European Nucleotide Archive

One archive, every base its federation accepts. The European Nucleotide Archive (ENA) is Europe's node of the International Nucleotide Sequence Database Collaboration - the INSDC arrangement under which Europe, Japan and the United States exchange every submission - so a run deposited anywhere in that federation resolves here, whether it entered through GenBank, DDBJ or directly. Holdings cover raw sequencing reads, assembled genomes and the annotation draped over both, with each record carrying its experiment and sample context.

Scale is the headline figure: a petabyte-scale store of raw reads spanning tens of millions of run records, with lineage back to the collaboration's earliest days in the early 1980s. Every row keys into a four-level hierarchy - study, sample, experiment, run - so "which reads sit behind this project" is one join rather than a literature hunt. Get a sample of this dataset and inspect real run and assembly records before committing pipeline time.

Three real records, straight from the archive - two runs and one assembly:

run_accession    : SRR017227
study_accession  : PRJDA36485
sample_accession : SAMN00002358
organism         : Homo sapiens
tax_id           : 9606
instrument_model : Illumina Genome Analyzer
library_layout   : SINGLE
read_count       : 2,296,574
base_count       : 57,414,350

run_accession    : DRR000013
study_accession  : PRJDA34559
sample_accession : SAMD00010993
organism         : Homo sapiens
instrument_model : Illumina Genome Analyzer
library_layout   : SINGLE
read_count       : 15,490,013

assembly_accession : GCA_007821485
study_accession    : PRJNA481794
assembly_name      : HG04217_prelim_1.0
assembly_level     : contig
last_updated       : 2019-07-31

Read the prefixes before the numbers. The SRR run and the DRR run were accepted by different INSDC partners, yet they land side by side in one row shape, because the federation shares accession namespaces across its members. Both carry the sizing pair - read_count and base_count - that lets a pipeline budget storage and compute from metadata alone. The third record is the archive's other half: an assembly, GCA-prefixed and graded contig, tied back to its parent study. Two runs, one assembly, three different grains - that spread is the norm, not the exception. The example rows are illustrative of the record shape; field-level detail is documented below.

Fourteen core fields, every one with a verified definition. The working shape of a record: accessions anchoring it at each level of the hierarchy, organism identity to route it, instrument and library descriptors explaining how the data was produced, read and base counts sizing it - and on the assembly side, an accession plus completeness level placing it among reference genomes. last_updated stamps recency onto each row.

Geography - global, because the federation is: submissions arrive from international sequencing centers and consortia on every continent, pooled through the partner archives. Organism, more than country, is the axis that matters here - coverage tracks the world's laboratories, and tax_id is how you slice them.

Temporal - deep by construction: records extend back to the early 1980s, when the collaboration began exchanging sequences, and run forward through current submissions. Each row carries last_updated, so recency is a field you filter on rather than a property you guess at.

Granularity - a strict four-level hierarchy. Studies (PRJ-prefixed) contain samples, samples carry experiments, experiments produce runs (SRR/DRR-style), and a parallel assembly layer (GCA) grades finished genomes by completeness level. Tens of millions of run records sit at the bottom of that stack, and every level above them is addressable by accession. Scale check from the same inventory: the raw-read store weighs petabytes, which is precisely why the count fields exist - to size work before starting it.

API, files, or your warehouse. Daily, weekly, or hourly.

• Reproducibility audits - study_accession retrieves the exact run set behind a published paper, and read_count/base_count on each row check the claimed sequencing depth against what was actually deposited. Discrepancy becomes a query, not a peer-review hunch.
• Sequencing-economics benchmarking - instrument_model joined with base_count turns tens of millions of runs into a yield-per-platform history: which machines produced how much sequence, when, for which organisms. Procurement and competitive-intel teams price capacity decisions off that curve.
• Biodiversity and surveillance rollups - scientific_name and tax_id aggregate the archive by organism, so "who is sequencing what, and how deeply" answers as a GROUP BY rather than a survey of papers.
• Reference-building and pan-genome work - the assembly layer supplies candidates with completeness already graded (contig, scaffold, chromosome), from draft assemblies like HG04217_prelim_1.0 upward.
• Retrieval and QC model training - run metadata arrives labeled and dense enough to train models that predict cost, quality or organism mix before any sequence is touched.

Bioinformaticians and computational biologists get the accession spine their pipelines already speak - study, sample, experiment, run - so reproducing a paper's inputs is a filter instead of a correspondence chain. Data scientists and ML engineers get tens of millions of labeled metadata rows for training retrieval, QC and forecasting models; the data scientists working in biotechnology briefing goes deeper. Developers building genomics products get permanent, globally unique accessions to key customer-facing features on - identifiers that outlive reorganizations; see developers building in biotechnology. Start from the biotechnology data hub or the best biotechnology datasets shortlist to see where this archive sits in the wider stack.

Three quirks worth knowing upfront. First, "a record" needs choosing: study, sample, experiment, run and assembly are distinct grains, and mixing them inflates counts fast - fix the grain before scoping the sample. Second, accession prefixes encode provenance (SRR versus DRR tells you which partner took the original submission), useful context but never a second identifier - the accession itself is unique across the federation. Third, weight: raw reads are petabyte-scale by design, which is why read_count and base_count ship on every run row - budget from metadata, then scope the request to the studies, taxa or assembly levels you actually plan to join.