Tooluniverse Literature Deep Research

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 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.

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.

---

Phase 3: Evidence Grading

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

---

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 -- 123 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

Literature Deep Research - Examples

Examples demonstrating target disambiguation, evidence grading, and biological model synthesis.

---

Example 1: Full Workflow for ATP6V1A (V-ATPase Target)

User Query

"What does the literature say about ATP6V1A?"

Phase 0: Clarification

Agent: I'll conduct comprehensive research on ATP6V1A. A few questions:

1. **Scope**: Comprehensive overview, or specific focus (druggability, disease, mechanism)?
2. **Known aliases**: Any specific names you use for this target?
3. **Methods appendix**: Do you want search methodology in a separate file?

User: Comprehensive. I know it's part of V-ATPase. No methods file needed.

Phase 1: Target Disambiguation

1.1 Resolve Identifiers

## Target Identity

| Identifier | Value | Source |
|------------|-------|--------|
| Official Symbol | ATP6V1A | HGNC via UniProt |
| UniProt | P38606 | UniProt_search |
| Ensembl Gene | ENSG00000114573 | UniProt_id_mapping |
| NCBI Gene ID | 523 | MyGene_get_gene_annotation |
| ChEMBL Target | CHEMBL2364682 | OpenTargets |

**Full Name**: V-type proton ATPase catalytic subunit A
**Synonyms**: ATP6A1, VPP2, Vma1, VA68, V-ATPase A subunit

1.2 Collision Check

Search: "ATP6V1A"[Title] - first 20 results reviewed.

### Naming Collision Assessment

**Collision risk**: LOW
- Symbol "ATP6V1A" is unambiguous
- Related but distinct: ATP6V0A1-4 (V0 domain vs V1 domain)
- "V-ATPase" is a complex name that may include V0 subunits

**Search strategy**: 
- Use exact symbol for precision: "ATP6V1A"
- For broader V-ATPase context: "V-ATPase" AND (V1 OR catalytic OR "A subunit")
- Exclude pure V0 papers when V1-specific: NOT "V0 domain" NOT "a1 subunit"

1.3 Protein Architecture

### Protein Architecture

| Domain | Position | InterPro ID | Function |
|--------|----------|-------------|----------|
| V-ATPase A, N-terminal | 1-90 | IPR022879 | ATP binding regulatory |
| V-ATPase A, catalytic | 91-490 | IPR005725 | Catalytic activity |
| V-ATPase A, C-terminal | 491-617 | IPR022878 | V1 complex assembly |

**Length**: 617 aa (canonical isoform)
**Isoforms**: 2
- P38606-1 (canonical): 617 aa
- P38606-2: Missing 1-45, 572 aa

**Key Sites**:
- Lys-168: ATP binding (active site)
- Glu-261: Catalytic glutamate

*Sources: InterPro, UniProt*

1.4 Expression Profile

### Baseline Tissue Expression (GTEx v8)

| Tissue | TPM | Rank | Interpretation |
|--------|-----|------|----------------|
| Kidney - Cortex | 145.3 | 1 | Elevated |
| Liver | 98.7 | 2 | Medium-high |
| Brain - Cerebellum | 87.2 | 3 | Medium |
| Lung | 76.4 | 4 | Medium |
| Ubiquitous baseline | ~50 | - | Broad |

**Tissue Specificity**: Low (τ = 0.28) - broadly expressed housekeeping gene
**Implication**: Essential cellular function; targeting may have broad effects

*Source: GTEx_get_median_gene_expression*

1.5 Functional Annotations

### GO Annotations

**Molecular Function**:
- Proton-transporting ATPase activity, rotational mechanism (GO:0046961) [IDA]
- ATP binding (GO:0005524) [IDA]
- ATP hydrolysis activity (GO:0016887) [IDA]

**Biological Process**:
- Lysosomal acidification (GO:0007041) [IMP]
- Autophagy (GO:0006914) [IMP]
- pH regulation (GO:0006885) [IBA]
- Bone resorption (GO:0045453) [IMP]

**Cellular Component**:
- Vacuolar proton-transporting V-type ATPase, V1 domain (GO:0000221) [IDA]
- Lysosomal membrane (GO:0005765) [IDA]

*Source: GO_get_annotations_for_gene, OpenTargets*

Phase 2: Literature Search (Internal - Not Shown)

Query Strategy Applied: 1. High-precision: "ATP6V1A"[Title] → 23 papers (core set) 2. Citation expansion from top 15 → 89 additional papers 3. Collision-filtered broad: "V-ATPase" AND (V1 OR "A subunit") AND (mechanism OR acidification) → 156 papers

Deduplication: 268 raw → 187 unique papers

Phase 3: Report Sections with Evidence Grading

## 6. Core Mechanisms

### 6.1 Molecular Function
ATP6V1A is the catalytic A subunit of the vacuolar H+-ATPase (V-ATPase), 
responsible for ATP hydrolysis that drives proton translocation across membranes 
[★★★ Mechanistic: PMID:12345678, crystal structure + biochemistry].

The V1 sector containing ATP6V1A performs ATP hydrolysis, while the V0 
sector mediates proton transport. Rotation of the central stalk couples 
ATP hydrolysis to proton pumping [★★★: PMID:23456789, cryo-EM structure].

**Evidence Quality**: Strong (8 mechanistic studies with direct biochemical evidence)

### 6.2 Biological Role

**Lysosomal Acidification**
V-ATPase acidifies lysosomes to pH 4.5-5.0, required for:
- Acid hydrolase activity [★★★: PMID:34567890]
- Autophagosome-lysosome fusion [★★☆: PMID:45678901, knockdown phenotype]
- Protein degradation [★★☆: PMID:56789012]

**mTORC1 Signaling**
ATP6V1A participates in lysosomal amino acid sensing:
- V-ATPase interacts with Ragulator complex [★★★: PMID:67890123]
- Required for mTORC1 recruitment to lysosome [★★☆: PMID:78901234]
- Bafilomycin A1 inhibits mTORC1 via V-ATPase [★★★: PMID:89012345]

**Bone Resorption**
Osteoclasts use V-ATPase for extracellular acidification:
- ATP6V1A mutation causes osteopetrosis [★★★: PMID:90123456, patient genetics + functional]
- Osteoclast-specific V-ATPase inhibitors in development [★★☆: PMID:01234567]

**Evidence Quality**: Strong (lysosome), Moderate (mTORC1), Strong (bone)

Theme Extraction Example

## 12. Research Themes

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

ATP6V1A is essential for lysosomal acidification, which is prerequisite for 
autophagy completion. Studies demonstrate that:
- V-ATPase activity required for autophagosome-lysosome fusion [★★★: PMID:xxx]
- Bafilomycin A1 blocks autophagic flux via V-ATPase inhibition [★★★: PMID:xxx]
- ATP6V1A knockdown accumulates LC3-II and p62 [★★☆: PMID:xxx]

**Representative papers** (≥3 required):
1. Forgac M (2007) - V-ATPase structure-function review [★★★: PMID:17428758]
2. Zoncu R et al (2011) - mTORC1 senses amino acids via V-ATPase [★★★: PMID:22153073]
3. Settembre C et al (2013) - TFEB controls lysosomal biogenesis [★★★: PMID:23332759]
4. Abu-Remaileh M et al (2017) - Lysosomal metabolomics [★★★: PMID:28893638]
5. [Additional papers...]

### 12.2 Cancer & Tumor Acidification (28 papers)
**Evidence Quality**: Moderate (8 functional, 15 association, 5 review)

V-ATPase contributes to tumor microenvironment acidification:
- Upregulated in metastatic cancers [★☆☆: PMID:xxx, expression correlation]
- V-ATPase inhibitors show anti-cancer effects [★★☆: PMID:xxx, xenograft study]
- Acidic pH promotes invasion and metastasis [★★☆: PMID:xxx]

**Representative papers**:
1. Neri D & Supuran CT (2011) - Tumor pH and V-ATPase [PMID:21677680]
2. [Additional papers...]

### 12.3 Bone Biology & Osteopetrosis (19 papers)
**Evidence Quality**: Strong (genetic + functional studies)

[Content with evidence grades...]

### 12.4 Viral Infection (12 papers)
**Evidence Quality**: Moderate

Multiple viruses exploit V-ATPase for entry:
- Influenza virus requires acidic endosomes [★★★: PMID:xxx]
- SARS-CoV-2 entry blocked by V-ATPase inhibitors [★★☆: PMID:xxx]
- Flavivirus membrane fusion pH-dependent [★★☆: PMID:xxx]

### 12.5 Neurodegenerative Disease (8 papers)
**Evidence Quality**: Limited (mostly association)

[Content with evidence grades...]

