Literature Deep Research

Systematic literature research: disambiguate, search with collision-aware queries, grade evidence, produce structured reports.

KEY PRINCIPLES: (1) Disambiguate first (2) Right-size deliverable (3) Grade every claim T1-T4 (4) All sections mandatory even if "limited evidence" (5) Source attribution for every claim (6) English-first queries, respond in user's language (7) Report = deliverable, not search log

LOOK UP, DON'T GUESS

Search PubMed/EuropePMC FIRST before reasoning. A published paper beats memory.

Factoid search strategy:

1. Extract KEY TERMS (most specific nouns/verbs)
2. EuropePMC_search_articles(query="term1 term2 term3", limit=5)
3. No results -> BROADEN (remove most restrictive term)
4. Too many -> NARROW (add specific terms)
5. Answer usually in abstract of top results
6. Failed query -> try DIFFERENT TERMS/synonyms, don't repeat

COMPUTE, DON'T DESCRIBE

When analysis requires computation (statistics, data processing, scoring, enrichment), write and run Python code via Bash. Don't describe what you would do — execute it and report actual results. Use ToolUniverse tools to retrieve data, then Python (pandas, scipy, statsmodels, matplotlib) to analyze it.

Workflow

Phase 0: Clarify + Mode Select → Phase 1: Disambiguate + Profile → Phase 2: Literature Search → Phase 3: Report

Phase 0: Mode Selection

Factoid Mode (Fast Path)

# [TOPIC]: Fact-check Report
## Question / ## Answer (with evidence rating) / ## Source(s) / ## Verification Notes / ## Limitations

Domain Detection

Cross-Skill Delegation

• Gene/protein deep-dive: tooluniverse-target-research
• Drug profile: tooluniverse-drug-research
• Disease profile: tooluniverse-disease-research

Use this skill for literature synthesis. Use specialized skills for entity profiling. For max depth, run both.

Phase 1: Subject Disambiguation + Profile

1.1 Biological Target Resolution

UniProt_search → UniProt_get_entry_by_accession → UniProt_id_mapping
ensembl_lookup_gene → MyGene_get_gene_annotation

1.2 Naming Collision Detection

Check first 20 results. If >20% off-topic, build negative filter: NOT [collision1] NOT [collision2]. Gene family: "ADAR" NOT "ADAR2" NOT "ADARB1". Cross-domain: add context terms.

1.3 Baseline Profile (Bio Targets)

InterPro_get_protein_domains, UniProt_get_ptm_processing_by_accession, HPA_get_subcellular_location,
GTEx_get_median_gene_expression, GO_get_annotations_for_gene, Reactome_map_uniprot_to_pathways,
STRING_get_protein_interactions, intact_get_interactions, OpenTargets_get_target_tractability_by_ensemblID

GPCR targets: delegate to tooluniverse-target-research.

1.5 Drug Disambiguation

Identity: OpenTargets_get_drug_chembId_by_generic_name, ChEMBL_get_drug, PubChem_get_CID_by_compound_name, drugbank_get_drug_basic_info_by_drug_name_or_id Targets: ChEMBL_get_drug_mechanisms, OpenTargets_get_associated_targets_by_drug_chemblId, DGIdb_get_drug_gene_interactions Safety: OpenTargets_get_drug_adverse_events_by_chemblId, OpenTargets_get_drug_indications_by_chemblId, search_clinical_trials

1.6 Disease Disambiguation

OpenTargets disease search → EFO/MONDO IDs
DisGeNET_get_disease_genes, DisGeNET_search_disease
CTD_get_disease_chemicals

1.7 Compound Queries (e.g., "metformin in breast cancer")

Resolve both entities, then cross-reference via CTD_get_chemical_gene_interactions, CTD_get_chemical_diseases, OpenTargets drug-target/drug-disease tools. Intersect shared targets/pathways.

1.8 General Academic / 1.9 Interdisciplinary

Non-bio: skip bio tools, use ArXiv/DBLP/OSF. Cross-domain: resolve bio entities with 1.1-1.3, search CS/general in parallel, merge and cross-reference.

Phase 2: Literature Search

Methodology stays internal. Report shows findings, not process.

