Literature Deep Research

Systematic approach to comprehensive literature research: disambiguate the subject, search with collision-aware queries, grade evidence, and produce a structured report.

KEY PRINCIPLES:

Disambiguate first - Resolve IDs, synonyms, naming collisions before literature search
Right-size the deliverable - Factoid mode for single questions; full report for deep research
Evidence grading - Grade every claim (T1 mechanistic → T4 mention)
Mandatory completeness - All sections must exist, even if "unknown/limited evidence"
Source attribution - Every claim traceable to database/tool
English-first queries - Use English for searches; respond in user's language
Report = deliverable - Show findings, not search process

Workflow Overview

User Query
  ↓
Phase 0: CLARIFY + MODE SELECT (factoid vs deep report)
  ↓
Phase 1: SUBJECT DISAMBIGUATION + PROFILE
  ├─ Detect domain (biological target / drug / disease / general academic)
  ├─ Resolve identifiers and gather synonyms/aliases
  ├─ Check for naming collisions
  └─ Gather baseline context via annotation tools (domain-specific)
  ↓
Phase 2: LITERATURE SEARCH (methodology kept internal)
  ├─ High-precision seed queries
  ├─ Citation network expansion from seeds
  ├─ Collision-filtered broader queries
  └─ Theme clustering + evidence grading
  ↓
Phase 3: REPORT SYNTHESIS (report-first pattern)
  ├─ Create [topic]_report.md with all section headers IMMEDIATELY
  ├─ Progressively fill sections as data arrives (update after each phase)
  ├─ Write Executive Summary LAST (after all sections complete)
  ├─ Generate [topic]_bibliography.json + .csv
  └─ Validate completeness checklist

Phase 0: Initial Clarification

Ask only what is needed; skip questions with obvious answers:

Subject type: Gene/protein, disease, drug, CS/ML topic, social science, or general?
Scope: Single factoid to verify, or comprehensive deep review?
Known aliases (if ambiguous): Specific names or symbols in use?
Constraints: Open access only? Include preprints? Specific organisms or date range?

Mode Selection

Mode	When to Use	Deliverable
Factoid / Verification	Single concrete question	`[topic]_factcheck_report.md` (≤1 page) + bibliography
Mini-review	Narrow topic	Short narrative report (1-3 pages)
Full Deep-Research	Comprehensive overview	Full 15-section report + bibliography

Heuristic: "Which antibiotic was X evolved to resist?" → Factoid. "What does the literature say about X?" → Full.

Factoid / Verification Mode (Fast Path)

Provide a correct, source-verified answer with explicit evidence attribution.

# [TOPIC]: Fact-check Report
*Generated: [Date]*

## Question
[User question]

## Answer
**[One-sentence answer]** [Evidence: ★★★/★★☆/★☆☆/☆☆☆]

## Source(s)
- [Primary citation: journal/year/PMID/DOI]

## Verification Notes
- [1-3 bullets: where the statement appears, key constraints]

## Limitations
- [Full text availability, evidence type caveats]

Prefer ToolUniverse literature tools over web browsing. Use EuropePMC_search_articles(extract_terms_from_fulltext=[...]) for OA snippet verification when possible.

Detect Subject Domain

Query Pattern	Domain	Phase 1 Action
Gene symbol (EGFR, TP53)	Biological target	Full bio disambiguation
Protein name ("V-ATPase")	Biological target	Full bio disambiguation
Drug name ("metformin")	Drug	Drug disambiguation (see 1.5)
Disease ("Alzheimer's")	Disease	Disease disambiguation (see 1.6)
CS/ML topic ("transformer architecture")	General academic	Literature-only (skip bio tools)
Method, concept, general topic	General academic	Literature-only (skip bio tools)
Cross-domain ("GNNs for drug discovery")	Interdisciplinary	Resolve each entity in its domain (see 1.9)

Cross-Skill Delegation

For deep entity-specific research beyond literature, delegate to specialized skills:

Gene/protein deep-dive (9-path profiling, druggability, GPCR data): use tooluniverse-target-research
Drug comprehensive profile (ADMET, FDA labels, formulations): use tooluniverse-drug-research
Disease comprehensive profile (ontologies, epidemiology, treatments): use tooluniverse-disease-research

Use this skill when the focus is literature synthesis and evidence grading. Use specialized skills when the focus is entity profiling with structured database queries. For maximum depth, run both in parallel.