### 12.6 Assays & Tools (15 papers)
**Evidence Quality**: Methodological

[Assay development papers...]

Biological Model & Hypotheses

## 14. Biological Model & Testable Hypotheses

### 14.1 Integrated Biological Model

ATP6V1A is the catalytic engine of the V-ATPase proton pump, essential for 
acidifying lysosomes and other intracellular compartments. Our literature 
synthesis supports the following model:

**Core Function**: ATP6V1A hydrolyzes ATP to power proton translocation, 
maintaining lysosomal pH at 4.5-5.0 [strong evidence from 15+ mechanistic studies].

**Regulatory Hub**: V-ATPase acts as a signaling hub, connecting nutrient 
status to mTORC1 via the Ragulator-Rag pathway. When amino acids are present, 
V-ATPase promotes mTORC1 lysosomal recruitment [moderate evidence, mechanism 
still debated].

**Disease Relevance**:
- Loss-of-function: Osteopetrosis (osteoclast defect), neurodegeneration 
  (impaired autophagy) [strong genetic evidence]
- Gain-of-function/upregulation: Cancer progression, tumor acidification 
  [moderate association evidence]
- Exploitation: Viral entry dependency [strong for influenza, moderate for others]

**Key Uncertainty**: The precise mechanism of V-ATPase-mTORC1 coupling, 
specifically whether V-ATPase ATPase activity vs. scaffolding function 
is required, remains debated [conflicting evidence from 4 studies].

### 14.2 Testable Hypotheses

| # | Hypothesis | Perturbation | Readout | Expected | Priority |
|---|------------|--------------|---------|----------|----------|
| 1 | ATP6V1A is required for autophagy completion | siRNA knockdown in HeLa | LC3-II levels, p62 accumulation, GFP-LC3 puncta | ↑LC3-II, ↑p62, puncta accumulation | HIGH |
| 2 | V-ATPase ATPase activity (not just presence) is required for mTORC1 signaling | Concanamycin A treatment vs knockdown | pS6K levels, mTOR localization | Both reduce pS6K; knockdown more severe if scaffolding matters | HIGH |
| 3 | ATP6V1A inhibition blocks SARS-CoV-2 entry | Bafilomycin A1 pretreatment | Viral RNA, spike protein, infection | ↓ infection in dose-dependent manner | MEDIUM |
| 4 | ATP6V1A expression correlates with cancer metastasis | TCGA pan-cancer analysis | Survival, metastasis-free survival | High ATP6V1A → worse prognosis in epithelial cancers | MEDIUM |
| 5 | V-ATPase inhibitor synergizes with autophagy inducers in cancer | Combination of bafilomycin + rapamycin | Cell death, xenograft growth | Synergistic cytotoxicity | MEDIUM |

### 14.3 Suggested Experiments

**For Hypothesis 1 (Autophagy)**:
- siRNA targeting ATP6V1A in HeLa or MEFs
- Treat ± bafilomycin A1 as positive control
- Western blot for LC3-II, p62
- Confocal microscopy for GFP-LC3 puncta
- Electron microscopy for autophagosome accumulation
- Expected: Knockdown phenocopies bafilomycin

**For Hypothesis 2 (mTORC1 mechanism)**:
- Compare: (a) Concanamycin A (inhibits activity), (b) ATP6V1A siRNA (removes protein)
- Readouts: pS6K, p4E-BP1, mTOR lysosomal localization (IF)
- Include rescue with catalytically-dead ATP6V1A mutant
- Expected: If scaffolding sufficient, siRNA more severe than inhibitor

---

Example 2: Handling Naming Collisions (TRAG)

Query: "What's known about TRAG?"

Collision Detection

### Naming Collision Assessment

**HIGH COLLISION RISK DETECTED**

Search `"TRAG"[Title]` reveals multiple meanings:
1. **T-cell regulatory gene (human TRAF1)** - Immune signaling
2. **Bacterial TraG** - Conjugation/plasmid transfer protein
3. **TraG-like domain** - Bioinformatics/structural studies

**Resolution**:
- User clarification needed: "Is this T-cell TRAG or bacterial TraG?"
- If T-cell: Use TRAF1 as primary symbol
- Apply filter: NOT plasmid NOT conjugation NOT "TraG domain"

After User Clarifies (T-cell immune)

### Resolved Query Strategy

**Target**: TRAF1 (also known as TRAG, EBI6)
**Collision filter**: NOT plasmid NOT conjugation NOT bacterial NOT "type IV secretion"

**High-precision queries**:
1. "TRAF1"[Title] AND (immune OR T-cell OR NF-kB)
2. "TRAF1" AND lymphoma
3. "TRAG" AND T-cell NOT plasmid NOT conjugation

**Citation expansion**: From TRAF1 signaling papers, expand citations

---

Example 3: Sparse Target with Citation-First Strategy

Query: "Literature on GPRC5D" (emerging CAR-T target)

Initial Search Results

"GPRC5D"[Title] → Only 8 papers
"GPRC5D" broader → 23 papers total

Citation-First Expansion

### Citation Network Strategy Applied

**Seed papers** (8 core papers from title search):
- PMID:31919438 - GPRC5D as multiple myeloma target
- PMID:33020169 - CAR-T targeting GPRC5D
- [6 more]

**Forward citations** from seeds:
- PubMed_get_cited_by → 67 additional papers
- EuropePMC_get_citations → 12 unique (not in PubMed)

**Backward citations** from seeds:
- EuropePMC_get_references → 45 additional (mostly myeloma biology)

**Final corpus**: 132 papers (from 8 initial)

**Note**: For this sparse target, citation expansion (124 papers) vastly 
outperformed keyword search (23 papers).

---

Example 4: Evidence Grading in Practice

Raw Claim (Before Grading)

"ATP6V1A is involved in autophagy and cancer."

With Evidence Grading

**Autophagy**:
ATP6V1A is essential for lysosomal acidification, which is required for 
autophagy completion [★★★ Mechanistic: PMID:22153073, biochemical + genetic]. 
Knockdown in cell lines causes LC3-II accumulation [★★☆ Functional: PMID:28893638]. 
Multiple reviews discuss V-ATPase-autophagy connection [★☆☆ Mention: PMIDs:xxx,yyy].

**Cancer**:
ATP6V1A is upregulated in several cancers [★☆☆ Association: TCGA data, PMID:xxx]. 
V-ATPase inhibitors show anti-proliferative effects in xenografts [★★☆ Functional: 
PMID:yyy]. The mechanistic role in cancer progression is hypothesized but not 
directly demonstrated [no ★★★ mechanistic studies identified].

**Summary**:
- Autophagy link: STRONG evidence (mechanistic studies)
- Cancer link: MODERATE evidence (functional + association, no mechanistic)

---

Example 5: Bibliography Output Format

ATP6V1A_bibliography.json

{
  "metadata": {
    "generated": "2026-02-04T15:30:00Z",
    "query": "ATP6V1A",
    "identifiers": {
      "symbol": "ATP6V1A",
      "uniprot": "P38606",
      "ensembl": "ENSG00000114573"
    },
    "total_raw": 268,
    "total_unique": 187,
    "collision_filter_applied": false
  },
  "papers": [
    {
      "pmid": "22153073",
      "doi": "10.1126/science.1207056",
      "title": "mTORC1 senses lysosomal amino acids through an inside-out mechanism that requires the vacuolar H+-ATPase",
      "authors": ["Zoncu R", "Bar-Peled L", "Efeyan A", "Wang S", "Sancak Y", "Sabatini DM"],
      "year": 2011,
      "journal": "Science",
      "source_databases": ["PubMed", "OpenAlex", "Crossref"],
      "evidence_tier": "T1",
      "evidence_tier_rationale": "Mechanistic study with biochemistry and genetics",
      "themes": ["mtorc1_signaling", "lysosomal_function"],
      "in_core_set": true,
      "citation_count": 1847,
      "oa_status": "bronze",
      "oa_url": null
    },
    {
      "pmid": "28893638",
      "doi": "10.1126/science.aan6298",
      "title": "Lysosomal metabolomics reveals V-ATPase- and mTOR-dependent regulation of amino acid efflux from lysosomes",
      "authors": ["Abu-Remaileh M", "Wyant GA", "Kim C", "Laqtom NN", "Abbasi M", "Chan SH", "Freinkman E", "Sabatini DM"],
      "year": 2017,
      "journal": "Science",
      "source_databases": ["PubMed", "EuropePMC"],
      "evidence_tier": "T1",
      "themes": ["lysosomal_function", "metabolism"],
      "in_core_set": true,
      "citation_count": 612,
      "oa_status": "gold",
      "oa_url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5704954/"
    }
  ],
  "theme_summary": {
    "lysosomal_function": {"count": 47, "evidence_quality": "strong"},
    "cancer": {"count": 28, "evidence_quality": "moderate"},
    "bone_biology": {"count": 19, "evidence_quality": "strong"},
    "viral_infection": {"count": 12, "evidence_quality": "moderate"},
    "neurodegenerative": {"count": 8, "evidence_quality": "limited"},
    "methodology": {"count": 15, "evidence_quality": "methodological"}
  }
}

---

Example 6: Drug-Centric Workflow (Metformin Cancer Repurposing)

User Query

"What evidence supports metformin for cancer treatment or prevention?"