2.1 Query Strategy

Step 1: Seeds (15-30 core papers): domain-specific title searches with date/sort filters. Step 2: Citation expansion: PubMed_get_cited_by, EuropePMC_get_citations/references, PubMed_get_related, SemanticScholar_get_recommendations, OpenCitations_get_citations. If the opt-in Noodle MCP is connected (noodle_*, needs NOODLE_MCP_URL), its bounded citation/semantic graph traversal is another angle on the same PubMed corpus -- a discovery signal, not evidence of causality or validity, same caveat as the others. Step 3: Collision-filtered broader queries: "[TERM]" AND ([context]) NOT [collision]

2.2 Literature Tools — core set + adaptive by domain

Run the core multi-field set on every review (catches what any single index misses), then add the domain rows that match the subject. Don't fire every source blindly — 6–10 well-chosen indexes beat 20 noisy ones.

ALWAYS run (core, all disciplines): PubMed_search_articles, EuropePMC_search_articles, openalex_search_works (query param search/query) or openalex_literature_search (query param search_keywords) — pick one and match its param; mixing them silently returns off-topic results — and SemanticScholar_search_papers

Then add by domain:

Multi-source: advanced_literature_search_agent (12+ DBs; needs Azure key -- fallback: query the core set individually). Citation impact: iCite_search_publications (RCR/APT), iCite_get_publications (by PMID), scite_get_tallies (support/contradict). PubMed-only; for CS use SemanticScholar.

Identifiers and reference lists: PubMed_convert_article_ids {"ids": ["23193287", "10.1093/nar/gks1195"]} converts between PMID, PMCID and DOI (auto-detected; only articles in PubMed Central convert, others return found: false) -- use it to turn a DOI into the PMCID that EuropePMC_get_full_text needs. PubMed_lookup_article_by_citation {"journal": "Nature", "year": 2015, "volume": 521, "first_page": 436, "author": "lecun y"} resolves a reference-list entry to a PMID (status: found / ambiguous / not_found; pass citations for a batch). The match is exact on the fields given, so an unfound citation is a signal to re-check the fields, not proof the paper is missing.

A domain-specific index returning 0 (e.g. ArXiv on a pure-clinical topic) is normal — only worry if the whole core set is empty.

2.3-2.4 Full-Text & PubMed Zero-Result Fallback

Full-text: see FULLTEXT_STRATEGY.md for three-tier strategy.

CRITICAL: PubMed returns 0 for ~30% of valid queries. Always retry with EuropePMC when PubMed returns empty. This is not optional.

2.5 Tool Failure / OA Handling

Retry once -> fallback tool. Key fallbacks: PubMed_get_cited_by -> EuropePMC_get_citations -> OpenCitations. OA: Unpaywall if configured, else Europe PMC/PMC/OpenAlex flags.

Last resort when every structured index above is empty (a brand-new preprint, a dataset page, a project site with no DOI): the opt-in exa_* tools (general neural web search, no key needed for casual use) can still find it, but it's general internet retrieval, not a scientific database -- verify anything it surfaces against a real source before citing, don't grade it T1-T4 as if it were literature.

2.6 Controlled Vocabulary, Text-Mined Annotations, Citation Cross-References, and Variant Literature

Four small tool families beyond core search/citation coverage — verified live, not schema-assumed:

MeSH (controlled vocabulary): MeSH_search_descriptors/MeSH_search_terms resolve a free-text term to NLM's standardized MeSH descriptor/entry-term IDs; MeSH_get_descriptor returns the descriptor's official label, type, and annotation. Use to broaden or standardize a query before searching (e.g. a user says "sugar disease" -> MeSH_search_terms finds the descriptor is actually filed under "Diabetes Mellitus") or to confirm two different-sounding search hits are actually indexed under the same concept.

EuroPMCAnnot (text-mined entity annotations): EuroPMCAnnot_get_article_annotations (all entity mentions in one article: genes, diseases, chemicals, organisms, etc.), EuroPMCAnnot_get_chemicals_from_article (chemical/compound mentions only, a filtered convenience view), EuroPMCAnnot_get_annotations_by_type (one annotation type across multiple articles at once). These extract what a paper mentions without you reading the full text — useful for a fast relevance check across many candidate papers, or for confirming a specific gene/chemical is actually discussed (not just present in an abstract keyword match). Live-verified example: EuroPMCAnnot_get_chemicals_from_article on PMC4353746 returned 19 real chemical mentions including "Resistin".

