mouse. Nat Commun. 2018;9(1):1366.
    doi:10.1038/s41467-018-03751-6