Phase 1: Drug Disambiguation

## Drug Identity

| Identifier | Value | Source |
|------------|-------|--------|
| Generic Name | Metformin | User query |
| ChEMBL ID | CHEMBL1431 | OpenTargets_get_drug_chembId_by_generic_name |
| DrugBank ID | DB00331 | drugbank_get_drug_basic_info_by_drug_name_or_id |
| PubChem CID | 4091 | PubChem_get_CID_by_compound_name |
| ATC Code | A10BA02 | ChEMBL_get_drug |

**Class**: Biguanide antidiabetic
**Approval Status**: Approved (FDA 1994)
**Primary Indication**: Type 2 diabetes mellitus
**Known Targets**: AMPK (indirect activation), Complex I (mitochondrial)

Phase 1b: Known Mechanisms & Safety

### Mechanism of Action
- Inhibits mitochondrial Complex I → ↑AMP/ATP ratio → activates AMPK
- Reduces hepatic glucose production [★★★: PMID:24851903]
- Targets: SLC22A1 (OCT1, primary uptake transporter)

*Source: ChEMBL_get_drug_mechanisms, drugbank_get_pharmacology_by_drug_name_or_drugbank_id*

### Safety Profile
- Well-tolerated with >60 years of clinical use
- Adverse events: GI disturbance (common), lactic acidosis (rare)
- No black box warnings for cancer indication

*Source: OpenTargets_get_drug_adverse_events_by_chemblId*

Phase 2: Literature Search Strategy

1. High-precision: "metformin"[Title] AND (cancer OR neoplasm) → 1,847 papers
2. Focused queries:
   - "metformin" AND "AMPK" AND cancer → 312 papers
   - "metformin" AND (clinical trial) AND cancer → 187 papers
   - "metformin" AND ("breast cancer" OR "colorectal cancer") → 445 papers
3. Citation expansion from top 20 most-cited → 234 additional unique

Phase 3: Report Excerpt with Evidence Grading

## Core Evidence for Anti-Cancer Activity

### Epidemiological Evidence
Retrospective studies consistently show reduced cancer incidence in metformin-treated
diabetic patients vs other antidiabetics [★☆☆ Association: PMID:16168273, large cohort].
A meta-analysis of 265 studies confirms ~30% risk reduction [★☆☆: PMID:24925038].

### Mechanistic Evidence
Metformin activates AMPK, inhibiting mTORC1 and reducing cell proliferation
[★★★ Mechanistic: PMID:16489903, direct biochemical evidence].
It also reduces insulin/IGF-1 levels systemically [★★☆ Functional: PMID:22893504].

### Clinical Trial Evidence
Multiple ongoing trials (NCT01101438, NCT02581137) testing metformin as adjuvant
in breast, colorectal, and pancreatic cancer [★★☆: active clinical data].
No phase III RCT has yet demonstrated overall survival benefit as primary endpoint
[gap identified].

**Summary**:
- Epidemiological: MODERATE (consistent association, confounding risk)
- Mechanistic: STRONG (AMPK/mTOR pathway well-characterized)
- Clinical: EMERGING (trials ongoing, no definitive phase III result)

---

Example 7: General Academic Topic (CS/ML - No Bio Tools)

User Query

"What does the literature say about retrieval-augmented generation (RAG)?"

Phase 0: Clarification

Agent: I'll conduct a deep literature review on RAG. Quick clarification:

1. **Scope**: Comprehensive overview, or focused on a specific aspect
   (architectures, evaluation, domain-specific applications)?
2. **Date range**: Include foundational work, or focus on recent (2023+)?

User: Comprehensive overview including foundational work.

Phase 1: Subject Disambiguation (General Academic)

No bio tools used. Proceed with literature-only disambiguation.

## Subject Identity & Scope

| Attribute | Value |
|-----------|-------|
| Full Name | Retrieval-Augmented Generation |
| Abbreviation | RAG |
| Domain | Natural Language Processing / Information Retrieval |
| Foundational Paper | Lewis et al. (2020), "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" |

**Related but distinct**:
- RAG (biology): Recombination Activating Gene (immunology)
- Collision filter applied: NOT "recombination activating" NOT "V(D)J" NOT "immunoglobulin"

**Key sub-topics**: dense retrieval, knowledge grounding, hallucination reduction,
chunking strategies, vector databases, evaluation frameworks

Phase 2: Literature Search Strategy

Tools used (no bio tools):
- ArXiv_search_papers(query="retrieval augmented generation", sort_by="submittedDate")
- SemanticScholar_search_papers(query="RAG retrieval augmented generation", sort="citationCount:desc")
- DBLP_search_publications(query="retrieval augmented generation")
- openalex_literature_search(search_keywords="retrieval augmented generation")

Collision filter: NOT "recombination activating" NOT "V(D)J"

1. High-precision: "retrieval-augmented generation"[Title] → 342 papers
2. Citation expansion from Lewis et al. 2020 (3,800+ citations) → top 100
3. Broader: "RAG" AND ("language model" OR "knowledge grounding") NOT "V(D)J" → 189 papers

Phase 3: Report Excerpt

## 6. Core Mechanisms

### 6.1 Architecture
RAG combines a parametric language model with a non-parametric retrieval component.
The original architecture uses DPR for retrieval and BART for generation
[★★★: arXiv:2005.11401, Lewis et al. 2020].

### 6.2 Retrieval Strategies
- Dense Passage Retrieval (DPR) [★★★: arXiv:2004.04906, Karpukhin et al.]
- ColBERT late interaction [★★★: arXiv:2004.12832, Khattab & Zaharia]
- Hybrid sparse+dense retrieval [★★☆: multiple studies]

### 6.3 Key Finding
RAG reduces hallucination by grounding generation in retrieved evidence
[★★☆: arXiv:2305.14627, multiple evaluation studies]. However, retrieval
noise can propagate errors [★★☆: arXiv:2401.15391].

## 12. Research Themes

### 12.1 Architectures & Training (87 papers)
**Evidence Quality**: Strong
[Dense retrieval, cross-encoders, late interaction, end-to-end training...]

### 12.2 Evaluation & Benchmarks (45 papers)
**Evidence Quality**: Moderate
[RAGAS, faithfulness metrics, attribution, citation generation...]

### 12.3 Domain-Specific RAG (62 papers)
**Evidence Quality**: Moderate
[Medical QA, legal, code generation, scientific literature...]

### 12.4 Chunking & Indexing (28 papers)
**Evidence Quality**: Limited (emerging area)
[Semantic chunking, hierarchical retrieval, multi-vector...]

## 14. Integrated Model & Hypotheses

| # | Hypothesis | Test | Readout | Expected | Priority |
|---|------------|------|---------|----------|----------|
| 1 | Iterative RAG outperforms single-pass | Multi-hop QA benchmark | Accuracy, faithfulness | 10-15% improvement | HIGH |
| 2 | Semantic chunking > fixed-size | Domain-specific QA | Answer quality, retrieval precision | Semantic chunking wins on complex docs | HIGH |
| 3 | RAG + fine-tuning > RAG alone | Specialized domain (medical) | F1, hallucination rate | Combined approach optimal | MEDIUM |

---

Example 8: Cross-Domain Query (GNN for Drug Discovery)

User Query

"What does the literature say about graph neural networks for drug discovery?"

Phase 1: Interdisciplinary Disambiguation

Identify domain components:

• CS/ML: Graph Neural Networks (GNN, GCN, GAT, MPNN)
• Bio/Drug: Drug discovery (molecular property prediction, virtual screening, ADMET)

CS component (no bio tools):

| Attribute | Value |
|-----------|-------|
| Full Name | Graph Neural Networks |
| Abbreviation | GNN |
| Sub-architectures | GCN, GAT, MPNN, SchNet, DimeNet |
| Foundational Paper | Kipf & Welling (2017), "Semi-Supervised Classification with GCNs" |

Bio component (resolve drug discovery entities):

OpenTargets_get_drug_chembId_by_generic_name → Not applicable (method, not drug)
Skip drug/target disambiguation — this is about a method applied to the domain

Collision check: "GNN" has low collision risk (not overloaded across domains). No negative filter needed.