Citation cross-references — naming collision, read carefully: src/tooluniverse/data/europepmc_citations_tools.json defines EPMC_get_citations, EPMC_get_references, and EuropePMC_get_article_datalinks — do NOT confuse these with the already-documented EuropePMC_get_citations/EuropePMC_get_references (core europe_pmc_tools.json, used above in Phase 2.1/TOOL_NAMES_REFERENCE.md). They are separate implementations with near-identical names hitting the same underlying Europe PMC REST endpoints. Live-verified as of this writing: the /references endpoint is down on BOTH implementations (EPMC_get_references and EuropePMC_get_references both return a real 503 "This API is temporarily unavailable due to maintenance" from www.ebi.ac.uk), and EuropePMC_get_article_datalinks independently confirmed broken via direct curl (33s response, HTTP 500). EPMC_get_citations/EuropePMC_get_citations (the forward-citation direction) both work normally. Practical guidance: prefer the already-documented EuropePMC_get_citations/EuropePMC_get_references names for citation work; if references-fetching fails with a 503, it is very likely this live outage, not a query problem — say so rather than reporting "no references found." EuropePMC_get_article_datalinks (when working) additionally surfaces what data/database records a paper deposited (GenBank accessions, PDB entries, clinical trial registrations) — complementary to, not a replacement for, citations/references.

Document conversion (any format to markdown): convert_to_markdown {"uri": "..."} accepts an http:/https:/file:/data: URI and converts it to markdown — verified live on an HTML page. Reach for this when a source is a PDF, DOCX, PPTX, XLSX, or HTML page rather than something the literature tools above already return as structured JSON (e.g. a supplementary-materials file, an institutional report, a non-indexed working paper someone hands you a link to) and you need its actual text content, not just metadata. This is a generic conversion utility, not a literature-search tool — use PubMed/EuropePMC/OpenAlex above whenever the source is already indexed there.

BGPT (structured critical-appraisal search): BGPT_search_paper_evidence is a literature search tool, not a generative model despite the name — live-verified, it returns real papers (doi, title, plus full-text-derived fields). Unlike PubMed/EuropePMC/OpenAlex (title+abstract only), each result also includes methods/techniques used, sample size and population, results, paper limitations and biases, conflicts of interest, data/code availability, and a how_to_falsify statement — reach for it when a claim needs quality-weighing (is this an RCT with n=12 or n=12,000? does the paper disclose a conflict of interest?) rather than just a citation. Caveat: these structured fields are model-generated summaries of the paper, not curated ground truth — treat them as an appraisal aid and still verify against the actual paper before citing a specific number from them. First 50 results/session are free; BGPT_API_KEY unlocks the paid tier after that.

LitVar (literature-derived variant mentions): LitVar_search_variants (find variants by rsID/gene/HGVS as discussed in the literature), LitVar_get_variant_publications (PMIDs mentioning a specific variant), LitVar_get_variant_details (structured record: gene, HGVS, ClinGen IDs). This answers "what does the literature say has been written about variant X" — distinct from clinical variant-classification databases (ClinVar, gnomAD, already in Phase 1.1/TOOL_NAMES_REFERENCE.md), which answer "what is variant X's clinical significance." Use LitVar to find the papers, then a clinical database to grade the variant itself. Live-verified: LitVar_search_variants(gene="BRCA1") and rsID-based lookups (rs328, rs7903146) all returned real hits; LitVar_get_variant_publications returned real PMID lists per variant.

Phase 3: Evidence Grading

Inline: Target X regulates Y [T1: PMID:12345678]. Per theme: summarize evidence distribution.

Triaging a large candidate set before reading in full: Consensus_search_papers returns study type and sample size per paper, a fast first pass for provisional tiering -- confirm against the actual paper before citing, its metadata is a starting point, not the grade itself.

Report Output

Progressive update: create report with all section headers immediately. Fill after each phase. Write Executive Summary LAST.

Use 15-section template from REPORT_TEMPLATE.md. Domain adaptations: bio (architecture/expression/GO/disease), drug (properties/MOA/PK/safety), disease (epi/patho/genes/treatments), general (history/theories/evidence/applications).

Communication

Brief progress updates only: "Resolving identifiers...", "Building paper set...", "Grading evidence..." Do NOT expose: raw tool outputs, dedup counts, search round details.

References

• TOOL_NAMES_REFERENCE.md -- 130+ tools with parameters
• REPORT_TEMPLATE.md -- template, domain adaptations, bibliography, completeness checklist
• FULLTEXT_STRATEGY.md -- three-tier full-text verification
• WORKFLOW.md -- compact cheat-sheet
• EXAMPLES.md -- worked examples