Integrate the SRAgent to retrieve datasets from Sequence Read Archive

[caufieldjh]

See https://github.com/ArcInstitute/SRAgent

SRAgent also includes agent tooling for Entrez tools. This may be a way to more reproducibly run queries against NCBI resources when populating Dismech pages with datasets.

[github-actions]

Summary: SRAgent Integration for SRA Dataset Retrieval

What is SRAgent?

SRAgent is an LLM-powered agentic toolkit from the Arc Institute designed to systematically extract, process, and annotate metadata from the Sequence Read Archive (SRA) and associated NCBI databases. It was developed to support the scBaseCount project — an AI-curated, uniformly processed single-cell data repository.

SRAgent is Claude-native (uses Anthropic API by default) and is structured as a collection of specialized agents.

---

Relevant Agents for Dismech

Agent	What it does	Dismech relevance
Entrez Agent	Calls NCBI Entrez services (esearch, efetch, esummary, elink); handles accession format conversions	Reproducible NCBI queries; replaces ad hoc scripts
Find-datasets Agent	Searches SRA by keyword query, then runs full metadata extraction	Discovering disease-relevant SRA datasets at scale
SRX-info / Metadata Agent	Extracts sequencing platform, library type, organism, tissue (Uberon IDs), disease (MONDO/PATO IDs) per SRX accession	Directly maps to Dismech ontology terms — provides MONDO/Uberon/PATO annotations automatically
Papers Agent	Locates publications linked to SRA accessions; downloads manuscripts from Europe PMC and preprint servers	Auto-linking datasets to PMID evidence items
Disease-Ontology Agent	Maps disease descriptions to MONDO terms	Consistent MONDO term binding
Tissue-Ontology Agent	Maps tissue descriptions to Uberon terms	Consistent UBERON term binding

---

Why This Matters for Dismech

Currently, populating disorder entries with datasets (SRA accessions, cell atlas entries, etc.) requires manual querying of NCBI resources. SRAgent would provide:

1. Reproducibility — structured, auditable agent-driven queries rather than manual web searches
2. Scale — batch discovery across all ~55+ disorder entries
3. Ontology alignment — SRX-info already outputs MONDO and Uberon IDs, which map directly to Dismech's ontology term system
4. Publication linking — the Papers Agent can automatically surface PMIDs associated with SRA datasets, feeding evidence items

---

Requirements

• ANTHROPIC_API_KEY (already in use for dismech curation agents)
• EMAIL for Entrez API (rate limit courtesy)
• Optional: NCBI_API_KEY for increased rate limits
• Python 3.12, sra-tools 3.1 (via conda or UV)

---

Related Issues

• Find CF related datasets #302 — Find CF-related datasets: A concrete near-term use case — SRAgent's find-datasets agent could automate this search for cystic fibrosis (and other diseases) with disease-matched MONDO terms and tissue Uberon IDs pre-populated.

---

Suggested Integration Path

1. Pilot: Use SRAgent's find-datasets CLI against a single disorder (e.g., Cystic Fibrosis per Find CF related datasets #302) to validate output quality and MONDO/Uberon alignment with Dismech schema
2. Schema addition: Formalize a datasets block in the Dismech schema (if not already present) to capture SRA accessions, tissue, technology, organism, linked PMIDs
3. Script wrapper: Write a scripts/fetch_sra_datasets.py wrapper around SRAgent calls that outputs Dismech-compatible YAML blocks
4. Batch curation: Run across all disorder entries, filtering for human single-cell or bulk RNA-seq studies

---

Notes

• SRAgent optionally tracks progress via a PostgreSQL database (GCP) — useful if scaling to batch curation across all disorders
• The tissue/disease ontology agents could serve as independent validators of Dismech's existing term annotations