Phase 1: Subject Disambiguation + Profile

1.1 Resolve Official Identifiers (Biological Targets)

UniProt_search → UniProt accession
UniProt_get_entry_by_accession → Full entry with cross-references
UniProt_id_mapping → Map between ID types
ensembl_lookup_gene → Ensembl gene ID, biotype
MyGene_get_gene_annotation → NCBI Gene ID, aliases, summary

1.2 Naming Collision Detection

Check the primary database for the domain (first 20 results). If >20% off-topic, build a negative filter:

Domain	Collision Check Syntax
Biomedical	PubMed: `"[TERM]"[Title]`
CS/ML	ArXiv: `ti:"[TERM]"` or SemanticScholar with `fieldsOfStudy` filter
General	OpenAlex or Crossref title search

Identify collision terms from off-topic results
Build negative filter: NOT [collision1] NOT [collision2]

Gene family disambiguation: Use official symbol with explicit exclusions. Example: "ADAR" NOT "ADAR2" NOT "ADARB1" for ADAR1-specific results.

Cross-domain collision: Some terms have different meanings across fields (e.g., "RAG" = Retrieval-Augmented Generation in CS, Recombination Activating Gene in biology). Add domain context terms to filter: "RAG" AND "language model" NOT "recombination activating".

1.3 Baseline Profile (Biological Targets)

Gather structural, functional, and expression context via annotation tools:

InterPro_get_protein_domains → Domain architecture
UniProt_get_ptm_processing_by_accession → PTMs, active sites
HPA_get_subcellular_location → Localization
GTEx_get_median_gene_expression → Tissue expression (use gtex_v8)
GO_get_annotations_for_gene → GO terms
Reactome_map_uniprot_to_pathways → Pathways
STRING_get_protein_interactions → Interaction partners
intact_get_interactions → Experimentally validated PPIs
OpenTargets_get_target_tractability_by_ensemblID → Druggability assessment

GPCR targets: If the target is a GPCR (~35% of approved drug targets), delegate to tooluniverse-target-research for specialized GPCRdb data (3D structures, ligands, mutations).

1.4 Baseline Profile Output

## Target Identity
| Identifier | Value | Source |
|------------|-------|--------|
| Official Symbol | [SYMBOL] | HGNC |
| UniProt | [ACC] | UniProt |
| Ensembl Gene | [ENSG...] | Ensembl |

**Synonyms**: [list]
**Collisions**: [assessment]

1.5 Drug-Centric Disambiguation

Skip protein architecture/expression/GO. Instead:

Resolve 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 & mechanisms: ChEMBL_get_drug_mechanisms, OpenTargets_get_associated_targets_by_drug_chemblId, DGIdb_get_drug_gene_interactions, drugbank_get_targets_by_drug_name_or_drugbank_id

Safety & indications: OpenTargets_get_drug_adverse_events_by_chemblId, OpenTargets_get_drug_indications_by_chemblId, search_clinical_trials

1.6 Disease-Centric Disambiguation

Resolve ontology IDs: Use OpenTargets_get_drug_chembId_by_generic_name or disease search tools to resolve EFO/MONDO IDs. Cross-reference ICD-10 and UMLS CUI when available from tool results.

