3  Session 3: Biological Databases

Note

Sources for this page: Petras Kundrotas’s KB8029 lecture “Databases of Sequence Information” (Feb 2026) — this page covers its databases and worked examples, reorganized around five questions and adapted for this book, not copied verbatim; Wikipedia, Biological database, Sequence database, International Nucleotide Sequence Database Collaboration, GenBank, UniProt, RefSeq, Accession number (bioinformatics), Genome browser, UCSC Genome Browser, Ensembl genome database project, PubMed, PubMed Central, MEDLINE, BioRxiv, Boolean search — CC BY-SA 4.0; NCBI, GenBank overview, About RefSeq, and RefSeq FAQ — public domain (US government work); EMBL-EBI Training, What is ENA? and The UniProt databases — CC BY 4.0. Every accession number, record and search-result count quoted below was retrieved live from the NCBI, UniProt, InterPro, RCSB, AlphaFold DB and IntAct APIs (2026-10-04); the notebook contains the code that produced them. The UniProt protein-existence levels follow UniProt’s documentation.

TipNaming things

Why do databases trust accession numbers more than gene names? Ask the gene Pokemon, renamed Zbtb7 in 2005 after Pokémon’s owners objected.

3.1 Learning goals

By the end of this session you should be able to:

  • follow one gene or protein across databases, from its genomic and transcript records to its protein entry, family, structure, interactions and literature, using the cross-references between them;
  • tell apart a gene symbol, a database accession, an accession with its version, and a database-specific entry name, and explain why one biological entity has many linked records;
  • say where a particular claim in a record comes from: data submitted by the people who produced it, curated by database staff, predicted by a program, or experimentally supported, and recognise the signs of each (RefSeq NM/NP vs. XM/XP, Swiss-Prot vs. TrEMBL, UniProt’s protein-existence level, an experimental structure vs. a predicted model);
  • build a PubMed query with Boolean operators and quotation marks, and predict its result before running it.

Rather than listing databases one by one, the session follows one protein, human beta-globin (HBB), through five questions: Where is the sequence? What is known about the protein? Which family does it belong to? Is there a structure? What does it interact with? Each database then appears as the answer to one of these questions.

3.2 Why coordinated, worldwide databases

Sequencing produces far more data than any single lab, or any single country, can curate. UniProtKB’s unreviewed section (TrEMBL) alone holds about 149 million protein entries. A private copy of “the” database would therefore go stale the moment anyone else submitted a sequence, so the databases have to be shared.

The field’s answer is the International Nucleotide Sequence Database Collaboration (INSDC), which joins GenBank (NCBI, USA), the European Nucleotide Archive (ENA, EMBL-EBI, Europe) and DDBJ (Japan). The three exchange all new data every day and follow shared annotation standards. A sequence submitted to any one of them can therefore be found through all three, under the same accession. Each presents it in its own way, though; GenBank and ENA, for example, use different flat-file formats and different web pages. Access has been free and unrestricted from the start, and since the 1990s most journals have required new sequences to be deposited in an INSDC database before publication. This is why a sequence you look up today reflects essentially everything published on it.

3.2.1 Further reading

3.3 Two ideas you need for every database

3.3.1 Identifiers: names, accessions, versions and cross-references

The most useful skill of this session is to know what kind of identifier you are holding, because one gene has many. For HBB alone, the table lists seven:

Table 3.1: Seven identifiers for the human HBB gene and its products.
Identifier Kind What it means
HBB approved gene symbol assigned by the HGNC (HUGO Gene Nomenclature Committee) for human genes, and widely used across databases (which may also list synonyms)
3043 NCBI Gene ID NCBI’s number for the gene record
NM_000518 accession one RefSeq transcript record; unique only within RefSeq
NM_000518.5 accession.version the fifth version of that record; the version changes whenever the sequence changes
P68871 UniProt accession the stable identifier of the UniProt entry
HBB_HUMAN UniProt entry name a readable label for the same entry, which can change
2HHB PDB ID one experimental structure; HBB appears in hundreds of them

An accession is unique only within its own database. The same gene has different accessions in RefSeq, Ensembl, UniProt and the PDB, and the records are tied together by explicit cross-references. The RefSeq protein record NP_000509.1, for example, states that it was translated from NM_000518.5, and the UniProt entry P68871 lists both. One biological entity thus has many linked records.

