Which public bio database should I search? Using AI to route your question

A quick primer before the walkthrough: Most life-science questions have one obvious public database behind them. Knowing which one saves hours; AI can help you choose and phrase the search, but the database itself is always the source of truth.

In plain English: You have a question. Instead of typing it into a general web search, you decide which public repository actually holds the answer, then search there. This tutorial shows you how to use an AI assistant to make that choice and to draft the search terms.

Who this is for: Anyone new to biology databases — scientists changing fields, and non-bench colleagues in ops, BD, or regulatory who need to look something up themselves.

What you need: A browser and any general AI assistant (ChatGPT, Claude, Copilot, or Gemini). No account beyond the free tier is required, and no coding.

Comfort level: No coding. Everything here uses public, non-confidential information only.

Key terms (defined again at the end): repository, accession, curated, abstract.

Step 1 — Know the six sources you will use most

What this step does: Gives you a mental map so you are not guessing.

Curated means a human editor has reviewed and organized the entry, which is why UniProt and NCBI Gene are more reliable for facts than a general web page.

Step 2 — Ask the assistant to route the question

What this step does: Turns "where do I even look?" into a named database plus a search string you can paste.

Role: You are a biomedical information specialist.
Task: For the question below, tell me (1) which public database is the best
primary source, (2) why, (3) an exact search string to paste into it, and
(4) one alternative source if the first returns nothing.
Rules:
- Choose only from: PubMed, UniProt, NCBI Gene, GEO, PubChem, ChEMBL, ClinicalTrials.gov.
- Do not answer the biological question itself.
- Do not invent accession numbers, identifiers, or citations.

Question: [YOUR QUESTION]

Telling it not to answer the question matters: you want a route, not a confident guess.

Worked example

Question: "Which human protein does the gene EGFR encode, and what are its main domains?"

A good response looks like this:

• Primary source: UniProt — it is the curated home for protein sequence and domain annotation.
• Why: Domain boundaries are annotated per protein entry with supporting references.
• Search string: gene:EGFR AND organism_id:9606 AND reviewed:true
• Alternative: NCBI Gene, if you need the genomic context rather than the protein.

Now do the check: open UniProt, run that search, and confirm the entry exists. The assistant suggested the route; the database supplied the answer.

Step 3 — Verify before you use anything

What this step does: Catches the one failure mode that matters — invented identifiers.

Three quick checks:

1. Does the identifier resolve? Paste any accession (a database's unique record ID, e.g. a UniProt accession) into the site's search box. If nothing comes up, it was fabricated.
2. Does the record actually say it? Read the entry, not the assistant's paraphrase.
3. How old is it? Note the last-updated date on the record; literature and annotations move.

Never cite a paper, accession, or trial number that you have not opened yourself.

Step 4 — Narrow a search that returns too much

What this step does: Turns 4,000 hits into a readable list.

Ask the assistant to add filters in the target database's own syntax:

Rewrite this PubMed search to return fewer, more relevant results.
Add filters for: humans, last 5 years, review articles excluded.
Return only the search string.

Then sanity-check the result count in the database. If it drops to zero, one filter is too strict — remove them one at a time.

If something goes wrong

• The assistant answers the biology question instead of routing it. Repeat the "do not answer the question itself" rule and paste the question again.
• A suggested accession or PMID does not exist. Assume any unverified identifier is invented. Search by name or keyword in the database instead.
• The search syntax is rejected. Field syntax differs by database and changes over time; paste the error back and ask for a plain keyword search, then apply filters through the site's own interface.
• You get zero results. Broaden first (drop the organism or date filter), then try the alternative source the assistant named.
• You are unsure whether you may paste your question. If it reveals an unpublished target, a program name, or a partner, do not — see Safer AI use at work: what to share and what to keep private.

What to try next

• Prompting ChatGPT for high-quality scientific paper summaries
• Deep literature reviews with ChatGPT for life-science teams

Key terms

• Repository — a public database that stores records of one kind (papers, proteins, datasets).
• Accession — the unique ID a repository assigns to a record.
• Curated — reviewed and annotated by human editors rather than deposited automatically.
• Abstract — the short summary at the top of a paper; often all you can read without a subscription.
• Metadata — the descriptive information about a dataset (samples, conditions, platform).
• PMID — PubMed's numeric identifier for a citation.

AI-suggested routes and search strings are drafting aids. The database record, read by you, is the citable source.