Bioinformatics

Bioinformatics Databases, Tested: How to Use 13 of Them on One Protein (API Code, Speed, Licences)

I followed one human protein, SOD1, through 13 bioinformatics databases by their APIs: what each one gives you, how to query it, how fast and reliably it answered, which release you get, and the licence.

·10 min read
#bioinformatics databases#UniProt#Ensembl#NCBI#PDB#AlphaFold#STRING#Reactome#KEGG#API#Python

Rewritten in October 2026. The earlier version of this page was a generic list. This one is a measured walkthrough: every result, timing and version below comes from API calls I made on 13 October 2026, and the code shown was run as printed.

Lists of "databases every researcher should know" are easy to find. What they rarely tell you is what you actually get back for your protein, how you ask for it, and what goes wrong when you do it at scale. So I picked one well-studied human protein, SOD1 (Cu/Zn superoxide dismutase; UniProt P00441, NCBI Gene 6647), and asked 13 public databases about it through their APIs.

The Short Version

DatabaseUse it forWhat came back for SOD1Median responseAnsweredRelease on 13 Oct 2026Licence
UniProtThe protein: sequence, function, IDs for everything else154 aa, entry version 279, 556 cross-references3.3 s10/102026_03CC BY 4.0
EnsemblGene coordinates, transcripts, orthologuesENSG00000142168, chr21:31,659,617–31,668,931 (GRCh38)8.1 s6/10116No restrictions
NCBI GeneGene summary and NCBI IDs"superoxide dismutase 1", 21q22.110.24 s10/10livePublic (NCBI policies)
ClinVarClinical variant interpretations397 records0.24 s10/10livePublic (NCBI policies)
RCSB PDBExperimental 3D structures157 entries0.40 s10/10weeklyCC0
AlphaFold DBPredicted structureAF-P00441-F1, model v6, mean pLDDT 97.90.09 s10/10v6CC BY 4.0
InterProDomains and families4 InterPro entries (Cu/Zn SOD domain, binding site, superfamily…)1.0 s10/10liveCC0
QuickGOGene Ontology annotations243 annotations1.1 s10/10liveCC BY 4.0
STRINGFunctional partnersCCS 0.999, SOD2 0.998, PARK7 0.996, CAT 0.985…0.70 s10/1012.5CC BY 4.0
ReactomeCurated pathways3 lowest-level pathways (detoxification of ROS, platelet degranulation, IL-12/JAK-STAT)1.2 s10/1097Data CC0
KEGGPathway maps8 pathways0.28 s9/10updated dailyAcademic, non-commercial only
GEOExpression studies mentioning the gene526 records (70 Series, 456 Samples)0.25 s10/10livePublic (NCBI policies)
PRIDEProteomics datasets57 projects matched "SOD1"7.3 s10/10livePer project (many CC0)

"Answered" counts successful replies out of 10 calls (two rounds of five, a few minutes apart, one call per second), made from Seoul late in the evening KST. "Median response" is over the successful calls only (Ensembl's four failures came back fast, in 1–2.5 s). Response times include the round trip to Europe or the US, so your numbers will differ; the ranking and the failures are the useful part.

The Three Things the Measurement Taught Me

1. Public APIs fail. Build retries in from the start. In the second round, Ensembl answered only 1 of 5 calls; the other four returned HTTP 500. KEGG had one call that hung for 20 seconds and failed. A few minutes earlier both had answered every call. A pipeline that queries 5,000 proteins once each, with no retry, will silently lose some of them.

2. Write down the release you used. STRING's API now answers from version 12.5, while the R package STRINGdb was set to 12.0 for the STRING background test. Reactome was on release 97, Ensembl on 116, UniProt on 2026_03. Results change between releases, so a methods section that says "we used STRING" is not reproducible.

3. Check the licence before you build a product on it. Most of these are open (CC0, CC BY, or no restrictions). KEGG is not: its API is for academic, non-commercial use, and commercial use or redistribution needs a paid licence. If your tool might ever be commercial, Reactome (data CC0) or WikiPathways are the safer pathway sources.

How to Use Each One

The examples below use only the Python standard library. Each was run on 13 October 2026.

UniProt: start here

UniProt is the hub. One entry gives you the sequence, the curated function, and cross-references to almost every other database (SOD1 has 556, including 156 PDB entries; RCSB's own search found 157 the same day, because PDB updates weekly and this UniProt release is from 2 September). If you only remember one ID per protein, make it the UniProt accession.

https://rest.uniprot.org/uniprotkb/P00441.json

The response headers tell you the release (X-UniProt-Release: 2026_03). For many proteins at once, use the stream or search endpoints with fields= to choose columns, rather than downloading full entries one by one.