Two habits follow from this. First, to move between records, follow an explicit cross-reference when one exists. Otherwise, search with a stable identifier or the approved symbol (and the organism), and check the organism and sequence of what you find. Second, when you report a sequence, give the accession with its version, because NM_000518 alone does not say which revision you used. The same holds for genome coordinates, which only mean something together with the genome build (GRCh37/hg19 or GRCh38/hg38). The lab shows the same gene at different coordinates in the two builds.

RefSeq accessions also encode the kind of record in their prefix:

Table 3.2: RefSeq accession prefixes and the kind of record they denote.
Prefix Record Example
NC_ genomic: a complete chromosome NC_000011.10
NM_ mRNA, not predicted (many curated) NM_000518.5
NP_ protein, not predicted (many curated) NP_000509.1
NR_ non-coding RNA, not predicted —
XM_ / XP_ mRNA / protein predicted by NCBI’s genome-annotation pipeline XP_011536997.1

3.3.2 Where does a record’s information come from?

Databases differ less in what they store than in who vouches for it. A claim in a record can come from four sources, listed below from the least to the most checked. These sources describe where a particular piece of information comes from, not four separate kinds of record. One Swiss-Prot entry, for example, mixes curated, experimental and predicted annotation, and a PDB entry is both submitted and experimental.

  • Submitted (archival). INSDC stores sequences close to how the submitters deposited them, together with their annotation. Nobody else revises it, so the quality varies, and the same gene may appear in many overlapping records.
  • Predicted. A program inferred it. An XM_/XP_ record says so itself: “MODEL REFSEQ: This record is predicted by automated computational analysis.” Most TrEMBL entries are translations of predicted coding sequences, annotated automatically. An AlphaFold model is a predicted structure.
  • Curated. Database staff have read the evidence and merged the records into one reviewed entry. RefSeq is a non-redundant collection built from INSDC data (it is not itself part of INSDC). Its NM_/NP_ records are the non-predicted class, many of them curated or supported by transcript and protein evidence. Its XM_/XP_ records, which sit side by side with them, are computational models. Swiss-Prot entries are reviewed by UniProt curators.
  • Experimentally supported. Somebody measured it. A PDB entry holds coordinates fitted to experimental data, an IntAct interaction comes from a published experiment, and UniProt marks the proteins whose existence has been observed directly.

UniProtKB makes the distinction between curated and predicted explicit, because it has two sections. Swiss-Prot holds about 576,000 entries reviewed by curators, in which a person has checked the function, domains, modifications and literature, with one entry per protein. TrEMBL holds about 149 million unreviewed, automatically annotated entries. Swiss-Prot is not a subset of TrEMBL. The two are separate sections, and when curators review a TrEMBL entry, it is integrated into Swiss-Prot and leaves TrEMBL. Every entry also has a protein-existence level, graded from the strongest evidence to the weakest: 1, experimental evidence at protein level; 2, at transcript level; 3, inferred from homology; 4, predicted; 5, uncertain. HBB (P68871) is at level 1.

The discussion problems practise both ideas. You name the level and kind of four RefSeq records from their prefixes, and then find a UniProt entry from an accession or from the gene symbol.

3.3.3 Further reading

3.4 Following one protein through five questions

3.4.1 Where is the sequence?

The first question is where the sequence is, and the answer depends on the level. Databases are organised by the central dogma, so DNA, RNA and protein each have their own records. For HBB, NCBI Gene (3043) places the gene on chromosome 11 (11p15.4). The chromosome itself is NC_000011.10, the curated transcript is NM_000518.5 and its protein is NP_000509.1. The raw submitted sequences behind them sit in INSDC, while RefSeq and Ensembl provide the annotated, non-redundant view of the genome, with the gene models marked out.

To see a gene in its genomic context, you use a genome browser. The UCSC Genome Browser and Ensembl show the sequence and its annotation as a function of position along a chromosome, for a chosen organism and genome build.

3.4.2 What is known about the protein?

UniProt is the central protein resource. P68871 (HBB_HUMAN) is a Swiss-Prot entry, so its function, the diseases its variants cause (such as sickle-cell anaemia), its modifications and its literature have been checked by curators. It also links to RefSeq, Ensembl, InterPro, the PDB, AlphaFold DB, IntAct and many more. In practice, the reviewed UniProt entry is therefore the best first stop for “what is known about this protein?”.