Phase 2: Hybrid Literature Search

CS tools:
- ArXiv_search_papers(query='graph neural network drug discovery', sort_by='submittedDate')
- SemanticScholar_search_papers(query='GNN molecular property prediction', sort='citationCount:desc')
- DBLP_search_publications(query='graph neural network drug')

Bio tools:
- PubMed_search_articles(query='"graph neural network" AND ("drug discovery" OR "molecular property")')
- EuropePMC_search_articles(query='"GNN" AND "drug discovery"')

Bridging:
- openalex_literature_search(search_keywords='graph neural network drug discovery')

Phase 3: Report Excerpt

## 6. Core Mechanisms

### 6.1 Molecular Graph Representations
Molecules are naturally represented as graphs where atoms are nodes and bonds
are edges. GNNs learn molecular representations through message-passing
[★★★: arXiv:1704.01212, Gilmer et al. 2017, MPNN framework].

### 6.2 Key Architectures
- **MPNN** (Message Passing Neural Network): Unified framework [★★★: arXiv:1704.01212]
- **SchNet**: Continuous-filter convolutions for 3D [★★★: arXiv:1706.08566]
- **AttentiveFP**: Graph attention for molecular properties [★★☆: PMID:31408336]

## 12. Research Themes

### 12.1 Molecular Property Prediction (89 papers)
**Evidence Quality**: Strong
ADMET, solubility, toxicity prediction using GNN encoders...

### 12.2 Virtual Screening & Drug-Target Interaction (52 papers)
**Evidence Quality**: Moderate
DTI prediction, binding affinity estimation...

### 12.3 De Novo Molecular Generation (38 papers)
**Evidence Quality**: Moderate
Generative models (VAE, flow-based) on molecular graphs...

### 12.4 Retrosynthesis & Reaction Prediction (21 papers)
**Evidence Quality**: Limited (emerging)

## 14. Integrated Model & Hypotheses

| # | Hypothesis | Test | Readout | Expected | Priority |
|---|------------|------|---------|----------|----------|
| 1 | 3D-aware GNNs outperform 2D for binding affinity | PDBbind benchmark | RMSE, Pearson r | 3D models win by 10-15% | HIGH |
| 2 | Pre-training on large molecular datasets improves low-data ADMET | Fine-tune on ToxCast | AUC, F1 | Pre-trained > random init | HIGH |
| 3 | GNN + physics-based scoring > either alone | CASF-2016 docking | Success rate, enrichment | Hybrid approach optimal | MEDIUM |

---

Key Principles Demonstrated

1. Disambiguation first - Resolve IDs/scope, check collisions before searching 2. Evidence grading everywhere - T1-T4 labels on all claims 3. Citation-first for sparse targets - Expand from seeds when keywords fail 4. Collision-aware queries - Apply NOT filters for ambiguous names 5. Mandatory sections - Include all 15 sections, even if "limited evidence" 6. Integrated model - Synthesize evidence into testable hypotheses 7. Scalable bibliography - Narrative in report, full data in JSON 8. Domain adaptation - Bio tools for bio queries, CS tools for CS queries, general tools for everything else

Full-Text Verification Strategy

Use when abstracts lack critical details (exact values, cell lines, concentrations, protocols, benchmark numbers, hyperparameters, dataset sizes).

---

Table of Contents

1. Tier 1: Auto-Snippet (Europe PMC) 2. Tier 2: Manual Two-Step 3. Tier 3: Manual Download 4. Decision Matrix 5. Best Practices

---

Tier 1: Auto-Snippet (Europe PMC) - FASTEST

Use for: Exploratory queries with 3-5 specific terms.

EuropePMC_search_articles(
    query="bacterial antibiotic resistance evolution",
    limit=10,
    extract_terms_from_fulltext=["ciprofloxacin", "meropenem", "A. baumannii", "MIC"]
)
→ Returns articles with fulltext_snippets[].term and fulltext_snippets[].snippet

• Single tool call (search + snippets)
• Bounded latency (max 3 OA articles, ~3-5 seconds)
• Terms processed in batches of 5 internally
• Only works for OA articles with fullTextXML (~30-40% coverage)

---

Tier 2: Manual Two-Step - TARGETED

Use for: Specific high-value papers identified from search.

Europe PMC Full-Text (broadest OA coverage)

EuropePMC_get_fulltext_snippets(
    article_id="PMC1234567",
    terms=["ADAR1", "MDA5", "interferon"],
    window_chars=300
)
→ Returns snippets from specific PMC article

EuropePMC_get_fulltext(article_id="PMC1234567")
→ Returns full-text XML

Semantic Scholar PDF

SemanticScholar_get_pdf_snippets(
    open_access_pdf_url="<url from search results>",
    terms=["SHAP", "gradient attribution"],
    window_chars=300
)
→ First search with SemanticScholar_search_papers, then use open_access_pdf_url from results

ArXiv (100% OA)

ArXiv_get_pdf_snippets(
    arxiv_id="2301.12345",
    terms=["attention mechanism", "self-attention", "layer normalization"],
    max_snippets_per_term=5
)
→ Works for any arXiv paper (100% coverage)

---

Tier 3: Manual Download - FALLBACK

Use for: Paywalled content via institutional access (last resort).

get_webpage_text_from_url(url="https://doi.org/10.1016/...")
→ Returns full page text (quality varies by publisher)

• Requires institutional access
• No snippet extraction (full HTML)
• Quality varies by publisher

---

Decision Matrix

---

Best Practices

1. Limit search terms to 3-5 specific keywords:

• Bio: ["ciprofloxacin 5 ug/mL", "HEK293 cells", "RNA-seq"]
• CS/ML: ["BLEU score", "F1 macro", "learning rate 3e-5"]
• Bad: ["drug", "method", "significant"]

2. Check OA status before extraction: Use isOpenAccess field from EuropePMC or open_access_pdf_url from SemanticScholar.

3. Adjust window size for context:

• Methods: 400-500 chars
• Quick verification: 150-200 chars
• Default: 220 chars

4. Handle failures gracefully: fall back to abstract or skip.

5. Document full-text sources in report:

   ## Methods Verification

   **Key detail** (verified from full text):
   - Study A: Value X [PMC12345, Methods section]
   - Study B: Value Y [arXiv:2301.12345, Experimental Design]

   *Full-text verification performed on 8/15 OA papers (53% coverage)*

Literature Deep Research - Quick Reference

Enhanced strategy with target disambiguation, evidence grading, and mandatory completeness.

---

Workflow Summary

1. CLARIFY → Target type? Scope? Aliases? Methods appendix needed?
2. DISAMBIGUATE → Resolve IDs, find collisions, gather baseline profile
3. SEARCH → High-precision seeds → Citation expansion → Collision-filtered broad
4. GRADE → Apply evidence tiers (T1-T4) to all claims
5. REPORT → Mandatory sections, biological model, testable hypotheses

---

Phase 1: Target Disambiguation (Default ON)

Tools to Use

Collision Detection

1. Search: "[SYMBOL]"[Title] in PubMed (first 20 results)
2. If >20% off-topic → identify collision terms
3. Build filter: NOT [collision1] NOT [collision2]

---

Phase 2: Query Strategy

Step 1: High-Precision Seeds

"[GENE_SYMBOL]"[Title] AND (mechanism OR function OR structure)
"[FULL_PROTEIN_NAME]"[Title]

Step 2: Citation Expansion (especially for sparse targets)

PubMed_get_cited_by(pmid) → Forward citations
EuropePMC_get_citations(pmid) → Fallback for forward
PubMed_get_related(pmid) → Related papers
EuropePMC_get_references(pmid) → Backward citations

Step 3: Collision-Filtered Broad

"[GENE]" AND ([pathway] OR [function]) NOT [collision_term]

---

Evidence Grading (Apply to ALL Claims)

In report:

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

---

Mandatory Report Sections (ALL Required)

---

Theme Extraction

Standard themes (adapt to target):

• Core function / Mechanism
• Disease relevance
• Signaling / Pathways
• Trafficking / Localization
• Genetics / Variants
• Infection / Immunity
• Methodology / Tools

---

Output Files

1. `[topic]_report.md` - Main narrative (DEFAULT) 2. `[topic]_bibliography.json` - Full deduplicated papers (ALWAYS) 3. `methods_appendix.md` - Methodology (ONLY if requested)

---

Tool Failure Handling

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

---

OA Handling

With Unpaywall email: Full OA check for all DOIs Without: Best-effort using:

• Europe PMC isOpenAccess field
• PMC papers (all OA)
• OpenAlex is_oa field

Label: *OA Status: Best-effort*

---

Completeness Checklist

Before delivery, verify ALL:

Identity

• [ ] UniProt, Ensembl, NCBI, ChEMBL IDs
• [ ] All synonyms documented
• [ ] Collisions handled

Biology

• [ ] Domains + architecture
• [ ] Localization
• [ ] Expression profile
• [ ] GO terms + pathways
• [ ] Interactors

Mechanism

• [ ] Core function (with evidence grades)
• [ ] Model organism data (or "none")
• [ ] Assays described

Disease

• [ ] Constraint scores + interpretation
• [ ] Variants (ClinVar, gnomAD)
• [ ] Disease links (graded by evidence)
• [ ] Pathogens (or "none")

Synthesis

• [ ] Themes with ≥3 papers each
• [ ] Open questions listed
• [ ] Biological model written
• [ ] ≥3 testable hypotheses
• [ ] Conclusions + confidence

Technical

• [ ] All claims have sources
• [ ] Evidence grades applied
• [ ] Bibliography file created
• [ ] Limitations noted

---

DO vs DON'T

---

Communication Style

Brief progress updates:

• "Resolving target identifiers..."
• "Building core paper set..."
• "Grading evidence and clustering themes..."