OpenTargets_get_diseases_phenotypes_by_target_ensembl → Disease associations
DisGeNET_get_disease_genes → Disease-gene associations
DisGeNET_search_disease → Disease search with ontology IDs
CTD_get_disease_chemicals → Chemical-disease links

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

Resolve both entities separately, then cross-reference:

CTD_get_chemical_gene_interactions → Chemical-gene links
CTD_get_chemical_diseases → Chemical-disease associations
OpenTargets_get_associated_targets_by_drug_chemblId → Drug targets
OpenTargets_get_associated_diseases_by_drug_chemblId → Drug-disease associations
→ Intersect to find shared targets/pathways

1.8 General Academic Topics (No Bio Tools)

For CS, social science, humanities, or other non-bio topics:

Skip all bio annotation tools (UniProt, InterPro, GTEx, etc.)
Proceed directly to Phase 2 literature search
Use domain-appropriate databases (ArXiv for CS/ML, DBLP for CS, OSF for social science)
Collision detection still applies (search term ambiguity)

1.9 Interdisciplinary / Cross-Domain Queries

For topics spanning multiple domains (e.g., "GNNs for drug discovery", "AlphaFold protein prediction"):

Identify each domain component separately (e.g., CS method + biological application)
Resolve bio entities using Phase 1.1-1.3 (targets, drugs, diseases)
Search CS/general literature using ArXiv, DBLP, SemanticScholar in parallel
Merge results — use both bio tools AND general academic tools in Phase 2
Cross-reference — find papers that bridge both domains (typically computational biology venues)

Phase 2: Literature Search

Methodology stays internal. The report shows findings, not process.

2.1 Query Strategy

Step 1: High-Precision Seeds (15-30 core papers)

Domain-specific seed queries:

Biomedical: "[TERM]"[Title] AND (mechanism OR function OR structure OR review)
CS/ML:      ti:"[TERM]" AND (architecture OR benchmark OR evaluation OR survey)
General:    "[TERM]" in title via OpenAlex/Crossref

Use date/sort filters for recency or impact:

PubMed: mindate, maxdate, sort="pub_date"
SemanticScholar: year="2023-2024", sort="citationCount:desc"
ArXiv: date_from, sort_by="submittedDate"

Step 2: Citation Network Expansion

PubMed_get_cited_by → Forward citations (primary)
EuropePMC_get_citations → Forward (fallback)
PubMed_get_related → Related papers
EuropePMC_get_references → Backward citations
SemanticScholar_get_recommendations → AI-similar papers
OpenCitations_get_citations → DOI-based citation data

Step 3: Collision-Filtered Broader Queries

"[TERM]" AND ([context1] OR [context2]) NOT [collision_term]

2.2 Literature Search Tools

Biomedical: PubMed_search_articles, PMC_search_papers, EuropePMC_search_articles, PubTator3_LiteratureSearch

CS/ML: ArXiv_search_papers, DBLP_search_publications, SemanticScholar_search_papers

General academic: openalex_literature_search, Crossref_search_works, CORE_search_papers, DOAJ_search_articles

Preprints: BioRxiv_get_preprint, MedRxiv_get_preprint, OSF_search_preprints, BioRxiv_list_recent_preprints (For preprint keyword search: EuropePMC_search_articles(source='PPR'))

Multi-source deep search: advanced_literature_search_agent (searches 12+ databases in parallel; requires Azure OpenAI key — if unavailable, replicate coverage by querying PubMed + ArXiv + SemanticScholar + OpenAlex individually)

Citation impact: iCite_search_publications (search + RCR/APT metrics), iCite_get_publications (metrics by PMID), scite_get_tallies (supporting/contradicting counts) Note: iCite and scite are PubMed-only. For CS/ML papers, use SemanticScholar_get_paper for citation counts and influence scores.

Author search: PubMed "Author[Author]", ArXiv "au:Name", SemanticScholar/OpenAlex as query text

2.3 Full-Text Verification

When abstracts lack critical details, use full-text snippet extraction. See FULLTEXT_STRATEGY.md for the three-tier strategy (Europe PMC auto-snippets → manual Semantic Scholar/ArXiv → manual download).