Ensembl: the gene, not the protein

Use Ensembl for genomic coordinates, transcripts, exons and orthologues across species. Look up by symbol or by Ensembl ID:

https://rest.ensembl.org/lookup/symbol/homo_sapiens/SOD1?content-type=application/json

It was the slowest and least reliable service in this test (median 8 s; four HTTP 500s in a row in the second round). Ensembl's REST service also limits you to 55,000 requests per hour, about 15 per second. For bulk work, the BioMart download or the release FTP files are far kinder than thousands of REST calls; see mapping orthologues with BioMart.

NCBI Gene, ClinVar and GEO: E-utilities

All three go through the same E-utilities endpoints (esearch, esummary, efetch), which were fast and steady here (0.24 s):

https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esummary.fcgi?db=gene&id=6647&retmode=json
https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=clinvar&term=SOD1[gene]&retmode=json

Without an API key NCBI allows 3 requests per second; a free key raises that to 10. A search like SOD1 AND "Homo sapiens"[orgn] in GEO returned 526 records, but only 70 were Series (studies); 456 were individual Samples. Add AND gse[ETYP] for a study-level list, and remember these are studies that mention the gene, not studies where it changes.

RCSB PDB and AlphaFold DB: structures

For experimental structures, the RCSB Search API takes a JSON query. Searching by UniProt accession avoids name ambiguity:

{"query": {"type": "terminal", "service": "text", "parameters": {
   "attribute": "rcsb_polymer_entity_container_identifiers.reference_sequence_identifiers.database_accession",
   "operator": "exact_match", "value": "P00441"}},
 "return_type": "entry", "request_options": {"return_all_hits": true}}

posted to https://search.rcsb.org/rcsbsearch/v2/query. SOD1 has 157 entries. For a predicted model, AlphaFold DB is one GET and was the fastest service in the test (0.09 s): https://alphafold.ebi.ac.uk/api/prediction/P00441 returns model v6 with a mean pLDDT of 97.9 and download links for PDB and mmCIF files. A high pLDDT says the model is confident, not that it matches the protein as it works: AlphaFold DB models are single chains without metal ions, while SOD1 is a Cu/Zn-binding homodimer with metal-bound and metal-free forms.

InterPro and QuickGO: what the protein is and does

InterPro gives domain and family assignments (/interpro/api/entry/interpro/protein/uniprot/P00441); QuickGO gives Gene Ontology annotations with evidence codes (/QuickGO/services/annotation/search?geneProductId=P00441). SOD1 has 243 GO annotations; if you use them for enrichment, check which evidence codes you include.

STRING: partners, with a caveat

https://string-db.org/api/json/interaction_partners?identifiers=SOD1&species=9606&limit=10

The top partners were CCS (its copper chaperone, 0.999), SOD2, PARK7, CAT and BCL2. STRING scores are functional association, not proof of physical binding, and the API version (12.5) may not match the R package you analyse with. For network and enrichment work see the STRING tutorial, and for the most common enrichment mistake, the background-set test.

Reactome and KEGG: pathways

Reactome maps a UniProt accession straight to its pathways (/ContentService/data/mapping/UniProt/P00441/pathways?species=9606): three lowest-level pathways for SOD1. KEGG's REST API is simple and fast (https://rest.kegg.jp/link/pathway/hsa:6647, 8 pathways) but, again, academic non-commercial use only. The two databases draw pathway boundaries differently, so "number of pathways" is not comparable between them.

PRIDE: proteomics datasets

PRIDE's search (/pride/ws/archive/v3/search/projects?keyword=SOD1) found 57 projects whose metadata mention SOD1. It was slow (median 7.3 s) and keyword matches include any project that names the protein, so read the project descriptions before you download raw files. Every dataset in my MaxQuant, FragPipe and DIA-NN benchmarks came from PRIDE.