Never expose:

• Raw tool outputs
• Deduplication stats
• Round-by-round search details
• Database failure logs

Report = Deliverable (not search log)

Report Templates & Output Formats

Templates for deep research reports, bibliography files, and completeness checklists.

---

Table of Contents

1. Full Deep-Research Report Template 2. Domain-Specific Adaptations 3. Bibliography Format 4. Theme Extraction Protocol 5. Completeness Checklist

---

Full Deep-Research Report Template

Use for Full Deep-Research Mode only. For factoid mode, use the short fact-check template in SKILL.md.

# [TARGET/TOPIC]: Comprehensive Research Report

*Generated: [Date]*
*Evidence cutoff: [Date]*
*Total unique papers: [N]*

---

## Executive Summary

[2-3 paragraphs synthesizing key findings across all sections]

**Bottom Line**: [One-sentence actionable conclusion]

---

## 1. Subject Identity & Scope
*[MANDATORY - clarify what is being researched]*

### 1.1 Official Identifiers
[Table of IDs, database entries, or scope definition]

### 1.2 Synonyms and Aliases
[All known names - critical for complete literature coverage]

### 1.3 Known Naming Collisions
[Document collisions and how they were handled]

---

## 2. Background & Context
*[MANDATORY - domain-specific background]*

For biological targets: protein architecture, domains, isoforms, key sites
For drugs: chemical structure, properties, drug class, formulation
For diseases: epidemiology, classification, current understanding
For general topics: historical context, scope, key definitions

---

## 3. Key Entities & Relationships
*[MANDATORY]*

For biological targets: complexes, interaction partners, scores
For drugs: known targets, mechanisms of action, binding sites
For diseases: associated genes, pathways, risk factors
For general topics: key actors, institutions, competing frameworks

---

## 4. Spatial/Temporal Context
*[MANDATORY]*

For biological targets: subcellular localization with confidence
For drugs: pharmacokinetics (ADME), tissue distribution
For diseases: affected tissues/organs, disease progression timeline
For general topics: geographic/temporal scope, key periods

---

## 5. Quantitative Profile
*[MANDATORY]*

For biological targets: expression profile (top tissues, specificity)
For drugs: clinical pharmacology (dosing, bioavailability, DDIs)
For diseases: prevalence, incidence, demographic patterns
For general topics: key metrics, benchmarks, trends over time

---

## 6. Core Mechanisms / Central Arguments
*[MANDATORY - heart of the report]*

### 6.1 Primary Function / Main Thesis
[Central finding with evidence grades]
**Evidence Quality**: [Strong/Moderate/Limited]

### 6.2 Supporting Evidence
[Role in broader context, pathways, causal chains]

### 6.3 Key Processes / Pathways
[Involvement in larger systems with evidence grades]

### 6.4 Regulation / Modulation
[How the subject is regulated, controlled, or influenced]

---

## 7. Experimental / Empirical Evidence
*[MANDATORY]*

### 7.1 Direct Evidence
[Controlled experiments, clinical trials, primary studies]

### 7.2 Model Systems
[Animal models, simulations, analogues]

### 7.3 Cross-Validation
[Conservation, replication, meta-analyses]

---

## 8. Variation & Heterogeneity
*[MANDATORY]*

For biological targets: human genetic variants, constraint scores, ClinVar, GWAS
For drugs: pharmacogenomics, responder vs non-responder profiles
For diseases: subtypes, genetic basis, population differences
For general topics: competing schools of thought, regional variation

---

## 9. Applied / Translational Relevance
*[MANDATORY - include evidence strength]*

### 9.1 Strong Evidence
[Claims with causal/mechanistic support]

### 9.2 Moderate Evidence
[Claims with functional/associative support]

### 9.3 Weak Evidence
[Claims with correlation/mention only]

### 9.4 Evidence Summary Table

| Claim/Link | Evidence Type | Key Papers | Grade |
|------------|---------------|------------|-------|
| [Claim 1] | Mechanistic | PMID:xxx | ★★★ |
| [Claim 2] | Association | PMID:yyy | ★★☆ |

---

## 10. External Factors
*[MANDATORY - state "None identified" if N/A]*

For biological targets: pathogen interactions, environmental factors
For drugs: drug-drug interactions, food effects, contraindications
For diseases: environmental triggers, comorbidities
For general topics: external influences, confounders, policy implications

---

## 11. Methods & Assays / Methodological Landscape
*[MANDATORY]*

### 11.1 Standard Approaches
[Established methods for studying this subject]

### 11.2 Emerging Methods
[New techniques, tools, technologies]

### 11.3 Key Resources
[Datasets, repositories, reference standards]

---

## 12. Research Themes
*[MANDATORY - structured theme extraction]*

### 12.1 [Theme 1 Name] (N papers)
**Evidence Quality**: [Strong/Moderate/Limited]
**Representative Papers**: [≥3 papers or state "insufficient"]

[Theme description with evidence-graded citations]

### 12.2 [Theme 2 Name] (N papers)
[Same structure]

[Continue for all themes - require ≥3 papers per theme, or state "limited evidence"]

---

## 13. Open Questions & Research Gaps
*[MANDATORY]*

### 13.1 Fundamental Unknowns
[What we don't understand]

### 13.2 Practical/Translational Unknowns
[What we don't know for applications]

### 13.3 Suggested Priority Questions
[Ranked list of important unanswered questions]

---

## 14. Integrated Model & Testable Hypotheses
*[MANDATORY - synthesis section]*

### 14.1 Integrated Model
[3-5 paragraph synthesis integrating all evidence into coherent model]

### 14.2 Testable Hypotheses

| # | Hypothesis | Test/Perturbation | Readout | Expected Result | Priority |
|---|------------|-------------------|---------|-----------------|----------|
| 1 | [Hypothesis] | [Experiment] | [Measure] | [Prediction] | HIGH |
| 2 | [Hypothesis] | [Experiment] | [Measure] | [Prediction] | HIGH |
| 3 | [Hypothesis] | [Experiment] | [Measure] | [Prediction] | MEDIUM |

### 14.3 Suggested Next Studies
[Brief description of key experiments/studies to test hypotheses]

---

## 15. Conclusions & Recommendations
*[MANDATORY]*

### 15.1 Key Takeaways
[Bullet points of most important findings]

### 15.2 Confidence Assessment
[Overall confidence: High/Medium/Low with justification]

### 15.3 Recommended Next Steps
[Prioritized action items]

---

## References

*[Summary reference list - full bibliography in separate file]*

### Key Papers (Must-Read)
1. [Citation with PMID/DOI] - [Why important] [Grade: ★★★]
2. ...

### By Theme
[Organized reference lists]

---

## Data Limitations

- [Any databases that failed or returned no data]
- [Any known gaps in coverage]
- [OA status method used]

*Full methodology available in methods_appendix.md upon request.*

---

Domain-Specific Adaptations

Biological Target Reports

Map generic sections to biology-specific content:

Drug Reports

Disease Reports

General Academic Topics (CS, Social Science, Humanities, etc.)

---

Bibliography Format

JSON ([topic]_bibliography.json)

{
  "metadata": {
    "generated": "2026-03-07",
    "query": "[TOPIC]",
    "identifiers": {},
    "total_raw": 268,
    "total_unique": 187
  },
  "papers": [
    {
      "pmid": "12345678",
      "doi": "10.1038/xxx",
      "arxiv_id": null,
      "dblp_key": null,
      "title": "Paper Title",
      "authors": ["Smith A", "Jones B"],
      "year": 2024,
      "journal": "Nature",
      "source_databases": ["PubMed", "OpenAlex"],
      "evidence_tier": "T1",
      "themes": ["theme_a", "theme_b"],
      "oa_status": "gold",
      "oa_url": "https://...",
      "citation_count": 45,
      "in_core_set": true
    }
  ]
}

Also generate [topic]_bibliography.csv with same data in tabular format.

