# GEO: Finding Public Gene Expression Data

> How to read Series, Samples, and accessions in NCBI GEO and download analysis files.

GEO (Gene Expression Omnibus) is an **NCBI repository where researchers publish the gene expression data and experimental descriptions used in papers**. Just as you find a project in a code repository, in GEO you find data and metadata through study-level accessions.

> **First step in GEO**: confirm that the research question and sample composition fit your analysis, then choose the file that matches the processing stage you need.

## An accession is the address of the data

Each GEO page has a stable identifier such as `GSE251845`. This is called an **accession**. Because it points to a dataset more precisely than a paper title, record the accession in searches, code, and analysis notes.

| Prefix | Unit | In a structure familiar to developers |
| --- | --- | --- |
| `GSE` | A study grouping multiple samples, called a Series | Project |
| `GSM` | One sample | Data object inside a project |
| `GPL` | Platform used for measurement | Build environment or equipment specification |

One GSE contains multiple GSMs. Read the overall study design on the GSE page, then inspect the tissue, treatment conditions, and files for a specific sample on its GSM page.

## Four things to check first on the page

1. Check the research question and comparison groups under **Summary** and **Overall design**.
2. Under **Samples**, check the sample count, tissue, conditions, and whether samples are matched by patient.
3. Under **Supplementary file**, check for processed files such as a count matrix.
4. Follow the **SRA link** to see whether the FASTQ files can be downloaded again.

Public availability alone does not make data comparable. You must also determine whether samples are tumour and adjacent normal tissue, tissues from different people, and whether they used the same sequencing protocol and genome build.

## Should you download the count matrix or FASTQ?

If the learning goal is differential-expression statistics, starting from a researcher-provided **[raw count matrix](/en/reference/htseq-counts/)** is usually best. The download is small and you can skip alignment and quantification.

To reproduce the QC, alignment, and quantification pipeline, start with **FASTQ from SRA**. This requires much more storage and computation, and you must apply the same pipeline to every sample. The workflow after FASTQ continues in [Checking FASTQ Quality and Trimming](/en/lessons/rna-seq-to-matrix/) and [Aligning RNA Reads to the Genome](/en/lessons/rna-seq-alignment/).

:::caution[Do not trust the word "processed" alone]
`processed data` can include raw counts, TPM, or a table that has already undergone statistical analysis. Open the file, inspect its row and column names and value ranges, and verify that it is the input required by your analysis tool.
:::

## What to record for reproducibility

- GSE and GSM accessions
- File names and download date
- Sample inclusion and exclusion criteria
- Rules used to construct condition and patient IDs
- Whether values are raw counts or normalized
- Genome build and quantification method reported in the paper and GEO page

### Official documentation

- [Introduction to NCBI GEO](https://www.ncbi.nlm.nih.gov/geo/info/overview.html)
- [GEO data structure and accessions](https://www.ncbi.nlm.nih.gov/geo/info/overview.html#DataOrganization)
- [Help for searching and downloading GEO data](https://www.ncbi.nlm.nih.gov/geo/info/download.html)