Navigate the key biological databases

In BIO-2.1 you learned to read the files. Now you need to know where they come from and which source to trust, because biology does not ship one canonical download. It ships a handful of public databases, each with its own scope, quality bar, and naming scheme, and a working bioinformatician moves between them all day. The goal here is not to memorize URLs. It is to build the two habits that separate a correct analysis from a plausible-looking wrong one: knowing what is authoritative, and keeping every identifier and coordinate consistent across your whole pipeline.

Databases are biology's package registries

Here is the analogy to carry through the lesson. These databases are the package registries and reference repositories of biology. NCBI's GenBank is the open, anyone-can-publish archive, the npm or PyPI of DNA sequence. RefSeq and Swiss-Prot are the curated standard library, one blessed version of each thing. dbSNP and ClinVar are catalogs of known variants, closer to a CVE database. And a genome build is the target platform you compile against: get it wrong and nothing links.

Accession numbers are the versioned package IDs. Just as you write numpy==1.26.4 to pin an exact release, biology writes NM_000207.3 to pin an exact record at an exact version. The name is stable and unique, the version bumps when the data is corrected, and the two together name one artifact for all time.

The analogy has a sharp failure edge, and it is the most important thing in this lesson, so we return to it below. Preview: a package manager refuses to link mismatched versions and throws an error. Biology's tools happily "link" a build mismatch and hand you fluent, wrong answers.

The sequence and annotation registries

Four names cover most of what you will touch.

NCBI (the US National Center for Biotechnology Information) is the giant, and it hosts two things beginners constantly confuse. GenBank is an archival database: researchers submit sequences, the submitter owns the record, and it is never silently overwritten. GenBank is redundant and can contain errors, because it is a faithful log of what people submitted, not a judgment about what is correct. It is authoritative as an archive, not as an answer. RefSeq (Reference Sequence) sits on top: NCBI curators distill the messy submissions plus the literature into one non-redundant, canonical record per molecule. When you want "the human insulin mRNA," you want the RefSeq, not a particular lab's GenBank deposit.

Ensembl (run by the European EMBL-EBI and the Wellcome Sanger Institute) is a genome browser and an independent gene annotation set. This matters: Ensembl and RefSeq are two different teams drawing gene boundaries on the same genome, and they sometimes disagree about where a transcript starts or which isoforms exist. Neither is "the truth." They are two well-built opinions, and a transcript from one does not automatically match the other.

UniProt (the Universal Protein Resource) is the protein hub: sequence plus curated functional annotation (what the protein does, where it lives, its domains). It splits into Swiss-Prot, manually reviewed, and TrEMBL, annotated automatically by software and not yet reviewed. Hold that split.

PDB (the Protein Data Bank) stores experimentally determined three-dimensional structures, solved by X-ray crystallography, NMR, or cryo-electron microscopy. A PDB entry measures real atoms in space. That differs in kind from an AlphaFold structure, which is a computational prediction, however good.

Curated versus predicted: not every record is equal

Notice the pattern that just repeated three times. RefSeq versus raw GenBank. Swiss-Prot versus TrEMBL. Experimental PDB versus predicted structure. Every major resource has a manually curated tier and a machine-generated tier living side by side, and they are not equally trustworthy. The machine tier is not junk. It is often the only data that exists, and for most genes in most species that is the situation. But it is a hypothesis, not an observation, and you must know which one you are holding.

Accession numbers are versioned package IDs

An accession is the stable handle you should carry between tools. Gene symbols are not stable: the committee that names human genes (HGNC) renames them fairly often, and famously renamed a batch (MARCH1 became MARCHF1, SEPT1 became SEPTIN1) specifically because spreadsheets kept auto-converting the symbols into dates. Accessions do not drift like that. Anatomy of one, for the human insulin gene:

NM_000207.3
| |     | |
| |     | +--  version: bumps when the record's sequence is corrected
| |     +----  (dot separates accession from version)
| +----------  accession: stable, unique, never reused for anything else
+------------  prefix: NM = curated mRNA (RefSeq)

other shapes you meet (all human insulin, INS, on chromosome 11):
  NP_000198        curated protein     (RefSeq)
  NC_000011.10     whole chromosome 11 (RefSeq, GRCh38 build)
  NC_000011.9      whole chromosome 11 (RefSeq, GRCh37 build)
  P01308           protein             (UniProt, Swiss-Prot = reviewed)
  ENSG00000254647  gene                (Ensembl)
  XM_...           an mRNA PREDICTED by software, not curated

Two things to read off that panel. First, the version suffix is doing real work: NM_000207.3 and a future NM_000207.4 are the same gene but potentially different bases, so citing an accession without its version is like pinning numpy with no version and hoping. (The exact digit you see today may already be higher than .3.) Second, the two chromosome-11 lines: NC_000011.9 and NC_000011.10 are the same chromosome in two different genome builds, and the build is baked into the version number. That is your bridge to the next section.

Genome builds: the same coordinate means different bases

A genome build (also called an assembly) is one agreed-upon reconstruction of a species' genome that everyone measures positions against. The current human builds are GRCh37 (Genome Reference Consortium human build 37, released 2009, also called hg19) and GRCh38 (2013, hg38). A newer complete assembly, T2T-CHM13, fills gaps the earlier builds left blank and is coming into use.

Why you cannot ignore this: a genomic coordinate is meaningless without its build. "Chromosome 1, position 100,000,000" points to a genuinely different base in GRCh37 than in GRCh38, because insertions and deletions between the assemblies shift everything downstream. Recall from BIO-2.1 that a VCF is a diff against a reference. That diff is only interpretable against the exact build it was called on. Mix a GRCh37 variant list into a GRCh38 annotation and you will tag real variants with the wrong genes, and nothing will complain.

The variant and clinical layer

Three resources dominate, and they answer three different questions. Do not blur them.

dbSNP (NCBI) is a catalog of observed short genetic variants, each given a stable rs identifier (like rs7903146). It tells you a variant has been seen and gives it a name. It makes no claim about whether the variant matters.

ClinVar (also NCBI) is the clinical-interpretation layer: it archives submitted assertions about whether a variant is pathogenic, benign, or of uncertain significance (VUS). Crucially, ClinVar is submitter-driven and labs can disagree, so every record carries a review status (rated from 0 to 4 stars). A one-star "pathogenic" from a single lab and a four-star expert-panel consensus are both "in ClinVar" and are not remotely the same evidence. Read the stars.

gnomAD (the Genome Aggregation Database, from the Broad Institute) reports how common each variant is across large populations of broadly healthy people. The logic is simple and powerful: a variant present in 5 percent of the general population is very unlikely to cause a rare severe disease, so allele frequency is a first-pass filter. gnomAD gives frequencies, not diagnoses, and each release is tied to a build (v2 on GRCh37, later versions on GRCh38), so the build discipline applies here too.

Where this leaves you

There is no single canonical download of biology. There is a small set of registries, each with a curated tier you can lean on and a predicted tier you must verify, all named by stable versioned accessions. The one discipline that prevents the worst class of error is consistency: every identifier and every coordinate in a pipeline must be on the same build and reference the same source, because the tools will not warn you when they are not. Carry the accession, pin the version, state the build. Next, in BIO-3, you will start comparing sequences directly, and every coordinate you produce will trace straight back to the choices you just learned to make.