---

Theme Extraction Protocol

Clustering Process

1. Extract keywords from titles and abstracts 2. Cluster into themes using semantic similarity 3. Require minimum N papers per theme (default N=3) 4. Label themes with standardized descriptive names

Theme Quality Requirements

Standard Theme Categories (adapt to domain)

Biological: Core function, Disease relevance, Signaling/Pathways, Trafficking/Localization, Genetics/Variants, Infection/Immunity, Methodology/Tools

Drug: Mechanism, Efficacy, Safety/Toxicity, Pharmacokinetics, Resistance, Repurposing, Clinical trials, Biomarkers

General academic: Theoretical foundations, Empirical studies, Methodology, Applications, Ethics/Policy, Historical development

---

Completeness Checklist

Verify ALL boxes before delivery. Mark "N/A" or "Limited evidence" where appropriate.

Identity & Context

• [ ] Official identifiers or scope definition resolved
• [ ] All synonyms/aliases documented
• [ ] Naming collisions identified and handled
• [ ] Background/context described (or N/A stated)

Core Content

• [ ] Core mechanisms/arguments with evidence grades
• [ ] Supporting evidence documented
• [ ] Experimental/empirical evidence (or "none found")
• [ ] Key entities and relationships listed
• [ ] Methods/assays described

Variation & Application

• [ ] Variation/heterogeneity documented
• [ ] Applied relevance with evidence strength grades
• [ ] External factors (or "none identified")

Synthesis

• [ ] Research themes clustered with ≥3 papers each (or noted as limited)
• [ ] Open questions/gaps articulated
• [ ] Integrated model synthesized
• [ ] ≥3 testable hypotheses with experiments
• [ ] Conclusions with confidence assessment

Technical

• [ ] All claims have source attribution
• [ ] Evidence grades applied throughout
• [ ] Bibliography file generated (JSON + CSV)
• [ ] Data limitations documented
• [ ] Code/data availability noted (for CS/ML topics with reproducibility expectations)

Literature Deep Research - Complete Tool Reference

All tool names for literature search + annotation databases used in enhanced research strategy.

---

Literature Search Tools

Biomedical & Life Sciences

'PubMed_search_articles'          # 35M+ papers - primary biomedical
'PMC_search_papers'               # Full-text biomedical archive (NIH)
'EuropePMC_search_articles'       # 42M+ European biomedical (use source='PPR' for preprints)
'BioRxiv_get_preprint'            # Get bioRxiv preprint by DOI (not keyword search)
'BioRxiv_list_recent_preprints'   # List recent preprints by subject area
'MedRxiv_get_preprint'            # Get medRxiv preprint by DOI (not keyword search)
# For preprint keyword search, use: EuropePMC_search_articles(source='PPR')

Computer Science

'DBLP_search_publications'        # 6M+ CS publications
'ArXiv_search_papers'             # Physics/CS/math preprints
'SemanticScholar_search_papers'   # 200M+ AI-ranked papers

Physics / High-Energy Physics

'InspireHEP_search_papers'        # 1.6M+ particle/astro physics (a:/t:/j:/cn: syntax)

General Academic

'openalex_literature_search'      # 250M+ works across fields
'Crossref_search_works'           # 140M+ DOI registry (use filter:"type:journal-article")
'Fatcat_search_scholar'           # Internet Archive
'DOAJ_search_articles'            # Open access journals
'CORE_search_papers'              # 200M+ aggregated open access papers
'openalex_search_works'           # OpenAlex with sort/filter support

Multi-Source Deep Search

'advanced_literature_search_agent' # Deep multi-source search agent (slower, thorough)

Regional/Specialized

'OpenAIRE_search_publications'    # EU-funded research (best-effort: API has been returning 400s; verify before relying on it)
'HAL_search_archive'              # French national archive (max_results required)
'OSF_search_preprints'            # Social science preprints
'Zenodo_search_records'           # Datasets, software, publications
'Figshare_search_articles'        # Datasets, figures, code, posters, theses (query param is `search_for`)

Reliability: CORE_search_papers needs CORE_API_KEY (HTTP 429 without it); OpenAIRE_search_publications has been returning HTTP 400 on basic queries — treat both as best-effort, not sole sources.

Text Mining / NER

'PubTator3_LiteratureSearch'      # PubTator literature search with entity annotations
'PubTator3_EntityAutocomplete'    # Entity name autocomplete (genes, diseases, chemicals)
'PubTator3_get_annotations'       # Get entity annotations for a PMID/PMCID

---

Citation & Metadata Tools (19)

PubMed Citation Tools

'PubMed_get_article'              # Complete metadata for PMID
'PubMed_get_cited_by'             # Forward citations (papers citing this)
'PubMed_get_related'              # Computationally related papers
'PubMed_get_links'                # External links (full-text)

Europe PMC Citation Tools (Fallbacks)

'EuropePMC_get_citations'         # Forward citations (fallback for PubMed)
'EuropePMC_get_references'        # Backward citations (reference list)
'EuropePMC_get_fulltext'          # Full-text XML for open access articles
'EuropePMC_get_fulltext_snippets' # Full-text snippet search within articles

Open Citation Tools

'OpenCitations_get_citations'     # Open citation data by DOI
'OpenCitations_get_references'    # Reference list by DOI
'OpenCitations_get_citation_count' # Citation count by DOI

Semantic Scholar

'SemanticScholar_get_paper'       # Paper metadata by DOI/PMID/S2 ID
'SemanticScholar_get_recommendations' # AI-recommended similar papers
'SemanticScholar_get_pdf_snippets' # PDF text extraction by DOI/PMID

Citation Impact & Metrics

'iCite_search_publications'       # PubMed search + iCite citation metrics (RCR, APT, NIH percentile)
'iCite_get_publications'          # iCite metrics by PMID (citation_count, RCR, APT)
'scite_get_tallies'               # Smart citation tallies (supporting/contradicting/mentioning)

Full-Text / PDF Tools

'ArXiv_get_pdf_snippets'          # ArXiv PDF text extraction

Open Access

'Unpaywall_check_oa_status'       # OA status by DOI (requires email)

---

Protein/Gene Annotation Tools

UniProt Tools

'UniProt_search'                  # Search UniProt by query
'UniProt_get_entry_by_accession'  # Full protein entry
'UniProt_id_mapping'              # Map between ID types
'UniProt_get_function_by_accession'  # Function description
'UniProt_get_sequence_by_accession'  # Protein sequence
'UniProt_get_recommended_name_by_accession'  # Official name
'UniProt_get_alternative_names_by_accession'  # Aliases
'UniProt_get_subcellular_location_by_accession'  # Localization
'UniProt_get_ptm_processing_by_accession'  # PTMs, active sites
'UniProt_get_disease_variants_by_accession'  # Disease variants

Domain/Structure Tools

'InterPro_get_protein_domains'    # Domain architecture
'alphafold_get_prediction'        # AlphaFold structure
'get_protein_metadata_by_pdb_id'  # PDB structure metadata
'proteins_api_get_protein'        # Alternative protein data

Gene Annotation

'MyGene_get_gene_annotation'      # NCBI gene info, aliases
'ensembl_lookup_gene'             # Ensembl gene details
'ensembl_get_xrefs'               # Cross-references

---

Expression Tools

GTEx (Tissue Expression)

'GTEx_get_gene_expression'        # Expression data
'GTEx_get_median_gene_expression' # Median TPM by tissue

Human Protein Atlas

'HPA_get_comprehensive_gene_details_by_ensembl_id'  # Full HPA data
'HPA_get_rna_expression_by_source'  # RNA expression
'HPA_get_subcellular_location'    # Subcellular localization
'HPA_get_protein_interactions_by_gene'  # DEPRECATED: returns error, use STRING/BioGRID instead
'HPA_get_cancer_prognostics_by_gene'  # Cancer prognosis

Single-Cell (if available)

'CELLxGENE_get_expression_data'   # Single-cell expression

---

Pathway & Function Tools

Gene Ontology

'GO_get_annotations_for_gene'     # GO terms (BP, MF, CC)
'GO_get_term_details'             # GO term details

Pathways

'Reactome_map_uniprot_to_pathways'  # Reactome pathways
'kegg_get_gene_info'              # KEGG gene/pathways
'WikiPathways_search'             # WikiPathways

Open Targets (Target-centric)

'OpenTargets_get_target_gene_ontology_by_ensemblID'  # GO via OT
'OpenTargets_get_target_tractability_by_ensemblID'   # Druggability
'OpenTargets_get_associated_drugs_by_target_ensemblID'  # Known drugs
'OpenTargets_get_target_safety_profile_by_ensemblID'   # Safety
'OpenTargets_get_diseases_phenotypes_by_target_ensembl'  # Disease links
'OpenTargets_get_publications_by_target_ensemblID'   # Publications

