How to Download Public Sequencing Data: SRA, GEO, and ENA Step by Step

Quick answer: Public sequencing data lives in three connected places — the Sequence Read Archive (SRA, mirrored by ENA in Europe and DDBJ in Japan) holds raw reads, GEO holds curated studies with processed supplementary files, and BioProject/BioSample hold the metadata that ties them together. To get data into your pipeline: resolve the study to its runs, install SRA Toolkit, then prefetch a run and convert it with fasterq-dump --split-files. Faster for most of the world: skip the toolkit entirely and pull the same FASTQ files over HTTPS from ENA’s portal API, which lists exact URLs, byte sizes and MD5 checksums in one request.

The hands-on companion to bioinformatics databases on this blog — that guide explains what each archive is for; this one gets the bytes onto your disk, ready for the NGS QC step that always comes next.

Know your accession: which string means what

Every download problem starts as an identifier problem. The prefixes are not interchangeable:

The rule: papers cite the study (PRJNA, GSE), tools download the run (SRR). The study-to-run mapping is public metadata — SRA Run Selector shows it as a table, and ENA’s API returns it as a file, which is the first thing to fetch before downloading anything. Taking two minutes to read the run table saves hours of downloading the wrong libraries: it tells you each run’s layout (single or paired), platform, library strategy (RNA-seq, WGS, amplicon), and size in gigabytes.

Route 1: SRA Toolkit

NCBI’s toolkit is the canonical path, installable from bioconda or the NCBI installer:

conda install -c bioconda sra-tools

# pull the raw SRA object, then convert to FASTQ
prefetch SRR1234567
fasterq-dump SRR1234567 --split-files --threads 4 --outdir fastq/

--split-files writes _1 and _2 for paired runs and a plain file for single-end. Two practical notes: fasterq-dump needs temporary space roughly equal to the output size, so point --temp at a disk that can take it; and for a whole project, loop the runs from the metadata table rather than guessing accessions one at a time. The older fastq-dump still appears in tutorials — it is slower and single-threaded, and new work should use fasterq-dump.

One caveat on scale: prefetching a hundred-gigabyte study one run at a time is slow. When the dataset is large, the next route is usually the better one.

Route 2: ENA’s API (usually faster)

ENA mirrors the same reads in Europe and publishes a plain HTTPS index for every accession — study, sample or run:

curl -s "https://www.ebi.ac.uk/ena/portal/api/filereport?accession=SRR1234567\
&result=read_run&fields=run_accession,fastq_ftp,fastq_bytes,fastq_md5" -o ena.tsv

The response contains direct ftp.sra.ebi.ac.uk URLs, per-file byte counts, and MD5 checksums — everything needed for a resumable, verifiable download:

# fastq_ftp holds ;-separated URLs; fetch them with curl and check sizes
awk -F'\t' 'NR>1 {n=split($2,u,";"); for(i=1;i<=n;i++) print u[i]}' ena.tsv |
while read -r url; do
  curl -L -C - -O "https://$url"
done

Because ENA returns the same runs as SRA (they are mirrors, not different data), this route is not a workaround — it is the standard method outside North America, and it sidesteps Toolkit’s temp-space requirement entirely. It also slots directly into workflows: the same API call inside a Nextflow or Snakemake process turns data acquisition into a reproducible pipeline step, as described in the workflows guide.

Route 3: GEO’s processed files

GEO is a curation layer, not a read archive. A GEO Series page carries three kinds of material:

• Processed data — count matrices, differential-expression tables, peak files, in each sample’s Supplementary file section or the series FTP directory. Often this is all an reanalysis needs.
• Raw reads — usually not stored in GEO itself; the record links out to its SRR runs in SRA. If a paper says “raw data available at GSE…”, follow the link table to the runs.
• Series matrix files — expression values plus sample annotations, downloadable from the series FTP folder for quick reprocessing without touching FASTQ.

# GEO supplementary files live under a bucketed FTP path (nnn = last 3 digits)
wget -O sample_suppl.tar.gz \
  "https://ftp.ncbi.nlm.nih.gov/geo/samples/GSM12345nnn/GSM1234567/suppl/sample_suppl.tar.gz"

Download the processed matrix first and inspect it; most “I need the raw data” requests turn out to need only the matrix plus the metadata table.

Verify before you analyse

Three checks, all cheap, all worth doing before a week of compute:

• Size. Compare downloaded bytes against fastq_bytes (ENA) or the Run Selector’s reported size. Truncated downloads are the most common silent failure — partial FASTQs parse surprisingly far before erroring.
• Checksum. ENA publishes MD5 per file; md5sum the file and compare. A mismatch means a corrupted transfer, not a strange sample.
• Record count. seqkit stats or a quick wc -l (divided by four for FASTQ) against the run’s spot count. Then run FastQC immediately — checking quality at download time tells you whether any later anomaly came from the data or from your analysis.

Storage discipline

Raw reads are the most expensive part of a project to keep. A single human RNA-seq lane is tens of gigabytes; a public project can be terabytes. Three habits that pay off: download only the runs you will use (the metadata table lets you filter by library strategy, layout and size); compress and archive only what passes QC, deleting intermediates that can be re-derived; and record the accession list with MD5s alongside your results — the moment your analysis stops being trivially re-runnable, that manifest becomes the difference between reproducible science and a folder of unexplained files.

Where the data comes from in a real pipeline

For most analyses the sequence is: resolve study → fetch run table → filter runs → download FASTQ with checksums → QC → process. The public archives make step one through four free and scriptable; the skills around them are the same terminal, file-format and pipeline skills the rest of this blog builds — the Linux for bioinformatics guide covers the shell mechanics these commands assume.

Key takeaways

• Papers cite studies (PRJNA, GSE); downloads use runs (SRR). Resolve one to the other before fetching bytes.
• SRA Toolkit (prefetch + fasterq-dump --split-files) is canonical; ENA’s portal API is usually faster and needs no temp space, returning URLs, sizes and MD5s in one call.
• GEO holds processed matrices and links to raw runs — check the matrix before downloading terabytes.
• Verify size, checksum and record count before analysis; a truncated FASTQ looks like a biology problem later.
• Human genomic data is often controlled-access (dbGaP, EGA): application required, not just a link.
• Always cite the accession in publications, and respect the original consent and any data-use conditions.

Challenge

Take a GEO accession from a paper in your field and, without clicking through any download buttons, produce a manifest: every run it contains, layout, platform, size, and an MD5 for each file — using nothing but ENA’s API and shell tools. Then download one small run, verify its checksum, and run FastQC on it. That manifest-and-verify loop is exactly what a reproducible reanalysis starts with, and it takes minutes once the API call is muscle memory.

Want guided practice turning public data into a defensible analysis? Ampersand Academy offers one-to-one mentoring in bioinformatics — part of Mahadhi, a development and digital marketing company.

FAQ