tooluniverse-literature-deep-research

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:

1. Disambiguate first - Resolve IDs, synonyms, naming collisions before literature search
2. Right-size the deliverable - Factoid mode for single questions; full report for deep research
3. Evidence grading - Grade every claim (T1 mechanistic → T4 mention)
4. Mandatory completeness - All sections must exist, even if "unknown/limited evidence"
5. Source attribution - Every claim traceable to database/tool
6. English-first queries - Use English for searches; respond in user's language
7. 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:

1. Subject type: Gene/protein, disease, drug, CS/ML topic, social science, or general?
2. Scope: Single factoid to verify, or comprehensive deep review?
3. Known aliases (if ambiguous): Specific names or symbols in use?
4. 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

1. Identify collision terms from off-topic results
2. 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"):

1. Identify each domain component separately (e.g., CS method + biological application)
2. Resolve bio entities using Phase 1.1-1.3 (targets, drugs, diseases)
3. Search CS/general literature using ArXiv, DBLP, SemanticScholar in parallel
4. Merge results — use both bio tools AND general academic tools in Phase 2
5. 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

---

Content truncated.