---

Interaction Tools (6)

'STRING_get_protein_interactions'  # STRING PPI network
'intact_get_interactions'         # IntAct experimental PPIs
'intact_get_complex_details'      # Complex membership
'OpenTargets_get_target_interactions_by_ensemblID'  # OT interactions
'BioGRID_get_interactions'        # BioGRID protein-protein interactions
'BioGRID_get_chemical_interactions'  # BioGRID chemical-protein interactions

---

Variant & Disease Tools

gnomAD

'gnomad_get_gene'                 # Population variants
'gnomad_get_gene_constraints'     # Constraint scores (pLI, LOEUF)

ClinVar

'ClinVar_search_variants'         # Clinical variants
'ClinVar_get_variant_details'     # Variant details

Disease

'OpenTargets_get_diseases_phenotypes_by_target_ensembl'  # Disease associations
'OpenTargets_get_target_constraint_info_by_ensemblID'  # Gene constraint info
'OpenTargets_get_biological_mouse_models_by_ensemblID'  # Mouse model phenotypes
'DGIdb_get_drug_gene_interactions'  # Drug-gene
'DGIdb_get_gene_druggability'     # Druggability categories
'CTD_get_chemical_gene_interactions'  # CTD chemical-gene interactions
'CTD_get_gene_diseases'           # CTD gene-disease associations
'CTD_get_disease_chemicals'       # Chemicals linked to a disease (CTD)
'CTD_get_chemical_diseases'       # Diseases linked to a chemical (CTD)
'DisGeNET_get_disease_genes'      # Gene-disease associations (DisGeNET score)
'DisGeNET_search_disease'         # Search diseases by name

---

Drug/Chemical Tools (24)

ChEMBL

'ChEMBL_search_drugs'            # Search drugs by name/synonym
'ChEMBL_get_drug'                # Get drug details by ChEMBL ID
'ChEMBL_get_drug_mechanisms'     # Drug mechanism of action
'ChEMBL_search_mechanisms'       # Search mechanisms by target/drug
'ChEMBL_search_targets'          # Search drug targets
'ChEMBL_get_target_activities'   # Bioactivity data for target
'ChEMBL_get_molecule_targets'    # Targets of a molecule by ChEMBL ID

DrugBank

'drugbank_get_drug_basic_info_by_drug_name_or_id'  # Basic drug info
'drugbank_get_indications_by_drug_name_or_drugbank_id'  # Drug indications
'drugbank_get_pharmacology_by_drug_name_or_drugbank_id'  # Pharmacology
'drugbank_get_drug_interactions_by_drug_name_or_id'  # Drug-drug interactions
'drugbank_get_targets_by_drug_name_or_drugbank_id'  # Drug targets

PubChem

'PubChem_get_CID_by_compound_name'  # Resolve compound name to CID
'PubChem_get_compound_properties_by_CID'  # Compound properties
'PubChem_get_compound_synonyms_by_CID'  # Compound synonyms

Open Targets (Drug-centric)

'OpenTargets_get_drug_chembId_by_generic_name'  # Resolve drug name to ChEMBL ID
'OpenTargets_get_drug_id_description_by_name'  # Drug description
'OpenTargets_get_drug_adverse_events_by_chemblId'  # Adverse events
'OpenTargets_get_drug_warnings_by_chemblId'  # Drug warnings
'OpenTargets_get_drug_mechanisms_of_action_by_chemblId'  # Mechanisms of action
'OpenTargets_get_associated_targets_by_drug_chemblId'  # Drug-target associations
'OpenTargets_get_associated_diseases_by_drug_chemblId'  # Drug-disease associations
'OpenTargets_get_drug_indications_by_chemblId'  # Approved indications

Clinical Trials

'search_clinical_trials'          # ClinicalTrials.gov search

---

Utility Tools (1)

'get_webpage_text_from_url'       # Fetch and extract text from a URL

---

Tool Categories by Use Case

Target Disambiguation (Phase 1)

# Resolve IDs
'UniProt_search'
'UniProt_id_mapping'
'ensembl_lookup_gene'

# Get baseline profile
'InterPro_get_protein_domains'
'HPA_get_subcellular_location'
'GTEx_get_median_gene_expression'
'GO_get_annotations_for_gene'
'Reactome_map_uniprot_to_pathways'

High-Precision Literature Seeds

'PubMed_search_articles'          # "[GENE]"[Title] queries
'EuropePMC_search_articles'       # Alternative

Citation Network Expansion

'PubMed_get_cited_by'             # Forward citations (primary)
'EuropePMC_get_citations'         # Forward (fallback)
'OpenCitations_get_citations'     # Open citation data by DOI
'OpenCitations_get_references'    # Backward citations by DOI
'PubMed_get_related'              # Related papers
'SemanticScholar_get_recommendations' # AI-recommended similar papers
'EuropePMC_get_references'        # Backward citations

Citation Impact Assessment

'iCite_search_publications'       # Search + citation metrics (RCR, APT)
'iCite_get_publications'          # Metrics by PMID
'scite_get_tallies'               # Supporting/contradicting/mentioning counts
'OpenCitations_get_citation_count' # Open citation count by DOI

Broad Search

'openalex_literature_search'      # Comprehensive
'Crossref_search_works'           # DOI-based (filter:"type:journal-article")
'SemanticScholar_search_papers'   # AI-ranked
'CORE_search_papers'              # Aggregated open access
'advanced_literature_search_agent' # Deep multi-source (slower, thorough)

OA Status Check

'Unpaywall_check_oa_status'       # If email provided
# Otherwise use OA fields from:
# - EuropePMC (isOpenAccess)
# - OpenAlex (is_oa)
# - PMC papers (all OA)

---

Fallback Chains

Note: HPA_get_protein_interactions_by_gene is DEPRECATED. Use STRING_get_protein_interactions or BioGRID_get_interactions instead.

---

Parameter Quick Reference

Literature Search

# PubMed
{'query': '"GENE"[Title]', 'limit': 100}

# ArXiv
{'query': 'ti:"transformer" AND abs:"attention mechanism"', 'limit': 50, 'sort_by': 'submittedDate'}

# DBLP (count param `limit` is REQUIRED, not `max_results`)
{'query': 'graph neural network', 'limit': 50}

# SemanticScholar
{'query': 'retrieval augmented generation', 'limit': 50, 'year': '2023-2024', 'sort': 'citationCount:desc'}

# OpenAlex (query alias `search`; count alias `per_page`)
{'search': 'term', 'per_page': 100, 'filter': 'from_publication_date:2020-01-01'}

# InspireHEP (query param is `q`, count is `size`)
{'q': 'higgs boson', 'size': 25, 'sort': 'mostcited'}

# Figshare (query param is `search_for`, not `query`)
{'search_for': 'crispr screen dataset', 'page_size': 20}

# Required count param (query alone is rejected): HAL/OpenAIRE/Fatcat/PubMed_Guidelines
{'query': 'term', 'max_results': 20}          # HAL, Fatcat
{'query': 'term', 'max_results': 20, 'type': 'publications'}  # OpenAIRE also needs type
{'query': 'hypertension', 'limit': 20}        # PubMed_Guidelines_Search

# With collision filter
{'query': '"TRAG" AND immune NOT plasmid NOT conjugation', 'limit': 50}

Citation Tools

{'pmid': '12345678', 'limit': 100}  # PubMed citations
{'article_id': 'MED:12345678', 'source': 'MED'}  # EuropePMC

Annotation Tools

# UniProt
{'accession': 'P38606'}
{'ids': ['P38606'], 'from_db': 'UniProtKB_AC-ID', 'to_db': 'Ensembl'}

# GTEx (gencode_id is an Ensembl gene ID; dataset_id defaults to gtex_v8)
{'gencode_id': 'ENSG00000114573', 'dataset_id': 'gtex_v8'}

# InterPro
{'uniprot_accession': 'P38606'}

---

Complete Alphabetical Tool List

Literature & Text Mining (24)

1. advanced_literature_search_agent (deep multi-source) 2. ArXiv_search_papers 3. ArXiv_get_pdf_snippets (PDF text extraction) 4. BioRxiv_get_preprint (DOI retrieval) 5. BioRxiv_list_recent_preprints (recent by subject) 6. CORE_search_papers (aggregated open access) 7. Crossref_search_works 8. DBLP_search_publications 9. DOAJ_search_articles 10. EuropePMC_search_articles 11. Fatcat_search_scholar 12. HAL_search_archive 13. MedRxiv_get_preprint (DOI retrieval) 14. OpenAIRE_search_publications 15. openalex_literature_search 16. openalex_search_works 17. OSF_search_preprints 18. PMC_search_papers 19. PubMed_search_articles 20. PubTator3_LiteratureSearch (literature search with entity annotations) 21. PubTator3_EntityAutocomplete (entity name autocomplete) 22. PubTator3_get_annotations (entity annotations for PMID/PMCID) 23. SemanticScholar_search_papers 24. Zenodo_search_records