2.4 Tool Failure Handling

Attempt 1 → fails → wait 2s → Attempt 2 → fails → wait 5s → Fallback tool

Primary	Fallback 1	Fallback 2
`PubMed_get_cited_by`	`EuropePMC_get_citations`	`OpenCitations_get_citations`
`PubMed_get_related`	`SemanticScholar_get_recommendations`	`SemanticScholar_search_papers`
`GTEx_get_median_gene_expression`	`HPA_get_rna_expression_by_source`	Document as unavailable
`Unpaywall_check_oa_status`	Europe PMC `isOpenAccess`	OpenAlex `is_oa`

2.5 Open Access Handling

With Unpaywall email: full OA check. Without: best-effort via Europe PMC, PMC, OpenAlex, DOAJ flags. Label: *OA Status: Best-effort (Unpaywall not configured)*

Phase 3: Evidence Grading

Grade every claim by evidence strength:

Tier	Label	Description	Bio Example	CS/ML Example
T1	★★★ Mechanistic	Direct experimental/formal evidence	CRISPR KO + rescue, RCT	Formal proof, controlled ablation with significance test
T2	★★☆ Functional	Functional study showing role	siRNA knockdown phenotype	Benchmark on standard dataset with baselines
T3	★☆☆ Association	Screen hit, correlation, observational	High-throughput screen, GWAS	Observational study, case study, anecdotal comparison
T4	☆☆☆ Mention	Review, text-mined, peripheral	Review article	Survey paper, blog post, workshop abstract

In report, label inline:

Target X regulates pathway Y [★★★: PMID:12345678] through direct
phosphorylation [★★☆: PMID:23456789].

Per theme, summarize evidence quality:

### Theme: Lysosomal Function (47 papers)
**Evidence Quality**: Strong (32 mechanistic, 11 functional, 4 association)

Report Output

Deliverables

File	Mode	Always?
`[topic]_report.md`	Full Deep-Research	Yes
`[topic]_factcheck_report.md`	Factoid	Yes
`[topic]_bibliography.json`	All modes	Yes
`[topic]_bibliography.csv`	All modes	Yes
`methods_appendix.md`	Any (only if requested)	No

Report-First Progressive Update Pattern

Create the report file immediately after Phase 0 with all 15 section headers (use template from REPORT_TEMPLATE.md). Then:

After Phase 1 (disambiguation): fill Sections 1-5
After Phase 2 (literature search): fill Sections 6-12
After evidence grading: fill Sections 13-14
Last: write Executive Summary and Section 15 (synthesizes everything)

This ensures partial results are saved even if the process is interrupted.

Report Template

Use the 15-section template from REPORT_TEMPLATE.md. Key sections adapt by domain:

Biological targets: protein architecture, expression, GO terms, disease links, pathogen involvement
Drugs: chemical properties, targets/MOA, pharmacokinetics, indications, safety
Diseases: epidemiology, pathophysiology, associated genes, treatments
General academic: historical context, key theories, empirical evidence, applications

See REPORT_TEMPLATE.md for full template, domain-specific adaptations, bibliography format, theme extraction protocol, and completeness checklist.

Communication

Brief progress updates (not search logs):

"Resolving subject identifiers..."
"Building core paper set..."
"Expanding via citation network..."
"Clustering themes and grading evidence..."

DO NOT expose: raw tool outputs, dedup counts, search round details, database-by-database results.

For factoid queries: ask (once) if user wants just the verified answer or a full report. Default to factoid mode.

References

TOOL_NAMES_REFERENCE.md — Complete list of 123 tools with parameters
REPORT_TEMPLATE.md — Full report template, domain adaptations, bibliography format, theme extraction, completeness checklist
FULLTEXT_STRATEGY.md — Three-tier full-text verification strategy
WORKFLOW.md — Compact workflow cheat-sheet
EXAMPLES.md — Worked examples (ATP6V1A, TRAG collision, sparse target, drug query)

tooluniverse-literature-deep-research