Biomedical Dataset Finder

Search public biomedical datasets from NCBI, NGDC, and CNGB by conversational query keywords.

Usage Trigger

User asks for datasets related to a disease/treatment/species/subtype/data type combination. Examples:

• "Find colon cancer dMMR immunotherapy single-cell data"
• "hepatocellular carcinoma PD-1 scRNA-seq baseline"
• "lung cancer immunotherapy single cell data"

Data Sources (Priority Order)

Workflow

Step 1 — Parse Query

Extract from user message:

• Disease/Cancer: e.g. colon cancer, hepatocellular carcinoma, lung cancer
• Treatment: e.g. immunotherapy, PD-1, chemotherapy, baseline therapy
• Species: human, mouse (defaults to human if unspecified)
• Pathology Subtype: e.g. dMMR, MSI-H, KRAS mutant
• Data Type: e.g. scRNA-seq, single-cell, RNA-seq, ChIP-seq, ATAC-seq

If any critical field is missing, ask the user to clarify.

Step 2 — NCBI Search (Primary)

Use NCBI E-utilities (free, no auth).

1. Search gds database (GEO Datasets, NOT gse) with combined keywords
2. For each result, pull accession (GSE prefix), title, summary, and pubmedids (list)
3. Fetch article info (authors, title, journal, year, DOI) for each PMID
4. For single-cell queries, also search sra database

Query: ({disease}) AND ({treatment}) AND ({species}) AND ({data_type})

Rate limit: ~3 requests/second.

Step 3 — NGDC Search (Primary)

API: https://ngdc.cncb.ac.cn/search/api/specific?q={keywords}&db=gsa&size=20

Requires User-Agent header. Filter response for type=="GSA" entries (CRA accessions).

Step 4 — CNGB Search (Secondary)

If CNGB token provided: search CNGBdb API. On auth error: ask user if they want to provide token or skip.

Step 5 — Output

Markdown table with bold dataset ID, article info (authors, title, journal, year, DOI), and direct links.

If no results: "No public datasets found matching your criteria. Try adjusting keywords or switching data sources."

Factuality Requirements (Critical — No Hallucinations)

This skill handles scientific research data. Fabricating a single dataset entry undermines the user's work.

Hard Rules

1. Dataset IDs: Only use IDs returned by actual API responses. Never invent, guess, or infer IDs.
2. Article info: Only populate from actual API/PubMed responses. Leave blank if no data returned.
3. Links: Build from verified accession patterns (e.g. https://.../acc.cgi?acc={GSE}). Never guess URLs.
4. "Not found" is valid: If a source returns 0 results, output the empty result — do not fabricate entries to fill the table.

Verification Checklist (before presenting results)

• Every Dataset ID is from an API response, not memory or guess
• Every Article Title + Authors + Journal is from a PubMed/API response, not reconstructed
• Every Link follows the confirmed URL pattern for that database
• If a field is empty in the API response, it must be blank - in the table — never fill with plausible text

Why This Matters

A researcher using wrong dataset IDs or fake article info could: waste weeks on non-existent data, cite non-existent papers, or compromise the validity of their research. The cost of hallucination here is far higher than in general conversation.

Security Notes

• User keywords are private — do NOT log the raw search query string to stderr/stdout. Log only counts (e.g. "Searching 5 keywords...").
• Token handling — CNGB token is passed via CLI arg only; never hardcode or log it.
• No external exfiltration — results table contains only public dataset metadata; no user-provided content is stored or transmitted elsewhere.

CLI Tool

python3 skills/biomed-dataset-finder/scripts/search_datasets.py \
  --disease "colon cancer" --treatment "immunotherapy" \
  --species human --subtype dMMR --type scRNA-seq --max-results 10

API Reference

See references/ncbi_api.md for NCBI E-utilities details. See references/ngdc_api.md for NGDC GSA API details. See references/cngb_api.md for CNGBdb API details.