Citation & Impact (19)

25. EuropePMC_get_citations 26. EuropePMC_get_references 27. EuropePMC_get_fulltext (full-text XML) 28. EuropePMC_get_fulltext_snippets (full-text snippet search) 29. iCite_get_publications (metrics by PMID) 30. iCite_search_publications (search + metrics) 31. OpenCitations_get_citations 32. OpenCitations_get_references 33. OpenCitations_get_citation_count 34. PubMed_get_article 35. PubMed_get_cited_by 36. PubMed_get_links 37. PubMed_get_related 38. scite_get_tallies 39. SemanticScholar_get_paper (paper metadata) 40. SemanticScholar_get_pdf_snippets (PDF text extraction) 41. SemanticScholar_get_recommendations 42. Unpaywall_check_oa_status

Protein/Gene (17)

43. alphafold_get_prediction 44. ensembl_get_xrefs 45. ensembl_lookup_gene 46. get_protein_metadata_by_pdb_id 47. InterPro_get_protein_domains 48. MyGene_get_gene_annotation 49. proteins_api_get_protein 50. UniProt_get_alternative_names_by_accession 51. UniProt_get_disease_variants_by_accession 52. UniProt_get_entry_by_accession 53. UniProt_get_function_by_accession 54. UniProt_get_ptm_processing_by_accession 55. UniProt_get_recommended_name_by_accession 56. UniProt_get_sequence_by_accession 57. UniProt_get_subcellular_location_by_accession 58. UniProt_id_mapping 59. UniProt_search

Expression (6)

60. CELLxGENE_get_expression_data 61. GTEx_get_gene_expression 62. GTEx_get_median_gene_expression 63. HPA_get_comprehensive_gene_details_by_ensembl_id 64. HPA_get_rna_expression_by_source 65. HPA_get_subcellular_location

Pathway/Function (9)

66. GO_get_annotations_for_gene 67. GO_get_term_details 68. kegg_get_gene_info 69. OpenTargets_get_associated_drugs_by_target_ensemblID 70. OpenTargets_get_publications_by_target_ensemblID 71. OpenTargets_get_target_gene_ontology_by_ensemblID 72. OpenTargets_get_target_safety_profile_by_ensemblID 73. OpenTargets_get_target_tractability_by_ensemblID 74. Reactome_map_uniprot_to_pathways 75. WikiPathways_search

Interaction (6)

76. BioGRID_get_interactions 77. BioGRID_get_chemical_interactions 78. intact_get_complex_details 79. intact_get_interactions 80. OpenTargets_get_target_interactions_by_ensemblID 81. STRING_get_protein_interactions

Variant/Disease (17, 1 deprecated)

82. ClinVar_get_variant_details 83. ClinVar_search_variants 84. CTD_get_chemical_diseases 85. CTD_get_chemical_gene_interactions 86. CTD_get_disease_chemicals 87. CTD_get_gene_diseases 88. DGIdb_get_drug_gene_interactions 89. DGIdb_get_gene_druggability 90. DisGeNET_get_disease_genes 91. DisGeNET_search_disease 92. gnomad_get_gene 93. gnomad_get_gene_constraints 94. HPA_get_cancer_prognostics_by_gene 95. HPA_get_protein_interactions_by_gene # DEPRECATED: use STRING/BioGRID 96. OpenTargets_get_biological_mouse_models_by_ensemblID 97. OpenTargets_get_diseases_phenotypes_by_target_ensembl 98. OpenTargets_get_target_constraint_info_by_ensemblID

Drug/Chemical (24)

99. ChEMBL_get_molecule_targets 100. ChEMBL_search_drugs 101. ChEMBL_get_drug 102. ChEMBL_get_drug_mechanisms 103. ChEMBL_search_mechanisms 104. ChEMBL_search_targets 105. ChEMBL_get_target_activities 106. drugbank_get_drug_basic_info_by_drug_name_or_id 107. drugbank_get_indications_by_drug_name_or_drugbank_id 108. drugbank_get_pharmacology_by_drug_name_or_drugbank_id 109. drugbank_get_drug_interactions_by_drug_name_or_id 110. drugbank_get_targets_by_drug_name_or_drugbank_id 111. OpenTargets_get_drug_chembId_by_generic_name 112. OpenTargets_get_drug_id_description_by_name 113. OpenTargets_get_drug_adverse_events_by_chemblId 114. OpenTargets_get_drug_warnings_by_chemblId 115. OpenTargets_get_drug_mechanisms_of_action_by_chemblId 116. OpenTargets_get_associated_targets_by_drug_chemblId 117. OpenTargets_get_associated_diseases_by_drug_chemblId 118. OpenTargets_get_drug_indications_by_chemblId 119. PubChem_get_CID_by_compound_name 120. PubChem_get_compound_properties_by_CID 121. PubChem_get_compound_synonyms_by_CID 122. search_clinical_trials

Utility (1)

123. get_webpage_text_from_url

---

Total Tools: 123 (122 active, 1 deprecated: HPA_get_protein_interactions_by_gene)

• Literature & Text Mining: 24
• Citation & Impact: 19
• Protein/Gene Annotation: 17
• Expression: 6
• Pathway/Function: 9
• Interaction: 6
• Variant/Disease: 17 (1 deprecated)
• Drug/Chemical: 24 (incl. ChEMBL target tools, clinical trials)
• Utility: 1

Last Updated: 2026-03-07

Literature Deep Research - Workflow Cheat-Sheet

---

Workflow Summary

1. CLARIFY → Subject type? Scope? Aliases? Domain?
2. DISAMBIGUATE → Resolve IDs, find collisions, gather baseline profile
3. SEARCH → High-precision seeds → Citation expansion → Collision-filtered broad
4. GRADE → Apply evidence tiers (T1-T4) to all claims
5. REPORT → 15-section template, integrated model, testable hypotheses

---

Phase 1: Subject Disambiguation

Biological Target Tools

Drug-Centric Tools

General Academic (No Bio Tools)

Skip bio annotation tools. Use domain-appropriate literature databases directly:

• CS/ML: ArXiv_search_papers, DBLP_search_publications, SemanticScholar_search_papers
• Social Science: OSF_search_preprints, openalex_literature_search
• General: openalex_literature_search, Crossref_search_works, CORE_search_papers

Interdisciplinary / Cross-Domain

For topics spanning multiple domains (e.g., "GNNs for drug discovery"): 1. Identify each domain component separately 2. Resolve bio entities using Phase 1 bio tools 3. Search CS/general literature using ArXiv, DBLP, SemanticScholar 4. Merge results from both bio and general tools in Phase 2 5. Cross-reference to find bridging papers

Collision Detection

1. Search primary database (first 20 results)
2. If >20% off-topic → identify collision terms
3. Build filter: NOT [collision1] NOT [collision2]

Gene family: "ADAR" NOT "ADAR2" NOT "ADARB1" for ADAR1-specific. Cross-domain: "RAG" AND "language model" NOT "recombination activating".

---

Phase 2: Query Strategy

Step 1: High-Precision Seeds

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

Step 2: Citation Expansion

PubMed_get_cited_by(pmid) → Forward citations
EuropePMC_get_citations(pmid) → Fallback for forward
OpenCitations_get_citations(doi) → Open citation data
PubMed_get_related(pmid) → Related papers
SemanticScholar_get_recommendations(pmid) → AI-recommended
EuropePMC_get_references(pmid) → Backward citations

Step 2b: Citation Impact (optional)

iCite_search_publications(query) → Search + RCR/APT/NIH percentile (PubMed-only)
iCite_get_publications(pmids) → Metrics by PMID (PubMed-only)
scite_get_tallies(doi) → Supporting/contradicting/mentioning counts (PubMed-only)
SemanticScholar_get_paper(paper_id) → Citation counts for any paper (CS/ML included)

Step 3: Collision-Filtered Broad

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

---

Evidence Grading

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

---

Report Sections (ALL Required)

---

Theme Extraction

---

Output Files

1. `[topic]_report.md` - Main narrative (DEFAULT) 2. `[topic]_bibliography.json` - Full deduplicated papers (ALWAYS) 3. `[topic]_bibliography.csv` - Tabular bibliography (ALWAYS) 4. `[topic]_factcheck_report.md` - Factoid/verification mode output 5. `methods_appendix.md` - Methodology (ONLY if requested)

---

Tool Failure Handling

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

---

DO vs DON'T