A third UniProt component, UniParc, stores every unique protein sequence once, whichever database it came from, under an identifier that is never reused. It is how you recognise that records in different databases hold the same sequence.

3.4.2.1 Further watching

Welcome to UniProt (UniProt’s own introduction to the resource):

3.4.3 Which family does it belong to?

A family or domain database tells you which group a sequence belongs to, which is often the fastest way to learn something about a new or poorly characterised protein. InterPro (EMBL-EBI) integrates protein family, domain and site signatures from over a dozen member databases. One of them is Pfam, whose own website was retired in 2023 and merged into InterPro. P68871 matches five InterPro entries, nested from broad to specific:

Table 3.3: The five InterPro entries that match β-globin (P68871), from broad to specific.
InterPro accession Name
IPR009050 Globin-like superfamily
IPR012292 Globin/Protoglobin
IPR000971 Globin
IPR050056 Hemoglobin and related oxygen transporters
IPR002337 Hemoglobin, beta-type

Most of these signatures are profile hidden Markov models, the method that Session 5 explains. For non-coding RNA, Rfam plays the same role. It uses covariance models, which score the conserved base-pairing pattern together with the sequence, because an RNA’s function depends on its fold. HBB is protein-coding and has no Rfam entry; Rfam’s tRNA family, for comparison, is RF00005.

3.4.4 Is there a structure?

The question of structure has two very different kinds of answer:

  • Experimental structures are stored in the Protein Data Bank (PDB). Entry 2HHB is human deoxyhaemoglobin, two alpha and two beta chains, determined by X-ray diffraction at 1.74 Å resolution (Fermi et al. 1984). Resolution describes the experimental data: the smaller the number, the finer the detail the data could resolve.
  • Predicted models are stored in the AlphaFold Protein Structure Database, for essentially every protein in UniProt. Each residue carries a confidence score, pLDDT (0-100). Above 90, the model is expected to be accurate down to the side chains, and above 70 the backbone is usually right. Below 50, the model confidence is very low, and intrinsic disorder is one common reason (Session 12). The model for P68871 has a mean pLDDT of 97.

A resolution and a pLDDT are therefore not the same kind of number. Resolution measures how much detail the experiment could see in a real molecule, whereas pLDDT is the model’s own estimate of how right its prediction is. For HBB, which has hundreds of experimental structures, the model adds little. AlphaFold DB matters most for the many proteins that have none. Session 6 treats these coordinates as data, aligning structures, measuring them and searching for structural relatives, and Session 12 shows how AlphaFold predicts them.

3.4.5 What does it interact with?

Interaction databases, too, differ in the kind of evidence they hold. They are listed here from the most to the least direct evidence:

  • IntAct (EMBL-EBI, part of the IMEx consortium) curates individual physical interactions from published experiments, such as that of HBB with HBZ (hemoglobin subunit zeta).
  • Complex Portal (IntAct’s sister resource) curates stable complexes with their composition. CPX-2158, “Hemoglobin HbA complex”, consists of two HBA1, two HBB and four haem groups, the same object whose atoms 2HHB shows.
  • STRING combines curated interactions, co-expression, conserved gene neighbourhood, predictions and text mining into one score per protein pair. It answers a broader question, “which proteins are functionally associated?”, with a mixture of experimental and predicted evidence. FunCoup (Stockholm University) builds a similar network in a different way, and a later course in this programme covers it.

3.4.6 The dossier

Put together, the answers to the five questions form a short database dossier for HBB. Each row is a different database’s record of the same biology, and each is reached from the previous one by a cross-reference:

Table 3.4: The databases that answer each question about β-globin, and the main evidence behind each answer.
Question Database Record Main evidence for the claim
Where is the sequence? NCBI Gene; RefSeq 3043; NC_000011.10 → NM_000518.5 → NP_000509.1 curated (RefSeq)
What is known about the protein? UniProt (Swiss-Prot) P68871 (HBB_HUMAN), protein existence level 1 curated; existence experimental
Which family? InterPro IPR000971 (Globin) and four more signatures built by curators, matches computed
Is there a structure? PDB; AlphaFold DB 2HHB (X-ray, 1.74 Å); AF-P68871-F1 (mean pLDDT 97) experimental; predicted
What does it interact with? IntAct; Complex Portal HBB-HBZ; CPX-2158 (HbA) experimental; curated
Where is it described? PubMed PMID 6726807 (Fermi et al. 1984, J Mol Biol 175:159), the paper behind 2HHB literature

The notebook fetches every one of these records live, and the lab ends with a dossier of your own, for human myoglobin.