One Protein Through Seven Databases, With Retries

This is the whole walkthrough as one script. The get() helper retries only what can recover (HTTP 429, 5xx, timeouts and dropped connections), honours Retry-After, and fails immediately on 400 or 404, so a mistyped accession doesn't cost you seven seconds of retries.

import json, time, http.client, urllib.request, urllib.error

def get(url, body=None, tries=4):
    """GET (or POST JSON) with retries. Retry only what can recover: 429, 5xx and network errors."""
    for i in range(tries):
        headers = {"User-Agent": "db-walkthrough/1.0"}
        if body:
            headers["Content-Type"] = "application/json"
        req = urllib.request.Request(url, data=json.dumps(body).encode() if body else None, headers=headers)
        try:
            with urllib.request.urlopen(req, timeout=30) as r:
                return r.read().decode()
        except urllib.error.HTTPError as err:
            if err.code != 429 and err.code < 500:
                raise                              # 400/404: a bad ID or query, retrying won't help
            wait = float(err.headers.get("Retry-After") or 2 ** i)
        except (urllib.error.URLError, OSError, http.client.HTTPException):
            wait = 2 ** i                          # timeouts, resets, dropped connections
        if i == tries - 1:
            raise RuntimeError(f"gave up after {tries} tries: {url}")
        time.sleep(wait)                           # 1, 2, 4 s unless the server says otherwise

acc = "P00441"                                      # human SOD1

u = json.loads(get(f"https://rest.uniprot.org/uniprotkb/{acc}.json"))
print("UniProt   ", u["uniProtkbId"], u["proteinDescription"]["recommendedName"]["fullName"]["value"],
      "|", u["sequence"]["length"], "aa")

e = json.loads(get("https://rest.ensembl.org/lookup/symbol/homo_sapiens/SOD1?content-type=application/json"))
print("Ensembl   ", e["id"], f"chr{e['seq_region_name']}:{e['start']}-{e['end']}", e["assembly_name"])

pdb = json.loads(get("https://search.rcsb.org/rcsbsearch/v2/query", {
    "query": {"type": "terminal", "service": "text", "parameters": {
        "attribute": "rcsb_polymer_entity_container_identifiers.reference_sequence_identifiers.database_accession",
        "operator": "exact_match", "value": acc}},
    "return_type": "entry", "request_options": {"return_all_hits": True}}))
print("RCSB PDB  ", pdb["total_count"], "entries, e.g.", ", ".join(h["identifier"] for h in pdb["result_set"][:3]))

af = json.loads(get(f"https://alphafold.ebi.ac.uk/api/prediction/{acc}"))[0]
print("AlphaFold ", af["entryId"], "v%s" % af["latestVersion"], "| mean pLDDT", af["globalMetricValue"])

s = json.loads(get("https://string-db.org/api/json/interaction_partners?identifiers=SOD1&species=9606&limit=5"))
print("STRING    ", ", ".join(f"{x['preferredName_B']} {x['score']}" for x in s))

r = json.loads(get(f"https://reactome.org/ContentService/data/mapping/UniProt/{acc}/pathways?species=9606"))
print("Reactome  ", "; ".join(p["displayName"].strip() for p in r))

c = json.loads(get("https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=clinvar&term=SOD1[gene]&retmode=json"))
print("ClinVar   ", c["esearchresult"]["count"], "records for SOD1")

Output on 13 October 2026:

UniProt    SODC_HUMAN Superoxide dismutase [Cu-Zn] | 154 aa
Ensembl    ENSG00000142168 chr21:31659617-31668931 GRCh38
RCSB PDB   157 entries, e.g. 1AZV, 1BA9, 1DSW
AlphaFold  AF-P00441-F1 v6 | mean pLDDT 97.94
STRING     CCS 0.999, SOD2 0.998, PARK7 0.996, CAT 0.985, BCL2 0.98
Reactome   Platelet degranulation; Detoxification of Reactive Oxygen Species; Gene and protein expression by JAK-STAT signaling after Interleukin-12 stimulation
ClinVar    397 records for SOD1

Counts like ClinVar records and PDB entries grow over time, so expect larger numbers if you run it later.

Practical Rules

  • Map everything to UniProt accessions first, then fan out. Gene symbols are ambiguous across species and over time.
  • Retry with backoff and cache what you fetched. Never query the same protein twice in one analysis.
  • Respect the limits (NCBI 3/s without a key, Ensembl about 15/s, KEGG 3/s) and set a descriptive User-Agent.
  • For more than a few hundred proteins, download instead of calling. UniProt, Reactome, STRING, InterPro and GO all publish bulk files.
  • Record releases and dates in your methods: "STRING 12.5 API, accessed 13 October 2026".
  • Read the licence before anything commercial. KEGG is the one most often used by mistake.

Limits

  • One protein, two rounds of five calls each, from one location on one evening. Latency and failures vary by hour and region; this is a snapshot, not a service-level measurement.
  • Counts are what each API returned for the queries shown; different query fields (for example, gene name instead of accession in PDB) return different numbers.
  • Licence summaries are from each provider's licence page on 13 October 2026; check them again before relying on them.

Code and outputs: the scripts and result tables for this post are on GitHub in sbmlab-benchmarks/bioinfo-db-check.

Related