3.4.7 Further reading

3.5 Searching: know your level, predict, then check

A search at the wrong level buries what you want. NCBI’s or EBI’s everything-at-once search for “hemoglobin beta” mixes genome, gene, transcript and protein records, whereas the same words in NCBI Gene or UniProt give a short list with the record you want near the top. Once you know the level, you should therefore use that database’s own search. When you need many records, or want a search to be reproducible, use the programmatic interfaces instead (NCBI’s E-utilities through Biopython’s Bio.Entrez, and the UniProt and PDB REST APIs), as the notebook does.

The same applies to the literature. PubMed, built by NCBI on the NLM’s MEDLINE database (started 1971), indexes the biomedical literature, with just over 41 million records. Its free full-text archive, PubMed Central (PMC), holds about 12.6 million, roughly a third. A PubMed hit is therefore not necessarily a free paper. When PMC has no copy, a preprint server such as bioRxiv often does.

A good habit for every database is to predict what a query will return, run it, and check whether the result makes sense. PubMed’s Boolean operators make this easy, because they behave like operations on sets:

Table 3.5: PubMed queries combined with Boolean operators.
Query Result count Prediction
hemoglobin 288,577 —
malaria 125,442 —
hemoglobin AND malaria 4,458 both terms present
hemoglobin OR malaria 409,561 288,577 + 125,442 − 4,458 (union)
hemoglobin NOT malaria 284,119 288,577 − 4,458 (difference)

Quotation marks are the second tool. The query hemoglobin beta is read as hemoglobin AND beta, that is, both words anywhere, in any order. The query "hemoglobin beta" requires the exact phrase, and it usually returns far fewer records. The third discussion problem asks you to predict the counts for a new pair of terms before running the searches.

3.5.1 Further watching

PubMed Overview (Johns Hopkins Welch Medical Library):

3.5.2 Further reading

These resources are not needed for this course, but you will meet them:

  • Variants: dbSNP and ClinVar (NCBI; ClinVar adds clinical significance), the European Variation Archive (EMBL-EBI), and the GWAS Catalog, which records statistical associations between variants and traits from genome-wide association studies. OMIM (Johns Hopkins) is the classic reference for genes and Mendelian disease.
  • Cancer genomics: TCGA, through NCI’s Genomic Data Commons, and COSMIC (Sanger Institute) for somatic mutations in cancer.
  • Systems: PRIDE (mass-spectrometry proteomics data), Reactome (curated pathways) and BioModels (computational models of pathways).

3.6 The course notebook

notebooks/session03-databases-in-code.ipynb builds the HBB dossier with live code. It retrieves the gene, mRNA and protein records from NCBI through Bio.Entrez, the UniProt entry with its cross-references and protein-existence level, the InterPro matches, PDB entry 2HHB through two different APIs (with its primary citation) and the AlphaFold model, the IntAct and Complex Portal records, and STRING. It then counts the Swiss-Prot and TrEMBL entries and runs the Boolean PubMed searches above, checking the set arithmetic in code. Open it in Google Colab via the badge at the top of the notebook page.

3.7 Check your understanding

3.8 The lab

The lab follows the same path on other genes and proteins. It covers identifiers and genome builds, predicted and curated records, a reviewed UniProt entry and its protein-existence level, InterPro domains, an experimental structure and an AlphaFold model, complexes and networks, and a literature search. It ends with a database dossier for human myoglobin, in which you also say what kind of information each record holds.

3.8.1 Submitting this lab

This lab is administered through Canvas, not through this book. See the course’s Canvas page for the current worksheet, the deadline and how to submit.

3.9 What’s next

Session 4 picks up where the cross-references leave off. Once you can retrieve two related sequences, the next question is how to line them up and score how similar they are, and that is sequence alignment.